Health Feature Extraction Method and Device for Isolated Switching Power Supply at the Front End of Aging Platform

Abstract

The embodiments of this specification provide a method and device for extracting health characteristics of a front-end isolated switching power supply of an aging platform. Among them, the method for extracting health characteristics of a front-end isolated switching power supply of an aging platform includes: obtaining multiple historical signal values of analog signals of the front-end isolated switching power supply of the aging platform; encoding the multiple historical signal values through a Transformer model encoder in a multi-layer model encoder to generate corresponding self-attention features. The multi-layer model encoder is a neural network structure that fuses context features constructed by a Transformer model encoder and a stacked autoencoder, including an input layer, an intermediate layer, and an output layer. The intermediate layer includes a sequentially connected Transformer model encoder structure and a stacked autoencoder; extracting the frequency domain features of the multiple historical signal values and fusing the self-attention features with the frequency domain features; encoding the fused features through a stacked autoencoder to generate the corresponding health characteristics of the front-end isolated switching power supply of the aging platform.

Description

Technical Field

[0001] The embodiments of this specification relate to the technical field of integrated circuit non-destructive reliability screening, and particularly to a method for extracting health characteristics of a front-end isolated switching power supply of an aging platform. Background Art

[0002] Integrated circuit high-temperature aging test platforms accelerate various physical and chemical reaction processes inside components by continuously applying a certain electrical stress to the components for a long time, prompting various potential faults inside the components to be exposed early, so as to eliminate early failure products and enable electronic components to enter a period with low failure rate and relatively stable state from the beginning of use. The current integrated circuit high-temperature aging test platforms 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 the machine fails, realizing after-sales maintenance based on fault data. However, due to the lack of active guarantee technologies for integrated circuit high-temperature aging test platforms, it is difficult to achieve the integrity of the test process and the consistency of test environment stress, and it is extremely easy to cause major property losses such as the destruction of millions of test devices due to the forced interruption of the test process caused by machine failures, or the aging test is recognized as a failure test due to adverse effects such as the introduction of additional stress during the test period caused by the degradation of machine performance, resulting in 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 is conducive to avoiding the shutdown of large-scale electronic systems such as new energy vehicles, civil airliners, and energy storage power transmission stations using the same integrated circuits due to integrated circuit failures.

[0003] The front-end isolated switching power supply provides the required power for the entire aging test platform to ensure that all parts of the test platform can work normally. As the core key equipment of the aging test platform, the front-end isolated switching power supply has a great impact on the overall reliability of the aging test platform. Once there are circuit degradation, failures or sudden faults, at best, the aging test platform stops the test and damages the object under test, and at worst, it causes voltage overload, triggers a fire, resulting in major property losses and safety hazards. Therefore, it is very necessary to conduct fault prediction and health assessment on the front-end isolated switching power supply of the aging test platform. And to conduct fault prediction and health assessment on the front-end isolated switching power supply of the aging test platform, it needs to be realized based on the health characteristics of the front-end isolated switching power supply of the aging test platform. Therefore, how to extract the health characteristics of the front-end isolated switching power supply of the aging test platform has become an urgent problem to be solved. Summary of the Invention

[0004] In view of this, an embodiment of this specification provides a method for extracting health characteristics of a front-end isolated switching power supply of an aging platform. One or more embodiments of this specification also relate to a device for extracting health characteristics of a front-end isolated switching power supply of an aging platform, a computing device, a computer-readable storage medium, and a computer program, so as to solve the technical defects existing in the prior art.

[0005] According to the first aspect of the embodiments of this specification, a method for extracting health characteristics of a front-end isolated switching power supply of an aging platform is provided, including:

[0006] Obtain multiple historical signal values of the analog signal of the front-end isolated switching power supply of the aging platform;

[0007] Encode the multiple historical signal values through the Transformer model encoder in the multi-layer model encoder to obtain the self-attention features in the multiple historical signal values. Among them, the multi-layer model encoder is a neural network structure that fuses context features constructed by the Transformer model encoder and the stacked autoencoder, including an input layer, an intermediate layer, and an output layer. The intermediate layer includes a Transformer model encoder structure and a stacked autoencoder connected in sequence;

[0008] Extract the frequency-domain features of the multiple historical signal values, and fuse the self-attention features and the frequency-domain features to generate target fusion features;

[0009] Encode the target fusion features through the stacked autoencoder in the multi-layer model encoder to generate the health characteristics corresponding to the front-end isolated switching power supply of the aging platform.

[0010] Optionally, the multi-layer model encoder further includes a multi-head attention layer, a frequency-domain feature extraction layer, and a fusion layer. The frequency-domain feature extraction layer is used to extract frequency-domain features, and the fusion layer is used to fuse the data features extracted by the Transformer model encoder and the frequency-domain feature extraction layer;

[0011] Correspondingly, the method further includes:

[0012] Obtain the transformation result generated by the linear transformation of the multiple historical signal values by the multi-head attention layer;

[0013] Encode the transformation result through the Transformer model encoder to generate the self-attention features in the multiple historical signal values;

[0014] Extract the frequency-domain features of the transformation result through the frequency-domain feature extraction layer to obtain the frequency-domain features in the multiple historical signal values output by the frequency-domain feature extraction layer;

[0015] Fuse the self-attention features and the frequency-domain features through the fusion layer to obtain the target fusion features output by the fusion layer.

[0016] Optionally, after obtaining the multiple historical signal values of the analog signal of the front-end isolated switching power supply of the aging platform, the method further includes:

[0017] Perform sliding window cutting on the multiple historical signal values according to a first window width and a first step length to generate a first sliding window cutting result, and determine an initial sample data set based on the first sliding window cutting result;

[0018] Perform normalization processing on the initial sample data set to generate a target sample data set;

[0019] Correspondingly, encoding the multiple historical signal values by the Transformer model encoder in the multi-layer model encoder includes:

[0020] Encoding the target sample data set by the Transformer model encoder in the multi-layer model encoder.

[0021] Optionally, extracting the frequency-domain features of the multiple historical signal values includes:

[0022] Perform sliding window cutting on the target sample data set according to a second window width and a second step length to generate a second sliding window cutting result, and extract the frequency-domain features of the multiple historical signal values based on the second sliding window cutting result.

[0023] Optionally, extracting the frequency-domain features of the multiple historical signal values based on the second sliding window cutting result includes:

[0024] Perform Fourier transform on each sample data set to be processed included in the second sliding window cutting result to generate a corresponding time-series energy spectrum for each sample data set to be processed;

[0025] Extract the eigenvalue corresponding to at least two initial frequency-domain features of the time-series energy spectrum respectively;

[0026] Perform normalization processing on the eigenvalues by using the min-max normalization algorithm to generate the frequency-domain features of the multiple historical signal values.

[0027] Optionally, the at least two initial frequency-domain features include: peak value, bandwidth, center frequency, shape factor, power spectral density.

[0028] Optionally, fusing the self-attention features and the frequency-domain features through the fusion layer to obtain the target fusion features output by the fusion layer includes:

[0029] The self-attention features and the frequency-domain features are concatenated through the fusion layer to obtain the target fusion features output by the fusion layer.

[0030] According to the second aspect of the embodiments of the present specification, a health feature extraction device for a front-end isolated switching power supply of an aging platform is provided, including:

[0031] An acquisition module configured to acquire a plurality of historical signal values of analog signals of a front-end isolated switching power supply of an aging platform;

[0032] A first encoding module configured to encode the plurality of historical signal values through a Transformer model encoder in a multi-layer model encoder to obtain self-attention features in the plurality of historical signal values, wherein the multi-layer model encoder is a neural network structure for fusing context features constructed by the Transformer model encoder and a stacked autoencoder, including an input layer, an intermediate layer, and an output layer, and the intermediate layer includes a sequentially connected Transformer model encoder structure and a stacked autoencoder;

[0033] A feature fusion module configured to extract frequency-domain features of the plurality of historical signal values, and fuse the self-attention features and the frequency-domain features to generate target fusion features;

[0034] A second encoding module configured to encode the target fusion features through the stacked autoencoder in the multi-layer model encoder to generate health features corresponding to the front-end isolated switching power supply of the aging platform.

[0035] According to the third aspect of the embodiments of the present specification, a computing device is provided, including:

[0036] A memory and a processor;

[0037] The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the steps of any one of the health feature extraction methods for the front-end isolated switching power supply of the aging platform.

[0038] According to the fourth aspect of the embodiments of the present specification, a computer-readable storage medium is provided, which stores computer-executable instructions, and when the instructions are executed by a processor, the steps of any one of the health feature extraction methods for the front-end isolated switching power supply of the aging platform are implemented.

[0039] According to the fifth aspect of the embodiments of the present specification, a computer program is provided, wherein when the computer program is executed on a computer, the computer is made to execute the steps of the above-mentioned health feature extraction method for the front-end isolated switching power supply of the aging platform.

[0040] In the embodiment of the present specification, multiple historical signal values of the analog signal of the isolated switching power supply at the front end of the aging table are obtained, and the multiple historical signal values are encoded by the Transformer model encoder in the multi-layer model encoder to obtain the self-attention features in the multiple historical signal values. The multi-layer model encoder is a neural network structure that fuses context features constructed by the Transformer model encoder and the stacked autoencoder, including an input layer, an intermediate layer, and an output layer. The intermediate layer includes a Transformer model encoder structure and a stacked autoencoder connected in sequence; the frequency domain features of the multiple historical signal values are extracted, and the self-attention features are fused with the frequency domain features to generate target fusion features; the target fusion features are encoded by the stacked autoencoder in the multi-layer model encoder to generate the health features corresponding to the isolated switching power supply at the front end of the aging table. In the embodiment of the present specification, the historical signal values of the analog signal are encoded once by the Transformer model encoder to obtain the self-attention features, then the frequency domain features of the historical signal values of the analog signal are extracted, and then the concatenation result of the self-attention features and the frequency domain features is encoded twice by the stacked autoencoder to achieve deep feature fusion. Through this processing method, the temporal dependence relationship and change trend between multiple historical signal values can be directly mapped into the hidden layer depth features, which is beneficial to ensuring the accuracy of the output result of the health features of the isolated switching power supply at the front end of the aging table. Description of the Drawings

[0041] Figure 1 is a flowchart of a method for extracting health features of an isolated switching power supply at the front end of an aging table provided by an embodiment of the present specification;

[0042] Figure 2 is a flowchart of a processing procedure for feature extraction provided by an embodiment of the present specification;

[0043] Figure 3 is a schematic structural diagram of a Transformer model encoder provided by an embodiment of the present specification;

[0044] Figure 4 is another flowchart of a processing procedure for feature extraction provided by an embodiment of the present specification;

[0045] Figure 5 is a schematic structural diagram of a stacked autoencoder provided by an embodiment of the present specification;

[0046] Figure 6 is a schematic structural diagram of a device for extracting health features of an isolated switching power supply at the front end of an aging table provided by an embodiment of the present specification;

[0047] Figure 7 It is a structural block diagram of a computing device provided by an embodiment of this specification. Detailed implementation manners

[0048] In the following description, many specific details are set forth in order to provide a thorough understanding of this specification. However, this specification can be implemented in many other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the connotation of this specification. Therefore, this specification is not limited by the specific implementations disclosed below.

[0049] The terms used in one or more embodiments of this specification are for the purpose of describing specific embodiments only and are not intended to limit one or more embodiments of this specification. The singular forms "a", "the", and "said" used in one or more embodiments of this specification and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term "and / or" used in one or more embodiments of this specification refers to and includes any or all possible combinations of one or more of the associated listed items.

[0050] It should be understood that although the terms first, second, etc. may be used in one or more embodiments of this specification to describe various information, such information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of one or more embodiments of this specification, the first may also be referred to as the second, and similarly, the second may also be referred to as the first. Depending on the context, the word "if" as used herein may be interpreted as "when" or "while" or "in response to determining".

[0051] First, the noun terms related to one or more embodiments of this specification are explained.

[0052] Stacked Auto-Encoders (SAE): It is a deep learning model, originating from the concept of traditional auto-encoders, and forming a deep neural network structure by stacking multiple simple auto-encoders layer by layer.

[0053] Current high-temperature aging products can achieve long-term monitoring of the test environment by means of increasing in-machine long-term monitoring, overstress protection mechanisms, etc., and can respond in a timely manner when the product fails, enabling after-the-fact maintenance based on failure data. However, due to the lack of active guarantee technologies for high-temperature aging products, it is difficult for existing aging 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 and the damage of millions of test devices, or the invalidation of the aging test due to adverse effects such as additional stress introduced during the test period caused by 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 failures of integrated circuits, greatly reduce the failure rate of integrated circuits, and avoid large-scale electronic system failures and outages such as new energy vehicles, civil airliners, and energy storage substations caused by integrated circuit failures.

[0054] Therefore, the development of an intelligent guarantee system for test quality not only has important economic value, but also can help China's high-end test equipment move towards the forefront of the world.

[0055] Currently, it is difficult for active guarantee technologies to adapt to high-temperature aging products, and the key difficulties mainly include: (1) it is difficult to determine the health benchmark due to inconsistent test environments; (2) it is difficult to calculate the performance degradation trends of different levels due to the multi-structural levels of products; (3) it is difficult to construct a fault self-healing strategy due to the complex composition of fault sources.

[0056] In addition, traditional time series extrapolation prediction methods usually adopt the strategy of time series decomposition, predict by decomposing the time series into trend terms, seasonal terms, residual terms, etc. respectively, and finally fuse the prediction results of each item to obtain the time series extrapolation prediction sequence of parameters. Although the output signal can reflect the degradation process of devices and even modules, this change is relatively weak. Considering the training and response time of the later model, the original signal is obviously not very suitable for the prediction model. Therefore, it is necessary to extract data that can represent fault characteristics from a large amount of original signals, that is, to perform dimensionality reduction, noise reduction, etc. on the original data. For switched-mode power supply circuits, due to the existence of tolerances, nonlinearities, etc., and high operating frequencies, both high-frequency and low-frequency signals are relatively rich, which poses relatively high requirements for feature extraction.

[0057] To address the above problems, in response to the urgent need for independent guarantee of the performance of high-end test equipment in key fields, by accurately monitoring the health characteristics of high-temperature aging products, to ensure the quality of the long-cycle operation environment of the test based on health characteristics and meet the batch aging requirements of large-scale integrated circuits, the relevant technical system can be effectively extended to the same type of aging systems to enhance the key scientific and technological strength of the integrated circuit testing industry.

[0058] Based on this, an embodiment of this specification proposes a feature extraction method based on dual auto-encoding time-frequency feature fusion. This method combines frequency-domain features and self-attention features, and realizes feature fusion through a quadratic auto-encoding mechanism, which can directly map the temporal dependence and change trend of the original parameters into the hidden layer depth characteristics, avoiding the problem of sequence decomposition in traditional methods, and providing a more practical method for the extrapolation prediction problem of the degradation time series of key parameters of the isolated switching power supply at the front end of the aging bench.

[0059] In this specification, a method for extracting health features of an isolated switching power supply at the front end of an aging bench is provided. This specification also relates to a device for extracting health features of an isolated switching power supply at the front end of an aging bench, a computing device, a computer-readable storage medium, and a computer program, which will be described in detail one by one in the following embodiments.

[0060] Figure 1 The flowchart of a method for extracting health features of an isolated switching power supply at the front end of an aging bench provided according to an embodiment of this specification is shown, which specifically includes the following steps.

[0061] Step 102: Obtain multiple historical signal values of the analog signals of the isolated switching power supply at the front end of the aging bench.

[0062] Specifically, the aging bench is an integrated circuit high-temperature aging test bench, which accelerates various physical and chemical reaction processes inside the components by continuously applying a certain electrical stress to the components for a long time, promotes the early exposure of various potential faults inside the components, so as to eliminate early failure products, and makes the electronic components enter a period with low failure rate and relatively stable from the beginning of use. And the front-end isolated switching power supply is used to provide the required power for the entire aging test bench.

[0063] Since the isolated switching power supply at the front end of the aging test bench generates monitoring signals such as current and voltage during operation, therefore, the analog signals described in the embodiments of this specification can include voltage signals and current signals, and the current signals can further include capacitor current signals and inductor current signals, and the voltage signals can further include average output voltage signals and ripple voltage signals.

[0064] When obtaining the analog signals of the isolated switching power supply at the front end of the aging bench, high-precision sensors are used to collect the output voltage and current signals, ensuring that the signal accuracy error is less than ±0.5%. For sensitive signals (such as capacitor current and inductor current), preprocessing is performed through a digital signal amplifier, and the gain coefficient is dynamically adjusted to enhance the weak signal characteristics and avoid feature loss caused by noise interference. In addition, according to the preset accuracy index (such as the power output stability requirement), the sampling frequency and signal amplification factor are adaptively adjusted to ensure that the signal-to-noise ratio of the key signals meets the adaptive threshold during the feature extraction stage.

[0065] In addition, the multiple historical signal values described in the embodiments of this specification may be historical time-series data of analog signals.

[0066] Step 104: Encode the multiple historical signal values through the Transformer model encoder in the multi-layer model encoder to obtain the self-attention features in the multiple historical signal values. The multi-layer model encoder is a neural network structure that fuses context features constructed by the Transformer model encoder and the stacked autoencoder, including an input layer, an intermediate layer, and an output layer. The intermediate layer includes a Transformer model encoder structure and a stacked autoencoder connected in sequence.

[0067] In an alternative embodiment, the multi-layer model encoder further includes a multi-head attention layer, a frequency-domain feature extraction layer, and a fusion layer. The frequency-domain feature extraction layer is used to extract frequency-domain features, and the fusion layer is used to fuse the data features extracted by the Transformer model encoder and the frequency-domain feature extraction layer;

[0068] Correspondingly, the method further includes:

[0069] Obtain the transformation result generated by the linear transformation of the multiple historical signal values by the multi-head attention layer;

[0070] Encode the transformation result through the Transformer model encoder to generate the self-attention features in the multiple historical signal values;

[0071] Extract the frequency-domain features of the multiple historical signal values output by the frequency-domain feature extraction layer by the frequency-domain feature extraction layer for the transformation result;

[0072] Fuse the self-attention features and the frequency-domain features through the fusion layer to obtain the target fusion features output by the fusion layer.

[0073] Among them, the fusing the self-attention features and the frequency-domain features through the fusion layer to obtain the target fusion features output by the fusion layer includes:

[0074] Concatenate the self-attention features and the frequency-domain features through the fusion layer to obtain the target fusion features output by the fusion layer.

[0075] Specifically, the multi-layer model encoder is a multi-layer Transformer model encoder, which includes an input layer, an intermediate layer, and an output layer. The intermediate layer includes a sequentially connected Transformer model encoder structure and a stacked autoencoder. Among them, the Transformer model encoder includes a multi-head attention sub-layer, a fully connected layer, and a feed-forward neural network. In addition, the multi-layer Transformer model encoder also includes a frequency domain feature extraction layer and a fusion layer. The multi-head attention sub-layer is used to extract multi-head attention features in historical time series data. The frequency domain feature extraction layer is used to extract frequency domain features in historical time series data. The fusion layer is used to fuse the multi-head attention features extracted by the multi-head attention sub-layer and the frequency domain features extracted by the frequency domain feature extraction layer to achieve the purpose of fully extracting features.

[0076] In an optional implementation manner, after obtaining multiple historical signal values of the analog signal of the isolated switching power supply at the front end of the aging station, the following steps are further included:

[0077] Perform sliding window cutting on the multiple historical signal values according to the first window width and the first step length to generate a first sliding window cutting result, and determine an initial sample data set based on the first sliding window cutting result;

[0078] Perform normalization processing on the initial sample data set to generate a target sample data set;

[0079] Correspondingly, encoding the multiple historical signal values through the Transformer model encoder in the multi-layer model encoder includes:

[0080] Encoding the target sample data set through the Transformer model encoder in the multi-layer model encoder.

[0081] A processing procedure flowchart for feature extraction provided by an embodiment of this specification is as Figure 2 shown. It includes the following steps.

[0082] Step 202: Obtain the fault prediction data of the isolated switching power supply at the front end of the aging station.

[0083] Step 204: Perform comprehensive preprocessing on the fault data.

[0084] Step 206: Perform sliding window cutting on the preprocessing result according to the first window width and the first step length to generate a first sliding window cutting result, and determine an initial sample data set based on the first sliding window cutting result.

[0085] Step 208: Perform normalization processing on the initial sample data set to generate a target sample data set.

[0086] Step 210: Construct a training dataset and a test dataset based on the target sample dataset.

[0087] Step 212: Perform multi-head attention feature extraction on the training dataset through the Transformer model encoder.

[0088] Step 214: Perform frequency-domain feature extraction on the training dataset.

[0089] During the frequency-domain feature extraction process, for sensitive frequency bands (such as the resonant frequency region of the isolated switching power supply at the front end of the aging station), the energy spectral density of the target frequency band is enhanced through a band-pass filter, and the feature significance is evaluated in combination with the power spectral entropy algorithm. The frequency-domain feature weights are adaptively selected. For example, in scenarios with high accuracy requirements, the weight ratios of the center frequency and the shape factor are increased to improve the discrimination of the degradation features.

[0090] Step 216: Perform deep feature fusion on the concatenated features through a stacked autoencoder to obtain the health features corresponding to the isolated switching power supply at the front end of the aging station.

[0091] Step 218: Output the health features.

[0092] The feature extraction process of the embodiments of this specification will be described in detail below.

[0093] For the acquisition of the fault prediction data of the isolated switching power supply at the front end of the aging station, the voltage and current signals of the isolated switching power supply at the front end in the integrated circuit high-temperature dynamic aging detection system can be monitored through sensors to obtain the original fault prediction data of the isolated switching power supply at the front end of the aging station, that is, the historical time-series data of the voltage and current signals.

[0094] For the comprehensive preprocessing of the fault data, the collected original fault prediction data can be sent to the fault data preprocessing unit for comprehensive processing to obtain a training dataset and a test dataset. Specifically, it can be achieved through the following steps:

[0095] Step a: Perform sliding window cutting on the historical time-series data to construct an initial sample dataset.

[0096] Specifically, the historical time-series data of any isolated switching power supply collected by the sensor is X, , and X is subjected to sliding window cutting to generate the corresponding initial sample dataset. When the first window width is W and the first step length is S, the number of samples generated is:

[0097]

[0098] Then the corresponding initial sample dataset generated is , for each sample in Take the data with a length of as the training data, and take the data with a length of W - as the prediction data corresponding to this training data.

[0099] Step b: Perform min - max normalization on the training data set.

[0100] To improve the data representation ability and accelerate the convergence speed of the subsequent model training, it is necessary to perform normalization on the training data set. mainly, the amplitude of the original parameters is scaled through the min - max normalization method to complete the linear transformation of the data. For a single sample data , the normalization process is achieved through the following formula to obtain the normalized sample data set .

[0101]

[0102] Step c: Construct the training data set and the test data set.

[0103] Select the first r% of the data from all the data as the training data set, and the remaining data as the test data set for verifying the model prediction performance. Generally speaking, r is generally taken as 60 - 80, and in the embodiments of this specification, r is taken as 70.

[0104] For the multi - head attention feature extraction of the processed fault data based on the Transformer model encoder, the training data set obtained after being processed by the comprehensive data pre - processing module can be sent into the Transformer model encoder and the frequency - domain feature extraction layer respectively to obtain self - attention features and frequency - domain features.

[0105] The structural schematic diagram of a Transformer model encoder provided by the embodiments of this specification is as Figure 3 shown.

[0106] The flow chart of the processing process of another feature extraction provided by the embodiments of this specification is as Figure 4 shown. Specifically, it includes the following steps.

[0107] Step 402: Obtain the training data set.

[0108] Step 404: Construct the Transformer model encoder.

[0109] Step 406: Train the constructed Transformer model encoder.

[0110] Step 408: Perform self - attention feature extraction on the training data set based on the trained Transformer model encoder.

[0111] Step 410: Perform sliding window cutting on the normalized target sample dataset.

[0112] Step 412: Extract at least two initial frequency domain features based on the sliding window cutting results.

[0113] Step 414: Perform normalization processing on at least two initial frequency domain features.

[0114] Step 416: Perform feature concatenation on the self-attention feature and the frequency domain feature.

[0115] When performing multi-head attention feature extraction on the processed fault data based on the Transformer model encoder, first, a Transformer model encoder can be constructed using the training dataset. Specifically, the model can be pre-trained using the training dataset. The Transformer model usually requires the input data to be three-dimensional, so a training dataset needs to be constructed. The purpose of constructing the training dataset is to meet the input requirements of the model. The method is to convert the data format of the training dataset into , where is the number of samples, is the data length of each sample, and 3 is the number of features, representing three features: the output voltage signal, the capacitor current signal, and the inductor current signal at each time point. After the training dataset is constructed, the constructed training sample dataset can be input into Figure 3 the Transformer model encoder shown.

[0116] The Transformer model is a model established based on the Seq-to-Seq framework. Compared with classical deep learning models, the most prominent advantage of the Transformer model is the use of the multi-head attention mechanism. The purpose of the multi-head attention sub-layer is to assign different importance to words / tokens in the sequence from multiple aspects. The Transformer model mainly consists of parts such as input, encoder, decoder, and output.

[0117] The position encoding layer in the Transformer model is to determine the position information of the sequence. Since there are no recursive layers and convolutional layers in RNN and CNN, and only relying on the self-attention mechanism cannot obtain the order information of the input, the order information of the sequence needs to be actively transmitted to the model. The Transformer model uses a combination of sine and cosine functions to perform position encoding on the sequence, and the calculation method is as follows:

[0118]

[0119]

[0120] Among them, pos is the position of the current sequence; i is the dimension; is the dimension of the input feature.

[0121] The multi-head attention mechanism in the Transformer model performs operations in parallel using multiple attention mechanisms, and then stitches together the operation results through a linear transformation. The core technology in the Transformer model is the multi-head attention mechanism, which uses the multi-head attention mechanism to extract the dependency relationship features between data, capture the correlation between data, and establish a context prediction model. The calculation method is as follows:

[0122]

[0123]

[0124]

[0125]

[0126]

[0127]

[0128] Among them, Q is the query matrix; K is the key matrix; V is the value matrix; 、 、 are trainable parameter matrices; X is the processed input; is the dimension of the key matrix; 、 、 、 are learnable parameter matrices.

[0129] In the Transformer model, the encoding part and the decoding part also include a feed-forward network and summation and normalization. The calculation formula of the feed-forward neural network is as follows:

[0130]

[0131] Among them, x is the input; 、 、 、 are parameters that can be obtained through training.

[0132] The calculation formula of summation and normalization is as follows:

[0133]

[0134] Among them, x is the input; is the result after being processed by the module.

[0135] The model processing flow is as follows: The input data is sent to the encoding part after position encoding. The output of the encoding part is flattened and then sent to the decoder to map the feature parameters in the high-dimensional hidden layer to the original input data, thereby training the feature extraction ability of the model.

[0136] Secondly, select appropriate number of iterations and loss function, and input the constructed dataset into the feature extraction model to repeatedly perform the forward propagation and backward propagation iterative calculation process; during this process, continuously adjust the model parameters of the embedding dimension, number of attention heads, number of encoder and decoder layers to complete the pre-training of the model.

[0137] Thirdly, take out the encoding layer and decoding layer of the pre-trained model, retain their weight parameters, and construct the encoder of the trained Transformer model.

[0138] Finally, based on the encoder of the Transformer model that has completed pre-training, perform self-attention feature extraction on the training dataset to obtain the self-attention feature set .

[0139] Step 106: Extract the frequency domain features of the multiple historical signal values, and fuse the self-attention features with the frequency domain features to generate target fusion features.

[0140] In an optional implementation manner, the extracting the frequency domain features of the multiple historical signal values includes:

[0141] Perform sliding window cutting on the target sample dataset according to the second window width and the second step length to generate a second sliding window cutting result, and extract the frequency domain features of the multiple historical signal values based on the second sliding window cutting result.

[0142] Further, the extracting the frequency domain features of the multiple historical signal values based on the second sliding window cutting result includes:

[0143] Perform Fourier transform on each sample dataset to be processed included in the second sliding window cutting result to generate the corresponding time series energy spectrum for each sample dataset to be processed;

[0144] Extract the eigenvalue corresponding to at least two initial frequency domain features of the time series energy spectrum respectively;

[0145] Use the maximum-minimum value normalization algorithm to normalize the eigenvalues to generate the frequency domain features of the multiple historical signal values.

[0146] Wherein, the at least two initial frequency domain features include: peak value, bandwidth, center frequency, shape factor, power spectral density.

[0147] Specifically, the frequency domain features of the multiple historical signal values are extracted. Specifically, the training data set generated by sliding window cutting is used for frequency domain feature extraction based on expert knowledge.

[0148] Furthermore, the target sample data set generated by sliding window cutting can be further subjected to sliding window cutting according to the second window width and the second step length to generate a second sliding window cutting result, and the frequency domain features of the multiple historical signal values are extracted based on the second sliding window cutting result.

[0149] Among them, when performing sliding window cutting on the target sample data set, if the second window width is and the second step length is 1, for the sample it can be cut into samples, and the length of each sample is , that is, is obtained.

[0150] After the sliding window cutting is completed, the Fourier transform can be performed on each to-be-processed sample data set included in the second sliding window cutting result to transform it into the frequency domain, and the corresponding frequency domain representation F′ (time series energy spectrum) is obtained.

[0151] For the frequency domain representation F′ of each to-be-processed sample data set, five initial frequency domain features of the peak value, bandwidth, center frequency, shape factor, and power spectral density of the energy spectrum are respectively extracted. Each initial frequency domain feature includes a frequency domain feature and the corresponding feature value of this frequency domain feature. Among them, for the window data the extracted initial frequency domain feature is , so for the sample the extracted initial frequency domain feature is . After the extraction is completed, the feature values of each initial frequency domain feature can be normalized by using the maximum-minimum value normalization algorithm to generate the frequency domain features of the multiple historical signal values.

[0152] After the frequency domain features and the self-attention features (high-dimensional hidden layer features) are extracted, the extracted frequency domain features and self-attention features can be feature-stitched.

[0153] For the training data set each sample in it is subjected to self-attention feature extraction and frequency domain feature extraction and feature fusion. Let the dimension of the fused feature be , then the training data set can be reorganized into two-dimensional fused feature matrix.

[0154] Step 108: Encode the target fusion feature through the stacked autoencoder in the multi-layer model encoder to generate the health feature corresponding to the isolated switching power supply at the front end of the aging platform.

[0155] Specifically, the autoencoder is an unsupervised learning method whose purpose is to learn an effective representation or encoding of data and achieve self-learning by reconstructing the input data. The stacked autoencoder (SAE) is an extension based on this. It maps the original data to an increasingly abstract and compact feature space through multi-layer non-linear transformations, thereby realizing deeper feature learning and representation learning, which is of great significance for the dimensionality reduction, feature extraction, and classification tasks of complex data.

[0156] In the embodiments of this specification, encoding the target fusion feature through the stacked autoencoder in the multi-layer model encoder specifically includes constructing the stacked autoencoder and decoder, training the stacked autoencoder and decoder, and using the stacked autoencoder to encode the target fusion feature.

[0157] The structural schematic diagram of a stacked autoencoder provided by an embodiment of this specification is as Figure 5 shown.

[0158] First, for constructing the stacked autoencoder and decoder, the model structure of the stacked autoencoder is as Figure 5 shown. The number of encoding layers is the same as the number of decoding layers, which can enable the model to have better secondary encoding ability for deep features.

[0159] Secondly, use the two-dimensional fusion feature matrix obtained in the previous step for pre-training the stacked autoencoder model. Take the two-dimensional fusion feature matrix as the input and output of the stacked autoencoder model, select an appropriate loss function and number of iterations, and complete the forward propagation and backward propagation iterative calculation process to make the model continuously reconstruct its own input. Finally, extract the encoding layer from the pre-trained stacked autoencoder model as the available stacked autoencoder.

[0160] Finally, perform secondary auto-encoding on the fusion feature based on the pre-trained stacked autoencoder to obtain a secondary encoding feature set , and this secondary encoding feature set is the health feature corresponding to the isolated switching power supply at the front end of the aging platform.

[0161] After obtaining multiple health features, binning processing can also be performed on the health features, and based on at least two binned data sets included in the binning processing results, feature vectors of the isolated switching power supply at the front end of the aging platform can be constructed respectively;

[0162] Construct a training sample set based on the feature vectors and the corresponding status labels of the feature vectors, and train the support vector machine diagnosis model to be trained based on the training sample set, where the trained support vector machine diagnosis model is used to detect and diagnose faults in the front-end isolated switching power supply of the aging platform.

[0163] Among them, the training sample set includes a training sample subset and a test sample subset, and the status labels include a healthy status label and a fault status label.

[0164] In an alternative embodiment, the healthy features can be binned by the K-means binning algorithm.

[0165] In another alternative embodiment, the reference information on the value range and the reference information on the change law corresponding to the healthy features can also be obtained;

[0166] According to the fault information of the front-end isolated switching power supply of the aging platform, determine the abnormal value interval corresponding to the healthy features;

[0167] Based on the reference information on the value range, the reference information on the change law, and the abnormal value interval, determine the binning interval corresponding to the healthy features;

[0168] Bin the healthy features according to the binning interval.

[0169] Specifically, based on the parameter distribution range of the front-end isolated switching power supply of the aging platform in the normal operating state (healthy state), determine the normal interval of the healthy features; combine the influence of possible fault scenarios in the actual scenario on the healthy features to delimit the abnormal interval of the healthy features, which is convenient for distinguishing the types of faults.

[0170] Through the above binning discretization process, continuous performance parameters can be converted into discrete performance parameters, improving the comparability of data and reducing the complexity in the model training process. The interval setting and discretization results of binning provide efficient and accurate input features for the subsequent support vector machine diagnosis model to be trained.

[0171] In an alternative embodiment, constructing the feature vectors of the front-end isolated switching power supply of the aging platform based on at least two binned data sets included in the binning processing results includes:

[0172] Based on the target healthy features included in the target binned data set, perform feature extraction on the target binned data set, and construct the feature vectors of the front-end isolated switching power supply of the aging platform based on the feature extraction results, where the target binned data set is each of the at least two binned data sets.

[0173] Further, the feature extraction from the target binned dataset based on the target health features included therein includes:

[0174] Based on the target health features included in the target binned dataset, feature extraction is performed on the target binned dataset to obtain the mean value, standard deviation, and peak value of the corresponding target health features.

[0175] Specifically, the target binned dataset is each of at least two binned datasets, and the health features included in the target binned dataset are the target health features.

[0176] Alternatively, the similarity between different degradation paths of the front-end isolated switching power supply of the aging platform can also be calculated, where different degradation paths are composed of health features corresponding to different historical time series;

[0177] Based on the similarity, clustering is performed on different degradation paths to generate at least two clustering results;

[0178] The health features corresponding to the degradation paths included in different clustering results are respectively used as training data and input into the corresponding fault prediction model to be trained to obtain the trained fault prediction model;

[0179] Among them, the number of fault prediction models to be trained is equal to the number of the at least two clustering results, and each fault prediction model to be trained corresponds to each clustering result one by one.

[0180] In practical applications, the fault prediction model to be trained is a multi-dimensional recurrent neural network model, including an input layer, an intermediate layer, and an output layer. The intermediate layer includes a sequentially connected nearest neighbor component analysis model structure and a multi-dimensional recurrent neural network model structure. The multi-dimensional recurrent neural network model structure includes a fully connected layer, a bidirectional long short-term memory network layer, a bidirectional gated recurrent unit layer, and a splicing layer.

[0181] Further, the nearest neighbor component analysis model structure can be used to process the health features corresponding to different degradation paths of the front-end isolated switching power supply of the aging platform to obtain the similarity between different degradation paths, where the nearest neighbor component analysis model structure processes the health features corresponding to different degradation paths of the front-end isolated switching power supply of the aging platform through the nearest neighbor component analysis algorithm.

[0182] Alternatively, the intermediate layer includes a sequentially connected nearest neighbor component analysis model structure, a Gaussian mixture model structure, and a multi-dimensional recurrent neural network model structure;

[0183] Based on this, different degradation paths can be clustered based on the similarity by the Gaussian mixture model structure, where the Gaussian mixture model structure clusters different degradation paths through the Gaussian mixture algorithm and based on the similarity.

[0184] The embodiments of this specification provide an active guarantee technology for an integrated circuit high-temperature aging test bench to achieve the integrity of the test process and the consistency of the test environment stress, and can reduce the significant property losses caused by the destruction of millions of test devices due to forced interruption during the test process caused by machine failures, or can reduce the adverse effects such as the introduction of additional stress during the test due to the degradation of machine performance, so as to avoid the aging test being recognized as a failure test, thereby facilitating the avoidance of waste of resources. At the same time, ensuring the quality of the high-temperature aging test can avoid over-aging and under-aging failures of integrated circuits, greatly reduce the failure rate of integrated circuits, and is conducive to avoiding the failure shutdown of large-scale electronic systems such as new energy vehicles, civil airliners, and energy storage substations using the same type of integrated circuits due to integrated circuit failures.

[0185] The embodiments of this specification obtain multiple historical signal values of the analog quantity signal of the front-end isolated switching power supply of the aging bench, and encode the multiple historical signal values through the Transformer model encoder in the multi-layer model encoder to obtain the self-attention features in the multiple historical signal values. The multi-layer model encoder is a neural network structure that fuses context features constructed by the Transformer model encoder and the stacked autoencoder, including an input layer, an intermediate layer, and an output layer. The intermediate layer includes a sequentially connected Transformer model encoder structure and a stacked autoencoder; extract the frequency domain features of the multiple historical signal values, and fuse the self-attention features with the frequency domain features to generate target fusion features; encode the target fusion features through the stacked autoencoder in the multi-layer model encoder to generate the health features corresponding to the front-end isolated switching power supply of the aging bench. The embodiments of this specification encode the historical signal values of the analog quantity signal once through the Transformer model encoder to obtain the self-attention features, then extract the frequency domain features of the historical signal values of the analog quantity signal, and then perform secondary encoding on the concatenation result of the self-attention features and the frequency domain features through the stacked autoencoder to achieve deep feature fusion. Through this processing method, the temporal dependence relationship and change trend between multiple historical signal values can be directly mapped into the hidden layer depth features, which is conducive to ensuring the accuracy of the output result of the health features of the front-end isolated switching power supply of the aging bench.

[0186] Corresponding to the above method embodiments, this specification also provides an embodiment of a device for extracting the health features of the front-end isolated switching power supply of the aging bench. Figure 6The figure shows a schematic structural diagram of a health feature extraction device for a front-end isolated switching power supply of an aging platform provided by an embodiment of this specification. As Figure 6 shown, the device includes:

[0187] An acquisition module 602, configured to acquire a plurality of historical signal values of analog quantity signals of a front-end isolated switching power supply of an aging platform;

[0188] A first encoding module 604, configured to encode the plurality of historical signal values through a Transformer model encoder in a multi-layer model encoder to obtain self-attention features in the plurality of historical signal values, wherein the multi-layer model encoder is a neural network structure that fuses context features constructed by the Transformer model encoder and a stacked autoencoder, and includes an input layer, an intermediate layer, and an output layer, and the intermediate layer includes a Transformer model encoder structure and a stacked autoencoder connected in sequence;

[0189] A feature fusion module 606, configured to extract frequency domain features of the plurality of historical signal values, and fuse the self-attention features and the frequency domain features to generate target fusion features;

[0190] A second encoding module 608, configured to encode the target fusion features through the stacked autoencoder in the multi-layer model encoder to generate health features corresponding to the front-end isolated switching power supply of the aging platform.

[0191] Optionally, the multi-layer model encoder further includes a multi-head attention layer, a frequency domain feature extraction layer, and a fusion layer. The frequency domain feature extraction layer is used to extract frequency domain features, and the fusion layer is used to fuse data features extracted by the Transformer model encoder and the frequency domain feature extraction layer;

[0192] Correspondingly, the device further includes a processing module 610, configured to:

[0193] Acquire a transformation result generated by linearly transforming the plurality of historical signal values by the multi-head attention layer;

[0194] Encode the transformation result through the Transformer model encoder to generate self-attention features in the plurality of historical signal values;

[0195] Extract frequency domain features of the plurality of historical signal values output by the frequency domain feature extraction layer by the frequency domain feature extraction layer;

[0196] The self-attention features and the frequency-domain features are fused through the fusion layer to obtain the target fusion features output by the fusion layer.

[0197] Optionally, the processing module 610 is further configured to:

[0198] Perform sliding window cutting on the multiple historical signal values according to a first window width and a first step length to generate a first sliding window cutting result, and determine an initial sample data set based on the first sliding window cutting result;

[0199] Perform normalization processing on the initial sample data set to generate a target sample data set;

[0200] Correspondingly, the first encoding module 604 is further configured to:

[0201] Encode the target sample data set through a Transformer model encoder in a multi-layer model encoder.

[0202] Optionally, the feature fusion module 606 is further configured to:

[0203] Perform sliding window cutting on the target sample data set according to a second window width and a second step length to generate a second sliding window cutting result, and extract the frequency-domain features of the multiple historical signal values based on the second sliding window cutting result.

[0204] Optionally, the feature fusion module 606 is further configured to:

[0205] Perform Fourier transform on each to-be-processed sample data set included in the second sliding window cutting result to generate a corresponding time-series energy spectrum for each to-be-processed sample data set;

[0206] Extract the eigenvalue corresponding to at least two initial frequency-domain features of the time-series energy spectrum respectively;

[0207] Normalize the eigenvalues by using the min-max normalization algorithm to generate the frequency-domain features of the multiple historical signal values.

[0208] Optionally, the at least two initial frequency-domain features include: peak value, bandwidth, center frequency, shape factor, power spectral density.

[0209] Optionally, the processing module 610 is further configured to:

[0210] Concatenate the self-attention features and the frequency-domain features through the fusion layer to obtain the target fusion features output by the fusion layer.

[0211] The above is a schematic solution of a health feature extraction device for a front-end isolated switching power supply of an aging platform in this embodiment. It should be noted that the technical solution of the health feature extraction device for the front-end isolated switching power supply of the aging platform belongs to the same concept as the technical solution of the above-mentioned health feature extraction method for the front-end isolated switching power supply of the aging platform. For the details not described in the technical solution of the health feature extraction device for the front-end isolated switching power supply of the aging platform, reference can be made to the description of the technical solution of the above-mentioned health feature extraction method for the front-end isolated switching power supply of the aging platform.

[0212] Figure 7 FIG. shows a block diagram of a computing device 700 provided according to an embodiment of the present specification. The components of the computing device 700 include, but are not limited to, a memory 710 and a processor 720. The processor 720 is connected to the memory 710 via a bus 730, and a database 750 is used to store data.

[0213] The computing device 700 further includes an access device 740, which enables the computing device 700 to communicate via one or more networks 760. Examples of these networks include the Public Switched Telephone Network (PSTN), Local Area Network (LAN), Wide Area Network (WAN), Personal Area Network (PAN), or a combination of communication networks such as the Internet. The access device 740 may include one or more of any type of wired or wireless network interfaces (e.g., Network Interface Card (NIC)), such as an IEEE802.11 Wireless Local Area Network (WLAN) wireless interface, a Worldwide Interoperability for Microwave Access (Wi-MAX) interface, an Ethernet interface, a Universal Serial Bus (USB) interface, a cellular network interface, a Bluetooth interface, a Near Field Communication (NFC) interface, and so on.

[0214] In an embodiment of the present specification, the above components of the computing device 700 and Figure 7 other components not shown in the figure may also be connected to each other, for example, via a bus. It should be understood that Figure 7 the block diagram of the computing device shown is only for illustrative purposes and is not a limitation on the scope of the present specification. Those skilled in the art can add or replace other components as needed.

[0215] The computing device 700 can be any type of stationary or mobile computing device, including mobile computers or mobile computing devices (e.g., tablet computers, personal digital assistants, laptop computers, notebook computers, netbooks, etc.), mobile phones (e.g., smartphones), wearable computing devices (e.g., smart watches, smart glasses, etc.) or other types of mobile devices, or stationary computing devices such as desktop computers or PCs. The computing device 700 can also be a mobile or stationary server.

[0216] Among them, the processor 720 is used to execute the following computer-executable instructions. When the computer-executable instructions are executed by the processor, the steps of the above-mentioned method for extracting the health characteristics of the isolated switching power supply at the front end of the aging platform are implemented.

[0217] The above is a schematic solution of a computing device according to this embodiment. It should be noted that the technical solution of this computing device and the technical solution of the above-mentioned method for extracting the health characteristics of the isolated switching power supply at the front end of the aging platform belong to the same concept. For the details not described in detail in the technical solution of the computing device, reference can be made to the description of the technical solution of the above-mentioned method for extracting the health characteristics of the isolated switching power supply at the front end of the aging platform.

[0218] An embodiment of this specification also provides a computer-readable storage medium, which stores computer-executable instructions. When the computer-executable instructions are executed by the processor, the steps of the above-mentioned method for extracting the health characteristics of the isolated switching power supply at the front end of the aging platform are implemented.

[0219] The above is a schematic solution of a computer-readable storage medium according to this embodiment. It should be noted that the technical solution of this storage medium and the technical solution of the above-mentioned method for extracting the health characteristics of the isolated switching power supply at the front end of the aging platform belong to the same concept. For the details not described in detail in the technical solution of the storage medium, reference can be made to the description of the technical solution of the above-mentioned method for extracting the health characteristics of the isolated switching power supply at the front end of the aging platform.

[0220] An embodiment of this specification also provides a computer program. Among them, when the computer program is executed on a computer, the computer is made to execute the steps of the above-mentioned method for extracting the health characteristics of the isolated switching power supply at the front end of the aging platform.

[0221] The above is a schematic solution of a computer program according to this embodiment. It should be noted that the technical solution of this computer program and the technical solution of the above-mentioned method for extracting the health characteristics of the isolated switching power supply at the front end of the aging platform belong to the same concept. For the details not described in detail in the technical solution of the computer program, reference can be made to the description of the technical solution of the above-mentioned method for extracting the health characteristics of the isolated switching power supply at the front end of the aging platform.

[0222] The above describes specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than in the embodiments and still achieve the desired result. Additionally, the processes depicted in the figures do not necessarily require the particular order or sequential order shown to achieve the desired result. In certain implementations, multitasking and parallel processing are also possible or may be advantageous.

[0223] The computer instructions include computer program code, which may be in the form of source code, object code, executable files, or some intermediate forms, etc. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately increased or decreased according to the requirements of legislation and patent practice within the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, the computer-readable medium does not include electrical carrier signals and telecommunication signals.

[0224] It should be noted that for the foregoing method embodiments, for the sake of simplicity of description, they are all expressed as a series of action combinations. However, those skilled in the art should know that the embodiments of this specification are not limited by the described action sequence, because according to the embodiments of this specification, certain steps can be performed in other sequences or simultaneously. Secondly, those skilled in the art should also know that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the embodiments of this specification.

[0225] In the above embodiments, the descriptions of each embodiment have their own emphases. For the parts not detailed in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0226] The preferred embodiments of this specification disclosed above are only used to help explain this specification. The alternative embodiments do not elaborate on all details and do not limit the invention to the specific embodiments described. Obviously, many modifications and changes can be made according to the content of the embodiments of this specification. These embodiments are selected and specifically described in this specification to better explain the principles and practical applications of the embodiments of this specification, so that those skilled in the art can understand and utilize this specification well. This specification is only limited by the claims and their full scope and equivalents.

Claims

1. A method for extracting health characteristics of a front-end isolated switching power supply of an aging platform, comprising: Obtaining multiple historical signal values of the analog signals of the front-end isolated switching power supply of the aging platform, and encoding the multiple historical signal values of the analog signals of the front-end isolated switching power supply of the aging platform through a Transformer model encoder in a multi-layer model encoder to obtain self-attention features in the multiple historical signal values of the analog signals of the front-end isolated switching power supply of the aging platform, wherein the analog signals include an average output voltage signal, a ripple voltage signal, a capacitor current signal, and an inductor current signal; the multi-layer model encoder is a neural network structure that fuses context features constructed by the Transformer model encoder and a stacked autoencoder, including an input layer, an intermediate layer, an output layer, a frequency-domain feature extraction layer, and a fusion layer, and the intermediate layer includes a sequentially connected Transformer model encoder structure and a stacked autoencoder; Among them, obtaining the self-attention features in the multiple historical signal values of the analog signals of the front-end isolated switching power supply of the aging platform includes: sending the input data into the encoding part after position encoding, flattening the output of the encoding part and sending it into the decoder to map the feature parameters in the high-dimensional hidden layer to the original input data, so as to train the feature extraction ability of the model; selecting appropriate iteration times and loss functions, and inputting the constructed data set into the feature extraction model to repeatedly perform forward propagation and backward propagation iterative calculation processes; during this process, continuously adjusting the model parameters of the embedding dimension, the number of attention heads, the number of encoder and decoder layers to complete the pre-training of the model; taking out the encoding layer and the decoding layer of the pre-trained model and retaining their weight parameters, and constructing them into a trained Transformer model encoder; based on the pre-trained Transformer model encoder, performing self-attention feature extraction on the training data set to obtain a self-attention feature set; Performing frequency-domain feature extraction on the multiple historical signal values of the analog signals of the front-end isolated switching power supply of the aging platform through the frequency-domain feature extraction layer in the multi-layer model encoder to obtain frequency-domain features in the multiple historical signal values of the analog signals of the front-end isolated switching power supply of the aging platform; Splicing the self-attention features of the analog signals of the front-end isolated switching power supply of the aging platform and the frequency-domain features of the analog signals of the front-end isolated switching power supply of the aging platform through the fusion layer in the multi-layer model encoder to obtain target fusion features of the analog signals of the front-end isolated switching power supply of the aging platform; Encoding the target fusion features of the analog signals of the front-end isolated switching power supply of the aging platform through the stacked autoencoder in the multi-layer model encoder to generate the corresponding health features of the front-end isolated switching power supply of the aging platform.

2. The method for extracting the health characteristics of the isolated switching power supply at the front end of the aging platform according to claim 1, wherein the multi-layer model encoder further includes a multi-head attention layer, the frequency domain feature extraction layer is used to extract frequency domain features, and the fusion layer is used to fuse the data features extracted by the Transformer model encoder and the frequency domain feature extraction layer; Correspondingly, the method further includes: Obtaining the transformation result generated by the linear transformation of the multi-head attention layer on the plurality of historical signal values; Encoding the transformation result through the Transformer model encoder in the multi-layer model encoder to generate self-attention features in the plurality of historical signal values; Performing frequency domain feature extraction on the transformation result through the frequency domain feature extraction layer to obtain the frequency domain features in the plurality of historical signal values output by the frequency domain feature extraction layer.

3. The method for extracting the health characteristics of the isolated switching power supply at the front end of the aging platform according to claim 1, after obtaining the plurality of historical signal values of the analog signal of the isolated switching power supply at the front end of the aging platform, further including: Performing sliding window cutting on the plurality of historical signal values according to the first window width and the first step length to generate a first sliding window cutting result, and determining an initial sample data set based on the first sliding window cutting result; Performing normalization processing on the initial sample data set to generate a target sample data set; Correspondingly, the encoding of the plurality of historical signal values through the Transformer model encoder in the multi-layer model encoder includes: Encoding the target sample data set through the Transformer model encoder in the multi-layer model encoder.

4. The method for extracting the health characteristics of the isolated switching power supply at the front end of the aging platform according to claim 3, wherein the extraction of the frequency domain features of the plurality of historical signal values includes: Performing sliding window cutting on the target sample data set according to the second window width and the second step length to generate a second sliding window cutting result, and extracting the frequency domain features of the plurality of historical signal values based on the second sliding window cutting result.

5. The method for extracting the health characteristics of the isolated switching power supply at the front end of the aging platform according to claim 4, wherein the extraction of the frequency domain features of the plurality of historical signal values based on the second sliding window cutting result includes: Performing Fourier transform on each to-be-processed sample data set included in the second sliding window cutting result to generate a corresponding time series energy spectrum for each to-be-processed sample data set; Respectively extracting the eigenvalue corresponding to at least two initial frequency domain features of the time series energy spectrum; Normalizing the eigenvalue by using the maximum-minimum normalization algorithm to generate the frequency domain features of the plurality of historical signal values.

6. The method for extracting the health characteristics of the front-end isolated switching power supply of the aging table according to claim 5, wherein the at least two initial frequency domain characteristics include: Peak value, bandwidth, center frequency, shape factor, power spectral density.

7. An apparatus for extracting the health characteristics of an isolated switching power supply at the front end of an aging platform, comprising: An acquisition module configured to acquire a plurality of historical signal values of an analog signal of an isolated switching power supply at the front end of an aging platform; the analog signal includes an average output voltage signal, a ripple voltage signal, a capacitor current signal, and an inductor current signal; The first encoding module is configured to encode multiple historical signal values of the analog signal of the isolated switching power supply at the front end of the aging platform through the Transformer model encoder in the multi-layer model encoder, so as to obtain the self-attention features in the multiple historical signal values of the analog signal of the isolated switching power supply at the front end of the aging platform. The multi-layer model encoder is a neural network structure that fuses context features constructed by the Transformer model encoder and the stacked autoencoder, and includes an input layer, an intermediate layer, an output layer, a frequency-domain feature extraction layer, and a fusion layer. The intermediate layer includes a sequentially connected Transformer model encoder structure and a stacked autoencoder. Among them, obtaining the self-attention features in the multiple historical signal values of the analog signal of the isolated switching power supply at the front end of the aging platform includes: sending the input data into the encoding part after position encoding, flattening the output of the encoding part and sending it into the decoder to map the feature parameters in the high-dimensional hidden layer to the original input data, so as to train the feature extraction ability of the model; selecting appropriate iteration times and loss functions, and inputting the constructed dataset into the feature extraction model to repeatedly perform the forward propagation and backward propagation iterative calculation processes; in this process, continuously adjusting the model parameters of the embedding dimension, the number of attention heads, the number of encoder and decoder layers to complete the pre-training of the model; taking out the encoding layer and the decoding layer of the pre-trained model and retaining their weight parameters, and constructing them into the trained Transformer model encoder; based on the pre-trained Transformer model encoder, performing self-attention feature extraction on the training dataset to obtain a self-attention feature set; The feature fusion module is configured to perform frequency-domain feature extraction on multiple historical signal values of the analog signal of the isolated switching power supply at the front end of the aging platform through the frequency-domain feature extraction layer in the multi-layer model encoder to obtain the frequency-domain features in the multiple historical signal values of the analog signal of the isolated switching power supply at the front end of the aging platform, and splicing the self-attention features of the analog signal of the isolated switching power supply at the front end of the aging platform and the frequency-domain features of the analog signal of the isolated switching power supply at the front end of the aging platform through the fusion layer in the multi-layer model encoder to obtain the target fusion features of the analog signal of the isolated switching power supply at the front end of the aging platform; The second encoding module is configured to encode the target fusion features of the analog signal of the isolated switching power supply at the front end of the aging platform through the stacked autoencoder in the multi-layer model encoder to generate the health features corresponding to the isolated switching power supply at the front end of the aging platform.

8. A computing device, comprising: a memory and a processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, the steps of the method for extracting the health features of the isolated switching power supply at the front end of the aging platform according to any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium stores computer-executable instructions, and when the computer-executable instructions are executed by a processor, the steps of the method for extracting the health characteristics of the front-end isolated switching power supply of the aging station described in any one of claims 1 to 6 are implemented.

Images