Dynamic adaptive based multi-modal fault diagnosis method for electrical equipment

Abstract

The present application relates to the technical field of electrical equipment fault diagnosis, and discloses a dynamic self-adaptive electrical equipment multi-modal fault diagnosis method, which comprises the following steps: obtaining original multi-modal signal streams containing vibration, temperature and current signals, inputting a dynamic self-adaptive diagnosis network, obtaining a fault feature matching result set and determining each modal result. The network is trained by historical fault data, the data contains time domain, frequency domain and fusion feature parameters extracted from continuous multi-modal signals, and corresponding fault type labels and confidence coefficients. The network contains cross-modal feature fusion and space-time feature coding modules, and triggers a three-level early warning when the confidence coefficients of at least two modal results exceed a dynamic threshold. The method improves the comprehensiveness, accuracy and real-time performance of diagnosis, and is suitable for electrical equipment fault diagnosis.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of electrical equipment fault diagnosis, in particular to an electrical equipment multi-modal fault diagnosis method based on dynamic self-adaptation. BACKGROUND

[0002] In modern power systems, the safe and stable operation of electrical equipment is crucial to ensuring power supply reliability and normal operation of the power system. With the development of power equipment towards high voltage, large capacity and intelligentization, its operating conditions are becoming increasingly complex, and the types of faults are also showing diversification and complexity. Traditional electrical equipment fault diagnosis methods usually only analyze single modal signals (such as vibration signals, temperature signals or current signals, etc.), which are difficult to comprehensively and accurately reflect the actual operating state of the equipment, and have problems such as low diagnosis accuracy, high rate of missed or misdiagnosis, etc.

[0003] On the one hand, the fault information provided by a single modal signal is limited. For example, vibration signals mainly reflect the operating state of mechanical components of the equipment and are sensitive to mechanical faults (such as bearing wear, gear box failure, etc.), but are relatively lagging or not obvious for electrical faults (such as winding overheating, partial discharge, etc.); temperature signals can directly reflect local overheating and other thermal faults of the equipment, but may not be able to capture subtle changes in temperature in a timely manner for some early mechanical or electrical faults; current signals are mainly related to electrical parameters and load conditions of the equipment and can be used to analyze faults in the electrical circuit, but have relatively weak ability for mechanical structure fault diagnosis. Therefore, relying solely on single modal signals for fault diagnosis can easily lead to misjudgment or missed judgment of complex faults.

[0004] On the other hand, most traditional diagnosis methods use fixed diagnosis models and parameters, lacking the ability to adapt to dynamic changes in the operating state of the equipment. The signal characteristics of electrical equipment change under different operating stages and different load conditions, and traditional methods cannot automatically adjust the diagnosis model according to real-time signal characteristics, resulting in a decline in diagnosis performance as the operating state of the equipment changes. For example, during the transition from normal operation to abnormal operation of the equipment, small changes in signal characteristics may not be effectively identified by a fixed model, thus delaying the fault diagnosis opportunity.

[0005] In addition, although existing multi-modal fault diagnosis methods integrate multiple signals, they mostly use simple feature splicing or fixed weight fusion methods, without fully considering the importance differences of different modal signals under different fault types. For example, in some fault types, vibration signals may carry more critical fault information, while in other fault types, temperature signals or current signals may be more important. Simple fusion methods cannot dynamically adjust the weights of each modal signal, resulting in the inability of the fused features to fully exploit the advantages of multi-modal signals, affecting the diagnosis accuracy.

[0006] Meanwhile, with the development of the intelligentization and digitization of power systems, a large amount of multi-modal data is generated in the operation process of electrical equipment, and the traditional diagnosis method is inefficient in data processing and feature extraction, which is difficult to meet the demand of real-time fault diagnosis. How to quickly and accurately extract effective fault features from a large amount of multi-modal data and construct a diagnosis model with dynamic adaptive ability has become a key problem to be solved in the field of electrical equipment fault diagnosis. SUMMARY

[0007] The purpose of the present application is to provide a multi-modal fault diagnosis method for electrical equipment based on dynamic adaptation to solve the problems raised in the background art.

[0008] To achieve the above purpose, the present application provides the following technical solution: a multi-modal fault diagnosis method for electrical equipment based on dynamic adaptation, the method comprising:

[0009] The method is applied to an electrical equipment fault diagnosis system, and the method comprises:

[0010] Obtaining an original multi-modal signal stream to be subjected to fault diagnosis, the original multi-modal signal stream comprising vibration signals, temperature signals and current signals;

[0011] Inputting the original multi-modal signal stream into a pre-trained dynamic adaptive diagnosis network to obtain a fault feature matching result set;

[0012] Determining, based on the fault feature matching result set, a fault feature matching result corresponding to each modal signal in the original multi-modal signal stream;

[0013] The dynamic adaptive diagnosis network is a parameterized network obtained by inputting historical fault data into an initial diagnosis network to be trained, the historical fault data comprising a batch of historical feature parameters extracted from a batch of historical multi-modal signals and a fault type label and a confidence coefficient corresponding to each historical feature parameter, the fault type label reflecting an abnormal state identifier of a historical feature parameter, the confidence coefficient reflecting a correction factor corresponding to the abnormal state identifier, the batch of historical multi-modal signals being a batch of continuously collected electrical equipment operation signals, and each historical feature parameter comprising a feature parameter determined by a historical multi-modal signal and an adjacent historical multi-modal signal of the historical multi-modal signal;

[0014] The fault feature matching result is used to represent a fault type identifier of the multi-modal signal stream.

[0015] Preferably, the method further comprises:

[0016] extracting time domain feature parameters from the batch of historical multi-modal signals to obtain a first group of feature parameters, wherein one feature parameter in the first group of feature parameters corresponds to one historical multi-modal signal in the batch of historical multi-modal signals;

[0017] extracting frequency domain feature parameters from the batch of historical multi-modal signals to obtain a second group of feature parameters, wherein a first feature parameter in the second group of feature parameters corresponds to a first historical multi-modal signal in the batch of historical multi-modal signals, an adjacent historical multi-modal signal of the first historical multi-modal signal is a second historical multi-modal signal, and the first feature parameter represents a frequency domain correlation feature between the first historical multi-modal signal and the second historical multi-modal signal;

[0018] fusing the first group of feature parameters and the second group of feature parameters respectively to obtain the batch of historical feature parameters.

[0019] Preferably, the extracting frequency domain feature parameters from the batch of historical multi-modal signals to obtain a second group of feature parameters comprises:

[0020] The frequency domain feature parameters are extracted from the batch of historical multi-modal signals by the following steps, wherein each time the historical multi-modal signal subjected to the frequency domain feature parameter extraction is regarded as a current historical multi-modal signal, the obtained feature parameter is regarded as a current feature parameter, and the second group of feature parameters comprises the current feature parameter;

[0021] performing a spectrum decomposition operation on the current historical multi-modal signal to determine a first group of frequency domain feature quantities, wherein the first group of frequency domain feature quantities are used to describe the main frequency band energy distribution in the current historical multi-modal signal;

[0022] performing a spectrum decomposition operation on an adjacent historical multi-modal signal of the current historical multi-modal signal to determine a second group of frequency domain feature quantities, wherein the second group of frequency domain feature quantities are used to describe the main frequency band energy distribution in the adjacent historical multi-modal signal;

[0023] determining the current feature parameter based on the first group of frequency domain feature quantities and the second group of frequency domain feature quantities, wherein the current feature parameter is used to represent the energy deviation amount of the corresponding frequency band in the first group of frequency domain feature quantities and the second group of frequency domain feature quantities.

[0024] Preferably, the method further comprises:

[0025] The fault type label and the confidence coefficient corresponding to each historical feature parameter in the batch of historical feature parameters are determined respectively by the following steps, wherein each time the historical feature parameter for determining the fault type label and the confidence coefficient is taken as a current historical feature parameter, and the fault type label corresponding to the current historical feature parameter is taken as a current fault type label and a current confidence coefficient:

[0026] The current historical feature parameter is input into the initial diagnosis network to obtain the current fault type label, wherein the initial diagnosis network is a parameterized network obtained by pre-executing initialization processing;

[0027] The current confidence coefficient is determined based on the current fault type label and a benchmark verification identifier corresponding to the current historical feature parameter, wherein the benchmark verification identifier carries a calibration label in a historical multi-modal signal corresponding to the current historical feature parameter, the calibration label includes a target fault type identifier of the multi-modal signal stream, and the confidence coefficient is used to indicate whether the current fault type label is consistent with the target fault type identifier.

[0028] Preferably, the step of inputting the current historical feature parameter into the initial diagnosis network to obtain the current fault type label comprises:

[0029] Each fault type identifier in a pre-stored fault type identifier set is input into a first classification algorithm branch in sequence together with the current historical feature parameter to obtain a first group of classification confidence indicators, wherein the first group of classification confidence indicators includes a confidence indicator corresponding to each fault type identifier;

[0030] A candidate fault type identifier corresponding to a first confidence indicator with the highest value in the first group of classification confidence indicators is determined as a current fault type identifier corresponding to the current fault type label.

[0031] Preferably, before the step of inputting each fault type identifier in the pre-stored fault type identifier set into the first classification algorithm branch in sequence together with the current historical feature parameter to obtain the first group of classification confidence indicators, the method further comprises:

[0032] A signal stability indicator is determined by a preset mode;

[0033] It is determined whether the signal stability indicator meets a set stability threshold;

[0034] On the basis that the signal stability indicator meets the set stability threshold, a candidate fault type identifier in the pre-stored fault type identifier set is dynamically selected as the current fault type label;

[0035] input the current historical feature parameter into the initial diagnosis network based on that the signal stability index does not meet the set stability threshold, to obtain the current fault type label.

[0036] Preferably, the candidate fault type identification corresponding to the first confidence index with the highest value in the first group of classification confidence indexes is determined as the current fault type identification corresponding to the current fault type label, comprising:

[0037] On the basis that the candidate fault type identification corresponding to the first confidence index is same as the target fault type identification, the current confidence coefficient is determined as a first confidence coefficient, wherein the first confidence coefficient represents that the initial diagnosis network successfully generates a matching identification;

[0038] On the basis that the candidate fault type identification corresponding to the first confidence index is different from the target fault type identification, the current confidence coefficient is determined as a second confidence coefficient, wherein the second confidence coefficient represents that the initial diagnosis network does not successfully generate a matching identification.

[0039] Preferably, the method further comprises:

[0040] screening a plurality of groups of historical joint fault records from the historical fault data, wherein the kth group of historical joint fault records in the plurality of groups of historical joint fault records comprises a kth historical feature parameter, a kth fault type label corresponding to the kth historical feature parameter, a kth confidence coefficient and a k+1th historical feature parameter, k is a positive integer;

[0041] training an initial diagnosis network to be trained based on the plurality of groups of historical joint fault records to obtain the dynamic adaptive diagnosis network, wherein on the basis that the number of iterations of training the initial diagnosis network reaches a set training threshold, the initial diagnosis network is determined as the dynamic adaptive diagnosis network, on the basis that the number of iterations of training the initial diagnosis network does not reach the set training threshold, the network parameters of the initial diagnosis network are updated according to a preset error function, and the input of each training process is a group of historical joint fault records in the plurality of groups of historical joint fault records.

[0042] Preferably, the dynamic adaptive diagnosis network comprises:

[0043] a cross-modal feature fusion module that dynamically calculates modal weight coefficients of vibration signals, temperature signals and current signals using an attention mechanism;

[0044] a space-time feature encoder that encodes time series features using a bidirectional LSTM structure;

[0045] The multi-granularity classifier comprises parallel convolutional neural network branches and full connection classification branches, wherein each branch output is fused through an adaptive weighting strategy to generate a final fault type identification.

[0046] The online updating module automatically triggers the incremental learning training of the network parameters when it is detected that the confidence coefficient of consecutive N signal samples is lower than a preset threshold.

[0047] Preferably, the method further comprises:

[0048] When the confidence coefficients of the matching results of the fault features of at least two modalities of the vibration signal, the temperature signal and the current signal simultaneously exceed the dynamic threshold, a three-level early warning mechanism is triggered, wherein the first-level early warning displays a yellow warning mark through a device operation interface and generates a diagnosis log, the second-level early warning superimposes an audible and visual alarm and automatically locks the device operation parameters, and the third-level early warning synchronously pushes an encrypted alarm message to a remote monitoring platform.

[0049] The dynamic threshold is dynamically adjusted according to the confidence degree distribution law of the same type of faults in the historical operation data of the device, and the threshold adjustment step of different modal signals is negatively correlated with the signal sampling frequency thereof.

[0050] Compared with the prior art, the present application has the following beneficial effects:

[0051] In the aspect of multi-modal signal processing, the time domain and frequency domain features of historical multi-modal signals are extracted and fused, so that the time and space features in the signals can be comprehensively captured. The time domain feature extraction can directly reflect the instantaneous characteristics and statistical characteristics of the signals, such as mean value, variance, peak value, etc., and effectively identify abnormal fluctuations of the signals. The frequency domain feature extraction analyzes the frequency components and energy distribution of the signals through spectrum decomposition operation, and can detect the frequency component changes caused by faults and the frequency domain correlation features between adjacent signals. For example, by calculating the energy deviation amount of the main frequency band of the current signal and the adjacent signals, the frequency feature change trend in the early stage of the fault can be sensitively captured, thereby providing a basis for early warning of the fault. The fusion of the time domain and frequency domain features can make full use of the complementarity of different domain features, form more comprehensive and rich fault feature descriptions, overcome the limitations of single domain features, and improve the feature representation ability for fault types.

[0052] The construction and training mechanism of the dynamic adaptive diagnostic network has significant advantages. The network uses a cross-modal feature fusion module, which dynamically calculates the weight coefficients of each modal signal using an attention mechanism. The importance of vibration, temperature, and current signals can be automatically adjusted according to different fault types. For example, in mechanical fault diagnosis, the weight coefficient of the vibration signal automatically increases, making it dominant in feature fusion, thereby more accurately identifying mechanical faults. In electrical fault diagnosis, the weight coefficient of the current signal increases, highlighting the key role of the current signal and improving the diagnostic accuracy of electrical faults. The spatio-temporal feature encoder uses a bidirectional LSTM structure, which can effectively capture the temporal dependence in time series signals and deeply encode the temporal features of the signals, making it suitable for analyzing electrical equipment operation signals with dynamic characteristics. The multi-granularity classifier contains parallel convolutional neural network branches and fully connected classification branches. Each branch can classify features at different granularities, and the branch outputs are fused through an adaptive weighting strategy to improve the reliability and robustness of the classification results. The online update module can automatically trigger incremental learning and training of network parameters when the confidence coefficients of consecutive N signal samples are detected to be lower than the preset threshold. This allows the network to update model parameters in real time based on new fault data, adapt to changes in device operating conditions, and avoid performance degradation due to device aging, environmental changes, and other factors.

[0053] The confidence coefficient and early warning mechanism during fault diagnosis further enhance the reliability and practicality of the diagnosis. By comparing the current fault type label with the reference verification label (carrying the calibration label), the confidence coefficient can be determined, which can quantitatively evaluate the reliability of the diagnostic results. When the confidence coefficients of the matching results of at least two modal fault features simultaneously exceed the dynamic threshold, a three-level early warning mechanism is triggered, which can take different measures according to the severity of the fault. Level one warning displays a yellow warning symbol on the device operation interface and generates a diagnostic log, enabling timely monitoring and recording of early-stage faults. Level two warning adds an audible and visual alarm and automatically locks the device operating parameters, allowing maintenance personnel to quickly identify and handle faults on site. Level three warning sends an encrypted alarm message to the remote monitoring platform, ensuring that remote rapid response and collaborative handling can be achieved in the case of serious faults. The dynamic threshold is dynamically adjusted based on the confidence distribution of similar faults in historical device operation data, and the threshold adjustment step of different modal signals is negatively related to their signal sampling frequency, allowing the early warning mechanism to adapt to changes in fault features of different devices and different operating stages, improving the accuracy and timeliness of early warning. BRIEF DESCRIPTION OF DRAWINGS

[0054] Figure 1 A working principle diagram of the dynamic adaptive electrical equipment multi-modal fault diagnosis method described in the present application;

[0055] Figure 2 A flowchart for historical feature parameter extraction;

[0056] Figure 3 Flow chart for frequency domain feature parameter extraction;

[0057] Figure 4 Flow chart for fault type label and confidence coefficient determination. DETAILED DESCRIPTION

[0058] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0059] Please refer to Figures 1-4 The dynamic adaptive-based electrical equipment multi-modal fault diagnosis method disclosed by the present application comprises the following specific implementation steps:

[0060] An original multi-modal signal stream to be subjected to fault diagnosis is acquired, and the original multi-modal signal stream comprises vibration signals, temperature signals and current signals. A sensor array arranged at a key part of an electrical equipment continuously acquires multi-dimensional real-time signals during operation of the equipment, wherein the vibration signals are acquired by an acceleration sensor, the temperature signals are acquired by a thermocouple or an infrared temperature measurement module, and the current signals are extracted from an electrical circuit by a current transformer. All the signals are synchronously acquired at a preset sampling frequency (such as 10 kHz) and transmitted to a signal preprocessing unit of a diagnosis system.

[0061] The original multi-modal signal stream is input into a pre-trained dynamic adaptive diagnosis network to obtain a fault feature matching result set. After the original signals are subjected to preliminary processing such as denoising and normalization by the signal preprocessing unit, the continuous signal stream is formed according to a time sequence and input into the dynamic adaptive diagnosis network. The fault features in the multi-modal signals are extracted, fused and matched by collaborative operation of components such as a cross-modal feature fusion module, a space-time feature encoder and a multi-granularity classifier in the network, and finally a result set containing the matching possibility of the fault features of each modal signal is output.

[0062] Based on the fault feature matching result set, a fault feature matching result corresponding to each modal signal in the original multi-modal signal stream is determined. According to the feature parameter distribution of each modal signal in the fault feature matching result set, the diagnosis system separates out the fault feature matching results corresponding to the vibration signals, the temperature signals and the current signals respectively, and each result contains a fault type identification and a corresponding confidence coefficient, thereby providing a basis for subsequent fault positioning and early warning.

[0063] The dynamic self-adaptive diagnosis network is a parameterized network obtained by training an initial diagnosis network to be trained by inputting historical fault data. The construction process of the historical fault data is as follows: first, a batch of continuous electrical equipment operation signals are collected as historical multi-modal signals, for each historical multi-modal signal, corresponding feature parameters are extracted in combination with its adjacent historical multi-modal signals to form a batch of historical feature parameters; at the same time, a corresponding fault type label and a confidence coefficient are labeled for each historical feature parameter, wherein the fault type label reflects the abnormal state identifier of the feature parameter, and the confidence coefficient reflects the correction factor corresponding to the abnormal state identifier.

[0064] The fault feature matching result is used to represent the fault type identifier of the multi-modal signal stream, and the classification and matching operation of the network is used to realize accurate identification of the fault type of the input signal.

[0065] The technical solutions of the present application will be further described in detail below in combination with specific embodiments.

[0066] Embodiment 1:

[0067] On the basis of the overall implementation scheme, the extraction process of the historical feature parameters is described in detail. When a batch of historical multi-modal signals are subjected to time domain feature parameter extraction, a plurality of statistical features are calculated in the time domain for each historical multi-modal signal (such as a combination of vibration, temperature and current signals collected at a certain time) to form a first group of feature parameters. Taking the vibration signal as an example, the mean value thereof needs to be calculated, which reflects the average energy level of the signal in the time domain and can be obtained by summing the values of all sampling points of the vibration signal and then dividing by the number of sampling points; the variance is calculated to reflect the fluctuation degree of the signal, and the greater the variance, the more dispersed the signal energy distribution and the more violent the fluctuation, and the calculation method is to square the average value of the difference between each sampling point value and the mean value; the peak value parameter can indicate whether there is an impact vibration, which is determined by identifying the maximum instantaneous value in the signal; the kurtosis is used to measure the sharpness of the signal amplitude distribution, and the skewness is used to describe the symmetry of the signal amplitude distribution, and the two features can be calculated by the corresponding probability statistical formula. For the temperature signal, the time domain feature parameters such as the mean value can reflect the current average temperature level of the device, and the variance can analyze the stationarity of the temperature variation, and if the variance is small, it means that the temperature fluctuation is small and the device running temperature is stable. The time domain feature parameters of the current signal such as the mean value can be used to judge the average size of the current, and the variance can be used to analyze the distortion of the current waveform, and if the variance abnormally increases, it may mean that the current waveform has irregular changes, indicating that the device may have an electrical fault. Each historical multi-modal signal corresponds to a feature parameter composed of statistical features of each modality in the time domain, and all such feature parameters of the historical multi-modal signals together constitute the first group of feature parameters.

[0068] When a batch of historical multi-modal signals are subjected to frequency domain feature parameter extraction to obtain a second set of feature parameters, the historical multi-modal signal subjected to each frequency domain feature parameter extraction is defined as a current historical multi-modal signal, and the corresponding feature parameter is a current feature parameter. In specific operations, first, a spectrum decomposition operation is performed on the current historical multi-modal signal, a time domain signal is converted into a frequency domain signal through fast Fourier transform (FFT), and then a first set of frequency domain feature quantities are determined, which are used to describe the energy distribution of the main frequency band in the current historical multi-modal signal. Taking a vibration signal as an example, the main frequency components and their energy proportions where the vibration energy is concentrated can be identified through spectrum decomposition. If there is a significant increase in energy at a certain specific frequency (such as a bearing fault characteristic frequency), it may indicate the presence of a corresponding mechanical fault. For temperature signals and current signals, frequency domain feature quantities can be used to analyze whether the energy distribution of the signals at different frequencies is abnormal. For example, if the energy of high-order harmonics other than the power frequency in the current signal increases, it may indicate the presence of harmonic interference or internal component failure. Next, the same spectrum decomposition operation is performed on the adjacent historical multi-modal signals (such as the signals at the previous or next time) of the current historical multi-modal signal, and a second set of frequency domain feature quantities are determined, which are used to describe the energy distribution of the main frequency band in the adjacent signals. Then, based on the first set of frequency domain feature quantities and the second set of frequency domain feature quantities, the energy deviation quantity of the corresponding frequency band is calculated, which is the current feature parameter. For example, if the energy of the main frequency band (such as 100-200 Hz) of the current vibration signal is E1, and the energy of the same frequency band of the adjacent vibration signal is E2, the energy deviation quantity can be represented as |E1-E2| / E1, which is used to reflect the change trend of the adjacent signal in the frequency domain. If the deviation quantity is large, it may mean that the fault has developed significantly in a short period of time and needs to be paid attention to. All current feature parameters constitute the second set of feature parameters.

[0069] After the extraction of the time domain feature parameters and the frequency domain feature parameters is completed, the first group of feature parameters and the second group of feature parameters need to be fused respectively to obtain a batch of historical feature parameters. The fusion process adopts a feature-level fusion strategy, that is, the time domain features and the frequency domain features of the same mode (vibration, temperature, and current) are combined according to the mode to form a feature vector containing multi-dimensional information. Taking the vibration mode as an example, the mean value, variance, peak value, kurtosis, skewness, and other features in the time domain are combined with the energy distribution features of the main frequency band and the energy deviation amount features of adjacent time points in the frequency domain to form complete feature parameters of the mode. The feature parameters contain the statistical characteristics of the vibration signal in the time domain and the energy distribution and change trend information in the frequency domain. Similarly, the same feature fusion operation is performed on the temperature mode and the current mode to combine the respective time domain features and frequency domain features into complete mode feature parameters. Finally, after all the historical multi-modal signals are processed, a batch of historical feature parameters containing the time-space features of each mode are obtained. These feature parameters can comprehensively and deeply reflect the state information of the electrical equipment during operation, provide rich and effective input data for the training of the subsequent dynamic self-adaptive diagnosis network, and enable the network to learn the feature expression and evolution law of different fault types in the multi-modal signals.

[0070] Embodiment 2

[0071] On the basis of completing the extraction of the historical feature parameters in Embodiment 1, this embodiment details the determination process of the fault type labels and the confidence coefficients corresponding to the historical feature parameters. The process realizes accurate labeling of the fault types and confidence evaluation by inputting and calculating the initial diagnosis network, in combination with the benchmark verification labels.

[0072] The historical feature parameters for determining the fault type labels and the confidence coefficients each time are defined as current historical feature parameters, and the fault type labels and the confidence coefficients corresponding thereto are respectively referred to as current fault type labels and current confidence coefficients. The initial diagnosis network is a parameterized network that is pre-processed by initialization. The structure of the network includes an input layer, a hidden layer, and an output layer. The neurons in the hidden layer perform signal transmission and calculation through preset weights and biases. When determining the current fault type labels, a multi-step classification algorithm input strategy is adopted.

[0073] In the first step, each fault type identifier in the pre-stored fault type identifier set is sequentially inputted into the first classification algorithm branch together with the current historical characteristic parameter. The pre-stored fault type identifier set covers various fault types that the electrical equipment can have, such as bearing wear, winding overheating, abnormal load, insulation aging, poor contact, etc. The first classification algorithm branch can use a classification algorithm such as support vector machine (SVM) or random forest to calculate the combination of the input characteristic parameter and the fault type identifier. Through the internal operation of the algorithm, a first set of classification confidence indicators is obtained, which includes the confidence indicators corresponding to each fault type identifier. These indicators reflect the possibility of the current historical characteristic parameter belonging to the fault type. For example, for the bearing wear fault type identifier, the algorithm can output a confidence indicator of 0.75, indicating that the current historical characteristic parameter has a 75% possibility of corresponding to the bearing wear fault.

[0074] In the second step, the candidate fault type identifier corresponding to the first confidence indicator with the highest value is selected from the first set of classification confidence indicators. Assuming that among all the confidence indicators corresponding to the fault type identifiers, the confidence indicator 0.85 corresponding to the bearing wear fault is the highest, the bearing wear is determined as the candidate fault type identifier, and further determined as the current fault type identifier corresponding to the current fault type label. This process ensures that the current fault type label is determined based on the maximum matching possibility between the historical characteristic parameter and the fault type.

[0075] Next, the current confidence coefficient is determined based on the current fault type label and the reference verification identifier corresponding to the current historical characteristic parameter. The reference verification identifier is a calibration label set in advance for the historical multi-modal signal corresponding to the current historical characteristic parameter. The label is determined by artificial diagnosis or other authoritative detection means, and contains the target fault type identifier of the multi-modal signal stream, i.e. the real fault type.

[0076] In determining the current confidence coefficient, the following judgment is made: if the candidate fault type identifier corresponding to the first confidence indicator is the same as the target fault type identifier in the reference verification identifier, it means that the initial diagnosis network has successfully generated an identifier matching the actual fault type. In this case, the current confidence coefficient is determined as the first confidence coefficient, which can be set to 1.0, indicating that the diagnosis result of the network is highly reliable. If they are different, it means that the initial diagnosis network has not successfully generated a matching identifier. In this case, the current confidence coefficient is determined as the second confidence coefficient, which can be set to 0.0, indicating that the diagnosis result of the network is unreliable. In this way, each historical characteristic parameter is labeled with a confidence coefficient, which is used to measure the accuracy of the initial diagnosis network in judging the fault type of the historical characteristic parameter, and provides an important reference for the training of the subsequent dynamic self-adaptive diagnosis network.

[0077] Before inputting the first classification algorithm branch, the signal stability needs to be judged to ensure the reliability of the input data and the accuracy of the diagnosis result. The signal stability index is determined by a preset mode, which can be designed based on the time domain statistical characteristics or frequency domain energy distribution of the signal. For example, the change amplitude of the time domain characteristic parameters (such as mean, variance) of the current historical multi-modal signal, or the fluctuation range of the energy of the main frequency band in the frequency domain, is calculated as an index to measure the signal stability.

[0078] Then, the determined signal stability index is compared with the set stability threshold. The stability threshold is set in advance according to the signal characteristic statistical law of the electrical equipment in normal operation, for example, the time domain variance change amplitude is not more than ±10% of the normal range, and the frequency domain main frequency band energy fluctuation is not more than ±5%.

[0079] If the signal stability index meets the set stability threshold, it means that the signal is in a stable state. In this case, the candidate fault type identifier in the pre-stored fault type identifier set is dynamically selected as the current fault type label. The dynamic selection strategy can prioritize the most likely fault type under the operating condition of the equipment according to factors such as the operating condition of the equipment and the historical fault frequency, thereby improving the diagnosis efficiency. For example, when the equipment is running at high load, the load-related fault type is given priority.

[0080] If the signal stability index does not meet the set stability threshold, it means that the signal may have abnormal fluctuations. At this time, the current historical characteristic parameters are input into the initial diagnosis network, and the current fault type label is obtained through the complex operation of the network to ensure accurate fault diagnosis even in the case of unstable signals. This process realizes adaptive judgment of signal stability, adopts different diagnosis strategies according to different situations, and improves the flexibility and reliability of the diagnosis system.

[0081] In the entire process of determining the fault type label and the confidence coefficient, multiple key steps and the cooperative work of algorithms are involved. First, the parameter setting of the initial diagnosis network has an important influence on the classification result, and the initial value of its weight and bias determines the initial response mode of the network to the input features. Second, the selection of the first classification algorithm branch is also crucial, as different classification algorithms (such as SVM, random forest) have different characteristics and application scenarios, and need to be selected reasonably according to the characteristics of the historical characteristic parameters and the distribution of the fault types. In addition, the accuracy of the benchmark verification identifier is directly related to the reliability of the confidence coefficient, so when setting the benchmark verification identifier, rigorous and accurate detection means need to be used to ensure that it can truly reflect the fault state of the equipment.

[0082] By the method of the embodiment, the corresponding fault type label and confidence coefficient are determined for each historical characteristic parameter, forming a complete historical fault data labeling system. These labeled data will be used as training samples to train the dynamic adaptive diagnosis network, so that the network can learn the mapping relationship between fault features and fault types, and how to adjust the diagnosis strategy according to the confidence of the characteristic parameters, thereby improving the accuracy and reliability of the multi-modal fault diagnosis of electrical equipment.

[0083] Embodiment 3:

[0084] Based on embodiment 2, this embodiment further illustrates the judgment and corresponding processing flow of signal stability before inputting the first classification algorithm branch. The flow determines the signal stability index through a preset mode, compares it with the set stability threshold, and adopts different fault type label determination strategies according to the comparison result, so as to realize adaptive processing of signal stability and improve the flexibility and reliability of the diagnosis system.

[0085] When determining the signal stability index through a preset mode, the preset mode is designed based on the time domain statistical characteristics or frequency energy distribution of the signal. For time domain statistical characteristics, the change amplitude of parameters such as mean, variance and peak value of the current historical multi-modal signal can be calculated. For example, for vibration signals, the mean value change rate in consecutive multiple sampling periods is calculated. If the mean value changes greatly in a short time, it may indicate that the signal has unstable factors. For temperature signals, the variance fluctuation is calculated. A sudden increase in variance may mean that the temperature has abnormal fluctuations and the signal stability decreases. For frequency energy distribution, the energy distribution change of the signal in different frequency bands can be analyzed. For example, after converting the signal to the frequency domain through fast Fourier transform, the energy proportion change of a specific frequency band (such as the fault feature frequency band) is calculated. If the energy proportion of a certain frequency band suddenly increases or decreases, it may indicate that the signal stability is affected. These parameter changes based on time domain and frequency domain together constitute the signal stability index, which is used to measure the stability degree of the current historical multi-modal signal.

[0086] After determining the signal stability index, it is compared with the set stability threshold. The stability threshold is set in advance according to the statistical rules of signal characteristics during normal operation of the electrical equipment. For the time domain parameter change amplitude threshold, for example, the mean value change rate of the vibration signal is set not to exceed ±5% of the normal range, and the variance fluctuation of the temperature signal is set not to exceed ±8% of the normal range. For the frequency energy distribution threshold, the energy proportion change of a specific frequency band is set not to exceed ±10% of the normal range. The setting of these thresholds is based on a large amount of historical data statistical analysis, ensuring that when the equipment is running normally, the signal stability index can remain within the threshold range, and when the equipment is abnormal or fails, the signal stability index will exceed the threshold.

[0087] On the basis that the signal stability index meets the set stability threshold, it is indicated that the signal is in a stable state. At this time, the candidate fault type identifier in the pre-stored fault type identifier set is dynamically selected as the current fault type label. The dynamic selection strategy comprehensively considers factors such as the operating condition of the equipment, the historical fault occurrence frequency and the like. For example, in the high-load operating condition of the equipment, the fault type identifiers related to the load (such as overload, winding overheating and the like) are preferentially selected as candidates, because in this condition, the possibility of occurrence of these fault types is relatively high. For another example, according to the historical fault record of the equipment, if a certain type of fault (such as bearing wear) frequently occurs in the past, then when the signal is stable, the probability of the fault type identifier being selected as a candidate is appropriately increased. Through this dynamic selection strategy, the most likely fault type label can be quickly determined when the signal is stable and the fault characteristics are obvious, thereby improving the diagnosis efficiency.

[0088] On the basis that the signal stability index does not meet the set stability threshold, it is indicated that the signal may have abnormal fluctuations. At this time, the current historical feature parameter is input into the initial diagnosis network, and the current fault type label is obtained through the complex operation of the network. The initial diagnosis network has strong non-linear mapping capability and feature extraction capability, and can process abnormal fluctuations and complex characteristics in the signal. When the signal is unstable, the network can analyze and process the input feature parameters in depth through its multi-layer structure and the connection between neurons, and mine the fault characteristics hidden in the unstable signal, so as to accurately judge the fault type. For example, when the vibration signal has a short-term abnormal fluctuation due to external interference, the initial diagnosis network can distinguish whether it is a real fault characteristic or an interference signal through comprehensive analysis of the time domain and frequency domain characteristics of the signal, thereby giving a reliable fault type label.

[0089] In the whole signal stability judgment and processing flow, multiple key links are involved. The design of the preset mode directly affects the accuracy of the signal stability index, and needs to be designed according to different types of signals (vibration, temperature, current) and the characteristics of the equipment. The setting of the stability threshold is also crucial, and a too high threshold value may lead to insensitivity to signal abnormalities, while a too low threshold value may lead to frequent misjudgment. The rationality of the dynamic selection strategy determines whether the fault type label can be quickly and accurately determined when the signal is stable, and the actual operation of the equipment and the historical data need to be fully considered. The performance of the initial diagnosis network directly affects the fault diagnosis capability when the signal is unstable, and needs to be ensured through reasonable network structure design and parameter training to ensure its accuracy and reliability.

[0090] Through the method of the embodiment, adaptive judgment and processing of signal stability are realized. When the signal is stable, a high-efficiency dynamic selection strategy is used to quickly determine the fault type label; when the signal is unstable, the powerful analysis capability of the initial diagnosis network is used for accurate diagnosis. This way of flexibly adjusting the diagnosis strategy according to the signal state improves the adaptability and reliability of the entire fault diagnosis system and can better cope with various complex signal situations in the operation of electrical equipment, providing a strong guarantee for the safe and stable operation of the equipment.

[0091] Embodiment 4:

[0092] Based on embodiment 3, this embodiment describes the training process and network structure of the dynamic adaptive diagnosis network. Through learning from historical fault data, the network can adaptively handle multi-modal fault diagnosis of electrical equipment, improving diagnosis accuracy and flexibility.

[0093] A number of sets of historical joint fault records are selected from historical fault data. Each set of records contains the kth historical feature parameter, the corresponding kth fault type label, the kth confidence coefficient, and the k+1th historical feature parameter (k is a positive integer). These records not only contain feature parameters and labels at a single time, but also consider the continuity of feature parameters at adjacent times, which helps the network learn the evolution law of fault features over time. For example, for bearing wear faults, the vibration signal features at adjacent times may show a gradual change trend. By including this time continuity information, the network can better capture the feature change pattern in the fault development process.

[0094] Based on the selected historical joint fault records, the initial diagnosis network to be trained is trained. The input of each round of training is a set of historical joint fault records, and the network calculates the output result through forward propagation. In this process, the input historical feature parameters are processed and transformed through each layer of the network, and finally the prediction of the fault type is output. Then, the error between the prediction result and the true label (i.e., the fault type label) is calculated according to the preset error function (such as the cross-entropy loss function). The error function measures the deviation between the network prediction result and the actual situation, and the smaller the error, the more accurate the network prediction.

[0095] The network parameters of the initial diagnosis network, such as weights and biases, are updated through the backpropagation algorithm. The backpropagation algorithm adjusts the weights of each connection and the biases of neurons in the network according to the gradient information of the error function, so that the value of the error function gradually decreases. This process is iterated until the number of iterations for training the initial diagnosis network reaches the set training threshold. At this time, it is considered that the network has learned the features and laws in the historical fault data sufficiently, and the initial diagnosis network is determined as the dynamic adaptive diagnosis network. If the number of iterations does not reach the set training threshold, the training continues until the condition is met.

[0096] The dynamic self-adaptive diagnosis network comprises a plurality of functional modules which work cooperatively to achieve accurate diagnosis of multi-modal fault signals.

[0097] The cross-modal feature fusion module dynamically calculates the modal weight coefficients of the vibration signal, the temperature signal and the current signal using an attention mechanism. The attention mechanism can automatically adjust the weight of each modality according to the importance of different modal signals in the current fault diagnosis. For example, in the bearing wear fault, the vibration signal usually contains more rich fault feature information, so the module will assign a higher weight to the vibration signal, thereby highlighting the features of this modality and improving the diagnosis accuracy. In the winding overheating fault, the temperature signal may be more critical, and the module will accordingly increase the weight of the temperature signal. This dynamic adjustment of the weight enables the network to adaptively focus on the most relevant modal signals according to the specific fault type.

[0098] The spatio-temporal feature encoder adopts a bidirectional LSTM structure to encode the time series features. The bidirectional LSTM can simultaneously capture past and future information of the signal, effectively extracting long-term dependencies in the time series. For the fault diagnosis of electrical equipment, many fault features need to be accurately identified by analyzing the signal changes over a period of time. For example, encoding the current signal at consecutive time points, the bidirectional LSTM can analyze the trend of current change and determine whether there is a gradually developing fault, such as the slow increase of current caused by insulation aging. Through the processing of the bidirectional LSTM, the network can better understand the evolution law of the signal in the time dimension and improve the ability to capture fault features.

[0099] The multi-granularity classifier comprises a convolutional neural network branch and a fully connected classification branch in parallel, where the outputs of each branch are fused through an adaptive weighting strategy to generate the final fault type label. The convolutional neural network branch is good at extracting local features of the signal, such as specific frequency components in the vibration signal. It extracts representative local feature patterns by sliding the convolution kernel over the signal. The fully connected classification branch can integrate global features for judgment, integrating and classifying all the extracted features. The adaptive weighting strategy dynamically adjusts the output weight of each branch according to its performance in different situations. For example, in some fault types, local features are more critical, and the weight of the convolutional neural network branch will be increased accordingly; while in other fault types, global features play a dominant role, and the weight of the fully connected classification branch will be greater. Through this fusion method, the multi-granularity classifier can fully utilize the advantages of different types of features to generate more reliable fault type labels.

[0100] The online updating module is an important part of the dynamic adaptive diagnostic network. When the confidence coefficients of consecutive N signal samples are detected to be lower than the preset threshold, it indicates that the current diagnostic performance of the network may have declined, and parameter updating is needed. At this time, the online updating module automatically triggers the incremental learning training of the network parameters. Incremental learning allows the network to update parameters in real time according to newly emerging fault data without forgetting previously learned knowledge. For example, when a new fault mode occurs in the equipment, through incremental learning, the network can incorporate this new fault feature into its knowledge system, thereby adapting to changes in the equipment operating state and maintaining the stability and advancement of diagnostic performance.

[0101] During the training and operation of the entire dynamic adaptive diagnostic network, the various modules cooperate with each other to form a complete fault diagnosis system. The cross-modal feature fusion module ensures that signals of different modalities can be reasonably integrated and utilized, the spatio-temporal feature encoder captures the time-dependent relationship of the signals, the multi-granularity classifier analyzes the features from different angles and generates diagnostic results, and the online updating module ensures that the network can adapt to the changing equipment state. This design enables the dynamic adaptive diagnostic network to efficiently and accurately handle the multi-modal fault diagnosis problem of electrical equipment, providing a strong guarantee for the safe operation of the equipment.

[0102] Embodiment 5:

[0103] Based on the dynamic adaptive diagnostic network constructed in Embodiment 4, this embodiment further designs a three-level early warning mechanism to achieve a graded response to electrical equipment faults through comprehensive analysis of multi-modal signals. This mechanism triggers different levels of early warning based on the comparison results of confidence coefficients and dynamic thresholds, ensuring that appropriate measures are taken at different stages of fault development.

[0104] When the confidence coefficients of the matching results of fault features in at least two modalities of vibration signals, temperature signals, and current signals simultaneously exceed the dynamic threshold, the three-level early warning mechanism is triggered. The dynamic threshold is dynamically adjusted according to the confidence distribution of similar faults in the historical operation data of the equipment. Specifically, by statistically analyzing the distribution of confidence coefficients of each modality when a certain type of fault occurs in the past, the dynamic threshold is set as:

[0105]

[0106] wherein, is the dynamic threshold, is the average value of historical confidence coefficients of similar faults, is the standard deviation, is the adjustment coefficient (usually 1.5-2.5). This formula ensures that the threshold can adapt to changes in equipment characteristics and operating environment, and the threshold for different equipment or the same equipment under different operating conditions will automatically adjust according to actual historical data.

[0107] The threshold adjustment step of different modal signals is negatively correlated with their signal sampling frequency. For vibration signals with high sampling frequency (e.g., 10 kHz), the threshold adjustment step is small (e.g., 0.02) to ensure high sensitivity to rapidly changing signal characteristics. For temperature signals with low sampling frequency (e.g., 1 Hz), the threshold adjustment step is large (e.g., 0.05) to avoid frequent triggering of warnings due to slow signal changes. This design allows the system to optimize the warning trigger logic based on the characteristics of each modal signal.

[0108] The specific implementation of the three-level warning mechanism is as follows:

[0109] Level 1 warning: When the trigger condition is met, the system first displays a yellow warning symbol through the device operation interface. This symbol uses a prominent color and flashing effect to ensure that the operator can promptly notice the device abnormal state. At the same time, the system automatically generates detailed diagnostic logs, recording information such as fault feature matching results, modal confidence coefficients, occurrence time, device operating parameters, etc. The diagnostic logs are stored in a structured format for subsequent tracing and analysis.

[0110] Level 2 warning: Based on the level 1 warning, the audible and visual alarm system is added. The audible and visual alarm is installed in the device site and control room, emitting a specific frequency alarm sound and red flashing light to attract the attention of the operator. At the same time, the system automatically locks the current device operating parameters to prevent further parameter fluctuations affecting device safety, providing stable reference data for fault troubleshooting. The locked parameters include current, voltage, temperature, vibration amplitude, and other key indicators, which are highlighted in special colors on the operation interface.

[0111] Level 3 warning: Based on the level 2 warning, the system synchronously pushes an encrypted alarm message to the remote monitoring platform. The alarm message is transmitted using a secure encryption protocol, containing device unique identifier, fault type, severity, real-time operating parameters, and other key information. After receiving the message, the remote monitoring platform immediately triggers SMS, email, or APP push notifications to inform the responsible person, achieving real-time fault reporting across regions.

[0112] The trigger logic of the entire warning mechanism also considers the duration of the signal. To avoid false alarms due to transient interference, only when the confidence coefficient exceeds the threshold for more than a set time (e.g., vibration signal for 30 seconds, temperature signal for 5 minutes) will the corresponding level of warning be triggered. This design effectively improves the reliability of the warning system.

[0113] In the implementation process of the early warning mechanism, the system also integrates the fault development trend prediction function. By analyzing the confidence coefficient change trend at multiple time points in succession, the possible development direction and severity of the fault are predicted. For example, if the confidence coefficient of a certain mode shows a sustained upward trend, it indicates that the fault may be intensifying, and the processing speed needs to be accelerated. The prediction results are displayed in the form of charts on the operation interface, providing decision support for the operators.

[0114] The design of the three-level early warning mechanism fully considers the diversity and complexity of electrical equipment faults. Through the synergistic effect of functions such as dynamic threshold adjustment, multi-modal signal comprehensive judgment, graded response, and trend prediction, the mechanism can provide timely and accurate early warning information under different fault scenarios, providing strong support for equipment maintenance and fault handling, and minimizing equipment downtime and maintenance costs.

[0115] It should be noted that in this paper, relational terms such as first and second are used only to distinguish one entity or action from another, and do not necessarily require or imply that there is any such actual relationship or order between these entities or actions. Moreover, the terms "include", "contain" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or device.

[0116] Although embodiments of the present application have been shown and described, it will be understood by those having ordinary skill in the art that various changes, modifications, substitutions and alterations can be made without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A multi-modal fault diagnosis method for electrical equipment based on dynamic self-adaptation, characterized in that: The method is applied to an electrical equipment fault diagnosis system, and the method comprises: Acquire an original multimodal signal stream for fault diagnosis, wherein the original multimodal signal stream includes a vibration signal, a temperature signal, and a current signal; Inputting the original multimodal signal stream into a pre-trained dynamic adaptive diagnosis network to obtain a fault feature matching result set; Determining a fault feature matching result corresponding to each modal signal in the original multimodal signal stream based on the fault feature matching result set; The dynamic adaptive diagnostic network is a parameterized network obtained by training an initial diagnostic network to be trained by inputting historical fault data into the historical fault data. The historical fault data includes a batch of historical characteristic parameters extracted from a batch of historical multimodal signals, as well as a fault type label and a confidence coefficient corresponding to each historical characteristic parameter. The fault type label reflects an abnormal state identifier of a historical characteristic parameter, and the confidence coefficient reflects a correction factor corresponding to the abnormal state identifier. The batch of historical multimodal signals is a batch of continuously collected electrical equipment operating signals. Each historical characteristic parameter includes a characteristic parameter determined by a historical multimodal signal and adjacent historical multimodal signals of the historical multimodal signal. The fault feature matching result is used to indicate a fault type identifier of the multimodal signal stream; The fault type label and confidence coefficient corresponding to each historical characteristic parameter in the batch of historical characteristic parameters are determined respectively by the following steps, wherein the historical characteristic parameter for which the fault type label and confidence coefficient are determined each time is used as the current historical characteristic parameter, and the fault type label and confidence coefficient corresponding to the current historical characteristic parameter are used as the current fault type label and current confidence coefficient: Inputting each fault type identifier in the pre-stored fault type identifier set together with the current historical characteristic parameter into the first classification algorithm branch in order to obtain a first set of classification confidence indicators, wherein the first set of classification confidence indicators includes the confidence indicator corresponding to each fault type identifier; Determine the candidate fault type identifier corresponding to the first confidence indicator with the highest value in the first group of classification confidence indicators as the current fault type identifier corresponding to the current fault type label; The current confidence coefficient is determined based on the current fault type label and a benchmark verification identifier corresponding to the current historical characteristic parameter, wherein the benchmark verification identifier carries a calibration label in the historical multimodal signal corresponding to the current historical characteristic parameter, and the calibration label includes a target fault type identifier of the multimodal signal stream, and the confidence coefficient is used to indicate whether the current fault type label is consistent with the target fault type identifier.

2. The method for multi-modal fault diagnosis of electrical equipment based on dynamic self-adaptation according to claim 1, characterized in that: The method further comprises: Extracting time-domain feature parameters from the batch of historical multimodal signals to obtain a first set of feature parameters, wherein a feature parameter in the first set of feature parameters corresponds to a historical multimodal signal in the batch of historical multimodal signals; Extracting frequency domain feature parameters from the batch of historical multimodal signals to obtain a second set of feature parameters, wherein a first feature parameter in the second set of feature parameters corresponds to a first historical multimodal signal in the historical multimodal signals, an adjacent historical multimodal signal of the first historical multimodal signal is a second historical multimodal signal, and the first feature parameter represents a frequency domain correlation feature between the first historical multimodal signal and the second historical multimodal signal; The first set of feature parameters and the second set of feature parameters are fused to obtain the batch of historical feature parameters.

3. The method for multi-modal fault diagnosis of electrical equipment based on dynamic self-adaptation according to claim 2, characterized in that: The extracting frequency domain feature parameters of the batch of historical multimodal signals to obtain a second set of feature parameters includes: Performing frequency domain feature parameter extraction on the batch of historical multimodal signals to obtain a second set of feature parameters through the following steps, wherein the historical multimodal signal subjected to the frequency domain feature parameter extraction each time is used as the current historical multimodal signal, and the obtained feature parameters are used as the current feature parameters, and the second set of feature parameters includes the current feature parameters; Performing a spectrum decomposition operation on the current historical multimodal signal to determine a first set of frequency domain feature quantities, wherein the first set of frequency domain feature quantities is used to describe a main frequency band energy distribution in the current historical multimodal signal; Performing a spectrum decomposition operation on adjacent historical multimodal signals of the current historical multimodal signal to determine a second set of frequency domain feature quantities, wherein the second set of frequency domain feature quantities is used to describe a main frequency band energy distribution in the adjacent historical multimodal signals; The current characteristic parameter is determined based on the first set of frequency domain feature quantities and the second set of frequency domain feature quantities, wherein the current characteristic parameter is used to represent the energy deviation of the corresponding frequency bands in the first set of frequency domain feature quantities and the second set of frequency domain feature quantities.

4. The method for multi-modal fault diagnosis of electrical equipment based on dynamic self-adaptation according to claim 1, characterized in that: Before inputting each fault type identifier in the pre-stored fault type identifier set together with the current historical characteristic parameters into the first classification algorithm branch in order to obtain a first set of classification confidence indicators, the method further includes: Determine signal stability indicators through preset modes; Determining whether the signal stability indicator meets a set stability threshold; On the basis that the signal stability index meets the set stability threshold, dynamically selecting a candidate fault type identifier from the pre-stored fault type identifier set as the current fault type label; On the basis that the signal stability index does not meet the set stability threshold, the current historical characteristic parameters are input into the initial diagnosis network to obtain the current fault type label.

5. The method for multi-modal fault diagnosis of electrical equipment based on dynamic self-adaptation according to claim 1, characterized in that: The step of determining the candidate fault type identifier corresponding to the first confidence indicator having the highest value in the first group of classification confidence indicators as the current fault type identifier corresponding to the current fault type label includes: Determining the current confidence coefficient as a first confidence coefficient based on the candidate fault type identifier corresponding to the first confidence indicator with the highest value in the first group of classification confidence indicators being the same as the target fault type identifier, wherein the first confidence coefficient indicates that the initial diagnostic network successfully generates a matching identifier; Based on the fact that the candidate fault type identifier corresponding to the first confidence indicator with the highest value in the first group of classification confidence indicators is different from the target fault type identifier, the current confidence coefficient is determined as the second confidence coefficient, wherein the second confidence coefficient indicates that the initial diagnostic network failed to successfully generate a matching identifier.

6. The method for multi-modal fault diagnosis of electrical equipment based on dynamic self-adaptation according to claim 1, characterized in that: The method further comprises: screening a plurality of groups of historical joint fault records from the historical fault data, wherein a kth group of historical joint fault records among the plurality of groups of historical joint fault records includes a kth historical characteristic parameter, a kth fault type label corresponding to the kth historical characteristic parameter, a kth confidence coefficient, and a k+1th historical characteristic parameter, where k is a positive integer; An initial diagnostic network to be trained is trained based on the several groups of historical joint fault records to obtain the dynamic adaptive diagnostic network, wherein, on the basis that the number of iterations of training the initial diagnostic network reaches a set training threshold, the initial diagnostic network is determined to be the dynamic adaptive diagnostic network, and on the basis that the number of iterations of training the initial diagnostic network does not reach the set training threshold, the network parameters of the initial diagnostic network are updated according to a preset error function, and the input of each round of training process is a group of historical joint fault records from the several groups of historical joint fault records.

7. The method for multi-modal fault diagnosis of electrical equipment based on dynamic self-adaptation according to claim 6, characterized in that: The dynamic adaptive diagnostic network comprises: The cross-modal feature fusion module uses the attention mechanism to dynamically calculate the modal weight coefficients of vibration signals, temperature signals, and current signals; The spatiotemporal feature encoder uses a bidirectional LSTM structure to encode time series features; A multi-granularity classifier, consisting of a parallel convolutional neural network branch and a fully connected classification branch, where the outputs of each branch are fused through an adaptive weighting strategy to generate the final fault type identification; The online update module automatically triggers incremental learning and training of network parameters when it detects that the confidence coefficient of N consecutive signal samples is lower than the preset threshold.

8. The method for multi-modal fault diagnosis of electrical equipment based on dynamic self-adaptation according to claim 1, characterized in that: The method further comprises: When the confidence coefficients of the fault feature matching results of at least two modes among the vibration signal, temperature signal, and current signal simultaneously exceed the dynamic threshold, a three-level warning mechanism is triggered. The first-level warning displays a yellow warning icon on the device operation interface and generates a diagnostic log. The second-level warning superimposes an audible and visual alarm and automatically locks the device operating parameters. The third-level warning simultaneously pushes an encrypted alarm message to the remote monitoring platform. The dynamic threshold is dynamically adjusted according to the confidence distribution law of similar faults in the historical operation data of the equipment, and the threshold adjustment step of different modal signals is negatively correlated with their signal sampling frequency.

Images