A small sample cross-domain classification prediction method for electromagnetic compatibility defects in battery simulators

Abstract

The present invention discloses a cross-domain classification prediction method for small samples of electromagnetic compatibility defects in battery simulators. The method collects high-frequency data of voltage, current and temperature of battery stacks, battery families and single cells from multiple sensors, performs fitting interpolation method to fill in missing data and continuous step judgment method to eliminate anomalies to form a training data set; extracts amplitude modulation morphology, envelope spectrum, frequency modulation information and multivariate secondary features of deep self-learning encoding of multi-sensor data; designs a meta-learning and knowledge transfer small sample model training method, establishes an asymmetric adversarial multi-source domain adaptive deep network model with mixed input of graphics and time series, and extracts high-dimensional fusion features of defects; uses KNN to perform similarity matching on high-dimensional fusion features, and gives a classification prediction of electromagnetic compatibility defects of the current circuit. The present invention classifies and predicts the anti-interference defects of battery simulators subjected to group pulse, static electricity, surge and power frequency radiation electromagnetic compatibility experiments, and discovers sub-healthy equipment.

Description

Technical Field

[0001] The present invention belongs to the field of big data fault diagnosis, and specifically relates to a method for classifying and predicting anti-interference defects in an electromagnetic compatibility experiment of a battery simulator. Background Art

[0002] The battery simulator is used for simulation testing of BMS functions and performance indicators, among which the stability and accuracy of the battery simulator are the main factors affecting the test performance. The battery simulator needs to be able to resist pulse groups, surges, static electricity and power frequency electromagnetic field radiation interference. For circuits with unstable performance, anti-interference defects will appear for different electromagnetic compatibility experiments. Some circuits will have anti-static defect characteristics, and some circuits will have anti-group pulse defect characteristics. These characteristics will be reflected in the high-frequency data of battery voltage and current. In order to enable the battery simulator to classify and predict the anti-interference defects of various magnetic compatibility experiments, the present invention proposes a small sample domain adaptive deep network model based on multi-sensor high-frequency sampling to predict the anti-interference defects of the circuit's electromagnetic compatibility experiment, screen out battery simulators with anti-interference defects from a large number of products, and ensure the reliability of the BMS system test performance. Summary of the Invention

[0003] The purpose of the present invention is to provide a classification, prediction and diagnosis method for anti-interference defects in an electromagnetic compatibility experiment of a battery simulator, so as to screen out sub-healthy devices with anti-interference defects from a large number of battery simulators.

[0004] To solve the above technical problems, the present invention provides a method for classifying and predicting anti-interference defects of a battery simulator under an electromagnetic compatibility test, comprising:

[0005] Collect multi-source high-frequency data on voltage, current, and temperature of battery stacks, battery families, and single cells, use the fitting interpolation method to fill in missing values, and the continuous step judgment method to eliminate anomalies to form a multi-source training data set;

[0006] Extracting the envelope morphology amplitude modulation information of the multi-source high-frequency training data set as the first feature, calculating the time series signal spectrum value of the envelope morphology to form an envelope spectrum sequence as the second feature, using the time-frequency analysis method to extract the frequency time variation information of the multi-source training data set as the third feature, and using the time series neural network to extract the deep network self-learning encoding of the original high-frequency waveform data as the fourth feature;

[0007] Based on four kinds of mixed feature information, a small sample knowledge transfer model training method is designed, and a rare small sample training model is based on the electromagnetic compatibility defect of the battery simulator;

[0008] Establish an asymmetric adversarial multi-source domain adaptive deep network model architecture for mixed input of graphics and time series multi-features to extract domain-invariant high-dimensional fusion features of defects;

[0009] The KNN nearest neighbor algorithm is used to perform similarity matching on the domain-invariant high-dimensional fusion features, and the anti-interference defect classification prediction of the current simulator circuit for four electromagnetic compatibility experiments: group pulse, static electricity, surge and power frequency radiation is given.

[0010] Optionally, the collecting and preprocessing of multi-source high-frequency data to form a multi-source data set for model training includes:

[0011] In order to extract the anti-interference defect characteristics of the battery simulator hardware circuit for various electromagnetic compatibility experiments and realize predictive defect classification, the present invention collects high-frequency data of current, voltage and temperature of battery families, battery stacks and single cells, and uses the fitting interpolation method and continuous step judgment method to fill in the missing data and eliminate anomalies to form a multi-source training data set.

[0012] Optionally, extracting multivariate secondary graph-time series mixed information from a multi-source training data set includes:

[0013] A Bayesian multi-optimization continuous piecewise fitting method is used to fit the waveform envelope to obtain the first feature of the envelope morphology of the original waveform. 2) The time series signal of the envelope morphology is collected and the FFT spectrum transform is performed to obtain the second feature of the envelope spectrum. 3) The wavelet packet transform time-frequency analysis method is used to extract the third feature of the frequency-time variation of the multi-sensor training dataset. 4) The LSTM time series network is used to extract the fourth feature of the deep model self-learning encoding of the original high-frequency data.

[0014] Optionally, the small sample knowledge transfer model training method includes:

[0015] 1) A generative adversarial network is used to generate new samples to expand the sample size and improve model training accuracy. 2) A meta-learning strategy is used to establish a multi-type, small-sample, and easily accessible circuit fault sample set, and the model is trained to acquire knowledge transfer capabilities. 3) For rare, anti-interference defects, semantics are used to define fault feature patterns. The spatial mapping relationship between semantics and features is then trained, and rare sample defects are identified through feature combination matching.

[0016] Optionally, the step of establishing an asymmetric adversarial multi-source domain adaptive deep network model architecture includes:

[0017] Taking the graphic-timing hybrid features as input, in order to achieve the defect recognition generalization capability of the small sample model across simulator circuits, the model needs to have the ability to extract domain-invariant defect features from working condition differences. The traditional domain adaptation method will change the trained source domain feature distribution through the two-way convergence strategy of the source domain and the target domain. The present invention adopts a domain adaptation method based on asymmetric adversarial, which improves the domain adversarial capability of the model by unidirectional movement from the target domain to the source domain, and at the same time adopts a multi-source domain adaptive model architecture to improve the cross-domain diagnostic capability of the model. The model outputs high-dimensional fusion features of different defects.

[0018] Optionally, the KNN nearest neighbor algorithm includes:

[0019] Based on the above method, high-dimensional fusion features of different electromagnetic compatibility experiments of various anti-interference defect circuits are extracted, and a supervised fusion feature set is established. At the same time, various electromagnetic compatibility experimental data of the target circuit are collected. The high-dimensional fusion features of the target domain are extracted through the above method, and the KNN algorithm is used to match the supervised features that are most similar to the target domain features, thereby realizing the prediction of the anti-interference defect type of the target circuit.

[0020] The present invention discloses a method for classifying and predicting electromagnetic compatibility defects in battery simulators based on a small sample domain adaptive deep network model. The method collects high-frequency data from multiple sensors for voltage, current, and temperature of battery stacks, battery families, and single cells, and uses a fitting interpolation method to fill in missing values ​​and a continuous step judgment method to eliminate anomalies to form a training data set; extracts the multivariate secondary features of the amplitude modulation morphology, envelope spectrum, frequency modulation information, and deep self-learning information of the multi-sensor data; designs a meta-learning and knowledge transfer small sample model training method, establishes a multi-source domain adaptive deep network model that counteracts the asymmetry of graphic and time series mixed inputs, and extracts high-dimensional fusion features of defects; uses KNN to perform similarity matching on the high-dimensional fusion features to provide a classification prediction of the electromagnetic compatibility defects of the current circuit. The present invention classifies and predicts the anti-interference defects of battery simulators when group pulses, static electricity, and electromagnetic radiation are applied, and screens out sub-healthy devices with anti-interference defects from a large number of newly produced battery simulators and predicts the defect type. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the embodiments of the present invention or the prior art solutions, the following briefly introduces the drawings required for use in the embodiments or the prior art descriptions. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0022] Figure 1 Flowchart of the cross-domain classification prediction method for a small sample of electromagnetic compatibility defects in a battery simulator provided by the present invention;

[0023] Figure 2 The meta-learning small sample model training method provided by the present invention;

[0024] Figure 3 The zero-sample semantic transfer model training method provided by the present invention;

[0025] Figure 4 The present invention provides an asymmetric adversarial multi-source domain adaptive model framework. DETAILED DESCRIPTION

[0026] In order to enable those skilled in the art to better understand the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the embodiments described are only some of the embodiments of the present invention, rather than all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0027] like Figure 1 As shown, Figure 1 The flow chart of the cross-domain classification prediction method for a small sample of electromagnetic compatibility defects in a battery simulator provided by the present invention includes five parts from S11 to S15.

[0028] S11: Collect high-frequency data of voltage, current, and temperature of battery stacks, battery families, and single cells from multiple sensors, and use the fitting interpolation method to fill in missing values ​​and the continuous step judgment method to eliminate anomalies to form a training data set.

[0029] A battery family is composed of multiple batteries connected in series, and a battery stack is composed of multiple battery families connected in series. For battery simulators with anti-interference defects, the voltage and current of single cells, battery families and battery stacks will be affected during electromagnetic compatibility experiments. The high-frequency waveform data contains modulation information of the interference, and the spectrum contains time-frequency variation information of the interference. In addition, when arc discharge, instantaneous short circuit and current fluctuation occur in the circuit, the local temperature of the circuit will fluctuate, and the real-time temperature morphological changes contain interference information. In order to classify and predict the anti-electromagnetic compatibility defects of the battery simulator circuit, it is necessary to collect high-frequency raw data of the voltage, current and temperature of single cells, battery families and battery stacks. The present invention uses a voltage divider resistor to reduce the voltage and then uses a high-speed AD to collect high-frequency voltage signals, a wide-band Hall current transformer to collect high-frequency current signals, and a PT100 to collect real-time temperature signals.

[0030] It should be noted that the purpose of introducing the fitting interpolation method and the continuous step judgment method in the present invention is that data missing and anomalies are inevitable during digital high-speed sampling. If the collected high-frequency signal is directly used for feature extraction, the result will be affected.

[0031] The basic idea of ​​the fitting interpolation method is to establish a prediction model for the collection points and predict and interpolate the missing data based on the data patterns learned by the model. The present invention adopts a multivariate linear prediction model and optimizes the model parameters through a data-driven method to achieve the goal of interpolating the missing data. Specifically, the following steps are included:

[0032] Step 1: Establish the full-pole model shown in equation (1). is the order, is the model parameter to be optimized, and formula (1) is Solve the model parameters when they are known , That is, the data point to be sampled, and For the previously collected data points;

[0033] (1)

[0034] Step 2: To solve the model parameters, the least squares objective function shown in formula (2) is used, and the self-supervised training method is used to transform the previous The collection points are used as input, and the rolling prediction The value of the data point, using the data-driven method to obtain the optimal statistical parameters ;

[0035] (2)

[0036] Step 3: Use the model (1) with optimal parameters based on the previous The missing values ​​are predicted by each data point, and the predicted values ​​are inserted into the missing positions to complete data interpolation.

[0037] The general outlier removal method has difficulty distinguishing between step changes and abnormal jumps. The continuous step judgment method performs a continuity test on the data with jumps to determine whether the jump is a step change or an abnormal data point, thereby achieving the goal of eliminating abnormal data. The specific steps include the following:

[0038] Step 1: Slide the data in a single window of k data points and calculate the mean of the data in each sliding window. and variance , when the value of the new data point entering the window is greater than When , the data point is marked as a mutation data point ;

[0039] Step 2, judge arrive There are n consecutive data points in total. If more than n consecutive data points are mutation points and the difference between them does not exceed , then all jump points are marked as step data, otherwise they are abnormal data and are eliminated;

[0040] Step 3: Perform the interpolation and anomaly elimination preprocessing on the multi-source high-frequency data of the anti-interference defective circuits of various electromagnetic compatibility experiments to form a multi-source training data set.

[0041] Based on the above discussion, in an optional embodiment of the present invention, the acquisition frequency is 10KHz, the order p is selected as 8, the sliding window k is selected as 10, and the outlier threshold m is selected as 6, that is, the classic 6 In principle, the number n of continuous jump points is selected as 20 according to the characteristics of the data of the present invention.

[0042] S12: Extracting multivariate quadratic graph-time hybrid features from multi-source training datasets.

[0043] It should be noted that the high-frequency raw data contains circuit anti-interference defect feature information, and it is necessary to extract the corresponding features of different defects through various signal processing methods. The present invention extracts four features.

[0044] First, feature extraction: The envelope shape of the original waveform contains defect feature information. To extract the waveform envelope shape, the present invention proposes a continuous segmented fitting boundary method. The number of segments and breakpoint positions are obtained by Bayesian multi-optimization algorithm. Specifically, the method includes the following steps:

[0045] Step 1: The multi-segment linear function is composed of multiple discrete linear functions. If each segment is connected end to end, it can describe any complex waveform. The function formula is as shown in formula (3). 、 、 , is the horizontal coordinate of each breakpoint, where , 、 、 Indicates the vertical coordinate of each breakpoint. 、 、 is the slope of each line segment. The function of formula (3) has n breakpoints and n-1 line segments.

[0046] (3)

[0047] Step 2: The number of segments and the breakpoint position of the continuous piecewise linear function are unknown parameters. To obtain the optimal number of segments and breakpoint position, the present invention uses the Bayesian global optimization algorithm to find the optimal number of segments, as shown in formula (4):

[0048] (4)

[0049] in, is the optimal number of segments, is the number of segments to be optimized, yes The variable range of . Bayesian optimization continuously approaches the objective function by calculating the mean and variance of any point. After the optimal number of segments is determined, the particle swarm optimization algorithm is used to obtain the optimal breakpoint position;

[0050] In step 3, the number of segments and breakpoint positions are first obtained according to the method in step 2, and then the segment boundary lines are fitted based on the big data according to step 1. The above process is repeated with the minimum sum of squared fitting errors as the optimization goal, and finally the original waveform data envelope morphological information is obtained as the first feature.

[0051] The second feature extraction is to sample the envelope line as discrete time series data and perform FFT transformation on the time series data to obtain the envelope spectrum. The envelope spectrum contains the envelope line fluctuation frequency information. This frequency information is the modulation signal, which is the interference information that the battery simulator hardware circuit cannot eliminate. The distribution of the spectrum values ​​at each frequency point in the envelope spectrum can be used to diagnose the circuit's anti-interference defects. This envelope spectrum is the second feature.

[0052] The third feature extraction is to extract the electromagnetic interference information contained in the high-frequency component. The present invention selects wavelet packet transform to extract the spectrum time variation information. The specific steps are as follows:

[0053] Step 1, wavelet basis selection, wavelet packet transform is to decompose the original waveform with wavelet basis functions of different frequencies and displacements to obtain the time distribution of frequency energy. The present invention selects discrete wavelet transform of formula (5), where is the displacement information, which represents the position information of the wavelet basis on the time axis, that is, the time information of the frequency. The width of the wavelet basis represents the frequency change. The original signal is decomposed by using wavelet basis functions of different frequencies and displacements to obtain the decomposed signal. ;

[0054] (5)

[0055] Step 2, parameter optimization. There are three levels of parameter selection for wavelet packet transform: ① basis function type, ② basis function frequency parameter, ③ basis function displacement parameter. To obtain the optimal three types of parameters, the present invention adopts a three-layer loop optimization method, with the cosine similarity between the wavelet reconstructed data and the original waveform data as the optimization target, and a series of wavelet basis function selections as the first layer loop, frequency parameter The variation within the constraint range is used as the second layer loop, the displacement parameter The changes within the constraint range are used as the third loop. The first loop adopts the sequential selection method, while the second and third loops use the Newton optimization method. Finally, the wavelet packet decomposition signal is obtained, which can perfectly reconstruct the original waveform using cosine similarity as the evaluation index.

[0056] Step 3: The section with the largest spectral energy in all wavelet decomposition signals is cut into a time series. All the cut time series are spliced ​​into long sequence data based on the time position information of the maximum spectral energy. The long sequence data contains the frequency-time distribution information of the interference feature. The long sequence spliced ​​data is the third feature.

[0057] Fourth, feature extraction. The aforementioned signal processing methods have significant limitations when extracting features; some circuit interference information cannot be extracted using these methods. Considering the powerful self-learning capabilities of deep learning algorithms, we used the anti-interference defect type as the supervisory label and selected a time-series deep neural network. The raw waveform time series data was used as input, and the model parameters were optimized using gradient descent. The feature vectors entering the fully connected classifier were used as self-learning codes, similar to the information compression function of an encoder. The resulting self-learning codes for the anti-interference defects in the battery simulator hardware circuit were used as the fourth feature.

[0058] Based on the above discussion, in an optional embodiment of the present invention, the wavelet basis selects the Morlet function, and the deep self-learning encoding model selects the LSTM temporal neural network with residual connection. All of the above methods select 8000 data points per second as training data through a sliding window.

[0059] S13: Design a small-sample deep model training method based on meta-learning and knowledge transfer.

[0060] It should be noted that the circuit anti-interference defects in different electromagnetic compatibility experiments are rare small samples. In order to realize the classification prediction of defect types by the model, the present invention proposes a small sample model training method, which specifically includes the following steps:

[0061] Step 1: Generate new samples using a generative adversarial network to increase the number of samples. The principle of the generative adversarial network objective function is as shown in Equation (6). This function consists of a generator G and a discriminator D, where x represents a real sample and z represents a Gaussian random distribution. Fixed D optimizes the parameters of G to minimize the objective function value, improving the ability of the generator G to make its generated samples close to the real samples; fixed G optimizes the parameters of D to maximize the objective function value, improving the ability of the discriminator D to find the difference between the generated samples and the real samples. Model training is completed by the continuous iteration of the above two processes. The adversarial model can generate generated samples with the same distribution as the real samples.

[0062] (6)

[0063] Step 2: The number of samples expanded by the above sample generation method is limited. In order to learn the defect classification features from the limited samples, the present invention adopts the meta-learning transfer training method, such as Figure 2 The method includes the following specific contents: ① Establishing a "few samples, multiple types" training set of easily accessible fault samples of battery simulator circuits, and using a supervised training method to enable the classification training model to obtain the feature extraction and classification capabilities of various circuit faults; ② Establishing a rare sample dataset of battery simulator circuit anti-electromagnetic compatibility defects, and based on the knowledge transfer logic, sharing the classification training model with the feature extraction model to extract various defect features to form a defect feature support set; ③ Sending various electromagnetic compatibility test data of the verification circuit as query samples into the feature extraction model to obtain the query sample features; ④ Using the Wasserstein distance to match the defect type corresponding to the query sample from the feature support set;

[0064] Step 3: For zero-shot defects where the support set cannot be established, the present invention proposes a zero-shot learning strategy based on expert semantic mapping, such as Figure 3 The method includes the following specific contents: ① Based on expert experience knowledge, feature description semantics is established; ② For the features Tz1 to Tzn of the semantic description, the two-dimensional graphic features corresponding to the semantic description are extracted from other circuit defects or faults and spliced ​​into a grid diagram; ③ A semantic-graphic mapping twin network is established, the semantic description is mapped to a high-dimensional space, and then the graphic features Tz1 to Tzn in the grid are matched to form high-dimensional features F1 to Fn of the semantic description. If the image feature of the semantic description exists, the probability value of the corresponding position is high. At the same time, the graphic features in the grid are mapped to a high-dimensional space and matched with the semantic description to form high-dimensional features S1 to Sn of the graphic features. If the semantic description of the graphic feature exists, the probability value of the corresponding position is high. The model is trained with the similarity of the probability distribution corresponding to the two mapping spaces as the objective function, so that the model has the transfer learning ability to convert expert semantic descriptions into fault graphic features; ④ Various graphic features are spliced ​​as input to the twin network model, and zero-shot defect prediction is achieved based on the matching degree between the visual features recognized by the model and the zero-shot semantic description.

[0065] Based on the above discussion, in an optional embodiment of the present invention, the backbone of all models uses a convolutional neural network, the classifier uses a fully connected network, the convolution method of the YOLO model is selected to extract the characteristic information of the spliced ​​grid image, and the meta-learning training set selects five easily accessible types: healthy circuit, cold solder joint, poor insulation, poor contact and device damage.

[0066] S14: Build an asymmetric adversarial multi-source domain adaptive deep network model architecture.

[0067] It should be noted that the distribution of the same defect feature exhibited by the battery simulator under different working conditions is different. In order to enhance the cross-working condition diagnostic capability of the prediction model, the present invention proposes an asymmetric adversarial multi-source domain adaptive deep network model architecture, as shown in Figure 4. The model architecture mainly consists of a hybrid domain feature extractor, a single source domain feature extractor, a defect type boundary amplifier, a defect classifier, and a defect feature reconstructor. The model establishment process specifically includes the following steps:

[0068] Step 1: The hybrid domain feature extractor maps the raw data of each source domain and target domain to the encoding space using deep self-learning, and extracts hybrid domain feature vectors that have a certain correlation with the defect;

[0069] Step 2: The distribution of the same defect samples in each domain is different. To extract the features of the defect samples in each source domain and target domain, a single source domain feature extractor is established. The balanced data of the source domain and target domain of the defect samples are subjected to secondary feature mapping to obtain the single source domain features of the defect.

[0070] Step 3: Using single-source domain features as input, the present invention proposes three objective functions to achieve domain-invariant feature extraction of defects: ① Ternary center training loss In order to reduce the intra-class distance of the source domain defect features, increase the inter-class distance, and avoid the defect features from gathering at the classification boundary and causing defect misclassification, the present invention selects the ternary center loss of formula (7) as the model objective function, where For the sample to be identified, are samples from different domains of the same defect type, For samples of different defect types, the training goal is to extract domain-invariant defect features, so that the feature distance of samples of the same defect type decreases, while the distance between features of samples of different defect types increases;

[0071] (7)

[0072] ② To extract cross-domain features of defects, a defect type classifier is established for each source domain data. During training, each classifier will have a cross entropy loss classification result. For samples of the same defect type in different domains, if multiple classifiers have the same classification results, the cosine similarity alignment loss is used. The smaller the defect classification features extracted, the more cross-domain migration capabilities are. ③ Asymmetric adversarial methods: the source domain labeled samples are trained to obtain the defect feature distribution. Traditional domain adversarial methods extract domain-invariant features of defects, which will change the already trained source domain defect distribution with classification capabilities, resulting in a smaller or even confused source domain classification boundary. In order to achieve a unidirectional shift of the target domain sample distribution toward the source domain, this paper proposes an asymmetric adversarial method to reduce domain confusion loss. The asymmetric method is based on the reconstructor principle. During classification training, the source domain data optimizes the classifier and reconstructor parameters. While the model implements defect classification, it can reconstruct the source domain data but not the target domain data. During domain adversarial training, the reconstructor parameters are fixed, and only the feature extractor parameters are optimized for the target domain data. When the reconstructor can reconstruct the target domain sample data, the model extracts the core features of the defect that span the source and target domains.

[0073] Based on the above discussion, in an optional embodiment of the present invention, the training set selects three working conditions as different source domain data, corresponding to temperature, voltage and humidity respectively, and the target domain selects the normal working environment of the battery simulator.

[0074] KNN defect type matching prediction.

[0075] It should be noted that the above-mentioned small sample model training method and multi-source domain adaptive deep network model can only extract circuit anti-interference defect features. To achieve predictive classification of defects, the present invention proposes a feature matching method based on the KNN nearest neighbor algorithm, which specifically includes the following steps:

[0076] Step 1: Extract anti-interference defect samples of various electromagnetic compatibility experimental circuits, input the above-mentioned asymmetric anti-domain migration model to obtain the domain-invariant features of the defects and mark the defect types, and establish the defect feature set shown in formula (8), where Defective samples The domain-invariant eigenvector of for Defect category label;

[0077] (8)

[0078] Step 2: Use formula (9) to calculate the unknown type of defect samples and The Euclidean distance of each known defect sample in;

[0079] (9)

[0080] Step 3, find the distance Recent feature vectors of known types, forming The nearest neighbor combination of ,exist The type with the largest number of corresponding samples is the defect type recognition result of KNN. .

[0081] It should be noted that, in this document, relational terms such as "1" and "2" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any actual relationship or order between these entities or operations. The principles and implementation methods proposed in the present invention are explained using specific cases. The description of the above embodiments is only intended to help understand the method of the present invention and its core ideas. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made to the present invention without departing from the principles of the present invention, and these improvements and modifications also fall within the scope of protection of the claims of the present invention.

Claims

1. A cross-domain classification prediction method for small samples of electromagnetic compatibility defects in battery simulators, characterized by: include: Collect multi-source high-frequency data on voltage, current, and temperature of battery stacks, battery families, and single cells, perform fitting and interpolation to fill missing values, and perform continuous step judgment to eliminate anomalies to form a multi-source training dataset; Extract the amplitude modulation morphology, envelope spectrum, frequency modulation information and deep self-learning coding multivariate quadratic graph-time series hybrid features of multi-sensor high-frequency data: ① Extract the envelope morphology of the multi-source training data set to obtain the amplitude modulation information represented by the envelope line of the original waveform as the first feature, ② Sample the envelope line into discrete time series data, perform FFT transformation on the time series data to obtain the envelope spectrum sequence as the second feature, ③ Use time-frequency analysis to extract the frequency-time variation information of the multi-source training data set as the third feature, ④ Use a time series neural network to extract the deep network self-learning coding of the original high-frequency waveform data as the fourth feature; Based on four hybrid features and a rare small sample training dataset of EMC defects of battery simulators, the model is trained using a knowledge transfer method. Taking multivariate mixed features of graphics and time series as input, adopting an asymmetric adversarial objective function and based on a multi-source domain adaptive model architecture, we extract domain-invariant high-dimensional fusion features of defects. The KNN nearest neighbor algorithm is used to perform similarity matching on the domain-invariant high-dimensional fusion features, and a classification prediction of the anti-interference defects of the current simulator circuit for group pulse, static electricity, surge and power frequency radiation experiments is given to achieve the screening of sub-healthy equipment.

2. The cross-domain classification prediction method for small samples of electromagnetic compatibility defects of battery simulators as described in claim 1 is characterized in that: Collect high-frequency data from multiple sensors, use fitting interpolation to fill missing values, and use continuous step judgment method to remove outliers to form a multi-source training data set, including: To predict the anti-interference defects of the battery simulator for different electromagnetic compatibilities, high-frequency data on voltage, current, and temperature of the battery stack, battery family, and single cell are collected. The voltage is measured by a high-speed AD after being stepped down by a resistor divider, the current is measured by a wide-band Hall current transformer, and the temperature is measured by a PT100. In order to obtain clean and defect-free data, the fitting interpolation method is used to fill the missing values ​​of the original waveform. At the same time, the continuous step judgment method is used to judge whether the data is abnormal based on the multi-point continuous stability after the signal step jump. After preprocessing, a multi-sensor training data set is formed for the circuit anti-interference defect classification prediction model.

3. The cross-domain classification prediction method for small samples of electromagnetic compatibility defects of battery simulators as described in claim 1 is characterized in that: Extracting the amplitude modulation morphology, envelope spectrum, frequency modulation information and deep network self-learning encoding of multivariate quadratic graph-time series mixed information from multi-sensor training datasets, including: 1) In order to extract the envelope morphological features of the waveform modulated by the interference information, a Bayesian multiple optimization continuous piecewise fitting method is used to fit the waveform boundary to obtain the envelope morphology of the modulated waveform as the first feature; 2) The envelope time series signal is subjected to FFT spectrum transformation to obtain the envelope spectrum sequence as the second feature; 3) The frequency time variation information of the multi-sensor training data set is extracted as the third feature using the wavelet packet transform time-frequency analysis method; 4) The deep network self-learning encoding of the original high-frequency waveform data is extracted as the fourth feature using the LSTM time series neural network.

4. The cross-domain classification prediction method for small samples of electromagnetic compatibility defects of battery simulators as described in claim 1 is characterized in that: Design a small sample deep model training method based on meta-learning and knowledge transfer, including: The following small sample model training methods are proposed: 1) Adopting adversarial generative methods to generate new samples; 2) Using meta-learning methods to establish the knowledge transfer capability of training models for multi-type, small-sample, easily accessible circuit fault samples; 3) For extremely rare anti-interference defects, semantics are used to define fault feature patterns, and then the spatial mapping relationship between semantics and features is trained. Finally, rare sample defects are identified through feature combination matching.

5. The cross-domain classification prediction method for a small sample of electromagnetic compatibility defects in a battery simulator as claimed in claim 1, characterized in that: Based on feature-semantic hybrid input, adopting an asymmetric adversarial objective function, and based on a multi-source domain adaptive deep network model architecture, it includes: In order to extract weak defect features, multivariate quadratic mixed features are used as model input, LSTM timing network is used to receive timing features, and CNN two-dimensional convolutional network is used to receive graphic features. Considering the hardware discrete differences of battery simulation circuits, different simulators have certain characteristic differences in the anti-interference defects of a certain electromagnetic compatibility experiment. In order to obtain the domain-invariant features of the anti-interference defects corresponding to the electromagnetic compatibility experiment, an asymmetric adversarial objective function and a multi-source domain adaptive model structure are established for the unidirectional movement of target domain sample features to source domain features, and the domain-invariant high-dimensional fusion features of the battery simulator circuit anti-interference defects are extracted.

6. The cross-domain classification prediction method for small samples of electromagnetic compatibility defects of battery simulators as described in claim 1 is characterized in that: The KNN nearest neighbor algorithm is used to achieve similarity matching of domain-invariant high-dimensional fusion features and classify and predict circuit anti-interference defects, including: Electromagnetic compatibility experiments are carried out on existing anti-interference defect circuits. The supervisory feature sets of various defects are extracted through deep network models. For the battery simulator in the experimental environment, high-dimensional fusion features of the target domain are extracted. The KNN nearest neighbor algorithm is used to perform similarity matching on the target domain features. The anti-interference defect types of the current simulator circuit for four electromagnetic compatibility experiments, namely group pulse, static electricity, surge and power frequency radiation, are matched from the supervisory feature set.

Images