A PCB analog circuit component fault prediction method based on improved large language model

Abstract

The present invention belongs to the technical field of circuit fault prediction, and discloses a PCB analog circuit component fault prediction method based on an improved large language model. Excitation is applied to the PCB analog circuit to extract the output voltage of the components to be predicted in the PCB analog circuit at different frequencies; data preprocessing is performed according to the Pearson correlation coefficient method; the improved large language model uses a general large language model as a base model, adds a linear layer, a cross-attention layer and a fully connected layer, and obtains a feature vector after cross-modal alignment; the prompt word of the improved general large language model is improved, and the prompt word vectorization of the improved general large language model and the feature vector after cross-modal alignment are added and connected as the input feature vector of the large model. The method proposed by the present invention reduces the loss of features during feature extraction, so that the prediction result is more generalized and reliable.

Description

Technical Field

[0001] The present invention relates to the technical field of circuit fault prediction, and in particular to a PCB analog circuit component fault prediction method based on an improved large language model. Background Art

[0002] PCB analog circuits are essential basic components in modern electronic devices and are widely used in aerospace equipment, industrial manufacturing, and household appliances. As PCB analog circuits become more integrated in the development of circuit manufacturing processes, once a failure occurs, it will cause a large-scale circuit system to be paralyzed. Therefore, PCB analog circuit fault prediction has become a crucial solution to identify the fault tendency of components in PCB analog circuits before the failure occurs, increase the operational stability of the circuit system, and reduce later operating costs.

[0003] The current research methods are widely used in prediction methods represented by machine learning and deep learning, such as "Ni Xianglong, Shi Changan, Ma Yueliang, et al. Research on electronic equipment fault prediction method based on Bi-LSTM [J]. Aviation Weapon, 2022, 29(06): 102-110.", which uses the Bi-LSTM model to learn historical PCB simulation circuit data to predict the time when failures will occur in the future.

[0004] However, the current research on PCB analog circuit fault prediction at home and abroad is still in the basic theoretical research stage, and there are many problems to be solved. The current research model will have overfitting problems in practical applications, resulting in low accuracy when predicting new data; the model structure is complex, the amount of data cannot be too large during training, the training speed is low, and the gradient is prone to disappearance or explosion, resulting in training failure; when PCB analog circuit data is extracted according to the deep learning method, there is inevitably feature loss. The more features are lost, the lower the model prediction accuracy. Summary of the invention

[0005] To solve the above problems, the present invention proposes a PCB analog circuit component fault prediction method based on an improved large language model. The method cuts and vectorizes the input time series data into time series vector blocks, linearly combines the word vectors pre-trained by the general large language model by constructing a linear layer, and cross-modally aligns the time series vector blocks and the linearly combined word vectors according to the multi-head attention mechanism by constructing a cross-attention layer and a fully connected layer, improves the prompt words of the general large language model, and adds and connects the improved prompt word vectorization and the feature vectors after cross-modal alignment and inputs them into the general large language model. The present invention reduces the loss of features when extracting features from the large model.

[0006] The technical solution of the present invention is as follows: A method for predicting faults of PCB analog circuit components based on an improved large language model, comprising the following steps:

[0007] Apply excitation to the PCB simulation circuit and extract the output voltage of the components to be predicted in the PCB simulation circuit at different frequencies;

[0008] The above frequencies and output voltages are preprocessed according to the Pearson correlation coefficient method. The components are set to change evenly during the aging process. The frequency sweep excitation is performed once at each time point. The data is processed into univariate time series data corresponding to "component health-time" to obtain the preprocessed data set.

[0009] Constructing an improved large language model; the improved large language model uses the general large language model as a base model, adds a linear layer, a cross attention layer and a fully connected layer, and improves the prompt words of the general large language model;

[0010] The univariate time series data corresponding to "component health - time" is cut and segmented to obtain time series blocks, and the time series blocks are vectorized to obtain a set of time series vector blocks;

[0011] The linear layer linearly combines the word vectors pre-trained by the general large language model in the original pre-trained word vector space according to a linear transformation. The number of word vectors after the linear combination is less than the number of word vectors pre-trained by the general large language model. The linear layer is used to reduce the number of word vectors to reduce the amount of calculation for the subsequent cross attention layer;

[0012] The cross attention layer performs cross-modal alignment of two different sequences, the time series vector block and the linearly combined word vector, according to the multi-head attention mechanism, so as to selectively focus the word vectors after different linear combinations on the data fluctuation characteristics in the time series vector block;

[0013] The fully connected layer introduces a ReLU activation function, and performs nonlinear fitting on the feature vector output by the cross attention layer through the ReLU activation function to obtain a feature vector after cross-modal alignment;

[0014] The improved prompt words of the general large language model are used to guide the general large language model to understand and perform the PCB analog circuit component fault prediction task, and the feature vectors after the improved prompt word vectorization and cross-modal alignment are added and connected, and the feature vectors after the addition and connection are used as the input feature vector of the general large language model to guide the general large language model to understand the task process of fault prediction;

[0015] The improved large language model is trained according to the preprocessed data set, the number of training iterations is set, and the training is stopped when the number of training iterations is reached, so as to obtain the trained improved large language model;

[0016] The input time dimension is represented as the health of the component to be predicted in the interval [0, T], and the health of the component to be predicted after time T is predicted based on the trained improved large language model; a threshold for the health of the component to be predicted is set to indicate that a fault occurs when the health of the component to be predicted is lower than the set threshold. When the health of the component to be predicted at a certain moment after time T in the prediction result is equal to the preset threshold, it indicates that the component to be predicted will fail at that moment.

[0017] The block segmentation process and sequential block vectorization specifically include:

[0018] The preprocessed data with a time step of T is cut into time series blocks with a fixed length, and the segmentation step is set to L. The number of time series blocks is , we get a set of timing blocks, denoted as ;

[0019] Each timing block Vectorized processing is , we get a set of time series vector blocks, denoted as ,definition , where N represents Quantity, D represents Vector dimensions.

[0020] The linear layer processing process is as follows:

[0021] The word vector pre-trained by the general large language model is denoted as m, and the definition , where M represents the number of pre-trained word vectors and P represents the dimension of pre-trained word vectors;

[0022] The linear layer performs a linear transformation on the pre-trained word vector. The linear transformation is calculated as follows:

[0023]

[0024] in, Represents the word vector after linear combination, definition , Represents the number of word vectors after linear combination, represents the constant bias term, Represents the weight parameter matrix.

[0025] The cross-modal alignment specifically includes:

[0026] All the attention heads in the multi-head attention mechanism are denoted as , where H represents the number of attention heads;

[0027] The query matrix of i attention heads is defined as ; The bond matrix is ​​defined as ; The value matrix is ​​defined as ;in, , , the cross attention is calculated as follows:

[0028]

[0029] in, represents the cross attention value of the i-th attention head, and we get , aggregate the cross attention values ​​of H attention heads to get ,get , represents the weight parameter matrix of the multi-head attention mechanism;

[0030] According to the calculation of the cross attention layer, the feature vector output by the cross attention layer is nonlinearly fitted through the ReLU activation function of the fully connected layer to obtain the feature vector after cross-modal alignment, which is recorded as ,definition .

[0031] The improved prompt words of the general large language model are prompt words based on the general large language model, with task scenario description, task instructions and input data statistics added, and stored in the Prompt template of the general large language model in JSON format;

[0032] The task scenario description includes the types of components to be predicted in PCB analog circuit fault prediction, the threshold at which the components will fail, and the meaning of the characteristic index corresponding to each data point;

[0033] The task instruction includes the number of time series blocks obtained by the block segmentation process, the time step of the time series blocks, the sequence of the time series blocks in the time dimension and the predicted time step;

[0034] The input data statistics include discrete data points of each time series block and feature statistics of each time series block;

[0035] The characteristic statistics of each time series block include the maximum value, minimum value, median value, and slope value of the data as a whole of the time series block;

[0036] The feature vector after the addition connection is the prompt word of the improved general large language model, and the text is vectorized to obtain the prompt word vector, which is recorded as ,Will and Each eigenvector in is added and connected, and the process is expressed as , aggregate i added feature vectors, denoted as , as the input vector of the general large language model.

[0037] The data preprocessing specifically includes:

[0038] Apply the swept frequency excitation source as the input source to the PCB analog circuit, and evenly extract the output voltage of a specific frequency in the entire response frequency band as the fault characteristic indicator. , where U represents the set of output voltages of the same component at different frequencies, and v represents the number of specific frequencies;

[0039] According to the Pearson correlation coefficient method, the correlation between the output voltage and the specific frequency in the extracted fault characteristic indicators is calculated:

[0040]

[0041] Where x represents the characteristic vector when the output voltage of the component is equal to the nominal value, The characteristic vector representing the health of the component at level i, and Respectively and The arithmetic mean of represents x and The value of the closeness of the linear relationship between them; when the components have no loss and faults, the performance parameters are nominal values, and the fault characteristic indicators extracted at specific frequencies are ; When the component deviates from the nominal value due to loss, the fault characteristic index extracted at a specific frequency is the i-th case Reflect the changing trend of the component's "frequency-output voltage" curve under the performance parameters;

[0042] It is used to represent the health of components and obtain the univariate time series data corresponding to "component health-time". When the components have no loss or failure, the fault characteristic index is equal to the nominal value. =1; When the fault characteristic index of a component gradually deviates from the nominal value over time at a specific frequency due to loss, the greater the deviation, The lower it is, the lower the health of the component, and the more likely the component is to fail. When the health of a component is lower than a certain value, it is considered that the component will fail. According to the type of component, such as resistors, capacitors, etc., the corresponding threshold is set to indicate that the component will fail when its health is lower than the set threshold.

[0043] Compared with the prior art, the present invention has the following beneficial effects:

[0044] The present invention divides and segments the univariate time series data corresponding to "component health-time" to obtain time series blocks, vectorizes the time series blocks, and cross-modally aligns the time series block vectors with the word vectors of a general large language model based on a multi-head attention mechanism, so as to align the data trends of the time series data such as "increase" and "decrease" with the natural language, effectively enhancing the ability of the large language model to capture the features of time series data and the reasoning ability of the large language model for time series prediction tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 A flow chart of a method for predicting faults of PCB analog circuit components based on an improved large language model used in an embodiment of the present invention;

[0046] Figure 2 A diagram of the improved large language model process obtained for use in an embodiment of the present invention;

[0047] Figure 3 This is a PCB simulation circuit diagram used in an embodiment of the present invention. DETAILED DESCRIPTION

[0048] In order to make the technical solution, purpose and advantages of the present invention more clear, the present invention is further described in detail below through the accompanying drawings and embodiments. It should be pointed out that the provided embodiments are only used to explain the present invention, rather than to limit the scope of protection of the present invention.

[0049] The following is a detailed description of a specific implementation method of a PCB analog circuit component fault prediction method based on an improved large language model, which is intended to help understand the technical solution and operating steps of the present invention.

[0050] See also Figure 1 , the method comprises the following steps:

[0051] S110, applying excitation to the PCB simulation circuit, and extracting output voltages of components to be predicted in the PCB simulation circuit at different frequencies;

[0052] S120, performing data preprocessing according to the Pearson correlation coefficient method, processing the extracted data into univariate time series data corresponding to "component health-time", and obtaining a preprocessed data set;

[0053] S130, constructing an improved large language model, wherein the improved large language model uses the general large language model as a base model, adds a linear layer, a cross attention layer and a fully connected layer, and improves the prompt words of the general large language model;

[0054] The univariate time series data corresponding to "component health-time" is cut and segmented to obtain time series blocks, and vectorized into a group of time series vector blocks; the linear layer linearly combines the word vectors pre-trained by the general large language model in the pre-trained word vector space based on linear transformation, and the number of word vectors after linear combination is much smaller than the number of pre-trained word vectors; the cross-attention layer cross-modally aligns the time series vector blocks and the linearly combined word vectors according to the multi-head attention mechanism, and the fully connected layer introduces the ReLU activation function, and the feature vector output by the cross-attention layer is nonlinearly fitted through the ReLU activation function to obtain the feature vector after cross-modal alignment; the prompt word of the general large language model is improved, the improved prompt word vectorization and the feature vector after cross-modal alignment are added and connected, and the feature vector after addition and connection is used as the input vector of the general large language model to guide the general large language model to understand the task process of fault prediction.

[0055] S140, training an improved large language model according to the preprocessed data set;

[0056] S150, input the health status of the component to be predicted in the interval [0, T] represented by the time dimension, predict the health status of the component to be predicted after time T according to the trained improved large language model, set a threshold for the health status of the component to be predicted, which is used to indicate that a failure will occur when the health status of the component to be predicted is lower than the set threshold, and when the health status of the component to be predicted at a certain moment after time T in the prediction result is equal to the preset threshold, it indicates that the component to be predicted will fail at that moment.

[0057] In this embodiment, the detailed implementation steps of step S110 include:

[0058] See also Figure 3 , take the Sallen-Key bandpass filter circuit as the PCB simulation circuit to be extracted, randomly select two components in the circuit, assume that the capacitor C1 and the resistor R1 in the Sallen-Key bandpass filter circuit are selected as the components to be tested, and extract the output voltages of the two components at different frequencies. Set the excitation start frequency of the Sallen-Key bandpass filter circuit to 1Hz and the end frequency to 50kHz. The resistance value of resistor R1 at the nominal voltage is 2kΩ, and the capacitance value of capacitor C1 at the nominal voltage is 5nF. In order to evaluate the health of the simulation circuit, capacitor C1 and resistor R1, it is assumed that the parameter value of each test component changes uniformly during the degradation process, and each change corresponds to a time point. At each time point, the Sallen-Key bandpass filter circuit is swept from the start frequency to the end frequency, that is, the output voltage of the component to be predicted is obtained at each time point at different frequencies.

[0059] In this embodiment, the detailed implementation steps of step S120 include:

[0060] The collected raw data is denoised, and high-frequency noise is removed based on the wavelet transform technology, retaining the original characteristics of the output voltage signal of capacitor C1 and resistor R1. According to the Pearson correlation coefficient method, the correlation between the output voltage and the specific frequency in the extracted fault characteristic indicators is calculated:

[0061]

[0062] Where x represents the characteristic vector when the output voltage of capacitor C1 and resistor R1 is equal to the nominal value, The characteristic vector representing the health of capacitor C1 and resistor R1 at level i, and Respectively and The arithmetic mean of represents x and The value of the closeness of the linear relationship between them.

[0063] In the experiment, the health data of each component was collected at 600 time points, the data of the first 400 time points were taken as the training set, and the data of the last 200 time points were taken as the validation set to obtain the preprocessed data set. When the fault characteristic index of the predicted component gradually deviates from the nominal value over time at a specific frequency due to loss, the greater the deviation, the lower the health, and the fault will occur when the capacitance and resistance parameter values ​​change by more than 50%.

[0064] In this embodiment, the detailed implementation steps of step S130 include:

[0065] See also Figure 2 , the improved large language model is divided into 4 steps in total.

[0066] S131, time series data block vectorization: In order to capture the data characteristics of the univariate time series data corresponding to "component health-time", the data of each time point with a length of 400 is processed in blocks.

[0067] In order to capture detailed features from shorter time point data. Set the segmentation step to 50 and get a set of time series blocks, denoted as , , each timing block Vectorized processing , we get a set of time series vector blocks, denoted as , .

[0068] S132. Build a linear layer based on the word vector pre-trained by the general large language model, the word vector is denoted as m, and define , where M represents the number of word vectors, P represents the word vector dimension, and the word vector is linearly transformed. The linear transformation calculation is:

[0069]

[0070] in, Represents the word vector after linear combination, definition ,in Represents the number of word vectors after linear combination, represents the constant bias term, represents the weight parameter matrix;

[0071] S133. The general large language model uses a large amount of data from various professional fields for pre-training, can effectively capture contextual information, and has excellent performance in generation tasks. However, in the time series prediction task, due to the difficulty in capturing the features of time series data, the large language model loses a lot of features when converting time series data into understandable natural language. Therefore, the present invention uses the cross-attention mechanism to capture the time series features and the large language model language for cross-modal alignment. The improved large language model can more efficiently capture the features of the "component health-time" univariate time series data in the PCB analog circuit fault prediction task.

[0072] Construct a cross attention layer and record all attention heads as , the query matrix of i attention heads is defined as , the bond matrix is ​​defined as , the value matrix is ​​defined as ,in , Perform cross-attention calculation:

[0073]

[0074] in, represents the cross attention value of the i-th attention head, and we get , aggregate the cross attention values ​​of H attention heads to get ,get , represents the weight parameter matrix of the multi-head attention mechanism;

[0075] According to the cross-attention layer calculation, a set of time series vector blocks Converted to the feature vector after cross-modal alignment, denoted as , ,definition .

[0076] S134, the step of improving the prompt word of the general large language model and linearly connecting the improved prompt word vectorization and the feature vector after cross-modal alignment specifically includes:

[0077] The task scenario description includes the types of components to be predicted in PCB analog circuit fault prediction, the threshold for component failure, and the meaning of the characteristic indicators corresponding to each data point: {"Task scenario description":"The types of components to be predicted in PCB analog circuit fault prediction are resistors. The concept of component health is: component health indicates the degree of degradation. When the component has no loss or failure, the health is 1. When the component's failure characteristic indicator gradually deviates from the nominal value over time at a specific frequency due to loss, the greater the deviation, the lower the health, indicating that the component is more likely to fail. The data in the input data statistics are the health values ​​at each time point. Assuming that you are an expert in the field of PCB analog circuit fault prediction, please predict the future trend of the curve based on the data in the input data statistics."}

[0078] The task instructions include the number of time series blocks obtained by slicing and dividing the input data, the time step of the time series blocks, the order of the time series blocks in the time dimension, and the predicted time step; {"Task Instructions":"The number of time series blocks is 8, the time step of each time series block is 50, and the time sequence is represented by time series blocks 1-8, representing continuous time from far to near. Please predict the data at each time point for the next 200 time steps."}

[0079] The input data statistics include the discrete data points of each time series block and the characteristic statistics of each time series block. The characteristic statistics of each time series block include the maximum value, minimum value, median value, and slope value of the data as a whole of the time series block: {"input data statistics":"<discrete data points of time series block 1>, ..., <discrete data points of time series block 8>, <characteristic statistics of time series block 1>, ..., <characteristic statistics of time series block 8>"}.

[0080] According to the prompt word of the improved general large language model, the prompt word text is vectorized to obtain the prompt word vector, which is recorded as ,Will and Each eigenvector in is added and connected, and the process is expressed as , aggregate i added feature vectors, denoted as , As the input vector of the general large language model.

[0081] In this embodiment, the detailed implementation steps of step S140 include:

[0082] According to steps S131-S134, an improved large language model is obtained, see Figure 2,The preprocessed data set is used to train all the improved layers to adjust the weight parameters of each layer to adapt to the PCB simulation circuit fault prediction task.

[0083] In this embodiment, the detailed implementation steps of step S150 include:

[0084] According to steps S110-S140, according to the trained improved large language model, new component health data to be predicted is input to obtain the final prediction result. The prediction result is compared with the health of the component under the nominal value. When the health value of the prediction result deviates from the nominal value by more than 50%, it means that the component will fail at this time.

[0085] The above disclosure is only the preferred embodiment of the present invention, but the present invention is not limited thereto. Based on the above content of the present invention, other embodiments obtained by those skilled in the art without departing from the principle of the present invention should all fall within the scope of protection of the present invention.

Claims

1. A PCB analog circuit component fault prediction method based on an improved large language model, characterized in that: The steps include: Apply excitation to the PCB simulation circuit and extract the output voltage of the components to be predicted in the PCB simulation circuit at different frequencies; The above frequencies and output voltages are preprocessed according to the Pearson correlation coefficient method. The components are set to change evenly during the aging process. The frequency sweep excitation is performed once at each time point. The data is processed into univariate time series data corresponding to "component health-time" to obtain the preprocessed data set. Constructing an improved large language model; the improved large language model uses the general large language model as a base model, adds a linear layer, a cross attention layer and a fully connected layer, and improves the prompt words of the general large language model; The univariate time series data corresponding to "component health - time" is cut and segmented to obtain time series blocks, and the time series blocks are vectorized to obtain a set of time series vector blocks; The linear layer linearly combines the word vectors pre-trained by the general large language model in the original pre-trained word vector space according to a linear transformation, and the number of word vectors after the linear combination is less than the number of word vectors pre-trained by the general large language model; The cross attention layer performs cross-modal alignment of two different sequences, the time series vector block and the linearly combined word vector, according to the multi-head attention mechanism, so as to selectively focus the word vectors after different linear combinations on the data fluctuation characteristics in the time series vector block; The fully connected layer introduces a ReLU activation function, and performs nonlinear fitting on the feature vector output by the cross attention layer through the ReLU activation function to obtain a feature vector after cross-modal alignment; The improved prompt words of the general large language model are used to guide the general large language model to understand and perform the PCB analog circuit component fault prediction task, and the feature vectors after the improved prompt word vectorization and cross-modal alignment are added and connected, and the feature vectors after the addition and connection are used as the input feature vector of the general large language model; The improved large language model is trained according to the preprocessed data set, the number of training iterations is set, and the training is stopped when the number of training iterations is reached, so as to obtain the trained improved large language model; The input time dimension is represented as the health of the component to be predicted in the interval [0, T], and the health of the component to be predicted after time T is predicted based on the trained improved large language model; a threshold for the health of the component to be predicted is set to indicate that a fault occurs when the health of the component to be predicted is lower than the set threshold. When the health of the component to be predicted at a certain moment after time T in the prediction result is equal to the preset threshold, it indicates that the component to be predicted will fail at that moment.

2. According to the method for predicting faults of PCB analog circuit components based on an improved large language model according to claim 1, it is characterized in that: The block segmentation process and sequential block vectorization specifically include: The preprocessed data with a time step of T is cut into time series blocks with a fixed length, and the segmentation step is set to L. The number of time series blocks is Get a set of timing blocks, denoted as Each time block c k Vectorized processing Get a set of time series vector blocks, denoted as definition Where N represents Quantity, D represents Vector dimensions.

3. A PCB analog circuit component fault prediction method based on an improved large language model according to claim 2, characterized in that: The linear layer processing process is as follows: The word vector pre-trained by the general large language model is denoted as m, and m∈R M×P , where M represents the number of pre-trained word vectors and P represents the dimension of pre-trained word vectors; The linear layer performs a linear transformation on the pre-trained word vector. The linear transformation is calculated as follows: m′=m·E T +b1 Among them, m′ represents the word vector after linear combination, and m′∈R is defined M′×P , M′ represents the number of word vectors after linear combination, b1 represents the constant bias term, E T Represents the weight parameter matrix.

4. The method for predicting faults of PCB analog circuit components based on an improved large language model according to claim 3, characterized in that: The cross-modal alignment specifically includes: All attention heads in the multi-head attention mechanism are denoted as h=[1,2,3,...,H],i∈[1,H], where H represents the number of attention heads; The query matrix of i attention heads is defined as The bond matrix is ​​defined as The value matrix is ​​defined as in, The cross attention is calculated as follows: in, represents the cross attention value of the i-th attention head, and we get Aggregate the cross attention values ​​of H attention heads to get A h , get A h ∈R N×D , represents the weight parameter matrix of the multi-head attention mechanism; According to the calculation of the cross attention layer, the feature vector output by the cross attention layer is nonlinearly fitted through the ReLU activation function of the fully connected layer to obtain the feature vector after cross-modal alignment, which is recorded as definition 5. The method for predicting faults of PCB analog circuit components based on an improved large language model according to claim 4, characterized in that: The improved prompt words of the general large language model are prompt words based on the general large language model, with task scenario description, task instructions and input data statistics added, and stored in the Prompt template of the general large language model in JSON format; The task scenario description includes the types of components to be predicted in PCB analog circuit fault prediction, the threshold at which the components will fail, and the meaning of the characteristic index corresponding to each data point; The task instruction includes the number of time series blocks obtained by the block segmentation process, the time step of the time series blocks, the sequence of the time series blocks in the time dimension and the predicted time step; The input data statistics include discrete data points of each time series block and feature statistics of each time series block; The characteristic statistics of each time series block include the maximum value, minimum value, median value, and slope value of the data as a whole of the time series block; The feature vector after the addition connection is the prompt word of the improved general large language model, and the text is vectorized to obtain the prompt word vector, which is recorded as Will and Each eigenvector in is added and connected, and the process is expressed as Aggregate k added feature vectors, denoted as As the input vector of the general large language model.

6. The method for predicting faults of PCB analog circuit components based on an improved large language model according to claim 5, characterized in that: The data preprocessing specifically includes: A swept frequency excitation source is applied to the PCB analog circuit as the input source, and the output voltage of a specific frequency is uniformly extracted in the entire response frequency band as the fault characteristic index. The fault characteristic index is denoted as U = [u1,u2,u3,...,u v ], where U represents the set of output voltages of the same component at different frequencies, and v represents the number of specific frequencies; According to the Pearson correlation coefficient method, the correlation between the output voltage and the specific frequency in the extracted fault characteristic indicators is calculated: Where x represents the characteristic vector when the output voltage of the component is equal to the nominal value, and y z The eigenvector representing the z-th level health of the component, and Respectively represent x j and zj The arithmetic mean of r(x,y z ) represents x and y z The value of the closeness of the linear relationship between them; when the components have no loss and fault, the performance parameters are nominal values, and the fault characteristic index x extracted at a specific frequency is x = [u1', u'2, u'3, ..., u' v ]; When the component deviates from the nominal value due to loss, the fault characteristic index y extracted at a specific frequency is z =[u' z,1 ,u' z,2 ,u' z,3 ,...,u' z,v ] reflects the changing trend of the component "frequency-output voltage" curve under the performance parameters; r(x,y z ) is used to represent the health of components, and the single variable time series data corresponding to "component health-time" is obtained; when the component has no loss and failure, the fault characteristic index is equal to the nominal value, at this time r(x,y z )=1; When the fault characteristic index of a component gradually deviates from the nominal value over time at a specific frequency due to wear and tear, the greater the deviation, the greater the error. z ) is lower, indicating that the health of the component is lower and the component is more likely to fail. When the health of a component is lower than a certain value, it is considered that the component will fail. According to the type of component, a corresponding threshold is set to indicate that a component will fail when its health is lower than the set threshold.

Images