Multimodal fault diagnosis and early warning method

Abstract

The invention discloses a multi-modal fault diagnosis and early warning method, which relates to the technical field of substation fault prediction, and comprises the following steps: respectively performing spectrum analysis on several modal signals used for fault prediction of historical substation main equipment to obtain a first frequency set, and using the G times frequency of the first frequency set as the initial sampling frequency of each modal signal to obtain an initial sampling frequency set; obtaining sampling overlap data and corresponding fault prediction data corresponding to each group of sampling frequency sets, and constructing a sampling frequency evaluation model; constructing several sampling frequency-evaluation effect curves with the output of the sampling frequency evaluation model and the corresponding sampling frequency set, and performing data analysis on the several sampling frequency-evaluation effect curves to obtain an improved sampling frequency set, and performing fault prediction according to the improved sampling frequency set. The application solves the problem of poor sampling frequency selection in the prior art due to the lack of a mechanism for dynamically adjusting the sampling frequency.

Description

Technical Field

[0001] The present invention relates to the technical field of transformer substation fault prediction, and more specifically, to a multi-modal fault diagnosis and early warning method. Background Art

[0002] Existing multimodal fault diagnosis and early warning methods usually include the following technical processes: First, the system collects data of different modes from multiple sensors or data sources, such as current, voltage, temperature, vibration, sound waves and other signals. Then, through data preprocessing steps such as denoising, normalization and dimensionality reduction, the quality and consistency of the data are ensured. Next, feature extraction methods (such as time-frequency analysis, wavelet transform or deep learning) are used to extract relevant features from each mode, and multimodal information is comprehensively processed through methods such as feature fusion or decision fusion to capture potential fault characteristics. Finally, based on the fused features, they are input into the prediction model or classifier for real-time fault diagnosis and early warning, and the health status and potential fault type of the equipment are output.

[0003] For example, the measurement and control equipment of the substation announced by the invention patent with announcement number: CN109901090B, the remote control terminal of the measurement and control screen, is used for traditional verification of the upgraded monitoring equipment and telecontrol equipment, wherein the upgraded monitoring equipment and telecontrol equipment are equipment used in conjunction with the substation, and the remote control terminal of the measurement and control screen includes: a disconnectable terminal block and a test jack; an indicator light, connected to the remote control terminal of the measurement and control screen, and when the remote control contact on the inner side of the remote control terminal of the measurement and control screen is turned on, a predetermined signal is sent out, wherein the predetermined signal is used to indicate the monitoring status. The present invention solves the technical problem that the method of monitoring the remote control outlet of the substation system in the related art has low safety.

[0004] For example, the invention patent with announcement number: CN108333449B discloses an online monitoring method and device for substation equipment. The vibration alarm obtains substation equipment information collected by sensors installed at different substation equipment. The vibration alarm determines whether the substation equipment information is normal based on the pre-stored standard substation equipment information. If the vibration alarm determines that the substation equipment information is abnormal, it sends a vibration signal to the information processing platform. The information processing platform receives the vibration signal and determines the substation equipment failure based on the vibration signal. After adopting the above scheme, the status of the substation equipment can be monitored through a variety of different sensors, and various hidden faults caused by long-term high-load operation of the substation equipment can be discovered in time, and handled in time, ensuring the safe and stable operation of the substation equipment.

[0005] In the above disclosed technical solutions, there are at least the following technical problems: different sampling frequencies have a significant impact on the fault prediction of multimodal data; higher sampling frequencies can capture more details and dynamic changes, making it easier for the model to identify potential abnormal patterns; however, too high a sampling frequency may lead to data redundancy, increase the computational burden, and may introduce noise, reducing the prediction accuracy; on the other hand, although lower sampling frequencies can reduce the amount of data and processing complexity, they may miss key fault signals, resulting in inaccurate predictions; however, due to the lack of a mechanism for dynamically adjusting the sampling frequency, the sampling frequency of the prior art is poorly selected. In view of the above problems, the present invention proposes a solution. Summary of the invention

[0006] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides a multimodal fault diagnosis and early warning method, which solves the problem of poor sampling frequency selection in the prior art due to the lack of a mechanism for dynamically adjusting the sampling frequency through sampling frequency screening.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] The multimodal fault diagnosis and early warning method comprises the following steps: performing spectrum analysis on several modal signals used for fault prediction of historical substation main equipment respectively to obtain a first frequency set, and taking the G times frequency of the first frequency set as the initial sampling frequency of each modal signal to obtain an initial sampling frequency set; performing random perturbations on the initial sampling frequency set several times to obtain several groups of different sampling frequency sets; sampling the modal signals corresponding to the substation main equipment according to the sampling frequency set to obtain multimodal data, and performing fault prediction based on ST-GCN; obtaining the sampling overlap data and the corresponding fault prediction data corresponding to each group of sampling frequency sets, and constructing a sampling frequency evaluation model; constructing several sampling frequency-evaluation effect curves with the output of the sampling frequency evaluation model and the corresponding sampling frequency set, and performing data analysis on the several sampling frequency-evaluation effect curves to obtain an improved sampling frequency set, and performing fault prediction based on the improved sampling frequency set.

[0009] In a preferred embodiment, the several modal signals include current signals, voltage signals, temperature signals, vibration signals and acoustic signals; the first frequency set is a set of frequencies obtained by spectral analysis of all modal signals; the initial sampling frequency set is composed of a vector consisting of G times the first frequency set and the corresponding modal signal category; the sampling overlap data includes a modal information overlap coefficient, and the fault prediction data includes a fault prediction deviation coefficient and a model efficiency evaluation coefficient.

[0010] In a preferred embodiment, the method for obtaining the several groups of different sampling frequency sets is as follows: taking the initial sampling frequency set as the basis, each modal signal corresponds to a fixed initial sampling frequency; based on a random number generator, introducing random disturbances to the sampling frequency of each modal signal; obtaining the preset upper and lower limits of the sampling frequency, and generating a random offset within the upper and lower limits of the sampling frequency, so that the sampling frequency of each mode fluctuates within a certain range, and obtaining several groups of different sampling frequency sets.

[0011] In a preferred embodiment, the specific method for fault prediction based on ST-GCN is as follows: according to different sampling frequency sets, each modal signal of the substation main equipment is synchronously collected to obtain multimodal data containing spatiotemporal information, and the multimodal data is stored in the form of a time series; the nodes of each modal data in the multimodal data are used as nodes in the hidden danger graph structure of the substation main equipment, and the spatiotemporal correlations between different modal signals are used as edges between nodes; the sampling values ​​of the modal signal are used as signal features, and the node features are represented by a feature matrix, and the features of each node are associated with its feature matrix based on ST-GCN. The features of adjacent nodes are aggregated to obtain the aggregated hidden danger map structure of the substation main equipment; the features of each node are aggregated with the features of its adjacent nodes, specifically: the feature matrix of each node is weighted summed with the feature matrix of its neighboring nodes, and the result of the weighted summation is linearly transformed based on a preset hidden danger aggregation formula to obtain the aggregated hidden danger map structure of the substation main equipment; the hidden danger map structure of the substation main equipment is aggregated multiple times to obtain the final hidden danger map structure of the substation main equipment, and the hidden danger map structure of the substation main equipment is processed based on a preset classification regression formula to output the fault prediction result of the equipment.

[0012] In a preferred embodiment, the specific method for obtaining the modal information overlap coefficient is as follows: extracting feature data from the multimodal signal of the main substation equipment; performing denoising, normalization and other processing on the modal signal, and calculating the mean and variance of the modal signal; comparing the mean and variance of each pair of modal signals respectively; obtaining the probability distribution function of the modal signal, and calculating the autoentropy of the modal signal, and calculating the cross entropy between each pair of modal signals based on the autoentropy of each modal signal; taking the minimum value of the comparison results of the two modes in terms of mean and variance, and adding these minimum values ​​and combining them with the cross entropy between the modal signals to calculate the modal information overlap coefficient.

[0013] In a preferred embodiment, the specific method for obtaining the fault prediction deviation coefficient is as follows: obtain the prediction result corresponding to the interval P time period within the time L at a specific sampling frequency; construct a prediction result waveform according to the prediction result, and perform Fourier transform, convert the prediction result waveform to the frequency domain, and compare it with the actual fault result; in the frequency domain, calculate the peak and valley values ​​of the difference between the prediction result waveform and the actual fault result, which respectively represent the peak and valley value positions of the prediction result waveform in the frequency spectrum; calculate the fault prediction deviation coefficient by the square of the Euclidean distance between the peak and the valley.

[0014] In a preferred embodiment, the specific method for obtaining the model efficiency evaluation coefficient is as follows: obtain the time complexity of the model; record the response time of the model after each command is received, and calculate the average response time; use the ratio of time complexity to average response time as the model processing efficiency; obtain the model space efficiency through square root calculation based on the known model space complexity and the number of model iterations; calculate the model efficiency evaluation coefficient based on the model space efficiency and the model processing efficiency.

[0015] In a preferred embodiment, the output of the sampling frequency evaluation model and the corresponding sampling frequency set are used to construct several sampling frequency-evaluation effect curves, specifically: obtaining the sampling frequency evaluation coefficients of the outputs of the sampling frequency sets under different settings from the sampling frequency evaluation model under different sampling frequency sets; establishing a rectangular coordinate system, with the horizontal axis representing the sampling frequency set and the vertical axis representing the sampling frequency evaluation coefficients; marking the sampling frequency evaluation coefficients corresponding to each sampling frequency set on the rectangular coordinate system to form data points; connecting the data points to form a sampling frequency-evaluation effect curve graph; respectively selecting a modal signal in the multi-modality as the direction of the horizontal axis, and constructing the same number of sampling frequency-evaluation effect curve graphs as the number of modal signals.

[0016] In a preferred embodiment, the data analysis is performed on several sampling frequency-evaluation effect curves to obtain an improved sampling frequency set, specifically: the peak value of each sampling frequency-evaluation effect curve is secondarily differentiated, and the sampling frequency set corresponding to the peak point of the sampling frequency-evaluation effect curve with the largest second derivative result is used as the improved sampling frequency set, and fault prediction is performed based on the improved sampling frequency set.

[0017] Technical effects and advantages of the multi-modal fault diagnosis and early warning method of the present invention:

[0018] 1. The present invention realizes the generation of the initial sampling frequency set by performing spectrum analysis on the multimodal signals of the historical substation main equipment, and obtains multiple sets of sampling frequency sets by random perturbation method. This process ensures the comprehensive collection of important modal data such as current, voltage, temperature, vibration and acoustic signals, reflecting the operating status and potential fault risks of the equipment. The method based on spatiotemporal graph convolutional network (ST-GCN) is used for fault prediction. By aggregating the characteristics of each node with the information of neighboring nodes, the hidden danger characteristics of the equipment are effectively extracted, and then the hidden danger graph structure is generated. The advantage of this technology is that through dynamic sampling frequency configuration and multimodal data fusion, the accuracy and reliability of fault prediction can be improved, the abnormal state of equipment operation can be identified in time, and strong technical support for the safe operation of the equipment is provided. Through multiple aggregation operations, the understanding of the health status of the equipment can be further deepened, thereby providing data basis for subsequent maintenance decisions.

[0019] 2. The present invention optimizes the fault prediction process by establishing a sampling frequency evaluation model. First, the method achieves a comprehensive analysis of multimodal signals by obtaining the modal information overlap coefficient, fault prediction deviation coefficient and model efficiency evaluation coefficient of each set of sampling frequencies. This process not only identifies the information redundancy between modal signals, but also reveals the unique contribution of each mode to fault prediction, ensuring the optimal performance of the model at different frequencies. Secondly, the analysis of the modal information overlap coefficient helps to reduce data redundancy, improve data processing efficiency, and thus improve prediction accuracy. At the same time, the evaluation of the fault prediction deviation coefficient provides intuitive feedback on model performance, helping researchers optimize sampling strategies and ensure the consistency of prediction results with actual conditions. In addition, the model efficiency evaluation coefficient comprehensively considers the response time and resource consumption of the model, making the fault detection process more efficient. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a structural schematic diagram of the multi-modal fault diagnosis and early warning method of the present invention. DETAILED DESCRIPTION

[0021] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0022] Embodiment 1, Figure 1 The present invention provides a multi-modal fault diagnosis and early warning method, which includes the following steps:

[0023] S1, perform spectrum analysis on several modal signals used for fault prediction of historical substation main equipment to obtain a first frequency set, and use G times the frequency of the first frequency set as the initial sampling frequency of each modal signal to obtain an initial sampling frequency set.

[0024] In this embodiment, the plurality of modal signals include but are not limited to current signals, voltage signals, temperature signals, vibration signals and acoustic signals;

[0025] Current signal: Current is an important indicator of substation operation. Too high or too low current may reflect equipment overload, short circuit or other abnormal conditions. By performing spectrum analysis on current signals, early signs of electrical faults can be detected.

[0026] Voltage signal: Voltage fluctuations are often closely related to the operating status of the equipment. Voltage abnormalities may indicate a decline in the insulation performance of the equipment, circuit failure, or imbalance in the power system. The frequency spectrum characteristics of the voltage signal can reflect the health status of the power equipment.

[0027] Temperature signal: Substation equipment generates heat during operation. Temperature signal is a key indicator to measure whether the equipment is overloaded or insulation aged. Abnormal temperature usually indicates internal failure or deterioration of the equipment.

[0028] Vibration signal: Mechanical components such as transformers and circuit breakers will generate vibrations during operation; spectral analysis of vibration signals can reveal mechanical failures of equipment, such as bearing wear or mechanical looseness;

[0029] Acoustic signals: The sound signals generated when the equipment is running can also be used for fault detection, especially in high-voltage power equipment. Partial discharge or other internal faults may be accompanied by noise of specific frequencies. These abnormal phenomena can be captured through acoustic signal analysis.

[0030] It should be noted that the process of spectrum analysis mainly includes the following steps: First, a multimodal signal set including current, voltage, temperature, vibration, acoustic and other signals needs to be extracted from historical data. Each signal needs to be preprocessed separately, such as removing noise, filling missing values, etc., to ensure the integrity and effectiveness of the signal. Next, for each modal signal, a spectrum analysis method such as Fourier transform (FFT) or short-time Fourier transform (STFT) is applied to convert the time domain signal into a frequency domain signal to obtain the frequency distribution characteristics of the signal. The frequency components in the frequency domain represent the periodicity and transient characteristics in the signal. Then, these frequency distributions are analyzed to extract specific frequencies or harmonic features that may be related to equipment failures. The spectrum results of each modal signal are summarized to form a multimodal frequency set to provide support for subsequent sampling frequency setting and fault prediction model training; the process of spectrum analysis has mature technology, and this application will not repeat it.

[0031] It should be noted that G is an artificially set mathematical value, which defaults to 2 and can be adjusted as needed.

[0032] In this embodiment, the first frequency set is a set of frequencies obtained by performing spectrum analysis on all modal signals; the initial sampling frequency set is composed of a vector consisting of G times the first frequency set and corresponding modal signal categories.

[0033] S2, perform several random perturbations on the initial sampling frequency set to obtain several different sampling frequency sets.

[0034] In this embodiment, the method for obtaining the plurality of different sampling frequency sets is as follows:

[0035] Taking the initial sampling frequency set as the basis, each modal signal corresponds to a fixed initial sampling frequency;

[0036] Based on the random number generator, random perturbations are introduced into the sampling frequency of each modal signal;

[0037] The preset upper and lower limits of the sampling frequency are obtained, and a random offset is generated within the upper and lower limits of the sampling frequency, so that the sampling frequency of each mode fluctuates within a certain range, and several groups of different sampling frequency sets are obtained.

[0038] S3, sample the modal signals corresponding to the main substation equipment according to the sampling frequency set to obtain multi-modal data, and perform fault prediction based on ST-GCN.

[0039] In this embodiment, the specific method for fault prediction based on ST-GCN is as follows:

[0040] According to different sampling frequency sets, each modal signal of the main substation equipment is synchronously collected to obtain multimodal data containing time and space information, and the multimodal data is stored in the form of a time series;

[0041] The nodes of each modal data in the multimodal data are used as nodes in the hidden danger graph structure of the main substation equipment, and the spatiotemporal correlation between different modal signals is used as the edge between the nodes;

[0042] The sampling value of the modal signal is used as the signal feature, and the node feature is represented by the feature matrix. Based on ST-GCN, the feature of each node is aggregated with the features of its adjacent nodes to obtain the aggregated hidden danger graph structure of the substation main equipment.

[0043] The feature of each node is aggregated with the features of its adjacent nodes, specifically:

[0044] The characteristic matrix of each node is weightedly summed with the characteristic matrix of its neighboring nodes, and the result of the weighted summation is linearly transformed based on a preset hidden danger aggregation formula to obtain the aggregated hidden danger graph structure of the main substation equipment;

[0045] The hidden danger graph structure of the main substation equipment is aggregated multiple times to obtain the final hidden danger graph structure of the main substation equipment. Based on the preset classification regression formula, the hidden danger graph structure of the main substation equipment is processed to output the fault prediction result of the equipment.

[0046] It should be noted that ST-GCN is a feasible method for fault prediction based on multimodal fusion. This application is also applicable to other fault prediction methods, such as multi-task learning, multimodal knowledge enhanced reasoning analysis, and multimodal prediction based on decision fusion, which will not be elaborated here.

[0047] It should be noted that the task of the preset classification and regression formula is to use the spatiotemporal features extracted by the graph convolution layer for classification or regression prediction. Different specific calculation methods can be selected for different situations. For classification tasks, the features are mapped to the probability distribution of fault categories. For regression tasks, continuous values ​​related to the health status of the equipment or the degree of fault are output. Therefore, the calculation method of the classification and regression layer will vary depending on the task type.

[0048] It should be noted that the preset hidden danger aggregation formula integrates the features of each node with the features of its adjacent nodes by weighted summation to extract the hidden danger information of the main substation equipment; the aggregated features are processed by a linear transformation to introduce more hidden danger feature information; finally, this process will generate a new feature representation, reflecting the relationship and potential hidden dangers between each node and its neighbors; through multiple such aggregation operations, a complete hidden danger graph structure of the main substation equipment can be gradually constructed to provide rich feature support for subsequent fault prediction. The hidden danger aggregation formula is the formula for any convolutional graph iteration.

[0049] It should be noted that the spatiotemporal correlation is defined based on the physical location of the device, the correlation of the signal, or the dependency between historical data; for example, signals at adjacent time points will form a temporal correlation, and different signals will form a spatial correlation.

[0050] This embodiment realizes the generation of the initial sampling frequency set by performing spectrum analysis on the multimodal signals of the historical substation main equipment, and obtains multiple sets of sampling frequency sets through the random perturbation method. This process ensures the comprehensive collection of important modal data such as current, voltage, temperature, vibration and acoustic signals, reflecting the operating status and potential fault risks of the equipment. The method based on spatiotemporal graph convolutional network (ST-GCN) is used for fault prediction. By aggregating the characteristics of each node with the information of neighboring nodes, the hidden danger characteristics of the equipment are effectively extracted, and then the hidden danger graph structure is generated. The advantage of this technology is that through dynamic sampling frequency configuration and multimodal data fusion, it can improve the accuracy and reliability of fault prediction, timely identify abnormal conditions in equipment operation, and provide strong technical support for the safe operation of the equipment. Through multiple aggregation operations, the understanding of the health status of the equipment can be further deepened, thereby providing data basis for subsequent maintenance decisions.

[0051] Embodiment 2, S4, obtaining sampling overlap data and corresponding fault prediction data corresponding to each sampling frequency set, and constructing a sampling frequency evaluation model.

[0052] The sampling overlap data includes a modal information overlap coefficient, and the fault prediction data includes a fault prediction deviation coefficient and a model efficiency evaluation coefficient.

[0053] The modal information overlap coefficient is a key indicator used to evaluate the degree of information overlap between multimodal signals and the overlap within each mode. Specifically, this coefficient reflects the information redundancy and complementarity of these modes in the fault prediction task by calculating the overlapping area of ​​different modal signals in the time and frequency domains. A higher modal information overlap coefficient indicates that there is a strong correlation between the modes, which may lead to information redundancy, while a lower coefficient indicates that each mode provides unique information, which helps to improve the prediction ability of the model. Through the evaluation of this coefficient, the sampling frequency and mode selection can be optimized to ensure that the advantages of each mode are fully utilized in the fault prediction process and enhance the accuracy and robustness of the overall model.

[0054] Analyzing the modal information overlap coefficient under different sampling frequency sets has the following advantages for analyzing the optimal sampling frequency set:

[0055] Identify information redundancy: By comparing the modal information overlap coefficients at different sampling frequencies, the redundant information between the modes can be effectively identified, helping to optimize data collection, reduce unnecessary data volume, and improve the processing efficiency of the model;

[0056] Improve prediction accuracy: By analyzing the overlap coefficient, the most complementary modal combination can be found, ensuring that the unique information of different modalities is fully utilized in fault prediction and improving the accuracy of the prediction model;

[0057] Optimize sampling strategy: According to the change of overlap coefficient, it can guide the dynamic adjustment of sampling frequency to obtain more representative and effective samples, thus forming a more reasonable sampling strategy;

[0058] Enhance model robustness: By selecting the best sampling frequency set, the risk of model overfitting caused by information overlap can be reduced, and the model's robustness to unknown faults can be enhanced;

[0059] Support multi-modal fusion: Overlap coefficient analysis at different sampling frequencies helps to effectively fuse multi-modal signals, ensuring that the fused features can more comprehensively reflect the device status, thereby improving the reliability of fault detection;

[0060] In summary, the analysis of modal information overlap coefficient provides an important theoretical basis and practical guidance for selecting the optimal sampling frequency set.

[0061] The specific method for obtaining the modal information overlap coefficient is as follows:

[0062] Extract characteristic data from multi-modal signals of main substation equipment;

[0063] Perform denoising and normalization on the modal signal, and calculate the mean and variance of the modal signal;

[0064] For each pair of modal signals, the mean and variance are compared pairwise;

[0065] Obtain the probability distribution function of the modal signal, calculate the autoentropy of the modal signal, and calculate the cross entropy between each pair of modal signals based on the autoentropy of each modal signal;

[0066] The minimum value of the comparison results of the two modes in terms of mean and variance is taken, and these minimum values ​​are added together and combined with the cross entropy between the modal signals to calculate the modal information overlap coefficient.

[0067] The specific calculation formula of the self-entropy of the modal signal is as follows:

[0068]

[0069] The specific calculation formula of the cross entropy between the modal signals is as follows:

[0070]

[0071] The specific calculation formula of the modal information overlap coefficient is as follows:

[0072]

[0073] In the formula, is the modal signal The self-entropy, is the modal signal value, is the modal signal and The cross entropy of is the number of combinations of selecting 2 from n modes, is the modal information overlap coefficient, is the total number of modal signal types, is the minimum value of the mean difference and the variance difference, is the sum of the mean and variance differences.

[0074] The fault prediction deviation coefficient is an important indicator for evaluating the performance of fault prediction models. It mainly reflects the difference between the model prediction results and the actual fault conditions. By quantifying this difference, the fault prediction deviation coefficient helps researchers and engineers understand the accuracy and reliability of the model. Specifically, the fault prediction deviation coefficient determines the accuracy of the prediction by calculating the difference between the model prediction value and the actual observation value. This coefficient is usually used to evaluate the performance of the model under different sampling frequencies or conditions, helping to identify possible deficiencies in the model and thus optimizing the fault prediction method. A lower deviation coefficient means that the model prediction results are relatively consistent with the actual situation, otherwise it means that there may be problems with the model's predictive ability.

[0075] Analyzing the fault prediction deviation coefficient under different sampling frequency sets has the following advantages for analyzing the optimal sampling frequency set:

[0076] Accuracy identification: By comparing the fault prediction deviation coefficients under different sampling frequency sets, it is possible to intuitively identify which frequency configuration provides higher prediction accuracy. A lower deviation coefficient indicates that the model can better reflect the actual fault situation.

[0077] Differences in modal signal performance: Reveals differences in the performance of different modal signals at different frequencies, helping to optimize data acquisition strategies and focus resources on signals that contribute most to fault prediction;

[0078] Model stability assessment: Evaluate the stability of the model at different frequencies to ensure that the fault prediction model can maintain high accuracy under various operating conditions and improve the safety and reliability of equipment operation;

[0079] Optimize resource allocation: Analyze the deviation coefficient and optimize the sampling frequency setting to reduce the data processing burden and improve the efficiency of fault prediction;

[0080] Feedback adjustment mechanism: Provides a feedback mechanism to help adjust the sampling frequency in actual operation, thereby continuously improving the performance of the fault prediction model;

[0081] In summary, the analysis of fault prediction deviation coefficient provides important theoretical basis and practical guidance for selecting the optimal sampling frequency set.

[0082] The specific method for obtaining the fault prediction deviation coefficient is as follows:

[0083] Get the prediction results corresponding to the time interval P within the time L for a specific sampling frequency;

[0084] Construct a prediction result waveform based on the prediction result, perform Fourier transform, convert the prediction result waveform into the frequency domain, and compare it with the actual fault result;

[0085] In the frequency domain, the peak and valley values ​​of the difference between the predicted result waveform and the actual fault result are calculated, which represent the peak and valley positions of the predicted result waveform in the frequency spectrum respectively;

[0086] The fault prediction deviation coefficient is calculated by the square of the Euclidean distance between the peak and the valley;

[0087] The calculation formula of the fault prediction deviation coefficient is as follows:

[0088] In the formula, It is the frequency corresponding to the maximum difference between the predicted waveform and the actual fault result. It is the frequency corresponding to the minimum difference between the predicted waveform and the actual fault result. Represents the frequency of the predicted result waveform spectrum, Indicates a point in time, It is the prediction result corresponding to the P time period interval within the time L for a specific sampling frequency, and i is the number of the P time period interval within the time L.

[0089] The model efficiency evaluation coefficient is a performance indicator used to measure the fault prediction model at a specific sampling frequency. It comprehensively considers factors such as the model's prediction accuracy, computing time, and resource consumption. By comparing the efficiency evaluation coefficients of the model at different sampling frequencies, its response speed and accuracy when processing multimodal data can be evaluated, thereby guiding the optimization of sampling strategies to achieve more efficient fault detection and early warning. At the same time, the coefficient can also reveal the stability and scalability of the model under different operating conditions, ensuring good reliability in practical applications.

[0090] Analyzing the model efficiency evaluation coefficients under different sampling frequency sets has the following advantages for analyzing the optimal sampling frequency set:

[0091] Comprehensive evaluation of performance: The model efficiency evaluation coefficient takes into account prediction accuracy, computing time, and resource consumption, providing a comprehensive view of the overall performance of the model;

[0092] Optimize sampling strategy: By comparing the efficiency evaluation coefficients under different sampling frequencies, the best performance of the model under specific frequency conditions can be identified, thereby optimizing the data sampling strategy;

[0093] Improve response speed: Analyzing the efficiency of the model at different sampling frequencies can effectively improve the response speed of fault detection and ensure timely warning;

[0094] Support decision making: The data-driven support provided helps to make scientific sampling frequency selection and provide a basis for fault management and maintenance decisions;

[0095] Enhance model stability: By evaluating the efficiency under different conditions, it is possible to identify the stability and reliability of the model in various environments and ensure consistency in practical applications;

[0096] In summary, the analysis of model efficiency evaluation coefficients provides important theoretical basis and practical guidance for selecting the optimal sampling frequency set.

[0097] The specific method for obtaining the model efficiency evaluation coefficient is as follows:

[0098] Get the time complexity of the model;

[0099] Record the model's response time each time it receives a command, and calculate the average response time;

[0100] The ratio of time complexity to average response time is used as the model processing efficiency;

[0101] Based on the known model space complexity and the number of model iterations, the model space efficiency is obtained by square root calculation;

[0102] The model efficiency evaluation coefficient is calculated based on the model space efficiency and model processing efficiency.

[0103] The specific calculation formula of the model efficiency evaluation coefficient is as follows:

[0104]

[0105] In the formula, is the model efficiency evaluation coefficient, is the number of responses recorded, is the jth response time, is the model space complexity, is the number of model iterations, and SJ is the time complexity of the model.

[0106] The obtained modal information overlap coefficient, fault prediction deviation coefficient and model efficiency evaluation coefficient are used to establish a resolution impact evaluation model to generate a sampling frequency evaluation coefficient;

[0107] The specific calculation formula of the sampling frequency evaluation coefficient is as follows:

[0108]

[0109] In the formula, is the sampling frequency evaluation coefficient, is the preset proportional coefficient of the modal information overlap coefficient, is the preset proportional coefficient of the fault prediction deviation coefficient, is the preset proportional coefficient of the model efficiency evaluation coefficient, is the modal information overlap coefficient, is the fault prediction deviation coefficient, is the model efficiency evaluation coefficient.

[0110] S5, constructing several sampling frequency-assessment effect curves based on the output of the sampling frequency assessment model and the corresponding sampling frequency set, and performing data analysis on the several sampling frequency-assessment effect curves to obtain an improved sampling frequency set, and performing fault prediction based on the improved sampling frequency set.

[0111] The output of the sampling frequency evaluation model and the corresponding sampling frequency set are used to construct several sampling frequency-evaluation effect curves, specifically:

[0112] Obtaining sampling frequency evaluation coefficients of outputs of sampling frequency sets under different settings from sampling frequency evaluation models under different sampling frequency sets;

[0113] A rectangular coordinate system is established, with the horizontal axis representing the sampling frequency set and the vertical axis representing the sampling frequency evaluation coefficient;

[0114] Mark the sampling frequency evaluation coefficient corresponding to each sampling frequency set on the rectangular coordinate system to form data points;

[0115] Connect the data points to form a sampling frequency-evaluation effect curve;

[0116] Select one modal signal from the multi-mode as the direction of the horizontal axis, and construct a sampling frequency-evaluation effect curve with the same number of modal signals.

[0117] The above mentioned method also performs data analysis on several sampling frequency-evaluation effect curves to obtain an improved sampling frequency set, specifically:

[0118] The peak value of each sampling frequency-evaluation effect curve is quadratically derived, and the sampling frequency set corresponding to the peak point of the sampling frequency-evaluation effect curve with the largest quadratic derivative result is used as the improved sampling frequency set, and fault prediction is performed based on the improved sampling frequency set.

[0119] This embodiment optimizes the fault prediction process by establishing a sampling frequency evaluation model. First, the method achieves a comprehensive analysis of multimodal signals by obtaining the modal information overlap coefficient, fault prediction deviation coefficient and model efficiency evaluation coefficient of each set of sampling frequencies. This process not only identifies the information redundancy between modal signals, but also reveals the unique contribution of each mode to fault prediction, ensuring the optimal performance of the model at different frequencies. Secondly, the analysis of the modal information overlap coefficient helps to reduce data redundancy, improve data processing efficiency, and thus improve prediction accuracy. At the same time, the evaluation of the fault prediction deviation coefficient provides intuitive feedback on model performance, helping researchers optimize sampling strategies and ensure consistency between prediction results and actual conditions. In addition, the model efficiency evaluation coefficient comprehensively considers the response time and resource consumption of the model, making the fault detection process more efficient.

[0120] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters in the formula are set by technicians in this field according to actual conditions.

[0121] The above embodiments may be implemented in whole or in part by software, hardware, firmware or any other combination. When implemented by software, the above embodiments may be implemented in whole or in part in the form of a computer program product.

[0122] Those of ordinary skill in the art will appreciate that the modules and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0123] In addition, each functional module in each embodiment of the present application may be integrated into one processing module, or each module may exist physically separately, or two or more modules may be integrated into one module.

[0124] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

[0125] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. Multimodal fault diagnosis and early warning method, It is characterized in that The steps include: Performing spectrum analysis on several modal signals used for fault prediction of historical substation main equipment to obtain a first frequency set, and using G times the frequency of the first frequency set as the initial sampling frequency of each modal signal to obtain an initial sampling frequency set; Perform several random perturbations on the initial sampling frequency set to obtain several groups of different sampling frequency sets; The modal signals corresponding to the main substation equipment are sampled according to the sampling frequency set to obtain multi-modal data, and fault prediction is performed based on ST-GCN; Obtain the sampling overlap data and corresponding fault prediction data corresponding to each sampling frequency set, and build a sampling frequency evaluation model; The output of the sampling frequency evaluation model and the corresponding sampling frequency set are used to construct several sampling frequency-evaluation effect curves, and data analysis is performed on the several sampling frequency-evaluation effect curves to obtain an improved sampling frequency set, and fault prediction is performed based on the improved sampling frequency set; The output of the sampling frequency evaluation model and the corresponding sampling frequency set are used to construct several sampling frequency-evaluation effect curves, specifically: Obtaining sampling frequency evaluation coefficients of outputs of sampling frequency sets under different settings from sampling frequency evaluation models under different sampling frequency sets; A rectangular coordinate system is established, with the horizontal axis representing the sampling frequency set and the vertical axis representing the sampling frequency evaluation coefficient; Mark the sampling frequency evaluation coefficient corresponding to each sampling frequency set on the rectangular coordinate system to form data points; Connect the data points to form a sampling frequency-evaluation effect curve; Select one modal signal from the multi-modal signals as the direction of the horizontal axis, and construct a sampling frequency-evaluation effect curve with the same number as the modal signal; The above mentioned method also performs data analysis on several sampling frequency-evaluation effect curves to obtain an improved sampling frequency set, specifically: The peak value of each sampling frequency-evaluation effect curve is quadratically derived, and the sampling frequency set corresponding to the peak point of the sampling frequency-evaluation effect curve with the largest quadratic derivative result is used as the improved sampling frequency set, and fault prediction is performed based on the improved sampling frequency set.

2. The multi-modal fault diagnosis and early warning method according to claim 1, It is characterized in that The plurality of modal signals include current signals, voltage signals, temperature signals, vibration signals and acoustic signals; The first frequency set is a set of frequencies obtained by spectrum analysis of all modal signals; the initial sampling frequency set is composed of a vector consisting of G times the first frequency set and the corresponding modal signal category; The sampling overlap data includes a modal information overlap coefficient, and the fault prediction data includes a fault prediction deviation coefficient and a model efficiency evaluation coefficient.

3. The multi-modal fault diagnosis and early warning method according to claim 2, It is characterized in that The method for obtaining the several groups of different sampling frequency sets is as follows: Taking the initial sampling frequency set as the basis, each modal signal corresponds to a fixed initial sampling frequency; Based on the random number generator, random perturbations are introduced into the sampling frequency of each modal signal; The preset upper and lower limits of the sampling frequency are obtained, and a random offset is generated within the upper and lower limits of the sampling frequency, so that the sampling frequency of each mode fluctuates within a certain range, and several groups of different sampling frequency sets are obtained.

4. The multi-modal fault diagnosis and early warning method according to claim 3, It is characterized in that The specific method for fault prediction based on ST-GCN is as follows: According to different sampling frequency sets, each modal signal of the main substation equipment is synchronously collected to obtain multimodal data containing time and space information, and the multimodal data is stored in the form of a time series; The nodes of each modal data in the multimodal data are used as nodes in the hidden danger graph structure of the main substation equipment, and the spatiotemporal correlation between different modal signals is used as the edge between the nodes; The sampling value of the modal signal is used as the signal feature, and the node feature is represented by the feature matrix. Based on ST-GCN, the feature of each node is aggregated with the features of its adjacent nodes to obtain the aggregated hidden danger graph structure of the substation main equipment. The feature of each node is aggregated with the features of its adjacent nodes, specifically: The characteristic matrix of each node is weightedly summed with the characteristic matrix of its neighboring nodes, and the result of the weighted summation is linearly transformed based on a preset hidden danger aggregation formula to obtain the aggregated hidden danger graph structure of the main substation equipment; The hidden danger graph structure of the main substation equipment is aggregated multiple times to obtain the final hidden danger graph structure of the main substation equipment. Based on the preset classification regression formula, the hidden danger graph structure of the main substation equipment is processed to output the fault prediction result of the equipment.

5. The multi-modal fault diagnosis and early warning method according to claim 4, It is characterized in that The specific method for obtaining the modal information overlap coefficient is as follows: Extract characteristic data from multi-modal signals of main substation equipment; Perform denoising and normalization on the modal signal, and calculate the mean and variance of the modal signal; For each pair of modal signals, the mean and variance are compared pairwise; Obtain the probability distribution function of the modal signal, calculate the autoentropy of the modal signal, and calculate the cross entropy between each pair of modal signals based on the autoentropy of each modal signal; The minimum value of the comparison results of the two modes in terms of mean and variance is taken, and these minimum values ​​are added together and combined with the cross entropy between the modal signals to calculate the modal information overlap coefficient.

6. The multi-modal fault diagnosis and early warning method according to claim 5, It is characterized in that The specific method for obtaining the fault prediction deviation coefficient is as follows: Get the prediction results corresponding to the time interval P within the time L for a specific sampling frequency; Construct a prediction result waveform based on the prediction result, perform Fourier transform, convert the prediction result waveform into the frequency domain, and compare it with the actual fault result; In the frequency domain, the peak and valley values ​​of the difference between the predicted result waveform and the actual fault result are calculated, which represent the peak and valley positions of the predicted result waveform in the frequency spectrum respectively; The fault prediction deviation coefficient is calculated by the square of the Euclidean distance between the peak and the valley.

7. The multi-modal fault diagnosis and early warning method according to claim 6, It is characterized in that The specific method for obtaining the model efficiency evaluation coefficient is as follows: Get the time complexity of the model; Record the model's response time each time it receives a command, and calculate the average response time; The ratio of time complexity to average response time is used as the model processing efficiency; Based on the known model space complexity and the number of model iterations, the model space efficiency is obtained by square root calculation; The model efficiency evaluation coefficient is calculated based on the model space efficiency and model processing efficiency.

8. The multi-modal fault diagnosis and early warning method according to claim 7, It is characterized in that The specific calculation formula of the self-entropy of the modal signal is as follows: The specific calculation formula of the cross entropy between the modal signals is as follows: The specific calculation formula of the modal information overlap coefficient is as follows: In the formula, is the modal signal The self-entropy, is the modal signal value, is the modal signal and The cross entropy of is the number of combinations of selecting 2 from n modes, is the modal information overlap coefficient, is the total number of modal signal types, is the minimum value of the mean difference and the variance difference, is the sum of the mean and variance differences.

Images