Circuit breaker fault diagnosis method based on multi-modal data fusion and few-sample learning

Abstract

The invention discloses a circuit breaker fault diagnosis method based on multi-modal data fusion and few-sample learning. The method comprises the following steps: 1) acquiring different modal data of a circuit breaker at the same time by using a data acquisition unit; 2) preprocessing the acquired different modal data by using a data preprocessing unit to obtain primary characteristics of each group of modal data; 3) fusing the primary features of each group of modal data by using a data fusion unit to obtain fused features; and 4) classifying the fusion features by using a fault diagnosis unit to realize fault diagnosis. According to the method, the multi-source sensor information is fused, and the support vector machine classifier is trained by using the MAML-based meta-learning algorithm, so that the fault diagnosis of the circuit breaker under a small number of samples is realized. The circuit breaker fault diagnosis method effectively solves the problems that an existing circuit breaker fault diagnosis method depends on a large amount of labeled data and is difficult to adapt to new fault types.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis, and in particular to a circuit breaker fault diagnosis method based on multimodal data fusion and few-sample learning. Background Art

[0002] As critical equipment in power systems, the stability of circuit breaker operation directly impacts the safe and reliable operation of the power grid. Therefore, real-time status monitoring and fault diagnosis of circuit breakers are crucial. Traditional circuit breaker monitoring methods typically rely on a single data source, such as vibration signals, current signals, or infrared images, to determine the circuit breaker's operating status. However, due to the complexity of circuit breaker operation, relying solely on a single type of data often fails to fully and accurately reflect its true state. For example, relying solely on vibration signals may fail to detect electrical performance degradation; relying solely on current signals may fail to detect loose mechanical structures; and relying solely on infrared images may fail to detect operational synchronization anomalies. These methods all have limitations and can easily lead to misjudgments or omissions of circuit breaker status.

[0003] Furthermore, existing fault diagnosis methods typically require large amounts of labeled data for model training. These methods build fault models by learning from large amounts of historical data, thereby enabling fault diagnosis for new data. However, in actual power systems, abnormal circuit breaker samples are often very scarce, making it difficult to obtain large amounts of labeled data. The scarcity of abnormal circuit breaker samples can easily lead to poor model training results and difficulty in accurately identifying different types of faults. This is especially true when faced with new circuit breakers or new fault modes, as the model's generalization ability is severely insufficient. This makes it difficult for existing fault diagnosis methods to meet the accuracy and efficiency requirements of power systems for circuit breaker status monitoring and fault diagnosis in practical applications. Summary of the Invention

[0004] The present invention aims to provide a circuit breaker fault diagnosis method based on multimodal data fusion and few-sample learning, comprising the following steps:

[0005] 1) Use the data acquisition unit to collect different modal data of the circuit breaker at the same time.

[0006] 2) Use the data preprocessing unit to preprocess the collected different modal data to obtain the primary features of each group of modal data.

[0007] 3) Use the data fusion unit to fuse the primary features of each group of modal data to obtain fused features.

[0008] 4) Use the fault diagnosis unit to classify the fused features and realize fault diagnosis.

[0009] Furthermore, the modal data includes sound signals, vibration signals, and current signals.

[0010] Furthermore, the data preprocessing unit performs the following steps on the collected data of different modalities:

[0011] 2.1) Use the cross-correlation method to align the modal data sets in time.

[0012] 2.2) Use marginal spectral energy entropy to extract primary features of sound and vibration signals.

[0013] 2.3) Use wavelet transform to extract the time-frequency characteristics of the current signal.

[0014] Furthermore, in step 2.2), the steps of extracting primary features using marginal spectrum energy entropy are as follows:

[0015] 2.2.1) The variable mode decomposition method is used to decompose the input signal into a series of eigenmode functions with specific center frequencies and limited bandwidths.

[0016] 2.2.2) Perform Hilbert transform on each eigenmode function to obtain the analytical function Z k (t), and then calculate the instantaneous amplitude A of each eigenmode function k (t) and instantaneous frequency f k (t).

[0017] 2.2.3) Obtain the marginal spectrum h by integral calculation k (f), as follows:

[0018] h k (f) = ∫A k (t)δ(ff k (t))dt (1)

[0019] Where t is time and f is frequency.

[0020] 2.2.4) Calculate the energy E of each eigenmode function k , as shown below:

[0021]

[0022] 2.2.5) Energy E k Normalization is performed and the marginal spectral energy entropy H is calculated. The obtained marginal spectral energy entropy H is the primary feature.

[0023] The marginal spectrum energy entropy H is as follows:

[0024] H=-ΣP k log2(Pk ) (3)

[0025] Where, P k is the normalized energy E k .

[0026] Furthermore, the steps of extracting the time-frequency characteristics of the current signal using wavelet transform are as follows:

[0027] 2.3.1) Use the db4 wavelet basis to perform wavelet packet decomposition on the current signal to obtain the wavelet coefficients of different frequency bands, as shown below:

[0028]

[0029] In the formula, j represents the number of wavelet packet layers, n represents the node number, It represents the wavelet coefficient at the nth node of the jth layer wavelet packet decomposition. h and g are low-pass and high-pass decomposition filters respectively. Represents the wavelet coefficient at the 2nth node of the j-1th layer wavelet packet decomposition. Represents the wavelet coefficient at the 2n+1th node of the j-1th layer wavelet packet decomposition. l represents the sampling point index on the time axis. k represents the filter length or convolution range.

[0030] 2.3.2) Calculate the time domain features and frequency domain features of the wavelet coefficients of each frequency band, and combine the time domain features and frequency domain features into a time-frequency domain feature vector.

[0031] The time domain features include the mean, variance, kurtosis, and skewness of the wavelet coefficients of each frequency band.

[0032] The frequency domain features include the energy of the wavelet coefficients of each frequency band.

[0033] 2.3.3) The time-frequency characteristics of the current signal are obtained by normalizing the time-frequency domain feature vector.

[0034] Furthermore, the data fusion unit fuses the primary features of each group of modal data in the following steps:

[0035] 3.1) Normalize the primary features of different dimensions and then map the normalized primary features to the same dimension.

[0036] 3.2) Using the concat method, the primary features in the same dimension are concatenated and fused to obtain preliminary fused features.

[0037] 3.3) Use principal component analysis to reduce the dimension of the preliminary fusion features and obtain the fusion features.

[0038] Furthermore, the fault diagnosis unit classifies the fused features using a support vector machine classifier based on meta-learning.

[0039] The training process of the support vector machine classifier based on meta-learning is as follows:

[0040] c1 initializes the parameters θ of the classifier based on the MAML algorithm, obtains the fused feature vector, and labels the fused feature vector with the fault type to construct the dataset.

[0041] c2 inputs the data set into the meta-learner, and the meta-learner randomly samples the data set to obtain the sampling result set T = {T1, T2, T3, ..., T h}, where h is the total number of samples.

[0042] c3 inputs the sampling result set T and parameter θ into the support vector machine for training. The support vector machine trains each sampling result T. i , and a model update parameter θ is obtained by training i ', where i = 1, 2, ..., h.

[0043] c4 uses the dataset to evaluate each model and update the parameters θ i ', and calculate the corresponding loss function

[0044] c5 calculates all sampling results T in the sampling result set T i The average loss And determine the average loss Whether it reaches the minimum value, if so, the optimal parameter θ of the classifier is obtained, if not, the parameter And return to step c2, where, is the gradient function. η is the learning rate.

[0045] Furthermore, the average loss As shown below:

[0046]

[0047] Where i is the sampling number and h is the total number of samples. is the sampling result T i The corresponding loss function

[0048] Furthermore, the data acquisition unit includes a sound sensor, a vibration sensor and a current sensor.

[0049] The sound sensor is used to collect sound signals from the circuit breaker.

[0050] The vibration sensor is used to collect vibration signals of the circuit breaker.

[0051] The current sensor is used to collect the current signal of the circuit breaker.

[0052] The technical effect of the present invention is unquestionable. The present invention realizes circuit breaker fault diagnosis with a small number of samples by fusing multi-source sensor information and using a MAML-based meta-learning algorithm to train a support vector machine classifier.

[0053] The present invention effectively solves the problems of existing circuit breaker fault diagnosis methods that rely on a large amount of labeled data and are difficult to adapt to new fault types, and has the following significant technical effects:

[0054] 1. Multimodal data fusion to improve diagnostic accuracy

[0055] This method integrates the circuit breaker's acoustic, vibration, and current signals. A data processing unit performs time-frequency analysis on the different modal data, and a feature fusion unit effectively fuses the features, providing a more comprehensive reflection of the circuit breaker's operating status. Compared to traditional single-data-source approaches, this method captures complementary information between different modal data, significantly improving the accuracy and reliability of fault diagnosis. Experimental results show that the average diagnostic accuracy for common circuit breaker faults increased by over 5% after implementing multimodal data fusion.

[0056] 2. Few-shot learning based on meta-learning reduces dependence on labeled data

[0057] This method uses a meta-learning algorithm based on MAML to train a support vector machine classifier. By learning how to generate good model initialization parameters, the meta-learner enables the support vector machine classifier to quickly adapt to new fault types with only a small number of labeled samples. This effectively addresses the scarcity of circuit breaker fault samples and reduces the reliance on large amounts of labeled data. Even with only five fault samples, this method still achieves a diagnostic accuracy of over 90%, compared to less than 75% for traditional methods.

[0058] 3. Quickly adapt to new fault types and improve model generalization capabilities

[0059] The MAML-based meta-learning algorithm enables the support vector machine classifier to be quickly fine-tuned for new fault types by learning a sensitive model initialization parameter. Compared to traditional support vector machines that require retraining the entire model, this method significantly shortens model training time and improves the model's generalization and practicality. Experimental results show that for new fault types, under the same computing power, this method only requires 3-5 minutes of fine-tuning to achieve a diagnostic accuracy of over 95%, while traditional methods require over 30 minutes to retrain the model and only achieve an accuracy of around 85%.

[0060] 4. Combining VMD and Hilbert transform to effectively extract fault features

[0061] This method uses VMD decomposition and Hilbert transform to process acoustic and vibration signals, effectively suppressing modal mixing and extracting marginal spectral energy entropy, which reflects the distribution of signal energy across different frequency bands, as a fault signature. Compared to traditional Fourier and wavelet transforms, this method more accurately extracts the characteristics of non-stationary and nonlinear signals, improving the sensitivity of fault diagnosis.

[0062] 5. End-to-end diagnostic process to improve efficiency

[0063] The present invention establishes a complete process from data acquisition, data preprocessing, data fusion to fault diagnosis, achieving end-to-end fault diagnosis. The modules are closely connected, with a high degree of automation, which reduces manual intervention and improves diagnostic efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 It is a schematic structural diagram of the circuit breaker fault diagnosis system of the present invention;

[0065] Figure 2 is a flow chart of a data processing unit of the present invention;

[0066] Figure 3 is a flow chart of the feature fusion unit of the present invention;

[0067] Figure 4 This is a flow chart of training and fine-tuning the support vector machine classifier based on meta-learning of the present invention;

[0068] In the figure, there are a data acquisition unit 1, a data preprocessing unit 2, a data fusion unit 3 and a fault diagnosis unit 4. DETAILED DESCRIPTION

[0069] The present invention will be further described below with reference to the following examples, but it should not be understood that the scope of the present invention is limited to the following examples. Without departing from the above technical ideas of the present invention, various substitutions and modifications can be made according to common technical knowledge and customary means in the art, and all should be included in the scope of protection of the present invention.

[0070] Example 1:

[0071] See also Figures 1 to 4 ,A circuit breaker fault diagnosis method based on multimodal data fusion and few-shot learning,,including the following steps:

[0072] 1) Use the data acquisition unit to collect different modal data of the circuit breaker at the same time.

[0073] 2) Use the data preprocessing unit to preprocess the collected different modal data to obtain the primary features of each group of modal data.

[0074] 3) Use the data fusion unit to fuse the primary features of each group of modal data to obtain fused features.

[0075] 4) Use the fault diagnosis unit to classify the fused features and realize fault diagnosis.

[0076] Example 2:

[0077] A circuit breaker fault diagnosis method based on multimodal data fusion and few-sample learning, the main technical content of which is shown in Example 1. Furthermore, the modal data includes sound signals, vibration signals, and current signals.

[0078] Example 3:

[0079] The circuit breaker fault diagnosis method based on multimodal data fusion and few-sample learning has the following main technical contents as shown in any one of Embodiments 1 to 2. Furthermore, the data preprocessing unit preprocesses the collected data of different modalities as follows:

[0080] 2.1) Use the cross-correlation method to align the modal data sets in time.

[0081] 2.2) Use marginal spectral energy entropy to extract primary features of sound and vibration signals.

[0082] 2.3) Use wavelet transform to extract the time-frequency characteristics of the current signal.

[0083] Example 4:

[0084] The circuit breaker fault diagnosis method based on multimodal data fusion and few-sample learning has the following main technical contents as shown in any one of Examples 1 to 3. Furthermore, in step 2.2), the steps of extracting primary features using marginal spectral energy entropy are as follows:

[0085] 2.2.1) The variable mode decomposition method is used to decompose the input signal into a series of eigenmode functions with specific center frequencies and limited bandwidths.

[0086] 2.2.2) Perform Hilbert transform on each eigenmode function to obtain the analytical function Z k (t), and then calculate the instantaneous amplitude A of each eigenmode function k (t) and instantaneous frequency f k (t).

[0087] 2.2.3) Obtain the marginal spectrum h by integral calculation k (f), as follows:

[0088] hk (f) = ∫A k (t)δ(ff k (t))dt (1)

[0089] Where t is time and f is frequency.

[0090] 2.2.4) Calculate the energy E of each eigenmode function k , as shown below:

[0091]

[0092] 2.2.5) Energy E k Normalization is performed and the marginal spectral energy entropy H is calculated. The obtained marginal spectral energy entropy H is the primary feature.

[0093] The marginal spectrum energy entropy H is as follows:

[0094] H=-ΣP k log2(P k ) (3)

[0095] Where, P k is the normalized energy E k .

[0096] Example 5:

[0097] The circuit breaker fault diagnosis method based on multimodal data fusion and few-sample learning has the following main technical contents as shown in any one of Examples 1 to 4. Furthermore, the step of extracting the time-frequency characteristics of the current signal using wavelet transform is as follows:

[0098] 2.3.1) Use the db4 wavelet basis to perform wavelet packet decomposition on the current signal to obtain the wavelet coefficients of different frequency bands, as shown below:

[0099]

[0100] In the formula, j represents the number of wavelet packet layers, n represents the node number, It represents the wavelet coefficient at the nth node of the jth layer wavelet packet decomposition. h and g are low-pass and high-pass decomposition filters respectively. Represents the wavelet coefficient at the 2nth node of the j-1th layer wavelet packet decomposition. Represents the wavelet coefficient at the 2n+1th node of the j-1th layer wavelet packet decomposition. l represents the sampling point index on the time axis. k represents the filter length or convolution range.

[0101] 2.3.2) Calculate the time domain features and frequency domain features of the wavelet coefficients of each frequency band, and combine the time domain features and frequency domain features into a time-frequency domain feature vector.

[0102] The time domain features include the mean, variance, kurtosis, and skewness of the wavelet coefficients of each frequency band.

[0103] The frequency domain features include the energy of the wavelet coefficients of each frequency band.

[0104] 2.3.3) The time-frequency characteristics of the current signal are obtained by normalizing the time-frequency domain feature vector.

[0105] Example 6:

[0106] The circuit breaker fault diagnosis method based on multimodal data fusion and few-sample learning has the following main technical contents as shown in any one of Examples 1 to 5. Furthermore, the data fusion unit fuses the primary features of each group of modal data as follows:

[0107] 3.1) Normalize the primary features of different dimensions and then map the normalized primary features to the same dimension.

[0108] 3.2) Using the concat method, the primary features in the same dimension are concatenated and fused to obtain preliminary fused features.

[0109] 3.3) Use principal component analysis to reduce the dimension of the preliminary fusion features and obtain the fusion features.

[0110] Example 7:

[0111] A circuit breaker fault diagnosis method based on multimodal data fusion and few-sample learning, the main technical content of which can be found in any one of Examples 1 to 6. Furthermore, the fault diagnosis unit uses a support vector machine classifier based on meta-learning to classify the fused features.

[0112] The training process of the support vector machine classifier based on meta-learning is as follows:

[0113] c1 initializes the parameters θ of the classifier based on the MAML algorithm, obtains the fused feature vector, and labels the fused feature vector with the fault type to construct the dataset.

[0114] c2 inputs the data set into the meta-learner, and the meta-learner randomly samples the data set to obtain the sampling result set T = {T1, T2, T3, ..., T h}, where h is the total number of samples.

[0115] c3 inputs the sampling result set T and parameter θ into the support vector machine for training. The support vector machine trains each sampling result T. i , and a model update parameter θ is obtained by training i ', where i = 1, 2, ..., h.

[0116] c4 uses the dataset to evaluate each model and update the parameters θ i ', and calculate the corresponding loss function

[0117] c5 calculates all sampling results T in the sampling result set T i The average loss And determine the average loss Whether it reaches the minimum value, if so, the optimal parameter θ of the classifier is obtained, if not, the parameter And return to step c2, where, is the gradient function. η is the learning rate.

[0118] Example 8:

[0119] A circuit breaker fault diagnosis method based on multimodal data fusion and few-sample learning, the main technical content of which is shown in any one of embodiments 1 to 7. Further, the average loss As shown below:

[0120]

[0121] Where i is the sampling number and h is the total number of samples. is the sampling result T i The corresponding loss function

[0122] Example 9:

[0123] The circuit breaker fault diagnosis method based on multimodal data fusion and few-sample learning, the main technical content of which is shown in any one of Examples 1 to 8. Furthermore, the data acquisition unit includes a sound sensor, a vibration sensor, and a current sensor.

[0124] The sound sensor is used to collect sound signals from the circuit breaker.

[0125] The vibration sensor is used to collect vibration signals of the circuit breaker.

[0126] The current sensor is used to collect the current signal of the circuit breaker.

[0127] Example 10:

[0128] See also Figures 1 to 4 ,A circuit breaker fault diagnosis method based on multimodal data fusion and few-shot learning,,including the following steps:

[0129] 1) Use the data acquisition unit to collect different modal data of the circuit breaker at the same time.

[0130] 2) Use the data preprocessing unit to preprocess the collected different modal data to obtain the primary features of each group of modal data.

[0131] 3) Use the data fusion unit to fuse the primary features of each group of modal data to obtain fused features.

[0132] 4) Use the fault diagnosis unit to classify the fused features and realize fault diagnosis.

[0133] Example 11:

[0134] A circuit breaker fault diagnosis method based on multimodal data fusion and few-sample learning, the main technical content of which is shown in Example 10. Furthermore, the modal data includes sound signals, vibration signals, and current signals.

[0135] Example 12:

[0136] The circuit breaker fault diagnosis method based on multimodal data fusion and few-sample learning has the following main technical contents as shown in any one of Embodiments 10 to 11. Furthermore, the data preprocessing unit preprocesses the collected data of different modalities as follows:

[0137] 2.1) The collected modal data may have a certain degree of time delay. The cross-correlation method is used to align the modal data groups in time.

[0138] 2.2) Use marginal spectral energy entropy to extract primary features of sound and vibration signals.

[0139] 2.3) Use wavelet transform to extract the time-frequency characteristics of the current signal.

[0140] Example 13:

[0141] The circuit breaker fault diagnosis method based on multimodal data fusion and few-sample learning has the following main technical contents as shown in any one of Examples 10 to 12. Furthermore, in step 2.2), the steps of extracting primary features using marginal spectral energy entropy are as follows:

[0142] 2.2.1) The variable mode decomposition (VMD) method is used to decompose the input signal into a series of intrinsic mode functions (IMFs) with specific center frequencies and limited bandwidths. The center frequency range is set from 0 Hz to 1 / 2 of the signal sampling frequency. The specific center frequency values ​​are determined based on the typical frequency components of the circuit breaker operation state: 50 Hz (grid fundamental frequency), 100 Hz (second harmonic), and 200 Hz to 300 Hz (mechanical vibration characteristic frequency) to fully capture the characteristic signals of electrical and mechanical faults. The limited bandwidth is set to 15% to 25% of the center frequency, that is, for a center frequency of 50 Hz, the bandwidth range is 7.5 Hz to 12.5 Hz; for a center frequency of 100 Hz, the bandwidth range is 15 Hz to 25 Hz; for a center frequency of 200 Hz to 300 Hz, the bandwidth range is 30 Hz to 75 Hz. This range ensures that the decomposed intrinsic mode functions can accurately capture the main characteristics of the signal while effectively avoiding modal aliasing.

[0143] 2.2.2) Perform Hilbert transform on each eigenmode function to obtain the analytical function Z k (t), and then calculate the instantaneous amplitude A of each eigenmode function k (t) and instantaneous frequency f k (t).

[0144] 2.2.3) Obtain the marginal spectrum h by integral calculation k (f), as follows:

[0145] h k (f) = ∫A k (t)δ(ff k (t))dt (1)

[0146] Where t is time and f is frequency.

[0147] 2.2.4) Calculate the energy E of each eigenmode function k , as shown below:

[0148]

[0149] 2.2.5) Energy E k Normalization is performed and the marginal spectral energy entropy H is calculated. The obtained marginal spectral energy entropy H is the primary feature.

[0150] The marginal spectrum energy entropy H is as follows:

[0151] H=-ΣP k log2(Pk ) (3)

[0152] Where, P k is the normalized energy E k .

[0153] Example 14:

[0154] The circuit breaker fault diagnosis method based on multimodal data fusion and few-sample learning has the following main technical contents as shown in any one of Embodiments 10 to 13. Furthermore, the step of extracting the time-frequency features of the current signal using wavelet transform is as follows:

[0155] 2.3.1) Use the db4 wavelet basis to perform wavelet packet decomposition on the current signal to obtain the wavelet coefficients of different frequency bands, as shown below:

[0156]

[0157] In the formula, j represents the number of wavelet packet layers, n represents the node number, It represents the wavelet coefficient at the nth node of the jth layer wavelet packet decomposition. h and g are low-pass and high-pass decomposition filters respectively. Represents the wavelet coefficient at the 2nth node of the j-1th layer wavelet packet decomposition. Represents the wavelet coefficient at the 2n+1th node of the j-1th layer of the wavelet packet decomposition. l represents the sampling point index on the time axis, which is used to indicate the position of the signal in the discrete time domain. k represents the filter length or convolution range. The low-pass filter h and high-pass filter g of the DB4 wavelet typically have 8 coefficients, so the value of K ranges from 0 to 7.

[0158] 2.3.2) Calculate the time domain features and frequency domain features of the wavelet coefficients of each frequency band, and combine the time domain features and frequency domain features into a time-frequency domain feature vector.

[0159] The time domain features include the mean, variance, kurtosis, and skewness of the wavelet coefficients of each frequency band.

[0160] The frequency domain features include the energy of the wavelet coefficients of each frequency band.

[0161] 2.3.3) The time-frequency characteristics of the current signal are obtained by normalizing the time-frequency domain feature vector.

[0162] Example 15:

[0163] The circuit breaker fault diagnosis method based on multimodal data fusion and few-sample learning has the following main technical contents as shown in any one of Embodiments 10 to 14. Furthermore, the data fusion unit fuses the primary features of each group of modal data as follows:

[0164] 3.1) Standardize the primary features of different dimensions to make their mean 0 and variance 1, and then map the standardized primary features to the same dimension.

[0165] 3.2) Using the concat method, the primary features in the same dimension are concatenated and fused to obtain preliminary fused features.

[0166] 3.3) Principal component analysis (PCA) is used to reduce the dimension of the preliminary fusion features to obtain fusion features, thereby obtaining more prominent features, reducing model complexity, and improving the generalization ability of the model.

[0167] Example 16:

[0168] A circuit breaker fault diagnosis method based on multimodal data fusion and few-sample learning, the main technical content of which can be found in any one of Examples 10 to 15. Furthermore, the fault diagnosis unit uses a support vector machine classifier based on meta-learning to classify the fused features.

[0169] The training process of the support vector machine classifier based on meta-learning is as follows:

[0170] c1 initializes the parameters θ of the classifier based on the MAML algorithm, obtains the fused feature vector, and labels the fused feature vector with the fault type to construct the dataset.

[0171] The dataset includes a training set and a validation set.

[0172] The training set is used to input the meta-learner for random sampling.

[0173] The validation set is used to evaluate each model update parameter θ i '.

[0174] c2 inputs the data set into the meta-learner, and the meta-learner randomly samples the data set to obtain the sampling result set T = {T1, T2, T3, ..., T h}, where h is the total number of samples.

[0175] c3 inputs the sampling result set T and parameter θ into the support vector machine for training. The support vector machine trains each sampling result T. i , and a model update parameter θ is obtained by training i ', where i = 1, 2, ..., h.

[0176] c4 uses the dataset to evaluate each model and update the parameters θ i ', and calculate the corresponding loss function

[0177] c5 calculates all sampling results T in the sampling result set T i The average loss And determine the average loss Whether it reaches the minimum value, if so, the optimal parameter θ of the classifier is obtained, if not, the parameter And return to step c2, where, is the gradient function. η is the learning rate.

[0178] By judging the average loss of two consecutive Is there any change in the calculation result of Reached minimum value.

[0179] Example 17:

[0180] A circuit breaker fault diagnosis method based on multimodal data fusion and few-sample learning, the main technical content of which is shown in any one of embodiments 10 to 16. Further, the average loss As shown below:

[0181]

[0182] Where i is the sampling number and h is the total number of samples. is the sampling result T i The corresponding loss function

[0183] Example 18:

[0184] The circuit breaker fault diagnosis method based on multimodal data fusion and few-sample learning, the main technical content of which can be found in any one of Examples 10 to 17. Furthermore, the data acquisition unit includes a sound sensor, a vibration sensor, and a current sensor.

[0185] The sound sensor is used to collect sound signals from the circuit breaker.

[0186] The vibration sensor is used to collect vibration signals of the circuit breaker.

[0187] The current sensor is used to collect the current signal of the circuit breaker.

[0188] Example 19:

[0189] A circuit breaker fault diagnosis system using the method described in any one of embodiments 10-18 includes: a data acquisition unit 1, a data preprocessing unit 2, a data fusion unit 3, and a fault diagnosis unit 4.

[0190] The data acquisition unit 1 acquires different modal data of the circuit breaker at the same time.

[0191] The data preprocessing unit 2 is used to preprocess the collected different modal data to obtain primary features of each group of modal data.

[0192] The data fusion unit 3 is used to fuse the primary features of each group of modal data to obtain a fused feature, which is used as a final feature vector for fault diagnosis.

[0193] The fault diagnosis unit 4 uses a support vector machine classifier based on meta-learning to classify the fused features to achieve fault diagnosis.

[0194] Example 20:

[0195] See also Figures 1 to 4 , a circuit breaker fault diagnosis method based on multimodal data fusion and few-shot learning, the main technical contents include:

[0196] like Figure 1 As shown, the circuit breaker fault diagnosis system proposed in the present invention mainly includes: a data acquisition unit 1, a data preprocessing unit 2, a data fusion unit 3 and a fault diagnosis unit 4.

[0197] Data acquisition unit 1: This unit includes a sound sensor, a vibration sensor, and a current sensor, and is responsible for collecting multimodal data of the circuit breaker.

[0198] Data processing unit 2: responsible for preprocessing the collected multimodal data to obtain the primary features of each modal data.

[0199] Feature fusion unit 3: responsible for fusing primary features to obtain fused features, which are used as the final feature vector for fault diagnosis.

[0200] Fault diagnosis unit 4: Use the support vector machine classifier based on meta-learning to classify the fused features and realize fault diagnosis.

[0201] like Figure 2 FIG. 1 is a flow chart of a data processing unit of the present invention, comprising the following steps:

[0202] In step 201 , the collected modal data may have a certain degree of time delay, and a cross-correlation method is used to achieve temporal alignment of the groups of multimodal data.

[0203] Step 202: For the sound signal and the vibration signal, the same method is used to extract features:

[0204] Firstly, the variational mode decomposition (VMD) is used to decompose the signal into a series of intrinsic mode functions (IMFs) with specific center frequencies and limited bandwidths.

[0205] Then, perform Hilbert transform on each IMF component to obtain the analytical function Z k (t), and thus calculate the instantaneous amplitude A of each IMF component k (t) and instantaneous frequency f k (t). Then, perform the integration operation to obtain the marginal spectrum h k (f):h k (f) = ∫A k (t)δ(ff k (t))dt.

[0206] Finally, the energy of each IMF component is calculated Then normalize and get Pk:P k =E k / ΣE i Finally, calculate the marginal spectrum energy entropy H: H = -ΣP k log2(P k ), and use it as the primary feature of sound and vibration signals.

[0207] Step 203: For the current signal, use wavelet transform to extract its time-frequency features.

[0208] First, the db4 wavelet basis is used to perform 5-layer wavelet packet decomposition on the current signal to obtain wavelet coefficients of different frequency bands. The wavelet packet coefficients on the nth node of the jth layer wavelet packet decomposition are

[0209] They are low-pass and high-pass analysis filters respectively.

[0210] Then, for each frequency band, the wavelet coefficients are calculated as the time domain statistical features, and the energy is calculated as the frequency domain features. These features are combined into a time-frequency domain feature vector and normalized.

[0211] like Figure 3 As shown, it is the feature fusion module of the present invention.

[0212] In step 301, the primary features of different dimensions are normalized to have a mean of 0 and a variance of 1. They are then mapped to the same dimension, with the target dimension taking the minimum value of the primary feature dimension.

[0213] Step 302: Use the concat method to merge the features in the same space.

[0214] In step 303, the principal component analysis (PCA) method is used to reduce the dimension of the fused features, thereby obtaining more prominent features, reducing the model complexity, and improving the generalization ability of the model.

[0215] like Figure 4 As shown in FIG, the support vector machine classifier training and fine-tuning process based on meta-learning of the present invention.

[0216] Step 401: perform a test on the labeled training set and validation set (both of which must contain all task types to achieve the best model effect). Figure 2 、 Figure 3 The operation is to obtain the fusion feature vector set (including labels), which is recorded as X = {X1, X2, X3, ..., X m} and Y={Y1,Y2,Y3,…Y n}. The label is the fault type.

[0217] Step 402: Based on the training set and the validation set, an initialization model of the classifier is obtained through iterative training. The specific steps are as follows:

[0218] 1. Input sets X and Y into the meta-learner based on the MAML algorithm,

[0219] 2. The meta-learner randomly samples X and obtains T = {T1, T2, T3, ..., T h}, pass the sampling result T and the initialization parameter θ of the meta-learner into the support vector machine.

[0220] 3. The support vector machine is trained based on the input parameters θ and T. For each T i ∈T, all trained to get the model update parameter θ i '.

[0221] 4. Using the entire validation set Y, i 'Perform performance evaluation and calculate each T i loss .

[0222] 5. According to the average loss of all samples in T Update the meta-learner parameters using gradient descent

[0223] 6. Repeat steps 2-5 until the average loss is minimized and the optimal initialization parameter θ of the classifier is obtained.

[0224] Step 403: perform the support set Figure 2 、 Figure 3 The operation is to obtain the fusion feature vector set (including labels) Z = {Z1, Z2, Z3, ..., ZN}.

[0225] Step 404: input Z into a support vector machine for training to achieve fine-tuning of model parameters.

[0226] Model training and fine-tuning are performed holistically. The technical solution focuses solely on the training and fine-tuning of the meta-learner-based classifier model within the fault diagnosis unit. In practice, sensor signals sequentially pass through the signal processing unit, feature fusion unit, and fault diagnosis unit to achieve circuit breaker status perception and fault diagnosis.

Claims

1. A circuit breaker fault diagnosis method based on multimodal data fusion and few-shot learning, characterized by: The following steps are involved: 1) Use the data acquisition unit to collect different modal data of the circuit breaker at the same time. 2) Using a data preprocessing unit to preprocess the collected different modal data to obtain primary features of each group of modal data; 3) Using the data fusion unit to fuse the primary features of each group of modal data to obtain fused features; 4) Use the fault diagnosis unit to classify the fused features and realize fault diagnosis.

2. The circuit breaker fault diagnosis method based on multimodal data fusion and few-sample learning according to claim 1 is characterized in that: The modal data includes sound signals, vibration signals, and current signals.

3. The circuit breaker fault diagnosis method based on multimodal data fusion and few-sample learning according to claim 1 is characterized in that: The steps of preprocessing the collected data of different modalities by the data preprocessing unit are as follows: 2.1) Using cross-correlation method to align the modal data sets in time; 2.2) Extract primary features of sound and vibration signals using marginal spectrum energy entropy; 2.3) Use wavelet transform to extract the time-frequency characteristics of the current signal.

4. The circuit breaker fault diagnosis method based on multimodal data fusion and few-sample learning according to claim 3 is characterized in that: In step 2.2), the steps for extracting primary features using marginal spectral energy entropy are as follows: 2.2.1) Decompose the input signal into a series of intrinsic mode functions with specific center frequencies and limited bandwidths using the variable mode decomposition method; 2.2.2) Perform Hilbert transform on each eigenmode function to obtain the analytical function Z k (t), and then calculate the instantaneous amplitude A of each eigenmode function k (t) and instantaneous frequency f k (t); 2.2.3) Obtain the marginal spectrum h by integral calculation k (f), as follows: h k (f)=∫A k (t)δ(f-f k (t))dt (1) Where t is time and f is frequency; 2.2.4) Calculate the energy E of each eigenmode function k , as shown below: 2.2.5) Energy E k Perform normalization and calculate the marginal spectrum energy entropy H. The obtained marginal spectrum energy entropy H is the primary feature; The marginal spectrum energy entropy H is as follows: H=-ΣP k log2(P k ) (3) Where, P k is the normalized energy E k .

5. The circuit breaker fault diagnosis method based on multimodal data fusion and few-sample learning according to claim 3, characterized in that the step of extracting the time-frequency characteristics of the current signal using wavelet transform is as follows: 2.3.1) Use the db4 wavelet basis to perform wavelet packet decomposition on the current signal to obtain the wavelet coefficients of different frequency bands, as shown below: In the formula, j represents the number of wavelet packet layers, n represents the node number, represents the wavelet coefficient at the nth node of the jth layer wavelet packet decomposition; h and g are low-pass and high-pass decomposition filters respectively; Represents the wavelet coefficient at the 2nth node of the j-1th layer wavelet packet decomposition; represents the wavelet coefficient at the 2n+1th node of the j-1th layer wavelet packet decomposition; l represents the sampling point index on the time axis; k represents the length of the filter or the convolution range; 2.3.2) Calculate the time domain features and frequency domain features of the wavelet coefficients of each frequency band, and combine the time domain features and frequency domain features into a time-frequency domain feature vector; The time domain features include the mean, variance, kurtosis and skewness of the wavelet coefficients of each frequency band; The frequency domain features include the energy of the wavelet coefficients of each frequency band; 2.3.3) The time-frequency characteristics of the current signal are obtained by normalizing the time-frequency domain feature vector.

6. The circuit breaker fault diagnosis method based on multimodal data fusion and few-sample learning according to claim 1, characterized in that: The steps of fusing the primary features of each group of modal data by the data fusion unit are as follows: 3.1) Normalize the primary features of different dimensions and then map the normalized primary features to the same dimension; 3.2) Using the concat method, the primary features under the same dimension are concatenated and fused to obtain the preliminary fused features; 3.3) Use principal component analysis to reduce the dimension of the preliminary fusion features and obtain the fusion features.

7. The circuit breaker fault diagnosis method based on multimodal data fusion and few-sample learning according to claim 1, characterized in that: The fault diagnosis unit classifies the fused features using a support vector machine classifier based on meta-learning; The training process of the support vector machine classifier based on meta-learning is as follows: c1 initializes the parameters θ of the classifier based on the MAML algorithm, obtains the fused feature vector, and labels the fused feature vector with the fault type to construct the data set; c2 inputs the data set into the meta-learner, and the meta-learner randomly samples the data set to obtain the sampling result set T = {T1, T2, T3, ..., T h }, where h is the total number of samples; c3 inputs the sampling result set T and parameter θ into the support vector machine for training. The support vector machine trains each sampling result T. i , and a model update parameter θ′ is obtained by training i , where i = 1, 2, ..., h; c4 uses the dataset to evaluate each model to update the parameters θ′ i , and calculate the corresponding loss function L Ti (θ′ i ); c5 calculates all sampling results T in the sampling result set T i The average loss And determine the average loss Whether it reaches the minimum value, if so, the optimal parameter θ of the classifier is obtained, if not, the parameter And return to step c2, where ▽ is the gradient function and η is the learning rate.

8. The circuit breaker fault diagnosis method based on multimodal data fusion and few-sample learning according to claim 7, characterized in that: The average loss As shown below: Where i is the sampling number, h is the total number of samples; L Ti (θ′ i ) is the sampling result T i The corresponding loss function L Ti (θ′ i ).

9. The circuit breaker fault diagnosis method based on multimodal data fusion and few-sample learning according to claim 1, characterized in that: The data acquisition unit includes a sound sensor, a vibration sensor and a current sensor; The sound sensor is used to collect sound signals from the circuit breaker; The vibration sensor is used to collect vibration signals of the circuit breaker; The current sensor is used to collect the current signal of the circuit breaker.

Images