Fault Prediction Method for the Isolated Switching Power Supply at the Front End of the Aging Bench

Abstract

The present invention discloses a fault prediction method for a front-end isolated switching power supply of an aging platform. The method includes: performing dimensionality reduction processing on preset timing signals of a plurality of switching power supplies of the aging platform, and extracting characteristic parameters; calculating the similarity of the performance degradation paths of each switching power supply of the aging platform according to the characteristic parameters of each preset timing signal; clustering each degradation path according to the similarity of the performance degradation paths of each switching power supply of the aging platform to obtain N degradation path sets; establishing N switching power supply life prediction models, and respectively training the N switching power supply life prediction models according to the N degradation path sets; predicting the life of the switching power supply of the aging platform based on the N trained switching power supply life prediction models to obtain the remaining life of the switching power supply. The fault prediction method for the front-end isolated switching power supply of the aging platform considering multiple degradation paths provided by this application can accurately predict the remaining service life of the switching power supply.

Description

Technical Field

[0001] The present invention relates to the technical field of integrated circuit non-destructive reliability screening, and particularly to a fault prediction method and device for a front-end isolated switching power supply of an aging station considering multiple degradation paths. Background Art

[0002] Existing high-temperature aging products can achieve long-term monitoring of the test environment by means of increasing in-machine long-term monitoring, over-stress protection mechanisms, etc., and can respond in a timely manner when a product fails, realizing after-sales maintenance based on fault data. However, due to the lack of active guarantee technologies for high-temperature aging products, it is difficult for related products to achieve the integrity of the test process and the consistency of test environment stress, which is extremely likely to lead to major property losses such as the forced interruption of the test process due to product failures resulting in the damage of millions of test devices (also known as devices to be detected), or the aging test being recognized as a failed test due to adverse effects such as additional stress introduced during the test due to product performance degradation, resulting in the ineffective waste of resources. At the same time, ensuring the quality of high-temperature aging tests can avoid over-aging and under-aging faults of integrated circuits, greatly reduce the failure rate of integrated circuits, and avoid the failure shutdown of large-scale electronic systems such as new energy vehicles, civil airliners, and energy storage power transmission stations due to integrated circuit failures.

[0003] In view of the fact that high-temperature aging products such as aging stations operate under harsh working conditions such as high temperature for a long time, the failure risk of their key equipment is relatively high. Therefore, the fault prediction and health management of the key equipment of the aging station have become an important research field. Effective health assessment and fault prediction can promote the development of reliable maintenance plans, thereby preventing potential faults. The key equipment of the aging station, such as the primary switching power supply, due to its core role in the system, once a failure occurs, it will have a serious impact on the entire system, not only resulting in test failures, but also possibly causing equipment damage and triggering safety accidents. Therefore, it is crucial to carry out accurate life prediction for the primary switching power supply of the aging station.

[0004] The prior art mainly focuses on life prediction for individual differences and degradation characteristics of equipment or systems, but there is little research on the impact of different degradation paths of the primary switching power supply of an aging platform on its life prediction results. However, in general, since the primary switching power supply system has unstable situations such as multiple degradation paths, there will be the following two problems with data-driven life prediction methods: First, multiple degradation paths: The primary switching power supply has multiple degradation paths, which are reflected as multiple performance degradation paths on the unexpired aging test bench. The impacts of different performance degradation paths on the remaining life are different, which affects life prediction. Second, long time series: The time series of the current and voltage parameters of the primary switching power supply is relatively long, and it is difficult to capture and learn the degradation information in the series, which affects life prediction. Due to the existence of the above two problems, it is currently impossible to accurately predict the life of the primary switching power supply of the aging platform. Summary of the Invention

[0005] The purpose of the embodiments of the present invention is to provide a fault prediction method, device, and electronic device for the front-end isolated switching power supply of an aging platform, which can solve the above problems existing in the prior art.

[0006] To solve the above technical problems, the present invention provides the following technical solutions:

[0007] The embodiments of the present invention provide a fault prediction method for the front-end isolated switching power supply of an aging platform, wherein the method includes:

[0008] Perform dimensionality reduction processing on the preset time series signals of multiple aging platform switching power supplies, and extract characteristic parameters;

[0009] Calculate the similarity of the performance degradation paths of each aging platform switching power supply according to the characteristic parameters of each preset time series signal;

[0010] Cluster each of the degradation paths according to the similarity of the performance degradation paths of each aging platform switching power supply to obtain N degradation path sets;

[0011] Establish N switching power supply life prediction models, and train the N switching power supply life prediction models respectively according to the N degradation path sets;

[0012] Based on the N trained switching power supply life prediction models, predict the life of the aging platform switching power supply to obtain the remaining life of the switching power supply.

[0013] Optionally, the characteristic parameters include: ripple voltage, average output voltage, peak-to-peak inductor current, peak inductor current, peak-to-peak capacitor current, peak capacitor current, circuit board temperature, heat flux density, heat flux rate, thermal conductivity, and thermal diffusivity, etc.

[0014] Optionally, the step of calculating the similarity of the performance degradation paths of the switching power supplies of each aging station according to the characteristic parameters of each of the preset timing signals includes:

[0015] Construct parameter vectors for the characteristic parameters of each of the preset timing signals respectively;

[0016] Construct a matrix grid according to each of the parameter vectors;

[0017] Based on the preset path boundary constraint conditions, path continuity constraint conditions, and path monotonicity constraint conditions, find the path with the minimum regularization cost passing through the matrix grid through multiple iterative adjustments, where the path with the minimum regularization cost can represent the similarity of the performance degradation paths of the switching power supplies of each aging station.

[0018] Optionally, the step of clustering each of the degradation paths according to the similarity of the performance degradation paths of the switching power supplies of each aging station to obtain N degradation path sets includes:

[0019] Randomly select K points from the path with the minimum regularization cost as the centers of the initial clusters;

[0020] For each point in the path with the minimum regularization cost, assign each point to the cluster closest to the center point;

[0021] Update the center point of each cluster;

[0022] For each point in the path with the minimum regularization cost, assign each point to the cluster closest to the updated center point;

[0023] Judge whether the clustering stop condition is reached. If not, return to execute the step of updating the center point of each cluster;

[0024] If it has been reached, take the clusters obtained by clustering the updated center points as the N degradation path sets.

[0025] Optionally, the step of updating the center point of each cluster includes:

[0026] For each cluster, calculate the average value of the points in the cluster;

[0027] Update the center point of the cluster to the position where the average value is located.

[0028] Optionally, the switching power supply life prediction model includes: an input layer, a Dense layer (i.e., a dense layer), a BiLSTM layer (i.e., a bidirectional long short-term memory network layer) for processing multi-sensor monitoring signals, a BiGRU layer (i.e., a bidirectional gated recurrent layer) for processing the transformation result after the operation condition data (i.e., the hidden representation), a Concatenate layer (i.e., a feature splicing layer) for fusing the outputs of the BiLSTM and BiGRU layers, a multi-dimensional feature map, and an output layer for outputting the RUL prediction value. Optionally, the step of establishing N switching power supply life prediction models and training the N switching power supply life prediction models respectively according to the N degradation path sets includes:

[0029] Establish N switching power supply life prediction models;

[0030] For each of the switching power supply life prediction models, use the characteristic parameters corresponding to the degradation paths in the corresponding degradation path set as the model input;

[0031] Perform a linear transformation on the input characteristic parameters in the fully connected layer and perform a hidden representation on the transformation result;

[0032] Construct a high-order vector based on the transformation result after the hidden representation and send the high-order vector to the bidirectional long short-term memory network layer and the bidirectional gated recurrent layer to obtain hidden feature mappings of different dimensions;

[0033] Input the feature vectors generated by combining the hidden features of different dimensions into two dense layers respectively to generate the RUL prediction value;

[0034] Use the error between the RUL prediction value and the actual RUL value as a loss function to calculate the loss;

[0035] Update the parameters of the switching power supply life prediction model according to the backpropagation of the loss;

[0036] Determine whether the switching power supply life prediction model after updating the parameters meets the preset accuracy rate;

[0037] If it meets the requirements, determine that the training of the switching power supply life prediction model is completed;

[0038] If it does not meet the requirements, return to execute the step of using the characteristic parameters corresponding to the degradation paths in the corresponding degradation path set as the model input.

[0039] An embodiment of the present invention further provides a front-end isolation type switching power supply fault prediction device for an aging platform. Among them, the device includes:

[0040] A feature extraction module for performing dimensionality reduction processing on the preset time series signals of multiple aging platform switching power supplies and extracting characteristic parameters;

[0041] A similarity determination module, configured to calculate the similarity of the performance degradation paths of the switching power supplies of each aging station according to the characteristic parameters of each preset timing signal;

[0042] A clustering module, configured to cluster the degradation paths according to the similarity of the performance degradation paths of the switching power supplies of each aging station, and obtain N degradation path sets;

[0043] A model training module, configured to establish N switching power supply life prediction models, and train the N switching power supply life prediction models respectively according to the N degradation path sets;

[0044] A model prediction module, configured to predict the life of the switching power supply of the aging station based on the N trained switching power supply life prediction models, and obtain the remaining life of the switching power supply.

[0045] Optionally, the characteristic parameters include: ripple voltage, average output voltage, peak-to-peak inductor current, peak inductor current, peak-to-peak capacitor current, peak capacitor current, circuit board temperature, heat flux density, heat flux rate, thermal conductivity, and thermal diffusivity, etc.

[0046] Optionally, the similarity determination module includes:

[0047] A first sub-module, configured to construct parameter vectors for the characteristic parameters of each preset timing signal respectively;

[0048] A second sub-module, configured to construct a matrix grid according to the parameter vectors;

[0049] A third sub-module, configured to find the path with the minimum regularization cost passing through the matrix grid through multiple iterative adjustments based on preset path boundary constraint conditions, path continuity constraint conditions, and path monotonicity constraint conditions, where the path with the minimum regularization cost can represent the similarity of the performance degradation paths of the switching power supplies of each aging station.

[0050] Optionally, the clustering module includes:

[0051] A fourth sub-module, configured to randomly select K points from the paths with the minimum regularization cost as the centers of the initial clusters;

[0052] A fifth sub-module, configured to assign each point in the paths with the minimum regularization cost to the cluster closest to the center point;

[0053] A sixth sub-module, configured to update the center point of each cluster;

[0054] A seventh sub-module, configured to assign each point in the paths with the minimum regularization cost to the cluster closest to the updated center point;

[0055] The eighth sub-module is used to determine whether the clustering stop condition is reached. If not, it returns to execute the sixth sub-module;

[0056] If the ninth sub-module has reached it, the clusters of the updated center points are used as the N sets of degradation paths.

[0057] Optionally, the sixth sub-module is specifically used for:

[0058] For each cluster, calculate the average value of each point in the cluster;

[0059] Update the center point of the cluster to the position where the average value is located.

[0060] Optionally, the switching power supply life prediction model includes:

[0061] An input layer, a Dense layer (i.e., a dense layer), a BiLSTM layer (i.e., a bidirectional long short-term memory network layer) for processing multi-sensor monitoring signals, a BiGRU layer (i.e., a bidirectional gated recurrent layer) for processing the transformation results after the operation condition data (i.e., hidden representations), a Concatenate layer (i.e., a feature concatenation layer) for fusing the outputs of the BiLSTM and BiGRU layers, a multi-dimensional feature map, and an output layer for outputting the RUL prediction value.

[0062] Optionally, the model training module is specifically used for:

[0063] Establish N switching power supply life prediction models;

[0064] For each of the switching power supply life prediction models, use the characteristic parameters corresponding to the degradation paths in the corresponding set of degradation paths as the model input;

[0065] Perform a linear transformation on the input characteristic parameters in the fully connected layer and perform a hidden representation on the transformation result;

[0066] Construct a high-order vector based on the transformation result after the hidden representation and send the high-order vector to the BLSTM layer and the BGRU layer to obtain hidden feature maps of different dimensions;

[0067] Respectively input the feature vectors generated by combining the hidden features of different dimensions into two dense layers (also known as linear regression dense layers) to generate the RUL prediction value;

[0068] Use the error between the RUL prediction value and the actual RUL value as a loss function to calculate the loss;

[0069] Update the parameters of the switching power supply life prediction model according to the backpropagation of the loss;

[0070] Determine whether the switching power supply life prediction model after updating the parameters meets the preset accuracy rate;

[0071] If satisfied, determine that the training of the switching power supply life prediction model ends;

[0072] If not satisfied, return to execute the operation of using the characteristic parameters corresponding to the degradation paths in the corresponding degradation path set as the model input.

[0073] An embodiment of the present invention also provides an electronic device, which is characterized by including a processor, a communication interface, a memory, and a communication bus. Among them, the processor, the communication interface, and the memory complete mutual communication through the communication bus; the memory is used to store a computer program; when the processor executes the program stored on the memory, it implements the flow of any of the above aging bench front-end isolated switching power supply fault prediction methods.

[0074] The aging bench front-end isolated switching power supply fault prediction solution disclosed in the embodiment of the present invention performs dimensionality reduction processing on the preset timing signals of multiple aging bench switching power supplies and extracts characteristic parameters; calculates the similarity of the performance degradation paths of each aging bench switching power supply according to the characteristic parameters of each preset timing signal; clusters each degradation path according to the similarity of the performance degradation paths of each aging bench switching power supply to obtain N degradation path sets; establishes N switching power supply life prediction models, and trains the N switching power supply life prediction models respectively according to the N degradation path sets; based on the trained N switching power supply life prediction models, predicts the life of the aging bench switching power supply to obtain the remaining life of the switching power supply. The aging bench front-end isolated switching power supply fault prediction solution disclosed in the embodiment of the present invention, on the one hand, the aging bench switching power supply life prediction model based on the deep neural network comprehensively considers the influence of the performance degradation path on the prediction of the remaining service life of the switching power supply, and can achieve high-precision life prediction; on the second hand, taking the failure mode as a consideration factor, identifying the performance degradation paths of the aging bench switching power supplies, and bringing the characteristic parameters of the aging bench switching power supplies with the same performance degradation path into the same model for training and prediction, which is beneficial to improving the accuracy of the aging bench switching power supply life prediction; on the third hand, since the fault of the aging bench front-end isolated switching power supply can be predicted, the damage to the device under test caused by the fault of the isolated switching power supply during the aging detection process can be effectively avoided. Description of the Drawings

[0075] Figure 1 It is a step flow chart showing an aging bench front-end isolated switching power supply fault prediction method according to an embodiment of the present application;

[0076] Figure 2 It is a schematic diagram showing the principle of an aging bench front-end isolated switching power supply fault prediction method according to an embodiment of the present application;

[0077] Figure 3It is a schematic diagram showing the DTW principle of the embodiment of the present application;

[0078] Figure 4 It is a flowchart showing the steps of a method for evaluating the health state of the secondary power supply of a high-temperature aging test device according to an embodiment of the present application;

[0079] Figure 5 It is a block diagram showing the structure of a fault prediction device for the front-end isolated switching power supply of an aging table according to an embodiment of the present application. Detailed implementation manners

[0080] To make the technical problems, technical solutions and advantages to be solved by the present invention clearer, the following will be described in detail with reference to the accompanying drawings and specific embodiments.

[0081] The following will, with reference to the accompanying drawings, through specific embodiments and their application scenarios, describe in detail the fault prediction method for the front-end isolated switching power supply of the aging table provided by the embodiments of the present application.

[0082] As shown in the attached Figure 1 figure, the fault prediction method for the front-end isolated switching power supply of the aging table in the embodiment of the present application includes the following steps:

[0083] Step 101: Perform dimensionality reduction processing on the preset timing signals of multiple aging table switching power supplies, and extract characteristic parameters.

[0084] The fault prediction method for the front-end isolated switching power supply of the aging table provided by the embodiments of the present application can be applied to an electronic device. An aging table switching power supply life prediction computer program is set in the electronic device. When the computer program is executed by a processor, the aging table switching power supply life prediction method in the embodiments of the present application is implemented to predict the remaining life of the switching power supply. The aging table switching power supply mentioned in the embodiments of the present application is the front-end isolated switching power supply of the aging table, and can also be a primary switching power supply, which can also be simply referred to as a switching power supply hereinafter. The aging table can also be referred to as a high-temperature aging table.

[0085] Through the aging table provided by the embodiments of the present application, non-destructive aging detection can be performed on the test piece to be detected. In the embodiments of the present application, the characteristic extraction and life prediction of the aging table switching power supply are to ensure the safety of the test piece to be detected.

[0086] Among them, the characteristic parameters include but are not limited to: ripple voltage, average output voltage, peak-to-peak value of inductor current, peak value of inductor current, peak-to-peak value of capacitor current, peak value of capacitor current, circuit board temperature, heat flux density, heat flux rate, thermal conductivity, thermal diffusivity, etc.

[0087] The preset timing signal can be a timing signal such as the current and voltage of the switching power supply. The multiple aging table switching power supplies are aging table switching power supplies with different decay rates and different degrees of fault starting.

[0088] Step 102: Calculate the similarity of the performance degradation paths of the switching power supplies of each aging station according to the characteristic parameters of each preset timing signal.

[0089] In the actual implementation process, the DTW (Dynamic Time Warping) method can be used to evaluate the similarity of the paths between the switching power supplies of the aging stations with multiple different decay rates and different degrees of fault initiation (i.e., the similarity of the performance degradation paths of the switching power supplies of each aging station). On this basis, the K-Means performance degradation path clustering method is introduced to achieve the rapid and accurate classification of the performance degradation paths of the switching power supplies of the aging stations (i.e., the aging test benches).

[0090] An optional way to calculate the similarity of the performance degradation paths of the switching power supplies of each aging station according to the characteristic parameters of each preset timing signal may include the following sub-steps:

[0091] Sub-step 1: Construct parameter vectors for the characteristic parameters of each preset timing signal respectively;

[0092] Sub-step 2: Construct a matrix grid according to each of the parameter vectors;

[0093] For the performance parameter vector of the primary switching power supply sample of the aging station , (a and b are sample numbers, and m and n are corresponding operation cycle numbers), a matrix grid is constructed, and the matrix element represents and the distance between two points , and each matrix element represents aligned with .

[0094] Sub-step 3: Based on the preset path boundary constraint conditions, path continuity constraint conditions, and path monotonicity constraint conditions, find the path with the minimum warping cost passing through the matrix grid through multiple iterative adjustments.

[0095] Among them, the path with the minimum warping cost can represent the similarity of the performance degradation paths of the switching power supplies of each aging station.

[0096] When specifically implementing Sub-step 3, the warping path passing through this grid can be found based on the path boundary constraint conditions, path continuity constraint conditions, and path monotonicity constraint conditions, and is represented by : 1. Then, based on these three constraint conditions, perform path warping iteratively multiple times until the path with the minimum warping cost is obtained , and the in the denominatorUsed to compensate for regular paths of different lengths.

[0097] The path boundary condition constraints can be set to , , ensuring that the selected path must start from the common starting point of the sequence and end at the common ending point of the sequence.

[0098] The path continuity constraint condition can be set as: If , then for the next point of the path needs to satisfy and , ensuring that and each coordinate in appears in

[0099] The path monotonicity constraint condition can be set as: If , then for the next point of the path needs to satisfy and , restricting the points above must progress monotonically with time.

[0100] Calculate the similarity of the characteristic time series trajectories such as ripple voltage, average output voltage, peak-to-peak inductor current, peak inductor current, peak-to-peak capacitor current, peak capacitor current, circuit board temperature, heat flux density, heat flux rate, thermal conductivity, and thermal diffusivity between different aging bench switching power supply samples (i.e., preset timing signals) through the DTW method. Subsequently, clustering of the time series trajectories can be achieved through the clustering method, so as to classify the aging bench switching power supplies according to the performance degradation trajectories.

[0101] Step 103: Cluster each degradation path according to the similarity of the performance degradation paths of each aging bench switching power supply to obtain N degradation path sets.

[0102] N is a positive integer greater than or equal to 2. In this embodiment of the application, N is taken as 2 for illustration.

[0103] In an alternative embodiment, the manner of clustering each degradation path according to the similarity of the performance degradation paths of each aging bench switching power supply to obtain N degradation path sets may include the following sub-steps:

[0104] Sub-step 1: Randomly select K points from the path with the minimum regularization cost as the centers of the initial clusters;

[0105] Sub-step 2: For each point in the path with the minimum regularization cost, assign each point to the cluster closest to the center point;

[0106] Sub-step 3: Update the center point of each cluster;

[0107] An exemplary way to update the center point of each cluster can be: for each cluster, calculate the average value of the points in the cluster; update the center point of the cluster to the position where the average value is located. The goal of updating the center point of the cluster is to minimize the within-cluster squared error.

[0108] Sub-step 4: For each point in the regularized cost minimum path, assign each point to the cluster closest to the updated center point;

[0109] Sub-step 5: Determine whether the clustering stop condition is reached. If not, return to execute the step of updating the center point of each cluster in sub-step 3;

[0110] If not, return to execute sub-step 3, sub-step 4, and sub-step 5, so as to update the center point again and cluster based on the updated center point.

[0111] The clustering stop condition can be flexibly set by those skilled in the art, and the embodiments of the present application do not make specific limitations thereto. For example: the clustering stop condition can be set as: the change in the position of the cluster center point before and after the update is less than a preset value (the position of the cluster center point no longer changes significantly before and after the update), and it can also be set as reaching the upper limit of the set number of iterations.

[0112] This optional degradation path clustering method utilizes the advantages of the time used for selecting the cluster centroid in K-Means and the spatial complexity of cluster overlap, and uses the truly existing optimal points in the dataset as its centroid. Therefore, the characteristics of the trajectory corresponding to the centroid can be found, and the system samples are clustered through the performance degradation path similarity matrix of the aging bench switching power supply, so as to realize the identification of the aging bench switching power supply according to the performance degradation trajectory time series (i.e., the preset timing signal).

[0113] Sub-step 6: If it has been reached, take the clusters obtained by clustering the updated center points as N degradation path sets.

[0114] Step 104: Establish N switching power supply life prediction models, and train the N switching power supply life prediction models respectively according to the N degradation path sets.

[0115] Among them, the switching power supply life prediction model includes:

[0116] An input layer, a Dense layer (i.e., a dense layer), a BiLSTM layer (i.e., a bidirectional long short-term memory network layer) for processing multi-sensor monitoring signals, a BiGRU layer (i.e., a bidirectional gated recurrent layer) for processing the transformation result after the operation condition data (i.e., the hidden representation), a Concatenate layer (i.e., a feature splicing layer) for fusing the outputs of the BiLSTM and BiGRU layers, a multi-dimensional feature map, and an output layer for outputting the RUL prediction value.

[0117] In an alternative embodiment, to establish N switching power supply life prediction models and train the N switching power supply life prediction models respectively according to N degradation path sets, the following sub-steps may be included:

[0118] Sub-step 1: Establish N switching power supply life prediction models.

[0119] When there are two performance degradation paths, two switching power supply life prediction models are established. The switching power supply life prediction model is used to predict the remaining life of the switching power supply.

[0120] Sub-step 2: For each switching power supply life prediction model, use the characteristic parameters corresponding to the degradation paths in the corresponding degradation path set as the model input.

[0121] Sub-step 3: Perform a linear transformation on the input characteristic parameters in the fully connected layer and perform a hidden representation on the transformation result;

[0122] Among them, the input characteristic parameters can be regarded as multi-sensory monitoring data X i .

[0123] Sub-step 4: Construct a high-order vector based on the transformation result after the hidden representation, and send the high-order vector to the BiLSTM layer and the BiGRU layer to obtain hidden feature maps of different dimensions;

[0124] Among them, the stacking layers of the BiLSTM layer and the BiGRU layer are the same.

[0125] Sub-step 5: Input the feature vectors generated by combining the hidden features of different dimensions into two linear regression dense layers respectively to generate RUL prediction values;

[0126] Sub-step 6: Use the error between the RUL prediction value and the actual RUL value as a loss function to calculate the loss;

[0127] Sub-step 7: Update the parameters of the switching power supply life prediction model according to the loss backpropagation;

[0128] Sub-step 8: Determine whether the switching power supply life prediction model after updating the parameters meets the preset accuracy rate; if it meets, determine that the training of the switching power supply life prediction model is completed; if it does not meet, return to execute the step of using the characteristic parameters corresponding to the degradation paths in the corresponding degradation path set as the model input.

[0129] The prediction accuracy rate is used to evaluate the performance of the prediction model. The accuracy rate of the switching power supply life prediction model can be calculated by the following formula:

[0130]

[0131] Among them, RUL is the actual RUL (Remaining Useful Life); is the predicted RUL estimated by the prediction model; TUL is the actual total service life of the target circuit, defined as the total number of test cycles when the switching power supply of the aging platform reaches the failure threshold.

[0132] The switching power supply life prediction model can be a multi-dimensional recurrent neural network (MDRNN) model. The multi-dimensional recurrent neural network uses momentum and an adaptive learning rate to accelerate the convergence speed and has been proven to be superior to the classical stochastic gradient descent optimizer, thus obtaining higher processing efficiency and prediction quality. The training process of the multi-dimensional recurrent neural network (MDRNN) model includes: S1: Extract the characteristic parameters of the switching power supply of the aging platform; S2: Establish an MDRNN multi-dimensional recurrent neural network model; S3: Initialize the parameters of the MDRNN multi-dimensional recurrent neural network model; S4: Based on the characteristic parameters in S1, iteratively train the MDRNN multi-dimensional recurrent neural network model multiple times, and calculate the loss function after each training is completed. Adjust the model parameters based on the calculated loss function; After multiple rounds of forward propagation and backpropagation training, when reaching the stop state, the trained MDRNN multi-dimensional recurrent neural network model can be used to predict the remaining life of the switching power supply.

[0133] Step 105: Based on the N trained switching power supply life prediction models, predict the life of the switching power supply of the aging platform to obtain the remaining life of the switching power supply.

[0134] When predicting the life of the switching power supply of the aging platform based on the N switching power supply life prediction models, the degradation path category to which it belongs can be determined based on the time series signal of the switching power supply to be predicted, and the switching power supply life prediction model matching this category can be called for accurate prediction. The specific prediction process can include: determining the time series signal characteristic parameters of the switching power supply of the aging platform to be predicted; determining the degradation category to which the switching power supply of the aging platform to be predicted belongs; inputting the characteristic parameters into the switching power supply life prediction model matching the category to predict the remaining service life of this switching power supply. The result of this prediction can also be used as the basic data for evaluating the accuracy of the prediction model of the switching unit in the future. Specifically, it can be: calculating the prediction error of the model based on the remaining service life obtained from this prediction and the real data of the remaining service life of this switching power supply; evaluating the accuracy of the model prediction result based on the prediction error.

[0135] The fault prediction method for the front-end isolated switching power supply of the aging platform provided by the embodiment of the present application performs dimensionality reduction processing on the preset timing signals of multiple aging platform switching power supplies and extracts characteristic parameters; calculates the similarity of the performance degradation paths of each aging platform switching power supply according to the characteristic parameters of each preset timing signal; clusters each degradation path according to the similarity of the performance degradation paths of each aging platform switching power supply to obtain N degradation path sets; establishes N switching power supply life prediction models, and trains the N switching power supply life prediction models respectively according to the N degradation path sets; predicts the life of the aging platform switching power supply based on the N trained switching power supply life prediction models to obtain the remaining life of the switching power supply. For the fault prediction method for the front-end isolated switching power supply of the aging platform, on the one hand, the switching power supply life prediction model of the aging platform based on the MDRNN multi-dimensional recurrent neural network model comprehensively considers the influence of the performance degradation path on the prediction of the remaining service life of the switching power supply, and can achieve high-precision life prediction; on the other hand, taking the fault mode as a consideration factor, identifying the performance degradation paths of the aging platform switching power supplies, and bringing the characteristic parameters of the aging platform switching power supplies with the same performance degradation path into the same model for training and prediction is beneficial to improving the accuracy of the life prediction of the aging platform switching power supply.

[0136] The following combines Figure 2 Take an example to illustrate the fault prediction scheme for the front-end isolated switching power supply of the aging platform provided by the present application, which specifically includes the following steps:

[0137] The fault prediction method for the front-end isolated switching power supply of the aging platform provided by this specific example evaluates the similarity of the paths between multiple front-end isolated switching power supplies of the aging platform (also called the primary switching power supply of the aging platform) with different decay rates and different fault starting degrees through the DTW method, and on this basis, introduces the K-Means performance degradation path clustering method, which can realize the rapid and accurate classification of the degradation paths of the primary switching power supply of the aging test bench; the prediction model of the remaining service life of the primary switching power supply of the aging platform based on the MDRNN multi-dimensional recurrent neural network model realizes high-precision life prediction by comprehensively considering the influence of the performance degradation path on the prediction of the remaining service life of the primary switching power supply of the aging platform; when training and predicting the multi-dimensional recurrent neural network model, taking the fault mode as a consideration factor, identifying the performance degradation paths of the primary switching power supply of the aging platform, and bringing the primary switching power supplies of the aging platform with the same performance degradation path into the same model for training and prediction is beneficial to improving the prediction accuracy of the remaining life of the primary switching power supply of the aging platform.

[0138] Such as Figure 2As shown in the schematic diagram of the principle of the fault prediction method for the front-end isolated switching power supply of the aging platform, it mainly includes the following parts: identifying and clustering the performance degradation paths of the primary switching power supply of the aging platform, establishing the MDRNN multi-dimensional recurrent neural network model, training the MDRNN multi-dimensional recurrent neural network model with category matching based on the clustered degradation paths, and predicting the remaining life of the primary switching power supply of the aging platform based on the trained MDRNN multi-dimensional recurrent neural network model.

[0139] The part of identifying and clustering the performance degradation paths of the primary switching power supply of the aging platform includes the following steps:

[0140] Sub-step 1: Dimension reduction and feature extraction.

[0141] In this specific example, feature parameters (also known as performance parameters) are extracted by dimension reduction from the time series signals such as current and voltage of the primary switching power supply. The extracted feature parameters include: ripple voltage, average output voltage, peak-to-peak inductor current, peak inductor current, peak-to-peak capacitor current, peak capacitor current, circuit board temperature, heat flux density, heat flux rate, thermal conductivity, thermal diffusivity, etc.

[0142] Sub-step 2: Calculation of DTW path similarity.

[0143] In this specific example, based on the extracted feature parameters, by taking advantage of the dynamic time warping (DTW) algorithm, the time axis of the unknown sequence is scaled to be the same length as the template sequence, so as to better grasp the similarity between the sample performance degradation paths. Among them, the principle of DTW is introduced as Figure 3 shown.

[0144] For the sample performance parameter vector of the primary switching power supply of the aging platform , (a, b are sample numbers, m, n are corresponding operation cycles), the main steps of the DTW method are as follows:

[0145] 1) Construct a matrix grid, and the matrix element represents and the distance between two points, and each matrix element represents aligned with .

[0146] 2) Find the warping path passing through this grid and represent it with : 1

[0147] 3) Path boundary condition constraint, that is, , , ensure that the selected path must start from the common starting point of the sequence and end at the common ending point of the sequence.

[0148] 4) Path continuity constraint, if , then for the next point of the path needs to satisfy and , ensure that and each coordinate in appears in

[0149] 5) Path monotonicity constraint, if , then for the next point of the path needs to satisfy and , restrict the points above must progress monotonically with time.

[0150] 6) Repeat (2)(3)(4) to obtain the path with the minimum regularization cost

[0151] , the in the denominator is used to compensate for the regularized paths of different lengths.

[0152] Calculate the similarity of the characteristic time series trajectories such as ripple voltage, average output voltage, peak-to-peak inductor current, peak inductor current, peak-to-peak capacitor current, peak capacitor current, circuit board temperature, heat flux density, heat flux rate, thermal conductivity, and thermal diffusivity between the primary switching power supply samples of different aging platforms through the DTW method. Subsequently, clustering of the time series trajectories can be achieved through the clustering method, and thus the primary switching power supplies of the aging platforms can be classified according to the performance degradation paths.

[0153] Sub-step 3: K-Means degradation path clustering.

[0154] Utilize the advantages of the time used for cluster centroid selection and the spatial complexity of cluster overlap in K-Means, and use the optimal points that truly exist in the dataset as its centroids. Therefore, the trajectories corresponding to the centroids can be found.

[0155] The steps of the K-Means algorithm are as follows:

[0156] 1) Initialize the center points:

[0157] Randomly select K points as the centers (Centroids) of the initial clusters.

[0158] 2) Assign clusters:

[0159] For each point in the dataset, calculate its distance to each centroid (usually using Euclidean distance) and assign it to the nearest cluster.

[0160] 3) Update the centroids:

[0161] For the points in each cluster, calculate their mean and update the centroid of the cluster to the mean position. The goal of K-Means is to minimize the within-cluster sum of squares, i.e.:

[0162]

[0163] 4) Repeat the iteration:

[0164] Repeat steps 2) and 3) until the cluster centroids no longer change significantly or the set number of iterations is reached.

[0165] Using the K-Means method, cluster the system samples through the similarity matrix of the performance degradation paths of the primary switching power supplies on the aging bench, so as to realize the identification of the primary switching power supplies on the aging bench according to the time series of the performance degradation paths (i.e., time series signals). As Figure 2 shown, the K-Means algorithm is used to cluster the degradation paths of multiple switching power supplies into two categories: Category 1 and Category 2. Subsequently, two MDRNN multi-dimensional recurrent neural network models will be constructed and trained based on the degradation paths of the two categories of switching power supplies respectively. The MDRNN multi-dimensional recurrent neural network model is hereinafter referred to as the MDRNN model for short.

[0166] The establishment of the MDRNN model, the training of the MDRNN model with category matching based on the clustered degradation paths, and the prediction of the remaining life of the primary switching power supply on the aging bench based on the trained MDRNN model are as follows:

[0167] The MDRNN multi-dimensional recurrent neural network extracts time series features by using the bidirectional LSTM (BiLSTM layer) and bidirectional GRU (BiGRU layer) layers respectively through the input of the degradation path data of two switching power supplies, and then splices the outputs of the two, and finally outputs the predicted remaining service life. Momentum and adaptive learning rate are used to accelerate the convergence speed and are proven to be superior to the classical stochastic gradient descent optimizer, thus obtaining higher processing efficiency and prediction quality.

[0168] The main framework of the MDRNN model is divided into four parts: the input layer, the processing layer, the fusion layer, and the output layer. The specific steps are as follows:

[0169] 1) Establish two lifetime prediction models corresponding to two performance degradation paths. Divide all power supply components into two categories according to the performance degradation paths, and use all the time series performance parameters (ripple voltage, average output voltage, peak-to-peak inductor current, peak inductor current, peak-to-peak capacitor current, peak capacitor current, circuit board temperature, heat flux density, heat flux rate, thermal conductivity, thermal diffusivity) of the primary switching power supply after preprocessing as the model inputs to construct the training set , where K is the number of training samples, represents the input data matrix obtained from the monitoring data time series according to the time window, with a size of N, where is the sensor measurement value of s selected sensors at time t.

[0170] 2) Concatenate the operating condition data in series to to construct a high-order vector . Then, feed the vector into the stacked BiLSTM layer and BiGRU layer to obtain hidden feature maps of different dimensions.

[0171] 3) Set the number of BiLSTM layers and BiGRU layers to be the same, that is, the total number of stacked BiGRU layers is also M. For the last BiLSTM layer and BiGRU layer, take the output of the last BiLSTM layer and BiGRU layer at the last time as the final output of the last BiLSTM layer and BiGRU layer.

[0172] 4) Concatenate the hidden feature vectors into a combined feature vector. Finally, input the combined feature vector into another two linear regression dense layers to generate the predicted RUL.

[0173] 5) Calculate the loss using the error between the predicted RUL value and the actual RUL value as the loss function, and update the relevant parameters in the network through backpropagation. Repeat the training until the loss meets the requirements.

[0174] For the time series of the performance parameters of the primary switching power supply on the aging bench , , n is the length of the time series, and each contains m power supply parameters, that is, an m-dimensional vector. In this case, m = 11. Perform positional encoding on , so that the input of the model is , where the PE calculation formula is as follows.

[0175]

[0176]

[0177] Among them, pos is the position of w in the sequence, and i is the position of the parameter in w.

[0178] The core of the MDRNN model lies in using a multi-layer neural network structure to combine multi-degradation path data, and using bidirectional LSTM (BiLSTM) and bidirectional GRU (BiGRU) layers to extract the long-term dependencies in the time series respectively. Then these features are fused, and finally the remaining useful life (RUL) of the device is predicted through a fully connected layer (Dense layer). The key to the whole process lies in feature extraction and fusion to achieve accurate prediction of the health state of the aging platform.

[0179] The prediction accuracy is used to evaluate the performance of the prediction model. In this patent, the definition of the prediction accuracy is as follows:

[0180]

[0181] Among them, RUL is the actual RUL (remaining useful life); is the predicted RUL estimated by the prediction model, and TUL is the actual total useful life of the target circuit, which is defined as the total number of test cycles when the aging platform switches the power supply once and reaches the failure threshold.

[0182] The method for predicting the failure of the front-end isolated switching power supply of the aging platform provided by the embodiments of the present application, on the one hand, the life prediction model of the switching power supply of the aging platform based on the deep neural network comprehensively considers the influence of the performance degradation path on the prediction of the remaining useful life of the switching power supply, and can achieve high-precision life prediction; on the other hand, taking the failure mode as a consideration factor, identifying the performance degradation path of the switching power supply of the aging platform, and bringing the characteristic parameters of the switching power supplies of the aging platforms with the same performance degradation path into the same model for training and prediction, which is beneficial to improving the life prediction accuracy of the switching power supply of the aging platform.

[0183] The embodiments of the present application also provide a method for evaluating the health state of the secondary power supply of a high-temperature aging test device. This method can be executed Figure 1 after the end of the failure prediction process of the front-end isolated switching power supply of the aging platform shown, or can be executed before it, or can be executed in parallel with it. The high-temperature aging test device is the high-temperature aging platform.

[0184] As shown in the appendix Figure 4 The method for evaluating the health state of the secondary power supply of the high-temperature aging test device according to the embodiments of the present application includes the following steps:

[0185] Step 401: Collect the monitoring signals of the secondary power supply of the high-temperature aging test device to be evaluated.

[0186] The method for evaluating the health status of the secondary power supply of the high-temperature aging test equipment provided by the embodiment of the present application can be applied to electronic devices. In the electronic device, a computer program for evaluating the health status of the secondary power supply of the high-temperature aging test equipment is set. When the computer program is executed by a processor, the method for evaluating the health status of the secondary power supply of the high-temperature aging test equipment in the embodiment of the present application is realized.

[0187] The aging test equipment provided by the present application can perform non-destructive aging detection on the test piece. The method for evaluating the health status of the secondary power supply in the embodiment of the present application can evaluate the health status of the secondary power supply of the aging test equipment, and the health evaluation of the secondary power supply of the aging test equipment can ensure the safety of the test piece.

[0188] In the actual implementation process, the directly measured monitoring signal data of the secondary power supply of the high-temperature aging test equipment is large in quantity and contains a large amount of redundant information. Therefore, preliminary feature extraction is performed on the directly measured data to obtain a feature sequence with high information density.

[0189] Step 402: Generate a multi-dimensional feature sequence according to the monitoring signal.

[0190] The monitoring signal includes but is not limited to current signal, voltage output signal, power supply ripple volatility, MOSFET case temperature to junction temperature, etc. The specific features extracted can be flexibly set by those skilled in the art according to actual needs, as long as the extracted features can characterize the state of the secondary power supply. Among them, the abbreviation of MOSFET is MOS, which is a metal-oxide-semiconductor field-effect transistor, a voltage-controlled semiconductor device, and refers to the voltage-controlled semiconductor device in the secondary power supply of the high-temperature aging test equipment in this embodiment.

[0191] Specifically, feature extraction is performed on the current signal and voltage output signal, and then the power supply ripple volatility and the degradation curve of the MOSFET device case temperature to junction temperature are mixed to obtain a multi-dimensional feature sequence.

[0192] Step 403: Input the multi-dimensional feature sequence into a pre-trained residual convolutional neural network model to obtain the degradation feature state of the secondary power supply of the high-temperature aging test equipment.

[0193] Among them, the degradation feature state includes: healthy state and degradation state. If the degradation feature state of the secondary power supply of the high-temperature aging test equipment is the healthy state, step 404 and subsequent steps do not need to be executed. If the degradation feature state of the secondary power supply of the high-temperature aging test equipment is the degradation state, it means that the secondary power supply of the high-temperature aging test equipment has degraded. By performing adaptive analysis on the multi-dimensional feature sequence data through the residual convolutional neural network model, the degradation progress of the secondary power supply of the high-temperature aging test equipment can be automatically identified.

[0194] In an alternative embodiment, the training process of the residual convolutional neural network model may include the following sub-steps:

[0195] Sub-step 1: Obtain the training time-series degradation feature samples of the secondary power supply of the high-temperature aging test equipment;

[0196] Among them, each training time-series degradation feature sample is labeled with a degradation feature state. The training time-series degradation feature samples of the secondary power supply of the high-temperature aging test equipment can be obtained by monitoring the temperature signal and the degradation feature state during the entire life cycle of the secondary power supply of a high-temperature aging test equipment.

[0197] In the actual implementation process, the training time-series degradation feature samples of the secondary power supply of the high-temperature aging test equipment can be expressed as , the maximum number of iterations max_iter. Mark the corresponding degradation feature state of the secondary power supply of the trained high-temperature aging test equipment , including the healthy state and the degraded state.

[0198] Sub-step 2: Input the training time-series degradation feature samples into the pre-established residual convolutional neural network model to obtain the first probability space point after mapping;

[0199] The residual convolutional neural network model includes: a convolutional layer, a batch normalization layer, an activation function (ReLU), a max pooling layer, multiple residual blocks (Residual blocks), a global average pooling layer, and a fully connected layer.

[0200] In the actual implementation process, taking as the input, as the output to train the ResNet model (i.e., the residual convolutional neural network model); obtain the iterative learning to update the ResNet model parameters. When the number of iterations reaches the maximum number of iterations max_iter, stop the iterative learning.

[0201] In the actual implementation process, the maximum number of iterations max_iter can be flexibly set by those skilled in the art, and no specific limitation is made in the embodiments of the present application.

[0202] Sub-step 3: Discriminate the degradation starting point of the secondary power supply of the high-temperature aging table according to the two-dimensional first probability space point, and generate the health baseline of the secondary power supply of the high-temperature aging table.

[0203] In an alternative embodiment, the training process of the residual convolutional neural network model may further include the following process:

[0204] Obtain the test timing degradation feature samples of the secondary power supply of the high-temperature aging test equipment; input the test timing degradation feature samples into the residual convolutional neural network model being trained to obtain the mapped second probability space points; calculate the distance between the second probability space and the healthy baseline based on a preset multi-distance metric algorithm.

[0205] The test timing degradation feature samples of the secondary power supply of the high-temperature aging test equipment can be expressed as: . During the training and testing of the residual convolutional neural network model, using and as the input of the ResNet algorithm model to obtain the mapped probability space points and .

[0206] Automatically discriminate the degradation starting point according to the two-dimensional probability space result, and use the weighted fusion multi-distance metric algorithm to calculate the distances between the healthy space and the degradation space of the secondary power supply of the training and testing high-temperature aging platforms . Traverse all failure modes to obtain the corresponding health assessment results for each failure mode. Finally, the trained residual convolutional neural network model outputs and The health assessment curves corresponding to the second power supply of each high-temperature aging platform.

[0207] More specifically, based on the preset multi-distance metric algorithm, the method for calculating the distance between the second probability space and the healthy baseline can be: calculate the distances between the second probability space and the healthy baseline based on the Mahalanobis distance algorithm, cosine similarity algorithm, and Manhattan distance algorithm respectively to obtain the first distance, second distance, and third distance; perform a weighted sum of the first distance, second distance, and third distance to obtain the corresponding health degree.

[0208] In the actual implementation process, it is not limited to using the above three distance algorithms to calculate the distance. It can also use any one of the above three distance algorithms, or use any two of the above three distance algorithms to obtain two distances and then perform a weighted sum of the two distances.

[0209] Introduce the test timing degradation feature samples of the secondary power supply of the high-temperature aging test equipment during the training process of the residual convolutional neural network model, and the accuracy of the prediction of the residual convolutional neural network model can be tested through the prediction results. In the actual implementation process, when the prediction accuracy of the residual convolutional neural network model reaches the accuracy threshold, it can be determined that the training of this residual convolutional neural network model is completed, and the subsequent trained residual convolutional neural network model can be used to predict the state degradation curve of the secondary power supply of the high-temperature aging platform to be evaluated.

[0210] Step 404: When the secondary power supply of the high-temperature aging test equipment is in a degraded state, determine the state degradation curve of the secondary power supply of the high-temperature aging test equipment to be evaluated based on the probability space points after mapping the multi-dimensional feature sequence and the health baseline in the residual convolutional neural network model.

[0211] Among them, the state degradation curve can characterize the health state of the secondary power supply of the high-temperature aging test equipment.

[0212] By analyzing the multi-dimensional features of the secondary power supply of the high-temperature aging test equipment to be evaluated through the residual convolutional neural network model, the health degree of the secondary power supply at the current moment can be obtained, and at the same time, the state degradation curve (also known as the health degree degradation curve) up to this moment is output, providing reliable data support for the further prediction research on the remaining service life of the secondary power supply of the high-temperature aging test equipment.

[0213] The method for evaluating the health state of the secondary power supply of the high-temperature aging test equipment disclosed in the embodiments of the present invention collects the monitoring signals of the secondary power supply of the high-temperature aging test equipment to be evaluated, and generates a multi-dimensional feature sequence based on the monitoring signals; inputs the multi-dimensional feature sequence into a pre-trained residual convolutional neural network model to obtain the degraded feature state of the secondary power supply of the high-temperature aging test equipment; when the secondary power supply of the high-temperature aging test equipment is in a degraded state, based on the probability space points after mapping the multi-dimensional feature sequence and the health baseline in the residual convolutional neural network model, determine the state degradation curve of the secondary power supply of the high-temperature aging test bench to be evaluated, so as to accurately determine the health state of the secondary power supply of the high-temperature aging test bench. The solution for evaluating the health state of the secondary power supply of the high-temperature aging test equipment provided by the embodiments of the present invention collects the monitoring signals of the secondary power supply and extracts the multi-dimensional feature sequence, and predicts the multi-dimensional feature sequence based on the trained model to obtain the state degradation curve of the secondary power supply of the high-temperature aging test equipment to be evaluated, so as to accurately determine the health state of the secondary power supply of the high-temperature aging test equipment.

[0214] Figure 5 It is a structural block diagram of the fault prediction device for the front-end isolated switching power supply of the aging bench in the embodiments of the present application.

[0215] The fault prediction device for the front-end isolated switching power supply of the aging bench provided by the embodiments of the present application includes the following functional modules:

[0216] The feature extraction module 501 is used to perform dimensionality reduction processing on the preset timing signals of multiple aging bench switching power supplies and extract feature parameters; among them, the aging bench switching power supply is the front-end isolated switching power supply of the aging bench.

[0217] The similarity determination module 502 is used to calculate the similarity of the performance degradation paths of each aging bench switching power supply according to the feature parameters of each preset timing signal.

[0218] The clustering module 503 is configured to cluster the degradation paths according to the similarity of the performance degradation paths of the switching power supplies of each aging station, and obtain N degradation path sets;

[0219] The model training module 504 is configured to establish N switching power supply life prediction models, and train the N switching power supply life prediction models respectively according to the N degradation path sets;

[0220] The model prediction module 505 is configured to predict the life of the switching power supply of the aging station based on the N trained switching power supply life prediction models, and obtain the remaining life of the switching power supply.

[0221] Optionally, the characteristic parameters include: ripple voltage, average output voltage, peak-to-peak inductor current, peak inductor current, peak-to-peak capacitor current, peak capacitor current, circuit board temperature, heat flux density, heat flux rate, thermal conductivity, thermal diffusivity, etc.

[0222] Optionally, the similarity determination module includes:

[0223] The first sub-module is configured to construct parameter vectors for the characteristic parameters of each of the preset timing signals;

[0224] The second sub-module is configured to construct a matrix grid according to the parameter vectors;

[0225] The third sub-module is configured to, based on the preset path boundary constraint conditions, path continuity constraint conditions, and path monotonicity constraint conditions, find the path with the minimum regularization cost passing through the matrix grid through multiple iterative adjustments, where the path with the minimum regularization cost can represent the similarity of the performance degradation paths of the switching power supplies of each aging station.

[0226] Optionally, the clustering module includes:

[0227] The fourth sub-module is configured to randomly select K points from the path with the minimum regularization cost as the centers of the initial clusters;

[0228] The fifth sub-module is configured to, for each point in the path with the minimum regularization cost, assign each point to the cluster closest to the center point;

[0229] The sixth sub-module is configured to update the center point of each cluster;

[0230] The seventh sub-module is configured to, for each point in the path with the minimum regularization cost, assign each point to the cluster closest to the updated center point;

[0231] The eighth sub-module is configured to determine whether the clustering stop condition is reached. If not, return to execute the sixth sub-module;

[0232] If the ninth sub-module is reached, the updated clusters of the center points are used as the set of N degradation paths.

[0233] Optionally, the sixth sub-module is specifically configured to:

[0234] For each cluster, calculate the average value of the points in the cluster;

[0235] Update the center point of the cluster to the position where the average value is located.

[0236] Optionally, the switching power supply life prediction model includes: an input layer, a Dense layer (i.e., a dense layer), a BiLSTM layer (i.e., a bidirectional long short-term memory network layer) for processing multi-sensor monitoring signals, a BiGRU layer (i.e., a bidirectional gated recurrent layer for processing the transformation result after the operation condition data, i.e., the hidden representation), a Concatenate layer (i.e., a feature concatenation layer) for fusing the outputs of the BiLSTM and BiGRU layers, a multi-dimensional feature map, and an output layer for outputting the RUL prediction value. Optionally, the model training module is specifically configured to:

[0237] Establish N switching power supply life prediction models;

[0238] For each of the switching power supply life prediction models, use the characteristic parameters corresponding to the degradation paths in the corresponding degradation path set as the model input;

[0239] Perform a linear transformation on the input characteristic parameters in the fully connected layer and perform a hidden representation on the transformation result;

[0240] Construct a high-order vector based on the transformation result after the hidden representation and send the high-order vector to the BLSTM layer and the BGRU layer to obtain hidden feature maps of different dimensions;

[0241] Input the feature vectors generated by combining the hidden features of different dimensions into two linear regression dense layers respectively to generate RUL prediction values;

[0242] Use the error between the RUL prediction value and the actual RUL value as a loss function to calculate the loss;

[0243] Update the parameters of the switching power supply life prediction model according to the backpropagation of the loss;

[0244] Determine whether the switching power supply life prediction model after updating the parameters meets the preset accuracy rate;

[0245] If it meets the requirements, determine that the training of the switching power supply life prediction model is completed;

[0246] If it does not meet the requirements, return to execute the operation of using the characteristic parameters corresponding to the degradation paths in the corresponding degradation path set as the model input.

[0247] The aging bench front-end isolated switching power supply fault prediction device provided by the embodiments of the present application Figure 5 shown can implement Figure 1 each process implemented by the method embodiments. To avoid repetition, details are not described herein again.

[0248] The aging bench front-end isolated switching power supply fault prediction device provided by the embodiments of the present application, on the one hand, the aging bench switching power supply life prediction model based on a deep neural network comprehensively considers the influence of the performance degradation path on the prediction of the remaining service life of the switching power supply, and can achieve high-precision life prediction; on the other hand, taking the fault mode as a consideration factor, identifying the performance degradation path of the aging bench switching power supply, and bringing the characteristic parameters of the aging bench switching power supplies with the same performance degradation path into the same model for training and prediction, which is beneficial to improving the life prediction accuracy of the aging bench switching power supply.

[0249] The embodiments of the present invention also provide an electronic device, including a processor, a communication interface, a memory, and a communication bus. Among them, the processor, the communication interface, and the memory complete mutual communication through the communication bus.

[0250] The memory is used to store a computer program;

[0251] The processor, when executing the program stored in the memory, implements the aging bench front-end isolated switching power supply fault prediction method considering multiple degradation paths shown in the above method embodiments.

[0252] The communication bus mentioned in the above terminal may be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into an address bus, a data bus, a control bus, etc.

[0253] The communication interface is used for communication between the above terminal and other devices.

[0254] The memory may include a Random Access Memory (RAM), and may also include a non-volatile memory, such as at least one disk memory. Optionally, the memory may also be at least one storage device located far from the aforementioned processor.

[0255] In another embodiment provided by the present invention, there is also provided a computer-readable storage medium storing instructions, which, when running on an electronic device, enable the electronic device to implement the aging bench front-end isolated switching power supply fault prediction method considering multiple degradation paths described in any one of the above embodiments.

[0256] In another embodiment provided by the present invention, there is also provided a computer program product containing instructions, which, when running on an electronic device, enable the electronic device to implement the aging bench front-end isolated switching power supply fault prediction method considering multiple degradation paths described in any one of the above embodiments.

[0257] It should be noted that in this text, the term "including", "comprising" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or further includes elements inherent to such process, method, article or device. Without further limitation, an element defined by the statement "including one..." does not exclude the existence of another identical element in the process, method, article or device including the element.

[0258] The above is the preferred embodiment of the present invention. It should be pointed out that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present invention.

Claims

1. A fault prediction method for a front-end isolated switching power supply of an aging table, characterized in that, The method includes: Performing dimensionality reduction processing on the preset timing signals of multiple aging bench switching power supplies, and extracting characteristic parameters; Calculating the similarity of the performance degradation paths of each aging bench switching power supply according to the characteristic parameters of each preset timing signal; Clustering each of the degradation paths according to the similarity of the performance degradation paths of each aging bench switching power supply to obtain N degradation path sets; Establishing N switching power supply life prediction models, and training the N switching power supply life prediction models respectively according to the N degradation path sets; it includes: Establishing N switching power supply life prediction models; wherein, the switching power supply life prediction model includes: an input layer, a dense layer, a bidirectional long short-term memory network layer, a bidirectional gated recurrent layer, a feature splicing layer, a multi-dimensional feature map, and an output layer; For each of the switching power supply life prediction models, using the characteristic parameters corresponding to the degradation paths in the corresponding degradation path set as the model input; Performing a linear transformation on the input characteristic parameters in the fully connected layer, and performing a hidden representation on the transformation result; Constructing a high-order vector according to the transformation result after the hidden representation, and sending the high-order vector to the bidirectional long short-term memory network layer and the bidirectional gated recurrent layer to obtain hidden feature mappings of different dimensions; Inputting the feature vectors generated by combining the hidden features of different dimensions into two dense layers respectively to generate RUL prediction values; Determining whether the updated switching power supply life prediction model meets a preset accuracy rate; If it is satisfied, determining that the training of the switching power supply life prediction model is completed; If it is not satisfied, returning to execute the step of using the characteristic parameters corresponding to the degradation paths in the corresponding degradation path set as the model input; Based on the N trained switching power supply life prediction models, predicting the life of the aging bench switching power supply to obtain the remaining life of the switching power supply.

2. The method according to claim 1, wherein The characteristic parameters include: ripple voltage, average output voltage, peak-to-peak inductor current, peak inductor current, peak-to-peak capacitor current, peak capacitor current, circuit board temperature, heat flux density, heat flux rate, thermal conductivity, and thermal diffusivity.

3. The method according to claim 1, characterized in that The step of calculating the similarity of the performance degradation paths of each aging bench switching power supply according to the characteristic parameters of each preset timing signal includes: Constructing parameter vectors for the characteristic parameters of each preset timing signal respectively; Constructing a matrix grid according to each parameter vector; Based on preset path boundary constraint conditions, path continuity constraint conditions, and path monotonicity constraint conditions, finding the path with the minimum regularization cost passing through the matrix grid through multiple iterative adjustments, where the path with the minimum regularization cost can represent the similarity of the performance degradation paths of each aging bench switching power supply.

4. The method according to claim 3, wherein The step of clustering each of the degradation paths according to the similarity of the performance degradation paths of each aging bench switching power supply to obtain N degradation path sets includes: Randomly selecting K points from the path with the minimum regularization cost as the centers of the initial clusters; For each point in the path with the minimum regularization cost, assigning each point to the cluster closest to the center point; Updating the center points of each cluster; For each point in the regularized minimum-cost path, assign each point to the cluster that is closest to the updated center point. Determine whether the clustering stop condition is reached. If not, return to execute the step of updating the center point of each cluster. If it has been reached, use the clusters obtained by clustering the updated center points as the N sets of degradation paths.

5. The method according to claim 4, characterized in that The step of updating the center point of each cluster includes: For each cluster, calculate the average value of the points in the cluster. Update the center point of the cluster to the position where the average value is located.

6. An aging table front-end isolated switching power supply fault prediction device, characterized in that, The device includes: A feature extraction module, configured to perform dimensionality reduction processing on the preset timing signals of multiple aging bench switch-mode power supplies and extract feature parameters. A similarity determination module, configured to calculate the similarity of the performance degradation paths of each aging bench switch-mode power supply according to the feature parameters of each preset timing signal. A clustering module, configured to cluster each of the degradation paths according to the similarity of the performance degradation paths of each aging bench switch-mode power supply to obtain N sets of degradation paths. A model training module, configured to establish N switch-mode power supply life prediction models, and train the N switch-mode power supply life prediction models respectively according to the N sets of degradation paths. A model prediction module, configured to predict the life of the aging bench switch-mode power supply based on the N trained switch-mode power supply life prediction models to obtain the remaining life of the switch-mode power supply. The model training module is specifically configured to establish N switch-mode power supply life prediction models; wherein, the switch-mode power supply life prediction model includes: an input layer, a dense layer, a bidirectional long short-term memory network layer, a bidirectional gated recurrent layer, a feature splicing layer, a multi-dimensional feature map, and an output layer; for each switch-mode power supply life prediction model, use the feature parameters corresponding to the degradation paths in the corresponding set of degradation paths as the model input; perform a linear transformation on the input feature parameters in the fully connected layer and perform a hidden representation on the transformation result; construct a high-order vector according to the transformation result after the hidden representation and send the high-order vector to the bidirectional long short-term memory network layer and the bidirectional gated recurrent layer to obtain hidden feature maps of different dimensions; input the feature vectors generated by combining the hidden features of different dimensions into two dense layers respectively to generate the RUL prediction value; determine whether the updated switch-mode power supply life prediction model meets the preset accuracy rate; if it meets, determine that the training of the switch-mode power supply life prediction model is completed; if it does not meet, return to execute the step of using the feature parameters corresponding to the degradation paths in the corresponding set of degradation paths as the model input.

7. The device according to claim 6, characterized in that, The feature parameters include: ripple voltage, average output voltage, peak-to-peak inductor current, peak inductor current, peak-to-peak capacitor current, peak capacitor current, circuit board temperature, heat flux density, heat flux rate, thermal conductivity, and thermal diffusivity.

8. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus. Among them, the processor, the communication interface, and the memory complete mutual communication through the communication bus. The memory is used to store computer programs. The processor is configured to implement the fault prediction method for the front-end isolated switch-mode power supply of the aging bench as described in any one of claims 1-5 when executing the program stored on the memory.

Images