CN110736968A - A Radar Abnormal State Diagnosis Method Based on Deep Learning - Google Patents

Abstract

The invention discloses a method for diagnosing abnormal state of radar based on deep learning. The characteristic parameters related to each type of fault are extracted by the method; for each type of fault, the characteristic parameter with the largest correlation coefficient among the characteristic parameters is taken as the reconstruction parameter target of the reconstruction model, and the long-short-term memory network LSTM model is used to build the reconstruction model. Use the feature parameters other than the feature parameter with the largest correlation coefficient to fit and reconstruct the feature parameter with the largest correlation coefficient to obtain the reconstructed value; quantify the difference between the reconstructed value and the measured value for each type of fault based on probability Standard; make time interval statistics on the quantitative results of each type of fault, integrate the diagnosis results of different models, obtain real-time diagnosis results of multiple faults and give early warnings, filter out false alarms, and obtain the final diagnosis results.

Description

A Radar Abnormal State Diagnosis Method Based on Deep Learning

technical field

The invention belongs to the technical field of radar systems, in particular to a method for diagnosing abnormal state of radar based on deep learning.

Background technique

The natural weather radar system currently in use is a relatively complex electronic system, including a transmitter subsystem, a receiver subsystem and a servo subsystem. There is no physical connection between the various electronic parameters of each subsystem. There is no physical connection between parameters, which makes it impossible to use traditional expert experience to diagnose and predict faults in natural weather radar systems.

SUMMARY OF THE INVENTION

In view of the above existing technical problems, the present invention provides a method for diagnosing abnormal state of radar based on deep learning.

In order to solve the above-mentioned technical problems, the present invention adopts the following technical scheme:

A method for diagnosing abnormal state of radar based on deep learning is applied to a weather radar system including a transmitter subsystem, a receiver subsystem and a servo subsystem, including the following steps:

Using the historical state data and alarm data of each sub-system of the weather radar system, using the alarm data as a label, classifying the faults, and using the stepwise regression method to extract the characteristic parameters related to each type of fault;

For each type of fault, the characteristic parameter with the largest correlation coefficient among the characteristic parameters is taken as the reconstruction parameter target of the reconstruction model, the long-short-term memory network LSTM model is used to build the reconstruction model, and the characteristic parameters except the characteristic parameter with the largest correlation coefficient are used. Fitting and reconstructing the characteristic parameter with the largest correlation coefficient to obtain the reconstructed value;

Make a probability-based quantification standard for the difference between the reconstructed value and the measured value of each type of fault;

Do time interval statistics on the quantitative results of each type of fault, integrate the diagnosis results of different models, obtain real-time diagnosis results for multiple faults and give early warnings, filter out false alarms, and obtain the final diagnosis results.

Preferably, for each type of fault, performing a probability-based quantification standard on the difference between the reconstructed value and the measured value further includes:

Analyze the distribution of measured values, if there are multiple operating states, it is the superposition of multiple Gaussian distributions;

Assuming that the predicted value and the measured value are independent, if the predicted value obeys a Gaussian distribution, assuming that the radar is operating normally, the measured value should also obey this Gaussian distribution, and the difference between the two obeys the N(0,2σ^2) distribution;

预测值等于测量值，则η＝1，预测值越偏离测量值，η越接近于0。The probability of the difference is normalized to obtain the final quantization result:

The predicted value is equal to the measured value, then η=1, and the more the predicted value deviates from the measured value, the closer η is to 0.

Preferably, for each type of fault, time interval statistics are performed on the quantification results, false alarms are filtered out, and the final diagnosis result obtained further includes:

Set the probability judgment threshold η _o , and η < η _o considers a fault state point;

M consecutive time points are used for time interval statistics. If the fault state point is greater than 0.3m, it is considered that a fault has occurred.

Preferably, the stepwise regression method is a forward introduction method, specifically: first, only one independent variable that alone explains the largest variation of the dependent variable is added to the model, and then another independent variable is added to check the factors that can be explained by the entire model. Whether the variation of the variable increases significantly, iterate repeatedly until there is no independent variable and then meet the conditions for joining the model.

Preferably, the stepwise regression method is a backward elimination method, specifically: put all variables into the model, then try to eliminate one of the independent variables from the model, and check whether the variation of the dependent variable explained by the entire model has changed significantly, The variable that will reduce the amount of explanation the least is eliminated; iteratively iterate until no independent variables meet the conditions for elimination.

Preferably, the stepwise regression method is a two-way elimination method, which is specifically as follows: first, add an independent variable that alone explains the largest variation of the dependent variable into the model, and then try to add another independent variable to test all variables in the entire model. If there is a significant increase in the dependent variable, the independent variable is retained and the variable with insignificant effect is eliminated, and iteratively iterates until an optimal variable combination is finally obtained.

Preferably, the process of building a reconstruction model using a long short-term memory network LSTM model is as follows:

The forget gate Γ _f reads the nonlinear activation a ^<t-1> of the previous cell and the input data x ^<t> of the current LSTM cell, and outputs a value between 0 and 1 for each LSTM cell state c ^< A number in ^t-1> , where 1 means "completely keep" and 0 means "completely discard";

更新的向量c^<t>由输入门和忘记门共同决定；The input gate Γ _u is the sigmoid layer, which is used to determine the value that needs to be updated, and use the tanh layer to create a new candidate value vector

The updated vector c ^<t> is jointly determined by the input gate and the forget gate;

The output gate Γ _o determines the output value, runs the sigmoid layer Γ to determine which part of the LSTM cell state will be output, processes the cell state c ^<t> through tanh to get a value between -1 and 1, and converts it It is multiplied by the output of the output gate Γ _o , and the final output determines the part of the output.

The invention has the following beneficial effects: for the failure of each sub-system of the natural weather radar, multi-fault diagnosis can be carried out, the combination of various failures can be judged, the possible failure can be predicted in advance, and the existing radar can be frequently and invalidly alarmed to filter out false alarms.

Description of drawings

FIG. 1 is a flowchart of steps of a method for diagnosing abnormal state of radar based on deep learning according to an embodiment of the present invention;

FIG. 2 is a flowchart of the diagnosis steps of the radar transmitter subsystem in the deep learning-based radar abnormal state diagnosis method according to the embodiment of the present invention

FIG. 3 is a schematic structural diagram in which the operating state of the diagnosis of the radar transmitter subsystem is the superposition of a plurality of Gaussian distributions;

FIG. 4 is a schematic flow chart of the LSTM algorithm for the diagnosis of the radar transmitter subsystem;

Figure 5 is a schematic diagram of the LSTM structure for the diagnosis of the radar transmitter subsystem.

Detailed ways

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

An embodiment of the present invention provides a deep learning-based radar abnormal state diagnosis method, which is applied to a weather radar system including a transmitter subsystem, a receiver subsystem, and a servo subsystem, including the following steps:

Using the historical state data and alarm data of each sub-system of the weather radar system, using the alarm data as a label, classifying the faults, and using the stepwise regression method to extract the characteristic parameters related to each type of fault;

For each type of fault, the characteristic parameter with the largest correlation coefficient among the characteristic parameters is taken as the reconstruction parameter target of the reconstruction model, the long-short-term memory network LSTM model is used to build the reconstruction model, and the characteristic parameters except the characteristic parameter with the largest correlation coefficient are used. Fitting and reconstructing the characteristic parameter with the largest correlation coefficient to obtain the reconstructed value;

Make a probability-based quantification standard for the difference between the reconstructed value and the measured value of each type of fault;

Do time interval statistics on the quantitative results of each type of fault, integrate the diagnosis results of different models, obtain real-time diagnosis results for multiple faults and give early warnings, filter out false alarms, and obtain the final diagnosis results.

Referring to FIG. 1, it is a schematic flow chart corresponding to the above-mentioned deep learning-based radar abnormal state diagnosis method according to the above-mentioned embodiment of the present invention. The equipment state time series data of each sub-system of the weather radar system is collected, and the first type of fault feature parameter extraction is performed respectively, The second type of fault feature parameter extraction, to the nth type of fault feature parameter extraction, and then for each type of fault, the feature parameter with the largest correlation coefficient among the feature parameters is taken as the reconstruction parameter target of the reconstruction model, and the long short-term memory network LSTM is used. The model builds a reconstruction model, obtains reconstruction model 1, reconstruction model 2 to reconstruction model N, and makes a probability-based quantification standard for the difference between the reconstructed value and the measured value of each type of fault, and obtains the classification result of fault discrimination. , that is, the first type of fault discrimination result, the second type of fault discrimination result to the Nth type of fault discrimination result, and then the overall identification result of the equipment status of the entire weather radar system is obtained according to the discrimination results of various types of faults.

The following takes the radar transmitter subsystem as an example to further describe the abnormal state diagnosis process, so that those skilled in the art can better understand the implementation process of the method of the present invention.

Referring to Fig. 2, shown is a flowchart of the steps of the fault diagnosis method of the radar transmitter subsystem, which includes the following steps:

S10, use the historical state data and alarm data of the radar transmitter subsystem, use the alarm data as a label, and use the stepwise regression method to extract a total of n characteristic parameters related to the fault of the transmitter subsystem. It is considered that if the characteristic parameters have changed significantly, Then the health state of the transmitter has changed, because the characteristic parameters are the key information extracted from the state variables of the system. These key information can accurately represent the state of the system. If the operating state of the system changes, it is directly reflected in the state variables. In terms of changes in , the characteristic parameters show these changes more significantly.

The historical state data may include the radar state data of each site in the country in the historical year, and a data point is collected at regular intervals, and each data point contains several variable parameters. Alarm data includes alarm time, alarm classification number. In a specific application example, the characteristic parameters may include: horizontal channel antenna peak power, transmitter peak power, transmit and antenna power ratio, horizontal channel antenna power zero adjustment, transmitter power zero adjustment, expected reflectivity value, short pulse system calibration constant, Long pulse system calibration constant, speed expectation value 4, speed measurement value 4, KD calibration expectation value, KD calibration measurement value, power before horizontal channel filtering, power after horizontal channel filtering.

In S10, the stepwise regression method may adopt the forward selection method (Forward selection), the backward elimination method (Backward elimination) and the bidirectional elimination method (Bidirectional elimination). The specific implementation process of the stepwise regression method is further described below.

If you use the forward introduction method, first add only one independent variable that alone explains the largest variation of the dependent variable to the model, and then try to add another independent variable to check whether the variation of the dependent variable explained by the entire model increases significantly (F-test, t-test, etc.), iterate repeatedly until no independent variables meet the conditions for joining the model.

If you use the backward elimination method, as opposed to the forward introduction method, at this time, all variables are put into the model, then try to eliminate one of the independent variables from the model, check whether the variation of the dependent variable explained by the whole model has changed significantly, and put The variable that reduces the amount of explanation the least is eliminated; iteratively iterates until no independent variables meet the conditions for elimination.

If you use the two-way elimination method, to combine the forward introduction method and the backward elimination method, first add an independent variable that alone explains the largest variation of the dependent variable to the model, and then try to add another independent variable to all variables in the entire model. If there is a significant increase in the dependent variable, the independent variable is retained and the variable with insignificant effect is eliminated, and iteratively iterates until an optimal combination of variables is finally obtained.

S20, take the characteristic parameter y with the largest correlation coefficient among the characteristic parameters (generally the transmitter peak power or the horizontal channel antenna peak power) as the reconstruction parameter target of the reconstruction model. Use the Long Short Term Memory Network (LSTM) model to build a reconstruction model, and use a total of n-1 feature parameters other than the feature parameter y to fit and reconstruct y to obtain the reconstructed value

输出门：Γ_o＝σ(W_o[a^<t-1>，x^<t>]+b_o)；更新值：非线性激活：a^<t>＝Γ_o×tanhc^<t>。其计算过程如下：As shown in Figure 4 and Figure 5, when using the LSTM model to build the reconstruction model, the input time series state data X={X ⁽¹⁾ ,...,X ^(m) }, through the long and short-term memory network LSTM, the output reconstruction parameters Value Y={Y ⁽¹⁾ ,...,Y ^(m) }. Where LSTM includes forget gate: Γ _f =σ(W _f [a ^<t-1> , x ^<t> ]+b _f ); input gate (update gate): Γ _u =σ(W _u [a ^{<t- 1>} , x ^<t> ]+b _u ); update the candidate value:

Output gate: Γ _o =σ(W _o [a ^<t-1> , x ^<t> ]+b _o ); update value: Nonlinear activation: a ^<t> =Γ _o ×tanhc ^<t> . Its calculation process is as follows:

The forget gate Γ _f will read the nonlinear activation a ^<t-1> of the previous cell and the input data x ^<t> of the current LSTM cell, and output a value between 0 and 1 for each state c in the LSTM cell Numbers in ^<t+1> . 1 means "completely reserved", 0 means "completely discarded".

更新的向量c^<t>由输入门和忘记门共同决定。The input gate Γ _u is a sigmoid layer that decides what values will be updated. Then, use a tanh layer to create a new vector of candidate values

The updated vector c ^<t> is jointly determined by the input gate and the forget gate.

The output gate _Γ determines the output value. First, a sigmoid layer Γ is run to determine which part of the LSTM cell state will be output. Then, the cell state c ^<t> is processed through tanh (to get a value between -1 and 1) and multiplied by the output of the output gate Γ _o , and finally only the part that determines the output is output.

和实测值y的差值做基于概率的量化标准。S30, on the reconstructed value

The difference from the measured value y is used as a probability-based quantitative standard.

In a specific application example, S30 may further include:

ⅰ. Analyze the distribution of y. If there are multiple operating states, it is the superposition of multiple Gaussian distributions. As shown in Figure 3, the distribution of the transmitter peak power is the superposition of two Gaussian distributions, less than 300 obeys (1) distribution, and greater than 300 obeys (2) distribution.

ii. Assuming that the predicted value and the measured value are independent, if the predicted value obeys the (1) distribution, assuming that the radar is operating normally, the measured value should also obey the (1) distribution, and the difference between the two obeys N(0,2σ^2) distributed.

iii. Normalize the probability of the difference to obtain the final quantization result: The predicted value is equal to the measured value, then η=1, and the more the predicted value deviates from the measured value, the closer η is to 0.

S40, perform time interval statistics on the quantification results, filter out false alarms, and obtain a final diagnosis result.

In a specific application example, S40 may further include:

ⅰ. Set the probability judgment threshold η ₀ , η < η ₀ considers the transmitter to have a fault state point

ⅱ. Do time interval statistics with m consecutive time points, if the fault state point is greater than 0.3m, it is considered that there is a fault.

It should be understood that the exemplary embodiments described herein are illustrative and not restrictive. Although one or more embodiments of the invention have been described in conjunction with the accompanying drawings, those of ordinary skill in the art will appreciate that various changes can be made without departing from the spirit and scope of the invention as defined by the appended claims. Changes in form and detail.

Claims (7)

1. The method for diagnosing the abnormal state of the radar based on deep learning is characterized by being applied to a meteorological radar system comprising a transmitter subsystem, a receiver subsystem and a servo subsystem and comprising the following steps of:

   the method comprises the steps of utilizing historical state data and alarm data of all subsystems of a meteorological radar system, using the alarm data as labels, classifying faults, and extracting characteristic parameters related to each types of faults by using a stepwise regression method;

   taking the characteristic parameter with the maximum correlation coefficient in the characteristic parameters of each -type faults as a reconstruction parameter target of a reconstruction model, building the reconstruction model by using a long-time memory network (LSTM) model, and performing fitting reconstruction on the characteristic parameter with the maximum correlation coefficient by using the characteristic parameters except the characteristic parameter with the maximum correlation coefficient to obtain a reconstruction value;

   making probability-based quantification standard for the difference value between the reconstructed value and the measured value of each types of faults;

   and carrying out time interval statistics on the quantitative result of each type fault, integrating the diagnosis results of different models to obtain a plurality of real-time fault diagnosis results, giving early warning, filtering false alarms and obtaining a final diagnosis result.
2. 2. The deep learning-based radar abnormal state diagnosis method of claim 1, wherein the probability-based quantization criterion for the difference between the reconstructed value and the measured value for each types of faults further steps comprise:

   analyzing the distribution condition of the measured values, and if a plurality of running states exist, superposing the plurality of Gaussian distributions;

   assuming that the predicted value and the measured value are independent, if the predicted value obeys Gaussian distribution, and assuming that the radar operating state is normal, the measured value should obey the Gaussian distribution, and the difference value between the two obeys N (0,2 sigma ^2) distribution;

   and (3) performing grouping processing on the probability of the difference to obtain a final quantization result:

   

   if the predicted value is equal to the measured value, η is equal to 1, and the more the predicted value deviates from the measured value, the closer η is to 0.
3. 3. The deep learning-based radar abnormal state diagnosis method as claimed in claim 1 or 2, wherein for each types of faults, time interval statistics is performed on the quantized result, false alarms are filtered out, and the step of obtaining a final diagnosis result further comprises:

   setting a probability decision threshold η_o，η＜η₀A fault condition point is considered to occur;

   and (5) performing time interval statistics by using m continuous time points, and determining that a fault occurs when the fault state point is greater than 0.3 m.
4. 4. The deep learning-based radar abnormal state diagnosis method as claimed in claim 1 or 2, wherein the stepwise regression method is a forward introduction method, and is characterized in that only independent variables with the largest interpretation dependent variable variation are added into the model firstly, then another independent variables are added in an attempt, whether the interpretation dependent variable variation of the whole model is increased remarkably or not is checked, and iteration is repeated until no independent variable meets the condition of adding the model.
5. 5. The deep learning-based radar abnormal state diagnosis method as claimed in claim 1 or 2, wherein the stepwise regression method is a backward elimination method, and is characterized in that all variables are put into a model, independent variables are tried to be eliminated from the model, whether variation of an explanation dependent variable of the whole model has significant variation is checked, variables with minimum reduction of the explanation variable are eliminated, and iteration is repeated until no independent variable meets the elimination condition.
6. 6. The method for diagnosing the abnormal state of the radar based on the deep learning of claim 1 or 2, wherein the stepwise regression method is a two-way elimination method, and is characterized in that independent arguments with the largest variation of the dependent variables are added into the model firstly, then additional arguments are added in an attempt, all the variables in the whole model are checked, if the dependent variables are increased remarkably, the independent variables are retained, the variables with the insignificant effects are eliminated, and iteration is repeated until optimal variable combinations are obtained finally.
7. 7. The deep learning-based radar abnormal state diagnosis method as claimed in claim 1 or 2, wherein the process of building the reconstruction model by using the long-time memory network LSTM model comprises the following steps:

   forgetting gamma_fNonlinear activation of cells read a^<t-1>And input data x of the current LSTM cell^<t> values between 0 and 1 are output to each of the LSTM cell states c^<t-1>Wherein 1 represents "completely retained", and 0 represents "completely discarded";

   input Γ_uFor the sigmoid layer, which is used to decide the values that need to be updated, the tanh layer is used to create new candidate value vectors

   

   Updated vector c^<t>Determined by both input and forget ;

   output Γ_oDetermining output value, operatingsigmoid layer Γ to determine which portion of the LSTM cell state will be output, cell state c^<t>Processing by tanh gives values between-1 and adds it to the output Γ_oThe outputs of (a) are multiplied to finally output a part of the determined output.

Images