Cloud-side collaborative power distribution intelligent terminal state monitoring method and device, terminal equipment and storage medium

Abstract

The invention discloses a cloud-side collaborative power distribution intelligent terminal state monitoring method and device, terminal equipment and a storage medium, and relates to the field of power distribution intelligent terminal monitoring. Comprise voltage, current, temperature, humidity, processor utilization rate, memory utilization rate, disk read-write rate and network bandwidth utilization rate; performing feature extraction on each feature data in the multi-modal feature data to obtain a corresponding feature vector; determining a query matrix, a key matrix and a value matrix according to all the feature vectors and a preset weight matrix, and further determining a weighted value matrix; and performing average pooling on the weighted value matrix to obtain a fusion feature vector, and sending the fusion feature vector to an anomaly detection model of the cloud, so that the model determines the state of the power distribution intelligent terminal according to the fusion feature vector. By implementing the method and the device, the problem of high bandwidth occupation when the edge device processes the multi-modal data in the prior art can be solved, and the response speed is improved.

Description

Technical Field

[0001] The present invention relates to the field of monitoring of distribution intelligent terminals, and particularly to a cloud-edge collaborative method, device, terminal device, and storage medium for monitoring the state of distribution intelligent terminals. Background Art

[0002] With the rapid development of Internet of Things and edge computing technologies, intelligent terminal devices play an important role in various application scenarios of smart grids. Currently, in the distribution system of smart grids, many key tasks, such as distributed energy management, fault detection and location, and intelligent load control, usually rely on cloud platforms for data processing and analysis. Existing technologies generally monitor data on edge devices and transmit the monitored data to the cloud platform for analysis through Internet of Things connections.

[0003] However, if the edge device simultaneously collects multi-modal data such as voltage, current, temperature, and humidity data of the device, and each modal data is sent to the cloud platform for data processing and analysis, it will occupy a large bandwidth, unable to respond to real-time monitoring requirements in a timely manner, or directly freeze and cannot work properly. Summary of the Invention

[0004] Embodiments of the present invention provide a cloud-edge collaborative method, device, terminal device, and storage medium for monitoring the state of distribution intelligent terminals, which can solve the problem of high bandwidth occupation of distribution intelligent terminals by edge devices when processing multi-modal data in the prior art and improve the response speed.

[0005] An embodiment of the present invention provides a cloud-edge collaborative method for monitoring the state of distribution intelligent terminals, including:

[0006] Obtain multi-modal feature data of the distribution intelligent terminal; the multi-modal feature data includes: voltage, current, temperature, humidity, processor usage rate, memory usage rate, disk read-write rate, and network bandwidth utilization rate;

[0007] Extract features from each feature data in the multi-modal feature data of the distribution intelligent terminal to obtain corresponding feature vectors;

[0008] Determine a query matrix, a key matrix, and a value matrix according to all the feature vectors and a preset weight matrix;

[0009] Determine a weighted value matrix according to the query matrix, the key matrix, and the value matrix;

[0010] Perform average pooling on the weighted value matrix to obtain a fused feature vector;

[0011] Send the fused feature vector to an anomaly detection model in the cloud so that the anomaly detection model determines the state of the distribution intelligent terminal according to the fused feature vector.

[0012] Further, after obtaining the multi-modal feature data of the distribution intelligent terminal, it further includes:

[0013] Perform filtering processing on the multi-modal feature data to obtain the filtered multi-modal feature data;

[0014] Perform normalization processing on the filtered multi-modal feature data to obtain the preprocessed multi-modal feature data.

[0015] Further, perform feature extraction on each feature data in the multi-modal feature data of the distribution intelligent terminal to obtain the corresponding feature vector, including:

[0016] For each feature data in the multi-modal feature data, perform local feature extraction on the feature data to generate a local feature vector;

[0017] Perform a non-linear transformation on the local feature vector to generate a non-linear feature vector;

[0018] Perform a pooling operation on the non-linear feature vector to generate a dimensionality-reduced feature vector;

[0019] Perform feature mapping on the dimensionality-reduced feature vector to generate the feature vector corresponding to the feature data.

[0020] Further, after obtaining the fused feature vector, it further includes:

[0021] Perform normalization processing on the fused feature vector to obtain the final fused feature vector.

[0022] Further, after determining the state of the distribution intelligent terminal, it further includes:

[0023] According to the state of the distribution intelligent terminal, adopt corresponding countermeasures.

[0024] Further, according to the fused feature vector, determine the state of the distribution intelligent terminal, including:

[0025] Through the built-in long short-term memory network module, perform feature extraction on the hidden layer of the fused feature vector to generate the hidden state of the last time step;

[0026] Through the built-in support vector machine regression module, perform optimization and solution on the hidden state of the last time step to generate a risk prediction probability result; the risk prediction probability result includes: low risk probability, medium risk probability, and high risk probability;

[0027] According to the risk prediction probability result, determine the state of the distribution intelligent terminal; the state includes: low risk, medium risk, or high risk.

[0028] Further, the anomaly detection model is determined by the following method:

[0029] Obtain a number of training samples; each training sample includes: a fused feature vector sample and its corresponding actual risk label;

[0030] Input a number of training samples into the anomaly detection model to be trained, so that the anomaly detection model to be trained generates a sample risk prediction probability result according to the fused feature vector sample in each training sample; the sample risk prediction probability result includes: sample low - risk probability, sample medium - risk probability, and sample high - risk probability; calculate the loss function according to the sample risk prediction probability result, the corresponding fused feature vector sample, and its corresponding actual risk label, and adjust the parameters of the anomaly detection model according to the loss function until the loss function converges to obtain a trained anomaly detection model.

[0031] Based on the above - mentioned method - item embodiments, the present invention correspondingly provides device - item embodiments, including: an edge - end data acquisition module, an edge - end feature extraction module, an attention matrix determination module, a weighted value matrix determination module, an edge - end feature fusion module, and a cloud - end detection module;

[0032] The edge - end data acquisition module is used to acquire multi - modal feature data of the distribution power intelligent terminal; the multi - modal feature data includes: voltage, current, temperature, humidity, processor usage rate, memory usage rate, disk read - write rate, and network bandwidth utilization rate;

[0033] The edge - end feature extraction module is used to perform feature extraction on each feature data in the multi - modal feature data of the distribution power intelligent terminal to obtain the corresponding feature vector;

[0034] The attention matrix determination module is used to determine a query matrix, a key matrix, and a value matrix according to all feature vectors and a preset weight matrix;

[0035] The weighted value matrix determination module is used to determine a weighted value matrix according to the query matrix, the key matrix, and the value matrix;

[0036] The edge - end feature fusion module is used to perform average pooling on the weighted value matrix to obtain a fused feature vector;

[0037] The cloud - end detection module is used to send the fused feature vector to the anomaly detection model in the cloud, so that the anomaly detection model determines the state of the distribution power intelligent terminal according to the fused feature vector.

[0038] Based on the above - mentioned method - item embodiments, the present invention correspondingly provides terminal - device - item embodiments, including: a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, the steps of the method for monitoring the state of the distribution power intelligent terminal with cloud - edge collaboration as described in the present invention are implemented.

[0039] Based on the above method embodiment, the present invention correspondingly provides a computer-readable storage medium embodiment, including: a stored computer program, which controls the device where the computer-readable storage medium is located to execute the steps of the cloud-edge collaborative power distribution intelligent terminal status monitoring method as described in the present invention when the computer program runs.

[0040] Compared with the prior art, the beneficial effects of the embodiment of the present solution are as follows:

[0041] The present invention obtains multi-modal feature data of a power distribution intelligent terminal, and the multi-modal feature data includes voltage, current, temperature, humidity, processor utilization rate, memory utilization rate, disk read-write rate, and network bandwidth utilization rate. Feature extraction is performed on each feature data in the multi-modal feature data of the power distribution intelligent terminal, converting the original high-dimensional data into feature vectors. Through feature extraction, key features in the data can be screened out, reducing the scale of the data to be processed subsequently. Then, the feature vectors of different modalities are converted into query matrices, key matrices, and value matrices, and different weights are assigned to different feature vectors to obtain a weighted value matrix, enabling the edge device to focus on key information. Finally, average pooling is performed to obtain a fused feature vector, and the fused feature vector can comprehensively reflect the information of different modality data, realizing efficient compression of multi-modal data and focusing on key information. Finally, the fused feature vector is sent to the anomaly detection model in the cloud so that the anomaly detection model determines the status of the power distribution intelligent terminal according to the fused feature vector. Since feature extraction and fusion have been completed on the edge device, the amount of data sent to the cloud is greatly reduced, reducing the demand for network bandwidth and also reducing the computing burden on the cloud, improving the processing efficiency of the entire system, thereby solving the problem of high bandwidth occupancy of the power distribution intelligent terminal when the edge device processes multi-modal data in the prior art. Description of the Drawings

[0042] Figure 1 is a schematic flowchart of the cloud-edge collaborative power distribution intelligent terminal status monitoring method provided by an embodiment of the present invention;

[0043] Figure 2 is a schematic flowchart of the training process of the anomaly detection model provided by an embodiment of the present invention;

[0044] Figure 3 is a schematic structural diagram of the cloud-edge collaborative power distribution intelligent terminal status monitoring device provided by an embodiment of the present invention. Detailed Embodiments

[0045] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0046] The smart grid system has strict requirements for "high real-time performance, low latency, and high reliability". During the operation of the smart grid, a large amount of data needs to be processed quickly and accurately to ensure the stable operation and efficient management of the power grid. Although edge computing technology can perform real-time processing on edge terminals close to the data source, significantly reducing latency and the demand for data transmission bandwidth, if all computing tasks are assigned to edge terminals for execution, it may lead to increased energy consumption of edge devices, and when dealing with high-concurrency tasks, the computing resources (such as CPUs and memory) of edge terminals may face the risk of overload, thereby affecting the overall performance and stability of the system.

[0047] As Figure 1 shown, to solve the problem of high bandwidth occupancy of distribution intelligent terminals by edge devices in processing multimodal data in the prior art, an embodiment of the present invention provides a cloud-edge collaborative method for monitoring the state of distribution intelligent terminals, which at least includes the following steps:

[0048] Step S1: Obtain multimodal feature data of the distribution intelligent terminal; the multimodal feature data includes: voltage, current, temperature, humidity, processor usage rate, memory usage rate, disk read-write rate, and network bandwidth utilization rate;

[0049] For step S1, first, monitor the data of the distribution intelligent terminal to collect the multimodal feature data of the distribution intelligent terminal; the multimodal feature data includes environmental data and device data. Specifically, the environmental data includes voltage, current, temperature, and humidity, which are obtained in real time through various high-precision sensors deployed around the distribution intelligent terminal and are used to evaluate the operating conditions of the device, timely discover potential overheating or humidity hazards, thereby ensuring the stable operation of the device; the device data includes the usage rate of the processor (CPU), memory usage rate, disk read-write rate (also known as disk I / O read-write rate), and network bandwidth utilization rate, which are obtained by real-time monitoring of the distribution intelligent terminal through specialized monitoring tools. The finally obtained multimodal feature data set of the distribution intelligent terminal can be expressed as M = {M1, M2, M3,..., M m}, where M1, M2, M3,..., M m respectively represent the 1st, 2nd, 3rd,..., mth type of feature data in the multimodal feature data, and m represents the number of feature data in the multimodal feature data.

[0050] In a preferred embodiment, after obtaining the multimodal feature data of the distribution intelligent terminal, it further includes:

[0051] Perform filtering processing on the multimodal feature data to obtain the filtered multimodal feature data;

[0052] Perform normalization processing on the filtered multimodal feature data to obtain the preprocessed multimodal feature data.

[0053] In an embodiment of the present invention, it should be noted that the present invention adopts a sampling strategy based on load frequency control to preprocess environmental data, and dynamically adjusts the sampling frequency by monitoring the change of the grid frequency. The grid frequency is a key parameter affecting the stable operation of the system, and the fluctuation of the frequency reflects the change of the grid load. When the grid frequency fluctuates greatly, the sampling frequency is increased to obtain more refined data, so as to ensure that the system can respond quickly. The calculation formula of the sampling frequency is:

[0054]

[0055] where f s represents the real-time sampling frequency, f base represents the basic sampling frequency, β is the sampling frequency adjustment coefficient, f rated represents the rated grid frequency, Δf represents the deviation between the current grid frequency and the rated frequency, that is, Δf = f current -f rated where f current is the currently monitored grid frequency.

[0056] Perform noise reduction preprocessing on the sampled data, that is, perform filtering processing to reduce the influence of redundant noise in data transmission. In this embodiment, Kalman filtering is used to implement noise reduction processing:

[0057] The formula in the prediction stage of the noise reduction processing is:

[0058]

[0059] The formula in the update stage of the noise reduction processing is:

[0060]

[0061] where represents the signal data after Kalman filtering processing at time t, represents the predicted value of the signal data at time t for time t - 1, Z(t) represents the sampled original signal data, t represents the current sampling time point, Z(t) can refer to the voltage parameter V(t), current parameter I(t), temperature parameter T(t), or humidity parameter A(t), F represents the state transition matrix, used to describe the transition relationship of the system state from time t - 1 to time t, Q represents the process noise covariance, R represents the observation noise covariance, and K(t) represents the Kalman gain. represents the estimated error covariance. represents the predicted value of the estimated error covariance at time t for time t - 1.

[0062] By combining the dynamic sampling strategy based on load frequency control and Kalman filtering, the efficiency and accuracy of data acquisition of distribution intelligent terminals can be effectively improved, while reducing the noise interference in data transmission, providing support for the stable operation of the smart grid.

[0063] Next, the filtered multi-modal feature data is normalized. Among them, the following formula is used to preprocess the environmental data by the maximum-minimum normalization method:

[0064]

[0065] where x i is the i-th data point in the original data, x min and x max are the minimum and maximum values in the original data respectively, and x i ′ is the normalized data.

[0066] The following formula is used to preprocess the device data by the Z-score normalization method:

[0067]

[0068] where x i is the i-th data point in the original data, μ X is the mean of the original data, σ X is the standard deviation of the original data, and x″ i is the normalized data, where is the mean of the set X, is the standard deviation of the set X.

[0069] In this embodiment, the collected voltage value set is V = {V1, V2,..., V n}, and V′1, V′2,..., V′ n represent the voltage sequences at the 1st, 2nd,..., n-th time steps respectively, where V nwhere n is the number of voltage sequences. Using the maximum-minimum normalization method, the minimum value V in this set is determined min and the maximum value V max . Then, according to the formula , each voltage value V i is normalized. The preprocessed voltage data set is V′ = {V′1, V′2,..., V′ n}, where V′ represents the preprocessed voltage data set, and V′1, V′2,..., V′ n represent the voltage sequences at the 1st, 2nd,..., nth time steps after preprocessing, respectively.

[0070] The collected current value set is I = {I1, I2,..., I n}, and I′1, I′2,..., I′ n represent the current sequences at the 1st, 2nd,..., nth time steps, respectively. Among them, n in I n is the number of current sequences. Using the maximum-minimum normalization method, the minimum value I min and the maximum value I max in this set are determined. Then, according to the formula , each current value I i is normalized. The preprocessed current data set is I′ = {I′1, I′2,..., I′ n}, where I′ represents the preprocessed current data set, and I′1, I′2,..., I′ n represent the current sequences at the 1st, 2nd,..., nth time steps after preprocessing, respectively.

[0071] The collected temperature value set is T = {T1, T2,..., T n}, and T1, T2,..., T n represent the temperature sequences at the 1st, 2nd,..., nth time steps, respectively. Among them, n in T n is the number of temperature sequences. Using the maximum-minimum normalization method, after determining the minimum value T min and the maximum value T max , the formula is used to normalize each temperature value T i . The preprocessed temperature data set is T′ = {T′1, T′2,..., T′ n}, where T′ represents the preprocessed temperature data set, and T′1, T′2,..., T′ n represent the temperature sequences at the 1st, 2nd,..., nth time steps after preprocessing, respectively.

[0072] The set of collected humidity values is AH = {AH1, AH2,..., AH n}, where AH1, AH2,..., AH n represent the humidity sequences at the 1st, 2nd,..., nth time steps respectively. Among them, n in AH n is the number of humidity sequences. Using the maximum-minimum normalization method, after determining the minimum value AH min and the maximum value AH max , the formula is used to normalize each humidity value AH i . The set of humidity data after preprocessing is AH′ = {AH′1, AH′2,..., AH′ n}, where AH′ represents the set of preprocessed humidity data, and AH′1, AH′2,..., AH′ n represent the humidity sequences at the 1st, 2nd,..., nth time steps after preprocessing respectively.

[0073] The set of collected CPU usage data is CPU i = {CPU1, CPU2, ·····, CPU n}, where CPU1, CPU2, …, CPU n represent the CPU usage sequences at the 1st, 2nd,..., nth time steps respectively. Among them, n in CPU n is the number of CPU usage sequences. Using the Z-Score method for normalization, calculate the mean μ CPU and the standard deviation σ CPU of the CPU data, and then according to the formula normalize each CPU usage value CPU i . The set of CPU usage data after preprocessing is CPU i ′ = {CPU1′, CPU2′,...., CPU n ′}, where CPU i ′ represents the set of preprocessed CPU usage data, and CPU1′, CPU2′,...., CPU n ′ represent the CPU usage sequences at the 1st, 2nd,..., nth time steps after preprocessing respectively.

[0074] The set of collected memory usage data is Mem = {Mem1, Mem2, ·····, Mem n}, where Mem1, Mem2, ······, Mem n represent the memory usage sequences at the 1st, 2nd,..., nth time steps respectively. Among them, Mem nwhere n is the number of memory usage rate sequences. The Z-Score method is used for normalization to calculate the mean μ of the memory usage rate data Mem and the standard deviation σ Mem , and then according to the formula each memory usage rate value Mem i is normalized. The preprocessed memory usage rate data set is Mem′ = {Mem′1, Mem′2,...., ·Mem′ n} where Mem′ represents the preprocessed memory usage rate data set, and Mem′1, Mem′2,...., ·Mem′ n represent the memory usage rate sequences at the 1st, 2nd,..., nth time steps after preprocessing respectively.

[0075] The collected disk I / O read / write rate data set is Disk = {Disk1, Disk2, …, Disk n}, where Disk1, Disk2, …, Disk n represent the disk I / O read / write rate sequences at the 1st, 2nd,..., nth time steps respectively. Here, n in Disk n is the number of disk I / O read / write rate sequences. The Z-Score method is used for normalization to calculate the mean μ Disk and the standard deviation σ Disk , and then according to the formula each disk I / O read / write rate value Disk i is normalized. The preprocessed disk I / O read / write rate data set is Disk′ = {Disk′1, Disk′2, …, Disk′ n}, where Disk′ represents the preprocessed disk I / O read / write rate data set, and Disk′1, Disk′2, …, Disk′ n represent the disk I / O read / write rate sequences at the 1st, 2nd,..., nth time steps after preprocessing respectively.

[0076] The collected network bandwidth utilization data set is η = {η1, η2, …, η n}, where η1, η2, …, η n represent the network bandwidth utilization sequences at the 1st, 2nd,..., nth time steps respectively. Here, n in η n is the number of network bandwidth utilization sequences. The Z-Score method is used for normalization to calculate the mean μ η and the standard deviation σ η , and then according to the formula each network bandwidth utilization value ηi Perform normalization processing. After preprocessing, the network bandwidth utilization data set is η′ = {η′1, η′2, …, η′ n}, where η′ represents the network bandwidth utilization data set after preprocessing, and η′1, η′2, …, η′ n respectively represent the network bandwidth utilization sequences at the 1st, 2nd, …, nth time steps after preprocessing.

[0077] To clearly represent the situation where the number of sequences of each feature data in the multi-modal feature data may be different, the symbol n k is introduced to represent the number of sequences of the kth feature data in the multi-modal feature data, where k = 1, 2, …, m, and m represents the number of feature data in the multi-modal feature data.

[0078] It should be noted that the multi-modal feature data includes voltage, current, temperature, humidity, processor usage rate, memory usage rate, disk read / write rate, and network bandwidth utilization; the above-listed data types are not all, and according to specific application scenarios and requirements, other types of data may also be included, such as light intensity, sound intensity, or vibration frequency, etc. At the same time, in actual applications, it may not be necessary to use all types of data simultaneously, and some data can be selected for monitoring and analysis according to specific requirements.

[0079] Step S2: Extract features from each feature data in the multi-modal feature data of the distribution intelligent terminal to obtain corresponding feature vectors;

[0080] In a preferred embodiment, extracting features from each feature data in the multi-modal feature data of the distribution intelligent terminal to obtain corresponding feature vectors includes:

[0081] For each feature data in the multi-modal feature data, perform local feature extraction on the feature data to generate local feature vectors;

[0082] Perform a non-linear transformation on the local feature vectors to generate non-linear feature vectors;

[0083] Perform a pooling operation on the non-linear feature vectors to generate dimensionality-reduced feature vectors;

[0084] Perform feature mapping on the dimensionality-reduced feature vectors to generate the feature vectors corresponding to the feature data.

[0085] For step S2, a series of convolutional layers, activation functions, pooling layers, and fully connected layers in a convolutional neural network (CNN) are used to gradually extract local features and patterns in the input data to obtain the feature vectors of each feature data in the multi-modal feature data.

[0086] Among them, the convolutional layer is the core part of the CNN and is used to extract local features from the input data. The convolution operation slides a convolution kernel (or filter) over the input data, calculates the dot product of the convolution kernel and the local region of the input data, and generates a feature map.

[0087] Let the input data be a d-dimensional vector z, and the convolutional layer uses k convolution kernels W of size f×d l (l = 1, 2, …, k) to perform the convolution operation. For the l-th convolution kernel, the output h of the convolution operation l The calculation formula is:

[0088]

[0089] where is the i-th element of the l-th convolution kernel, y i+s is the (i + s)-th element of the input data y starting from position s, and b l is the bias term of the l-th convolution kernel. The convolution operation slides over the input data with a stride of s, obtaining a series of convolution results.

[0090] The activation function is used to introduce non-linearity so that the network can learn complex features. In this embodiment, the activation function is the ReLU function, which performs a non-linear transformation on the convolution result to obtain the output of the activated convolutional layer. The ReLU function is defined as:

[0091] a l = max(0, h l )

[0092] where a l is the final output of the convolutional layer, and h l is the output of the convolution operation.

[0093] The pooling layer is applied to reduce the spatial dimension of the feature map, reduce the amount of computation and the number of parameters, while retaining important features. In this embodiment, the pooling method is max pooling. The pooling window size is p×p and the stride is s p . For the feature map a l output by the convolutional layer, the output p l of the max pooling operation is calculated as:

[0094]

[0095] where R is the region covered by the pooling window, is the element at position (i, j) in the feature map a l .

[0096] The fully connected layer integrates the local features extracted by the convolutional layer and the pooling layer to generate a global feature representation. The output of the pooling layer is flattened into a one-dimensional vector and then input into the fully connected layer. The input vector of the fully connected layer is v, and the output vector is X k , and the weight matrix of the fully connected layer is W f , and the bias vector is b f , then the output calculation formula of the fully connected layer is:

[0097] X k =W f ·v + b f

[0098] Among them, X k is the output vector of the fully connected layer, W f is the weight matrix of the fully connected layer, v is the input vector of the fully connected layer, and b f is the bias vector of the fully connected layer.

[0099] The output after passing through the fully connected layer is the extracted feature vector, that is, the feature vector sequence after feature extraction for each feature data in the multi-modal feature data is:

[0100]

[0101] Among them, X k represents the set of feature vectors after feature extraction for the k-th feature data in the multi-modal feature data, k = 1, 2,..., m, where m represents the number of feature data in the multi-modal feature data, represents the feature vector at the i-th time step of the k-th feature data in the multi-modal feature data, and n k represents the number of sequences of the k-th feature data in the multi-modal feature data.

[0102] It should be noted that for some simple feature extraction tasks, such as monitoring the real-time operating status of equipment (such as fluctuations in current and voltage), it can be directly completed by the CNN model on the distribution intelligent terminal, providing real-time data support for power grid monitoring quickly; for complex feature extraction tasks, such as in-depth analysis of long-term performance indicators of equipment (such as CPU usage rate, memory usage rate), more representative feature vectors need to be extracted, and then the extracted features are fed back to the distribution intelligent terminal for further processing.

[0103] Step S3: Determine the query matrix, key matrix, and value matrix according to all feature vectors and the preset weight matrix;

[0104] For step S3, after obtaining m feature vectors obtained by the CNN processing the multi-modal feature data, for each feature vector sequence, the query vector Q k and the key vector Kk Sum vector V k :

[0105] Q k = X k W Q K k = X k W K V k = X k W V

[0106] where W Q represents the weight matrix of the query vector, W K represents the weight matrix of the key vector, W V represents the weight matrix of the value vector. These weight matrices are obtained through model training and learning. Q k represents the query vector of the k-th feature vector sequence, K k represents the key vector of the k-th feature vector sequence, V k represents the value vector of the k-th feature vector sequence.

[0107] Concatenate the query vectors of all sequences into a query matrix Q, and the key vectors into a key matrix K:

[0108] Q = [Q 1 ; Q 2 ;...; Q m K = [K 1 ; K 2 ;...; K m

[0109] It should be noted that the query matrix Q, the key matrix K, and the value matrix V are the core components of the attention mechanism. Through these matrices, the model can effectively process and fuse multi-modal feature data, thereby improving the overall performance. The present invention can more effectively capture long-range dependency relationships and complex feature interactions in the data by introducing the attention mechanism.

[0110] Step S4: Determine the weighted value matrix according to the query matrix, the key matrix, and the value matrix;

[0111] For step S4, calculate the attention score matrix according to the attention score formula;

[0112] A = {A 1,1 , A 1,2 ,..., A i,j}

[0113]

[0114] where A represents the attention score matrix, A i,j ​represents an element in the attention score matrix, representing the attention score between the $i$-th row of the query matrix $Q$ and the $j$-th row of the key matrix $K$. represents the transpose of the $i$-th row of the query vector $Q$, $k$ j represents the $j$-th row of the key vector $K$, $d$ k represents the dimension of the key vector $K$, $i$ represents the row index of the query matrix $Q$, $i = 1, 2, \ldots, m$, $j$ represents the row index of the key matrix $K$, $j = 1, 2, \ldots, m$, and $m$ represents the number of feature vectors.

[0115] In multimodal data processing, data of different modalities may have different feature representations. By calculating the similarity between the query matrix and the key matrix, the features of different modalities can be aligned and fused together, which helps the model to understand and process multimodal data more comprehensively.

[0116] Next, the attention score matrix $A$ is normalized by sequence grouping through the following formula to obtain the attention weight matrix For the query position $i$ belonging to the $k$-th sequence, its attention weight is:

[0117]

[0118] where represents an element in the attention weight matrix, representing the normalized attention score between the $i$-th row of the query matrix $Q$ and the $j$-th row of the key matrix $K$, $\exp(\cdot)$ represents the exponential function, which is used to calculate the softmax normalization, $seq$ k represents the $k$-th sequence grouping, and $j'$ represents the index in the sequence grouping.

[0119] Concatenate all value vectors into the value matrix $V$:

[0120] $V = [V$ 1 ; $V$ 2 ; \ldots; $V$ m

[0121] According to the attention weight matrix and the value matrix, the weighted value matrix of the multimodal feature data of the distribution intelligent terminal is:

[0122]

[0123] $Z = [Z$ 1 ; $Z$ 2 ; \ldots; $Z$ m

[0124] where $Z$ represents the weighted value matrix, represents the attention weight matrix, and $V$ represents the value matrix;

[0125] ​​Step S5: Perform average pooling on the weighted value matrix to obtain a fused feature vector;

[0126] For step S5, the fused feature vector of the multimodal data of the distribution power intelligent terminal is obtained by performing average pooling on the weighted value vector Z through the following formula:

[0127]

[0128] where f is the fused feature vector, and n k represents the sequence number of the k-th feature data in the multimodal feature data, k = 1, 2,..., m, m represents the number of feature data in the multimodal feature data, and z i represents the i-th row vector of the weighted value matrix.

[0129] Through the average pooling operation, the data information of different modalities can be integrated, so that the final fused feature vector can more comprehensively reflect the state of the distribution power intelligent terminal. In addition, average pooling reduces the dimension of the data by calculating the average value of the local area, thereby reducing the computational complexity.

[0130] In a preferred embodiment, after obtaining the fused feature vector, it further includes:

[0131] Perform normalization processing on the fused feature vector to obtain the final fused feature vector.

[0132] In an embodiment of the present invention, the maximum value max(f) and the minimum value min(f) in the fused feature vector f are calculated;

[0133] Each element f in the fused feature vector f is normalized through the following formula i for normalization:

[0134]

[0135] where, represents the i-th element in the final fused feature vector, min(f) and max(f) represent the minimum value and the maximum value in the fused feature vector, and f i represents the i-th element in the fused feature vector. Finally, the final fused feature vector f new is formed.

[0136] By performing normalization processing on the fused feature vector, all features can be scaled to the same scale, making the feature distribution more uniform, which helps the model to better learn the relationship between features, thereby improving the prediction performance of the model.

[0137] Step S6: Send the fused feature vector to the anomaly detection model in the cloud, so that the anomaly detection model determines the state of the distribution power intelligent terminal according to the fused feature vector.

[0138] In an embodiment of the present invention, determining the state of the distribution intelligent terminal according to the fusion feature vector includes:

[0139] Extracting the features of the hidden layer from the fusion feature vector through the built-in long short-term memory network module to generate the hidden state of the last time step;

[0140] Optimally solving the hidden state of the last time step through the built-in support vector machine regression module to generate a risk prediction probability result; the risk prediction probability result includes: low risk probability, medium risk probability, and high risk probability;

[0141] Determining the state of the distribution intelligent terminal according to the risk prediction probability result; the state includes: low risk, medium risk, or high risk.

[0142] For step S6, the distribution intelligent terminal sends the fusion feature vector to the anomaly detection model deployed in the cloud, and through the built-in long short-term memory network module (LSTM), calculates according to the following formula at each time step t:

[0143] Input gate: i t = σ(W ii f norm,t + W hi h t-1 + b i )

[0144] Forget gate: f t = σ(W if f norm,t + W hf h t-1 + b f )

[0145] Cell state update:

[0146]

[0147] Output gate: o t = σ(W io f norm,t + W ho h t-1 + b o )

[0148] Hidden state update: h t = o t ⊙ tanh(C t )

[0149] where σ(·) represents the sigmoid activation function, i tRepresents the output value of the input gate at time step t, f norm,t Represents the value of the normalized feature vector at time step t, W ii Represents the input feature f norm,t To the weight matrix of the input gate, h t-1 Represents the hidden state of the LSTM cell at time step t - 1, W hi Represents the hidden state h at the previous moment t-1 To the weight matrix of the input gate, b i Represents the bias vector of the input gate, f t Represents the output value of the forget gate at time step t, W if Represents the input feature f norm,t To the weight matrix of the forget gate, W hf Represents the hidden state h at the previous moment t-1 To the weight matrix of the forget gate, b f Represents the bias vector of the forget gate, The candidate cell state at time step t, maps the input to the interval (-1, 1) through the hyperbolic tangent function tanh(·), W ic Represents the input feature f norm,t To the weight matrix of the candidate cell state, W hc Represents the hidden state h at the previous moment t-1 To the weight matrix of the candidate cell state, b c Represents the bias vector of the candidate cell state, C t Represents the updated cell state at time step t, o t Represents the output value of the output gate at time step t, W io Represents the input feature f norm,t To the weight matrix of the output gate, W ho Represents the hidden state h at the previous moment t-1 To the weight matrix of the output gate, b o Represents the bias vector of the output gate, h t Represents the hidden state of the LSTM cell at time step t, which is the final output feature representation.

[0150] Take the hidden state h of the last time step of the long short-term memory network module (LSTM) final As the extracted feature, and input the hidden state h of the last time step final Into the built-in support vector machine regression module (SVR), and output the risk prediction probability results. The risk prediction probability results include low-risk probability, medium-risk probability, and high-risk probability. Finally, according to the probability magnitudes in the risk prediction probability results, determine the state of the distribution intelligent terminal. The states of the distribution intelligent terminal are divided into low risk (R = 1), medium risk (R = 2), and high risk (R = 3).

[0151] In a preferred embodiment, after determining the state of the distribution intelligent terminal, it further includes:

[0152] According to the state of the distribution intelligent terminal, corresponding countermeasures are taken.

[0153] In an embodiment of the present invention, when the state of the distribution intelligent terminal is determined to be a low risk (R = 1), the cloud performs regular inspections and maintenance on the distribution terminal equipment and the power system where it is located according to the first preset time interval (such as once a week) to ensure the stable operation of the equipment. At the same time, continuously collect new multi-modal feature data for subsequent model update and optimization to improve the accuracy of model prediction.

[0154] When the state of the distribution intelligent terminal is determined to be a medium risk (R = 2), the cloud increases the monitoring frequency and performs regular inspections and maintenance on the distribution terminal equipment and the power system where it is located according to the second preset time interval (such as once a day or once every half day). It should be noted that the second preset time interval is less than the first preset time interval; secondly, perform a performance evaluation on the distribution terminal equipment and the power system where it is located, and analyze potential risk factors, such as judging whether there is an overload trend in the equipment and whether the insulation performance has decreased through data analysis. Organize an expert team in the power field to conduct discussions and formulate targeted optimization plans. When necessary, adjust some parameters, such as adjusting the protection setting value of the equipment, to ensure the safe operation of the equipment.

[0155] When the state of the distribution intelligent terminal is determined to be a high risk (R = 3), the cloud immediately activates the emergency plan and takes emergency shutdown measures to avoid further expansion of equipment failures and cause greater losses. Conduct a comprehensive inspection and fault troubleshooting on the distribution terminal equipment and the power system where it is located, and use professional inspection equipment and technical means to find the fault point and the cause of the fault. After repairing the fault, re-collect data and input it into the model for risk assessment to ensure that the system returns to a safe state before resuming normal operation.

[0156] Next, a detailed description of the training of the anomaly detection model will be given:

[0157] As Figure 2 shown, the training of the above anomaly detection model includes the following steps:

[0158] Step S61: Obtain a number of training samples; each training sample includes: a fused feature vector sample and its corresponding actual risk label;

[0159] For step S61, a number of training samples are obtained. Each training sample includes a fused feature vector sample and its corresponding actual risk label. Among them, the fused feature vector sample is obtained by feature extraction of historical multi-modal feature data. The actual risk label includes low risk, medium risk or high risk, and is the label corresponding to each fused feature vector sample, indicating the actual risk level of the sample.

[0160] Step S62: Input a number of training samples into the anomaly detection model to be trained, so that the anomaly detection model to be trained generates sample risk prediction probability results according to the fused feature vector samples in each training sample. The sample risk prediction probability results include: sample low risk probability, sample medium risk probability and sample high risk probability. Calculate the loss function according to the sample risk prediction probability results, the corresponding fused feature vector samples and their corresponding actual risk labels, and adjust the parameters of the anomaly detection model according to the loss function until the loss function converges to obtain the trained anomaly detection model.

[0161] For step S62, first, an anomaly detection model to be trained is constructed in the cloud. This model is a prediction model composed of a long short-term memory network (LSTM) and a support vector machine regression (SVR), so it is also called an LSTM-SVR combined prediction model. Multi-modal feature data often has time series characteristics. For example, data such as temperature and humidity at different times have a sequence and dependence relationship. LSTM is specifically used to process time series data. It can capture long-term dependence relationships in the data through its unique gating mechanism, avoiding the problem of gradient disappearance or gradient explosion that traditional recurrent neural networks encounter when processing long sequences. And SVR is good at regression prediction in the case of small samples. Even when the amount of training data is limited, it can map the data to a high-dimensional space through a kernel function, find an optimal hyperplane for regression fitting, and thus achieve accurate prediction.

[0162] Input a number of training samples into the anomaly detection model to be trained in the cloud. Normalize the fused feature vector samples in the number of training samples, and input the feature vector samples after normalization into the built-in long short-term memory network module. Process time series data through the forget gate, input gate, output gate, memory unit and hidden state. Through the coordinated work of these gates, LSTM can effectively capture long-term dependence relationships in time series data and extract useful features.

[0163] Specifically, for the forget gate: the calculation formula of the forget gate is f t =σ(W f ·[h t-1 ,x t +b f ). Among them, W fis the weight matrix, whose role is to perform a linear transformation on the hidden state h at the previous moment of the input t-1 and the input x at the current moment t to adjust the influence degree of different input features on the output of the forget gate; b f is the bias vector, adding a constant offset to the result of the linear transformation to make the model have stronger expressive ability; σ is the sigmoid function, which compresses the result of the linear transformation between 0 and 1. The output value f t represents the proportion of information discarded from the cell state C t-1 . For example, when processing temperature data, if the temperature at the current moment has little correlation with the temperature at a previous moment, the forget gate may output a value close to 0, thus discarding the cell state information corresponding to the temperature at that previous moment.

[0164] For the input gate: The calculation formula of the input gate is i t =σ(W i ·[h t-1 , x t +b i ). Similarly, W i and b i are the weight matrix and the bias vector respectively, and the σ function maps the result of the linear transformation between 0 and 1. The output value i t controls the proportion of new information input to the cell state at the current moment. For example, when new humidity data is input, the input gate will decide whether to add the relevant information of the humidity data to the cell state according to the input at the current moment and the hidden state at the previous moment.

[0165] For the output gate: The calculation formula of the output gate is o t =σ(W o ·[h t-1 , x t +b i ). The roles of W o and b o are similar to those above. o t determines which information in the cell state C t will be output to the hidden state h t . For example, after processing a series of multi-modal data, the output gate will screen out the most valuable information in the cell state for the current prediction and output it to the hidden state as the basis for subsequent processing.

[0166] For the memory unit: The update formula of the memory unit is C t =f t *C t-1 +i t *tanh(W c ·[h t-1 , xt +b c ). Here, f t *C t-1 represents the part of the previous cell state C that the forget gate allows to be retained, and i t-1 *tanh(W t ·[h c ,x t-1 +b t ) represents adding the newly input information at the current time after processing to the cell state. The tanh function maps the result of the linear transformation to the range between -1 and 1. After element-wise multiplication with the input gate output i c , it is added to the retained cell state of the previous time step to obtain the updated cell state C t . In this way, the memory cell can continuously update and store the historical information of the data. t .

[0167] For the hidden state: The calculation formula for the hidden state is h t = o t *tanh(C t ). Through the output gate o t , the updated cell state C t is filtered to output the hidden state h t at the current time, which is used as the output result of the LSTM at the current time and also as one of the inputs to the LSTM unit at the next time. The hidden state contains the feature extraction and understanding results of the LSTM for the input time series data up to the current time.

[0168] It is trained using the Backpropagation Through Time (BPTT) algorithm. The first loss function is:

[0169]

[0170] where y i represents the actual risk label of the i-th sample, n represents the number of training samples, and h final,i represents the hidden state of the last time step of the i-th sample.

[0171] Next, the hidden state of the last time step in each training sample extracted is input into the built-in support vector machine regression module. By solving the following optimization problem:

[0172]

[0173] The constraint conditions are:

[0174] where w represents the weight vector of the support vector machine regression module, w Trepresents the transpose of the weight vector of the support vector machine regression module, C represents the penalty parameter. The larger the value, the more severe the penalty for error, and the model may be more complex. n represents the number of samples, ξ i and represents the slack variable, y i represents the true abnormal label, φ(h final,i ) represents the input feature of the support vector machine regression module. In this embodiment, the hidden state h of the last time step of LSTM is final,i The feature vector transformed by the mapping function φ(·), ∈ represents the parameter of the insensitive loss function, and b represents the bias term of the support vector machine regression module, which is used to adjust the baseline value of linear prediction.

[0175] Determine the parameters of the support vector machine regression module and obtain the prediction function:

[0176]

[0177] in, Represents the support vector machine regression module for the input feature h final The predicted value of i and represents the Lagrange multiplier, b represents the bias term of the support vector machine regression module, K(h final ,h final,i ) represents the kernel function, which is used to calculate the input feature h final With the training sample feature h final,i The similarity between .

[0178] In this embodiment, according to actual data characteristics and experience, the penalty parameter C, the parameter ∈ of the insensitive loss function, etc. are set so that the support vector machine regression module can accurately predict the input features.

[0179] In the cloud, the multimodal feature vectors of a large number of samples are input into the constructed LSTM-SVR combined prediction model. The model predicts based on the input feature vectors and outputs the sample risk prediction probability results, including the sample low risk probability, sample medium risk probability and sample high risk probability. Based on the sample multimodal feature vectors, sample risk prediction probability results and the corresponding actual risk labels, the second loss function is calculated using the cross entropy loss function combined with the regularization term:

[0180]

[0181] Among them, L represents the loss function value, which is used to measure the difference between the model prediction result and the true label, N is the number of samples, and y ij Represents the true risk label of the i-th sample, which is a ternary label. If it is a low risk y i1 =1,yi2 = 0, y i3 = 0; If it is medium risk, y i1 = 0, y i2 = 1, y i3 = 0; If it is high risk, y i1 = 0, y i2 = 0, y i3 = 1, p ij represents the probability that the i-th sample is predicted as the j-th risk, j = 1, 2, 3, and λ is the regularization coefficient used to control the weight of the regularization term ||θ k || 2 to prevent the model from overfitting, and θ k represents the parameters in the model.

[0182] Through optimization algorithms such as stochastic gradient descent, continuously adjust the parameters of the LSTM-SVR combined prediction model according to the second loss function until the model converges, and obtain a trained anomaly detection model.

[0183] Preferably, the present invention can also evaluate the state of the distribution intelligent terminal through the following method:

[0184] Use the predicted value as the input, and by constructing multiple decision trees, calculate the path length of each data point in the tree. The path length can reflect the complexity and uncertainty of the data point in the decision tree. Calculate the anomaly score according to the path length. The longer the path length, usually the more complex or abnormal the data point is. If the anomaly score exceeds the set threshold, it is determined as abnormal.

[0185] Alternatively, in the cloud, according to the anomaly recognition result, combined with historical data and domain knowledge, use a logistic regression model. Take the abnormal data and other relevant features as the input, and use the formula to calculate the risk occurrence probability, where ω represents the weight vector, x represents the input feature vector, and b represents the bias. Determine the state of the distribution intelligent terminal according to the risk occurrence probability through the following formula:

[0186]

[0187] Among them, R = 1 means that the state of the distribution intelligent terminal is at a low risk level, and at this time, conventional monitoring and maintenance measures are taken; R = 2 means that the state of the distribution intelligent terminal is at a medium risk level, and it is necessary to pay close attention and adjust and optimize in time; the risk level R = 3 means that the state of the distribution intelligent terminal is at a high risk level, and emergency measures such as shutdown for maintenance and activation of emergency plans should be taken immediately.

[0188] Such as Figure 3 shown, on the basis of the above method item embodiments, corresponding device item embodiments are provided;

[0189] An embodiment of the present invention provides a cloud-edge collaborative power distribution intelligent terminal status monitoring device, including: an edge-end data acquisition module, an edge-end feature extraction module, an attention matrix determination module, a weighted value matrix determination module, an edge-end feature fusion module, and a cloud detection module;

[0190] The edge-end data acquisition module is used to acquire multi-modal feature data of the power distribution intelligent terminal; the multi-modal feature data includes: voltage, current, temperature, humidity, processor utilization rate, memory utilization rate, disk read-write rate, and network bandwidth utilization rate;

[0191] The edge-end feature extraction module is used to perform feature extraction on each feature data in the multi-modal feature data of the power distribution intelligent terminal to obtain corresponding feature vectors;

[0192] The attention matrix determination module is used to determine a query matrix, a key matrix, and a value matrix according to all the feature vectors and a preset weight matrix;

[0193] The weighted value matrix determination module is used to determine a weighted value matrix according to the query matrix, the key matrix, and the value matrix;

[0194] The edge-end feature fusion module is used to perform average pooling on the weighted value matrix to obtain a fused feature vector;

[0195] The cloud detection module is used to send the fused feature vector to an anomaly detection model in the cloud, so that the anomaly detection model determines the status of the power distribution intelligent terminal according to the fused feature vector.

[0196] It can be understood that the above device item embodiment corresponds to the method item embodiment of the present invention, and it can implement the cloud-edge collaborative power distribution intelligent terminal status monitoring method provided by any one of the above method item embodiments of the present invention.

[0197] It should be noted that the above-described device embodiments are merely illustrative, and some or all of the modules can be selected according to actual needs to achieve the purpose of the present embodiment solution. In addition, in the accompanying drawings of the device embodiments provided by the present invention, the connection relationships between the modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those of ordinary skill in the art can understand and implement without creative efforts.

[0198] Based on the embodiments of the above-mentioned cloud-edge collaborative power distribution intelligent terminal status monitoring method, another embodiment of the present invention provides a terminal device, which includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the cloud-edge collaborative power distribution intelligent terminal status monitoring method of any embodiment of the present invention.

[0199] Exemplarily, in this embodiment, the computer program may be divided into one or more modules. The one or more modules are stored in the memory and executed by the processor to complete the present invention. The one or more modules may be a series of computer program instruction segments capable of performing specific functions, and these instruction segments are used to describe the execution process of the computer program in the terminal device.

[0200] The terminal device may be a computing device such as a desktop computer, a notebook, a palm computer, and a cloud server. The terminal device may include, but is not limited to, a processor and a memory.

[0201] The so-called processor may be a central processing unit (CPU), or may also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc. The processor is the control center of the terminal device, and uses various interfaces and lines to connect various parts of the entire terminal device.

[0202] Based on the above method item embodiments, another embodiment is provided: A storage medium provided by another embodiment of the present invention includes a stored computer program, wherein when the computer program runs, it controls the device where the storage medium is located to execute the cloud-edge collaborative power distribution intelligent terminal status monitoring method described in any one of the above method item embodiments of the present invention.

[0203] Among them, for the modules / units integrated in the cloud-edge collaborative power distribution intelligent terminal status monitoring device / terminal equipment, when 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 such an understanding, to implement all or part of the processes in the above-mentioned embodiment methods of the present invention, it can also be completed by instructing relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by a processor, the steps of the above-mentioned various method embodiments can be realized. Among them, the computer program includes computer program code, and the computer program code can be in the form of source code, object code, executable file or some intermediate form, etc. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electrical carrier signal, telecommunication signal, and software distribution medium, etc.

[0204] The above is the preferred embodiment of the present invention. It should be noted that for those of ordinary skill in the art of the present technology, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements are also regarded as the protection scope of the present invention.

Claims

1. A method for monitoring the status of a distribution intelligent terminal with cloud-edge collaboration, characterized in that, Including: Obtain multi-modal feature data of the distribution intelligent terminal; The multi-modal feature data includes: voltage, current, temperature, humidity, processor usage rate, memory usage rate, disk read / write rate, and network bandwidth utilization rate; Extract features from each feature data in the multi-modal feature data of the distribution intelligent terminal to obtain corresponding feature vectors; Determine the query matrix, key matrix, and value matrix according to all feature vectors and a preset weight matrix; Determine the weighted value matrix according to the query matrix, key matrix, and value matrix; Perform average pooling on the weighted value matrix to obtain a fused feature vector; Send the fused feature vector to the anomaly detection model in the cloud so that the anomaly detection model determines the state of the distribution intelligent terminal according to the fused feature vector.

2. The cloud-edge collaborative power distribution intelligent terminal status monitoring method according to claim 1, wherein After obtaining the multi-modal feature data of the distribution intelligent terminal, it further includes: Perform filtering processing on the multi-modal feature data to obtain filtered multi-modal feature data; Perform normalization processing on the filtered multi-modal feature data to obtain preprocessed multi-modal feature data.

3. The cloud-edge collaborative power distribution intelligent terminal status monitoring method according to claim 1, characterized in that Extract features from each feature data in the multi-modal feature data of the distribution intelligent terminal to obtain corresponding feature vectors, including: For each feature data in the multi-modal feature data, perform local feature extraction on the feature data to generate local feature vectors; Perform a non-linear transformation on the local feature vectors to generate non-linear feature vectors; Perform a pooling operation on the non-linear feature vectors to generate dimensionality-reduced feature vectors; Perform feature mapping on the dimensionality-reduced feature vectors to generate feature vectors corresponding to the feature data.

4. The cloud-edge collaborative power distribution intelligent terminal status monitoring method according to claim 1, wherein After obtaining the fused feature vector, it further includes: Perform normalization processing on the fused feature vector to obtain the final fused feature vector.

5. The cloud-edge collaborative power distribution intelligent terminal status monitoring method according to claim 1, wherein, After determining the state of the distribution intelligent terminal, it further includes: Adopt corresponding coping strategies according to the state of the distribution intelligent terminal.

6. The cloud-edge collaborative power distribution intelligent terminal status monitoring method according to claim 1, wherein Determine the state of the distribution intelligent terminal according to the fused feature vector, including: Extract features of the hidden layer from the fused feature vector through the built-in long short-term memory network module to generate the hidden state of the last time step; Optimize and solve the hidden state of the last time step through the built-in support vector machine regression module to generate a risk prediction probability result; the risk prediction probability result includes: low-risk probability, medium-risk probability, and high-risk probability; Determine the state of the distribution intelligent terminal according to the risk prediction probability result; the state includes: low risk, medium risk, or high risk.

7. The cloud-edge collaborative power distribution intelligent terminal status monitoring method according to claim 6, characterized in that, The anomaly detection model is determined by the following method: Obtain a number of training samples; each training sample includes: a fused feature vector sample and its corresponding actual risk label; Input a number of training samples into the anomaly detection model to be trained, so that the anomaly detection model to be trained generates sample risk prediction probability results according to the fused feature vector samples in each training sample; the sample risk prediction probability results include: sample low-risk probability, sample medium-risk probability, and sample high-risk probability; calculate the loss function according to the sample risk prediction probability results, the corresponding fused feature vector samples, and their corresponding actual risk labels, and adjust the parameters of the anomaly detection model according to the loss function until the loss function converges to obtain a trained anomaly detection model.

8. A state monitoring device for a distribution intelligent terminal with cloud-edge collaboration, characterized in that, including: an edge-side data acquisition module, an edge-side feature extraction module, an attention matrix determination module, a weighted value matrix determination module, an edge-side feature fusion module, and a cloud detection module; The edge-side data acquisition module is used to acquire multi-modal feature data of the distribution intelligent terminal; The multi-modal feature data includes: voltage, current, temperature, humidity, processor utilization rate, memory utilization rate, disk read-write rate, and network bandwidth utilization rate; The edge-side feature extraction module is used to extract features from each feature data in the multi-modal feature data of the distribution intelligent terminal to obtain corresponding feature vectors; The attention matrix determination module is used to determine a query matrix, a key matrix, and a value matrix according to all feature vectors and a preset weight matrix; The weighted value matrix determination module is used to determine a weighted value matrix according to the query matrix, the key matrix, and the value matrix; The edge-side feature fusion module is used to perform average pooling on the weighted value matrix to obtain a fused feature vector; The cloud detection module is used to send the fused feature vector to the anomaly detection model in the cloud, so that the anomaly detection model determines the state of the distribution intelligent terminal according to the fused feature vector.

9. A terminal device, characterized in that, including: a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, the cloud-edge collaborative distribution intelligent terminal status monitoring method according to any one of claims 1-7 is implemented.

10. A storage medium, characterized in that, including: a stored computer program, wherein when the computer program runs, it controls the device where the storage medium is located to execute the cloud-edge collaborative distribution intelligent terminal status monitoring method according to any one of claims 1-7.

Images