Nothing Special   »   [go: up one dir, main page]

WO2021051596A1 - 模型预测优化方法、装置、设备及可读存储介质 - Google Patents

模型预测优化方法、装置、设备及可读存储介质 Download PDF

Info

Publication number
WO2021051596A1
WO2021051596A1 PCT/CN2019/118263 CN2019118263W WO2021051596A1 WO 2021051596 A1 WO2021051596 A1 WO 2021051596A1 CN 2019118263 W CN2019118263 W CN 2019118263W WO 2021051596 A1 WO2021051596 A1 WO 2021051596A1
Authority
WO
WIPO (PCT)
Prior art keywords
data
neural network
deep neural
prediction
prediction result
Prior art date
Application number
PCT/CN2019/118263
Other languages
English (en)
French (fr)
Inventor
王健宗
Original Assignee
平安科技(深圳)有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 平安科技(深圳)有限公司 filed Critical 平安科技(深圳)有限公司
Publication of WO2021051596A1 publication Critical patent/WO2021051596A1/zh

Links

Images

Classifications

    • HELECTRICITY
    • H03ELECTRONIC CIRCUITRY
    • H03MCODING; DECODING; CODE CONVERSION IN GENERAL
    • H03M13/00Coding, decoding or code conversion, for error detection or error correction; Coding theory basic assumptions; Coding bounds; Error probability evaluation methods; Channel models; Simulation or testing of codes
    • H03M13/03Error detection or forward error correction by redundancy in data representation, i.e. code words containing more digits than the source words
    • H03M13/05Error detection or forward error correction by redundancy in data representation, i.e. code words containing more digits than the source words using block codes, i.e. a predetermined number of check bits joined to a predetermined number of information bits
    • H03M13/11Error detection or forward error correction by redundancy in data representation, i.e. code words containing more digits than the source words using block codes, i.e. a predetermined number of check bits joined to a predetermined number of information bits using multiple parity bits
    • H03M13/1102Codes on graphs and decoding on graphs, e.g. low-density parity check [LDPC] codes
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/04Architecture, e.g. interconnection topology
    • G06N3/045Combinations of networks

Definitions

  • This application relates to the field of artificial intelligence technology, and in particular to a method, device, device, and readable storage medium for model prediction and optimization.
  • the existing communication system generally includes a sending end and a receiving end, and the receiving end needs to perform data post-processing on the signal from the sending end.
  • the post-processing process it is necessary to perform error correction processing on the signal.
  • error correction codes for error correction processing, such as repeated accumulation of RA codes and polygonal LDPC codes.
  • the inventor realized that the RA code and the polygonal LDPC code have a relatively high word error rate, and under a very low signal-to-noise ratio, the RA code has poor efficiency, resulting in lower transmission performance of the communication system. How to optimize the prediction model to improve the prediction effect of the data to be transmitted by the model is an urgent problem to be solved by those skilled in the art.
  • the main purpose of this application is to solve the technical problem in the prior art that the model has a low accuracy rate for predicting the result of the data to be transmitted.
  • the first aspect of the present application provides a model prediction optimization method, including: inputting first training sample data into a deep neural network set pre-deployed in the machine learning framework tensorflow, after passing the first training sample
  • the neurons in each deep neural network are randomly hidden according to the preset hiding ratio through a random algorithm, and the deep neural network after the hidden neurons is obtained and the respective depths are output
  • the first prediction result of the neural network the first prediction result output by the deep neural network after the hidden neurons is normalized by a normalization algorithm to obtain a prediction result set, and the prediction result set includes multiple A first prediction result; calculate the accuracy of each first prediction result in the prediction result set according to the first actual result, and set the weight for the deep neural network behind each hidden neuron based on the accuracy to obtain the first prediction Model; predict the first data to be predicted by the first prediction model to obtain a second prediction result; determine whether the second prediction result meets the preset accuracy rate according to the second actual result;
  • the weights of each deep neural network are adjusted through the back propagation algorithm until the second prediction result Satisfy the preset accuracy rate; predict the second data to be predicted by the second prediction model to obtain the third prediction result; determine whether the amount of error data in the third prediction result is greater than that of the preset error correction code for the error data Processing amount; if the amount of error data in the current data is greater than the amount of error data processed by the preset error correction code, the current data is classified through the support vector machine model pre-deployed in the machine learning framework tensorflow to obtain the error data And the correct data and the correct data is transmitted to the data receiving end, if the amount of error data in the current data is less than or equal to the processing amount of the error data by the preset error correction code, the third prediction result is transmitted to the data receiving end end.
  • the second aspect of the present application provides a model prediction optimization device, including: a training output module for inputting first training sample data into a deep neural network set pre-deployed in the machine learning framework tensorflow, after passing the first training
  • a training output module for inputting first training sample data into a deep neural network set pre-deployed in the machine learning framework tensorflow, after passing the first training
  • the neurons in each deep neural network are randomly hidden according to the preset hiding ratio through a random algorithm to obtain the deep neural network after the hidden neurons and output the various The first prediction result of the deep neural network
  • a normalization calculation module for normalizing the first prediction result output by the deep neural network after the hidden neuron through a normalization algorithm to obtain a prediction result set,
  • the prediction result set includes a plurality of first prediction results;
  • the calculation module is configured to calculate the accuracy of each first prediction result in the prediction result set according to the first actual result, and provide each hidden nerve based on the accuracy
  • the post-metastatic deep neural network sets the weight
  • the backpropagation algorithm is used to adjust the weight of each deep neural network until the second prediction result meets the preset accuracy rate; the second prediction module is used to pass the second The prediction model predicts the second data to be predicted to obtain the third prediction result; the second judgment module is used to judge whether the amount of error data in the third prediction result is greater than the processing amount of the error data by the preset error correction code; The module is used to classify the current data through the support vector machine model pre-deployed in the machine learning framework tensorflow if the amount of error data in the current data is greater than the amount of error data processed by the preset error correction code to obtain the error The data and correct data and the correct data are transmitted to the data receiving end. If the amount of error data in the current data is less than or equal to the processing amount of the error data by the preset error correction code, the third prediction result is transmitted to the data Receiving end.
  • the third aspect of the present application provides a model prediction optimization device, including: a memory, a processor, and a model prediction optimization program stored in the memory and capable of running on the processor, and the model prediction optimization program is When the processor is executed, the steps of the model prediction optimization method as described in any one of the above are implemented.
  • a fourth aspect of the present application provides a computer-readable storage medium that stores instructions in the computer-readable storage medium, and when the instructions run on a computer, the computer executes the method described in the first aspect.
  • This application predicts the post-processing part of the communication signal through a predictive model, which solves the problem of low efficiency in negotiating error correction.
  • the prediction result shows that the amount of erroneous data has exceeded the error correction capability of the error correction code, then The data is classified, the wrong data and the correct data are classified, and the correct data is transmitted to the data receiving end, which improves the accuracy of data transmission, optimizes the training process of the model, and improves the prediction effect of the data to be transmitted by the model.
  • FIG. 1 is a schematic structural diagram of a model prediction and optimization equipment operating environment involved in a solution of an embodiment of this application;
  • FIG. 2 is a schematic flowchart of a first embodiment of a model prediction optimization method according to this application;
  • FIG. 3 is a schematic flowchart of a second embodiment of a model prediction optimization method according to this application.
  • FIG. 4 is a schematic flowchart of a third embodiment of a model prediction optimization method according to this application.
  • FIG. 5 is a detailed flowchart of step S180 in FIG. 4;
  • FIG. 6 is a schematic flowchart of a fourth embodiment of a model prediction optimization method according to this application.
  • FIG. 7 is a detailed flowchart of step S210 in FIG. 6;
  • FIG. 8 is a schematic diagram of functional modules of the first embodiment of the model prediction and optimization device according to the present application.
  • FIG. 9 is a schematic diagram of functional modules of a second embodiment of a model prediction and optimization device according to the present application.
  • the embodiments of the application provide a model prediction optimization method, device, equipment, and readable storage medium.
  • the post-processing part of the communication signal is predicted through the prediction model, which well solves the problem of low efficiency in negotiation error correction.
  • the prediction result shows that the amount of error data has exceeded the error correction capability of the error correction code, the data is classified, the error data and the correct data are classified, and the correct data is transmitted to the data receiving end, which improves the accuracy of data transmission Optimized the training process of the model to improve the prediction effect of the model to be transmitted.
  • This application provides a model prediction optimization device.
  • FIG. 1 is a schematic structural diagram of the operating environment of the model prediction and optimization equipment involved in the solution of the embodiment of the application.
  • the model prediction optimization device includes: a processor 1001, such as a CPU, a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005.
  • the communication bus 1002 is used to implement connection and communication between these components.
  • the user interface 1003 may include a display screen (Display) and an input unit such as a keyboard (Keyboard), and the network interface 1004 may optionally include a standard wired interface and a wireless interface (such as a WI-FI interface).
  • the memory 1005 may be a high-speed RAM memory, or a non-volatile memory (non-volatile memory), such as a magnetic disk memory.
  • the memory 1005 may also be a storage device independent of the aforementioned processor 1001.
  • model prediction optimization device does not constitute a limitation on the model prediction optimization device, and may include more or less components than those shown in the figure, or a combination of certain components, Or different component arrangements.
  • the memory 1005 which is a computer-readable storage medium, may include an operating system, a network communication module, a user interface module, and a model prediction optimization program.
  • the operating system is a program that manages and controls model prediction and optimization equipment and software resources, and supports the operation of model prediction and optimization programs and other software and/or programs.
  • the network interface 1004 is mainly used to access the network; the user interface 1003 is mainly used to detect and confirm instructions and edit instructions.
  • the processor 1001 may be used to call the model prediction optimization program stored in the memory 1005, and execute the operations of the following various embodiments of the model prediction optimization method.
  • FIG. 2 is a schematic flowchart of a first embodiment of a model prediction optimization method according to the present application.
  • the model prediction optimization method includes the following steps:
  • Step S10 Input the first training sample data into the deep neural network set pre-deployed in the machine learning framework tensorflow, and when training each deep neural network in the deep neural network set through the first training sample data, a random algorithm is used According to a preset hiding ratio, randomly hide neurons in each deep neural network, obtain a deep neural network after the hidden neurons, and output the first prediction result of each deep neural network;
  • Step S20 Perform normalization processing on the first prediction result output by the deep neural network behind the hidden neurons by a normalization algorithm to obtain a prediction result set, the prediction result set including a plurality of first prediction results ;
  • the first prediction result output by the deep neural network after the hidden neurons is normalized by a normalization algorithm to obtain a prediction result set.
  • the prediction result set includes a plurality of first prediction results.
  • the prediction result because the neurons in the deep neural network are randomly hidden according to the preset ratio, so the prediction results output by each deep neural network may be different, because the output prediction results are not exactly the same. Therefore, the occurrence of over-fitting can be greatly reduced, and the accuracy of the prediction result can be improved.
  • the method of processing training samples by multiple deep neural networks improves the speed of model training.
  • the value obtained is positive 0.8 and negative 0.2
  • the value obtained by the second deep neural network is positive 0.6 and negative 0.4
  • the value obtained by the third deep neural network is 0.3 positive and 0.7 negative.
  • Step S30 calculating the accuracy of each first prediction result in the prediction result set according to the first actual result, and setting weights for the deep neural network behind each hidden neuron based on the accuracy, to obtain a first prediction model;
  • the first actual result is preset, and the accuracy of each first prediction result in the prediction result set is calculated according to the first actual result.
  • the first actual result refers to the data that is known to be masked.
  • the verification data of what data is used to verify the accuracy. For example, the "apple” in “I like to eat apples” is preset as the wrong data, and the "apple” is covered up, but people know what is being covered up is “ “Apple”, here "Apple” is the verification data. Since the deep neural network after hiding the neurons does not know that the hidden data is "Apple”, after the prediction is made by the deep neural network after hiding the neurons, it will Obtain the prediction result.
  • the prediction result here may be "apple” or “banana”. By calculating the ratio of the number of occurrences of "apple” in the prediction result to all the prediction results, the accuracy can be obtained.
  • the single neural network model considered is limited in its ability to predict data. Therefore, the combination of multiple deep neural network algorithms solves the limitations of a single deep neural network algorithm for data processing and predicts data.
  • the purpose is to check whether the amount of erroneous data in the current data has exceeded the number of error correction codes that can be corrected. If so, it means that the information receiver has received the wrong information. In order to ensure the safe transmission of information, it is necessary to correct The error data in the data is predicted in real time.
  • Step S40 Predict the first to-be-predicted data through the first prediction model to obtain a second prediction result
  • the first to-be-predicted data is predicted by the first prediction model to obtain the second prediction result
  • the first training sample data is input into the deep neural network set pre-deployed in the machine learning framework tensorflow.
  • the neurons in each deep neural network are randomly hidden according to a preset hiding ratio through a random algorithm to obtain a deep neural network after the hidden neurons
  • output the first prediction results of each of the deep neural networks and normalize the first prediction results output by the deep neural network after the hidden neurons through a normalization algorithm to obtain a set of prediction results.
  • An actual result calculates the accuracy of each first prediction result in the prediction result set, and based on the accuracy, sets the weight for the deep neural network behind each hidden neuron to obtain the first prediction model.
  • the neural network passes After training, the ability to predict is already available, so the first to-be-predicted data can be predicted by the first prediction model to obtain the second prediction result.
  • Step S50 judging whether the second prediction result meets the preset accuracy rate according to the second actual result
  • the second actual result is preset, which is the verification data of the data that is known to be covered up.
  • the second prediction result refers to the prediction of the data to be predicted by the first prediction model, Obtained text data that is pre-masked.
  • the pre-masked data is "evaluation”
  • the pre-masked data must be known to check whether the second prediction result output by the first prediction model is "evaluation”.
  • the second actual result it is judged whether the second prediction result meets the preset accuracy rate.
  • Step S60 if the second prediction result does not meet the preset accuracy rate, return to step S10, and adjust the weights occupied by each deep neural network through the back propagation algorithm until the second prediction result meets the preset accuracy rate;
  • a second prediction model refers to the first prediction model that is fully trained, and can perform normal prediction on the prediction data to be predicted. If the first prediction result does not meet the preset accuracy rate, it means that the weight of the deep neural network that outputs the wrong prediction result is too high. Therefore, it is necessary to adjust the weight of each deep neural network through the backpropagation algorithm until the value is reached. The second prediction result meets the preset accuracy rate.
  • Step S70 If the second prediction result meets the preset accuracy rate, a second prediction model is obtained;
  • the second prediction result meets the preset accuracy rate, it means that the prediction result output by the first prediction model already meets the preset accuracy rate, for example, 98%, then the second prediction model is obtained.
  • Step S80 Predict the second to-be-predicted data through the second prediction model to obtain a third prediction result
  • the second prediction model outputs a vector with four rows and one column.
  • the fourth column of y 2 has the largest element, then y
  • the category is 4, that is, the data corresponding to the element in the fourth column is obtained.
  • Step S90 judging whether the amount of error data in the third prediction result is greater than the amount of error data processed by the preset error correction code
  • the second prediction result it is judged according to the second prediction result whether the amount of error data in the current data is greater than the processing amount of the preset error correction code for the error data.
  • the prediction result of the first data is the error data, and the first data is incorrect.
  • the prediction result of the two data is wrong data until the prediction result of the hundredth data is also wrong data.
  • the error correction code can only process 99 wrong data, which means whether the amount of wrong data in the current data is greater than Preset error correction code to deal with erroneous data.
  • Step S100 if the amount of error data in the current data is greater than the amount of error data processed by the preset error correction code, the current data is classified by the support vector machine model pre-deployed in the machine learning framework tensorflow to obtain the error data And correct data and transmit the correct data to the data receiving end;
  • the current data is classified through the support vector machine model pre-deployed in the machine learning framework tensorflow to obtain The wrong data and the correct data and the correct data are transmitted to the data receiving end.
  • Step S110 If the amount of error data in the current data is less than or equal to the processing amount of the error data by the preset error correction code, the third prediction result is transmitted to the data receiving end.
  • the prediction model is used to predict the post-processing part of the communication signal, which solves the problem of low efficiency in negotiating error correction.
  • the prediction result shows that the amount of erroneous data has exceeded the error correction of the error correction code Ability, the data is classified, the wrong data and the correct data are classified, and the correct data is transmitted to the data receiving end, which improves the accuracy of data transmission.
  • Fig. 3 is a schematic flowchart of a second embodiment of a model prediction optimization method according to the present application. In this embodiment, before step S10 in FIG. 2, the following steps are further included:
  • Step S120 according to a preset strategy of randomly hiding neurons, randomly hiding neurons of the same deep neural network through a random algorithm
  • the neurons of the same deep neural network are randomly hidden through a random algorithm.
  • the strategy of randomly hiding neurons refers to randomly hiding neurons of the deep neural network.
  • the law can be randomly hiding 10% of neurons, randomly hiding 20% of neurons, randomly hiding 30% of neurons to randomly hiding 100% of neurons.
  • Step S130 using the second training sample data to train the deep neural network after the hidden neurons to obtain an initial deep neural network
  • the second training sample data is used to train the deep neural network after the hidden neurons to obtain the initial deep neural network.
  • the first neural network randomly hides 10% of the neurons
  • the second neural network randomly hides 20% of the neurons
  • the third neural network randomly hides 30% of the neurons
  • uses the same training sample to train the deep neural network after the hidden neurons to obtain three initial Deep neural network.
  • Step S140 Predict the third data to be predicted through the initial deep neural network to obtain an initial prediction result set, where the initial prediction result set includes a plurality of initial prediction results;
  • the third data to be predicted is predicted by the initial deep neural network to obtain an initial prediction result set
  • the initial prediction result set includes a plurality of initial prediction results, according to a preset strategy for randomly hiding neurons , Hide the neurons of the same deep neural network randomly through a random algorithm, and then use the second training sample data to train the deep neural network after the hidden neurons to obtain the initial deep neural network, in order to check whether the initial deep neural network has certain For example, 97% of the prediction accuracy, the third to-be-predicted data needs to be predicted by the initial deep neural network to obtain an initial prediction result set, and the initial prediction result set includes multiple initial prediction results.
  • step S150 the initial prediction result with the highest accuracy rate is selected from the initial prediction result set according to the preset artificial prediction result, and the preset hidden ratio among hidden neurons is determined based on the initial prediction result with the highest accuracy rate.
  • a strategy for randomly hiding neurons needs to be set in advance.
  • the first training sample data is used to train 10% of neurons randomly hidden by a random algorithm, and after training for a thousand times, the 10% hidden neurons are then hidden.
  • the deep neural network behind the neurons predicts the second sample data and obtains the first prediction result; and then uses the first training sample data to train the deep neural network that hides 20% of the neurons randomly through a random algorithm, and trains one
  • the second sample data is predicted, and the second prediction result is obtained until the deep neural network that hides 90% of the neurons is trained through the first training sample data.
  • the second sample The data is predicted, and the ninth prediction result is obtained.
  • the correct rate of the nine prediction results is tested, and the prediction result with the largest correct rate is selected, so that the optimal proportion of hidden neurons can be determined.
  • FIG. 4 is a schematic flowchart of a third embodiment of a model prediction optimization method according to the present application.
  • step S40 in FIG. 2 the following steps are further included:
  • Step S160 receiving an optical signal, and performing a time-domain sampling operation on the optical signal to obtain multiple frames of sub-optical signals
  • a dynamic polarization controller can perform linear birefringence on the photons in the channel, change the phase difference of the incident optical signal, and realize the polarization state conversion. Then the polarized optical signal is split by a polarization beam splitter to form two signal lights with the same properties and adjustable light intensity. The two optical signals are phase modulated to form two in-phase and orthogonal beams.
  • Step S170 converting the sub-optical signal into a digital signal
  • the sub-optical signal is further processed by the analog-to-digital converter to convert the sub-optical signal into a digital signal that can carry photon information. This step completes the conversion from analog signal to digital signal.
  • Step S180 Perform standardization processing on the digital signal to obtain first data to be predicted.
  • a frame of digital signal includes 100 data: X 1 , X 2 , X 3 ??X 100 , calculate the average value M of these 100 data, and find the The maximum value max and minimum value min are then substituted into the formula (X i -M)/(max-min) to calculate 100 discrete data and compress the 100 discrete data to the range of [-1,1] as the first A data to be predicted. This step can improve the prediction accuracy.
  • step S180 includes the following steps:
  • Step S1801 Calculate the average value M of the n data included in the digital signal, and find the maximum value max and the minimum value min in the n data, where each data is denoted as X i , i is the value of different data Logo
  • the average value of the digital signal includes M n data of n data and find the maximum value max and minimum value min, where each data denoted by X i, i for different The identity of the data.
  • step S1802 n discrete data are calculated by the formula (X i -M)/(max-min);
  • n discrete data are calculated by the formula (X i -M)/(max-min).
  • Step S1803 Compress the n discrete data to obtain the first data to be predicted.
  • a frame of digital signal includes 100 data: X 1 , X 2 , X 3 ??X 100 , calculate the average value M of these 100 data, and find the The maximum value max and minimum value min are then substituted into the formula (X i -M)/(max-min) to calculate 100 discrete data and compress the 100 discrete data to the range of [-1,1] as the first A data to be predicted. This step can improve the prediction accuracy.
  • Fig. 6 is a schematic flowchart of a fourth embodiment of a model prediction optimization method according to the present application. In this embodiment, before step S10 in FIG. 2, the following steps are further included:
  • Step S190 Obtain an optical signal used for training, and perform a time-domain sampling operation on the optical signal used for training to obtain multiple frames of sub-optical signals used for training;
  • an optical signal used for training is acquired, and a time-domain sampling operation is performed on the optical signal used for training to obtain multiple frames of sub-optical signals used for training.
  • Step S200 converting the multiple frames of sub-optical signals used for training into multiple frames of digital signals used for training
  • Step S210 labeling the multiple frames of digital signals used for training to obtain labeling information of the digital signals used for training in each frame;
  • each frame of data is sequentially mapped to the N-order quadrature amplitude modulation constellation according to the transformation.
  • label the nearest constellation point as the output label.
  • the number of corresponding classification results required is N.
  • Step S220 Perform standardization processing on the multiple frames of digital signals used for training to obtain multiple frames of standard data
  • the training is to be carried out by the deep neural network
  • the data is standardized before the data is input.
  • a frame of digital signal used for training includes 100 data: X 1 , X 2 , X 3 ??X 100 , calculate the average value M of these 100 data, and find the maximum value max and minimum value min in these 100 data, and then substitute the formula:
  • Step S230 based on the label information, divide the multi-frame standard data into multiple groups, wherein the standard data in each group has the same label information;
  • the label information is divided into four types: the first type, the second type, the third type, and the fourth type.
  • the label information the standard data of multiple frames can be divided into four groups, and the label information of the standard data in each group is the same.
  • Step S240 Obtain standard data of the same order of magnitude from each group to form a training sample.
  • standard data of the same order of magnitude (for example, 50) are obtained from each group to form training data, so that data characteristics can be better learned to achieve a more accurate model training effect.
  • step S210 includes the following steps:
  • step S2101 the digital signal used for training in each frame is mapped to a quadrature amplitude modulation constellation through quadrature amplitude modulation QAM;
  • the digital signal used for training in each frame is mapped to the quadrature amplitude modulation constellation through quadrature amplitude modulation QAM, and the digital signal is mapped to the quadrature amplitude modulation constellation through quadrature amplitude modulation QAM, for example, by QAM maps the digital signal to a modulation constellation for transmission.
  • the digital signal is mapped according to the orthogonality of the trigonometric function to generate different constellation points.
  • Step S2102 and determine the closest point on the quadrature amplitude modulation constellation diagram to the digital signal used for training in each frame according to the distribution of the digital signal used for training in each frame in the Hilbert space;
  • the point on the quadrature amplitude modulation constellation diagram that is closest to the digital signal used for training in each frame is determined.
  • Step S2103 Obtain the label information of the digital signal used for training in each frame according to the coordinate information of the nearest point.
  • each frame of data is sequentially mapped to the fourth-order quadrature amplitude modulation constellation diagram according to the transformation.
  • label the nearest constellation point as the output label.
  • the number of corresponding classification results required is N.
  • model prediction optimization method in the embodiment of the present application is described above, and the model prediction optimization device in the embodiment of the present application is described below.
  • FIG. 8 is a schematic diagram of the functional modules of the first embodiment of the model prediction and optimization apparatus according to the present application.
  • the model prediction optimization device includes:
  • the training output module 10 is configured to input the first training sample data into a deep neural network set pre-deployed in the machine learning framework tensorflow, and when training each deep neural network in the deep neural network set through the first training sample data , Using a random algorithm to randomly hide neurons in each deep neural network according to a preset hiding ratio, obtain a deep neural network after the hidden neurons, and output the first prediction result of each deep neural network;
  • the normalization calculation module 20 is used to normalize the first prediction result output by the deep neural network after the hidden neurons by a normalization algorithm to obtain a prediction result set, and the prediction result set includes multiple First prediction result;
  • the calculation module 30 is configured to calculate the accuracy of each first prediction result in the prediction result set according to the first actual result, and set weights for the deep neural network behind each hidden neuron based on the accuracy to obtain the first prediction model;
  • the first prediction module 40 is configured to predict the first data to be predicted through the first prediction model to obtain a second prediction result
  • the first judgment module 50 is configured to judge whether the second prediction result meets the preset accuracy rate according to the second actual result
  • the adjustment module 60 is configured to obtain a second prediction model if the second prediction result meets the preset accuracy rate, and if the second prediction result does not meet the preset accuracy rate, adjust each depth nerve through a back propagation algorithm The weight of the network until the second prediction result meets the preset accuracy rate;
  • the second prediction module 70 is configured to predict the second to-be-predicted data through the second prediction model to obtain a third prediction result
  • the second judgment module 80 is configured to judge whether the amount of error data in the third prediction result is greater than the processing amount of the error data by the preset error correction code;
  • the classification module 90 is configured to classify the current data through a support vector machine model pre-deployed in the machine learning framework tensorflow if the amount of error data in the current data is greater than the processing amount of the error data by the preset error correction code, Obtain error data and correct data and transmit the correct data to the data receiving end. If the amount of error data in the current data is less than or equal to the processing amount of the error data by the preset error correction code, the third prediction result is transmitted To the data receiving end.
  • the problem of low efficiency in negotiating error correction is solved by the classification module. If the prediction result shows that the amount of error data has exceeded the number that can be corrected by the error correction code, the data is classified, and the error data and the correct data are classified. Data, and the correct data is transmitted to the data receiving end, which improves the accuracy of data transmission.
  • FIG. 9 is a schematic diagram of the functional modules of the second embodiment of the model prediction and optimization device according to the present application.
  • the model prediction optimization device includes:
  • the receiving module 100 is configured to receive an optical signal, and perform a time-domain sampling operation on the optical signal to obtain multiple frames of sub-optical signals;
  • the conversion module 110 is used to convert the sub-optical signal into a digital signal; the standardization processing module 120 is used to perform a standardization process on the digital signal to obtain the first to-be-predicted data.
  • Figures 7-8 above describe in detail the model prediction and optimization device in the embodiment of the present application from the perspective of modular functional entities, and the following describes the model prediction and optimization device in the embodiment of the present application in detail from the perspective of hardware processing.
  • the present application also provides a computer-readable storage medium.
  • the computer-readable storage medium may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium.
  • the computer-readable storage medium stores computer instructions, and when the computer instructions run on a computer, the computer executes the above-mentioned model prediction optimization method.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Probability & Statistics with Applications (AREA)
  • Evolutionary Computation (AREA)
  • Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

一种模型预测优化方法、装置、设备及计算机可读存储介质,涉及人工智能技术领域,包括以下步骤:通过隐藏神经元后的深度神经网络输出第一预测结果;对第一预测结果归一化处理,得到预测结果集合;基于预测结果的准确度为各个深度神经网络设置权重,对第一待预测数据进行预测得到第二预测结果;判断第二预测结果是否满足预置准确率;若否,则调节各个深度神经网络所占的权重,得到第三预测结果;判断数据中的错误数据量是否大于预置纠错码对错误数据的处理量。所述方法提高了模型对待预测数据的预测效果。

Description

模型预测优化方法、装置、设备及可读存储介质
本申请要求于2019年09月19日提交中国专利局、申请号为201910884629.X、发明名称为“模型预测优化方法、装置、设备及可读存储介质”的中国专利申请的优先权,其全部内容通过引用结合在申请中。
技术领域
本申请涉及人工智能技术领域,尤其涉及模型预测优化方法、装置、设备及可读存储介质。
背景技术
目前,现有的通信系统一般包括发送端和接收端,接收端需要对来自发送端的信号进行数据后处理。在后处理过程中需要对信号进行纠错处理,目前普遍依赖于纠错码进行纠错处理,例如重复累加RA码和多边型LDPC码。然而,发明人意识到RA码和多边型LDPC码具有相当高的字错误率,并且在非常低的信噪比下,RA编码具有较差的效率,从而导致通信系统的传输性能较低。如何对预测模型进行优化,以提高模型对待传输数据的预测效果,是目前本领域技术人员亟待解决的问题。
发明内容
本申请的主要目的在于解决现有技术中,模型对待传输数据预测结果的准确率低的技术问题。
为实现上述目的,本申请第一方面提供了一种模型预测优化方法,包括:将第一训练样本数据输入预先部署在机器学习框架tensorflow内的深度神经网络集,在通过所述第一训练样本数据训练所述深度神经网络集中的各个深度神经网络时,通过随机算法按照预置隐藏比例,随机隐藏各个深度神经网络中的神经元,得到隐藏神经元后的深度神经网络以及输出所述各个深度神经网络的第一预测结果;通过归一化算法对所述隐藏神经元后的深度神经网络所输出的第一预测结果进行归一化处理,得到预测结果集合,所述预测结果集合中包括多个第一预测结果;根据第一实际结果计算所述预测结果集合中各个第一预测结果的准确度,以及基于所述准确度为各个隐藏神经元后的深度神经网络设置权重,得到第一预测模型;通过所述第一预测模型对第一待预测数据进行预测,得到第二预测结果;根据第二实际结果判断所述第二预测结果是否满足预置准确率;若所述第二预测结果满足预置准确率,则得第二预测模型,若所述第二预测结果不满足预置准确率,则通过反向传播算法调节各个深度神经网络所占的权重,直至所述第二预测结果满足预置准确率;通过所述第二预测模型对第二待预测数据进行预测,得到第三预测结果;判断所述第三预测结果中错误数据量是否大于预置纠错码对错误数据的处理量;若当前数据中的错误数据量大于预置纠错码对错误数据的处理量,则通过预先部署在机器学习框架tensorflow内的支持向量机模型对所述当前数据进行分类,得到错误数据与正确数据以及将所述正确数据传送至数据接收端,若当前数据中的错误数据量小于或等于预置纠错码对错误数据的处理量,则将所述第三预测结果传送至数据接收端。
本申请第二方面提供了一种模型预测优化装置,包括:训练输出模块,用于将第一训练样本数据输入预先部署在机器学习框架tensorflow内的深度神经网络集,在通过所述第 一训练样本数据训练所述深度神经网络集中的各个深度神经网络时,通过随机算法按照预置隐藏比例,随机隐藏各个深度神经网络中的神经元,得到隐藏神经元后的深度神经网络以及输出所述各个深度神经网络的第一预测结果;归一计算模块,用于通过归一化算法对所述隐藏神经元后的深度神经网络所输出的第一预测结果进行归一化处理,得到预测结果集合,所述预测结果集合中包括多个第一预测结果;计算模块,用于根据第一实际结果计算所述预测结果集合中各个第一预测结果的准确度,以及基于所述准确度为各个隐藏神经元后的深度神经网络设置权重,得到第一预测模型;第一预测模块,用于通过所述第一预测模型对第一待预测数据进行预测,得到第二预测结果;第一判断模块,用于根据第二实际结果判断所述第二预测结果是否满足预置准确率;调节模块,用于若所述第二预测结果满足预置准确率,则得第二预测模型,若所述第二预测结果不满足预置准确率,则通过反向传播算法调节各个深度神经网络所占的权重,直至所述第二预测结果满足预置准确率;第二预测模块,用于通过所述第二预测模型对第二待预测数据进行预测,得到第三预测结果;第二判断模块,用于判断所述第三预测结果中错误数据量是否大于预置纠错码对错误数据的处理量;分类模块,用于若当前数据中的错误数据量大于预置纠错码对错误数据的处理量,则通过预先部署在机器学习框架tensorflow内的支持向量机模型对所述当前数据进行分类,得到错误数据与正确数据以及将所述正确数据传送至数据接收端,若当前数据中的错误数据量小于或等于预置纠错码对错误数据的处理量,则将所述第三预测结果传送至数据接收端。
本申请第三方面提供了一种模型预测优化设备,包括:存储器、处理器以及存储在所述存储器上并可在所述处理器上运行的模型预测优化程序,所述模型预测优化程序被所述处理器执行时实现如上述任一项所述的模型预测优化方法的步骤。
本申请的第四方面提供了一种计算机可读存储介质,所述计算机可读存储介质中存储有指令,当所述指令在计算机上运行时,使得计算机执行上述第一方面所述的方法。
本申请通过预测模型对后处理部分的通信信号进行预测,很好地解决了协商纠错中效率低下的难题,另外,若预测结果表明错误数据的量已经超过纠错码的纠错能力,则对数据进行分类,分类出错误数据与正确数据,并将正确数据传送至数据接收端,提高了数据传输的准确率,对模型的训练过程进行了优化,提高了模型对待传输数据的预测效果。
附图说明
图1为本申请实施例方案涉及的模型预测优化设备运行环境的结构示意图;
图2为本申请模型预测优化方法第一实施例的流程示意图;
图3为本申请模型预测优化方法第二实施例的流程示意图;
图4为本申请模型预测优化方法第三实施例的流程示意图;
图5为图4中步骤S180的细化流程示意图;
图6为本申请模型预测优化方法第四实施例的流程示意图;
图7为图6中步骤S210的细化流程示意图;
图8为本申请模型预测优化装置第一实施例的功能模块示意图;
图9为本申请模型预测优化装置第二实施例的功能模块示意图。
具体实施方式
本申请实施例提供了一种模型预测优化方法、装置、设备及可读存储介质,通过预测模型对后处理部分的通信信号进行预测,很好地解决了协商纠错中效率低下的难题,另外,若预测结果表明错误数据的量已经超过纠错码的纠错能力,则对数据进行分类,分类出错误数据与正确数据,并将正确数据传送至数据接收端,提高了数据传输的准确率,对模型的训练过程进行了优化,提高了模型对待传输数据的预测效果。
本申请提供一种模型预测优化设备。
参照图1,图1为本申请实施例方案涉及的模型预测优化设备运行环境的结构示意图。
如图1所示,该模型预测优化设备包括:处理器1001,例如CPU,通信总线1002、用户接口1003,网络接口1004,存储器1005。其中,通信总线1002用于实现这些组件之间的连接通信。用户接口1003可以包括显示屏(Display)、输入单元比如键盘(Keyboard),网络接口1004可选的可以包括标准的有线接口、无线接口(如WI-FI接口)。存储器1005可以是高速RAM存储器,也可以是稳定的存储器(non-volatile memory),例如磁盘存储器。存储器1005可选的还可以是独立于前述处理器1001的存储装置。
本领域技术人员可以理解,图1中示出的模型预测优化设备的硬件结构并不构成对模型预测优化设备的限定,可以包括比图示更多或更少的部件,或者组合某些部件,或者不同的部件布置。
如图1所示,作为一种计算机可读存储介质的存储器1005中可以包括操作系统、网络通信模块、用户接口模块以及模型预测优化程序。其中,操作系统是管理和控制模型预测优化设备和软件资源的程序,支持模型预测优化程序以及其它软件和/或程序的运行。
在图1所示的模型预测优化设备的硬件结构中,网络接口1004主要用于接入网络;用户接口1003主要用于侦测确认指令和编辑指令等。而处理器1001可以用于调用存储器1005中存储的模型预测优化程序,并执行以下模型预测优化方法的各实施例的操作。
基于上述模型预测优化设备硬件结构,提出本申请模型预测优化方法的各个实施例。
为便于理解,下面对本申请实施例的具体流程进行描述。
参照图2,图2为本申请模型预测优化方法第一实施例的流程示意图。本实施例中,所述模型预测优化方法包括以下步骤:
步骤S10,将第一训练样本数据输入预先部署在机器学习框架tensorflow内的深度神经网络集,在通过所述第一训练样本数据训练所述深度神经网络集中的各个深度神经网络时,通过随机算法按照预置隐藏比例,随机隐藏各个深度神经网络中的神经元,得到隐藏神经元后的深度神经网络以及输出所述各个深度神经网络的第一预测结果;
本实施例中,在训练深度神经网络的时候,会遇到两个问题:一是容易出现过拟合现象,即当待预测数据发生变化时,训练好的模型无法根据变化后的数据输出准确的预测结果,二是比较费时,在实际训练过程中往往需要大量的训练样本才能训练出符合预测标准的模型。为了解决这上述问题,因此通过随机算法按照预置隐藏比例,随机隐藏各个深度神经网络中的神经元,得到隐藏神经元后的深度神经网络以及输出所述各个深度神经网络的第一预测结果。
步骤S20,通过归一化算法对所述隐藏神经元后的深度神经网络所输出的第一预测结果进行归一化处理,得到预测结果集合,所述预测结果集合中包括多个第一预测结果;
本实施例中,通过归一化算法对所述隐藏神经元后的深度神经网络所输出的第一预测结果进行归一化处理,得到预测结果集合,所述预测结果集合中包括多个第一预测结果,由于是按照预置比例,随机性地隐藏了深度神经网络中的神经元,因此每个深度神经网络输出的预测结果可能不尽相同,就因为输出的预测结果存在不完全相同的情况,因此可以极大减少过拟合现象的发生,进而可提高预测结果的准确性,另外,由于采用的是由多个深度神经网络对训练样本进行处理的方式,提高了模型训练的速度。例如,通过归一化算法对第一深度神经网络输出的结果的进行归一化处理后,得到的值是正为0.8,负为0.2,第二深度神经网络到的值是正为0.6,负为0.4,第三深度神经网络到的值是正为0.3,负为0.7。
步骤S30,根据第一实际结果计算所述预测结果集合中各个第一预测结果的准确度,以及基于所述准确度为各个隐藏神经元后的深度神经网络设置权重,得到第一预测模型;
本实施例中,第一实际结果是预置的,根据第一实际结果计算所述预测结果集合中各个第一预测结果的准确度,第一实际结果指的是已知被掩盖掉的数据是什么数据的校验数据,用于检验准确度,例如,将“我喜欢吃苹果”中的“苹果”预先设置为错误数据,并将“苹果”掩盖掉,但是人是知道被掩盖的是“苹果”,此处是“苹果”就是校验数据,由于隐藏神经元后的深度神经网络并不知道被掩盖的数据是“苹果”,因此通过隐藏神经元后的深度神经网络进行预测后,会得到预测结果,此处的预测结果可能是“苹果”,也可能是“香蕉”,通过计算预测结果中“苹果”的出现次数与所有预测结果的比值,即可得到准确度。
考虑的单一的神经网络模型对数据进行预测的能力是有限的,因此采用多个深度神经网络算法相互结合的方式,解决了单一的深度神经网络算法对数据处理的局限性,对数据进行预测的目的是,检验当前数据中的错误数据的量是否已经超过纠错码可纠错的数量,若是,则说明信息接收方存在接收到错误的信息的情况,为了保证信息的安全传输,因此需要对数据中的错误数据进行实时预测。
步骤S40,通过所述第一预测模型对第一待预测数据进行预测,得到第二预测结果;
本实施例中,通过所述第一预测模型对第一待预测数据进行预测,得到第二预测结果,将第一训练样本数据输入预先部署在机器学习框架tensorflow内的深度神经网络集,在通过所述第一训练样本数据训练所述深度神经网络集中的各个深度神经网络时,通过随机算法按照预置隐藏比例,随机隐藏各个深度神经网络中的神经元,得到隐藏神经元后的深度神经网络以及输出所述各个深度神经网络的第一预测结果,通过归一化算法对所述隐藏神经元后的深度神经网络所输出的第一预测结果进行归一化处理,得到预测结果集合,根据第一实际结果计算所述预测结果集合中各个第一预测结果的准确度,以及基于所述准确度为各个隐藏神经元后的深度神经网络设置权重,得到第一预测模型,此时的神经网络通过训练后已经具备了预测的能力,因此可以通过所述第一预测模型对第一待预测数据进行预测,得到第二预测结果。
步骤S50,根据第二实际结果判断所述第二预测结果是否满足预置准确率;
本实施例中,第二实际结果是预置的,是已知被掩盖掉的数据是什么数据的校验数据,第二预测结果指的是,通过第一预测模型对待预测数据进行预测后,得到到的被预先掩盖掉的文本数据。例如,预先掩盖掉的数据为“评估”,必须要知道被预先掩盖掉的数据才可以检验通过第一预测模型输出的第二预测结果是否为“评估”。根据第二实际结果判断所述第二预测结果是否满足预置准确率,为了检验第二预测结果是否满足预置准确率,因此需要预先准备好对第一待预测数据的人工预测结果,例如通过标注方式对不同的数据进行标注。
步骤S60,若所述第二预测结果不满足预置准确率,则返回步骤S10,通过反向传播算法调节各个深度神经网络所占的权重,直至所述第二预测结果满足预置准确率;
本实施例中,若所述第二预测结果满足预置准确率,则得第二预测模型,第二预测模型指的是训练完整的第一预测模型,且可以对待预测数据进行正常预测。若所述第一预测结果不满足预置准确率,则说明输出错误预测结果的深度神经网络所占的权重太高,因此需要通过反向传播算法调节各个深度神经网络所占的权重,直至所述第二预测结果满足预置准确率。
步骤S70,若所述第二预测结果满足预置准确率,则得第二预测模型;
本实施例中,若所述第二预测结果满足预置准确率,说明第一预测模型输出的预测结果已经满足预置准确率,例如,98%,则得第二预测模型。
步骤S80,通过所述第二预测模型对第二待预测数据进行预测,得到第三预测结果;
本实施例中,将第二待预测数据输入第二预测模型后,第二预测模型输出一个四行一列的向量,假设该向量为y=[-0.5,0.1,0,0.3],首先取以自然数e的幂,变成以下形式:y 1=[exp(-0.5),exp(0),exp(0.1),exp(0.3)]=[0.6065,1.1051,1.0,1.3498],对y 1所有元素求和得到y 1_sum=0.6065+1.1051+1.0+1.3498=4.061,记y2=y 1/y 1_sum=[0.149,0.272,0.246,0.332],y 2中第4列元素最大,那么y的类别就是4,即得到与第四列元素相对应的数据。
步骤S90,判断所述第三预测结果中错误数据量是否大于预置纠错码对错误数据的处理量;
本实施例中,根据所述第二预测结果判断当前数据中的错误数据量是否大于预置纠错码对错误数据的处理量,例如,对第一个数据的预测结果是错误数据,对第二个数据的预测结果为错误数据直至对第一百个数据的预测结果也为错误数据,其中,纠错码仅能处理九十九个错误数据,则说明当前数据中的错误数据量是否大于预置纠错码对错误数据的处理量。
步骤S100,若当前数据中的错误数据量大于预置纠错码对错误数据的处理量,则通过预先部署在机器学习框架tensorflow内的支持向量机模型对所述当前数据进行分类,得到错误数据与正确数据以及将所述正确数据传送至数据接收端;
本实施例中,若当前数据中的错误数据量大于预置纠错码对错误数据的处理量,则通过预先部署在机器学习框架tensorflow内的支持向量机模型对所述当前数据进行分类,得到错误数据与正确数据以及将所述正确数据传送至数据接收端。
步骤S110,若当前数据中的错误数据量小于或等于预置纠错码对错误数据的处理量,则将所述第三预测结果传送至数据接收端。
本实施例中,通过预测模型对后处理部分的通信信号进行预测,很好地解决了协商纠错中效率低下的难题,另外,若预测结果表明错误数据的量已经超过纠错码的纠错能力,则对数据进行分类,分类出错误数据与正确数据,并将正确数据传送至数据接收端,提高了数据传输的准确率。
参照图3,图3为本申请模型预测优化方法第二实施例的流程示意图。本实施例中,在图2的步骤S10之前,还包括以下步骤:
步骤S120,根据预先设置的随机隐藏神经元的策略,通过随机算法随机隐藏同一深度神经网络的神经元;
本实施例中,根据预先设置的随机隐藏神经元的策略,通过随机算法随机隐藏同一深度神经网络的神经元,随机隐藏神经元的策略指的是,随机隐藏深度神经网络的神经元,隐藏的规律可以是随机隐藏10%的神经元、随机隐藏20%的神经元、随机隐藏30%的神经元至随机隐藏100%的神经元。
步骤S130,采用第二训练样本数据训练所述隐藏神经元后的深度神经网络,得到初始深度神经网络;
本实施例中,采用第二训练样本数据训练所述隐藏神经元后的深度神经网络,得到初始深度神经网络,例如,目前有三个神经网络,第一个神经网络随机隐藏掉10%的神经元,第二个神经网络随机隐藏掉20%的神经元,第三个神经网络随机隐藏掉30%的神经元,然后采用同一训练样本训练所述隐藏神经元后的深度神经网络,得到三个初始深度神经网络。
步骤S140,通过所述初始深度神经网络对第三待预测数据进行预测,得到初始预测结果集合,所述初始预测结果集合包括多个初始预测结果;
本实施例中,通过所述初始深度神经网络对第三待预测数据进行预测,得到初始预测结果集合,所述初始预测结果集合包括多个初始预测结果,根据预先设置的随机隐藏神经元的策略,通过随机算法随机隐藏同一深度神经网络的神经元,然后,采用第二训练样本数据训练所述隐藏神经元后的深度神经网络,得到初始深度神经网络,为了检验初始深度神经网络是否具备了一定的预测准确率,例如97%,则需要通过所述初始深度神经网络对第三待预测数据进行预测,得到初始预测结果集合,初始预测结果集合包括多个初始预测结果。
步骤S150,根据预置人工预测结果从所述初始预测结果集合中筛选出正确率最高的初始预测结果,以及基于所述正确率最高的初始预测结果确定隐藏神经元之间的预置隐藏比例。
本实施例中,需要预先设置随机隐藏神经元的策略,例如,先通过第一训练样本数据去训练通过随机算法随机隐藏10%的神经元,训练一千次后,再通过所述隐藏10%的神经元 后的深度神经网络对第二样本数据进行预测,得到第一个预测结果;再通过第一训练样本数据去训练通过随机算法随机隐藏20%的神经元后的深度神经网络,训练一千次后,对第二样本数据进行预测,得到第二个预测结果直至通过通过第一训练样本数据去训练隐藏90%的神经元后的深度神经网络,训练一千次后,对第二样本数据进行预测,得到第九个预测结果。通过预先准备好的人工预测结果检验上述九个预测结果的正确率的大小,筛选出正确率最大预测结果,从而可以确定隐藏的神经元所占的最佳比例。
参照图4,图4为本申请模型预测优化方法第三实施例的流程示意图。本实施例中,在图2的步骤S40之前,还包括以下步骤:
步骤S160,接收光信号,并对所述光信号进行时域采样操作,得到多帧子光信号;
本实施例中,为了将光纤信道中传输的光信号转为计算机可识别的数字信号,在信号接收端需要首先需要对光信号进行时域的采样操作,得到多帧子光信号。具体的,可通过动态偏振控制器对信道中的光子进行线性双折射,改变入射光信号的相位差,实现偏振态转换。然后通过偏振光分束器对偏振态的光信号进行分束,形成两束性质一致光强可调的信号光,对这两束光信号进行相位调制,形成同相、正交两支光束。
步骤S170,将所述子光信号转换为数字信号;
本实施例中,通过模数转换器对子光信号进一步处理,将子光信号变为可以携带光子信息的数字信号。该步完成了由模拟信号至数字信号的转换。
步骤S180,对所述数字信号进行标准化处理,得到第一待预测数据。
本实施例中,假设一帧数字信号包括100个数据:X 1、X 2、X 3......X 100,计算这100个数据的平均值M,并查找这100个数据中的最大值max和最小值min,然后代入公式(X i-M)/(max-min),计算得到100个离散数据,并将100个离散数据压缩至[-1,1]的范围,作为第一待预测数据。该步骤可提高预测准确率。
参照图5,图5为图4中步骤S180的细化流程示意图。本实施例中,步骤S180包括以下步骤:
步骤S1801,计算所述数字信号包括的n个数据的平均值M,并查找所述n个数据中的最大值max和最小值min,其中,每个数据记作X i,i为不同数据的标识;
本实施例中,计算所述数字信号包括的n个数据的平均值M,并查找所述n个数据中的最大值max和最小值min,其中,每个数据记作X i,i为不同数据的标识。
步骤S1802,通过公式(X i-M)/(max-min)计算得到n个离散数据;
本实施例中,通过公式(X i-M)/(max-min)计算得到n个离散数据。
步骤S1803,对所述n个离散数据进行压缩,得到第一待预测数据。
本实施例中,假设一帧数字信号包括100个数据:X 1、X 2、X 3......X 100,计算这100个数据的平均值M,并查找这100个数据中的最大值max和最小值min,然后代入公式(X i-M)/(max-min),计算得到100个离散数据,并将100个离散数据压缩至[-1,1]的范围,作为第一待预测数据。该步骤可提高预测准确率。
参照图6,图6为本申请模型预测优化方法第四实施例的流程示意图。本实施例中,在图2的步骤S10之前,还包括以下步骤:
步骤S190,获取用于训练的光信号,并对所述用于训练的光信号进行时域采样操作,得到多帧用于训练的子光信号;
本实施例中,获取用于训练的光信号,并对所述用于训练的光信号进行时域采样操作,得到多帧用于训练的子光信号。
步骤S200,将所述多帧用于训练的子光信号转换为多帧用于训练的数字信号;
本实施例中,为了将光纤信道中传输的光信号转为计算机可识别的数字信号,在信号接收端需要首先需要对用于训练的光信号进行时域的采样操作。然后通过模数转换器对信号进一步处理变为可以携带光子信息的用于训练的数字信号。
步骤S210,对所述多帧用于训练的数字信号进行标注,得到每帧用于训练的数字信号的标注信息;
本实施例中,将每一帧数据依次按照变换映射至N阶正交调幅星座图上。根据数据在希尔伯特空间的分布情况,标注与之最近的一个星座图的点作为输出的label。对于N阶的正交调幅星座图的信号分类问题来说,所需要相应的分类结果数为N。
步骤S220,对所述多帧用于训练的数字信号进行标准化处理,得到多帧标准数据;
本实施例中,由于要通过深度神经网络进行训练,所以在数据在输入前会对数据进行标准化处理,具体为:假设一帧用于训练的数字信号包括100个数据:X 1、X 2、X 3......X 100,计算这100个数据的平均值M,并查找这100个数据中的最大值max和最小值min,然后代入公式:
(X i-M)/(max-min),计算得到100个离散数据,将离散数据映射压缩至[-1,1]的范围,以便在前馈网络学习和反向传播权重更新的过程中避免发生特征的偏移。
步骤S230,基于所述标注信息,将所述多帧标准数据分为多组,其中,每组中的标准数据的标注信息相同;
本实施例中,若采用四阶正交调幅星座图,则标注信息分为四种:第一类、第二类、第三类以及第四类。则根据标注信息,可将多帧标准数据分为四组,且每组中的标准数据的标注信息相同。
步骤S240,从每组中获取相同数量级的标准数据,组成训练样本。
本实施例中,从每组中获取相同数量级(例如50个)的标准数据,组成训练数据,这样可以更好的对数据特征进行学习,以达到更准确的模型训练效果。
参照图7,图7为图6中步骤S210的细化流程示意图。本实施例中,步骤S210包括以下步骤:
步骤S2101,通过正交振幅调制QAM将每一帧用于训练的数字信号映射至正交调幅星座图;
本实施例中,通过正交振幅调制QAM将每一帧用于训练的数字信号映射至正交调幅星座图,通过正交振幅调制QAM实现将数字信号映射至正交调幅星座图,例如,通过QAM将数字信号映射到一个调制星座,以便进行传输,根据所应用的函数,例如,根据三角函数的正交性对数字信号进行映射,将产生不同的星座点。
步骤S2102,并根据每一帧用于训练的数字信号在希尔伯特空间的分布情况,确定所述正交调幅星座图上与每一帧用于训练的数字信号最近的点;
本实施例中,并根据每一帧用于训练的数字信号在希尔伯特空间的分布情况,确定所述正交调幅星座图上与每一帧用于训练的数字信号最近的点。
步骤S2103,根据所述最近的点的坐标信息,得到每一帧用于训练的数字信号的标注信息。
本实施例中,将每一帧数据依次按照变换映射至四阶正交调幅星座图上。根据数据在希尔伯特空间的分布情况,标注与之最近的一个星座图的点作为输出的label。对于N阶的正交调幅星座图的信号分类问题来说,所需要相应的分类结果数为N。
上面对本申请实施例中模型预测优化方法进行了描述,下面对本申请实施例中模型预测优化装置进行描述。
参照图8,图8为本申请模型预测优化装置第一实施例的功能模块示意图。本实施例中,所述模型预测优化装置包括:
训练输出模块10,用于将第一训练样本数据输入预先部署在机器学习框架tensorflow内的深度神经网络集,在通过所述第一训练样本数据训练所述深度神经网络集中的各个深度神经网络时,通过随机算法按照预置隐藏比例,随机隐藏各个深度神经网络中的神经元,得到隐藏神经元后的深度神经网络以及输出所述各个深度神经网络的第一预测结果;
归一计算模块20,用于通过归一化算法对所述隐藏神经元后的深度神经网络所输出的第一预测结果进行归一化处理,得到预测结果集合,所述预测结果集合中包括多个第一预测结果;
计算模块30,用于根据第一实际结果计算所述预测结果集合中各个第一预测结果的准确度,以及基于所述准确度为各个隐藏神经元后的深度神经网络设置权重,得到第一预测模型;
第一预测模块40,用于通过所述第一预测模型对第一待预测数据进行预测,得到第二预测结果;
第一判断模块50,用于根据第二实际结果判断所述第二预测结果是否满足预置准确率;
调节模块60,用于若所述第二预测结果满足预置准确率,则得第二预测模型,若所述第二预测结果不满足预置准确率,则通过反向传播算法调节各个深度神经网络所占的权重,直至所述第二预测结果满足预置准确率;
第二预测模块70,用于通过所述第二预测模型对第二待预测数据进行预测,得到第三预测结果;
第二判断模块80,用于判断所述第三预测结果中错误数据量是否大于预置纠错码对错误数据的处理量;
分类模块90,用于若当前数据中的错误数据量大于预置纠错码对错误数据的处理量,则通过预先部署在机器学习框架tensorflow内的支持向量机模型对所述当前数据进行 分类,得到错误数据与正确数据以及将所述正确数据传送至数据接收端,若当前数据中的错误数据量小于或等于预置纠错码对错误数据的处理量,则将所述第三预测结果传送至数据接收端。
本实施例中,通过分类模块解决了协商纠错中效率低下的难题,若预测结果表明错误数据的量已经超过纠错码可纠错的数量,则对数据进行分类,分类出错误数据与正确数据,并将正确数据传送至数据接收端,提高了数据传输的准确率。
参照图9,图9为本申请模型预测优化装置第二实施例的功能模块示意图。本实施例中,所述模型预测优化装置包括:
接收模块100,用于接收光信号,并对所述光信号进行时域采样操作,得到多帧子光信号;
转换模块110,用于将所述子光信号转换为数字信号;标准化处理模块120,用于对所述数字信号进行标准化处理,得到第一待预测数据。
上面图7-8从模块化功能实体的角度对本申请实施例中的模型预测优化装置进行详细描述,下面从硬件处理的角度对本申请实施例中模型预测优化设备进行详细描述。
本申请还提供一种计算机可读存储介质,该计算机可读存储介质可以是非易失性计算机可读存储介质,还可以是易失性计算机可读存储介质。计算机可读存储介质存储有计算机指令,当所述计算机指令在计算机上运行时,使得计算机执行上述模型预测优化方法。
所属领域的技术人员可以清楚地了解到,为描述的方便和简洁,上述描述的系统,装置和单元的具体工作过程,可以参考前述方法实施例中的对应过程,在此不再赘述。

Claims (20)

  1. 一种模型预测优化方法,包括:
    将第一训练样本数据输入预先部署在机器学习框架tensorflow内的深度神经网络集,在通过所述第一训练样本数据训练所述深度神经网络集中的各个深度神经网络时,通过随机算法按照预置隐藏比例,随机隐藏各个深度神经网络中的神经元,得到隐藏神经元后的深度神经网络以及输出所述各个深度神经网络的第一预测结果;
    通过归一化算法对所述隐藏神经元后的深度神经网络所输出的第一预测结果进行归一化处理,得到预测结果集合,所述预测结果集合中包括多个第一预测结果;
    根据第一实际结果计算所述预测结果集合中各个第一预测结果的准确度,以及基于所述准确度为各个隐藏神经元后的深度神经网络设置权重,得到第一预测模型;
    通过所述第一预测模型对第一待预测数据进行预测,得到第二预测结果;
    根据第二实际结果判断所述第二预测结果是否满足预置准确率;
    若是,则得第二预测模型,若否,则通过反向传播算法调节各个深度神经网络所占的权重,直至所述第二预测结果满足预置准确率;
    通过所述第二预测模型对第二待预测数据进行预测,得到第三预测结果;
    判断所述第三预测结果中错误数据量是否大于预置纠错码对错误数据的处理量;
    若是,则通过预先部署在机器学习框架tensorflow内的支持向量机模型对所述当前数据进行分类,得到错误数据与正确数据以及将所述正确数据传送至数据接收端,若否,则将所述第三预测结果传送至数据接收端。
  2. 如权利要求1所述的模型预测优化方法,在所述将第一训练样本数据输入预先部署在机器学习框架tensorflow内的深度神经网络集,在通过所述第一训练样本数据训练所述深度神经网络集中的各个深度神经网络时,通过随机算法按照预置隐藏比例,随机隐藏各个深度神经网络中的神经元,得到隐藏神经元后的深度神经网络以及输出所述各个深度神经网络的第一预测结果的步骤之前,还包括以下步骤:
    根据预先设置的随机隐藏神经元的策略,通过随机算法随机隐藏同一深度神经网络的神经元;
    采用第二训练样本数据训练所述隐藏神经元后的深度神经网络,得到初始深度神经网络;
    通过所述初始深度神经网络对第三待预测数据进行预测,得到初始预测结果集合,所述初始预测结果集合包括多个初始预测结果;
    根据预置人工预测结果从所述初始预测结果集合中筛选出正确率最高的初始预测结果,以及基于所述正确率最高的初始预测结果确定隐藏神经元之间的预置隐藏比例。
  3. 如权利要求1所述的模型预测优化方法,在所述通过所述第一预测模型对第一待预测数据进行预测,得到第二预测结果的步骤之前,还包括以下步骤:
    接收光信号,并对所述光信号进行时域采样操作,得到多帧子光信号;
    将所述子光信号转换为数字信号;
    对所述数字信号进行标准化处理,得到第一待预测数据。
  4. 如权利要求3所述的模型预测优化方法,所述对所述数字信号进行标准化处理,得到第一待预测数据包括以下步骤:
    计算所述数字信号包括的n个数据的平均值M,并查找所述n个数据中的最大值max和最小值min,其中,每个数据记作X i,i为不同数据的标识;
    通过公式(X i-M)/(max-min)计算得到n个离散数据;
    对所述n个离散数据进行压缩,得到第一待预测数据。
  5. 如权利要求1所述的模型预测优化方法,在将第一训练样本数据输入预先部署在机器学习框架tensorflow内的深度神经网络集,在通过所述第一训练样本数据训练所述深度神经网络集中的各个深度神经网络时,通过随机算法按照预置隐藏比例,随机隐藏各个深度神经网络中的神经元,得到隐藏神经元后的深度神经网络以及输出所述各个深度神经网络的第一预测结果的步骤之前,还包括以下步骤:
    将所述多帧用于训练的子光信号转换为多帧用于训练的数字信号;
    对所述多帧用于训练的数字信号进行标注,得到每帧用于训练的数字信号的标注信息;
    对所述多帧用于训练的数字信号进行标准化处理,得到多帧标准数据;
    基于所述标注信息,将所述多帧标准数据分为多组,其中,每组中的标准数据的标注信息相同;
    从每组中获取相同数量级的标准数据,组成训练样本。
  6. 如权利要求5所述的模型预测优化方法,所述对所述多帧用于训练的数字信号进行标注,得到每帧用于训练的数字信号的标注信息包括以下步骤:
    通过正交振幅调制QAM将每一帧用于训练的数字信号映射至正交调幅星座图;
    并根据每一帧用于训练的数字信号在希尔伯特空间的分布情况,确定所述正交调幅星座图上与每一帧用于训练的数字信号最近的点;
    根据所述最近的点的坐标信息,得到每一帧用于训练的数字信号的标注信息。
  7. 一种模型预测优化装置,所述模型预测优化装置包括:
    训练输出模块,用于将第一训练样本数据输入预先部署在机器学习框架tensorflow内的深度神经网络集,在通过所述第一训练样本数据训练所述深度神经网络集中的各个深度神经网络时,通过随机算法按照预置隐藏比例,随机隐藏各个深度神经网络中的神经元,得到隐藏神经元后的深度神经网络以及输出所述各个深度神经网络的第一预测结果;
    归一计算模块,用于通过归一化算法对所述隐藏神经元后的深度神经网络所输出的第一预测结果进行归一化处理,得到预测结果集合,所述预测结果集合中包括多个第一预测结果;
    计算模块,用于根据第一实际结果计算所述预测结果集合中各个第一预测结果的准确度,以及基于所述准确度为各个隐藏神经元后的深度神经网络设置权重,得到第一预测模型;
    第一预测模块,用于通过所述第一预测模型对第一待预测数据进行预测,得到第二预测结果;
    第一判断模块,用于根据第二实际结果判断所述第二预测结果是否满足预置准确率;
    调节模块,用于若所述第二预测结果满足预置准确率,则得第二预测模型,若所述第二预测结果不满足预置准确率,则通过反向传播算法调节各个深度神经网络所占的权重,直至所述第二预测结果满足预置准确率;
    第二预测模块,用于通过所述第二预测模型对第二待预测数据进行预测,得到第三预测结果;
    第二判断模块,用于判断所述第三预测结果中错误数据量是否大于预置纠错码对错误数据的处理量;
    分类模块,用于若当前数据中的错误数据量大于预置纠错码对错误数据的处理量,则通过预先部署在机器学习框架tensorflow内的支持向量机模型对所述当前数据进行分类,得到错误数据与正确数据以及将所述正确数据传送至数据接收端,若当前数据中的错误数据量小于或等于预置纠错码对错误数据的处理量,则将所述第三预测结果传送至数据接收端。
  8. 如权利要求7所述的模型预测优化装置,所述模型预测优化装置还包括以下模块:
    预置模块,用于根据预先设置的随机隐藏神经元的策略,通过随机算法随机隐藏同一深度神经网络的神经元;
    深度神经网络训练模块,用于采用第二训练样本数据训练所述隐藏神经元后的深度神经网络,得到初始深度神经网络;
    第三待预测数据预测模块,用于通过所述初始深度神经网络对第三待预测数据进行预测,得到初始预测结果集合,所述初始预测结果集合包括多个初始预测结果;
    预置隐藏比例确定模块,用于根据预置人工预测结果从所述初始预测结果集合中筛选出正确率最高的初始预测结果,以及基于所述正确率最高的初始预测结果确定隐藏神经元之间的预置隐藏比例。
  9. 如权利要求7所述的模型预测优化装置,所述模型预测优化装置还包括:
    接收模块,用于接收光信号,并对所述光信号进行时域采样操作,得到多帧子光信号;
    转换模块,用于将所述子光信号转换为数字信号;
    第一标准化处理模块,用于对所述数字信号进行标准化处理,得到第一待预测数据。
  10. 如权利要求9所述的第一标准化处理模块,包括:
    第一计算单元,用于计算所述数字信号包括的n个数据的平均值M,并查找所述n个数据中的最大值max和最小值min,其中,每个数据记作X i,i为不同数据的标识;
    第二计算单元,用于通过公式(X i-M)/(max-min)计算得到n个离散数据;
    压缩单元,用于对所述n个离散数据进行压缩,得到第一待预测数据。
  11. 如权利要求7所述的模型预测优化装置,所述模型预测优化装置还包括:
    转换模块,用于将所述多帧用于训练的子光信号转换为多帧用于训练的数字信号;
    标注模块,用于对所述多帧用于训练的数字信号进行标注,得到每帧用于训练的数字信号的标注信息;
    第二标准化处理模块,用于对所述多帧用于训练的数字信号进行标准化处理,得到多帧标准数据;
    分组模块,用于基于所述标注信息,将所述多帧标准数据分为多组,其中,每组中的标准数据的标注信息相同;
    获取模块,用于从每组中获取相同数量级的标准数据,组成训练样本。
  12. 如权利要求11所述的模型预测优化装置,所述标注模块包括:
    映射单元,用于通过正交振幅调制QAM将每一帧用于训练的数字信号映射至正交调幅星座图;
    筛选单元,用于根据每一帧用于训练的数字信号在希尔伯特空间的分布情况,确定所述正交调幅星座图上与每一帧用于训练的数字信号最近的点;
    输出单元,用于根据所述最近的点的坐标信息,得到每一帧用于训练的数字信号的标注信息。
  13. 一种模型预测优化设备,包括存储器、处理器及存储在所述存储器上并可在所述处理器上运行的计算机程序,所述处理器执行所述计算机程序时实现如下步骤:
    将第一训练样本数据输入预先部署在机器学习框架tensorflow内的深度神经网络集,在通过所述第一训练样本数据训练所述深度神经网络集中的各个深度神经网络时,通过随机算法按照预置隐藏比例,随机隐藏各个深度神经网络中的神经元,得到隐藏神经元后的深度神经网络以及输出所述各个深度神经网络的第一预测结果;
    通过归一化算法对所述隐藏神经元后的深度神经网络所输出的第一预测结果进行归一化处理,得到预测结果集合,所述预测结果集合中包括多个第一预测结果;
    根据第一实际结果计算所述预测结果集合中各个第一预测结果的准确度,以及基于所述准确度为各个隐藏神经元后的深度神经网络设置权重,得到第一预测模型;
    通过所述第一预测模型对第一待预测数据进行预测,得到第二预测结果;
    根据第二实际结果判断所述第二预测结果是否满足预置准确率;
    若是,则得第二预测模型,若否,则通过反向传播算法调节各个深度神经网络所占的权重,直至所述第二预测结果满足预置准确率;
    通过所述第二预测模型对第二待预测数据进行预测,得到第三预测结果;
    判断所述第三预测结果中错误数据量是否大于预置纠错码对错误数据的处理量;
    若是,则通过预先部署在机器学习框架tensorflow内的支持向量机模型对所述当前数据进行分类,得到错误数据与正确数据以及将所述正确数据传送至数据接收端,若否,则将所述第三预测结果传送至数据接收端。
  14. 根据权利要求13所述的模型预测优化设备,所述处理器执行所述计算机程序实现所述将第一训练样本数据输入预先部署在机器学习框架tensorflow内的深度神经网络集,在通过所述第一训练样本数据训练所述深度神经网络集中的各个深度神经网络时,通过随机算法按照预置隐藏比例,随机隐藏各个深度神经网络中的神经元,得到隐藏神经元后的深度神经网络以及输出所述各个深度神经网络的第一预测结果的步骤之前,还包括以下步骤:
    根据预先设置的随机隐藏神经元的策略,通过随机算法随机隐藏同一深度神经网络的神经元;
    采用第二训练样本数据训练所述隐藏神经元后的深度神经网络,得到初始深度神经网络;
    通过所述初始深度神经网络对第三待预测数据进行预测,得到初始预测结果集合,所述初始预测结果集合包括多个初始预测结果;
    根据预置人工预测结果从所述初始预测结果集合中筛选出正确率最高的初始预测结果,以及基于所述正确率最高的初始预测结果确定隐藏神经元之间的预置隐藏比例。
  15. 根据权利要求13所述的模型预测优化设备,所述处理器执行所述计算机程序实现所述通过所述第一预测模型对第一待预测数据进行预测,得到第二预测结果的步骤之前,还包括以下步骤:
    接收光信号,并对所述光信号进行时域采样操作,得到多帧子光信号;
    将所述子光信号转换为数字信号;
    对所述数字信号进行标准化处理,得到第一待预测数据。
  16. 根据权利要求15所述的模型预测优化设备,所述处理器执行所述计算机程序实现所述对所述数字信号进行标准化处理,得到第一待预测数据时,包括以下步骤:
    计算所述数字信号包括的n个数据的平均值M,并查找所述n个数据中的最大值max和最小值min,其中,每个数据记作X i,i为不同数据的标识;
    通过公式(X i-M)/(max-min)计算得到n个离散数据;
    对所述n个离散数据进行压缩,得到第一待预测数据。
  17. 根据权利要求13所述的模型预测优化设备,所述处理器执行所述计算机程序实现所述在将第一训练样本数据输入预先部署在机器学习框架tensorflow内的深度神经网络集,在通过所述第一训练样本数据训练所述深度神经网络集中的各个深度神经网络时,通过随机算法按照预置隐藏比例,随机隐藏各个深度神经网络中的神经元,得到隐藏神经元后的深度神经网络以及输出所述各个深度神经网络的第一预测结果的步骤之前,包括以下步骤:
    将所述多帧用于训练的子光信号转换为多帧用于训练的数字信号;
    对所述多帧用于训练的数字信号进行标注,得到每帧用于训练的数字信号的标注信息;
    对所述多帧用于训练的数字信号进行标准化处理,得到多帧标准数据;
    基于所述标注信息,将所述多帧标准数据分为多组,其中,每组中的标准数据的标注信息相同;
    从每组中获取相同数量级的标准数据,组成训练样本。
  18. 根据权利要求17所述的模型预测优化设备,所述处理器执行所述计算机程序实现所述对所述多帧用于训练的数字信号进行标注,得到每帧用于训练的数字信号的标注信息时,包括以下步骤:
    通过正交振幅调制QAM将每一帧用于训练的数字信号映射至正交调幅星座图;
    并根据每一帧用于训练的数字信号在希尔伯特空间的分布情况,确定所述正交调幅星座图上与每一帧用于训练的数字信号最近的点;
    根据所述最近的点的坐标信息,得到每一帧用于训练的数字信号的标注信息。
  19. 一种计算机可读存储介质,所述计算机可读存储介质中存储计算机指令,当所述计算机指令在计算机上运行时,使得计算机执行如下步骤:
    将第一训练样本数据输入预先部署在机器学习框架tensorflow内的深度神经网络集,在通过所述第一训练样本数据训练所述深度神经网络集中的各个深度神经网络时,通过随机算法按照预置隐藏比例,随机隐藏各个深度神经网络中的神经元,得到隐藏神经元后的深度神经网络以及输出所述各个深度神经网络的第一预测结果;
    通过归一化算法对所述隐藏神经元后的深度神经网络所输出的第一预测结果进行归一化处理,得到预测结果集合,所述预测结果集合中包括多个第一预测结果;
    根据第一实际结果计算所述预测结果集合中各个第一预测结果的准确度,以及基于所述准确度为各个隐藏神经元后的深度神经网络设置权重,得到第一预测模型;
    通过所述第一预测模型对第一待预测数据进行预测,得到第二预测结果;
    根据第二实际结果判断所述第二预测结果是否满足预置准确率;
    若是,则得第二预测模型,若否,则通过反向传播算法调节各个深度神经网络所占的权重,直至所述第二预测结果满足预置准确率;
    通过所述第二预测模型对第二待预测数据进行预测,得到第三预测结果;
    判断所述第三预测结果中错误数据量是否大于预置纠错码对错误数据的处理量;
    若是,则通过预先部署在机器学习框架tensorflow内的支持向量机模型对所述当前数据进行分类,得到错误数据与正确数据以及将所述正确数据传送至数据接收端,若否,则将所述第三预测结果传送至数据接收端。
  20. 如权利要求19所述的一种计算机可读存储介质,所述计算机可读存储介质中存储计算机指令,当所述计算机指令在计算机上运行时,使得计算机执行如下步骤:
    根据预先设置的随机隐藏神经元的策略,通过随机算法随机隐藏同一深度神经网络的神经元;
    采用第二训练样本数据训练所述隐藏神经元后的深度神经网络,得到初始深度神经网络;
    通过所述初始深度神经网络对第三待预测数据进行预测,得到初始预测结果集合,所述初始预测结果集合包括多个初始预测结果;
    根据预置人工预测结果从所述初始预测结果集合中筛选出正确率最高的初始预测结果,以及基于所述正确率最高的初始预测结果确定隐藏神经元之间的预置隐藏比例。
PCT/CN2019/118263 2019-09-19 2019-11-14 模型预测优化方法、装置、设备及可读存储介质 WO2021051596A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201910884629.XA CN110798227B (zh) 2019-09-19 2019-09-19 模型预测优化方法、装置、设备及可读存储介质
CN201910884629.X 2019-09-19

Publications (1)

Publication Number Publication Date
WO2021051596A1 true WO2021051596A1 (zh) 2021-03-25

Family

ID=69427341

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/118263 WO2021051596A1 (zh) 2019-09-19 2019-11-14 模型预测优化方法、装置、设备及可读存储介质

Country Status (2)

Country Link
CN (1) CN110798227B (zh)
WO (1) WO2021051596A1 (zh)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113837383A (zh) * 2021-10-18 2021-12-24 中国联合网络通信集团有限公司 模型训练方法、装置、电子设备及存储介质
CN114925920A (zh) * 2022-05-25 2022-08-19 中国平安财产保险股份有限公司 离线位置预测方法、装置、电子设备及存储介质
CN115642972A (zh) * 2022-12-23 2023-01-24 鹏城实验室 动态信道通信检测方法、装置、设备及可读存储介质

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114254724A (zh) * 2020-09-11 2022-03-29 华为技术有限公司 一种数据处理方法、神经网络的训练方法以及相关设备
CN112507855A (zh) * 2020-12-04 2021-03-16 国网浙江省电力有限公司武义县供电公司 一种基于瞬时包络等势星球图的通信辐射源个体识别方法
CN113177074B (zh) * 2021-04-02 2023-09-29 北京科技大学 一种提升环境自适应性的光学性能监测系统及方法

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US9053431B1 (en) * 2010-10-26 2015-06-09 Michael Lamport Commons Intelligent control with hierarchical stacked neural networks
CN109146076A (zh) * 2018-08-13 2019-01-04 东软集团股份有限公司 模型生成方法及装置、数据处理方法及装置
US20190156183A1 (en) * 2018-12-27 2019-05-23 David M. Durham Defending neural networks by randomizing model weights
CN109919304A (zh) * 2019-03-04 2019-06-21 腾讯科技(深圳)有限公司 神经网络搜索方法、装置、可读存储介质和计算机设备

Family Cites Families (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102622418B (zh) * 2012-02-21 2013-08-07 北京联合大学 一种基于bp神经网络的预测装置及设备
JP6954082B2 (ja) * 2017-12-15 2021-10-27 富士通株式会社 学習プログラム、予測プログラム、学習方法、予測方法、学習装置および予測装置
CN109905271B (zh) * 2018-05-18 2021-01-12 华为技术有限公司 一种预测方法、训练方法、装置及计算机存储介质
CN109408583B (zh) * 2018-09-25 2023-04-07 平安科技(深圳)有限公司 数据处理方法及装置、计算机可读存储介质、电子设备

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US9053431B1 (en) * 2010-10-26 2015-06-09 Michael Lamport Commons Intelligent control with hierarchical stacked neural networks
CN109146076A (zh) * 2018-08-13 2019-01-04 东软集团股份有限公司 模型生成方法及装置、数据处理方法及装置
US20190156183A1 (en) * 2018-12-27 2019-05-23 David M. Durham Defending neural networks by randomizing model weights
CN109919304A (zh) * 2019-03-04 2019-06-21 腾讯科技(深圳)有限公司 神经网络搜索方法、装置、可读存储介质和计算机设备

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
DAVID W OPITZ, JUDE W SHAVLIK: "Generating Accurate and Diverse Members of a Neural-Network Ensemble", ADVANCES IN NEURAL INFORMATION PROCESSING SYSTEMS 8 (NIPS 1995), 31 December 1995 (1995-12-31), pages 532 - 541, XP007903583 *

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113837383A (zh) * 2021-10-18 2021-12-24 中国联合网络通信集团有限公司 模型训练方法、装置、电子设备及存储介质
CN113837383B (zh) * 2021-10-18 2023-06-23 中国联合网络通信集团有限公司 模型训练方法、装置、电子设备及存储介质
CN114925920A (zh) * 2022-05-25 2022-08-19 中国平安财产保险股份有限公司 离线位置预测方法、装置、电子设备及存储介质
CN114925920B (zh) * 2022-05-25 2024-05-03 中国平安财产保险股份有限公司 离线位置预测方法、装置、电子设备及存储介质
CN115642972A (zh) * 2022-12-23 2023-01-24 鹏城实验室 动态信道通信检测方法、装置、设备及可读存储介质
CN115642972B (zh) * 2022-12-23 2023-03-21 鹏城实验室 动态信道通信检测方法、装置、设备及可读存储介质

Also Published As

Publication number Publication date
CN110798227A (zh) 2020-02-14
CN110798227B (zh) 2023-07-25

Similar Documents

Publication Publication Date Title
WO2021051596A1 (zh) 模型预测优化方法、装置、设备及可读存储介质
CN107342962B (zh) 基于卷积神经网络的深度学习智能星座图分析方法
CN110233661B (zh) 长短期记忆神经网络训练方法,信道参数调整系统及方法
CN107977710B (zh) 用电异常数据检测方法和装置
CN110728328B (zh) 分类模型的训练方法和装置
KR101968449B1 (ko) 데이터 생산성 향상을 위한 ai 학습 기반의 레이블 타입 데이터 자동 검수 시스템 및 그 방법
CN111224779B (zh) 基于码本的物理层密钥生成方法、装置、存储介质及终端
US20210241441A1 (en) Methods, Apparatus and Computer-Readable Mediums Relating to Detection of Cell Conditions in a Wireless Cellular Network
US11907090B2 (en) Machine learning for taps to accelerate TDECQ and other measurements
WO2021088465A1 (zh) 基于多分布测试数据融合的多层感知器快速调制识别方法
CN109246495A (zh) 一种面向多层次、多指标的光网络业务质量评估方法
CN116597461B (zh) 基于人工智能的题目知识点关联方法及系统
US20150331062A1 (en) Failure Detection Method and Detection Device for Inverter
CN115706607A (zh) 使用机器学习的组合tdecq测量和发射器调谐
CN114925720A (zh) 基于时空混合特征提取网络的小样本调制信号识别方法
Gao et al. Joint baud-rate and modulation format identification based on asynchronous delay-tap plots analyzer by using convolutional neural network
WO2023061303A1 (zh) 大尺度衰落的建模及估计方法、系统、设备和存储介质
CN104104389B (zh) 一种信号重建方法及设备
Li et al. Effects of measurement dependence on generalized Clauser-Horne-Shimony-Holt Bell test in the single-run and multiple-run scenarios
CN115422977A (zh) 基于cnn-bls网络的雷达辐射源信号识别方法
CN111612159A (zh) 特征重要性测量方法、设备及可读存储介质
CN114785433A (zh) 信道场景识别的方法、网络设备及存储介质
CN111914923B (zh) 一种基于聚类特征提取的目标分布式识别方法
CN117914378B (zh) 一种5g直放站信号处理方法及系统
CN115496200B (zh) 神经网络量化模型训练方法、装置及设备

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 19946105

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 19946105

Country of ref document: EP

Kind code of ref document: A1