Transformer multi-modal fault diagnosis method based on weighted naive bayes

Abstract

The application relates to the technical field of fault prediction diagnosis, in particular to a transformer multi-modal fault diagnosis method based on a weighted naive Bayes, which comprises the following steps: acquiring a multi-modal feature matrix of each preset sliding time window in a transformer operation process; acquiring an abnormality recognition degree of each physical quantity signal row vector in the multi-modal feature matrix, combining abnormal mutation characteristics of each physical quantity signal, and constructing a compensation adjustment coefficient to correct and compensate information gain, so that an information gain correction value of each physical quantity signal row vector in the multi-modal feature matrix is obtained, then the weight coefficient of a multi-modal feature vector in a weighted naive Bayes model is determined, and the weighted naive Bayes model is used to identify the fault type of the transformer. According to the application, the weight coefficient determined by the information gain is accurately corrected and compensated, and the accuracy of diagnosing the fault type of the transformer is improved.

Description

Technical Field

[0001] This application relates to the field of fault prediction and diagnosis technology, specifically to a transformer multimodal fault diagnosis method based on weighted Naive Bayes. Background Technology

[0002] During long-term operation, transformers are subjected to the combined effects of various factors, including electrical stress, thermal stress, mechanical stress, operating conditions, and meteorological environment. Once a transformer malfunctions, it is crucial to accurately identify the fault promptly and efficiently to prevent further escalation and ensure the reliability of the power supply system. Therefore, to guarantee the safety and reliability of the power supply system, timely fault diagnosis and health management of transformers are necessary to improve the operational efficiency of transformer equipment.

[0003] To diagnose transformer equipment fault types, existing technologies utilize multimodal sensing data from the transformer equipment and employ a weighted Naive Bayes model for fault type identification. In this model, the weighting coefficients of the multimodal feature parameters are typically determined using information gain. However, this information gain-based method does not adequately consider the data distribution characteristics of the multimodal feature parameters. Consequently, the weighting coefficients fail to accurately represent the contribution of these parameters to the classification of transformer equipment fault types, resulting in poor accuracy in fault type diagnosis and impacting the power supply security and reliability of the power system. Summary of the Invention

[0004] To address the aforementioned technical problems, this application provides a transformer multimodal fault diagnosis method based on weighted Naive Bayes, thereby resolving the existing issues.

[0005] The transformer multimodal fault diagnosis method based on weighted Naive Bayes in this application adopts the following technical solution: One embodiment of this application provides a transformer multimode fault diagnosis method based on weighted Naive Bayes, including the following steps: Data of various physical quantity signals during transformer operation are acquired to construct multimodal feature matrices for each preset sliding time window, where each row vector corresponds to a physical quantity signal; The trend component and residual component of the multimodal feature matrix of each sliding time window are extracted, and the anomaly identification degree of each physical quantity signal row vector in the multimodal feature matrix is ​​obtained by the change of the trend component and the degree of random fluctuation of the residual component. Based on the anomaly recognition degree and combined with the abnormal change characteristics of each physical quantity signal, a compensation adjustment coefficient is constructed. Based on the compensation adjustment coefficient, the information gain is corrected and compensated to obtain the information gain correction value of each physical quantity signal row vector in the multimodal feature matrix. The weight coefficients of the multimodal eigenvectors in the weighted naive Bayes model are determined by using the information gain correction value, and the fault types of transformers are identified by using the weighted naive Bayes model.

[0006] Preferably, the row vectors of each physical quantity signal are constructed into a Hankel matrix using a singular spectrum analysis algorithm, and the Hankel matrix is ​​subjected to singular spectrum analysis to obtain each trend component and each residual component in the Hankel matrix.

[0007] Preferably, before obtaining the anomaly identification degree of each physical quantity signal row vector, the anomaly trend degree of each trend component is calculated: , Let be the degree of abnormal trend of the i-th trend component. and Let i be the maximum and minimum values ​​of the i-th trend component. This is the preset first parameter.

[0008] Preferably, before obtaining the anomaly identification degree of each physical quantity signal row vector, the product of the mean of the permutation entropy of all residual components and the dispersion of the permutation entropy of all residual components is calculated as the anomaly randomness of each physical quantity signal row vector.

[0009] Preferably, the anomaly detection rate of each physical quantity signal row vector in the multimodal feature matrix is ​​obtained as follows: In the formula, Let be the anomaly detection score of the row vector of the m-th physical quantity signal. The maximum value of the abnormal trend degree of all trend components in the Hankel matrix constructed for the row vector of the m-th physical quantity signal. Let be the degree of anomalous randomness of the row vector of the m-th physical quantity signal. It is an exponential function with the natural constant as its base. This is a function that takes the minimum value.

[0010] Preferably, before constructing the compensation adjustment coefficient, the absolute value of each residual component is taken to form an energy fluctuation sequence. The energy fluctuation sequence is used as the input of the Bernaola Galvan segmentation algorithm to obtain the number of mutation points in each residual component. The ratio of the number of mutation points in each residual component to the total number is calculated. The mean of the ratios calculated for all residual components is used as the anomalous burstiness of each physical quantity signal row vector in the multimodal feature matrix.

[0011] Preferably, the compensation adjustment coefficient is obtained as follows: In the formula, Let be the compensation adjustment coefficient for the row vector of the m-th physical quantity signal. Let be the anomaly detection score of the row vector of the m-th physical quantity signal. For preset weighting coefficients, It is the sum of the absolute differences between the abnormal bursts of the m-th physical quantity signal row vector and the abnormal bursts of all other physical quantity signal row vectors.

[0012] Preferably, the information gain correction value for each physical quantity signal row vector in the multimodal feature matrix is ​​obtained as follows: In the formula, This is the information gain correction value for the row vector of the m-th physical quantity signal. Let be the compensation adjustment coefficient for the row vector of the m-th physical quantity signal. The basic information gain of the m-th physical quantity signal row vector is obtained offline through the information gain algorithm.

[0013] Preferably, the vector composed of the RMS effective value, variance, mean absolute deviation, energy entropy, kurtosis and skewness of each physical quantity signal row vector is used as the multimode feature vector of the transformer.

[0014] Preferably, the weight coefficients of the multimodal feature vectors in the weighted Naive Bayes model are the normalized results of the information gain correction values ​​of each physical quantity signal row vector in the multimodal feature matrix.

[0015] This application has at least the following beneficial effects: This application fully considers the data distribution characteristics of physical quantity time-series signals. By performing singular spectrum analysis on the physical quantity signals, trend components and residual components are extracted. Based on the analysis of trend components and residual components, the cumulative effect of long-term operation of transformers and random fault events are measured. This accurately measures the effectiveness of different physical quantities in identifying transformer fault types, which is beneficial to improving the accuracy of transformer fault type diagnosis in the future. This application uses information gain to determine the weighting coefficients for accurate correction and compensation. By analyzing the sudden anomalies of physical quantity signals during transformer operation and the effectiveness of identifying transformer fault types based on physical quantities, the compensation adjustment coefficient of each physical quantity signal row vector is measured more accurately, thereby more accurately correcting and compensating the weighting coefficients determined by information gain. This application fully considers the data distribution characteristics of multimodal feature parameters and corrects and compensates the weighting coefficients determined by the information gain method through compensation adjustment coefficients, thereby more accurately reflecting the contribution of multimodal feature parameters to the classification of transformer equipment fault types and improving the accuracy of transformer fault type diagnosis. Attached Figure Description

[0016] To more clearly illustrate the technical solutions and advantages in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 The flowchart of the steps of the transformer multimodal fault diagnosis method based on weighted Naive Bayes provided in this application. Detailed Implementation

[0018] To further illustrate the technical means and effects adopted by this application to achieve the intended inventive objective, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of the transformer multimodal fault diagnosis method based on weighted Naive Bayes proposed in this application. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0019] Unless otherwise defined, terms such as “comprising,” “including,” or any other variations thereof are intended to cover a non-exclusive inclusion, such that a circuit structure, article, or device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such an article or device. Without further limitation, an element defined by the phrase “comprising one…” does not exclude the presence of other identical elements in the article or device that includes said element. Furthermore, the term “and / or” as used herein includes any and all combinations of one or more of the associated listed items. All technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0020] The following description, in conjunction with the accompanying drawings, details the specific scheme of the transformer multimodal fault diagnosis method based on weighted Naive Bayes provided in this application.

[0021] This application provides an embodiment of a transformer multimode fault diagnosis method based on weighted Naive Bayes. For details, please refer to [link to specific implementation details]. Figure 1 This includes the following steps: Step 1: Acquire data of various physical quantity signals during transformer operation to construct multimodal feature matrices for each preset sliding time window, where each row vector corresponds to a physical quantity signal.

[0022] To avoid affecting the power supply safety and reliability of the power system, this invention needs to fully consider the data distribution characteristics of multimodal characteristic parameters and accurately correct and compensate the weight coefficients determined by the information gain method. This will more accurately reflect the contribution of multimodal characteristic parameters to the classification of transformer equipment fault types and improve the accuracy of transformer fault type diagnosis.

[0023] Therefore, in this embodiment, the required transformer multimodal parameters are first collected. The data acquisition equipment includes a vibration acceleration sensor, an optical fiber temperature sensor, a microphone array, an ultra-high frequency partial discharge sensor, a core / clamping grounding current sensor, and a DC bias monitoring device. All data acquisition equipment is deployed at the transformer site. The physical quantity signals of the transformer's operating status are collected in real time and synchronously through the data acquisition equipment. The physical quantity signals include vibration signals, temperature signals, sound signals, partial discharge signals, grounding current signals, and DC polarization signals. Then, the real-time synchronously collected physical quantity signals are transmitted to the edge computing gateway. Clock alignment is performed on the edge computing gateway, and a sliding time window with a time length of 30 seconds and a sliding step size of 10 seconds is set to obtain the physical quantity signals within each sliding time window.

[0024] To eliminate the influence of the dimensions of various physical quantity signals on subsequent analysis and reduce the impact of external noise interference, maximum value normalization is performed on each physical quantity signal within each sliding time window. The normalized physical quantity signals are then used as input to a wavelet denoising algorithm. The algorithm selects the db4 function as the wavelet basis function, presets a decomposition level of 3, and uses a hard thresholding type. The wavelet denoising algorithm is then used to obtain the denoised physical quantity signals within each sliding time window. Furthermore, based on all the denoised physical quantity signals within each sliding time window, a multimodal feature matrix for each sliding time window is constructed. Where M is the number of physical quantity signals, The number of data sample points within the sliding time window is used to analyze the data distribution characteristics of each physical quantity signal in all physical quantity signals. The weight coefficients determined by the information gain method are accurately corrected and compensated, so as to more accurately reflect the contribution of the multimodal feature parameters to the classification of transformer equipment fault types and improve the accuracy of transformer fault type diagnosis.

[0025] Step 2: Extract the trend component and residual component of the multimodal feature matrix for each sliding time window, and obtain the anomaly identification degree of each physical quantity signal row vector in the multimodal feature matrix by the changes in the trend component and the degree of random fluctuation of the residual component.

[0026] Because the information gain calculation method is based on the mutual information between multimodal feature parameters and fault categories to calculate the weight coefficients of multimodal feature parameters in the weighted Naive Bayes model, it does not fully consider the data distribution characteristics of multimodal feature parameters, which can easily lead to poor accuracy in diagnosing transformer fault types. Therefore, in order to ensure that the weight coefficients of multimodal feature parameters can accurately represent the contribution of multimodal feature parameters to the classification of transformer equipment fault types, it is necessary to fully consider the data distribution characteristics of multimodal feature parameters and accurately correct and compensate for the weight coefficients determined by the information gain method.

[0027] Therefore, in order to accurately analyze the data distribution characteristics of multimodal feature parameters, for each sliding time window of the multimodal feature matrix, the row vector of each physical quantity signal in the multimodal feature matrix is ​​used as the input to the Singular Spectrum Analysis algorithm, where the preset window length is... , To achieve the floor function, the grouping strategy uses a weighted correlation graph for grouping. The singular spectrum analysis algorithm is used to construct a Hankel matrix from the row vectors of each physical quantity signal. Singular spectrum analysis is then performed on the Hankel matrix to obtain each trend component and each residual component, reflecting the data distribution characteristics of each physical quantity signal in the multimodal characteristic matrix.

[0028] Generally, the physical quantity signals of transformer operating status exhibit regular periodic variations. Trend variations can measure the short-term offset trend and transient fluctuations of the transformer's operating status within the current sliding time window, while residual variations can indicate random fault events in transformer operation, such as cooling system failures or sudden partial discharges. Therefore, the more significant the trend and residual variations in the data distribution characteristics of the physical quantity signals of transformer operating status, the better they reflect the fault type characteristics of the transformer operation, and the more helpful it is in identifying the contribution of each physical quantity signal to the classification of transformer equipment fault types.

[0029] Furthermore, to measure the fault type characteristics of transformer operation, the degree of anomalous trend of each trend component in the Hankel matrix is ​​calculated: , Let be the degree of abnormal trend of the i-th trend component. and Let i be the maximum and minimum values ​​of the i-th trend component. To preset the first parameter and avoid the denominator being 0, the value range is 0.01-0.1, and in this embodiment, the value is 0.1.

[0030] Among them, the anomaly trend degree can reflect the cumulative effect of long-term operation of transformers. The greater the anomaly trend degree, the more significant the fault types with cumulative fault characteristics in the long-term operation of transformers are, which helps to identify the contribution of each physical quantity signal to the classification of transformer equipment fault types.

[0031] Simultaneously, the permutation entropy of each residual component is calculated. Permutation entropy reflects the magnitude of random fluctuation characteristics within each residual component. The dispersion of the permutation entropy of all residual components is also calculated. Dispersion can be variance, standard deviation, or coefficient of variation, reflecting the complexity of random fluctuations within each residual component. Therefore, considering both the average level and dispersion of random fluctuations, the random fault characteristics of each physical quantity signal are measured. The product of the mean of the permutation entropy of all residual components and the dispersion is denoted as the anomalous randomness of the row vector of each physical quantity signal. Anomalous randomness reflects random fault events during transformer operation. A higher anomalous randomness indicates a higher degree of significance of random fault events during transformer operation, which also helps to identify the contribution of each physical quantity signal to the classification of transformer equipment fault types.

[0032] Based on the above analysis, and combining the anomaly trend degree and anomaly randomness degree, the anomaly identification degree of each physical quantity signal row vector in the multimodal feature matrix is ​​calculated: In the formula, Let be the anomaly detection score of the row vector of the m-th physical quantity signal. The maximum value of the abnormal trend degree of all trend components in the Hankel matrix constructed for the row vector of the m-th physical quantity signal. Let be the degree of anomalous randomness of the row vector of the m-th physical quantity signal. It is an exponential function with the natural constant as its base. This is a function that takes the minimum value.

[0033] This embodiment is based on existing feature measurement methods. Since the abnormal trend degree and abnormal randomness can reflect the characteristics of different types of abnormal faults during transformer operation, in order to effectively identify the types of abnormal faults in the long-term operation of transformers, the formula uses a ratio to compare the exponentially normalized values ​​of the abnormal trend degree and abnormal randomness. The greater the difference, the more singular the data distribution characteristics of the physical quantity signal, which helps to accurately distinguish the types of transformer faults in the future.

[0034] Among them, the anomaly identification degree reflects the effectiveness of this physical quantity in identifying transformer fault types. The higher the anomaly identification degree, the higher the effectiveness of this physical quantity in identifying transformer fault types. Therefore, this feature physical quantity contributes more to the subsequent use of weighted Naive Bayes to distinguish fault types, which is beneficial to improving the accuracy of transformer fault type diagnosis.

[0035] Step 3: Based on the anomaly recognition degree and combined with the abnormal change characteristics of each physical quantity signal, construct the compensation adjustment coefficient, and correct and compensate the information gain based on the compensation adjustment coefficient to obtain the information gain correction value of each physical quantity signal row vector in the multimodal feature matrix.

[0036] Meanwhile, if there are many abrupt change points in the residual components of the Hankel matrix constructed from the row vectors of a certain physical quantity signal, indicating a significant sudden anomaly in that physical quantity signal during transformer operation, then in the process of classifying transformer equipment fault types, the greater the difference in sudden anomaly between that physical quantity signal and the other physical quantity signals, the more helpful that physical quantity signal is in reducing the uncertainty in distinguishing transformer fault types. Therefore, when using a weighted Naive Bayes model to diagnose transformer fault types, in order to accurately correct and compensate for the weight coefficients determined by the information gain method, it is necessary to combine the sudden anomalies of different physical quantity signals during transformer operation.

[0037] Furthermore, to analyze the sudden anomalies of different physical quantity signals during transformer operation, the absolute values ​​of each residual component are taken to form an energy fluctuation sequence. This energy fluctuation sequence is used as the input to the Bernaola-Galvan segmentation algorithm, with a significance level of 0.05 and a minimum segment length of 30. The Bernaola-Galvan segmentation algorithm is used to obtain the number of abrupt change points in each residual component, and the ratio of the number of abrupt change points in each residual component to the total number is calculated. The mean of the ratios corresponding to abrupt change points in all residual components is denoted as the abrupt anomaly of each physical quantity signal row vector in the multimodal feature matrix. This reflects the sudden anomalies of physical quantity signals during transformer operation. The greater the difference in sudden anomalies between a certain physical quantity signal and the other physical quantity signals, the more helpful the physical quantity signal is in reducing the uncertainty of distinguishing transformer fault types, and the more necessary it is to increase the weight coefficient of this feature physical quantity in the weighted Naive Bayes model.

[0038] Therefore, based on the above analysis, the compensation adjustment coefficient of each physical quantity signal row vector in the multimodal characteristic matrix is ​​calculated: In the formula, Let be the compensation adjustment coefficient for the row vector of the m-th physical quantity signal. The preset weighting coefficient ranges from 0.4 to 0.6; in this embodiment, it is set to 0.5. The sum of the absolute differences between the abnormal bursts of the m-th physical quantity signal row vector and all other physical quantity signal row vectors is given. norm() is the maximum and minimum value normalization function. The specific normalization process is existing technology and will not be described in detail in this embodiment.

[0039] Understandably, the greater the anomaly identification rate, the more effective the physical quantity is in identifying transformer fault types; and the greater the sum of differences in anomalous sudden changes, the more the physical quantity helps to reduce the uncertainty in distinguishing transformer fault types. Therefore, the formula uses a weighted summation method to fuse the features of the sum of differences in anomaly identification rate and anomalous sudden changes, so as to more accurately measure the compensation adjustment coefficient.

[0040] The compensation adjustment coefficient reflects the strength of the correction and compensation for the weight coefficients determined by the information gain method. The larger the compensation adjustment coefficient, the stronger the correction and compensation for the weight coefficients determined by the information gain method. This more accurately reflects the contribution of multimodal characteristic parameters to the classification of transformer equipment fault types and improves the accuracy of transformer fault type diagnosis.

[0041] Therefore, in order to accurately diagnose transformer fault types using a weighted Naive Bayes model, the set of transformer fault types includes core loosening, winding deformation, poor contact resistance, sudden partial discharge, floating discharge, insulation aging, core overheating, clamp overheating, and cooling system failure. Based on a pre-established historical multimodal feature sample library of transformers containing known fault types, including data samples of vibration signals, temperature signals, sound signals, partial discharge signals, grounding current signals, and DC polarization signals, the basic information gain of each physical quantity signal row vector is pre-calculated offline using an information gain algorithm. Basic information gain The larger the value, the greater the contribution of the row vector representing the physical quantity signal to the differentiation of transformer fault types.

[0042] Furthermore, based on the compensation adjustment coefficient, the weight coefficients determined by the information gain method are corrected and compensated, and the information gain correction value of each physical quantity signal row vector in the multimodal feature matrix is ​​calculated: In the formula, This is the information gain correction value for the m-th physical quantity signal row vector.

[0043] Therefore, a compensation adjustment coefficient is used to correct and compensate the weighting coefficients determined by the information gain method. The larger the compensation adjustment coefficient, the greater the correction and compensation force for the weighting coefficients determined by the information gain method, resulting in a larger corrected information gain. This more accurately reflects the contribution of multimodal feature parameters to the classification of transformer equipment fault types, improving the accuracy of transformer fault type diagnosis. By introducing a compensation adjustment coefficient based on real-time short-term window calculation, the basic information gain calculated offline based on historical global samples is dynamically fine-tuned, achieving feature fusion between the static global baseline and real-time local fluctuations.

[0044] Step 4: Use the information gain correction value to determine the weight coefficients of the multimodal eigenvectors in the weighted Naive Bayes model, and use the weighted Naive Bayes model to identify the fault type of the transformer.

[0045] Furthermore, to diagnose transformer equipment fault types, the RMS effective value, variance, mean absolute deviation, energy entropy, kurtosis, and skewness of each physical quantity signal row vector in the multimodal feature matrix of the transformer within the current sliding time window are calculated. These values ​​constitute a preprocessed feature vector for each physical quantity signal. This preprocessed feature vector is then used as the transformer's multimodal feature vector and input into a pre-trained WNB weighted Naive Bayes model. The normalized result of the information gain correction value of each physical quantity signal row vector in the multimodal feature matrix is ​​used as the weight coefficient of the multimodal feature vector. The WNB weighted Naive Bayes model is used to obtain the posterior probability of each fault type in the transformer's fault type set. The fault type corresponding to the maximum posterior probability is the fault type that occurred in the transformer within the current sliding time window. It should be noted that, preferably, the maximum value normalization method is used for normalization in this embodiment. The specific maximum value normalization process is existing technology and will not be described in detail in this embodiment.

[0046] It is understood that references to "one embodiment" or "some embodiments" in this specification mean that one or more embodiments of this application include the specific features, structures, or characteristics described in connection with that embodiment. Therefore, the appearance of phrases such as "in one embodiment," "in some embodiments," "in other embodiments," or "in still other embodiments" in different parts of this specification does not necessarily refer to the same embodiment, but rather means "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof all mean "including but not limited to," unless otherwise specifically emphasized.

[0047] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous. Moreover, the sequence numbers of the steps in the embodiments do not imply a specific order of execution; the execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments in this specification.

[0048] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A transformer multimodal fault diagnosis method based on weighted Naive Bayes, characterized in that, Includes the following steps: Data of various physical quantity signals during transformer operation are acquired to construct multimodal feature matrices for each preset sliding time window, where each row vector corresponds to a physical quantity signal; The trend component and residual component of the multimodal feature matrix of each sliding time window are extracted, and the anomaly identification degree of each physical quantity signal row vector in the multimodal feature matrix is ​​obtained by the change of the trend component and the degree of random fluctuation of the residual component. Based on the anomaly recognition rate and combined with the anomalous suddenness of each physical quantity signal, a compensation adjustment coefficient is constructed. Based on the compensation adjustment coefficient, the information gain is corrected and compensated to obtain the information gain correction value of each physical quantity signal row vector in the multimodal feature matrix. The weight coefficients of the multimodal eigenvectors in the weighted naive Bayes model are determined by the information gain correction value, and the fault types of transformers are identified by the weighted naive Bayes model. The anomaly detection rate of each physical quantity signal row vector in the multimodal feature matrix is ​​obtained as follows: In the formula, Let be the anomaly detection score of the row vector of the m-th physical quantity signal. The maximum value of the abnormal trend degree of all trend components in the Hankel matrix constructed for the row vector of the m-th physical quantity signal. Let be the degree of anomalous randomness of the row vector of the m-th physical quantity signal. It is an exponential function with the natural constant as its base. To find the minimum value function, This is the preset first parameter; Before constructing the compensation adjustment coefficient, the absolute value of each residual component is taken to form an energy fluctuation sequence. The energy fluctuation sequence is used as the input of the Bernaola Galvan segmentation algorithm to obtain the number of mutation points in each residual component and calculate the ratio of the number of mutation points in each residual component to the total number. The mean of the ratios calculated for all residual components is used as the anomalous burstiness of each physical quantity signal row vector in the multimodal feature matrix. The compensation adjustment coefficient is obtained as follows: In the formula, The compensation adjustment coefficient is the row vector of the m-th physical quantity signal. Let be the anomaly detection score of the row vector of the m-th physical quantity signal. For preset weighting coefficients, It is the sum of the absolute differences between the abnormal bursts of the m-th physical quantity signal row vector and the abnormal bursts of all other physical quantity signal row vectors, and norm() is the maximum and minimum value normalization function. The information gain correction value for each physical quantity signal row vector in the multimodal feature matrix is ​​obtained as follows: In the formula, This is the information gain correction value for the row vector of the m-th physical quantity signal. Let be the compensation adjustment coefficient for the row vector of the m-th physical quantity signal. The basic information gain of the m-th physical quantity signal row vector is obtained offline through the information gain algorithm.

2. The transformer multi-mode fault diagnosis method based on weighted Naive Bayes as described in claim 1, characterized in that, The singular spectrum analysis algorithm is used to construct the row vector of each physical quantity signal into a Hankel matrix, and singular spectrum analysis is performed on the Hankel matrix to obtain each trend component and each residual component in the Hankel matrix.

3. The transformer multi-mode fault diagnosis method based on weighted Naive Bayes as described in claim 2, characterized in that, Before obtaining the anomaly detection score of each physical quantity signal row vector, calculate the anomaly trend score of each trend component: , Let be the degree of abnormal trend of the i-th trend component. and These are the maximum and minimum values ​​in the i-th trend component.

4. The transformer multimode fault diagnosis method based on weighted Naive Bayes as described in claim 3, characterized in that, Before obtaining the anomaly identification degree of each physical quantity signal row vector, the product of the mean of the permutation entropy of all residual components and the dispersion of the permutation entropy of all residual components is calculated as the anomaly randomness of each physical quantity signal row vector.

5. The transformer multi-mode fault diagnosis method based on weighted Naive Bayes as described in claim 1, characterized in that, The vector composed of the RMS effective value, variance, mean absolute deviation, energy entropy, kurtosis, and skewness of each physical quantity signal row vector is used as the multimode characteristic vector of the transformer.

6. The transformer multi-mode fault diagnosis method based on weighted Naive Bayes as described in claim 1, characterized in that, In the weighted Naive Bayes model, the weight coefficients of the multimodal feature vectors are the normalized results of the information gain correction values ​​of each physical quantity signal row vector in the multimodal feature matrix.

Images