Intelligent monitoring method and system for power distribution cabinet

Through small sample learning and edge dynamic neural network technology, combined with multimodal data fusion, the problems of inaccurate fault identification, insufficient dynamic regulation and real-time response lag in traditional distribution cabinet monitoring systems are solved, and the efficiency of distribution cabinets is achieved. Accurate and intelligent monitoring improves operational safety and management efficiency.

CN119051268AInactive Publication Date: 2024-11-29GUANGZHOU SHAOCHUANG ELECTRONIC TECHNOLOGY CO LTD
View PDF 0 Cites 14 Cited by

Patent Information

Application Number
CN202411211879.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-30
Publication Date
2024-11-29
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

Traditional distribution cabinet monitoring systems lack accurate fault identification capabilities, insufficient dynamic regulation capabilities, delayed real-time response, and risk of data transmission delays and single point failures, making it difficult to adapt to complex power load changes and multi-source data integration.

Method used

Using small sample learning, edge dynamic neural network and multimodal data fusion technology, dynamic fault prediction model and load regulation strategy are constructed through asymmetric differential compression, adaptive support vector machines, multiple time scale convolution and recursive neural networks to realize Real-time fault detection and load regulation, and improve system flexibility and reliability through edge computing and multi-model collaborative updates.

Benefits of technology

It improves the accuracy of fault detection and dynamic control capabilities of the distribution cabinet, shortens the fault response time, reduces the risk of single point of faults, enhances the adaptability and intelligent management capabilities of the system, and improves operational safety and management efficiency.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119051268A_ABST
    Figure CN119051268A_ABST
Patent Text Reader

Abstract

The invention provides an intelligent monitoring method and system for a power distribution cabinet, and the method comprises the steps: obtaining the original data of a sensor, carrying out the preprocessing of the original data, constructing a feature vector, and enabling the input feature vector to serve as the input of a self-adaptive small sample learning model; feature fusion: performing time sequence feature extraction on the feature vectors to obtain final time sequence feature representation, and taking the final time sequence feature representation as input of a dynamic fault prediction model; a dynamic load condition real-time monitoring and adjusting model is constructed according to the obtained time sequence feature representation and the prediction result of the dynamic fault prediction model, and real-time execution is performed through an edge calculation unit; fault signals are detected, corresponding response strategies are automatically generated according to fault classification, and finally normal operation of the system is recovered through a dynamic adjustment gradual recovery mechanism. According to the invention, the system is updated online by adopting the adaptive learning rate, and long-term stable operation of the system is realized, so that efficient monitoring of the power distribution cabinet is completed.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of intelligent monitoring, and in particular relates to an intelligent monitoring method and system for a power distribution cabinet. Background Art

[0002] As an important basic equipment in the power system, the distribution cabinet is widely used in various power transmission and distribution scenarios, and undertakes the distribution and protection functions of electric energy. With the continuous growth of power demand and the complexity of the power system, the reliability and intelligent monitoring of the distribution cabinet have become particularly important. The traditional distribution cabinet monitoring system mainly relies on simple electrical parameter detection, such as voltage, current, temperature, etc., and triggers alarms and protection mechanisms through preset thresholds. Although these systems can realize basic monitoring functions to a certain extent, they have many limitations.

[0003] First, traditional monitoring systems lack the ability to accurately identify faults. In actual operation, the types of faults in distribution cabinets are complex and diverse, especially some rare or early faults are often difficult to effectively detect through simple threshold settings. This detection method may not only lead to missed reports, but also cause false alarms, further increasing the difficulty of operation and maintenance. In addition, the data processing and decision-making capabilities of traditional systems rely on central servers. After data collection, they need to be transmitted to remote servers for analysis. Such an architecture cannot guarantee sufficient response speed when facing sudden failures. At the same time, the centralized processing mode that relies on the central server is also prone to single point failure risks. Once the central server fails, it will directly affect the operation of the entire monitoring system.

[0004] Secondly, with the increasing complexity of the power distribution system, the volatility and diversity of power loads have put forward higher requirements for the intelligent control of the distribution cabinet. The traditional static monitoring mode is difficult to adapt to the needs of such dynamic changes and cannot adjust the operating status of the distribution cabinet according to the real-time situation, which leads to resource waste and reduced operating efficiency. During peak load periods, traditional systems are prone to overload or overload, while during low load periods, resources may be idle, which not only affects the overall stability of the power system, but also increases operating costs.

[0005] Furthermore, traditional fault handling methods mostly rely on manual inspection and post-maintenance, making it difficult to achieve predictive maintenance of faults. Although some systems have achieved preliminary automation through simple rule engines, these methods are often based on historical experience data, lack sufficient flexibility and intelligence, and cannot adapt to the rapidly changing actual environment. With the rise of Internet of Things technology, although solutions based on remote monitoring have emerged, these solutions also have problems with data transmission delays and poor real-time performance. In addition, traditional systems are relatively simple in terms of data integration and analysis, and it is difficult to effectively utilize multi-source data from different sensors, resulting in difficulty in making comprehensive analysis and accurate decisions in the face of complex environments. Summary of the invention

[0006] The purpose of the present invention is to propose an intelligent monitoring method and system for distribution cabinets. By integrating small sample learning, edge dynamic neural network and multimodal data fusion technology, an efficient, accurate and intelligent distribution cabinet monitoring solution is provided. The problems of inaccurate rare fault detection, insufficient dynamic control capability and delayed real-time response in traditional systems are successfully overcome, and the operation safety and management efficiency of distribution cabinets are greatly improved.

[0007] In order to achieve the above object, a first aspect of the present invention provides an intelligent monitoring method for a power distribution cabinet, the method comprising:

[0008] S1. Deploy sensors in the power distribution cabinet to collect raw data S i (t), for the original data S i (t) Preprocessing to obtain data S″′ i (t), and then use asymmetric differential compression method to compress the data S″′ i (t) transmit and store;

[0009] S2, for data S″′ i (t) Perform feature extraction and input the extracted features into the feature vector X i (t), according to the input feature vector X i (t) Establishing an adaptive small sample learning model of an adaptive support vector machine based on meta-learning to classify the identified faults, designing an adaptive regularization term for the distribution cabinet environment and adding it to the training of the adaptive small sample learning model, using the trained adaptive small sample learning model to perform online fault detection on real-time data, and designing a feedback optimization mechanism to regularly perform adaptive adjustments on the adaptive small sample learning model;

[0010] S3, combining the results of online fault detection with data S″′ i (t) Design a method based on adaptive correlation weighted fusion to fuse the sensor features according to the correlation weights and obtain the fused feature vector X f (t), and then use the method based on multiple time scale convolution to transform the feature vector X f (t) Extract time series features and obtain the final time series feature representation Z f (t), then a dynamic fault prediction model is proposed by constructing a recurrent neural network structure based on time-dependent enhancement, and the dynamic fault prediction model is adaptively updated online;

[0011] S4. Express Z according to the obtained time series characteristics f(t) and the prediction results of the dynamic fault prediction model are used to build a dynamic load condition real-time monitoring and adjustment model, and the dynamic load condition real-time monitoring and adjustment model is used to perform adaptive load control on the load state vector L(t), and the edge computing unit executes it in real time, and the control strategy is adaptively updated and optimized regularly;

[0012] S5. According to the load state vector L(t) and the adaptive load control strategy, an adaptive threshold fault detection algorithm is constructed to detect fault signals, and a feature-enhanced multimodal fault classification algorithm is designed to classify faults. According to the fault classification, a corresponding response strategy is automatically generated, and finally a dynamic adjustment step-by-step recovery mechanism is used to restore the normal operation of the system.

[0013] S6. Adopt adaptive learning rate to update steps S1-S5 online, and adopt rolling window retraining mechanism to retrain them regularly. Design multi-model collaborative update mechanism for collaborative update, and design long-term trend monitoring and global adjustment mechanism to perform long-term intelligent monitoring of distribution cabinets.

[0014] Preferably, the asymmetric differential compression method performs different processing on abnormal points in differential data, and the processing is as follows:

[0015]

[0016] Among them, ΔS i (t) is the updated outlier point, Δ syn (t) is the standard symmetric difference compression result, ∈ is a threshold used to identify the normal fluctuation range, κ is the adjustment factor, R(t) is the regression residual based on the historical data of the distribution cabinet, and Δt is the time difference between adjacent data points.

[0017] Preferably, the adaptive small sample learning model uses an adaptive support vector machine based on meta-learning for classification, wherein the adaptive small sample learning model loss function L(w,λ,t) is expressed as follows:

[0018]

[0019] Among them, w is the weight vector of the classifier, is the transpose of w, X i (t) is the input feature vector of the i-th sample, y i is the label of sample i, y i ∈{-1,1}, C is the penalty parameter in the loss function, λ is the adjustment factor of the adaptive regularization term, ρ j (t) is the adaptive coefficient, n is the number of sensors, b is the bias, m is the number of distribution cabinets, and j is the current distribution cabinet;

[0020] Use the trained adaptive small sample learning model to perform online fault detection on real-time data, where for each new data point S″′ i (t), according to the input feature vector X of the i-th sample i (t), calculate the output of the classification function, expressed as follows:

[0021]

[0022] if It is determined that a fault signal occurs at the current time t.

[0023] Preferably, the adaptive adjustment means that every time T r , remix the collected fault samples and normal samples for training, and dynamically adjust the adaptive coefficient ρ in the regularization term j (t), the adjustment formula is as follows:

[0024]

[0025] Where η is the learning rate, is the predicted label of the jth sample under the current adaptive small sample learning model, y j is the true label of sample j, n s is the number of new samples in time step t.

[0026] Preferably, the adaptive correlation weighted fusion represents that the features of multiple sensors are correlated and analyzed to calculate different types of sensor data X i (t) and X j The dynamic correlation weight ω of (t) ij (t), the formula is as follows:

[0027]

[0028] Among them, ω ij (t) represents the correlation weight between sensors i and j at time t, Σ ij is the covariance matrix between features, which is used to measure the correlation of sensor data, X i (t) represents the input feature vector of the i-th sample, X j (t) Input feature vector of the jth sample; The features of multiple sensors are weighted and fused according to the correlation weights to obtain the fused feature vector X f (t), expressed as follows:

[0029]

[0030] Where n is the number of sensors.

[0031] Preferably, the state update formula of the recursive neural network based on time-series dependency enhancement is:

[0032] h t =σ(W h ·Z f (t)+U h ·h t-1 +b h )

[0033] Among them, h t is the hidden state at time t, h t-1 is the hidden state at time t-1, W h and U h are the weight matrices of the input and hidden layers, respectively, and b h is the bias vector and σ is the activation function.

[0034] Preferably, the dynamic load real-time monitoring and adjustment model is constructed as follows:

[0035] First, the key power parameters are extracted to obtain the time series P(t), which is then compared with the time series feature representation Z f (t) to establish the load state representation vector L(t), which is expressed as follows:

[0036] L(t)=P(t)⊙Z f (t)

[0037] Where P(t) is the power parameter sequence, including the characteristics of current and voltage, ⊙ represents the element-by-element multiplication operation, and L(t) is the fused load state vector;

[0038] The control strategy is designed according to the load state vector L(t), which is expressed as follows:

[0039]

[0040] Where C(t) is the load regulation strategy vector at time t, λ s is a regularization parameter that controls the smoothness of regulation and is used to limit the severity of regulation changes. 2 represents the L2 norm;

[0041] After executing the control strategy, the real-time monitored power parameters are fed back for detection. If the deviation exceeds the set threshold, the strategy is fine-tuned immediately. The feedback adjustment formula is:

[0042]

[0043] Among them, C′(t) is the strategy vector after immediate feedback adjustment, P(t) is the actual monitored power parameter, and P′(t) is the power parameter expected under the current strategy. is the adjustment coefficient, which controls the amplitude of feedback adjustment.

[0044] Preferably, the feature-enhanced multi-modal fault classification algorithm specifically includes:

[0045] First, generate the fault feature vector F(t), which contains the time series feature representation Z f (t) and the current load characteristic L(t), the fault characteristic vector is defined as:

[0046] F(t)=Z f (t)⊙L(t)

[0047] Among them, ⊙ is the element-by-element multiplication operation, and then the fault classifier uses a multimodal classification decision function based on an enhanced regularization term to perform fault classification, which is expressed as follows:

[0048]

[0049] in, is the fault type identification result, f i (F(t)) is a classification function based on multimodal features, α i (t) is the classification weight, γ s is the coefficient of the enhanced regularization term, F j (t) is the jth eigenvalue, F j (t-1) is the jth eigenvalue at time t-1, k s is the number of feature components, ∥∥ 2 represents the L2 norm, m is the number of features in fault classification;

[0050] The automatic generation of corresponding response strategies according to fault classification specifically includes:

[0051] The optimization objective function R(t) of the designed response strategy is:

[0052]

[0053] Among them, W r is a weight matrix associated with the fault response strategy, weighing different response objectives. is a regularization parameter used to smooth the response policy changes, η s is the coefficient that controls the complexity of the response strategy, is the complexity function of the response strategy, R represents the response strategy variable that needs to be optimized, and R(t-1) represents the response strategy at the previous time point t-1, which is used as a reference point for the current strategy optimization.

[0054] Preferably, the multi-model collaborative update mechanism includes:

[0055] Through a joint loss function L c The weights of multiple models are updated synchronously, and the joint loss function is defined as:

[0056]

[0057] Among them, w i is the weight vector of the i-th model, w1 corresponds to the fault classifier model, w2 corresponds to the control strategy generator model, w3 corresponds to the recovery mechanism model, j represents the index between models, ∥∥ 2 represents the L2 norm, L i (w i ) is the loss function of the i-th model, α i is the weight of the loss function of each model, λ c is the weight of the collaborative regularization term, and the weight update formula of the model is:

[0058]

[0059] Among them, w i (t+1) represents the weight vector of the i-th model at time t+1, Represents partial differential.

[0060] In a second aspect of the present invention, an intelligent monitoring system for a power distribution cabinet is provided, the system comprising:

[0061] Data acquisition module, used to deploy sensors in the power distribution cabinet to collect raw data S i (t), for the original data S i (t) Preprocessing to obtain data S″′ i (t), and then use asymmetric differential compression method to compress the data S″′ i (t) transmit and store;

[0062] Fault detection module, used for data S″′ i (t) Perform feature extraction and input the extracted features into the feature vector X i (t), according to the input feature vector X i (t) Establishing an adaptive small sample learning model of an adaptive support vector machine based on meta-learning to classify the identified faults, designing an adaptive regularization term for the distribution cabinet environment and adding it to the training of the adaptive small sample learning model, using the trained adaptive small sample learning model to perform online fault detection on real-time data, and designing a feedback optimization mechanism to regularly perform adaptive adjustments on the adaptive small sample learning model;

[0063] Data fusion and fault prediction module, used to combine data S″′ based on the results of online fault detection i(t) Design a method based on adaptive correlation weighted fusion to fuse the sensor features according to the correlation weights and obtain the fused feature vector X f (t), and then use the method based on multiple time scale convolution to transform the feature vector X f (t) Extract time series features and obtain the final time series feature representation Z f (t), then a dynamic fault prediction model is proposed by constructing a recurrent neural network structure based on time-dependent enhancement, and the dynamic fault prediction model is adaptively updated online;

[0064] The strategy execution module is used to represent Z according to the obtained time series characteristics. f (t) and the prediction results of the dynamic fault prediction model are used to build a dynamic load condition real-time monitoring and adjustment model, and the dynamic load condition real-time monitoring and adjustment model is used to perform adaptive load control on the load state vector L(t), and the edge computing unit executes it in real time, and the control strategy is adaptively updated and optimized regularly;

[0065] The local control module is used to construct an adaptive threshold fault detection algorithm to detect fault signals based on the load state vector L(t) and the adaptive load control strategy, and to design a feature-enhanced multi-modal fault classification algorithm to classify faults. It automatically generates corresponding response strategies based on the fault classification, and finally uses a dynamic adjustment and gradual recovery mechanism to restore the normal operation of the system.

[0066] The system optimization module is used to update the system online using an adaptive learning rate, and to retrain it regularly using a rolling window retraining mechanism. A multi-model collaborative update mechanism is designed for collaborative updating. A long-term trend monitoring and global adjustment mechanism is designed for long-term intelligent monitoring of the distribution cabinet.

[0067] The beneficial technical effects of the present invention are at least as follows:

[0068] (1) The present invention introduces a small sample learning algorithm, which can effectively identify rare or new fault types by extracting features from a small amount of labeled data, especially when data is scarce, and can still ensure the accuracy of detection. The application of this technology greatly improves the coverage and sensitivity of the distribution cabinet monitoring system in early fault detection, and reduces the occurrence of missed reports and false alarms.

[0069] (2) The present invention designs an edge dynamic neural network module, which can adaptively adjust the monitoring and control strategies according to the real-time operation data of the power distribution cabinet to ensure that the system can still operate stably under high load or abnormal conditions. By implementing dynamic neural networks on edge devices, the present invention not only ensures real-time response, but also improves the flexibility and adaptability of the system, effectively avoiding delays and single-point failure risks in the centralized processing mode.

[0070] (3) The present invention integrates edge AI technology, enabling the system to complete data processing and decision-making in a localized environment, greatly shortening the fault response time. By deploying edge computing units on-site in the power distribution cabinet, the system can still operate independently and handle faults when the network connection is unstable or disconnected, thereby improving the reliability and autonomy of the system.

[0071] (4) The present invention utilizes multimodal data fusion technology to comprehensively analyze multi-source data from different sensors, effectively improving the accuracy of fault prediction, optimizing maintenance plans, and enhancing the intelligent management capabilities of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] The present invention is further described using the accompanying drawings, but the embodiments in the accompanying drawings do not constitute any limitation to the present invention. A person skilled in the art can obtain other drawings based on the following drawings without creative work.

[0073] Figure 1 The present invention is a flow chart of an intelligent monitoring method for a power distribution cabinet.

[0074] Figure 2 The present invention is a framework diagram of an intelligent monitoring system for a power distribution cabinet. DETAILED DESCRIPTION

[0075] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0076] like Figure 1 As shown, an embodiment of the present invention provides an intelligent monitoring method for a power distribution cabinet, the method comprising:

[0077] S1. Deploy sensors in the distribution cabinet to collect raw data S i (t), for the original data S i (t) Preprocessing to obtain data S″′ i (t), and then use asymmetric differential compression method to compress the data S″′i (t) Transmit and store.

[0078] The sensors include voltage sensors, temperature sensors, humidity sensors, vibration sensors and current sensors. The data recorded by each sensor is expressed as S i (t), where i represents the sensor type and t represents the sampling time. In order to ensure high-quality data collection, the sensor is initialized and calibrated to ensure that the readings of all sensors are consistent under the same physical environment. Specifically, the voltage sensor is used to measure the voltage of each node in the distribution cabinet, the temperature sensor is used to monitor the internal temperature of the distribution cabinet, the humidity sensor monitors the humidity level to prevent humidity from causing electrical failures, the vibration sensor is used to monitor the mechanical vibration of the cabinet and predict structural problems, and the current sensor is used to monitor the current load and detect load abnormalities.

[0079] Furthermore, in said S1, the specific steps are:

[0080] S101, a denoising algorithm based on a multi-level threshold is used to collect the original data S i (t) is subjected to denoising, and the specific formula is as follows:

[0081] θ i (t) = μ i (t)+α·σ i (t)+γ·e -β·Δt

[0082] Among them, μ i (t) is the data S i (t) in the window W(t), the formula is:

[0083]

[0084] Among them, σ i (t) is the standard deviation of the data, α is the noise sensitivity adjustment factor, γ is the dynamic noise adjustment term, β is the parameter that controls the exponential decay rate, and Δt is the time difference between adjacent data points; if a data point S i (t) satisfies |S i (t)-μ i (t)|>θ i (t), the point is determined to be a noise point and replaced by the mean of the data at the adjacent time points;

[0085] S102, the data S′ after denoising i (t), adaptive nonlinear normalization is performed, and the normalization formula is:

[0086]

[0087] Among them, λi It is a nonlinear factor adjusted for different sensor data distribution characteristics, by adaptively adjusting λ i , so that the normalized data S″ i (t) distribute more evenly in the interval [0,1], especially appropriately amplify abnormal data points for subsequent detection;

[0088] S103, using a cross-coupling interpolation algorithm to time-align the data, and selecting the optimal interpolation point by analyzing the physical correlation between data from different sensors. The formula is as follows:

[0089]

[0090] Where: η ij It is data S i (t) at its historical data point t j The interpolation coefficient, ρ ik Other sensors S k (t) and S i The cross-coupling coefficient between (t) is used to introduce the data influence of related sensors, thereby enhancing the accuracy of the interpolation result;

[0091] Specifically, cross-coupling interpolation not only interpolates data from a single sensor, but also takes into account the influence of other sensors, making it particularly suitable for processing complex environmental data coupled with multiple sensors;

[0092] S104, asymmetric differential compression performs different processing on abnormal points in differential data, and the processing is as follows:

[0093]

[0094] Among them, ΔS i (t) is the updated outlier point, Δ sym (t) is the standard symmetric difference compression result, ∈ is a threshold used to identify the normal fluctuation range, κ is the adjustment factor, R(t) is the regression residual based on the historical data of the distribution cabinet, and Δt is the time difference between adjacent data points.

[0095] The asymmetric differential compression retains more critical data related to the fault, while reducing the amount of data transmitted under normal circumstances, ensuring efficient data compression and transmission;

[0096] S105, the data ΔS after the above processing i (t) is stored in the cache of the edge computing unit and adaptively batch-transmitted according to bandwidth conditions, as shown below:

[0097]

[0098] Among them, W(t) is the transmission window size at the current time, γ is a constant based on network bandwidth regulation, B(t) is the current network bandwidth, and δ is a parameter that controls transmission stability and delay. By adaptively adjusting the transmission window, the system can maintain efficient data transmission under bandwidth fluctuations and avoid network congestion. This coherent process ensures the quality and real-time nature of the data and provides accurate input for intelligent monitoring of the distribution cabinet. Through the sophisticated data collection and processing process, solid data support is provided for the efficient operation of the small sample learning model, which is the key to realizing intelligent monitoring of the distribution cabinet.

[0099] S2. Extract features from the data and combine the extracted features into input feature vectors. According to the input feature vectors, an adaptive small sample learning model of an adaptive support vector machine based on meta-learning is established to classify the identified faults. At the same time, an adaptive regularization term for the distribution cabinet environment is designed and added to the training of the adaptive small sample learning model. The trained adaptive small sample learning model is used to perform online fault detection on real-time data, and a feedback optimization mechanism is designed to regularly adaptively adjust the adaptive small sample learning model.

[0100] Feature extraction is performed on the acquired high-quality data, and small sample learning technology is used to specifically identify rare or new faults, which makes up for the shortcomings of traditional fault detection in the case of data scarcity. The detection results are passed to the next step as an important basis for fault trend analysis.

[0101] Specifically, in S2, the specific steps are:

[0102] S201, using a sliding feature extraction method based on a time window to extract features from each sensor data, setting the time window length to T w , for each sensor data sequence S″′ i (t), in the time window [tT w ,t] to extract the following features:

[0103] Mean:

[0104] variance:

[0105] Kurtosis:

[0106] The above features constitute the input feature vector Fault detection model for the next step;

[0107] S202, the adaptive small sample learning model uses an adaptive support vector machine based on meta-learning for classification, and the adaptive small sample learning model loss function L(w,λ,t) is expressed as follows:

[0108]

[0109] Among them, w is the weight vector of the classifier, X i (t) is the input feature vector of the i-th sample, y i is the label of the sample, y i ∈{-1,1}, C is the penalty parameter in the loss function, λ is the adjustment factor of the adaptive regularization term, ρ j (t) is the adaptive coefficient, n is the number of sensors, which is dynamically adjusted over time to target different failure modes of the distribution cabinet; b is the bias, m is the number of distribution cabinets, j is the current distribution cabinet, R j (w,t) is a regularization term designed specifically for rare faults, usually containing weight sparsity control or a measure of the difference between fault modes;

[0110] S203, using the trained adaptive support vector machine model to perform online fault detection on real-time data, for each new data point S″′ i (t), according to the extracted feature vector X i (t), calculate the output of the classification function:

[0111]

[0112] if It is determined that a fault signal has occurred at the current time t. In order to improve the robustness of detection, when a fault signal is detected, the sample and its neighboring samples are saved and incremental training and update are performed to ensure that the adaptive support vector machine model can dynamically adapt to new fault modes;

[0113] S204, adaptive adjustment means that every time T r , remix the collected fault samples and normal samples for training, and dynamically adjust the adaptive coefficient ρ in the regularization term j (t), the adjustment formula is as follows:

[0114]

[0115] Where η is the learning rate, is the predicted label of the jth sample under the current adaptive small sample learning model, y j is the true label of the sample, n s is the number of new samples in time step t.

[0116] S3, combining the results of online fault detection with data S″′ i (t) Design a method based on adaptive correlation weighted fusion to fuse the sensor features according to the correlation weights and obtain the fused feature vector Xf (t), and then use the method based on multiple time scale convolution to transform the feature vector X f (t) Extract time series features and obtain the final time series feature representation Z f (t), then a dynamic fault prediction model is proposed based on the structure of a recursive neural network enhanced by timing dependency, and the dynamic fault prediction model is adaptively updated online.

[0117] Further integration of multiple sensor data (including environmental data), comprehensive analysis through multimodal deep learning models, mining potential correlations between data, and predicting fault trends will not only deepen the understanding of faults, but also provide more accurate prediction information for the next step of real-time regulation.

[0118] Specifically, the specific steps of S3 include:

[0119] S301: After the first two steps, a variety of pre-processed sensor data S″′ has been obtained. i (t) and its related features X i (t). In order to more accurately predict the potential failure of the power distribution cabinet, it is first necessary to perform correlation analysis on these different types of sensor data and correlate the characteristics of multiple sensors.

[0120] Specifically, for different types of sensor data X i (t) and X j (t), calculate the dynamic correlation weight ω between them ij (t), the formula is as follows:

[0121]

[0122] Among them, ω ij (t) represents the correlation weight between sensors i and j at time t, Σ ij is the covariance matrix between features, which is used to measure the correlation of sensor data, X i (t) represents the input feature vector of the i-th sample, X j (t) Input feature vector of the jth sample; The features of multiple sensors are weighted and fused according to the correlation weights to obtain the fused feature vector X f (t), expressed as follows:

[0123]

[0124] Where n is the number of sensors. The fused feature vector can better capture the implicit relationship between different sensors and provide a global and more relevant feature basis for subsequent fault prediction.

[0125] S302, using the fused feature vector X f (t) to extract time series features. A time series feature extraction method based on "multiple time scale convolution" is proposed. Specifically, multi-layer convolution operations are used to extract multiple time scale information in the feature sequence to cope with the time characteristics of different fault modes in the distribution cabinet. For each time step t, the convolution operation is defined as follows:

[0126]

[0127] Among them, Z f (t) is the temporal feature representation extracted after the convolution operation, W k is the convolution kernel matrix, k represents different time scales, * represents the convolution operation, and K is the number of convolution kernels. Through the convolution operation of multiple time scales, the fusion feature vector X is extracted f (t) to obtain the final temporal feature representation Z f (t);

[0128] S303, the dynamic fault prediction model is specially designed for the multimodal data and time series characteristics of the distribution cabinet, and the state update formula of the recursive neural network based on time series dependency enhancement is:

[0129] h t =σ(W h ·Z f (t)+U h ·h t-1 +b h )

[0130] Among them, h t is the hidden state at time t, W h and U h are the weight matrices of the input and hidden layers, respectively, and b h is the bias vector, σ is the activation function;

[0131] By dynamically updating the hidden state h t , captures the temporal dependency of the fusion features at the current moment, and uses the final hidden state h t Perform fault prediction. If the prediction result exceeds the set threshold, it is marked as a potential fault and an early warning mechanism is triggered in the system;

[0132] S304. Design an adaptive tuning and online update mechanism to ensure the long-term adaptability of the model in practical applications. By monitoring the prediction accuracy during system operation, dynamically adjust the parameters of the time-dependent enhanced recursive neural network model, and calculate the prediction error e t , fine-tune the weights of the model, the formula is as follows:

[0133]

[0134] Among them, e t is the prediction error at the current moment, and η is the learning rate. Through this online update mechanism, the model can be adaptively tuned when facing different operating conditions and emerging fault modes, ensuring that it still maintains efficient fault prediction capabilities during long-term operation.

[0135] S4. Express Z according to the obtained time series characteristics f The real-time monitoring and adjustment model of dynamic load conditions is constructed based on the prediction results of the dynamic fault prediction model. The load state vector L(t) is adaptively controlled by the real-time monitoring and adjustment model of dynamic load conditions, and the model is executed in real time through the edge computing unit. Meanwhile, the control strategy is adaptively updated and optimized regularly.

[0136] Specifically, the specific steps of S4 include:

[0137] S401, the construction of the real-time monitoring and adjustment model of dynamic load conditions, firstly extract the features of key power parameters (such as current and voltage) to obtain the time series P(t), and then compare it with the time series feature representation Z f (t) to establish the load state representation vector L(t), which is expressed as follows:

[0138] L(t)=P(t)⊙Z f (t)

[0139] Among them, P(t) is the power parameter sequence, including the characteristics of current and voltage, ⊙ represents the element-by-element multiplication operation, and L(t) is the fused load state vector, which is used to reflect the current operating status of the distribution cabinet;

[0140] S402, design a control strategy according to the load state vector L(t), which is expressed as follows:

[0141]

[0142] Where C(t) is the load regulation strategy vector at time t, λ s It is a regularization parameter that controls the smoothness of regulation and is used to limit the severity of regulation changes. The optimal regulation strategy C(t) is calculated based on the real-time load state L(t). This strategy will serve as a direct basis for executing operations to ensure load balancing of the distribution cabinet and reduce energy loss.

[0143] S403: After executing the control strategy, feedback detection is performed on the real-time monitored power parameters. If the deviation exceeds the set threshold, the strategy is immediately fine-tuned. The feedback adjustment formula is:

[0144]

[0145] Among them, C′(t) is the strategy vector after immediate feedback adjustment, P(t) is the actual monitored power parameter, and P′(t) is the power parameter expected under the current strategy. is the adjustment coefficient, which controls the amplitude of feedback adjustment.

[0146] S404: Based on the historical data and the current running status, the rolling window optimization algorithm is used to optimize the time window [tT w ,t], the effectiveness of the control strategy is re-evaluated, and the optimization process updates the control strategy by minimizing the cumulative error of historical operation data. The formula is as follows:

[0147]

[0148] Where C(t+1) is the update policy vector at time t+1, η2 is the learning rate, and T w is the length of the rolling window, which controls the impact of historical data on strategy updates.

[0149] Through this optimization algorithm, the control strategy can be gradually improved over time and as the system changes, adapting to emerging operating conditions and load patterns, thereby improving the long-term stability and efficiency of the system.

[0150] S5. According to the load state vector L(t) and the adaptive load control strategy, an adaptive threshold fault detection algorithm is constructed to detect fault signals, and a feature-enhanced multimodal fault classification algorithm is designed to classify faults. According to the fault classification, the corresponding response strategy is automatically generated, and finally the dynamic adjustment step-by-step recovery mechanism is used to restore the normal operation of the system.

[0151] Specifically, the specific steps of S5 include:

[0152] S501. Detect and dynamically isolate real-time fault signals based on an adaptive threshold fault detection algorithm. The difference measure of the fault signal is defined as:

[0153] D(t)=L(t)-B(t)

[0154] Where L(t) represents the current load state vector, and B(t) represents the baseline state vector of the system under normal operation. The conditions for detecting faults are:

[0155] ∥D(t)∥2>δ(t)

[0156] Where δ(t) is the dynamically adjusted fault detection threshold, defined as:

[0157]

[0158] Among them, δ0 is the basic threshold, which is set to a fixed value calculated based on historical data; α and β are adjustment coefficients that control the dynamic adjustment range of the threshold; L i (t) and B i (t) denotes the i-th component of the load state and the baseline state, respectively.

[0159] It can be understood that when ∥D(t)∥2 exceeds the threshold δ(t), the system immediately determines that a fault has occurred and automatically isolates the corresponding circuit. At this time, the system starts the backup line to ensure the continuity of power supply and the safe operation of the distribution cabinet.

[0160] S502, feature-enhanced multi-modal fault classification algorithm, specifically including:

[0161] First, generate the fault feature vector F(t), which contains the time series feature representation Z f (t) and the current load characteristic L(t), the fault characteristic vector is defined as:

[0162] F(t)=Z f (t)⊙L(t)

[0163] Among them, ⊙ is an element-by-element multiplication operation, and then the fault classifier uses a multimodal classification decision function based on an enhanced regularization term to perform fault classification, which is expressed as follows:

[0164]

[0165] in, is the fault type identification result, f i (F(t)) is a classification function based on multimodal features, α i (t) is the classification weight, γ s is the coefficient of the enhanced regularization term, F j (t) is the jth eigenvalue, F j (t-1) is the jth eigenvalue at time t-1, k s is the number of characteristic components;

[0166] The automatic generation of corresponding response strategies according to fault classification specifically includes:

[0167] The optimization objective function R(t) of the designed response strategy is:

[0168]

[0169] Among them, W r is a weight matrix associated with the fault response strategy, weighing different response objectives. is a regularization parameter used to smooth the response policy changes, η s is the coefficient that controls the complexity of the response strategy, is the complexity function of the response strategy.

[0170] S503. Design a dynamic adjustment and gradual recovery mechanism to ensure that the system returns to a stable state. In this mechanism, the system continuously monitors the changing trends of various key parameters during the recovery process. During the gradual recovery process, the system monitors and optimizes the current recovery state through the following formula:

[0171]

[0172] Among them, M(t) is the monitoring indicator of the current recovery state, reflecting the degree of system recovery; T r is the time window length for restoring monitoring, and ξ is the coefficient of the control strategy adjustment amplitude.

[0173] Specifically, by monitoring the dynamic changes of M(t), the system can promptly detect abnormal conditions during the recovery process and make adjustments to ensure that the system can smoothly transition to normal operation. During the recovery period, if the recovery is not as expected, the system will re-evaluate the fault and may execute the fault response strategy again to ensure the robustness of the recovery process. The dynamic adjustment and gradual recovery mechanism ensures that in the complex recovery process, the system can dynamically adjust the response strategy and recovery pace to avoid the occurrence of secondary faults and ensure the long-term stable operation of the entire distribution cabinet system.

[0174] S6. Adopt adaptive learning rate to update steps S1-S5 online, and adopt rolling window retraining mechanism to retrain them regularly. Design multi-model collaborative update mechanism for collaborative update, and design long-term trend monitoring and global adjustment mechanism to perform long-term intelligent monitoring of distribution cabinets.

[0175] Specifically, the specific steps of S6 include:

[0176] S601. The core model in the system is continuously updated online based on the adaptive learning rate online update algorithm. The algorithm dynamically adjusts the learning rate η(t) according to the real-time performance of the model to balance the convergence speed and adaptability of the model. The specific formula is:

[0177]

[0178] Among them, η0 is the initial learning rate, which reflects the learning rate at the beginning of the model; κ a is the adjustment amplitude coefficient, which is used to control the range of change of the learning rate; a is a constant that adjusts the sensitivity and determines how quickly the learning rate responds to changes in error; E i (t) represents the error metric of the i-th model at time t, such as prediction error, classification error rate, etc.

[0179] The adaptive learning rate online update algorithm adaptively adjusts the learning rate according to the changes in real-time errors, ensuring that the model can quickly adapt to new working conditions and environmental changes in long-term operation. The system can improve its response to sudden failures and new failure modes while maintaining stability.

[0180] S602, based on the rolling window retraining mechanism, by setting the time window T w , retrain the model regularly so that the model can absorb the latest data features and update its weights. The specific steps are as follows:

[0181] In the time window T w The data D(t) collected by the system is stored as a retraining data set. At the end of each time window, the weight w(t) of the model is updated by minimizing the loss function L(w,D(t)) of the model on the retraining set:

[0182]

[0183] Among them, w(t) is the weight vector of the model at time t, L(w,D(t)) is the loss function of the model on the retraining set, which can be in the form of classification error, regression error, etc., depending on the specific goal of the model; η(t) is the learning rate dynamically adjusted according to the adaptive learning rate online update algorithm. The rolling window retraining mechanism uses the data of the recent period of time for model retraining, so that the model can better adapt to environmental changes and ensure that it still maintains efficient fault detection and processing capabilities when facing new fault modes and working conditions.

[0184] S603. Design a multi-model collaborative update mechanism, specifically including:

[0185] Through a joint loss function L c The weights of multiple models are updated synchronously, and the joint loss function is defined as:

[0186]

[0187] Among them, w i is the weight vector of the ith model (such as the weights of the fault classifier, control strategy generator, and recovery mechanism model), L i (w i ) is the loss function of the i-th model, reflecting the performance index of the model, α i is the weight of the loss function of each model, which is used to balance the importance of different models, c is the weight of the co-regularization term, ensuring that the weight vectors of different models remain consistent during the update process.

[0188] Furthermore, the weight update formula of the model is:

[0189]

[0190] The multi-model collaborative update mechanism can ensure the consistency of different models when making decisions through collaborative updates, thereby improving the coordination and efficiency of the entire system when dealing with complex fault scenarios, and avoiding fault misjudgment or response delays due to inconsistencies between models.

[0191] S604. Based on the long-term trend monitoring and global adjustment mechanism, the long-term performance trends of each model are monitored to identify possible performance degradation or inconsistency problems, and the system is globally optimized and adjusted. The long-term trend monitoring indicator T(t) is defined as:

[0192]

[0193] Among them, T(t) is the m The long-term trend monitoring indicator within the time period reflects the changing trend of the model performance; E(t′) is the performance metric of the model at time t′, such as classification accuracy, response time, etc.; T m It is the length of the monitoring time window, which is usually longer and is used to capture the performance change trend of the model in the long run.

[0194] Understandably, if T(t) significantly exceeds the preset threshold over a period of time, the system will trigger global retraining and parameter optimization, adjusting the weights and strategies of each model to ensure the long-term robustness and efficiency of the entire system. Long-term trend monitoring and global adjustment mechanism can not only respond to sudden environmental changes, but also effectively deal with performance problems that gradually accumulate in long-term operation, ensuring that the system always maintains the best state throughout the entire operation cycle.

[0195] like Figure 2 As shown, an embodiment of the present invention further provides an intelligent monitoring system for a power distribution cabinet, the system comprising:

[0196] The data acquisition module 10 is used to deploy sensors in the power distribution cabinet to collect raw data S i (t), for the original data S i (t) Preprocessing to obtain data S″′ i (t), and then use asymmetric differential compression method to compress the data S″′ i (t) transmit and store;

[0197] Fault detection module 11, used for data S″′ i (t) Perform feature extraction and input the extracted features into the feature vector X i (t), according to the input feature vector Xi (t) Establishing an adaptive small sample learning model of an adaptive support vector machine based on meta-learning to classify the identified faults, designing an adaptive regularization term for the distribution cabinet environment and adding it to the training of the adaptive small sample learning model, using the trained adaptive small sample learning model to perform online fault detection on real-time data, and designing a feedback optimization mechanism to regularly perform adaptive adjustments on the adaptive small sample learning model;

[0198] The data fusion and fault prediction module 12 is used to combine the data S″′ according to the results of online fault detection i (t) Design a method based on adaptive correlation weighted fusion to fuse the sensor features according to the correlation weights and obtain the fused feature vector X f (t), and then use the method based on multiple time scale convolution to transform the feature vector X f (t) Extract time series features and obtain the final time series feature representation Z f (t), then a dynamic fault prediction model is proposed by constructing a recurrent neural network structure based on time-dependent enhancement, and the dynamic fault prediction model is adaptively updated online;

[0199] The strategy execution module 13 is used to represent Z according to the obtained time series characteristics. f (t) and the prediction results of the dynamic fault prediction model are used to build a dynamic load condition real-time monitoring and adjustment model, and the dynamic load condition real-time monitoring and adjustment model is used to perform adaptive load control on the load state vector L(t), and the edge computing unit executes it in real time, and the control strategy is adaptively updated and optimized regularly;

[0200] The local control module 14 is used to construct an adaptive threshold fault detection algorithm to detect fault signals according to the load state vector L(t) and the adaptive load control strategy, and to design a feature-enhanced multi-modal fault classification algorithm to classify faults, automatically generate corresponding response strategies according to the fault classification, and finally use a dynamic adjustment step-by-step recovery mechanism to restore the normal operation of the system;

[0201] The system optimization module 15 is used to update the system online using an adaptive learning rate, and to retrain it regularly using a rolling window retraining mechanism, and to design a multi-model collaborative update mechanism for collaborative update, and to design a long-term trend monitoring and global adjustment mechanism for long-term intelligent monitoring of the distribution cabinet;

[0202] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0203] In the several embodiments provided in the present application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a division of logical functions. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, and the indirect coupling or communication connection of the device or unit can be electrical, mechanical or other forms.

[0204] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server or a network device, etc.) to perform all or part of the steps of the methods described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0205] Although embodiments of the present invention have been shown and described, those skilled in the art will appreciate that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the claims and their equivalents.

Claims

1. An intelligent monitoring method for a power distribution cabinet, characterized in that: The method comprises: S1. Deploy sensors in the power distribution cabinet to collect raw data S i (t), for the original data S i (t) Preprocessing to obtain data S″′ i (t), and then use asymmetric differential compression method to compress the data S″′ i (t) transmit and store; S2, for data S″′ i (t) Perform feature extraction and input the extracted features into the feature vector X i (t), according to the input feature vector X i (t) Establishing an adaptive small sample learning model of an adaptive support vector machine based on meta-learning to classify the identified faults, designing an adaptive regularization term for the distribution cabinet environment and adding it to the training of the adaptive small sample learning model, using the trained adaptive small sample learning model to perform online fault detection on real-time data, and designing a feedback optimization mechanism to regularly perform adaptive adjustments on the adaptive small sample learning model; S3, combining the results of online fault detection with data S″′ i (t) Design a method based on adaptive correlation weighted fusion to fuse the sensor features according to the correlation weights and obtain the fused feature vector X f (t), and then use the method based on multiple time scale convolution to transform the feature vector X f (t) Extract time series features and obtain the final time series feature representation Z f (t), then a dynamic fault prediction model is proposed by constructing a recurrent neural network structure based on time-dependent enhancement, and the dynamic fault prediction model is adaptively updated online; S4. Express Z according to the obtained time series characteristics f (t) and the prediction results of the dynamic fault prediction model are used to build a dynamic load condition real-time monitoring and adjustment model, and the dynamic load condition real-time monitoring and adjustment model is used to perform adaptive load control on the load state vector L(t), and the edge computing unit executes it in real time, and the control strategy is adaptively updated and optimized regularly; S5. According to the load state vector L(t) and the adaptive load control strategy, an adaptive threshold fault detection algorithm is constructed to detect fault signals, and a feature-enhanced multimodal fault classification algorithm is designed to classify faults. According to the fault classification, a corresponding response strategy is automatically generated, and finally a dynamic adjustment step-by-step recovery mechanism is used to restore the normal operation of the system. S6. Adopt adaptive learning rate to update steps S1-S5 online, and adopt rolling window retraining mechanism to retrain them regularly. Design multi-model collaborative update mechanism for collaborative update, and design long-term trend monitoring and global adjustment mechanism to perform long-term intelligent monitoring of distribution cabinets.

2. The intelligent monitoring method for a power distribution cabinet according to claim 1, characterized in that: The asymmetric differential compression method performs different processing on abnormal points in differential data, and the processing is as follows: Among them, ΔS i (t) is the updated outlier point, Δ sym (t) is the standard symmetric difference compression result, ∈ is a threshold used to identify the normal fluctuation range, κ is the adjustment factor, R(t) is the regression residual based on the historical data of the distribution cabinet, and Δt is the time difference between adjacent data points.

3. The intelligent monitoring method for a power distribution cabinet according to claim 1, characterized in that: The adaptive small sample learning model uses an adaptive support vector machine based on meta-learning for classification, wherein the adaptive small sample learning model loss function L(w,λ,t) is expressed as follows: Among them, w is the weight vector of the classifier, w T is the transpose of w, X i (t) is the input feature vector of the i-th sample, y i is the label of sample i, y i ∈{-1,1}, C is the penalty parameter in the loss function, λ is the adjustment factor of the adaptive regularization term, ρ j (t) is the adaptive coefficient, n is the number of sensors, b is the bias, m is the number of distribution cabinets, and j is the current distribution cabinet; Use the trained adaptive small sample learning model to perform online fault detection on real-time data, where for each new data point S″′ i (t), according to the input feature vector X of the i-th sample i (t), calculate the output of the classification function, expressed as follows: if It is determined that a fault signal occurs at the current time t.

4. The intelligent monitoring method for a power distribution cabinet according to claim 3, characterized in that: The adaptive adjustment means that every time T r , remix the collected fault samples and normal samples for training, and dynamically adjust the adaptive coefficient ρ in the regularization term j (t), the adjustment formula is as follows: Where η is the learning rate, is the predicted label of the jth sample under the current adaptive small sample learning model, y j is the true label of sample j, n s is the number of new samples in time step t.

5. The intelligent monitoring method for a power distribution cabinet according to claim 1, characterized in that: The adaptive correlation weighted fusion means that the features of multiple sensors are correlated and analyzed to calculate different types of sensor data X. i (t) and X j The dynamic correlation weight ω of (t) ij (t), the formula is as follows: Among them, ω ij (t) represents the correlation weight between sensors i and j at time t, Σ ij is the covariance matrix between features, which is used to measure the correlation of sensor data, X i (t) represents the input feature vector of the i-th sample, X j (t) Input feature vector of the jth sample; The features of multiple sensors are weighted and fused according to the correlation weights to obtain the fused feature vector X f (t), expressed as follows: Where n is the number of sensors.

6. The intelligent monitoring method for a power distribution cabinet according to claim 5, characterized in that: The state update formula of the recurrent neural network based on temporal dependency enhancement is: h t =σ(W h ·Z f (t)+U h ·h t-1 +b h ) Among them, h t is the hidden state at time t, h t-1 is the hidden state at time t-1, W h and U h are the weight matrices of the input and hidden layers, respectively, and b h is the bias vector and σ is the activation function.

7. The intelligent monitoring method for a power distribution cabinet according to claim 1, characterized in that: The construction of the real-time monitoring and adjustment model of dynamic load conditions is as follows: First, the key power parameters are extracted to obtain the time series P(t), which is then compared with the time series feature representation Z f (t) to establish the load state representation vector L(t), which is expressed as follows: L(t)=P(t)⊙Z f (t) Where P(t) is the power parameter sequence, including the characteristics of current and voltage, ⊙ represents the element-by-element multiplication operation, and L(t) is the fused load state vector; The control strategy is designed according to the load state vector L(t), which is expressed as follows: Where C(t) is the load regulation strategy vector at time t, λ s is a regularization parameter that controls the smoothness of regulation and is used to limit the severity of regulation changes. 2 is the L2 norm; After executing the control strategy, the real-time monitored power parameters are fed back for detection. If the deviation exceeds the set threshold, the strategy is fine-tuned immediately. The feedback adjustment formula is: Among them, C′(t) is the strategy vector after immediate feedback adjustment, P(t) is the actual monitored power parameter, and P′(t) is the power parameter expected under the current strategy. is the adjustment coefficient, which controls the amplitude of feedback adjustment.

8. The intelligent monitoring method for a power distribution cabinet according to claim 1, characterized in that: The feature-enhanced multi-modal fault classification algorithm specifically includes: First, generate the fault feature vector F(t), which contains the time series feature representation Z f (t) and the current load characteristic L(t), the fault characteristic vector is defined as: F(t)=Z f (t)⊙L(t) Among them, ⊙ is the element-by-element multiplication operation, and then the fault classifier uses a multimodal classification decision function based on an enhanced regularization term to perform fault classification, which is expressed as follows: in, is the fault type identification result, f i (F(t)) is a classification function based on multimodal features, α i (t) is the classification weight, γ s is the coefficient of the enhanced regularization term, F j (t) is the jth eigenvalue, F j (t-1) is the jth eigenvalue at time t-1, k s is the number of feature components, ∥∥ 2 is the L2 norm, m is the number of features in fault classification; The automatic generation of corresponding response strategies according to fault classification specifically includes: The optimization objective function R(t) of the designed response strategy is: Among them, W r is a weight matrix associated with the fault response strategy, weighing different response objectives. is a regularization parameter used to smooth the response policy changes, η s is the coefficient that controls the complexity of the response strategy, is the complexity function of the response strategy, R represents the response strategy variable that needs to be optimized, and R(t-1) represents the response strategy at the previous time point t-1, which is used as a reference point for the current strategy optimization.

9. The intelligent monitoring method for a power distribution cabinet according to claim 1, characterized in that: The multi-model collaborative update mechanism includes: Through a joint loss function L c The weights of multiple models are updated synchronously, and the joint loss function is defined as: Among them, w i is the weight vector of the i-th model, w1 corresponds to the fault classifier model, w2 corresponds to the control strategy generator model, w3 corresponds to the recovery mechanism model, j is the index between models, ∥∥ 2 is the L2 norm, L i (w i ) is the loss function of the i-th model, α i is the weight of the loss function of each model, λ c is the weight of the collaborative regularization term, and the weight update formula of the model is: Among them, w i (t+1) is the weight vector of the i-th model at time t+1, It is a partial differential.

10. Based on an intelligent monitoring system for a power distribution cabinet, characterized in that: The system comprises: Data acquisition module, used to deploy sensors in the power distribution cabinet to collect raw data S i (t), for the original data S i (t) Preprocessing to obtain data S″′ i (t), and then use asymmetric differential compression method to compress the data S″′ i (t) transmit and store; Fault detection module, used for data S″′ i (t) Perform feature extraction and input the extracted features into the feature vector X i (t), according to the input feature vector X i (t) Establishing an adaptive small sample learning model of an adaptive support vector machine based on meta-learning to classify the identified faults, designing an adaptive regularization term for the distribution cabinet environment and adding it to the training of the adaptive small sample learning model, using the trained adaptive small sample learning model to perform online fault detection on real-time data, and designing a feedback optimization mechanism to regularly perform adaptive adjustments on the adaptive small sample learning model; Data fusion and fault prediction module, used to combine data S″′ based on the results of online fault detection i (t) Design a method based on adaptive correlation weighted fusion to fuse the sensor features according to the correlation weights and obtain the fused feature vector X f (t), and then use the method based on multiple time scale convolution to transform the feature vector X f (t) Extract time series features and obtain the final time series feature representation Z f (t), then a dynamic fault prediction model is proposed by constructing a recurrent neural network structure based on time-dependent enhancement, and the dynamic fault prediction model is adaptively updated online; The strategy execution module is used to represent Z according to the obtained time series characteristics. f (t) and the prediction results of the dynamic fault prediction model are used to build a dynamic load condition real-time monitoring and adjustment model, and the dynamic load condition real-time monitoring and adjustment model is used to perform adaptive load control on the load state vector L(t), and the edge computing unit executes it in real time, and the control strategy is adaptively updated and optimized regularly; The local control module is used to construct an adaptive threshold fault detection algorithm to detect fault signals based on the load state vector L(t) and the adaptive load control strategy, and to design a feature-enhanced multi-modal fault classification algorithm to classify faults. It automatically generates corresponding response strategies based on the fault classification, and finally uses a dynamic adjustment and gradual recovery mechanism to restore the normal operation of the system. The system optimization module is used to update the system online using an adaptive learning rate, and to retrain it regularly using a rolling window retraining mechanism. A multi-model collaborative update mechanism is designed for collaborative updating. A long-term trend monitoring and global adjustment mechanism is designed for long-term intelligent monitoring of the distribution cabinet.

Citation Information

Cited By

  • Fault detection method and system for power distribution cabinet

    CN119720030A

  • A fault detection method and system for power distribution cabinet

    CN119720030B

  • Interlocking remote control system and method based on low-voltage power distribution cabinet

    CN119727132A

  • Low-voltage power distribution cabinet interlocking remote control system and method thereof

    CN119727132B

  • Intelligent electric cabinet electrical fault detection system and device

    CN119966067A