Time-frequency characteristic deep fusion method for secondary power supply of high-temperature aging test equipment

Abstract

The invention discloses a high-temperature aging test equipment secondary power supply time-frequency characteristic deep fusion method. The method comprises the following steps: acquiring a fault data set of a high-temperature aging test equipment device; preprocessing the fault data set to obtain a training data set and a test data set; performing short-time Fourier transform on each key parameter time sequence data in the training data set to obtain a time-frequency graph; performing random masking on each time-frequency graph and inputting the time-frequency graphs into a preset model for training to generate a feature extraction model; testing the feature extraction model according to the test data set; and after the test is passed, inputting key parameter time sequence data of a to-be-predicted device into the feature extraction model, and predicting to obtain a target time-frequency domain feature. According to the method provided by the invention, the characteristics of the secondary switching power supply can be accurately extracted, so that the health state of the secondary switching power supply of the high-temperature aging test equipment can be evaluated.

Description

Technical Field

[0001] The present invention relates to the technical field of non-destructive reliability screening of integrated circuits, and in particular to a method and device for deep fusion of time-frequency characteristics of a secondary power supply of a high-temperature aging test device. Background Art

[0002] Existing high-temperature aging products can achieve long-term monitoring of the test environment by adding long-term monitoring in the machine and overstress protection mechanisms, and can respond in time when the product fails, realizing post-maintenance based on fault data. However, due to the lack of active protection technology for high-temperature aging products, existing high-temperature aging products make it difficult for related products to achieve the integrity of the test process and the consistency of the test environment stress. It is very easy to cause the test process to be interrupted due to product failure, resulting in major property losses such as damage to millions of test devices (also known as test pieces), or the aging test is identified as a failure test due to adverse effects such as additional stress introduced during the test due to product performance degradation, 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 failures in integrated circuits, greatly reduce the failure rate of integrated circuits, and avoid large-scale electronic system failures and shutdowns such as new energy vehicles, civil airliners, and energy storage transmission stations due to integrated circuit failures.

[0003] High-temperature aging products such as integrated circuit high-temperature aging test equipment accelerate various physical and chemical reaction processes inside components by continuously applying certain electrical stress to components over a long period of time, prompting various potential faults inside components to be exposed early, thereby eliminating early failure products and making electronic components enter a period of low failure rate and relative stability from the beginning of use. The secondary switching power supply provides stable voltage and current for integrated circuits during high-temperature dynamic aging tests to ensure the consistency of test conditions; it can flexibly adjust the output voltage and current according to different test requirements to meet the testing needs of different types of integrated circuits.

[0004] The secondary switching power supply plays a vital role in the high temperature aging test equipment. The failure of the secondary switching power supply may cause failures in other parts of the system, such as the drive control detection board, further affecting the normal operation of the system. The failure of the secondary switching power supply may cause safety hazards such as overheating and short circuit, endangering the safety of operators and the integrity of the equipment. Therefore, it is very necessary to perform fault prediction and health assessment on the secondary switching power supply of the aging test bench.

[0005] In the prior art, a time series extrapolation prediction method is used to predict faults of a secondary switching power supply. This method usually adopts a time series decomposition strategy, by decomposing the time series into trend terms, seasonal terms, residual terms, etc., and predicting them separately, and finally integrating the prediction results to obtain a time series extrapolation prediction sequence of parameters. Although the output signal can reflect the degradation process of the device or even the module, this change is relatively weak. At the same time, considering the training and reaction time of the later model, the original signal is obviously not suitable for the prediction model. Therefore, it is necessary to extract data that can represent the fault characteristics from a large number of original signals, that is, to perform dimensionality reduction, denoising and other processing on the original data. For the secondary switching power supply circuit, due to the existence of tolerance, nonlinearity and other conditions, the operating frequency is high, so the high and low frequency signals are rich, and the feature extraction requirements are high.

[0006] It can be seen that there is an urgent need for technical personnel in this field to provide a solution for extracting the characteristics of the secondary switching power supply of high-temperature aging test equipment. Summary of the invention

[0007] The purpose of the embodiments of the present invention is to provide a method and device for deep fusion of time-frequency characteristics of secondary power supply of high-temperature aging test equipment, which can solve the above-mentioned problems existing in the prior art.

[0008] In order to solve the above technical problems, the present invention provides the following technical solutions: The embodiment of the present invention provides a method for deep fusion of time-frequency characteristics of secondary power supply of high-temperature aging test equipment, wherein the method comprises: Collecting a fault data set of a high temperature aging test equipment component; wherein the fault data includes a plurality of key parameter timing data; Preprocessing the fault data set to obtain a training data set and a test data set; Performing short-time Fourier transform on the time series data of each key parameter in the training data set to obtain a time-frequency diagram; Performing random masking on each of the time-frequency graphs and inputting the mask into a preset model for training to generate a feature extraction model; Testing the feature extraction model according to the test data set; After the test is passed, the key parameter timing data of the device to be predicted is input into the feature extraction model to predict the target time-frequency domain features.

[0009] Optionally, the step of preprocessing the fault data set to obtain a training data set and a test data set includes: For each key parameter in the fault data set, the key parameter time series data is subjected to sliding window cutting to construct a sample data set; Normalizing each data sample in the sample data set to obtain a normalized sample data set; A preset percentage of sample data is selected from the normalized sample data set as a training data set, and the remaining sample data is selected as a test data set; wherein the training data set is used to train the preset model, and the test data set is used to verify the prediction accuracy of the trained feature extraction model.

[0010] Optionally, the step of performing short-time Fourier transform on each time series data in the training data set to obtain a time-frequency diagram includes: For each time series data in the training data set, a window function of finite time length is used to intercept the vibration signal in the time series data; Perform Fourier transform on the intercepted vibration signal to obtain a local spectrum in a small range of the corresponding time period; By moving the window function on the time axis, each vibration signal band is gradually analyzed to obtain a set of local frequency spectra of the vibration signal, and a time-frequency diagram is generated.

[0011] Optionally, the window function is a Hanning window function.

[0012] Optionally, the step of randomly masking each of the time-frequency graphs and inputting them into a preset model for training to generate a feature extraction model includes: The time-frequency graph after random masking is input into the preset model, and is sent into the encoding module after position encoding, and the output of the encoding module is sent into the decoder, and the feature parameters in the high-dimensional hidden layer are restored to the time-frequency graph before masking through the decoder; The preset model is trained multiple times iteratively using the restored time-frequency graph to adjust the model parameters, wherein the model parameters include: embedding dimension, number of attention heads, number of encoder layers, and number of decoder layers; Take out the encoder layer and decoder layer of the model after multiple iterations and keep the weight parameters of each layer; A feature extraction model is generated based on each of the encoder layers, each of the decoder layers, and weight parameters of each layer.

[0013] Optionally, the device includes: a switching power supply of a high temperature aging test equipment, a high temperature test chamber of a high temperature aging test equipment, and a high temperature and high humidity test chamber of a high temperature aging test equipment.

[0014] The embodiment of the present invention further provides a device for deep fusion of time-frequency characteristics of secondary power supply of high-temperature aging test equipment, wherein the device comprises: A collection module, used to collect fault data sets of high temperature aging test equipment components; wherein the fault data includes multiple key parameter time series data; A preprocessing module, used to preprocess the fault data set to obtain a training data set and a test data set; A transformation module, used for performing short-time Fourier transformation on the time series data of each key parameter in the training data set to obtain a time-frequency diagram; A mask training module, used to randomly mask each of the time-frequency graphs and input the mask into a preset model for training to generate a feature extraction model; A testing module, used for testing the feature extraction model according to the test data set; The prediction module is used to input the key parameter timing data of the device to be predicted into the feature extraction model after the test is passed, and predict the target time-frequency domain features.

[0015] Optionally, the preprocessing module includes: The first submodule is used to perform sliding window cutting on the key parameter time series data for each key parameter in the fault data set to construct a sample data set; The second submodule is used to perform normalization processing on each data sample in the sample data set to obtain a normalized sample data set; The third submodule is used to select a preset percentage of sample data from the normalized sample data set as a training data set, and the remaining sample data as a test data set; wherein the training data set is used to train the preset model, and the test data set is used to verify the prediction accuracy of the trained feature extraction model.

[0016] Optionally, the transformation module includes: A fourth submodule is used to intercept the vibration signal in each time series data in the training data set by using a window function of a finite time length; The fifth submodule is used to perform Fourier transform on the intercepted vibration signal to obtain a local spectrum in a small range of the corresponding time period; The sixth submodule is used to gradually analyze each vibration signal band to obtain a set of local frequency spectra of the vibration signal by moving the window function on the time axis, and generate a time-frequency diagram.

[0017] Optionally, the window function is a Hanning window function.

[0018] Optionally, the mask training module includes: The seventh submodule is used to input the randomly masked time-frequency graph into the preset model, send it to the encoding module after position encoding, and send the output of the encoding module to the decoder, and restore the feature parameters in the high-dimensional hidden layer to the time-frequency graph before masking through the decoder; An eighth submodule is used to iteratively train the preset model multiple times using the restored time-frequency graph to adjust the model parameters, wherein the model parameters include: embedding dimension, number of attention heads, number of encoder layers, and number of decoder layers; The ninth submodule is used to extract the encoder layer and decoder layer of the model after multiple iterations and adjust, and retain the weight parameters of each layer; The tenth submodule is used to generate a feature extraction model based on each of the encoder layers, each of the decoder layers and the weight parameters of each layer.

[0019] Optionally, the device includes: a switching power supply of a high temperature aging test equipment, a high temperature test chamber of a high temperature aging test equipment, and a high temperature and high humidity test chamber of a high temperature aging test equipment.

[0020] An embodiment of the present invention also provides an electronic device, characterized in that it includes a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other through the communication bus; the memory is used to store computer programs; the processor is used to execute the program stored in the memory to implement any one of the above-mentioned methods for deep fusion of time-frequency characteristics of the secondary power supply of high-temperature aging test equipment.

[0021] The deep fusion scheme of time-frequency characteristics of secondary power supply of high temperature aging test equipment disclosed in the embodiment of the present invention collects the fault data set of high temperature aging test equipment components; preprocesses the fault data set to obtain a training data set and a test data set; performs short-time Fourier transform on the time series data of each key parameter in the training data set to obtain a time-frequency diagram; randomly masks each time-frequency diagram and inputs it into a preset model for training to generate a feature extraction model; tests the feature extraction model according to the test data set; after the test is passed, the key parameter time series data of the device to be predicted is input into the feature extraction model to predict the target time-frequency domain characteristics. The scheme provided by the embodiment of the present invention fully coordinates and refines a large amount of fault data of high temperature aging test equipment components, and randomly masks the time-frequency diagram obtained by short-time Fourier transform of the segmented time series data, and uses the time-frequency diagram after random masking to train the preset model, which can fully refine the local semantic information of various parts of the high temperature aging test equipment components, overcome the constraints of data scarcity and shallow annotation on the training model, and the trained feature extraction model predicts the time-frequency characteristics more accurately and reliably. Furthermore, the device health status of the high temperature aging test equipment obtained based on the predicted time-frequency characteristics is more reliable. Since the device health status of the high temperature aging test equipment can be accurately analyzed, timely repairs can be performed when device failures are found, thereby avoiding damage to the device under test due to device failures during the aging test. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 is a flowchart showing the steps of a method for deep fusion of time-frequency characteristics of a secondary power supply of a high-temperature aging test equipment according to an embodiment of the present application; Figure 2 is a schematic diagram of a time-frequency diagram after random masking in an embodiment of the present application; Figure 3 Schematic diagram of a Transformer-based autoencoder model according to an embodiment of the present application; Figure 4 is a flowchart showing the steps of a method for deep fusion of time-frequency characteristics of a secondary power supply of a high-temperature aging test equipment according to an embodiment of the present application; Figure 5 is a flowchart showing the steps of a method for evaluating the health status of a high temperature aging test chamber according to an embodiment of the present application; Figure 6 It is a structural block diagram of a device for deep fusion of time-frequency characteristics of a secondary power supply of a high-temperature aging test equipment according to an embodiment of the present application. DETAILED DESCRIPTION

[0023] In order to make the technical problems, technical solutions and advantages to be solved by the present invention more clear, a detailed description will be given below with reference to the accompanying drawings and specific embodiments.

[0024] In combination with the accompanying drawings, the method for deep fusion of time-frequency characteristics of the secondary power supply of high-temperature aging test equipment provided in the embodiment of the present application is described in detail through specific embodiments and their application scenarios.

[0025] As attached Figure 1 As shown, the method for deep fusion of time-frequency characteristics of secondary power supply of high temperature aging test equipment in the embodiment of the present application includes the following steps: Step 101: Collect a fault data set of components of a high temperature aging test equipment.

[0026] The high-temperature aging test equipment provided in the present application can perform non-destructive aging testing on the test parts. The time-frequency domain feature processing method used in the embodiments of the present application can extract features of the status data, health data, etc. of the components of the high-temperature aging test equipment. The extracted features can be used for health assessment, fault prediction, etc. of the components of the high-temperature aging test equipment, thereby ensuring the safety of the test parts.

[0027] Among them, the fault data includes multiple key parameter timing data, and the key parameters are different for different devices. For example: when the device is a secondary switching power supply, the key parameters can be the power supply voltage, current signal, and the thermal resistance characteristics from the MOSFET shell temperature to the junction temperature synchronously monitored by the infrared thermal imager and the embedded temperature sensor. When the device is a test chamber, the key parameters can be temperature monitoring signals, etc. These monitoring signals can be used to evaluate the health status of the high-temperature aging test equipment devices, but in the actual implementation process, the amount of directly measured key parameter data is large and contains a lot of redundant information. Therefore, it is necessary to use the feature extraction model trained by the deep fusion method of the time-frequency characteristics of the secondary power supply of the high-temperature aging test equipment in the embodiment of the present application to perform preliminary feature extraction on the directly measured data to obtain a feature sequence with high information density.

[0028] The fault data set of the high-temperature aging test equipment device collected in the embodiment of the present application can be a set of fault data from the beginning of degradation to the cessation of operation (i.e., the end of the remaining life cycle) of the device during its entire life cycle.

[0029] When collecting fault data sets of high-temperature aging test equipment, it is necessary to simultaneously monitor the load change characteristics of the secondary switching power supply, including the output voltage and current transient response under dynamic load switching. For the load fluctuation when the device under test is connected, record the power supply ripple fluctuation rate caused by it, and the calculation formula is: Ripple volatility = / )×100% in, is the peak-to-peak output voltage, is the average value of the output voltage. The ripple waveform is captured by a high-precision oscilloscope, and its spectrum characteristics are analyzed to identify abnormal harmonic components. Devices may include but are not limited to: high temperature aging test equipment switching power supply, high temperature aging test equipment high temperature test chamber, high temperature aging test equipment high temperature and high humidity test chamber, etc. The high temperature aging test equipment switching power supply can be further subdivided into: primary switching power supply, secondary switching power supply and secondary switching power supply MOSFET switch tube. In addition, high temperature aging test equipment devices may also include: drive control detection board.

[0030] The specific components included in the high-temperature aging test equipment can be flexibly adjusted according to the type of the high-temperature aging test equipment. The specific type of the high-temperature aging test equipment is not limited in the embodiments of the present application. For example, the high-temperature aging test equipment can be a high-temperature, high-humidity, and high-temperature aging test equipment, a constant temperature and constant pressure high-temperature aging test equipment, or a high-temperature, high-temperature aging test equipment, etc.

[0031] Step 102: pre-process the fault data set to obtain a training data set and a test data set.

[0032] An optional method of preprocessing the fault data set to obtain a training data set and a test data set may include the following sub-steps: Sub-step 1: For each key parameter in the fault data set, perform sliding window cutting on the key parameter time series data to construct a sample data set; When constructing a sample data set by sliding window cutting, the data in the load mutation period needs to be adaptively segmented. For intervals where the load change rate exceeds the threshold (such as ±10% / ms), a smaller window width and step size are used to capture the detailed characteristics of the transient response. At the same time, dynamic baseline calibration is performed on the current and voltage output data to eliminate the baseline drift caused by the connection of the device under test.

[0033] Sub-step 2: normalize each data sample in the sample data set to obtain a normalized sample data set; Sub-step 3: Select a preset percentage of sample data from the normalized sample data set as the training data set, and use the remaining sample data as the test data set.

[0034] The training data set is used to train the preset model, and the test data set is used to verify the prediction accuracy of the feature extraction model after training. The preset percentage can be flexibly set by those skilled in the art, and is not specifically limited in the embodiments of the present application, for example: set to 60%-80%, preferably 70%.

[0035] Step 103: Perform short-time Fourier transform on the time series data of each key parameter in the training data set to obtain a time-frequency diagram.

[0036] Traditional time domain and frequency domain analysis methods are mostly for stationary signals, and can only obtain information in one aspect of the time domain or frequency domain. In actual devices such as secondary switching power supply applications, the collected signals are mostly non-stationary, and analysis in one aspect can no longer meet the needs, so it is necessary to determine the relationship between the signal frequency and time. The joint time-frequency analysis method is a very effective tool in the current process of processing non-stationary signals. Therefore, short-time Fourier transform is used in this application to process variable time series and non-stationary signals. The basic principle of STFT (i.e. short-time Fourier transform) is: use a finite time window function h(t) to intercept the vibration signal, perform Fourier transform on the obtained signal, obtain a small range of local spectrum in this time period, and through the movement of the window function h(t) on the time axis, gradually analyze the signal band to obtain a set of local "spectra" of the signal. STFT is essentially a transformation of the basis function.

[0037] An optional method of performing short-time Fourier transform on each time series data in the training data set to obtain a time-frequency diagram may include the following sub-steps: Sub-step 1: for each time series data in the training data set, a window function with a finite time length is used to intercept the vibration signal in the time series data; Sub-step 2: Perform Fourier transform on the intercepted vibration signal to obtain a local spectrum in a small range of the corresponding time period; The conversion formula of the signal collected by high temperature aging test equipment components such as secondary power supply can be:

[0038] Where: is the source signal, is the analysis window function; is the time spectrum at time t.

[0039] Sub-step 3: By moving the window function on the time axis, each vibration signal band is gradually analyzed to obtain a set of local spectra of the vibration signal and generate a time-frequency diagram.

[0040] Preferably, the window function is a Hanning window function, which has good smoothness and is suitable for reducing spectrum leakage. The expression of the Hanning window function is as follows:

[0041] Step 104: Randomly mask each time-frequency graph and input it into a preset model for training to generate a feature extraction model.

[0042] Through steps 101 to 104, time-frequency diagrams corresponding to different time segments of multiple key parameters can be obtained. For each time-frequency diagram, a random masking method can be used to mask the time-frequency diagram. The schematic diagram of the time-frequency diagram after random masking is shown in FIG. Figure 2 As shown. The time-frequency diagram is different from the time series characteristics and frequency domain characteristics. The time series characteristics and frequency domain characteristics are one-dimensional data, while the time-frequency diagram is a two-dimensional picture. The horizontal axis represents time, the vertical axis represents frequency, and the depth of color represents the difference in frequency amplitude.

[0043] The randomly masked time-frequency graphs are sent to preset models such as the Transformer autoencoder for training, and a feature extraction model is generated after the training is completed. Figure 3 It is a schematic diagram of a Transformer-based autoencoding model according to an embodiment of the present application.

[0044] Transformer is a model built on the Seq-to-Seq framework. Compared with classic deep learning models, the most prominent advantage of Transformer is the use of a multi-head attention mechanism. The purpose of the multi-head attention layer is to give different importance to words / tokens in a sequence from multiple aspects. The Transformer model is mainly composed of input, encoder, decoder and output.

[0045] Position encoding: The position encoding layer is used to determine the position information of the sequence. Since there is no recursive layer and convolution layer in RNN and CNN, the input order information cannot be obtained by relying solely on the self-attention mechanism. The order information of the sequence needs to be actively passed to the model. The Transformer model uses a combination of sine and cosine functions to positionally encode the sequence. The calculation method is as follows:

[0046]

[0047] Where: is the position of the current sequence; for dimension; is the dimension of the input features.

[0048] Multi-head attention mechanism: The multi-head attention mechanism uses multiple attention mechanisms to perform operations in parallel, and then splices the operation results together through linear transformation. The core technology in Transformer is the multi-head attention mechanism, which uses the multi-head attention mechanism to extract the dependency features between data, capture the correlation between data and establish a context prediction model. The calculation method is as follows:

[0049]

[0050]

[0051]

[0052]

[0053]

[0054] Where: is the query matrix; is the key matrix; is the value matrix; , , is a trainable parameter matrix; is the processed input; is the dimension of the key matrix; , , , is the learnable parameter matrix.

[0055] Feedforward network, summation and normalization: In the Transformer model, the encoding part and the decoding part also contain feedforward network, summation and normalization. The calculation formula of the feedforward neural network is as follows:

[0056] Where: For input; , , , are parameters that can be obtained through training.

[0057] The calculation formula for summation and normalization is as follows:

[0058] Where: For input; This is the result after module processing.

[0059] In an optional embodiment, each time-frequency graph is randomly masked and input into a preset model for training, and a method of generating a feature extraction model may include the following sub-steps: Sub-step 1: Input the randomly masked time-frequency graph into the preset model, send it into the encoding module after position encoding, and send the output of the encoding module into the decoder, and restore the feature parameters in the high-dimensional hidden layer into the time-frequency graph before masking through the decoder; Sub-step 2: using the restored time-frequency graph to iteratively train the preset model multiple times to adjust the model parameters; Among them, the model parameters include: embedding dimension, number of attention heads, number of encoder layers, and number of decoder layers.

[0060] Sub-step 3: Take out the encoder layer and decoder layer of the model after multiple iterations of adjustment, and retain the weight parameters of each layer.

[0061] Sub-step 4: Generate a feature extraction model based on each encoder layer, each decoder layer and each layer weight parameter.

[0062] Step 105: Test the feature extraction model based on the test data set.

[0063] After training the preset model based on the training data set (also called pre-training) to generate a feature extraction model, the feature extraction model can be tested based on the test data in the test data set to determine the accuracy of the prediction results of the feature extraction model.

[0064] Based on the pre-trained Transforme autoencoder model, the training data set Perform self-attention feature extraction to obtain the self-attention feature set .

[0065] Step 106: After the test is passed, the key parameter timing data of the device to be predicted is input into the feature extraction model to predict the target time-frequency domain features.

[0066] The method for deep fusion of time-frequency features of secondary power supply of high-temperature aging test equipment provided in the embodiment of the present application collects the fault data set of high-temperature aging test equipment components; preprocesses the fault data set to obtain a training data set and a test data set; performs short-time Fourier transform on the time series data of each key parameter in the training data set to obtain a time-frequency diagram; randomly masks each time-frequency diagram and inputs it into a preset model for training to generate a feature extraction model; tests the feature extraction model according to the test data set; after the test passes, inputs the key parameter time series data of the device to be predicted into the feature extraction model to predict the target time-frequency domain features. The method provided in the embodiment of the present invention fully coordinates and refines a large amount of fault data of high-temperature aging test equipment components, and randomly masks the time-frequency diagram obtained by short-time Fourier transform of the segmented time series data, and uses the time-frequency diagram after random masking to train the preset model, which can fully refine the local semantic information of various parts of the high-temperature aging test equipment components, overcome the constraints of data scarcity and shallow annotation on the training model, and the trained feature extraction model predicts the time-frequency features more accurately and reliably. Furthermore, the health status of the components of the high-temperature aging test equipment obtained based on the predicted time-frequency characteristics analysis is more reliable.

[0067] Combine the following Figure 4 Taking a specific example, the deep fusion solution of time-frequency characteristics of the secondary power supply of high-temperature aging test equipment provided in this application is explained.

[0068] In this specific example, the processing of the time-frequency characteristics of the secondary switching power supply of the high-temperature aging test equipment is used as an example for explanation. The secondary switching power supply of the high-temperature aging test equipment will generate monitoring signals such as current and voltage during operation. In order to accurately extract the time-frequency domain characteristics of the secondary switching power supply, a deep fusion method of the time-frequency characteristics of the secondary power supply of the high-temperature aging test equipment is proposed in this specific example. This method fully coordinates and refines the local semantic information of the fault data of the high-temperature aging test equipment, overcomes the constraints of data scarcity and shallow annotation on the training network, and further improves the performance required by the task. It provides a more practical method for the extrapolation prediction problem of the degradation timing of the key parameters of the secondary switching power supply of the high-temperature aging test equipment (also known as the aging test bench).

[0069] like Figure 4 The feature extraction method shown includes several important steps: Step 1: Acquire the fault data of the secondary switching power supply of the aging test bench; Step 2: Perform comprehensive preprocessing on the fault data; Step 3: Perform short-time Fourier transform on the input time series data to obtain a time-frequency diagram; Step 4: Perform random masking on the time-frequency diagram; Step 5: Put the randomly masked time-frequency diagram into the Transformer autoencoder for model training; Step 6: Input the collected key parameter monitoring signal of the secondary switching power supply to be predicted into the Transformer autoencoder model for feature extraction, and output the extracted features.

[0070] Step 1: Obtaining fault data of the secondary switching power supply of the aging test bench like Figure 4 This link includes two steps: obtaining the key parameters of the secondary switching power supply of the aging test bench and obtaining the historical timing data of the parameters to be predicted (i.e. the key parameters to be predicted). In the actual timing process, a complete set of integrated circuit high-temperature dynamic aging detection system is used to monitor the voltage and current signals of the secondary switching power supply through sensors to obtain the original fault data of the secondary switching power supply of the aging test bench.

[0071] Step 2: Comprehensive preprocessing of fault data The collected raw fault data is sent to the fault data preprocessing unit for comprehensive processing to obtain training data set and test data set, such as Figure 4 This link specifically includes the following steps: Step 1: Data sliding window cutting generates the corresponding sample data set.

[0072] In this step, sliding window cutting is performed on the key parameter time series data to construct a sample data set.

[0073] For example, the key parameter timing data of any secondary switching power supply collected by the sensor is X, , perform sliding window cutting on X to generate the corresponding sample data set. When the window width is When the step size is S, the number of samples generated by cutting is:

[0074] The corresponding data set is generated as .

[0075] Step 2: Perform maximum and minimum value normalization on the training data set; In order to improve the data expression ability and speed up the convergence of subsequent model training, it is necessary to normalize the training data set, mainly by scaling the amplitude of the original parameters through the maximum and minimum normalization method to complete the linear transformation of the data. , through the formula:

[0076] Implement normalization processing to obtain a normalized sample data set .

[0077] Step 3: Construct training data set and test data set; Select the previous one from all the data The data is used as the training data set, and the remaining data is used as the test data set to verify the model prediction performance. Generally speaking, r is usually 60-80, and here it is 70.

[0078] Step 3: Perform short-time Fourier transform on the input time series data (i.e., training data set) to obtain a time-frequency diagram.

[0079] Traditional time domain and frequency domain analysis methods are mostly for stationary signals, and can only obtain information in one aspect of the time domain or frequency domain. In actual secondary power supply applications, the collected signals are mostly non-stationary, and analysis in one aspect can no longer meet the needs. We need to know the relationship between the signal frequency and time. The joint time-frequency analysis method is currently a very effective tool in the process of processing non-stationary signals. Therefore, this paper uses short-time Fourier transform to process variable time series and non-stationary signals. The basic principle of STFT, or short-time Fourier transform, is to use a finite time window function h(t) to intercept the vibration signal, perform Fourier transform on the obtained signal, obtain the local spectrum of a small range in the time period, and gradually analyze the signal band through the movement of the window function h(t) on the time axis to obtain a set of local "spectra" of the signal. STFT is essentially a transformation of the basis function.

[0080] The transformation formula of the secondary switching power supply acquisition signal can be:

[0081] Where: is the source signal; is the analysis window function; is the time spectrum at time t Preferably, the window function selects the Hanning window function, which has good smoothness and is suitable for reducing spectrum leakage. The expression of the Hanning window function is as follows:

[0082] Step 4: Randomly mask the time-frequency graph.

[0083] The schematic diagram of the masked time-frequency diagram generated after the random masking operation on the time-frequency diagram is as follows Figure 2 shown.

[0084] Step 5: Put the randomly masked time-frequency graph into the Transformer autoencoder for model training (i.e. training the Transformer autoencoder).

[0085] The randomly masked time-frequency graphs are sent to the Transformer autoencoder (i.e., the preset model described above) for training. The Transformer-based autoencoder model is as follows: Figure 3 shown.

[0086] Transformer is a model built on the Seq-to-Seq framework. Compared with classic deep learning models, the most prominent advantage of Transformer is the use of a multi-head attention mechanism. The purpose of the multi-head attention layer is to give different importance to words / tokens in a sequence from multiple aspects. The Transformer model is mainly composed of input, encoder, decoder and output.

[0087] Position encoding: The position encoding layer is used to determine the position information of the sequence. Since there is no recursive layer and convolution layer in RNN and CNN, the input order information cannot be obtained by relying solely on the self-attention mechanism. The order information of the sequence needs to be actively passed to the model. The Transformer model uses a combination of sine and cosine functions to positionally encode the sequence. The calculation method is as follows:

[0088]

[0089] Where: is the position of the current sequence; for dimension; is the dimension of the input features.

[0090] Multi-head attention mechanism: The multi-head attention mechanism uses multiple attention mechanisms to perform operations in parallel, and then splices the operation results together through linear transformation. The core technology in Transformer is the multi-head attention mechanism, which uses the multi-head attention mechanism to extract the dependency features between data, capture the correlation between data and establish a context prediction model. The calculation method is as follows:

[0091]

[0092]

[0093]

[0094]

[0095]

[0096] Where: is the query matrix; is the key matrix; is the value matrix; , , is a trainable parameter matrix; is the processed input; is the dimension of the key matrix; , , , is the learnable parameter matrix.

[0097] Feedforward network, summation and normalization: In the Transformer model, the encoding part and the decoding part also contain feedforward network, summation and normalization. The calculation formula of the feedforward neural network is as follows:

[0098] Where: For input; , , , are parameters that can be obtained through training.

[0099] The calculation formula for summation and normalization is as follows:

[0100] Where: For input; This is the result after module processing.

[0101] The model processing flow is as follows: the masked data is sent to the encoding part after position encoding, and the output of the encoding part is sent to the decoder to restore the feature parameters in the high-dimensional hidden layer to the original unmasked data, thereby training the model's feature extraction ability.

[0102] Secondly, select the appropriate number of iterations and loss function, input the constructed data set into the feature extraction model to repeatedly perform forward propagation and back propagation iterative calculation processes; in this process, the model parameters of the embedding dimension, number of attention heads, encoder and decoder layers are continuously adjusted to complete the pre-training of the model.

[0103] Again, take out the encoding layer and decoding layer of the pre-trained model, retain their weight parameters, and construct them into the trained Transformer autoencoder model.

[0104] Finally, based on the pre-trained Transformer autoencoder model, the training dataset Perform self-attention feature extraction to obtain the self-attention feature set .

[0105] Step 6: Input the collected key parameter monitoring signal of the secondary switching power supply to be predicted into the Transformer autoencoder model for feature extraction and output the extracted features.

[0106] After completing the prediction performance test of the trained Transformer autoencoder model (i.e., the feature extraction model described above) based on the training data set, if the test passes, it can be used to extract the health status characteristics of the secondary switching power supply of any aging test bench at any time to obtain a multi-dimensional feature sequence for subsequent evaluation of the health status of the secondary switching power supply.

[0107] The deep fusion method of time-frequency characteristics of the secondary power supply of high-temperature aging test equipment provided in this application can be used to extract features from key parameter monitoring signals of high-temperature aging test equipment components to obtain a multi-dimensional feature sequence. The obtained multi-dimensional feature sequence can be used to train a preset network model, and the trained network model can be used to evaluate the health status of high-temperature aging test equipment components. Of course, when evaluating the health status of a certain high-temperature aging test equipment component at any time, the multi-dimensional feature sequence input into the trained network model can also be extracted.

[0108] Taking the high-temperature aging test chamber as an example, the high-temperature aging test chamber is used to extract the multi-dimensional feature sequence of the high-temperature aging test chamber by using the deep fusion method of the time-frequency characteristics of the secondary power supply of the high-temperature aging test chamber. The extracted multi-dimensional feature sequence is input into the pre-trained time convolutional network model to obtain the degradation characteristic state of the high-temperature aging test chamber. The health status of the high-temperature aging test chamber is evaluated by the obtained degradation characteristic state.

[0109] As attached Figure 5 As shown, the high temperature aging test chamber health status assessment method of the embodiment of the present application includes the following steps: Step 501: Collect monitoring signals of the high-temperature aging test chamber to be evaluated.

[0110] Among them, the monitoring signals include: nine-point temperature measurement signals of the high-temperature aging test chamber, compressor vibration signals and fan speed signals.

[0111] The high-temperature aging test chamber health status assessment method provided by the present application can be applied to electronic devices. A high-temperature aging test chamber health status assessment computer program is provided in the electronic device. When the computer program is executed by the processor, the high-temperature aging test chamber health status assessment method of the present application embodiment is implemented.

[0112] The high-temperature aging test equipment provided in the present application can perform non-destructive aging testing on the test pieces, and the high-temperature aging test chamber health status assessment method provided in the embodiments of the present application can detect the health assessment status of the high-temperature aging test chamber, thereby ensuring the safety of the test pieces when the high-temperature aging test equipment is testing the test pieces.

[0113] In the actual implementation process, the temperature monitoring signal of the high-temperature aging test chamber directly measured has a large amount of data and contains a lot of redundant information. Therefore, preliminary feature extraction is performed on the directly measured data to obtain a feature sequence with high information density.

[0114] Step 502: extract features from the monitoring signal to obtain a multi-dimensional feature sequence.

[0115] The extracted features include at least two of the following: temperature fluctuation, temperature control accuracy, temperature uniformity, temperature overshoot, heating and cooling rate, vibration signal time, frequency domain features, and fan speed time domain features, etc. The specific features extracted can be flexibly set by those skilled in the art according to actual needs, ensuring that the extracted features can characterize the state of the test box.

[0116] Step 503: Input the multi-dimensional feature sequence into the pre-trained temporal convolutional network model to obtain the degradation feature state of the high temperature and high aging test chamber.

[0117] The degradation characteristic state includes: healthy state and degraded state. If the degradation characteristic state of the high temperature aging test chamber is healthy state, there is no need to execute step 504 and subsequent steps. If the degradation characteristic state of the high temperature aging test chamber is degraded state, it means that the high temperature aging test chamber has been degraded. The degradation progress of the high temperature aging test chamber can be automatically identified by adaptively analyzing the multi-dimensional feature sequence data through the time convolution network model.

[0118] In an optional embodiment, the training process of the temporal convolutional network model may include the following sub-steps: Sub-step 1: Obtaining a sample of the timing degradation characteristics for training in a high-temperature aging test chamber; Each training time series degradation feature sample is marked with a degradation feature state. The high temperature aging test chamber training time series degradation feature sample can be obtained by monitoring the temperature signal and degradation feature state of a high temperature aging test chamber during its entire life cycle.

[0119] In the actual implementation process, the high temperature aging test chamber training time series degradation feature samples can be expressed as , the maximum number of iterations max_iter. Mark the degradation characteristic state corresponding to the high temperature aging test chamber for training , including healthy and degraded states.

[0120] Sub-step 2: input the training temporal degradation feature samples into the pre-established temporal convolutional network model to obtain the first probability space point after mapping; In the actual implementation process, For input, Train the TCN model (i.e., the temporal convolutional network model) for output; obtain the TCN model parameters obtained through iterative learning. When the number of iterations reaches the maximum number of iterations max_iter, stop iterative learning.

[0121] 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 imposed on this in the embodiments of the present application.

[0122] Sub-step 3: Determine the degradation starting point of the high temperature aging test chamber according to the two-dimensional first probability space point, and generate a health baseline of the high temperature aging test chamber.

[0123] In an optional embodiment, the training process of the temporal convolutional network model may also include the following process: Acquire a sample of the test timing degradation feature of a high-temperature aging test chamber; input the test timing degradation feature sample into the time convolutional network model in training to obtain a mapped second probability space point; and calculate the distance between the second probability space and a healthy baseline based on a preset multi-distance metric algorithm.

[0124] The high temperature aging test chamber test timing degradation characteristic sample can be expressed as: In the training and testing process of the temporal convolutional network model, and Input the TCN algorithm model to obtain the mapped probability space points and .

[0125] The degradation starting point is automatically identified based on the two-dimensional probability space results, and the distance between the healthy space and the degradation space of the training and testing high-temperature aging test chamber is calculated using a weighted fusion multi-distance metric algorithm. Traverse all failure modes and obtain the health assessment results corresponding to each failure mode. The final trained temporal convolutional network model outputs and Health assessment curves corresponding to each high temperature aging test chamber.

[0126] More specifically, based on the preset multi-distance measurement algorithm, the distance between the second probability space and the health baseline can be calculated as follows: the distance between the second probability space and the health baseline is calculated based on the Mahalanobis distance algorithm, the cosine similarity algorithm and the Manhattan distance algorithm respectively to obtain the first distance, the second distance and the third distance; the first distance, the second distance and the third distance are weightedly summed to obtain the corresponding health level.

[0127] In the actual implementation process, it is not limited to using the above three distance algorithms to calculate the distance. Any one of the above three distance algorithms can also be used, or any two of the above three distance algorithms can be used to obtain two distances and then perform weighted summation of the two distances.

[0128] In the process of training the temporal convolutional network model, the high temperature aging test chamber test time series degradation feature samples are introduced, and the prediction accuracy of the temporal convolutional network model can be tested through the prediction results. In the actual implementation process, when the prediction accuracy of the temporal convolutional network model reaches the accuracy threshold, it can be determined that the training of this temporal convolutional network model is completed. After the subsequent training is completed, the temporal convolutional network model can be used to predict the state degradation curve of the high temperature aging test chamber to be evaluated.

[0129] Step 504: when the high temperature aging test chamber is in a degraded state, a state degradation curve of the high temperature aging test chamber to be evaluated is determined based on the probability space points after multi-dimensional feature sequence mapping and the health baseline in the time convolution network model.

[0130] Among them, the state degradation curve can characterize the health state of the high temperature aging test chamber.

[0131] By analyzing the multi-dimensional characteristics of the high-temperature aging test chamber to be evaluated through the time convolution network model, the health of the test chamber at the current moment can be obtained, and the state degradation curve up to that moment (also called the health degradation curve) can be output, providing reliable data support for further remaining service life prediction research of the high-temperature aging test chamber.

[0132] The high-temperature aging test chamber health status assessment method disclosed in the embodiment of the present invention collects the temperature monitoring signal of the high-temperature aging test chamber to be evaluated; performs feature extraction on the temperature monitoring signal to obtain a multidimensional feature sequence; inputs the multidimensional feature sequence into a pre-trained time convolutional network model to obtain the degradation feature state of the high-temperature aging test chamber; when the high-temperature aging test chamber is in a degraded state, based on the probability space points after the multidimensional feature sequence is mapped and the health baseline in the time convolutional network model, the state degradation curve of the high-temperature aging test chamber to be evaluated is determined to determine the health state of the high-temperature aging test chamber. The high-temperature aging test chamber health status assessment method provided in the embodiment of the present invention collects the temperature monitoring signal of the test chamber and extracts the multidimensional feature sequence, and predicts the multidimensional feature sequence based on the trained time convolutional network model, and can determine the state degradation curve of the high-temperature aging test chamber to be evaluated, thereby accurately determining the health state of the high-temperature aging test chamber.

[0133] The following is a specific example to illustrate the high temperature aging test chamber health status assessment solution provided by this application.

[0134] Provided in this specific example is a method for assessing the health status of a high-temperature aging test chamber based on degradation point identification using a temporal convolutional network and fusion of multiple distance metrics. The core of the method is to monitor the temperature of the high-temperature aging test chamber during the aging test. The amount of data from the directly measured monitoring signal is large and contains a large amount of redundant information. First, a preliminary feature extraction is performed on the directly measured data, including features such as temperature fluctuation, temperature control accuracy, temperature uniformity, temperature overshoot, and heating and cooling rates, to obtain a feature sequence with high information density. The feature sequence data is adaptively analyzed using a temporal convolutional network to automatically identify the degradation starting point of the high-temperature aging test chamber. The distance metric algorithm is used to quantitatively calculate the health status of the feature sequence in the degradation stage to complete the health assessment of the high-temperature aging test chamber.

[0135] The process of the high temperature aging test chamber health status assessment method based on temporal convolutional network degradation point identification and multi-distance metric fusion can specifically include the following process: S1: High temperature aging test chamber temperature monitoring: During the aging test, the temperature signal of the high-temperature aging test chamber is monitored by sensors, and large-scale monitoring data is obtained for the status evaluation of the test chamber.

[0136] S2: Preliminary feature extraction.

[0137] Since the amount of directly measured monitoring signal data is large and contains a lot of redundant information, it is necessary to first extract features from the direct monitoring data, including temperature fluctuation, temperature control accuracy, temperature uniformity, temperature overshoot, heating and cooling rates, etc., to obtain a high-information-density feature sequence, namely the multi-dimensional degradation characteristics of the high-temperature aging test chamber.

[0138] S3: Input the multi-dimensional degradation characteristics of the high-temperature aging test chamber into the temporal convolutional network model to automatically identify the degradation starting point and obtain the health baseline and degradation characteristics.

[0139] The health assessment process is further realized based on the extracted feature sequence: First, the TCN (Temporal Convolutional Network) model is used to identify the degradation starting point of the temperature data and feature data of the high-temperature aging test chamber, eliminating the instability caused by artificially setting the health baseline to construct health indicators. The TCN model is based on causal convolution, strictly ensuring that the output of each time step depends only on the input before the time step, and then introduces dilated convolution on this basis. By inserting intervals between the convolution kernel elements, the receptive field is increased according to specific rules, so that it can effectively capture the long-term dependencies in the time series data. The TCN model reasonably stacks multiple modules containing causal convolution, dilated convolution, and residual connections, so that the time series data can extract rich and valuable feature information after layer-by-layer processing. Then, it is connected to the fully connected layer, which integrates the previously extracted features and maps them to a two-dimensional space to realize the identification of the degradation starting point of the high-temperature aging test chamber.

[0140] Specifically, during the training phase of the TCN model, training data with health status labels (i.e., training time series degradation feature samples and degradation feature state labels marked by each sample) are input. ,in is the feature sample, for The corresponding labels, 0 and 1, represent the healthy state and the degraded state respectively, and n is the number of feature samples obtained by sliding window resampling of the entire degradation curve. In this process, the TCN model continuously optimizes the mapping parameters of the input data to the low-dimensional feature space, strengthens the classification ability of the healthy state data and the degraded state data, and enables the TCN model to automatically segment the healthy baseline of the test data.

[0141] In the testing phase of the TCN model, the test data (i.e., the test chamber test timing degradation feature samples) are input. The TCN model can optimize the learned model parameters according to the training process, map the healthy baseline data and the degradation data into a two-dimensional probability space, and complete the identification of the degradation starting point of the test data. , avoiding the blindness of manually dividing the health baseline.

[0142] S4: A multi-distance metric fusion method is used to calculate the distance between the health baseline and the degradation characteristics, and the state degradation curve and the current health of the high-temperature aging test chamber are determined.

[0143] In order to further realize the health status assessment of the high-temperature aging test chamber at the current moment, the multi-distance metric fusion method is used to comprehensively consider the Mahalanobis distance, cosine similarity and Manhattan distance between sample data and perform weighted fusion to obtain the health status at the current moment. At the same time, the health degradation curve up to that moment is output, providing reliable data support for further remaining service life prediction research of the high-temperature aging test chamber.

[0144] Figure 6 A structural block diagram of a device for deep fusion of time-frequency characteristics of a secondary power supply of a high-temperature aging test equipment in an embodiment of the present application.

[0145] The device for deep fusion of time-frequency characteristics of secondary power supply of high-temperature aging test equipment provided in the embodiment of the present application includes the following functional modules: The acquisition module 601 is used to acquire the fault data set of the high temperature aging test equipment components; wherein the fault data includes a plurality of key parameter time series data; A preprocessing module 602 is used to preprocess the fault data set to obtain a training data set and a test data set; A transformation module 603 is used to perform a short-time Fourier transform on each key parameter time series data in the training data set to obtain a time-frequency diagram; The mask training module 604 is used to randomly mask each of the time-frequency graphs and input the mask into a preset model for training to generate a feature extraction model; A testing module 605 is used to test the feature extraction model according to the test data set; The prediction module 606 is used to input the key parameter timing data of the device to be predicted into the feature extraction model after the device passes the test, and predict the target time-frequency domain features.

[0146] Optionally, the preprocessing module includes: The first submodule is used to perform sliding window cutting on the key parameter time series data for each key parameter in the fault data set to construct a sample data set; The second submodule is used to perform normalization processing on each data sample in the sample data set to obtain a normalized sample data set; The third submodule is used to select a preset percentage of sample data from the normalized sample data set as a training data set, and the remaining sample data as a test data set; wherein the training data set is used to train the preset model, and the test data set is used to verify the prediction accuracy of the trained feature extraction model.

[0147] Optionally, the transformation module includes: A fourth submodule is used to intercept the vibration signal in each time series data in the training data set by using a window function of a finite time length; The fifth submodule is used to perform Fourier transform on the intercepted vibration signal to obtain a local spectrum in a small range of the corresponding time period; The sixth submodule is used to gradually analyze each vibration signal band to obtain a set of local frequency spectra of the vibration signal by moving the window function on the time axis, and generate a time-frequency diagram.

[0148] Optionally, the window function is a Hanning window function.

[0149] Optionally, the mask training module includes: The seventh submodule is used to input the randomly masked time-frequency graph into the preset model, send it to the encoding module after position encoding, and send the output of the encoding module to the decoder, and restore the feature parameters in the high-dimensional hidden layer to the time-frequency graph before masking through the decoder; An eighth submodule is used to iteratively train the preset model multiple times using the restored time-frequency graph to adjust the model parameters, wherein the model parameters include: embedding dimension, number of attention heads, number of encoder layers, and number of decoder layers; The ninth submodule is used to extract the encoder layer and decoder layer of the model after multiple iterations and adjust, and retain the weight parameters of each layer; The tenth submodule is used to generate a feature extraction model based on each of the encoder layers, each of the decoder layers and the weight parameters of each layer.

[0150] Optionally, the device includes: a switching power supply of a high temperature aging test equipment, a high temperature test chamber of a high temperature aging test equipment, and a high temperature and high humidity test chamber of a high temperature aging test equipment.

[0151] The embodiments of the present application provide Figure 6 The high temperature aging test equipment secondary power supply time-frequency characteristics deep fusion device shown can achieve Figure 1 To avoid repetition, the various processes implemented by the method embodiment are not described here.

[0152] The method and device for deep fusion of time-frequency features of secondary power supply of high-temperature aging test equipment provided in the embodiment of the present application fully coordinates and extracts a large amount of fault data of high-temperature aging test equipment components, and randomly masks the time-frequency graph obtained by short-time Fourier transform of the segmented time series data, and uses the randomly masked time-frequency graph to train the preset model, which can fully extract the local semantic information of various parts of the high-temperature aging test equipment components, overcome the constraints of data scarcity, shallow annotation, etc. on the training model, and the trained feature extraction model predicts the time-frequency features more accurately and reliably. Furthermore, the health status of the high-temperature aging test equipment components obtained based on the predicted time-frequency feature analysis is more reliable.

[0153] An embodiment of the present invention further provides an electronic device, including a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other via the communication bus.

[0154] Memory, used to store computer programs; The processor is used to implement the method for deeply integrating the time-frequency characteristics of the secondary power supply of the high-temperature aging test equipment and the method for evaluating the health status of the high-temperature aging test chamber shown in the above method embodiment when executing the program stored in the memory.

[0155] 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. The communication bus may be divided into an address bus, a data bus, a control bus, etc.

[0156] The communication interface is used for communication between the above terminal and other devices.

[0157] The memory may include a random access memory (RAM) or a non-volatile memory, such as at least one disk memory. Optionally, the memory may also be at least one storage device located away from the aforementioned processor.

[0158] In another embodiment provided by the present invention, a computer-readable storage medium is also provided, which stores instructions. When the computer-readable storage medium is run on an electronic device, the electronic device implements the method for deep fusion of time-frequency characteristics of secondary power supply of high-temperature aging test equipment or the method for health status assessment of high-temperature aging test chamber described in any of the above embodiments.

[0159] In another embodiment provided by the present invention, a computer program product containing instructions is also provided. When the computer program product is run on an electronic device, the electronic device implements the method for deep fusion of time-frequency characteristics of the secondary power supply of high-temperature aging test equipment or the method for health status assessment of the test chamber of high-temperature aging test equipment described in any of the above embodiments.

[0160] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or device including the element.

[0161] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for deep fusion of time-frequency characteristics of secondary power supply of high-temperature aging test equipment, characterized in that: The method comprises: Collecting a fault data set of a high temperature aging test equipment component; wherein the fault data includes a plurality of key parameter timing data; Preprocessing the fault data set to obtain a training data set and a test data set; Performing short-time Fourier transform on the time series data of each key parameter in the training data set to obtain a time-frequency diagram; Performing random masking on each of the time-frequency graphs and inputting the mask into a preset model for training to generate a feature extraction model; Testing the feature extraction model according to the test data set; After the test is passed, the key parameter timing data of the device to be predicted is input into the feature extraction model to predict the target time-frequency domain features.

2. The method according to claim 1, characterized in that The step of preprocessing the fault data set to obtain a training data set and a test data set includes: For each key parameter in the fault data set, the key parameter time series data is subjected to sliding window cutting to construct a sample data set; Normalizing each data sample in the sample data set to obtain a normalized sample data set; A preset percentage of sample data is selected from the normalized sample data set as a training data set, and the remaining sample data is selected as a test data set; wherein the training data set is used to train the preset model, and the test data set is used to verify the prediction accuracy of the trained feature extraction model.

3. The method according to claim 1, characterized in that The step of performing short-time Fourier transform on each time series data in the training data set to obtain a time-frequency diagram includes: For each time series data in the training data set, a window function of finite time length is used to intercept the vibration signal in the time series data; Perform Fourier transform on the intercepted vibration signal to obtain a local spectrum in a small range of the corresponding time period; By moving the window function on the time axis, each vibration signal band is gradually analyzed to obtain a set of local frequency spectra of the vibration signal, and a time-frequency diagram is generated.

4. The method according to claim 3, characterized in that The window function is a Hanning window function.

5. The method according to claim 1, characterized in that: The step of randomly masking each of the time-frequency graphs and inputting them into a preset model for training to generate a feature extraction model includes: The time-frequency graph after random masking is input into the preset model, and is sent into the encoding module after position encoding, and the output of the encoding module is sent into the decoder, and the feature parameters in the high-dimensional hidden layer are restored to the time-frequency graph before masking through the decoder; The preset model is trained multiple times iteratively using the restored time-frequency graph to adjust the model parameters, wherein the model parameters include: embedding dimension, number of attention heads, number of encoder layers, and number of decoder layers; Take out the encoder layer and decoder layer of the model after multiple iterations and keep the weight parameters of each layer; A feature extraction model is generated based on each of the encoder layers, each of the decoder layers, and weight parameters of each layer.

6. The method according to claim 1, characterized in that The devices include: a switching power supply of a high temperature aging test equipment, a high temperature test box of a high temperature aging test equipment, and a high temperature and high humidity test box of a high temperature aging test equipment.

7. A device for deep fusion of time-frequency characteristics of secondary power supply of high-temperature aging test equipment, characterized in that: The device comprises: A collection module, used to collect fault data sets of high temperature aging test equipment components; wherein the fault data includes multiple key parameter time series data; A preprocessing module, used to preprocess the fault data set to obtain a training data set and a test data set; A transformation module, used for performing short-time Fourier transformation on the time series data of each key parameter in the training data set to obtain a time-frequency diagram; A mask training module, used to randomly mask each of the time-frequency graphs and input the mask into a preset model for training to generate a feature extraction model; A testing module, used for testing the feature extraction model according to the test data set; The prediction module is used to input the key parameter timing data of the device to be predicted into the feature extraction model after the test is passed, and predict the target time-frequency domain features.

8. The device according to claim 7, characterized in that The preprocessing module comprises: The first submodule is used to perform sliding window cutting on the key parameter time series data for each key parameter in the fault data set to construct a sample data set; The second submodule is used to perform normalization processing on each data sample in the sample data set to obtain a normalized sample data set; The third submodule is used to select a preset percentage of sample data from the normalized sample data set as a training data set, and the remaining sample data as a test data set; wherein the training data set is used to train the preset model, and the test data set is used to verify the prediction accuracy of the trained feature extraction model.

9. The device according to claim 7, characterized in that The transformation module comprises: A fourth submodule is used to intercept the vibration signal in each time series data in the training data set by using a window function of a finite time length; The fifth submodule is used to perform Fourier transform on the intercepted vibration signal to obtain a local spectrum in a small range of the corresponding time period; The sixth submodule is used to gradually analyze each vibration signal band to obtain a set of local frequency spectra of the vibration signal by moving the window function on the time axis, and generate a time-frequency diagram.

10. An electronic device, characterized in that: It includes a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other through the communication bus; Memory, used to store computer programs; The processor is used to implement the method for deep fusion of time-frequency characteristics of the secondary power supply of high-temperature aging test equipment as described in any one of claims 1 to 6 when executing the program stored in the memory.

Images