Fault online prediction method and system based on timing feature selection

Abstract

The application discloses a kind of based on timing feature selection's fault online prediction method and system, the original data collected by module monitoring point is preprocessed, and the state monitoring time series data set of M monitoring points is obtained;Divide training set, validation set and test set;Respectively, the fault feature quantity of each monitoring point of module is established fault prediction model, and the timing feature selection of fault prediction model is carried out, and the best input timing feature of each fault prediction model is obtained;The prediction model after timing feature selection is trained, and the prediction performance of model is improved;Respectively, the fault feature quantity of each monitoring point is predicted online by the trained fault prediction model.The input timing feature of the fault prediction model of the present application is selected, the redundant input feature is reduced, the space complexity of prediction model is reduced, the prediction accuracy of prediction model is improved, has good generalization performance, can provide basis for preventive maintenance of radar equipment.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of radar, and particularly relates to a fault online prediction method and system based on time sequence feature selection. BACKGROUND

[0002] Modern radars have higher and higher degrees of integration and digitization, and their structures are more and more complex, which brings a variety of faults. Once a radar equipment has a fault, it may bring serious consequences. At present, a radar system establishes monitoring points on modules, and monitors fault feature quantities (voltage, current, temperature, etc.) of the modules through the monitoring points. When the fault feature quantities of the modules exceed the set threshold, it is considered that the modules have faults, and real-time alarm information is generated. The radar system generates a large amount of fault feature quantity state monitoring data when performing state monitoring. The state monitoring data itself has time sequence, and if the fault feature quantities of each monitoring point are predicted according to historical time sequence data, it can be judged in advance whether the fault feature quantities will exceed the set threshold, a fault warning is generated in advance, and then a maintenance plan and a maintenance strategy of the equipment are formulated and generated in advance, and the maintenance support efficiency of the equipment is improved.

[0003] The BP neural network model is a multi-layer feedforward network trained according to the error back propagation algorithm. Its learning is divided into two parts: one is the forward propagation of information, the training samples of the input layer are processed by the transfer function and given to the output layer, and the output layer gives the error and result. The other is the back propagation of training error, which distributes the training error to each neuron, and each neuron adjusts the connection weight and threshold value according to the training error. When the number of training times reaches the set value or the training error reaches the target value, the training is completed and the result is output.

[0004] The time sequence feature in the fault prediction model is defined as follows: when the fault feature quantity at time t is predicted by the fault feature quantity at time t-i (i=1, 2, …), the time sequence data of the fault feature quantity at time t-i is called an input time sequence feature of the prediction model. The time sequence feature selection of the prediction model means that the time sequence feature set formed by the input time sequence data of different time lags of the model is selected by an algorithm, the input time sequence features of the prediction model are reduced, the input dimension of the prediction model is reduced, and the prediction accuracy of the model is improved.

[0005] The methods currently used for equipment failure prediction mainly include support vector regression (SVR) model, recurrent neural network (RNN), CNN-BiLSTM-Attention model, BP neural network, least square method, differential wavelet neural network, etc. For the SVR model, the failure prediction methods that can be found in the public literature are as follows: divide the training sample set into training set and test set; train the model by genetic algorithm to obtain the optimized model parameters; establish the SVR model; extract the fault feature quantity, and use the first four time values to predict the value of the next time. For the recurrent neural network model, the failure prediction methods that can be found in the public literature are as follows: data preprocessing is performed on the original data, and then the training set and test set are divided, the RNN network model is constructed, the parameters of the RNN network model are set, the RNN is trained through the training set until the iteration end condition is reached, and finally the trained RNN is tested through the test set to obtain the predicted data. For the CNN-BiLSTM-Attention model, the failure prediction methods that can be found in the public literature are as follows: the working failure log of high-performance computing system is processed and analyzed, the failure log is cleaned to obtain the failure time series data set C, the Kmeans method is used to cluster C to obtain K clusters corresponding to K fault types; then the CNN-BiLSTM-Attention model is established, the K clusters are used to train the failure prediction model, and the trained failure prediction model is used to predict the failure occurrence time and failure node position. For the BP neural network model, the failure prediction methods that can be found in the public literature are as follows: the number of neurons in each layer of the BP neural network model is determined, the neural network model is trained through the training set divided in advance to obtain the optimal parameters, and then the trained BP neural network model is used to predict the failure of the analog circuit. For the least square method, the failure prediction methods that can be found in the public literature are as follows: a quadratic curve fitting mathematical model is used, the least square mathematical method is used to analyze the historical data, and a curve consistent with the historical data is fitted, and then the future health status of the equipment for a period of time is predicted according to the curve.The fault prediction method capable of being found from the public literature for the differential wavelet neural network model is as follows: screening effective fault cases; calculating the fault interval time, constructing the original chaotic interval time sequence; constructing training and prediction samples, performing differential preprocessing on the training and prediction samples, constructing the training differential time sequence and the prediction differential time sequence; constructing the input matrix and the expected output matrix of the training network through phase space reconstruction; constructing the fault prediction algorithm based on the differential wavelet neural network; completing the training of the differential wavelet neural network, constructing a prediction system; inputting the prediction differential time sequence, predicting the latest fault differential time; converting the output differential prediction value into the latest fault interval time prediction value of the software through the differential restoration method; adding the predicted differential time to the prediction differential time sequence, and realizing multi-step fault prediction through iteration.

[0006] In the existing prediction model or method, the value at the next time is directly predicted by the values at the previous time, on the one hand, because the input time sequence characteristics of the model are not selected, the input feature redundancy of the model is caused, and the model space complexity is large; on the other hand, the model trained by using the redundant time sequence characteristics cannot achieve the optimal prediction performance.

[0007] Therefore, it is necessary to design a fault prediction method capable of selecting the input time sequence characteristics of the fault prediction model, reducing the input feature redundancy and improving the prediction performance of the model, which is an important problem to be solved in radar fault prediction. SUMMARY

[0008] The present application aims at solving the problems in the background art, and provides a fault online prediction method and system based on time sequence characteristic selection, which can be used for predicting potential faults of radar equipment.

[0009] In order to achieve the object of the present application, the present application discloses a fault online prediction method and system based on time sequence characteristic selection, which comprises the following steps:

[0010] Step 1: removing the repeated records in the original data set collected through the module monitoring point, when the state monitoring data of the fault characteristic quantity of the monitoring point at a certain time is lost, the average value of the fault characteristic quantity of the adjacent time of the monitoring point is taken as the interpolation value, then the state monitoring data of each monitoring point is sorted in chronological order, and the state monitoring time sequence data set D of the M monitoring points is obtained by normalizing the state monitoring data of each monitoring point in the data set m (m = 1, 2, …, M);

[0011] Step 2: forming a training set by each monitoring point in proportion Validation set Test set

[0012] Step 3: Establish a backpropagation BP neural network prediction model P for the fault characteristic quantities of each monitoring point of the module. m (m=1,2,…,M), the mean absolute percentage error (MAPE) of the prediction model is used as the performance metric of the prediction model. A time series feature selection algorithm is constructed to select time series features for the fault prediction model and obtain the best input time series features for each fault prediction model.

[0013] Step 4: Construct the input vector, actual output, and expected output of the prediction model. Train the fault prediction model selected by time series features. By iteratively updating the connection weights between hidden layer-output layer and input layer-hidden layer neurons, as well as the thresholds of hidden layer and output layer neurons, better model parameters are obtained to improve the prediction performance of the model.

[0014] Step 5: Perform online prediction of the fault characteristics of each monitoring point using the trained fault prediction model to obtain the normalized predicted value of the fault characteristics. Then, through inverse normalization, obtain the actual predicted value of the fault characteristics.

[0015] Furthermore, in step 1, the specific calculation formula for normalizing the status monitoring data of each monitoring point in the dataset is as follows:

[0016]

[0017] in It is the fault characteristic state monitoring value of monitoring point m (m=1,2,…,M) at time t. It is the historical maximum value of the fault characteristic quantity at monitoring point m; It is the normalized value of the fault characteristic quantity state monitoring at monitoring point m at time t.

[0018] Furthermore, in step 1, the state monitoring time series dataset D m It contains 4 fields: module number, monitoring point number, monitoring time, and fault characteristic quantity.

[0019] Furthermore, in step 1, the fault characteristic quantities of the monitoring point at a nearby time include voltage, current, and temperature.

[0020] Furthermore, in step 2, each monitoring point forms a training set. Validation set test set The ratio is 8:1:1.

[0021] Furthermore, in step 3, the BP neural network model consists of three layers of neurons: an input layer, a hidden layer, and an output layer. The number of neurons in the hidden layer and the output layer are 10 and 1, respectively.

[0022] Further, in step 3, the proposed time series feature selection algorithm specifically includes the following sub-steps:

[0023] Step 3-1, forming time series feature set F from time series data of each monitoring point m at time t-i to predict the failure feature quantity at time t, m (m = 1, 2, …, M), wherein represents the time series feature formed by the time series data of the failure feature quantity at time t-i of monitoring point m when predicting the failure feature quantity at time t, wherein Q m represents the number of test set samples of monitoring point m;

[0024] Step 3-2, forming the output sequence Y of the prediction model corresponding to each monitoring point through the training set of each monitoring point m wherein

[0025] Step 3-3, initializing the input parameters of the time series feature selection algorithm, including: time series feature set F m , output sequence Y m , prediction model P m , time series feature subset validation set stop control parameter Z, in the present application, Z takes the value of 6;

[0026] Step 3-4, training the prediction model P using the time series feature subset m , and using the validation set to calculate the mean absolute percentage error E MAPE of the trained prediction model, MAPE taking E m as the performance measure of the prediction model;

[0027] Step 3-5, initializing z = 1, and initializing the optimal time series feature subset

[0028] Step 3-6, comparing the size of z and Z, if z < Z, go to step 3-7, otherwise go to step 3-15;

[0029] Step 3-7, generating a set containing feature subset m and adding one more feature in time series feature set F

[0030] Step 3-8, let a = 0, and calculate the number A of non-empty subsets of set ;

[0031] Step 3-9: Compare the size of a with that of A. If a < A, go to step 3-10; otherwise, go to step 3-14.

[0032] Steps 3-10, in the set Non-repeating subset search via subset Training prediction model P m and using the validation set To calculate the mean absolute percentage error E of the trained prediction model d ;

[0033] Step 3-11, Compare E d With E MAPE The size of E d <E MAPE Proceed to step 3-12 if the above steps are correct, otherwise proceed to step 3-13.

[0034] Step 3-12, let E MAPE =E d a = a + 1, go to step 3-9;

[0035] Step 3-13: Let a = a + 1, then go to step 3-9;

[0036] Step 3-14: Let z = z + 1, then go to step 3-6;

[0037] Step 3-15: Output the optimal temporal feature subset

[0038] Furthermore, in step 4, the input vector of the prediction model m is Where d represents the dimension of the input, i.e., the number of neurons in the input layer of the BP neural network, and d is determined by the temporal feature selection algorithm. The actual output of the prediction model in the g-th iteration is: The expected output of the prediction model m is y m .

[0039] Furthermore, step 4, training the prediction model specifically includes the following sub-steps:

[0040] Step 4-1: Initialize the connection weights between the input layer and the hidden layer. Connection weights between hidden and output layers Hidden layer neuron threshold Output layer neuron threshold Expected training error α m Learning rate λ m Maximum number of iterations G m Activation function of neurons in

[0041]

[0042] Step 4-2: Let n = 1 and g = 1;

[0043] Step 4-3: Select the nth input sample and expected output y m (n) is provided to the prediction model;

[0044] Step 4-4: Calculate the outputs of neurons in the hidden layer and output layer:

[0045]

[0046] Steps 4-5: Using y m (n) and h m (n) Calculate the partial derivatives of the error function with respect to the output layer neurons.

[0047]

[0048] Steps 4-6: Utilize the connection weights between the hidden layer and the output layer Partial derivatives and hidden layer output Calculate the partial derivatives of the error function with respect to the hidden layer neurons.

[0049]

[0050] Steps 4-7: Utilize the partial derivatives of the output layer neurons and hidden layer neuron output To adjust the connection weights and thresholds:

[0051]

[0052] Steps 4-8: Utilize the partial derivatives of hidden layer neurons and input layer neurons input To adjust the connection weights and thresholds:

[0053]

[0054] Step 4-9: If n < N, where N is the number of samples, let n = n + 1 and go to step 4-3; otherwise, go to step 4-10.

[0055] Step 4-10: Calculate the total error E m :

[0056]

[0057] Step 4-11, if Em <α m or g > G m If yes, then the training is finished; otherwise g = g + 1, n = 1, and return to step 4-3.

[0058] To achieve the object of the present application, the present application further discloses a fault online prediction system based on time sequence feature selection, comprising a data processing module, a sample division module, a time sequence feature selection module, a training prediction model module, and a fault feature quantity prediction module.

[0059] The data processing module eliminates repeated records in an original data set collected by the monitoring points, and when the fault feature quantity state monitoring data of the monitoring points at a certain moment is lost, the average value of the fault feature quantity of the monitoring points at adjacent moments is taken as an interpolation value, and then the state monitoring data of each monitoring point is sorted in chronological order, and the state monitoring data of each monitoring point in the data set is normalized to obtain a state monitoring time sequence data set D of M monitoring points m (m = 1, 2, …, M).

[0060] The sample division module proportionally forms a training set a validation set and a test set

[0061] The time sequence feature selection module respectively establishes a back propagation BP neural network prediction model P for the fault feature quantity of each monitoring point of the module m (m = 1, 2, …, M), takes the mean absolute percentage error MAPE of the prediction model as the performance measure of the prediction model, constructs a time sequence feature selection algorithm, performs time sequence feature selection on the fault prediction model, and obtains the best input time sequence feature of each fault prediction model.

[0062] The training prediction model module constructs an input vector, an actual output, and an expected output of the prediction model, trains the fault prediction model after time sequence feature selection, updates the connection weights between the hidden layer-output layer and the input layer-hidden layer neurons and the threshold values of the hidden layer and the output layer neurons through repeated iterations to obtain better model parameters and improve the prediction performance of the model.

[0063] The fault feature quantity prediction module respectively performs online prediction on the fault feature quantity of each monitoring point through the trained fault prediction model, obtains the normalized prediction value of the fault feature quantity, and obtains the actual prediction value of the fault feature quantity through the inverse normalization processing.

[0064] Compared with the prior art, the significant progress of the present application is that the proposed time sequence feature selection algorithm performs feature selection on the input time sequence features of the fault prediction model, reduces the redundant input features, reduces the spatial complexity of the prediction model, improves the prediction accuracy of the prediction model, has good generalization performance, and thus can provide a basis for preventive maintenance of radar equipment.

[0065] To make the functional characteristics and structural parameters of the present application clearer, the following further describes the present application in conjunction with the accompanying drawings and specific embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0066] The accompanying drawings, which are included to provide a further understanding of the present application, constitute a part of this application, illustrate embodiments of the present application and specific examples thereof, and serve to explain the present application, and do not constitute an improper limitation on the present application. In the drawings:

[0067] Figure 1 is a flowchart of the method of the present application;

[0068] Figure 2 is a flowchart of the time sequence feature selection algorithm;

[0069] Figure 3 is a flowchart of the training of the fault prediction model. DETAILED DESCRIPTION

[0070] The technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the accompanying drawings of the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments of the present application; based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0071] The implementation process and flow of the fault online prediction method based on time sequence feature selection are as shown in Figure 1 , and the specific steps are as follows:

[0072] Step 1: pre-process the original data collected through the module monitoring points to obtain the state monitoring time sequence data set D of M monitoring points (m = 1, 2, …, M). m The pre-processing of the original data collected through the monitoring points in step 1 specifically includes the following steps:

[0073] Step 1-1: eliminate duplicate records in the data set;

[0074] Step 1-2: when the state monitoring data of a monitoring point at a certain time is missing, the average value of the fault feature quantity of the adjacent time of the monitoring point is taken as the interpolation value;

[0075] Step 1-3: Sort the state monitoring data of each monitoring point in chronological order respectively;

[0076] Step 1-4: Normalize the state monitoring data of each monitoring point in the data set to obtain the state monitoring time series data set D m (m = 1, 2, …, M), the specific calculation formula is as follows:

[0077]

[0078] Wherein is the state monitoring value of the fault characteristic quantity of monitoring point m (m = 1, 2, …, M) at time t, is the historical maximum value of the fault characteristic quantity of monitoring point m; is the normalized value of the fault characteristic quantity of monitoring point m at time t.

[0079] Wherein the state monitoring time series data set D m contains 4 fields, and the 4 fields are: module number, monitoring point number, monitoring time, and fault characteristic quantity. The fault characteristic quantity includes voltage, current, and temperature.

[0080] Step 2: Divide the training set, validation set and test set, wherein the training set of each monitoring point The ratio of the validation set and the test set is 8:1:1.

[0081] Step 3: Establish a fault prediction model for the fault characteristic quantity of each monitoring point of the module, select the time sequence features of the fault prediction model through the proposed time sequence feature selection algorithm, and obtain the best input time sequence features of each fault prediction model. Wherein the fault prediction model P m (m = 1, 2, …, M) for each monitoring point of the module is a BP neural network model, and the BP neural network model is composed of input layer, hidden layer and output layer. The number of neurons in the hidden layer and the output layer is 10 and 1 respectively. The process of the proposed time sequence feature selection algorithm is as shown in Figure 2 , which specifically includes the following sub-steps:

[0082] Step 3-1: Form time sequence feature set F from the training set m of each monitoring point according to different time lags, Wherein indicates that when monitoring point m predicts the fault characteristic quantity at time t through the fault characteristic quantity at time t-i, the time sequence features formed by the time sequence data of the fault characteristic quantity at time t-i, Wherein Q mThis indicates the number of test set samples at monitoring point m;

[0083] Step 3-2: Using the training set of each monitoring point The output sequence Y of the prediction model corresponding to each monitoring point is formed. m ,in

[0084] Step 3-3: Initialize the input parameters of the time series feature selection algorithm, including: the time series feature set F m Output sequence Y m Prediction model P m Temporal feature subset Validation set The stop control parameter Z is set to 6 in this invention.

[0085] Steps 3-4: Using temporal feature subsets Training prediction model P m and using the validation set To calculate the mean absolute percentage error E of the trained prediction model MAPE E MAPE As a performance metric for predictive models;

[0086] Steps 3-5: Initialize z = 1, initialize the optimal temporal feature subset

[0087] Step 3-6: Compare the size of z with Z. If z < Z, go to step 3-7; otherwise, go to step 3-15.

[0088] Steps 3-7: In the time series feature set F m Generate a subset containing features And add one more feature set

[0089] Steps 3-8: Let a = 0, calculate the set The number A of non-empty subsets;

[0090] Step 3-9: Compare the size of a with that of A. If a < A, go to step 3-10; otherwise, go to step 3-14.

[0091] Steps 3-10: In the set Non-repeating subset search via subset Training prediction model P m and using the validation set To calculate the mean absolute percentage error E of the trained prediction model d ;

[0092] Step 3-11: Compare E dWith E MAPE of size E d < E MAPE , go to step 3-12, otherwise go to step 3-13;

[0093] Step 3-12: Let E E MAPE = E d , a = a + 1, go to step 3-9;

[0094] Step 3-13: Let a = a + 1, go to step 3-9;

[0095] Step 3-14: Let z = z + 1, go to step 3-6;

[0096] Step 3-15: Output the optimal timing feature subset

[0097] Step 4: Train the fault prediction model after timing feature selection to improve the prediction performance of the model. The input vector of the prediction model m is where d represents the dimension of the input (the number of input layer neurons of the BP neural network), d is determined by the timing feature selection algorithm, and the actual output of the gth iteration of the prediction model m is The expected output of the prediction model m is y m . The training process of the prediction model is shown in Figure 3 , which specifically includes the following sub-steps:

[0098] Step 4-1: Initialize the connection weights of the input layer and the hidden layer the connection weights of the hidden layer and the output layer the threshold value of the hidden layer neurons the threshold value of the output layer neurons the expected training error a m , the learning rate l m , the maximum number of iterations G m , the activation function of the neurons where

[0099]

[0100]

[0101] Step 4-2: Let n = 1, g = 1;

[0102] Step 4-3: Select the nth input sample and the expected output y m (n) to provide to the prediction model;

[0103] Step 4-4: Calculate the output of the hidden layer and the output layer neurons:

[0104]

[0105] Step 4-5: Calculate the partial derivative of the error function with respect to the output layer neuron using y m (n) and h m (n) Calculate the partial derivative of the error function with respect to the output layer neuron using y

[0106]

[0107] Step 4-6: Modify the connection weights and thresholds using the partial derivative of the error function with respect to the output layer neuron partial derivative and the output of the hidden layer Calculate the partial derivative of the error function with respect to the hidden layer neuron

[0108]

[0109] Step 4-7: Modify the connection weights and thresholds using the partial derivative of the error function with respect to the output layer neuron and the output of the hidden layer neuron

[0110]

[0111] Step 4-8: Modify the connection weights and thresholds using the partial derivative of the error function with respect to the hidden layer neuron and the input of the input layer neuron

[0112]

[0113]

[0114] Step 4-9: If n < N (N is the number of samples), set n = n + 1 and go to Step 4-3, otherwise go to Step 4-10;

[0115] Step 4-10: Calculate the total error E m :

[0116]

[0117] Step 4-11: If E m < α m or g > G m , then the training is complete; otherwise, set g = g + 1, n = 1, and return to Step 4-3.

[0118] Step 5: Perform online prediction of the fault characteristic quantity for each monitoring point using the trained fault prediction model. Since the fault characteristic quantity is normalized before prediction, the actual predicted value of the fault characteristic quantity is obtained by the following formula:​​

[0119]

[0120] wherein h m (t) is a normalized prediction value of the failure characteristic quantity at time t obtained by the prediction model, is a prediction value of the failure characteristic quantity at time t.

[0121] It is to be noted that the relational terms herein, such as first and second, and the like, are used solely to distinguish one from another entity or action without necessarily requiring or implying any actual relationship or order between or among such entities or actions. Moreover, the terms "comprises", "comprising", or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can also include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0122] While the embodiments of the present application have been illustrated and described, it will be understood by those skilled in the art that various changes, modifications, substitutions, and alterations can be made therein without departing from the spirit and scope of the application, which is defined by the appended claims and their equivalents.

Claims

1. A method for online failure prediction based on time-series feature selection, characterized in that, The online fault prediction method based on time sequence feature selection comprises the following steps: Step 1, remove the duplicate records in the original data set collected by the module monitoring point, when the monitoring point fails at a certain moment, the mean value of the fault characteristic quantity of the adjacent moment of the monitoring point is used as the interpolation value, then the state monitoring data of each monitoring point is sorted in chronological order, and the state monitoring data of each monitoring point in the data set is normalized to obtain the state monitoring time series data set D of M monitoring points m (m = 1, 2, …, M); Step 2, Form training set in proportion to each monitoring point Validation set Test set Step 3, a back propagation BP neural network prediction model P is established for each monitoring point of the module respectively m (m = 1, 2, …, M), taking the mean absolute percentage error MAPE of the prediction model as the performance measure of the prediction model, constructing a time series feature selection algorithm, performing time series feature selection on the fault prediction model, and obtaining the best input time series feature of each fault prediction model; Step 4, input vector, actual output and expected output of the prediction model are constructed, the fault prediction model after time sequence feature selection is trained, the connection weight values between the hidden layer-output layer and the input layer-hidden layer neurons and the threshold values of the hidden layer and the output layer neurons are updated through repeated iterations to obtain better model parameters and improve the prediction performance of the model; Step 5, the fault feature quantity of each monitoring point is predicted online by the trained fault prediction model respectively, the normalized prediction value of the fault feature quantity is obtained, and the actual prediction value of the fault feature quantity is obtained through the inverse normalization processing. 2.The method according to claim 1, characterized in that, In step 1, the specific calculation formula of the normalization processing of the state monitoring data of each monitoring point in the data set is as follows: wherein is a failure characteristic quantity state monitoring value of the monitoring point m (m = 1, 2, …, M) at the t time point, is a failure characteristic quantity historical maximum value of the monitoring point m; is a failure characteristic quantity state monitoring normalized value of the monitoring point m at the t time point.

3. The method according to claim 1, wherein, In step 1, the state monitoring time series dataset D m Comprise 4 fields, 4 fields are: module number, monitoring point number, monitoring time, fault characteristic quantity.

4. The method according to claim 1, wherein, In step 1, the fault feature quantity of the adjacent time of the monitoring point includes voltage, current and temperature.

5. The method according to claim 1, wherein, In step 2, each monitoring point forms a training set Validation set Test set The ratio is 8:1:

1.

6. The method according to claim 1, wherein, In step 3, the BP neural network model is composed of three layers of neurons of input layer, hidden layer and output layer, and the number of neurons of the hidden layer and the output layer is 10 and 1 respectively.

7. The method according to claim 1, wherein, In step 3, the time sequence feature selection algorithm proposed specifically comprises the following sub-steps: Step 3-1, training set of each monitoring point Forming time series feature set F according to different time lags m (m = 1, 2, …, M), Wherein The time series feature formed by the time series data of the fault feature quantity at the time t-i when the monitoring point m predicts the fault feature quantity at the time t, Wherein Q m Indicates the number of test set samples of the monitoring point m; Step 3-2, training set by each monitoring point forming the output sequence Y of the prediction model corresponding to each monitoring point m wherein Step 3-3, initialize input parameters of the time series feature selection algorithm, including: time series feature set F m , output sequence Y m , prediction model P m , time series feature subset validation set stop control parameter Z, Z takes value 6; Step 3-4, by the subset of timing features Training the prediction model P m and using the validation set to compute the mean absolute percentage error E MAPE of the trained prediction model P MAPE as a performance measure of the prediction model P. Step 3-5, initialize z = 1, initialize optimal timing feature subset Step 3-6, compare z and Z, if z < Z, go to step 3-7, otherwise go to step 3-15; Step 3 - 7, in the temporal feature set F m producing a feature subset and a set of features Step 3 - 8, let a = 0, count the number of non-empty subsets A of the set Step 3-9, compare a and A, if a < A, go to step 3-10, otherwise go to step 3-14; Step 3 - 10, in the collection do not repeat the search subset by subset Train the prediction model P m and use the validation set to calculate the mean absolute percentage error E d of the trained prediction model; Step 3-11, Compare E d With the size of E MAPE If E d < E MAPE Go to Step 3-12, otherwise go to Step 3-13; Step 3-12, Let E MAPE = E d , a = a + 1, go to Step 3-9; Step 3-13, let a = a + 1, go to step 3-9; Step 3-14, let z = z + 1, go to step 3-6; Step 3-15, output optimal timing feature subset 8. The method according to claim 1, wherein, In step 4, the input vector of the prediction model m is where d denotes the dimension of the input, i.e., the number of neurons in the input layer of the BP neural network, which is determined by the time series feature selection algorithm, and the actual output of the prediction model m in the gth iteration is The expected output of the prediction model m is y m .

9. The method according to claim 1, wherein, In step 4, the training of the prediction model specifically comprises the following sub-steps: Step 4-1, initialize the connection weights between input layer and hidden layer connection weights between hidden layer and output layer hidden layer neuron threshold output layer neuron threshold expected training error α m learning rate λ m maximum iteration number G m neuron activation function wherein Step 4-2, let n = 1, g = 1; Step 4-3, select the nth input sample and the desired output y m (n) provided to the prediction model; Step 4-4, the output of the hidden layer and the output layer neurons is calculated: Step 4 - 5, utilize y m (n) and h m (n) Compute the partial derivative of the error function with respect to the output layer neuron Step 4 - 6, using the connection weights of the hidden layer to the output layer partial derivative and the hidden layer output computing the partial derivative of the error function with respect to the hidden layer neuron Step 4-7, using the partial derivatives of the output layer neurons and hidden layer neuron outputs to correct the connection weights and thresholds: Step 4-8, using hidden layer neuron partial derivatives and input layer neuron inputs to revise connection weights and thresholds: Step 4-9, if n < N, N is the number of samples, let n = n + 1, go to step 4-3, otherwise go to step 4-10; Step 4 - 10, calculate total error E m : Step 4-11, if E m < α m or g > G m then training is finished; otherwise g = g + 1, n = 1, return to step 4-3.

10. A system for online failure prediction based on any one of claims 1-9, characterized in that, The data processing module, the sample division module, the time sequence feature selection module, the training prediction model module and the fault feature quantity prediction module are included. The data processing module eliminates repeated records in the original data set collected by the monitoring points, and when the fault characteristic state monitoring data of the monitoring point is lost at a certain moment, the average value of the fault characteristic of the adjacent moment of the monitoring point is taken as the interpolation value, and then the state monitoring data of each monitoring point is sorted in chronological order, and the state monitoring data of each monitoring point in the data set is normalized to obtain the state monitoring time series data set D of the M monitoring points m (m = 1, 2, …, M); A sample division module, each monitoring point forms a training set in proportion Validation set Test set A time series feature selection module respectively establishes a back propagation (BP) neural network prediction model P for each monitoring point of the module m (m = 1, 2, …, M), taking the mean absolute percentage error (MAPE) of the prediction model as the performance measure of the prediction model, constructing a time series feature selection algorithm, performing time series feature selection on the fault prediction model, and obtaining the optimal input time series feature of each fault prediction model. The training prediction model module, the input vector, the actual output and the expected output of the prediction model are constructed, the fault prediction model after time sequence feature selection is trained, the connection weight values between the hidden layer-output layer and the input layer-hidden layer neurons and the threshold values of the hidden layer and the output layer neurons are updated through repeated iterations to obtain better model parameters and improve the prediction performance of the model; The fault feature quantity prediction module, the fault feature quantity of each monitoring point is predicted online by the trained fault prediction model respectively, the normalized prediction value of the fault feature quantity is obtained, and the actual prediction value of the fault feature quantity is obtained through the inverse normalization processing.

Images