CN110446173B - Efficient and energy-saving satellite-borne wireless sensor network data compression method - Google Patents
Efficient and energy-saving satellite-borne wireless sensor network data compression method Download PDFInfo
- Publication number
- CN110446173B CN110446173B CN201910700088.0A CN201910700088A CN110446173B CN 110446173 B CN110446173 B CN 110446173B CN 201910700088 A CN201910700088 A CN 201910700088A CN 110446173 B CN110446173 B CN 110446173B
- Authority
- CN
- China
- Prior art keywords
- network
- data
- cbn
- vae
- crbm
- Prior art date
- Legal status (The legal status 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 status listed.)
- Expired - Fee Related
Links
- 238000000034 method Methods 0.000 title claims abstract description 61
- 238000013144 data compression Methods 0.000 title claims abstract description 22
- 238000007906 compression Methods 0.000 claims abstract description 37
- 230000006835 compression Effects 0.000 claims abstract description 37
- 238000004364 calculation method Methods 0.000 claims abstract description 29
- 238000012549 training Methods 0.000 claims abstract description 18
- 238000007781 pre-processing Methods 0.000 claims abstract description 5
- 210000002569 neuron Anatomy 0.000 claims description 41
- 238000011176 pooling Methods 0.000 claims description 30
- 239000011159 matrix material Substances 0.000 claims description 23
- 238000013138 pruning Methods 0.000 claims description 18
- 238000005070 sampling Methods 0.000 claims description 18
- 238000004422 calculation algorithm Methods 0.000 claims description 15
- 238000011156 evaluation Methods 0.000 claims description 6
- 230000002159 abnormal effect Effects 0.000 claims description 4
- 238000013508 migration Methods 0.000 claims description 4
- 230000005012 migration Effects 0.000 claims description 4
- 238000005520 cutting process Methods 0.000 claims description 3
- 238000009826 distribution Methods 0.000 claims description 3
- 238000013136 deep learning model Methods 0.000 claims description 2
- 238000013507 mapping Methods 0.000 claims description 2
- 238000005265 energy consumption Methods 0.000 abstract description 16
- 238000004891 communication Methods 0.000 abstract description 9
- 238000003860 storage Methods 0.000 abstract description 4
- 238000009966 trimming Methods 0.000 description 12
- 238000011160 research Methods 0.000 description 9
- 238000012360 testing method Methods 0.000 description 9
- 238000005516 engineering process Methods 0.000 description 6
- 238000010801 machine learning Methods 0.000 description 6
- 238000013528 artificial neural network Methods 0.000 description 4
- 238000013527 convolutional neural network Methods 0.000 description 4
- 238000011161 development Methods 0.000 description 4
- 238000011084 recovery Methods 0.000 description 4
- 230000009467 reduction Effects 0.000 description 4
- PEDCQBHIVMGVHV-UHFFFAOYSA-N Glycerine Chemical compound OCC(O)CO PEDCQBHIVMGVHV-UHFFFAOYSA-N 0.000 description 3
- 238000010586 diagram Methods 0.000 description 3
- 238000005259 measurement Methods 0.000 description 3
- 238000000513 principal component analysis Methods 0.000 description 3
- 230000008569 process Effects 0.000 description 3
- 238000013526 transfer learning Methods 0.000 description 3
- 238000013459 approach Methods 0.000 description 2
- 238000013473 artificial intelligence Methods 0.000 description 2
- 238000013135 deep learning Methods 0.000 description 2
- 230000007123 defense Effects 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 238000012545 processing Methods 0.000 description 2
- 241000854291 Dianthus carthusianorum Species 0.000 description 1
- 230000002776 aggregation Effects 0.000 description 1
- 238000004220 aggregation Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000002457 bidirectional effect Effects 0.000 description 1
- 230000005540 biological transmission Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000000205 computational method Methods 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 230000007613 environmental effect Effects 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 230000006870 function Effects 0.000 description 1
- 239000000463 material Substances 0.000 description 1
- 230000001617 migratory effect Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 238000010606 normalization Methods 0.000 description 1
- 238000012827 research and development Methods 0.000 description 1
- 238000013215 result calculation Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/06—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons
- G06N3/061—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons using biological neurons, e.g. biological neurons connected to an integrated circuit
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L69/00—Network arrangements, protocols or services independent of the application payload and not provided for in the other groups of this subclass
- H04L69/04—Protocols for data compression, e.g. ROHC
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04W—WIRELESS COMMUNICATION NETWORKS
- H04W28/00—Network traffic management; Network resource management
- H04W28/02—Traffic management, e.g. flow control or congestion control
- H04W28/06—Optimizing the usage of the radio link, e.g. header compression, information sizing, discarding information
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04W—WIRELESS COMMUNICATION NETWORKS
- H04W4/00—Services specially adapted for wireless communication networks; Facilities therefor
- H04W4/30—Services specially adapted for particular environments, situations or purposes
- H04W4/38—Services specially adapted for particular environments, situations or purposes for collecting sensor information
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D30/00—Reducing energy consumption in communication networks
- Y02D30/70—Reducing energy consumption in communication networks in wireless communication networks
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Theoretical Computer Science (AREA)
- Computer Networks & Wireless Communication (AREA)
- Signal Processing (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- General Engineering & Computer Science (AREA)
- Neurology (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Computing Systems (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Microelectronics & Electronic Packaging (AREA)
- Computer Security & Cryptography (AREA)
- Mobile Radio Communication Systems (AREA)
Abstract
The invention discloses a high-efficiency energy-saving satellite-borne wireless sensor network data compression method, which comprises the following steps: step one, collecting streaming data of each terminal sensing node of a wireless sensing network; secondly, preprocessing the streaming data; step three, constructing a calculation method of a D-CRBM network calculation layer; step four, combining the D-CRBM network computing layer with a variational hybrid encoder to construct a CBN-VAE network; training a CBN-VAE network to obtain model parameters and constructing a compression model; and step six, compressing the data of the wireless sensor network by adopting a compression model. The invention effectively reduces the node communication energy consumption, the storage energy consumption and the calculation energy consumption of the wireless sensor network.
Description
Technical Field
The invention relates to the technical field of artificial intelligence. More specifically, the invention relates to a high-efficiency and energy-saving data compression method for a satellite-borne wireless sensor network.
Background
A Wireless Sensor Network (WSN) is an intelligent device integrating sensing, computing and communication capabilities. The WSN is considered to be an important future aviation detection technology due to the characteristics of simple deployment, automatic data acquisition, real-time data processing, ad hoc network multi-hop wireless communication, good adaptability to severe environments and the like. The development of deep space exploration has higher technical difficulty and higher risk compared with the activities of near-earth and even moon exploration due to the fact that the deep space exploration is far away from the earth and the environment is complex, and a batch of new core technologies need to be further broken through, wherein autonomous navigation control, energy and propulsion, measurement and control communication, descending and landing entering, novel data acquisition equipment and the like are key technologies which need to be broken through and mastered urgently, and a data acquisition technology for scattering wireless sensing network nodes on the surface of a planet through a spacecraft to carry out environmental information is also one of research hotspots of the aerospace technology in the current stage of China. China, one of the major spacecrafts, is still blank in the deep space exploration field farther than the moon, and the field faces not only a great gap from the traditional aerospace strong countries such as America, Europe and Russia, but also the reality surpassed by emerging aerospace countries such as India, the key link starting before and after the development process from the moon to the planet in the deep space exploration field is also a necessary way for further deep space in the future, at present, a plurality of military and scientific and technological strong countries in the world are actively exploring the application of WSN in the military aerospace field, a plurality of universities and government organizations in Europe and America, such as the United states department of defense advanced research project, the European space technology center, the American aerospace agency, the American scientist, the American department of defense and the like, have been invested with a great deal of manpower and material resources for research, and under the vigorous development situation of aerospace in recent years, the application research requirements of WSN in the fields of aerospace, military and the like are more and more urgent. The research and development of WSN spacecrafts with reliability, effectiveness and instantaneity become one of the very important key technologies in the development of military affairs and aerospace careers in China, in most cases, a sensor node onboard radio transceiver is the main reason of energy consumption, and the energy problem is always the bottleneck limiting the wide use of a wireless sensor network, so that the reduction of communication energy consumption becomes one of the hot spots of wireless sensor network research.
Data compression can effectively reduce the data volume and communication energy consumption of the WSN, and some data compression algorithms focus on time-series-based de-approximation data, convert data samples into a set of coefficients to simplify data representation, such as Fast Fourier Transform (FFT) and Wavelet Transform (WT), the performance of the compression algorithm depends on the number of coefficients needed for encoding input data, and the more the algorithm coefficients, the better the performance, but the higher the computation energy consumption; a lightweight time compression algorithm (LTC) is an effective and simple lossy compression technique, suitable for habitat monitoring, which introduces a small amount of error in each reading limited by a control knob, the larger the margin of the error, the larger the compression savings; marcelloni et al propose an improved Differential Pulse Code Modulation (DPCM) scheme to compress the sensing data, which has better compression effect than LTC; the Compressed Sensing (CS) method proposed by Donoho provides a new direction for data compression in a wireless sensor network, when original data are sparse on the basis, the CS method can recover a large amount of original data by using less measurement, due to the use of a sparse binary matrix, the CS can greatly reduce the system cost, CS requires that signals are sparse or compressible at a certain level, otherwise, the signals cannot be reconstructed; there are currently many signal recovery algorithms for fast reconstruction and reliable accuracy, such as base-tracking (BP), Orthogonal Matching Pursuit (OMP) and segmented OMP (StOMP), BP having high computational complexity and not being usable for large-scale applications, OMP and StOMP using a bottom-up approach in signal recovery, which is much less complex than BP, but they require more measurements and lack recovery guarantees.
Machine Learning (ML) is an artificial intelligence technique with excellent mathematical fitting capabilities. In recent years, Convolutional Neural Networks (CNNs) have shown remarkable capabilities in various fields, facilitating the widespread use of ML in various fields, Variational Automatic Encoders (VAEs) and constrained boltzmann machines (RBMs) are data generation models designed from ML, which use computational methods to improve model performance by detecting and describing consistency and patterns in training data, CNNs can extract deeper and richer data hiding information through multi-layer iterative convolution; the Principal Component Analysis (PCA) algorithm is a dimensionality reduction technique from machine learning and can be used for compressing cluster data of a cluster head of a wireless sensor network. Masiero et al combines PCA compression with data aggregation to combine information from many sensor nodes; mouswavi et al discusses the combination of the ML method and CS, which uses a feed-forward deep neural network structure to assist in CS signal reconstruction; qiu et al propose a data compression algorithm that combines a Stacked Auto Encoder (SAE) with a cluster routing protocol; liu et al studied the combination of a Restricted Boltzmann Machine (RBM) and an auto-encoder and proposed a Stacked RBM-AE compression scheme. Large deep convolutional networks have also been applied to data compression, but most of the research is currently limited to the field of image compression, most people use RBM or full connectivity layer for wireless sensor networks, and few people research into compressing sensing data using convolutional networks. One important reason for the lack of research is that the computation of the deep convolutional network is very computationally expensive and difficult to apply to sensor nodes with limited computational power, Yildirim et al use the deep convolutional network to compress electrocardiogram signals, but the network requires a large amount of computation, and in conclusion, it is necessary to find an efficient and energy-saving compression model using convolutional neural network computation to apply to WSN data compression.
Disclosure of Invention
An object of the present invention is to solve at least the above problems and to provide at least the advantages described later.
The invention also aims to provide an efficient and energy-saving data compression method for the satellite-borne wireless sensor network, which effectively reduces the node communication energy consumption, the storage energy consumption and the calculation energy consumption of the wireless sensor network.
To achieve these objects and other advantages in accordance with the present invention, there is provided an energy-efficient data compression method for a wireless sensor network on a satellite, comprising the steps of:
step one, collecting flow data of each sensor node of a wireless sensor network;
secondly, preprocessing the streaming data;
the method specifically comprises the following steps: reserving temperature data with a temperature interval between-5 ℃ and 45 ℃, removing abnormal temperature data by a triple standard deviation method, and then mapping the removed reserved temperature data to a [0,1] interval in a normalized mode;
step three, constructing a calculation method of a D-CRBM network calculation layer, wherein the D-CRBM is a convolution RBM with down sampling;
the method specifically comprises the following steps: input size Xw×Xh×CiVia CRBM calculation, and then performing maximum pooling (max-pooling) on the CRBM calculation result, wherein, the convolution step setting is the same as the size of the convolution kernel during the CRBM calculation, the size of the pooled region is 2 multiplied by 1, the convolution step setting is the same as the width of the convolution kernel, the redundancy caused by the repeated calculation of the same data by the convolution kernel can be reduced, after max-posing, constructing an index matrix to store the max-posing result to form a position index of the max-posing result, wherein the index matrix is used for recovering data before maximum pooling when reconstructing the data, the index matrix is a binary matrix and consists of 0 and 1, the size of the index matrix is the same as the size of the output size of the maximum pooling, for the reconstruction part of the maximum pooling, sequentially recovering data before the maximum pooling according to the value in the maximum pooling output and the corresponding index value in the index matrix;
step four, combining the D-CRBM network computing layer with a variational hybrid encoder to construct a CBN-VAE network;
the method specifically comprises the following steps: sequentially using a plurality of D-CRBM network computing layers, a maximum pooling layer and a full-link layer for coding, and generating the input of a decoding network by the mean value and the variance of Gaussian distribution output by the coding network through variational sampling, wherein the decoding network of the CBN-VAE is obtained by turning over the coding network to form the CBN-VAE network;
training a CBN-VAE network to obtain model parameters and constructing a compression model;
the method specifically comprises the following steps: taking the preprocessed flow data forming sequence as a training data set of the CBN-VAE network, and finely adjusting the weight parameters of the deep learning model by using a BP algorithm to obtain model parameters;
and step six, compressing the data of the wireless sensor network by adopting a compression model.
Preferably, the streaming data in the first step is temperature streaming data, the temperature streaming data is ambient temperature data information collected by a sensor node, and the sensor node collects a timestamp every 31 seconds.
Preferably, after the fifth step, before the sixth step, the method further comprises: and detecting the data reconstruction precision of the trained CBN-VAE network.
Preferably, after the fifth step, before the sixth step, the method further comprises: verifying the migration learning capability of the compression models on different sensor nodes, wherein the method comprises the following steps: and applying the compression model of one sensor node to other sensor nodes to reconstruct data, and detecting the difference value between the reconstructed data on the other sensor nodes and the reconstructed data by using the self compression models of the other sensor nodes.
Preferably, after the fifth step, before the sixth step, the method further comprises: carrying out importance evaluation on neurons in the CBN-VAE network, cutting off the neurons with low importance, and simplifying the network, wherein the importance evaluation method comprises the following steps: and calculating the importance scores of all neurons in the CBN-VAE network, marking the neurons with the neuron importance scores lower than a pruning threshold value as the neurons with low importance by the neurons, and clipping the neurons.
The invention at least comprises the following beneficial effects: the invention effectively reduces the node communication energy consumption, the storage energy consumption and the calculation energy consumption of the wireless sensor network, compared with the traditional convolution, the designed D-CRBM network calculation layer can effectively reduce the parameter and the calculation amount of the convolution network, and the CBN-VAE network has good compression ratio and robustness and improves the life working cycle of the wireless sensor network.
Additional advantages, objects, and features of the invention will be set forth in part in the description which follows and in part will become apparent to those having ordinary skill in the art upon examination of the following or may be learned from practice of the invention.
Drawings
FIG. 1 is a schematic flow chart of a data compression method according to an embodiment of the present invention;
FIG. 2 is a diagram illustrating a process of temperature flow data of one of the sensor nodes according to one embodiment of the present invention;
FIG. 3 is a D-CRBM network computing layer according to one embodiment of the present invention;
FIG. 4 is a diagram of a data recovery scheme after maximum pooling and pooling dimensionality reduction in one embodiment of the present invention;
FIG. 5 is a CBN-VAE network structure according to one embodiment of the present invention;
FIG. 6 is a reconstruction result of a CBN-VAE network according to one embodiment of the present invention;
FIG. 7 illustrates the migratory learning capabilities of a CBN-VAE network according to one embodiment of the present invention;
fig. 8 is a schematic diagram of a reconstruction error change of a network after neuron trimming is performed on a CBN-VAE network at different ratios by using multiple trimming methods according to one embodiment of the present invention.
Detailed Description
The present invention is further described in detail below with reference to the drawings and examples so that those skilled in the art can practice the invention with reference to the description.
The embodiment provides a high-efficiency energy-saving satellite-borne wireless sensor network data compression method, as shown in fig. 1, including the following steps:
step one, collecting streaming data of each terminal sensing node of a wireless sensing network;
the flow data is temperature flow data, the temperature flow data is environment temperature data information collected by the sensor node, and the sensor node collects a timestamp every 31 seconds;
the temperature flow data of the present embodiment is derived from 3308442 temperature data collected by the research team of the wireless sensor network of california university from 54 sensor nodes placed in the laboratory from 2.28.4.5.2004.
Secondly, preprocessing the streaming data;
the flow data is temperature flow data, and the preprocessing method of the temperature flow data comprises the following steps: the temperature data of the sensor node 7 is used as a training set of the model, abnormal data with a data value smaller than minus 5 ℃ and larger than 45 ℃ are removed, temperature data with a temperature interval between minus 5 ℃ and 45 ℃ is reserved, then the abnormal temperature data are removed by a triple standard deviation method, the temperature data reserved after removal are mapped to a [0,1] interval in a normalized mode, as shown in figure 2, the difference of orders of magnitude between the temperature data input to the compression model is reduced through normalization processing, and the algorithm is converged more quickly.
Step three, constructing a down-sampling-convolution RBM, namely a D-CRBM network computing layer, wherein the structure of the D-CRBM network computing layer is shown in figure 3;
the D-CRBM network computing layer is an important unit for learning hidden mathematical characteristics of original sensor network data and reducing network parameters, the D-CRBM is a Convolution RBM (CRBM) with down sampling, the basic structure of the CRBM is a standard RBM, which is an undirected graph structure and consists of two layers: the difference between the input layer v and the hidden layer h, the CRBM and the standard RBM is the method of calculating the neuron states of the input layer and the hidden layer. For the standard RBM, the neuron states of the input layer and the hidden layer are obtained by directly multiplying the input and the weight matrix, and are unique since the number of the weight matrix is 1; in the CRBM, a method of calculating the neuron states of the input layer and the hidden layer is convolution calculation, the standard RBM calculates the neuron states using only one weight matrix, and the CRBM calculates using a plurality of convolution kernels, so that after calculating the plurality of convolution kernels, the CRBM generates a plurality of corresponding neuron states, which is opposite to the state result of the standard RBM, and thus, the convolution result needs to be processed. The present embodiment connects a plurality of convolution results into one and uses it as a neuron state, and retains the undirected graph structure of the CRBM, while minimizing the loss of useful information of convolution, and at the same time, the undirected graph property of the CRBM allows the CRBM to perform bidirectional transmission of information, that is, the convolution input can also be obtained by convolution result calculation, which cannot be realized by the standard convolution layer. For input size Xw×Xh×CiAnd a convolution kernel size of Kw×Kh×Ci×CoThe standard convolutional layer of (1), the size of the convolution result output by the standard convolutional layer is Ow×Oh×CoIn which O isw＝Xw-Kw，Oh＝Xh-KhThe convolution step size is 1 and for a CRBM net meter with the same parameters as the standard convolution layerCalculation layer, convolution result size of CRBM output is Ocw×OhX 1, wherein Ocw＝Ow×CoThe convolution step is 1.
The calculation method of the D-CRBM network calculation layer comprises the following steps: input size Xw×Xh×CiThe network parameters are calculated by the CRBM, and then the maximum pooling (max-pooling) is carried out on the calculation result of the CRBM, wherein the convolution step size setting in the CRBM calculation is the same as the size of a convolution kernel, the size of the pooled area is 2 multiplied by 1, and the convolution step size setting is the same as the width of the convolution kernel, so that the redundancy caused by the repeated calculation of the same data by the convolution kernel can be reduced. After max-pooling, constructing an index matrix to store the max-pooling result to form a position index of the max-pooling result, wherein the index matrix is used for recovering data before maximum pooling when reconstructing the data, the index matrix is a binary matrix and consists of 0 and 1, the use method of the index matrix is as shown in FIG. 4, the size of the index matrix is the same as the size of the output size of the maximum pooling, and for the reconstruction part of the maximum pooling, recovering the data before the maximum pooling sequentially according to the value in the output of the maximum pooling and the corresponding index value in the index matrix.
Step four, constructing a CBN-VAE network by combining a D-CRBM network computing layer;
the method comprises the steps of sequentially using a plurality of D-CRBM network computing layers, a maximum pooling layer and a full connection layer for coding, generating the input of a decoding network by the mean value and the variance of Gaussian distribution output by the coding network through variational sampling, turning over the decoding network part of the CBN-VAE from the coding network part, wherein the parameters of the decoding network are the same as the parameters of the coding network, and constructing the coding and decoding part of the CBN-VAE as shown in figure 5.
Training a CBN-VAE network to obtain model parameters and constructing a compression model;
when the CBN-VAE network is trained, the temperature sequence normalized by the sensor node 7 in the first step is used as a training data set of the CBN-VAE network, the CBN-VAE network is trained by using a BP algorithm to obtain model parameters, a compression model is constructed according to the model parameters, and the BP algorithm flow is as follows:
in the algorithm, xiIs an input training data set, xi' is the network output,. mu.2And log σ2Is the mean and variance of the output of the coding network.
And step six, compressing the data of the wireless sensor network by adopting a compression model.
After the fifth step, before the sixth step, three operations of detecting the reconstruction precision of the trained CBN-VAE network, verifying the transfer learning capability of the model parameters, further optimizing the CBN-VAE network and the like are also included, and the three operations are respectively as follows:
(1) detecting data reconstruction accuracy of trained CBN-VAE network
Selecting temperature data of the sensor node 7 as a test set, iterating the CBN-VAE network training for 50 times, and after the CBN-VAE network training is completed, selecting 6000 sample points from the test set of the temperature data of the sensor node 7 to test the performance of the CBN-VAE network;
the test result is shown in fig. 6, the dotted line in fig. 6 is the original temperature data in the test set, and the solid line is the temperature data reconstructed by the CBN-VAE network, although the CBN-VAE network inputs the total number of the sampling points in the test set in a small batch of 120 sampling points for training, after feature extraction is performed by convolution kernels of different sizes, the CBN-VAE network can still separately and approximately fit the values of all the sampling points, thereby avoiding the reduction of the fitting performance of the network to a single sampling point due to the oversize size of the input data; in this embodiment, the reconstruction results of all sampling points in the test set of the sensor node 7 are recorded, the total number of the sampling points in the test set is 8520, the maximum reconstruction error of the CBN-VAE network is 1.2301 ℃, the minimum reconstruction error is less than 0.0001 ℃, the average reconstruction error is 0.0678 ℃ for all the sampling points in the test set of the sensor node 7, for the 8520 sampling points, the reconstruction error values of most of the sampling points are less than 0.1 ℃, only 18 sampling points with reconstruction errors exceeding 1.0 ℃ are provided, only 868 sampling points with reconstruction errors exceeding 0.1 ℃, the reconstructed data closely approach the trend and the numerical value of the original data, and it is indicated that the CBN-VAE network has high reconstruction accuracy.
(2) Verifying transfer learning capability of compression model on different sensor nodes in the embodiment
In order to verify the generalization performance of the CBN-VAE network and the spatial correlation among different nodes, training is carried out by adopting temperature data of the sensor node 7 to obtain model parameters, then other sensor nodes in the wireless sensor network are tested by using a compression model of the model parameters, which is called migration learning in deep learning, meanwhile, each sensor node in the wireless sensor network is trained independently, namely each sensor node has the model parameters corresponding to the flow data of the sensor node, the optimal reconstruction error of the CBN-VAE network at each sensor node is calculated, the model parameters of the sensor node 7 and the sensor node 7 are applied to the reconstruction errors of each sensor node obtained by other sensor nodes as comparison, and the migration learning capability of the compression model corresponding to the model parameters on different nodes is verified;
the CBN-VAE network is iteratively trained for 50 times, the experimental result is shown in FIG. 7, the optimal compression performance is obtained at the sensor node 2, the average reconstruction error is 0.0387 ℃, the model parameters obtained by training the temperature data of the sensor node 7 can be directly used for compression models of other sensor nodes even if the sensor node is not trained independently, and meanwhile, the hidden mathematical features of deep learning of the CBN-VAE network are common to adjacent data of the same class, and the CBN-VAE network has good compression performance for all the sensor nodes. For reconstruction errors of all sensor nodes, most of CBN-VAE network reconstruction errors using corresponding parameters of the sensor nodes are lower than 0.1 ℃; for the network parameters using the sensor node itself, the compression performance is usually slightly better than the network parameters using the sensor node 7; for all the sensor nodes, the line where the circle mark is located in fig. 7 represents the reconstruction data of the CBN-VAE network to which the sensor node 7 is applied, and the line where the triangle mark is located represents the reconstruction data of the CBN-VAE network corresponding to the self node, and the minimum error and the maximum error of the numerical difference are 0.0003 ℃ and 0.067 ℃; for nodes located near the sensor node 7, such as the sensor nodes 4-10, a CBN-VAE network corresponding to CBN-VAE network parameters which do not correspond to the sensor node is used, and reconstruction errors are not reduced remarkably; the results prove that the CBN-VAE network has good transfer learning capability, and when the CBN-VAE network is applied, only a compression model of one sensor node needs to be trained, and the sensor node is applied to all other sensor nodes, so that the calculation consumption during the training of the sensor node can be further reduced.
(3) And (4) carrying out importance evaluation on the neurons in the CBN-VAE network, and cutting off the neurons with low importance to obtain a simplified network.
For the CBN-VAE network, network parameters are usually redundant, and the CBN-VAE network needs to be further simplified, in the embodiment, a neuron pruning method is used to further reduce the number of network parameters and the calculation consumption, the embodiment equates the network parameter pruning problem with the neuron classification problem, and classifies all neurons in the network into two types: the method for pruning the neurons is guided by the importance of the neurons to the whole neural network, and compared with other neuron pruning methods, the method for pruning the neurons needs to take the importance of the neurons to the whole neural network as a guide, and the flow of a pruning algorithm is as follows:
ACC in an algorithmoIs the original network reconstruction error, ACCpIs when N isi,jReconstructing errors of the pruned network, calculating a pruning threshold, recording the importance scores of neurons outside the pruning threshold as neurons with low importance, and pruning the neurons to obtain a simplified network;
comparing the trimming result of the neuron trimming method of the embodiment with other common methods under the same trimming rate, recording reconstruction errors corresponding to the CBN-VAE network trimming at different trimming rates, wherein the result of the compression model reconstruction errors is shown in FIG. 8, the temperature data of the sensor node 7 is trained by using a compression model, the trimmed CBN-VAE network obtained by each trimming method is retrained, the iteration times of the trimming and trimming process are 5 times, Random represents Random trimming, and Mag represents trimming according to the amplitude of the weight in FIG. 8;
as can be seen from fig. 8, when the clipping rate is 50%, the reconstruction error of the clipping method of the present embodiment is 0.0971 ℃, and the network reconstruction errors of Random and Mag are 0.6145 ℃ and 0.3624 ℃, respectively; when the pruning rate is 80%, the network reconstruction error of the pruning method of the embodiment is only 0.3032 ℃, and the reconstruction errors of Random and Mag both exceed 1.5 ℃; the result shows that the pruning method of the embodiment has obvious advantages over the Random and Mag methods, the pruning method of the embodiment can accurately identify the redundant neurons in the CBN-VAE network, 40% of network parameters pruned by using the pruning method of the embodiment do not affect the reconstruction precision of the CBN-VAE network, and the network parameters and the calculation consumption can be further reduced by using the pruning method of the embodiment to prune the CBN-VAE network.
In summary, this embodiment provides an efficient and energy-saving data compression method for a satellite-borne wireless sensor network, which includes combining a high-efficiency convolution structure D-CRBM with a down-sampling function with a variational hybrid encoder (VAE) to construct a data compression network CBN-VAE, performing iterative training using a processed sensing data set as a training set of the network, improving data compression ratio and reconstruction accuracy, and performing importance evaluation on neurons in the network to cut off low-importance neurons, thereby obtaining a cut-out simplified network. The method effectively reduces the node communication energy consumption, the storage energy consumption and the calculation energy consumption of the wireless sensor network, compared with the traditional convolution, the designed D-CRBM can effectively reduce the parameter and the calculation amount of the convolution network, the CBN-VAE has good compression ratio and robustness, the life working cycle of the wireless sensor network is improved, and the provided neuron pruning method can better prune the neural network.
While embodiments of the invention have been described above, it is not limited to the applications set forth in the description and the embodiments, which are fully applicable to various fields of endeavor for which the invention may be embodied with additional modifications as would be readily apparent to those skilled in the art, and the invention is therefore not limited to the details shown and described herein without departing from the generic concept as defined by the claims and their equivalents.
Claims (5)
1. An efficient and energy-saving satellite-borne wireless sensor network data compression method is characterized by comprising the following steps:
step one, collecting flow data of each sensor node of a wireless sensor network;
secondly, preprocessing the streaming data;
the method specifically comprises the following steps: reserving temperature data with a temperature interval between-5 ℃ and 45 ℃, removing abnormal temperature data by a triple standard deviation method, and then mapping the removed reserved temperature data to a [0,1] interval in a normalized mode;
step three, constructing a calculation method of a D-CRBM network calculation layer, wherein the D-CRBM is a convolution RBM with down sampling;
the method specifically comprises the following steps: input size Xw×Xh×CiThe network parameter of (1) is calculated by CRBM, and then maximum pooling (max-pooling) is carried out on the result of CRBM calculation, wherein, the setting of convolution step length in CRBM calculation is the same as the size of convolution kernel, the size of the pooled area is 2 x 1, the setting of convolution step length is the same as the width of convolution kernel, which can reduce redundancy caused by the repeated calculation of the same data by convolution kernel, after max-pooling, an index matrix is constructed to store the max-pooling result to form a position index of the max-pooling result, the index matrix is used for recovering the data before maximum pooling when reconstructing data, the index matrix is a binary matrix composed of 0 and 1, the size of the index matrix is the same as the size of the maximum pooled output, and for the reconstruction part of maximum pooling, the value in the maximum pooled output and the corresponding index value in the index matrix are dependently calculated according to the value in the maximum pooled output and the corresponding index value in the index matrixRecovering data before maximal pooling;
step four, combining the D-CRBM network computing layer with a variational hybrid encoder to construct a CBN-VAE network;
the method specifically comprises the following steps: sequentially using a plurality of D-CRBM network computing layers, a maximum pooling layer and a full-link layer for coding, and generating the input of a decoding network by the mean value and the variance of Gaussian distribution output by the coding network through variational sampling, wherein the decoding network of the CBN-VAE is obtained by turning over the coding network to form the CBN-VAE network;
training a CBN-VAE network to obtain model parameters and constructing a compression model;
the method specifically comprises the following steps: taking the preprocessed flow data forming sequence as a training data set of the CBN-VAE network, and finely adjusting the weight parameters of the deep learning model by using a BP algorithm to obtain model parameters;
and step six, compressing the data of the wireless sensor network by adopting a compression model.
2. The energy-efficient data compression method for the satellite-borne wireless sensor network according to claim 1, wherein the streaming data in the first step is temperature streaming data, the temperature streaming data is environment temperature data information collected by the sensor node, and the sensor node collects a time stamp every 31 seconds.
3. An energy-efficient satellite-borne wireless sensor network data compression method as claimed in any one of claims 1-2, characterized in that after step five, before step six, further comprising: and detecting the data reconstruction precision of the trained CBN-VAE network.
4. An energy-efficient satellite-borne wireless sensor network data compression method as claimed in any one of claims 1-2, characterized in that after step five, before step six, further comprising: verifying the migration learning capability of the compression models on different sensor nodes, wherein the method comprises the following steps: and applying the compression model of one sensor node to other sensor nodes to reconstruct data, and detecting the difference value between the reconstructed data on the other sensor nodes and the reconstructed data by using the self compression models of the other sensor nodes.
5. An energy-efficient satellite-borne wireless sensor network data compression method as claimed in any one of claims 1-2, characterized in that after step five, before step six, further comprising: carrying out importance evaluation on neurons in the CBN-VAE network, cutting off the neurons with low importance, and simplifying the network, wherein the importance evaluation method comprises the following steps: and calculating the importance scores of all neurons in the CBN-VAE network, marking the neurons with the neuron importance scores lower than a pruning threshold value as the neurons with low importance by the neurons, and clipping the neurons.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910700088.0A CN110446173B (en) | 2019-07-31 | 2019-07-31 | Efficient and energy-saving satellite-borne wireless sensor network data compression method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910700088.0A CN110446173B (en) | 2019-07-31 | 2019-07-31 | Efficient and energy-saving satellite-borne wireless sensor network data compression method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110446173A CN110446173A (en) | 2019-11-12 |
| CN110446173B true CN110446173B (en) | 2021-08-27 |
Family
ID=68432394
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910700088.0A Expired - Fee Related CN110446173B (en) | 2019-07-31 | 2019-07-31 | Efficient and energy-saving satellite-borne wireless sensor network data compression method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110446173B (en) |
Families Citing this family (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111630787B (en) * | 2020-04-07 | 2022-12-20 | 东莞理工学院 | MIMO multi-antenna signal transmission and detection technology based on deep learning |
| CN112153659B (en) * | 2020-08-21 | 2022-05-03 | 中国地质大学（武汉） | Efficient and energy-saving construction method of data compression model of satellite-borne wireless sensor network |
| CN112192318B (en) * | 2020-09-28 | 2022-07-19 | 北京天泽智云科技有限公司 | Machining tool state monitoring method and system |
| CN112532400B (en) * | 2020-11-27 | 2022-06-07 | 中国工程物理研究院电子工程研究所 | Single-layer plane ad hoc network method based on UDP multicast communication protocol |
| CN117394353B (en) * | 2023-12-08 | 2024-05-14 | 国网天津市电力公司电力科学研究院 | Power distribution network load transferring and recovering method and device |
Family Cites Families (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105245608B (en) * | 2015-10-23 | 2018-05-08 | 同济大学 | The screening of car networking network node and its access route constructing method based on autoencoder network |
| US10432953B2 (en) * | 2016-02-05 | 2019-10-01 | Deepmind Technologies Limited | Compressing images using neural networks |
| JP6318211B2 (en) * | 2016-10-03 | 2018-04-25 | 株式会社Ｐｒｅｆｅｒｒｅｄ Ｎｅｔｗｏｒｋｓ | Data compression apparatus, data reproduction apparatus, data compression method, data reproduction method, and data transfer method |
| CN107977703A (en) * | 2016-10-21 | 2018-05-01 | 辉达公司 | For trimming neutral net to realize the system and method for the effective reasoning of resource |
| US12073308B2 (en) * | 2017-01-04 | 2024-08-27 | Stmicroelectronics International N.V. | Hardware accelerator engine |
| CN109525598A (en) * | 2018-12-26 | 2019-03-26 | 中国地质大学（武汉） | A kind of fault-tolerant compression method of wireless sense network depth and system based on variation mixing |
| CN109766843A (en) * | 2019-01-14 | 2019-05-17 | 河海大学常州校区 | EMG Feature Extraction based on improved limited Boltzmann machine |
-
2019
- 2019-07-31 CN CN201910700088.0A patent/CN110446173B/en not_active Expired - Fee Related
Non-Patent Citations (2)
| Title |
|---|
| Compressed auto-encoder building block for deep learning network;Qiying Feng 等;《2016 3rd International Conference on Informative and Cybernetics for Computational Social Systems (ICCSS)》;20161231;全文 * |
| 基于受限玻尔兹曼机的深度学习模型及其应用;张艳霞;《中国优秀硕士学位论文全文数据库（电子期刊）信息科技辑》;20170228;全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110446173A (en) | 2019-11-12 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110446173B (en) | Efficient and energy-saving satellite-borne wireless sensor network data compression method | |
| CN110532859B (en) | Remote sensing image target detection method based on deep evolution pruning convolution net | |
| Li et al. | An energy-efficient data collection scheme using denoising autoencoder in wireless sensor networks | |
| CN109670548B (en) | Multi-size input HAR algorithm based on improved LSTM-CNN | |
| Zhang et al. | Self-organizing map as a new method for clustering and data analysis | |
| CN112200244A (en) | Intelligent detection method for anomaly of aerospace engine based on hierarchical countermeasure training | |
| CN110519233B (en) | Satellite-borne sensor network data compression method based on artificial intelligence | |
| CN109884419A (en) | A kind of wisdom grid power quality on-line fault diagnosis method | |
| CN107786959A (en) | The compressed data collection method based on adaptive measuring in wireless sensor network | |
| CN112819082B (en) | Satellite spectrum sensing data reconstruction method based on deep learning | |
| CN113065649A (en) | Complex network topology graph representation learning method, prediction method and server | |
| CN106375940A (en) | Agricultural perception data sparse vector acquiring and space coupling method | |
| CN115662142B (en) | Deep learning traffic state estimation method and system based on sparse information of Internet of vehicles | |
| CN114820828A (en) | Vision transform-based image compressed sensing reconstruction method and system | |
| CN117034189A (en) | Multi-source internet of things data fusion method | |
| CN112749663A (en) | Agricultural fruit maturity detection system based on Internet of things and CCNN model | |
| CN116863247A (en) | Multi-mode remote sensing data classification method integrating global information and local information | |
| CN112215490B (en) | Power load cluster analysis method based on correlation coefficient improved K-means | |
| CN117237432A (en) | Mask uncertainty hyperspectral image reconstruction method based on self-attention | |
| CN113705695A (en) | Power distribution network fault data identification method based on convolutional neural network | |
| CN114969237B (en) | Automatic address analyzing and matching method for geographic information system | |
| CN116170874A (en) | Robust WiFi fingerprint indoor positioning method and system | |
| CN115631371A (en) | Extraction method of electroencephalogram signal core network | |
| CN114677545A (en) | Lightweight image classification method based on similarity pruning and efficient module | |
| CN112257807B (en) | Dimension reduction method and system based on self-adaptive optimization linear neighborhood set selection |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20210827 |