Cluster type electrical potential safety hazard pre-judgment method, system and equipment and storage medium

Abstract

The invention discloses a cluster type electrical potential safety hazard pre-judgment method, system and device and a storage medium. The method comprises the steps of obtaining a real-time data stream, performing data processing on the real-time data stream to obtain a first data stream, performing feature extraction by using an edge computing node, and screening abnormal features to obtain an abnormal feature set; analyzing the abnormal feature set by using a deep learning model, modeling time sequence features in the abnormal feature set through a neural network model to obtain a potential risk probability of electrical hidden dangers, and optimizing deep learning model parameters by using an update algorithm to obtain updated model parameters; extracting neural network model parameters related to the edge end computing node, returning the neural network model parameters to the edge end computing node to obtain an optimized edge model, and reasoning the real-time data stream to obtain a real-time hidden danger early warning result; and a real-time hidden danger early warning result is judged to obtain an electrical potential safety hazard early warning instruction, so that a powerful guarantee is provided for safe and stable operation of a power system.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of electrical safety, and in particular to a cluster type electrical safety hidden danger prediction method, system, device and storage medium. BACKGROUND

[0002] Electrical safety hidden danger prediction is a key field to ensure the stable operation of power systems and public safety. Its importance lies in identifying potential risks in advance to avoid equipment failures, fires and other serious accidents. With the popularization of smart grid and Internet of Things technology, network-based cluster monitoring has become an important direction of electrical safety management. However, existing methods often face low computational efficiency and slow response speed when dealing with large-scale real-time data, especially in complex network environments, where data transmission and processing delays are significant, making it difficult to meet the millisecond-level warning needs. The limitations of current solutions mainly lie in the inefficiency of data processing and model updating. Traditional methods rely on centralized computing, which requires the transmission of massive monitoring data to the cloud, resulting in high network bandwidth pressure and high latency. At the same time, model updating usually requires retraining on the full amount of data, which consumes a lot of computing resources and is difficult to adapt to rapidly changing electrical parameters. These problems limit the warning capability of existing systems in dynamic and continuous monitoring scenarios. The core challenge lies in the dual demands of real-time and efficiency. The continuous generation of massive electrical parameters requires the system to have fast processing capability, but the centralized architecture leads to large data transmission volume and overburdened computing nodes. As a result, the computational tasks need to be reasonably distributed between the edge and the cloud to reduce the transmission burden. However, the limited computing resources at the edge make it difficult to efficiently perform complex analysis tasks, which further challenges the model updating mechanism. The traditional full-retraining method cannot quickly adapt to new data, affecting the real-time performance of the warning.

[0003] Therefore, how to optimize the computing architecture and algorithm design to achieve efficient data processing at the edge and collaborative computing in the cloud, while using incremental model updating to speed up the response, has become a key problem in network cluster-based electrical safety hidden danger prediction. SUMMARY

[0004] In view of the above existing problems, the present application is proposed.

[0005] Therefore, the present application provides a cluster type electrical safety hidden danger prediction method, system, device and storage medium to solve the problems of low computational efficiency, slow response speed and difficulty in meeting the millisecond-level warning needs of existing electrical safety hidden danger prediction methods when dealing with large-scale real-time data.

[0006] To solve the above technical problems, the present application provides the following technical solutions:

[0007] In a first aspect, the present application provides a cluster type electrical safety hidden danger prediction method, comprising:

[0008] Obtaining a real-time data stream, performing data processing on the real-time data stream to obtain a first data stream;

[0009] According to the first data stream, using an edge computing node to perform feature extraction and filter abnormal features to obtain an abnormal feature set;

[0010] Using a deep learning model to analyze the abnormal feature set, modeling the time series features in the abnormal feature set through a neural network model to obtain a potential risk probability of an electrical hidden danger;

[0011] According to the potential risk probability, using an update algorithm to optimize the deep learning model parameters to obtain updated model parameters;

[0012] Extracting the neural network model parameters related to the edge computing node in the updated model parameters and returning to the edge computing node to obtain an optimized edge model;

[0013] According to the optimized edge model, performing inference on the real-time data stream to obtain a real-time hidden danger early warning result;

[0014] Judging the real-time hidden danger early warning result, and obtaining an electrical safety hidden danger early warning instruction according to the judgment result.

[0015] As a preferred scheme of the cluster type electrical safety hidden danger pre-judgment method, wherein the data processing on the real-time data stream comprises:

[0016] Obtaining original data from the real-time data stream, pre-processing the original data using a first data compression algorithm to obtain an encoded data stream;

[0017] Using a first quantization coding technology to reduce the data packets in the encoded data stream to below a first threshold to obtain a first data stream.

[0018] As a preferred scheme of the cluster type electrical safety hidden danger pre-judgment method, wherein the abnormal feature set comprises:

[0019] Obtaining data packets from the first data stream, decompressing the data packets through an edge computing node to obtain an electrical parameter sequence;

[0020] Using a first neural network model to perform time series analysis on the electrical parameter sequence to obtain a feature vector set, and if any feature in the feature vector set exceeds a second threshold, it is a potential abnormal point to obtain an abnormal candidate set;

[0021] The first abnormality feature set is obtained by performing time sequence correlation analysis on the abnormality candidate set through the first flow control technology, and the abnormality type is classified by using the first classifier to obtain an abnormality category;

[0022] The high-confidence abnormality is extracted from the abnormality category, and the abnormality situation is distinguished by using clustering analysis to obtain an abnormality detection result, generate time sequence distribution data of the abnormality feature, and output the final abnormality feature set.

[0023] As a preferred scheme of the cluster type electrical safety hidden danger prediction method, the potential risk probability of the electrical hidden danger includes:

[0024] The high-dimensional feature vector is output by performing feature extraction on the abnormality feature set through the pre-trained deep learning model of the cloud computing node;

[0025] The time sequence of the high-dimensional feature vector is modeled by using the neural network model to generate a dynamic change trend of the time sequence;

[0026] According to the dynamic change trend, if the fluctuation amplitude of the high-dimensional feature vector exceeds a third threshold value, the potential risk probability of the electrical hidden danger is calculated by probability;

[0027] The risk factor is extracted from the potential risk probability, and the risk factor is classified by using a clustering analysis method to obtain a high-risk factor set;

[0028] For the high-risk factor set, whether there is a match with a known electrical hidden danger mode is judged by comparing the historical data, and a matching result is output;

[0029] According to the matching result, the type and severity of the electrical hidden danger are classified by using a decision tree algorithm to obtain a hidden danger classification result.

[0030] The beneficial effects of the preferred technical scheme are that the electrical hidden danger risk probability can be accurately identified by deep learning and clustering analysis, the warning accuracy and real-time performance are improved, and the safe operation of the power system is effectively ensured.

[0031] As a preferred scheme of the cluster type electrical safety hidden danger prediction method, the updated model parameters include:

[0032] The risk assessment value is obtained by calculating the risk probability of the data sample by using the first training method through the online learning technology from the real-time data stream;

[0033] If the risk assessment value is greater than a fourth threshold value, the deep learning model parameters are adjusted by using the first update algorithm, the inference of the deep learning model is performed, the prediction result of the current data stream is generated, and the model output value is determined;

[0034] By comparing the prediction result with the actual data, the model convergence error is calculated to determine whether the preset convergence condition is met, and an error evaluation result is obtained.

[0035] If the error evaluation result does not meet the convergence condition, the sample selection strategy is adjusted, the data samples in the real-time data stream are reacquired, and a new training sample set is obtained.

[0036] According to the new training sample set, the first updating algorithm is repeatedly executed to optimize the deep learning model, and the operation is repeated to continuously monitor the model convergence error until the preset convergence condition is met, and the final updated model parameters are obtained.

[0037] The beneficial effects of the preferred technical solution are that online learning and dynamic optimization can quickly adapt to new data, improve model prediction performance and convergence speed, and enhance the real-time performance and accuracy of the system.

[0038] As a preferred scheme of the cluster type electrical safety hidden danger prediction method, the optimized edge model comprises:

[0039] The original model parameters are obtained from the edge computing node, and the parameters are processed by distributed computing architecture to obtain neural network model parameters.

[0040] The neural network model parameters are encrypted by a secure encryption protocol, and encrypted parameters are generated by using an encryption algorithm.

[0041] If the encrypted parameters pass the integrity check, the encrypted parameters are sent to the edge computing node through the parameter transmission process to obtain the transmission successful parameters.

[0042] According to the transmission successful parameters, model deployment update is performed on the edge computing node to generate a preliminary optimized model.

[0043] For the preliminary optimized model, edge device performance data is obtained, and if the performance data meets the fifth threshold, real-time parameter processing is performed to obtain a performance adaptation model.

[0044] Through the interaction of the performance adaptation model and the edge computing node, model parameter fine-tuning is performed by using a distributed computing architecture to obtain an optimized edge model.

[0045] The beneficial effects of the preferred technical solution are that through encrypted transmission and distributed optimization, the model parameter security update is ensured, the edge device performance is adapted, and the edge inference efficiency and accuracy are improved.

[0046] As a preferred scheme of the cluster type electrical safety hidden danger prediction method, the real-time hidden danger early warning result comprises:

[0047] The time series data segment is acquired, the time series data segment is segmented to obtain a segmented time series, the optimized edge model is used for reasoning on each data segment to generate a segment-level feature vector, and a feature representation is obtained;

[0048] If the feature representation matches a preset hidden danger mode, the feature vector is compared with a sixth threshold value to determine whether there is a potential hidden danger, and a hidden danger judgment result is obtained;

[0049] According to the hidden danger judgment result, a continuous monitoring mechanism of the time series is adopted, a moving step of a sliding window is combined, a subsequent data segment is acquired, and an updated time series is obtained;

[0050] Through the updated time series, the optimized edge model is used to perform rapid reasoning again to generate a new feature vector, and a continuous feature representation is obtained;

[0051] If the continuous feature representation shows that the hidden danger trend continues, a real-time hidden danger early warning result is generated through a preset early warning rule, a warning level is determined, an output mechanism of the edge computing node is used to transmit the early warning result to a target system, and the real-time hidden danger early warning result is output;

[0052] If the real-time hidden danger early warning result shows that the risk probability exceeds a seventh threshold value, the real-time hidden danger early warning result is transmitted to a control center, and a final electrical safety hidden danger early warning instruction is obtained.

[0053] In a second aspect, the present application provides a cluster type electrical safety hidden danger prediction system, comprising:

[0054] A data processing module is configured to acquire a real-time data stream, process the real-time data stream, and obtain a first data stream;

[0055] A feature extraction module is configured to extract features from the first data stream using an edge computing node, and filter abnormal features to obtain an abnormal feature set;

[0056] A feature modeling module is configured to analyze the abnormal feature set using a deep learning model, model time series features in the abnormal feature set through a neural network model, and obtain a potential risk probability of an electrical hidden danger;

[0057] A parameter updating module is configured to optimize the parameters of the deep learning model using an updating algorithm according to the potential risk probability, and obtain updated model parameters;

[0058] A model optimization module is configured to extract neural network model parameters related to the edge computing node from the updated model parameters, and return the parameters to the edge computing node to obtain an optimized edge model;

[0059] An inference module is configured to perform inference on the real-time data stream according to the optimized edge model, and obtain a real-time hidden danger early warning result.

[0060] A judgment module is configured to judge the real-time hidden danger early warning result, and obtain an electrical safety hidden danger early warning instruction according to a judgment result.

[0061] In a third aspect, the present application provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and wherein the processor implements the steps of the cluster type electrical safety hidden danger pre-judgment method when executing the computer program.

[0062] In a fourth aspect, the present application provides a computer readable storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement the steps of the cluster type electrical safety hidden danger pre-judgment method.

[0063] Compared with the prior art, the present application has the following beneficial effects: by compressing and performing preliminary analysis on the massive real-time data stream through edge computing, the rapid detection of electrical parameter abnormalities is realized; when potential abnormalities are detected, the feature data is transmitted to the cloud for deep learning analysis to obtain a hidden danger risk probability. Based on this, the present application uses an incremental algorithm to continuously optimize the cloud model, and the optimized lightweight model parameters are returned to the edge node to realize real-time monitoring and early warning based on the cooperation of the edge and the cloud; through multi-level data processing and model optimization, the early warning accuracy and real-time performance of the electrical safety hidden danger are effectively improved, thereby providing a strong guarantee for the safe and stable operation of the power system. BRIEF DESCRIPTION OF DRAWINGS

[0064] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiment description. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can be obtained without creative labor based on these drawings.

[0065] Figure 1 A whole flow logic diagram of a cluster type electrical safety hidden danger pre-judgment method provided by an embodiment of the present application. DETAILED DESCRIPTION

[0066] In order to make the above-mentioned purposes, features and advantages of the present application more apparent and easy to understand, the specific embodiments of the present application will be described in detail below with reference to the drawings in the specification. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should be within the protection scope of the present application.

[0067] Embodiment 1, reference Figure 1 For an embodiment of the present application, a cluster type electrical safety hidden danger prediction method is provided, comprising:

[0068] S100: Obtain a real-time data stream, and perform data processing on the real-time data stream to obtain a first data stream;

[0069] S200: According to the first data stream, use an edge computing node to extract features, and filter abnormal features to obtain an abnormal feature set;

[0070] S300: Use a deep learning model to analyze the abnormal feature set, model the time series features in the abnormal feature set through a neural network model, and obtain a potential risk probability of an electrical hidden danger;

[0071] S400: According to the potential risk probability, use an updating algorithm to optimize the deep learning model parameters to obtain updated model parameters;

[0072] S500: Extract the neural network model parameters related to the edge computing node in the updated model parameters, and return them to the edge computing node to obtain an optimized edge model;

[0073] S600: According to the optimized edge model, infer the real-time data stream to obtain a real-time hidden danger early warning result;

[0074] S700: Judge the real-time hidden danger early warning result, and obtain an electrical safety hidden danger early warning instruction according to the judgment result.

[0075] It should be noted that through the edge-cloud collaborative architecture, the whole process optimization from data acquisition to hidden danger early warning is realized. By using data compression, edge computing, deep learning and model updating technology, the detection efficiency and early warning accuracy of electrical safety hidden danger are significantly improved, which meets the real-time demand and provides strong guarantee for the safe operation of power system.

[0076] In the embodiment of the present application, the above step S100 comprises the following sub-steps A1-A2;

[0077] In A1: Obtain original data from the real-time data stream, and use a first data compression algorithm to pre-process the original data to obtain an encoded data stream;

[0078] In A2: Use a first quantization coding technology to reduce the data packets in the encoded data stream to below a first threshold to obtain a first data stream.

[0079] In an alternative embodiment, the first data compression algorithm can be wavelet transform, which divides the continuous sampling point data into data blocks; applies wavelet transform to each data block to decompose the signal into coefficients of different frequency levels; retains the main coefficients of the low frequency part and truncates the minor coefficients of the high frequency part; reconstructs the data block through inverse wavelet transform to realize data compression;

[0080] In an alternative embodiment, the first data compression algorithm can be principal component analysis, which divides the continuous sampling point data into data blocks; performs standardization processing on each data block to make the data mean value 0 and variance 1; calculates the covariance matrix of the data block and solves its eigenvalues and eigenvectors; selects the eigenvectors corresponding to the first k largest eigenvalues as the principal components, and projects the original data onto these principal components to obtain the reduced dimension data; reduces the data amount by retaining the main components;

[0081] In the embodiment of the application, the first data compression algorithm includes a compression algorithm based on discrete cosine transform for preprocessing, which is expressed as:

[0082]

[0083] wherein x n is a sampling point in the original data stream, X k is a frequency domain coefficient, and N is the length of the data block;

[0084] Specifically, real-time data stream is obtained from the sensor of the electrical equipment, the data is collected through a preset sampling frequency to obtain an original data stream; the original data stream is preprocessed by using a Huffman coding algorithm, an encoding table is generated by counting the data frequency to obtain an encoded data stream; the encoded data stream is processed by using quantization coding technology, if the data packet size exceeds a preset threshold, the quantization step is adjusted to obtain a first data stream, i.e. a compressed data stream; the data is reorganized by using a sliding window technology according to the blocking characteristics of the compressed data stream to obtain a reorganized data stream; if the transmission rate of the reorganized data stream is lower than the preset bandwidth requirement, the transmission priority is dynamically adjusted to obtain an optimized data stream; the integrity of the optimized data stream is verified by using a preset verification algorithm to obtain a verified data stream; the verified data stream is encapsulated by using a data stream transmission protocol to obtain a transmission data stream.

[0085] Specifically, the real-time data stream obtained from the sensor of the electrical equipment is usually generated at a frequency of 1000 sampling points per second, each sampling point contains a 32-bit floating point value, and the original data amount reaches 4MB per second; 100 continuous sampling points are divided into a data block, the time domain signal is converted into frequency domain coefficients by DCT transform, the first 20 main frequency components are retained and the remaining components are truncated to realize 85% data amount reduction.

[0086] In an alternative embodiment, the first quantization coding technique can be uniform quantization, determining the dynamic range of the data, uniformly dividing the entire range into several quantization intervals; each quantization interval is allocated the same number of quantization bits, and the quantization step of all frequency components is the same; the value of each frequency component is mapped to the nearest quantization interval center value, thereby completing quantization.

[0087] In an alternative embodiment, the first quantization coding technique can be non-uniform quantization, dividing the data range into different quantization intervals according to the energy distribution of the frequency components, allocating more quantization bits to the intervals with high energy, and allocating fewer quantization bits to the intervals with low energy, and combining an adaptive quantization strategy to dynamically adjust the quantization intervals and bit allocation according to the energy distribution of the real-time data.

[0088] In an embodiment of the present application, the first quantization coding technique includes a dynamic bit allocation technique represented as:

[0089]

[0090] where x is the frequency domain coefficient, Δ is the quantization step, and q(x) is the quantized value.

[0091] Specifically, for each frequency component, 4 to 8 bits of quantization bits are allocated according to its energy distribution, wherein coefficients with energy greater than 0.5 are allocated 8 bits, coefficients with energy from 0.1 to 0.5 are allocated 6 bits, and coefficients with energy less than 0.1 are allocated 4 bits. Finally, each data block is compressed from the original 3200 bits to not more than a first threshold of 512 bits. During the compression process, the data quality is evaluated by calculating the signal-to-noise ratio (SNR) to ensure that the SNR after compression is always maintained above 40 dB. For abnormal data points (such as values exceeding 3 times the standard deviation), a linear interpolation algorithm is used for smoothing processing, and an abnormal flag bit is recorded for subsequent analysis. The compressed data stream is added with a 2-byte check code through a CRC-16 check algorithm to form a final transmission data packet, and the overall compression ratio reaches 6.25:1, meeting the real-time transmission bandwidth requirements.

[0092] It should be noted that through data compression, quantization coding, abnormal processing and checking, etc., the data transmission amount is significantly reduced, the transmission efficiency is improved, and at the same time the data integrity and quality are ensured, meeting the bandwidth requirements of real-time monitoring.

[0093] In an embodiment of the present application, the above step S200 includes the following sub-steps B1-B4.

[0094] In B1, the data packet is obtained from the first data stream, and the data packet is decompressed by the edge computing node to obtain the electrical parameter sequence.

[0095] In B2: the first neural network model is used for time series analysis on the electrical parameter sequence to obtain a feature vector set, and if any feature in the feature vector set exceeds a second threshold, it is a potential abnormal point, and an abnormal candidate set is obtained;

[0096] In B3: the first flow control technology is used for time series correlation analysis on the abnormal candidate set to obtain a first abnormal feature set, and a first classifier is used for classifying the abnormal type to obtain an abnormal category;

[0097] In B4: high-confidence abnormalities are extracted from the abnormal category, and clustering analysis is used to distinguish abnormal conditions to obtain an abnormal detection result, generate time series distribution data of abnormal features, and output a final abnormal feature set.

[0098] In an optional embodiment, the first flow control technology can be a moving average method, and the data points in the abnormal candidate set are subjected to moving average processing to calculate the average value of each time point and several time points before and after it; the moving average is used to smooth the data and reduce the influence of short-term fluctuations, and highlight long-term trends; the change of the moving average value is detected to judge the continuity of the abnormal point; if consecutive moving average values exceed a preset threshold, it is considered that there is an abnormal trend;

[0099] In an optional embodiment, the first flow control technology can be Kalman filtering, and the data points in the abnormal candidate set are regarded as observation values, a state space model is established, and state variables and observation models are defined; the Kalman filtering algorithm is used to recursively estimate the state variables, combine prior knowledge and observation data, and dynamically adjust the estimated value; the change of the filtered state variables is analyzed to judge the continuity and periodicity of the abnormal point. If the filtered state variables continuously deviate from the normal range, it is considered that there is an abnormality;

[0100] In the embodiment of the application, the first flow control technology includes a sliding window algorithm;

[0101] Specifically, data packets are obtained from the compressed data stream, decompressed and preprocessed by the edge computing node to obtain a preliminary electrical parameter sequence; a first neural network model, which is a lightweight convolutional neural network, is used for time series analysis on the electrical parameter sequence to extract time domain and frequency domain features and obtain a feature vector set; if any feature value in the feature vector set exceeds a preset threshold, it is marked as a potential abnormal point, and an abnormal candidate set is obtained; a sliding window mechanism is used for time series correlation analysis on the abnormal candidate set to judge the continuity and periodicity of the abnormal point, and an abnormal feature set is obtained;

[0102] For example, by using a sliding window algorithm, the window size is set to 128 data points, and the step size is 64. The voltage and current signals are processed in segments. The data in each window is converted to frequency domain features by fast Fourier transform, and the main frequency components and their amplitudes are extracted to form the initial feature vector. The time series data is converted to frequency domain feature representation as follows:

[0103]

[0104] where x n is the sampling point in the time series, X k is the frequency domain feature, and N is the length of the data block.

[0105] Then, a lightweight convolutional neural network (CNN) model is used to analyze these feature vectors. The model uses a 1D convolutional layer with a kernel size of 3 and a step size of 1, and the activation function is ReLU. Through two layers of convolution and pooling operations, local features in the time series are gradually extracted. In the training process, the cross-entropy loss function and the Adam optimizer are used, and the learning rate is set to 0.001. After 100 iterations, the model can accurately identify abnormal features in electrical parameters. The convolution operation is represented as follows:

[0106]

[0107] where K(m,n) is the convolution kernel, I(i,j) is the input feature map, and F(i,j) is the output feature map.

[0108] When an abnormal fluctuation with a frequency of 50Hz appears in the voltage signal, the model can mark it as a potential fault point. Finally, through the abnormal feature set output by the model, combined with the running state of the electrical equipment, further fault diagnosis and prediction can be carried out to ensure the stable operation of the system.

[0109] In an alternative embodiment, the first classifier can be a random forest classifier, which constructs multiple decision trees, each trained using a randomly selected subset of features and data samples. When classifying each abnormal feature data, the data is input into all decision trees, each tree outputs a classification result, and the final classification result is determined by the majority voting mechanism;

[0110] In an alternative embodiment, the first classifier can be a deep neural network classifier, which constructs a deep neural network model containing multiple hidden layers. The input layer receives the abnormal feature data, and the output layer outputs the classification result. The network is trained using the backpropagation algorithm, and the weights and biases are adjusted to minimize the loss function. After training is complete, the abnormal feature data is input into the trained DNN model, and the model outputs the classification result;

[0111] In the embodiment of the present application, the first classifier includes a support vector machine algorithm to classify the abnormal feature data to obtain classified feature data; and according to the classified feature data, a final analysis result is obtained through a distributed computing framework of a cloud computing node.

[0112] Specifically, the abnormal feature set is transmitted to the cloud computing node through a low-latency network protocol to obtain feature data received by the cloud; the K-means algorithm is used to cluster the feature data according to the feature data received by the cloud to obtain clustered feature data; if there is an abnormal cluster in the clustered feature data, the category of the abnormal feature is determined by comparing the distribution of the abnormal cluster and the normal cluster to obtain an abnormal category; the pre-established abnormal processing rule library is obtained according to the abnormal category, and the processing mode of the abnormal feature is determined through rule matching to obtain a processing mode; the support vector machine algorithm is used to classify the abnormal feature data through the processing mode to obtain classified feature data; and a final analysis result is obtained through a distributed computing framework of a cloud computing node according to the classified feature data.

[0113] For example, in the preliminary abnormal feature set, the feature values include CPU usage, memory occupancy and network traffic, and the preset second threshold values are 90%, 85% and 100 Mbps respectively; when it is detected that the CPU usage at a certain moment reaches 95%, the memory occupancy is 88%, and the network traffic is 120 Mbps, these feature values all exceed the preset abnormal threshold value; the system immediately transmits these feature sets to the cloud computing node through a low-latency network protocol such as UDP or QUIC; after the cloud computing node receives the data, the machine learning-based anomaly detection algorithm such as Isolation Forest or Local Outlier Factor is used for further analysis.

[0114] The Isolation Forest algorithm detects abnormal points by constructing multiple random trees, and assumes that in the construction process of 100 trees, the abnormal score of a certain feature set reaches 0.8, which is much higher than the normal value of 0.2, so it is determined that the feature set is abnormal; the system stores these feature data to be further analyzed in a high-performance database such as Redis or Cassandra for subsequent in-depth analysis and processing; the whole process from feature detection to data transmission to cloud analysis is completed within milliseconds, ensuring real-time and high efficiency.

[0115] It should be noted that through edge computing and a lightweight convolutional neural network model, time-frequency domain features of electrical parameters are quickly extracted, abnormalities are accurately identified, and the prediction efficiency and accuracy are improved, providing a reliable basis for subsequent in-depth analysis.

[0116] In the embodiment of the present application, the above step S300 includes the following sub-steps C1-C6.

[0117] In C1: feature extraction is performed on the abnormal feature set by a pre-trained deep learning model of a cloud computing node, and a high-dimensional feature vector is output;

[0118] In C2: a neural network model is used to model the time series of the high-dimensional feature vector, and a dynamic change trend of the time series is generated;

[0119] In C3: according to the dynamic change trend, if the fluctuation amplitude of the high-dimensional feature vector exceeds a third threshold, the potential risk probability of the electrical hidden danger is calculated by probability;

[0120] In C4: risk factors are extracted from the potential risk probability, and a clustering analysis method is used to classify the risk factors to obtain a high-risk factor set;

[0121] In C5: for the high-risk factor set, whether there is a match with a known electrical hidden danger mode is judged by comparing historical data, and a matching result is output;

[0122] In C6: according to the matching result, a decision tree algorithm is used to classify the type and severity of the electrical hidden danger, and a hidden danger classification result is obtained.

[0123] Specifically, when deploying a pre-trained deep learning model on a cloud computing node, ResNet-50 can be used as the basic architecture, the input dimension is set to 256x256 of the electrical equipment infrared thermal image, the spatial features are extracted by 3x3 convolution kernel, and the ReLU activation function is used to process the nonlinear relationship; for time series feature analysis, an LSTM network containing 128 hidden units is used, the input window is set to 30-minute interval of 24-dimensional temperature, current, vibration and other multi-source sensor data, and the tanh function is used to control the threshold state; when the three-phase current data of a power distribution cabinet exceeds the third threshold of 0.85 in the abnormal probability output by the LSTM, the system automatically triggers the early warning mechanism;

[0124] For example, the model outputs a risk probability value of 0.92 when the A-phase current harmonic distortion rate is monitored to exceed 7.2% for 5 time steps in a row, combined with the ultrasonic signal of the partial discharge amount of 0.21 pC / s;

[0125] At this time, the cloud will call the associated operation and maintenance work order system to generate a maintenance task containing GPS coordinates, and store 1-minute granularity historical data into the time series database InfluxDB for traceability analysis; in the feature fusion stage, the attention mechanism is used to weight the spatial features and time features, and the weight coefficients are normalized to 0.34 and 0.66 through the softmax function, and the final hidden danger identification accuracy reaches 93.7%; the entire process is containerized through Kubernetes, and each computing node is configured with 4-core CPU and 16GB memory resources to ensure that real-time analysis is completed within 200ms.

[0126] It should be noted that through the deep learning model and the long short-term memory network, the electrical hidden danger is accurately modeled and risk evaluated, the historical data comparison and feature fusion are combined, and the accuracy and real-time performance of hidden danger identification are significantly improved.

[0127] In the embodiment of the present application, the above step S400 includes the following sub-steps D1-D5.

[0128] In D1, through online learning technology, new data is obtained from real-time data stream, the risk probability of the data sample is calculated by using the first training method, and a risk evaluation value is obtained;

[0129] In D2, if the risk evaluation value is greater than the fourth threshold value, the first updating algorithm is used to adjust the parameters of the deep learning model, the inference of the deep learning model is performed, the prediction result of the current data stream is generated, and the model output value is determined;

[0130] In D3, by comparing the prediction result with the actual data, the model convergence error is calculated, whether the preset convergence condition is met is judged, and an error evaluation result is obtained;

[0131] In D4, if the error evaluation result does not meet the convergence condition, the sample selection strategy is adjusted, the data sample in the real-time data stream is reacquired, and a new training sample set is obtained;

[0132] In D5, according to the new training sample set, the first updating algorithm is repeatedly executed to optimize the deep learning model, the operation is repeated, the model convergence error is continuously monitored, until the preset convergence condition is met, and the final updated model parameters are obtained.

[0133] In an optional embodiment, the first updating algorithm can be federated learning, the model is trained on multiple edge computing nodes, each node uses local data for training to obtain local model parameters. All local model parameters are sent to the cloud, and the cloud computing node aggregates these parameters to obtain global model parameters; the cloud returns the updated global model parameters to the edge computing node, and the edge computing node continues to train using the new global model parameters; repeat the above process until the model parameters converge;

[0134] In an optional embodiment, the first updating algorithm can be transfer learning, obtaining a pre-trained model from an existing similar task that has been trained on a large amount of data and has good feature extraction capability; using the parameters of the pre-trained model as initial parameters to fine-tune the cloud deep learning model and only update the last few layers of the model; using new data on the cloud computing node to train the model and adjust the model parameters to adapt to the new task; repeating the above process until the performance of the model on the new task reaches the expectation;

[0135] In an optional embodiment, the first updating algorithm includes an incremental model updating algorithm.

[0136] Specifically, in the optimization process of the cloud deep learning model, first, the performance fluctuation of the model under different data distributions is identified through potential risk probability analysis, for example, the accuracy of the model may decrease from 95% to 85% in a certain scenario; in order to solve this problem, an incremental model updating algorithm is used, and the parameter updating is specifically performed using an online gradient descent method, which is expressed as:

[0137]

[0138] wherein θ t is the current model parameter, η is the learning rate, is the gradient of the loss function;

[0139] Suppose the amount of new data is 1000, the amount of data for each small batch training is 100, and the learning rate is set to 0.01; in each training, the model calculates the loss function according to the data of the current batch, and updates the parameters through the gradient descent method, for example, the weight matrix W is updated from the initial value W0 to W1, wherein Through multiple iterations, the model gradually adapts to the new data, and the accuracy on the test set is finally improved to 92%;

[0140] In order to further optimize the model, regularization techniques such as L2 regularization are used, and the regularization coefficient λ is set to 0.001 to prevent overfitting; in the updating process, the change trend of the model parameters is monitored through the loss function curve to ensure the stability of the training process; finally, the updated model is deployed in the cloud and can process new data in real time and maintain high prediction performance.

[0141] It should be noted that through online learning and incremental updating algorithm, the model parameters are optimized in real time, the new data is quickly adapted, the accuracy and response speed of the model are improved, and the efficiency and reliability of the early warning system are ensured.

[0142] In the embodiment of the present application, the step S500 comprises the following sub-steps E1-E6.

[0143] In E1, the original model parameters are obtained from the edge computing node, the distributed computing architecture is used for sharding processing of the parameters, and the neural network model parameters are obtained.

[0144] In E2, the neural network model parameters are encrypted through a secure encryption protocol, and the encrypted parameters are generated by using an encryption algorithm.

[0145] In E3, if the encrypted parameters pass the integrity check, the encrypted parameters are sent to the edge computing node through the parameter transmission process, and the transmission successful parameters are obtained.

[0146] In E4, according to the transmission successful parameters, the model deployment update is performed on the edge computing node, and the preliminary optimized model is generated.

[0147] In E5, the edge device performance data is obtained for the preliminary optimized model, if the performance data meets the fifth threshold value, the real-time parameter processing is performed, and the performance adaptation model is obtained.

[0148] In E6, through the interaction of the performance adaptation model and the edge computing node, the model parameter fine-tuning is performed by using the distributed computing architecture, and the optimized edge model is obtained.

[0149] In an optional embodiment, the encryption algorithm can be a homomorphic encryption algorithm, the gradient parameters are encrypted by using the homomorphic encryption algorithm, the edge node sends the homomorphic encrypted local calculation gradient to the aggregation server, the server performs weighted average operation on the encrypted gradient without decryption, the server sends the aggregated encrypted gradient back to the edge node, and the edge node uses the private key to decrypt and update the local model.

[0150] In an optional embodiment, the encryption algorithm can be differential privacy, noise is added to the gradient in the gradient calculation process, the edge node adds differential privacy noise after calculating the local gradient, and then sends the noise gradient to the aggregation server; the server aggregates the noise gradient, and even if the aggregated gradient is leaked, the original data cannot be restored due to the addition of noise; the edge node decrypts and updates the local model after receiving the aggregated gradient.

[0151] In the embodiment of the present application, the encryption algorithm comprises a Paillier homomorphic encryption algorithm.

[0152] In the model parameter updating process, first, the lightweight sub-model parameters are extracted through gradient clipping and quantization compression technology, for example, the convolution layer weight of ResNet18 is quantized to 8-bit integer (range-128 to 127), and the original 32-bit floating point parameter is compressed from 11MB to 2.75MB; the fifth threshold adopts the threshold-based parameter screening algorithm, and the convolution kernel parameters with L1 norm higher than 1e-4 (about 63% of the total) are retained; at the same time, K-means clustering (k=16) is used for clustering compression of the full connection layer parameters; the parameter feedback is realized through secure multi-party computation protocol, and Paillier homomorphic encryption algorithm (key length 2048 bits) is used to encrypt the gradient parameters, and federal average algorithm (learning rate 0.01, local iteration 5 times) is used for model aggregation on the edge node side; according to the difference of the computing power of the edge device, the model slicing strategy is dynamically adjusted, such as Raspberry Pi 4B is allocated MobileNetV2 reverse residual module (expansion factor 6, channel number 96), and Jetson Xavier is allocated complete EfficientNet-B0 module; the encryption transmission process adopts AES-256-GCM mode (group length 128 bits, Nonce length 96 bits) to ensure data security, and the actual transmission delay is controlled within 300ms (bandwidth 10Mbps); finally, the knowledge of the teacher model (accuracy 92.3%) is migrated to the student model (accuracy 89.1% after reducing 58% of the parameter quantity) through knowledge distillation (temperature coefficient T=2, KL divergence loss weight 0.7) at the edge, and the edge model optimization is completed, and the knowledge of the teacher model is migrated to the student model, which is expressed as:

[0153] L=αL CE +(1-α)L KL

[0154] Wherein, L CE is the cross-entropy loss, L KL is the KL divergence loss, and α is the weight coefficient.

[0155] It should be noted that through lightweight, encrypted transmission and distributed optimization, efficient model parameter updating and safe feedback are realized, which adapts to the performance of edge devices and improves the inference efficiency and accuracy of edge models.

[0156] In the embodiment of the application, the above steps S600, S700 include the following sub-steps F1-F6;

[0157] In F1: obtain a time series data segment, segment the time series data segment to obtain a segmented time series, and infer each data segment using the optimized edge model to generate a segment-level feature vector and obtain a feature representation;

[0158] In F2, if the feature representation matches the preset hazard pattern, the feature vector is compared with the sixth threshold value to determine whether there is a potential hazard, and a hazard judgment result is obtained;

[0159] In F3, according to the hazard judgment result, a time series continuous monitoring mechanism is adopted, combined with the moving step length of the sliding window, to obtain a subsequent data segment, and an updated time series is obtained;

[0160] In F4, through the updated time series, the optimized edge model is used to perform fast reasoning again to generate a new feature vector, and a continuous feature representation is obtained;

[0161] In F5, if the continuous feature representation shows that the hazard trend continues, the preset warning rule is used to generate a real-time hazard warning result, determine the warning level, and use the output mechanism of the edge computing node to transmit the warning result to the target system, and output the real-time hazard warning result;

[0162] In F6, if the real-time hazard warning result shows that the risk probability exceeds the seventh threshold value, it is transmitted to the control center to obtain the final electrical safety hazard warning instruction.

[0163] Specifically, on the edge computing node, the optimized edge model is used to perform fast reasoning on the real-time data stream, first the sliding window technology is used to continuously monitor the time series data; the size of the sliding window is set to 10 seconds, and it is slid every 1 second, and the window contains 1000 data points;

[0164] The LSTM (Long Short Term Memory) algorithm is used to extract features and identify patterns from the data in the window, the hidden layer of LSTM is set to 128 units, and the input layer is 10 feature dimensions; the trained model is used to infer the data in each window, and the hazard warning probability is output; when the warning probability exceeds the seventh threshold value of 0.8, real-time warning is triggered; at the same time, the Kalman filter algorithm is used to smooth the inference result to reduce noise interference and ensure the accuracy of the warning result; the state transition matrix of Kalman filter is set to the unit matrix, the observation matrix is the diagonal matrix, and the initial state covariance is 0.1;

[0165] It should be noted that the system can monitor and warn potential hazards in real time, providing a reliable basis for subsequent decision-making.

[0166] Specifically, the running data of the electrical equipment is collected through the sensor network to generate a real-time hazard data stream;

[0167] If the risk probability in the real-time hidden danger data stream exceeds the preset threshold, a support vector machine algorithm is used to classify the data to determine a high-risk state; a low-delay communication interface is used to transmit a warning signal of the high-risk state to a control center to generate a preliminary warning instruction; historical hidden danger data is obtained from the control center database according to the preliminary warning instruction, and a random forest algorithm is used to analyze the data mode to determine the hidden danger type; if the hidden danger type is determined to be an electrical safety hidden danger, a final warning instruction is generated through a preset rule and output to an execution terminal; feedback data of the execution terminal is obtained according to the final warning instruction, and time series analysis is used to judge the instruction execution effect; if the instruction execution effect does not reach the preset standard, an adjustment signal is returned to the control center through the communication interface to generate an optimized warning instruction;

[0168] For example, the real-time hidden danger warning system collects temperature, current, harmonic and other parameters through the sensor network deployed on the electrical equipment, the sampling frequency is set to 1 kHz, the sliding window algorithm is used to perform fast Fourier transform analysis on the sampling data of every 10 ms to analyze the harmonic distortion rate, and when the 5th harmonic content exceeds 4.2%, the preliminary warning is triggered.

[0169] The system uses a risk probability calculation model based on a Bayesian network, inputs six characteristic parameters such as temperature gradient change rate (e.g. 2.3°C per minute increase) and current fluctuation standard deviation (more than 15% of the rated value) into a pre-trained neural network for multi-dimensional fusion analysis, and when the output probability value exceeds the preset threshold of 0.78, a warning message containing the device ID, risk level (classified into 1-5 levels), and positioning coordinates is transmitted to the control center through the RS-485 bus at a baud rate of 115200 bps within 50 ms.

[0170] The decision engine of the control center uses an improved D-S evidence theory algorithm to comprehensively infer the historical fault database (containing more than 3000 cases) and the real-time topology analysis result, if three or more adjacent nodes simultaneously appear three or more warnings, the final instruction containing the isolation instruction (such as disconnecting the circuit breaker CB-1024) and the load transfer scheme (switching 150kW load to standby line L-208) is generated, and is transmitted to the on-site execution unit through the fiber ring network with a delay of less than 10ms; the joint inference of the historical data and the real-time data is represented as:

[0171]

[0172] Wherein, S i is the similarity of historical data and real-time data.

[0173] It should be noted that the total delay of the entire processing flow from data acquisition to instruction generation is strictly controlled within 200ms, ensuring millisecond-level response to rapidly developing hidden dangers such as arc faults.

[0174] The above is a schematic scheme of the cluster type electrical safety hidden danger prediction method of the embodiment. It should be noted that the technical scheme of the cluster type electrical safety hidden danger prediction system belongs to the same concept as the technical scheme of the cluster type electrical safety hidden danger prediction method described above. The technical scheme of the cluster type electrical safety hidden danger prediction system in the embodiment is not described in detail. Details can be seen from the description of the technical scheme of the cluster type electrical safety hidden danger prediction method described above.

[0175] The cluster type electrical safety hidden danger prediction system in the embodiment comprises:

[0176] The data processing module is configured to obtain a real-time data stream, perform data processing on the real-time data stream, and obtain a first data stream.

[0177] The feature extraction module is configured to perform feature extraction using an edge computing node according to the first data stream, and filter abnormal features to obtain an abnormal feature set.

[0178] The feature modeling module is configured to analyze the abnormal feature set using a deep learning model, model time series features in the abnormal feature set through a neural network model, and obtain a potential risk probability of an electrical hidden danger.

[0179] The parameter updating module is configured to optimize the deep learning model parameters using an updating algorithm according to the potential risk probability, and obtain updated model parameters.

[0180] The model optimization module is configured to extract neural network model parameters related to the edge computing node in the updated model parameters, and return the neural network model parameters to the edge computing node to obtain an optimized edge model.

[0181] The inference module is configured to perform inference on the real-time data stream according to the optimized edge model, and obtain a real-time hidden danger early warning result.

[0182] The judgment module is configured to judge the real-time hidden danger early warning result, and obtain an electrical safety hidden danger early warning instruction according to a judgment result.

[0183] The embodiment also provides a computer device suitable for cluster type electrical safety hidden danger prediction, comprising:

[0184] The memory is configured to store computer executable instructions, and the processor is configured to execute the computer executable instructions to implement a cluster type electrical safety hidden danger prediction method as described in the above embodiment.

[0185] The embodiment also provides a storage medium, which stores a computer program, and the computer program is executed by a processor to implement the cluster electrical safety hidden danger prediction method proposed in the above embodiment.

[0186] The storage medium proposed in the embodiment and the cluster electrical safety hidden danger prediction method proposed in the above embodiment belong to the same inventive concept, and the technical details not described in the embodiment can be referred to the above embodiment, and the embodiment has the same beneficial effects as the above embodiment.

[0187] Through the above description of the embodiments, those skilled in the art can clearly understand that the present application can be realized by means of software and necessary general hardware, and of course can also be realized by hardware. Based on such understanding, the technical solutions of the present application or the part that contributes to the prior art can be embodied in the form of a software product. The computer software product can be stored in a computer readable storage medium, such as a floppy disk, a read-only memory (ROM), a random access memory (RAM), a FLASH memory, a hard disk, or an optical disc, and includes a number of instructions to make a computing device (which can be a personal computer, a server, or a network device, etc.) execute the methods of various embodiments of the present application.

[0188] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application but not limit the present application. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced by equivalents without departing from the spirit and scope of the present application, and all of them should be covered in the scope of the claims of the present application.

Claims

1. A cluster type electrical safety hazard pre-judgment method, characterized in that, The method comprises the following steps: acquiring a real-time data stream, performing data processing on the real-time data stream, and obtaining a first data stream; performing feature extraction on the first data stream by using an edge computing node, screening abnormal features, and obtaining an abnormal feature set; analyzing the abnormal feature set by using a deep learning model, modeling time sequence features in the abnormal feature set by using a neural network model, and obtaining a potential risk probability of an electrical hidden danger; optimizing the parameters of the deep learning model by using an updating algorithm according to the potential risk probability, and obtaining updated model parameters; extracting neural network model parameters related to the edge computing node in the updated model parameters, and returning the neural network model parameters to the edge computing node to obtain an optimized edge model; performing reasoning on the real-time data stream according to the optimized edge model, and obtaining a real-time hidden danger early warning result; judging the real-time hidden danger early warning result, and obtaining an electrical safety hidden danger early warning instruction according to a judgment result.

2. The method of claim 1, wherein the method comprises: The data processing on the real-time data stream comprises the following steps: acquiring original data from the real-time data stream, pre-processing the original data by using a first data compression algorithm, and obtaining an encoded data stream; reducing data packets in the encoded data stream to below a first threshold by using a first quantization coding technology, and obtaining the first data stream.

3. The method of claim 2, wherein the method further comprises: The abnormal feature set comprises the following steps: acquiring data packets from the first data stream, decompressing the data packets by using the edge computing node, and obtaining an electrical parameter sequence; performing time sequence analysis on the electrical parameter sequence by using a first neural network model, obtaining a feature vector set, and if any feature in the feature vector set exceeds a second threshold, obtaining an abnormal candidate set; performing time sequence correlation analysis on the abnormal candidate set by using a first flow control technology, obtaining a first abnormal feature set, classifying abnormal types by using a first classifier, and obtaining an abnormal category; extracting high-confidence abnormalities from the abnormal category, distinguishing abnormal conditions by using clustering analysis, obtaining an abnormal detection result, generating time sequence distribution data of abnormal features, and outputting a final abnormal feature set.

4. The cluster electrical safety hazard pre-judgment method of claim 3, wherein, The potential risk probability of the electrical hidden danger comprises the following steps: performing feature extraction on the abnormal feature set by using a pre-trained deep learning model of a cloud computing node, and outputting a high-dimensional feature vector; performing time sequence modeling on the high-dimensional feature vector by using a neural network model, and generating a dynamic change trend of the time sequence; if the fluctuation amplitude of the high-dimensional feature vector exceeds a third threshold, calculating the potential risk probability of the electrical hidden danger according to the dynamic change trend; extracting risk factors from the potential risk probability, classifying the risk factors by using a clustering analysis method, and obtaining a high-risk factor set; judging whether there is a match with a known electrical hidden danger mode by comparing the high-risk factor set with historical data, and outputting a matching result; classifying the types and severity of the electrical hidden danger according to the matching result by using a decision tree algorithm, and obtaining a hidden danger classification result.

5. The cluster electrical safety hazard pre-judgment method of claim 4, wherein, The updated model parameters comprise the following steps: acquiring new data from the real-time data stream by using an online learning technology, calculating a risk probability of a data sample by using a first training method, and obtaining a risk assessment value; If the risk evaluation value is greater than the fourth threshold value, the first updating algorithm is used to adjust the deep learning model parameters, the inference of the deep learning model is performed, the prediction result of the current data stream is generated, and the model output value is determined; By comparing the prediction result with the actual data, the model convergence error is calculated, and it is judged whether the preset convergence condition is met to obtain an error evaluation result; If the error evaluation result does not meet the convergence condition, the sample selection strategy is adjusted, the data samples in the real-time data stream are re-acquired, and a new training sample set is obtained; According to the new training sample set, the first updating algorithm is repeatedly executed to optimize the deep learning model, and the operation is repeated to continuously monitor the model convergence error until the preset convergence condition is met, and the final updated model parameters are obtained.

6. The cluster type electrical safety hazard pre-judgment method according to claim 5, characterized in that, The optimized edge model includes: Obtain the original model parameters from the edge computing node, perform sharding processing on the parameters using a distributed computing architecture, and obtain the neural network model parameters; The neural network model parameters are encrypted through a secure encryption protocol, and an encryption algorithm is used to generate encrypted parameters; If the encrypted parameters pass the integrity check, the encrypted parameters are sent to the edge computing node through the parameter transmission process to obtain the transmission successful parameters; According to the transmission successful parameters, the model deployment update is performed on the edge computing node to generate a preliminary optimized model; For the preliminary optimized model, the edge device performance data is obtained, and if the performance data meets the fifth threshold value, real-time parameter processing is performed to obtain a performance adaptation model; Through the interaction of the performance adaptation model and the edge computing node, the model parameter fine-tuning is performed using the distributed computing architecture to obtain the optimized edge model.

7. The cluster type electrical safety hazard pre-judgment method according to claim 6, characterized in that, The real-time hidden danger early warning result includes: Obtain the time series data segment, segment the time series data segment to obtain a segmented time series, and use the optimized edge model to perform inference on each data segment to generate a segment-level feature vector to obtain a feature representation; If the feature representation matches the preset hidden danger mode, compare the feature vector with the sixth threshold value to determine whether there is a potential hidden danger, and obtain a hidden danger judgment result; According to the hidden danger judgment result, a continuous monitoring mechanism of the time series is used in combination with the moving step of the sliding window to obtain subsequent data segments to obtain an updated time series; Through the updated time series, the optimized edge model is used to perform rapid inference again to generate a new feature vector to obtain a continuous feature representation; If the continuous feature representation shows that the hidden danger trend continues, the real-time hidden danger early warning result is generated through the preset early warning rule, the warning level is determined, the output mechanism of the edge computing node is used to transmit the warning result to the target system, and the real-time hidden danger early warning result is output; If the real-time hidden danger early warning result shows that the risk probability exceeds the seventh threshold value, it is transmitted to the control center to obtain the final electrical safety hidden danger early warning instruction.

8. A cluster electrical safety hazard pre-judgment system, applying a cluster electrical safety hazard pre-judgment method according to any one of claims 1-7, characterized in that, It includes: A data processing module is configured to obtain a real-time data stream, process the real-time data stream, and obtain a first data stream; A feature extraction module is configured to extract features from the first data stream using an edge computing node and filter abnormal features to obtain an abnormal feature set; The feature modeling module is configured to analyze the abnormal feature set by using a deep learning model, model time series features in the abnormal feature set by using a neural network model, and obtain a potential risk probability of the electrical hidden danger. The parameter updating module is configured to optimize the deep learning model parameters by using an updating algorithm according to the potential risk probability, and obtain updated model parameters. The model optimization module is configured to extract neural network model parameters related to the edge computing node from the updated model parameters, and return the neural network model parameters to the edge computing node to obtain an optimized edge model. The inference module is configured to perform inference on the real-time data stream according to the optimized edge model, and obtain a real-time hidden danger early warning result. The judgment module is configured to judge the real-time hidden danger early warning result, and obtain an electrical safety hidden danger early warning instruction according to a judgment result.

9. A computer device, comprising: The memory and the processor, the memory stores a computer program, the processor executes the computer program to realize the steps of the cluster type electrical safety hidden danger pre-judgment method in any one of claims 1 to 7. The computer program is stored on the memory and executed by the processor to realize the steps of the cluster type electrical safety hidden danger pre-judgment method in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, ​

Images