Electric vehicle charging safety early warning method based on cloud side-end cooperation

Abstract

The invention discloses an electric vehicle charging safety early warning method based on cloud side end collaboration, and relates to the technical field of electric vehicle charging safety. A cloud side end collaboration architecture is adopted, and a charging terminal sensor cluster collects charging data in real time and transmits the charging data to an edge end; the edge end processes data through an intelligent charging data analysis algorithm to obtain abnormal data and uploads the abnormal data to the cloud end, and the cloud end establishes a charging safety early warning model for global evaluation and early warning. According to the invention, real-time collection and transmission of charging data are realized, and the data processing efficiency is improved; edge end preprocessing reduces cloud pressure, and network bandwidth consumption and data delay are reduced; abnormal features are accurately extracted through an intelligent algorithm, and the early warning accuracy is improved; the cloud global evaluation realizes the overall control of the safety condition of the charging network; the three-level alarm mechanism ensures that potential safety hazards of different degrees can be effectively processed in time, the real-time performance, accuracy and reliability of electric vehicle charging safety early warning are comprehensively improved, and the safety of the charging process is guaranteed.

Description

Technical Field

[0001] The present invention relates to the technical field of electric vehicle charging safety, and more specifically, to an electric vehicle charging safety warning method based on cloud-edge-terminal collaboration. Background Art

[0002] Currently, with the popularization of electric vehicles, the problem of charging safety has become increasingly prominent. Most of the existing charging safety warning methods rely on a single cloud server or local device for data processing and analysis, resulting in problems such as large data transmission delays, low processing efficiency, and untimely warnings. Although the cloud server has powerful computing and storage capabilities, uploading a large amount of charging data to the cloud in real time will cause great pressure on network bandwidth, and there is a certain delay in data processing, making it difficult to achieve real-time warning of charging safety. On the other hand, the processing ability of local devices is limited, and they are unable to comprehensively analyze and judge complex charging data, easily resulting in misjudgment or missed judgment.

[0003] Therefore, there is an urgent need for a cloud-edge-terminal collaborative electric vehicle charging safety warning method that can combine the advantages of the cloud, edge, and terminal to improve the accuracy and timeliness of charging safety warnings. Summary of the Invention

[0004] In view of this, the present invention provides an electric vehicle charging safety warning method based on cloud-edge-terminal collaboration to solve the problems in the background art.

[0005] To achieve the above object, the present invention adopts the following technical solutions:

[0006] An electric vehicle charging safety warning method based on cloud-edge-terminal collaboration includes:

[0007] Using a charging terminal sensor cluster to collect charging data in real time and transmit the charging data to the edge terminal through wireless communication;

[0008] The edge terminal receives the charging data transmitted by the charging terminal sensor cluster, processes the charging data through an intelligent charging data analysis algorithm deployed in the edge device to obtain abnormal data, and the edge terminal uploads the abnormal data to the cloud;

[0009] The cloud receives the data uploaded by the edge terminal, stores the data in the cloud database, extracts the characteristics of the abnormal data, establishes a charging safety warning model in the cloud, globally evaluates and warns the safety status of the entire charging network according to the extracted characteristics, and when a potential safety hazard is found, sends a global warning signal to the relevant edge terminals and terminals through the cloud to notify relevant personnel to take corresponding safety measures.

[0010] Optionally, it further includes performing interpolation processing on the collected charging data, and the formula is as follows:

[0011]

[0012] Among them, L(x) is the calculated value of the missing eigenvalue to be inserted in the charging data, x is the value of the time dimension in the charging data where the missing eigenvalue needs to be inserted, and x j is the value of the time dimension of the j-th charging data, and y j is the value of the recorded eigenvalue of the j-th charging data, and ω j is the centroid weight, k is the number of recorded eigenvalues in the charging data participating in the calculation of the missing eigenvalue L(x) during the interpolation process, and 0 ≤ j ≤ k.

[0013] Optionally, the specific process of the intelligent charging data analysis algorithm is as follows:

[0014] In the native data processing network structure, a lightweight feature extraction model is used as the backbone network to replace the traditional fully connected layer or convolutional layer for extracting key features in the charging data. At the same time, a channel attention module is embedded in the backbone network to enhance the sensitivity to key features of the charging state through adaptive weight allocation;

[0015] In the feature fusion layer of the native network, a hierarchical dynamic attention module is designed and deployed before multi-dimensional data fusion; this module adaptively allocates weights to the feature maps of different sensor data to strengthen the fusion accuracy of charging anomaly features and weaken the interference of irrelevant background data, thereby more accurately extracting the abnormal feature region in the charging state;

[0016] In the analysis and decision-making layer of the native network, a variable convolution module is used to replace the standard convolution. By dynamically adjusting the sampling position of the convolution kernel, it adapts to the detection requirements of irregular abnormal patterns in the charging data; this module learns the local feature offset of the charging data, flexibly expands the receptive field range, enhances the ability to capture unstructured abnormal data, and improves the robustness and accuracy of charging state analysis.

[0017] Optionally, the charging safety warning model in the cloud is specifically as follows:

[0018] Extract the charging boundary features of the charging data collected in the cloud and form a high-dimensional matrix of the charging boundary features;

[0019] Reduce the dimension of the high-dimensional matrix through a machine learning algorithm, extract the low-dimensional feature matrix of the high-dimensional matrix, calculate the current failure probability at each moment during this period based on the low-dimensional feature matrix, and compare it with the preset failure probability threshold to determine whether there is a potential safety hazard currently;

[0020] When a potential safety hazard is detected, decompress the low-dimensional matrix, calculate the contribution value of each dimension in the high-dimensional matrix to the failure probability, and identify the electric vehicle corresponding to the boundary feature with the highest failure probability contribution as the single unit prone to safety hazards.

[0021] Analyze the charging data of the single unit prone to safety hazards, combine it with the deep learning prediction model, calculate the deviation degree between the current state and the predicted state, compare the charging data of the single unit prone to safety hazards with the three-level alarm threshold, and issue an alarm according to the corresponding level of the deviation degree.

[0022] Optionally, the failure probability is obtained by calculating the Hotelling T 2 statistic and the SPE statistic of the low-dimensional matrix, thereby calculating the comprehensive index φ and regressing it to the probability distribution to obtain the failure probability.

[0023] Optionally, the Hotelling T 2 statistic and the SPE statistic are calculated based on the dimensionality reduction process matrix of the charging data training set; among them, the Hotelling T 2 statistic is calculated according to the following formula:

[0024]

[0025] The SPE statistic is calculated according to the following formula:

[0026]

[0027] In the formula, X i is the boundary feature matrix of the i-th module at a certain moment, P k is the dimensionality reduction transformation matrix, composed of the principal eigenvectors, S is the diagonal matrix composed of the principal eigenvalues of the training sample set, I is the identity matrix, and k is the number of principal components.

[0028] Optionally, based on the solution of the Hotelling T 2 and the SPE statistic, according to the comprehensive index obtain the system failure rate function of the sample at the current moment:

[0029]

[0030] In the formula, δ 2 are the control limits of the Hotelling T 2 and SPE respectively, φ is a symmetric positive definite matrix, and the comprehensive index conforms to the chi-square distribution with degrees of freedom h and coefficient g. According to The probability distribution function obtains the current failure probability function The failure probability h is calculated according to the following formula:

[0031]

[0032] Optionally, the degree of deviation is quantified using the relative deviation, and the formula is:

[0033]

[0034] In the formula, the current measured value is X real , and the deep learning predicted value is X pred .

[0035] Optionally, alarms are given according to the corresponding levels of the degree of deviation, specifically:

[0036] The preset three-level alarm thresholds are Thres1, Thres2, and Thres3 respectively, where Thres1 < Thres2 < Thres3. The specific alarm rules are as follows:

[0037] Trigger condition for the first-level alarm: ΔX rel ≥Thres1 and ΔX rel <Thres2

[0038] Alarm strategy:

[0039] Send a warning signal to the charging terminal or the edge side, and light up the yellow warning light;

[0040] Push real-time alarm information to the operation and maintenance personnel through the cloud platform, including deviation parameters, current values, predicted values, and deviation ratios; automatically increase the data collection frequency of this terminal and continuously monitor the trend;

[0041] Trigger condition for the second-level alarm: ΔX rel ≥Thres2 and ΔX rel <Thres3

[0042] Alarm strategy:

[0043] Trigger the orange sound and light alarm, and synchronize the alarm between the local terminal and the cloud;

[0044] Automatically generate an on-site inspection work order and assign nearby operation and maintenance personnel to check the charging equipment connection, battery appearance, and charging terminal sensor cluster;

[0045] Third-level alarm

[0046] Trigger condition: ΔX rel ≥Thres3

[0047] Alarm strategy:

[0048] Immediately send a power-off instruction to the edge device through the cloud to cut off the charging circuit; trigger a red emergency alarm to notify the vehicle owner, the charging station administrator, and the safety supervision department;

[0049] Automatically save the complete charging data for 10 minutes before and after the current moment as the original basis for fault analysis.

[0050] As can be seen from the above technical solutions, compared with the prior art, the present invention discloses a method for early warning of electric vehicle charging safety based on cloud-edge collaboration. It adopts a cloud-edge collaboration architecture. The charging terminal sensor cluster collects charging data in real time and transmits it to the edge device. After the edge device processes the data through an intelligent charging data analysis algorithm to obtain abnormal data, it uploads the data to the cloud. The cloud establishes a charging safety early warning model for global evaluation and early warning. The present invention realizes the real-time collection and transmission of charging data, improves the data processing efficiency; the preprocessing at the edge device reduces the pressure on the cloud, reduces the network bandwidth consumption and data latency; the intelligent algorithm accurately extracts abnormal features, improves the accuracy of early warning; the global evaluation by the cloud realizes the overall control of the safety status of the charging network; the data interpolation processing ensures the integrity of the data; the three-level alarm mechanism ensures that safety hazards of different levels can be handled in a timely and effective manner, comprehensively improving the real-time performance, accuracy, and reliability of the early warning of electric vehicle charging safety and ensuring the safety of the charging process. Brief Description of the Drawings

[0051] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only the embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on the provided drawings.

[0052] Figure 1 It is a schematic flowchart of the method provided by the present invention. Detailed Embodiments

[0053] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the 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. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0054] The embodiment of the present invention discloses a method for early warning of electric vehicle charging safety based on cloud-edge collaboration, as Figure 1 shown, including:

[0055] Utilize the charging terminal sensor cluster to collect charging data in real time and transmit the charging data to the edge device through wireless communication;

[0056] The edge side receives the charging data transmitted by the charging terminal sensor cluster, processes the charging data through the intelligent charging data analysis algorithm deployed in the edge device to obtain abnormal data, and the edge side uploads the abnormal data to the cloud;

[0057] The cloud receives the data uploaded by the edge side, stores the data in the cloud database, extracts the characteristics of the abnormal data, establishes a charging safety warning model in the cloud, globally evaluates and warns the safety status of the entire charging network according to the extracted characteristics, and when a potential safety hazard is found, sends a global warning signal to the relevant edge side and terminal through the cloud to notify relevant personnel to take corresponding safety measures.

[0058] In a specific embodiment, it further includes using the barycentric Lagrange interpolation method to interpolate the collected charging data to solve the situation of incomplete data and missing partial data caused by reasons such as equipment and storage. The formula is as follows:

[0059]

[0060] Among them, L(x) is the calculated value of the missing characteristic value to be inserted in the charging data, x is the value of the time dimension where the missing characteristic value needs to be inserted in the charging data, x j is the value of the time dimension of the jth charging data, y j is the recorded characteristic value of the jth charging data, ω j is the barycentric weight, k is the number of recorded characteristic values in the charging data participating in the calculation of the missing characteristic value L(x) during the interpolation process, and 0 ≤ j ≤ k.

[0061] The advantage of using the barycentric Lagrange interpolation method is that only the function values and barycentric coordinates at the interpolation nodes need to be calculated. Dividing ω j by (x - x j ) can obtain the new barycentric weight, and its computational complexity is O(n), which is simpler than the computational complexity C(n 2 ) before improvement.

[0062] In a specific embodiment, the specific process of the intelligent charging data analysis algorithm is as follows:

[0063] In the native data processing network structure, a lightweight feature extraction model is used as the backbone network to replace the traditional fully connected layer or convolutional layer for extracting key features in the charging data. At the same time, a channel attention module is embedded in the backbone network to enhance the sensitivity to key features of the charging state through adaptive weight allocation;

[0064] In the feature fusion layer of the native network, a hierarchical dynamic attention module is designed and deployed before multi-dimensional data fusion. This module adaptively assigns weights to the feature maps of different sensor data, strengthens the fusion accuracy of charging anomaly features, weakens the interference of irrelevant background data, and thus more accurately extracts the abnormal feature regions in the charging state.

[0065] In the analysis and decision-making layer of the native network, a variable convolution module is used to replace the standard convolution. By dynamically adjusting the sampling position of the convolution kernel, it adapts to the detection requirements of irregular abnormal patterns in the charging data. This module learns the local feature offset of the charging data, flexibly expands the receptive field range, enhances the ability to capture unstructured abnormal data, and improves the robustness and accuracy of charging state analysis.

[0066] Specifically:

[0067] Backbone network: lightweight feature extraction model and channel attention module

[0068] Replace traditional layers: Use lightweight models (such as MobileNet, ShuffleNet) to replace the fully connected layer or convolutional layer, reducing the computational amount while retaining the feature extraction ability.

[0069] Embed channel attention: Add a channel attention module (such as the SE module) to the backbone network to strengthen the key feature channels through adaptive weight assignment.

[0070] Lightweight feature extraction: Reduce the computational load at the edge and adapt to the resource limitations of edge devices.

[0071] Enhance channel sensitivity: Through the weight matrix weight the channel features The formula is: F′ = F × σ(W c · GlobalAvgPool(F)), where σ is the Sigmoid function. By suppressing low-contribution channels (such as temperature fluctuation noise), the sensitivity of key features such as voltage and current is enhanced.

[0072] Feature fusion layer: hierarchical dynamic attention module

[0073] Deployment before multi-dimensional data fusion: For sensor cluster data (such as voltage, current, temperature, humidity), introduce a dynamic attention mechanism before feature map fusion.

[0074] Adaptive weight assignment: Calculate the attention weight α n for the feature maps {F1, F2, …, F i}, and the formula is: where Q is the query vector, and K i is the key vector of the i-th sensor.

[0075] Abnormal feature enhancement: through the weight α i Increase the fusion weight of abnormal features (such as overcharge voltage) and suppress the interference of irrelevant background data (such as normal fluctuations in ambient temperature).

[0076] Fusion accuracy improvement: output the fused feature F fusion = ∑ i α i ·F i , ensuring that the abnormal feature area (such as the signal in the early stage of battery thermal runaway) is accurately extracted.

[0077] Analysis and decision-making layer: deformable convolution module

[0078] Replace standard convolution: adopt deformable convolution, and learn the offset Δp n Dynamically adjust the local feature offset learning of the convolution kernel sampling: for the input feature The sampling points of the standard convolution are {p n}, and the sampling points of the deformable convolution are {p n +Δp n}, where Δp n is predicted by an additional convolution layer.

[0079] Irregular anomaly detection: adapt to the irregular anomaly patterns in the charging data (such as the non-periodic voltage fluctuations of internal micro-short circuits in the battery), and the formula is: where (i,j) are the coordinates of the output feature map, and W n is the convolution kernel weight.

[0080] Dynamic expansion of the receptive field: expand the local receptive field through the offset Δp n Enhance the ability to capture unstructured abnormal data and improve the analysis robustness.

[0081] In a specific embodiment, the charging safety warning model in the cloud is specifically:

[0082] Extract the charging boundary features of the charging data collected in the cloud, and form a high-dimensional matrix with the charging boundary features; specifically, it can include the voltage Nernst coefficient, voltage change rate, differential value of voltage with respect to temperature, etc. of the battery. Among them, the boundary features can take statistical parameters such as the maximum value, sub-maximum value, standard deviation, and range of the parameters related to thermal runaway.

[0083] Dimensionality reduction is performed on a high-dimensional matrix through a machine learning algorithm to extract a low-dimensional feature matrix of the high-dimensional matrix. Based on the low-dimensional feature matrix, the current failure probability at each moment within this time period is calculated. By comparing the preset failure probability threshold with the current failure probability, it is determined whether there is a potential safety hazard currently; among them, the machine learning algorithm includes linear dimensionality reduction algorithms represented by PCA and non-linear deep learning dimensionality reduction algorithms represented by autoencoders. The low-dimensional features are set according to the required dimensionality reduction effect, and the lowest-dimensional features are selected when the dimensionality reduction effect meets the actual requirements. For example, in the PCA algorithm, the low-dimensional features corresponding to 90% (determined according to the actual dimensionality reduction requirements) of the information occupied by the eigenvalues of the covariance matrix are retained; the low-dimensional features corresponding to the reasonable loss function before and after dimensionality reduction in the autoencoder algorithm.

[0084] When it is determined that there is a potential safety hazard, the low-dimensional matrix is decompressed, the contribution value of each dimension in the high-dimensional matrix to the failure probability is calculated, and the electric vehicle corresponding to the boundary feature with the largest contribution to the failure probability is determined as the single entity prone to potential safety hazards.

[0085] Analyze the charging data of the single entity prone to potential safety hazards, combine it with the deep learning prediction model, calculate the deviation degree between the current state and the predicted state, compare the charging data of the single entity prone to potential safety hazards with the three-level alarm threshold, and give an alarm according to the corresponding level of the deviation degree. Among them, the deep learning prediction model uses a known prediction model.

[0086] In a specific embodiment, the failure probability is calculated by calculating the Hotelling T 2 statistic and the SPE statistic of the low-dimensional matrix, thereby calculating the comprehensive index φ and regressing it to the probability distribution to obtain the failure probability.

[0087] In a specific embodiment, the Hotelling T 2 statistic and the SPE statistic are calculated based on the dimensionality reduction process matrix of the charging data training set; among them, the Hotelling T 2 statistic is calculated according to the following formula:

[0088]

[0089] The SPE statistic is calculated according to the following formula:

[0090]

[0091] In the formula, X i is the boundary feature matrix of the i-th module at a certain moment, and P kis a dimensionality reduction transformation matrix, which is composed of principal eigenvectors. S is a diagonal matrix composed of the principal eigenvalues of the training sample set, I is the identity matrix, and k is the number of principal components. The principal eigenvalue is a well-known term in the art, specifically referring to selecting the corresponding number of eigenvalues of the matrix according to the dimensionality reduction dimension, and selecting the eigenvalues in descending order. The principal component refers to the number of features for dimensionality reduction.

[0092] In a specific embodiment, based on Hotelling T 2 and the solution of the SPE statistic, according to the comprehensive index obtain the system failure rate function of the sample at the current moment:

[0093]

[0094] In the formula, δ 2 are the control limits of Hotelling T 2 and SPE respectively. φ is a symmetric positive definite matrix. The comprehensive index conforms to the chi-square distribution with a degree of freedom of h and a coefficient of g. According to the probability distribution function, obtain the current failure probability function The failure probability h is calculated according to the following formula: The failure probability h is calculated as follows:

[0095]

[0096] In a specific embodiment, the degree of deviation is quantified using the relative deviation, and the formula is:

[0097]

[0098] In the formula, the current measured value is X real , and the deep learning predicted value is X pred .

[0099] In a specific embodiment, an alarm is given according to the corresponding level of the degree of deviation, specifically:

[0100] The preset three-level alarm thresholds are Thres1, Thres2, and Thres3 respectively, where Thres1 < Thres2 < Thres3. The specific alarm rules are as follows:

[0101] Trigger condition for the first-level alarm: ΔX rel ≥Thres1 and ΔX rel <Thres2

[0102] Alarm strategy:

[0103] Send a warning signal to the charging terminal or the edge side and turn on the yellow warning light;

[0104] Push real-time alarm information to the operation and maintenance personnel through the cloud platform, the content includes deviation parameters, current values, predicted values, and deviation ratios; automatically increase the data collection frequency of the terminal and continuously monitor the trend;

[0105] Secondary alarm trigger condition: ΔX rel ≥Thres2 and ΔX rel <Thres3

[0106] Alarm strategy:

[0107] Trigger orange audible and visual alarms, and synchronize the alarms between the local terminal and the cloud;

[0108] Automatically generate on-site inspection work orders and assign nearby operation and maintenance personnel to check the charging equipment connection, battery appearance, and charging terminal sensor cluster;

[0109] Tertiary alarm

[0110] Trigger condition: ΔX rel ≥Thres3

[0111] Alarm strategy:

[0112] Immediately send a power-off command to the edge device through the cloud to cut off the charging circuit; trigger a red emergency alarm and notify the vehicle owner, charging station administrator, and safety supervision department;

[0113] Automatically save the complete charging data for 10 minutes before and after the current moment as the original basis for fault analysis.

[0114] Suppose the predicted value of the battery temperature during the charging of an electric vehicle is 35°C and the measured value is 40°C, then:

[0115] Relative deviation:

[0116] If the thresholds are set as Thres1 = 5%, Thres2 = 10%, and Thres3 = 15%, then this deviation belongs to the secondary alarm, triggering an orange alarm and starting an on-site inspection.

[0117] Through the above quantitative formulas and classification strategies, it is possible to accurately identify and respond to charging safety hazards, balance the warning sensitivity and false alarm rate, and improve the safety and operation and maintenance efficiency of the charging network.

[0118] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the description of the method part.

[0119] The foregoing description of the disclosed embodiments enables those skilled in the art to practice or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Thus, the present invention is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An electric vehicle charging safety warning method based on cloud-edge-end collaboration, characterized in that, Including: Utilize the charging terminal sensor cluster to collect charging data in real-time and transmit the charging data to the edge side via wireless communication; The edge side receives the charging data transmitted by the charging terminal sensor cluster, processes the charging data through the intelligent charging data analysis algorithm deployed in the edge device to obtain abnormal data, and the edge side uploads the abnormal data to the cloud; The cloud receives the data uploaded by the edge side, stores the data in the cloud database, extracts the characteristics of the abnormal data, establishes a charging safety early warning model in the cloud, globally evaluates and warns the safety status of the entire charging network according to the extracted characteristics. When a potential safety hazard is detected, a global warning signal is sent to the relevant edge sides and terminals through the cloud to notify relevant personnel to take corresponding safety measures.

2. The method for warning of electric vehicle charging safety based on cloud-edge-terminal collaboration according to claim 1, wherein, It also includes interpolating the collected charging data, and the formula is as follows: Among them, L(x) is the calculated value of the missing feature value to be inserted in the charging data, x is the value of the time dimension where the missing feature value needs to be inserted in the charging data, and x j is the value of the time dimension of the j-th charging data, and y j is the value of the recorded feature value of the j-th charging data, and ω j is the barycentric weight, k is the number of recorded feature values in the charging data participating in the calculation of the missing feature value L(x) during the interpolation process, and 0 ≤ j ≤ k.

3. A method for early warning of electric vehicle charging safety based on cloud-edge-end collaboration according to claim 1, characterized in that, The specific process of the intelligent charging data analysis algorithm is as follows: In the native data processing network structure, a lightweight feature extraction model is used as the backbone network to replace the traditional fully connected layer or convolutional layer for extracting key features in the charging data. At the same time, a channel attention module is embedded in the backbone network to enhance the sensitivity to key features of the charging state through adaptive weight allocation; In the feature fusion layer of the native network, a hierarchical dynamic attention module is designed and deployed before multi-dimensional data fusion; this module adaptively allocates weights to the feature maps of different sensor data to strengthen the fusion accuracy of charging abnormal features and weaken the interference of irrelevant background data, so as to more accurately extract the abnormal feature regions in the charging state; In the analysis and decision-making layer of the native network, a variable convolution module is used to replace the standard convolution. By dynamically adjusting the sampling position of the convolution kernel, it adapts to the detection requirements of irregular abnormal patterns in the charging data; this module learns the local feature offset of the charging data, flexibly expands the receptive field range, enhances the ability to capture unstructured abnormal data, and improves the robustness and accuracy of charging state analysis.

4. A method for warning of electric vehicle charging safety based on cloud-edge-end collaboration according to claim 1, characterized in that The charging safety early warning model in the cloud is specifically: Extract the charging boundary features of the charging data collected by the cloud and form a high-dimensional matrix with the charging boundary features; Reduce the dimension of the high-dimensional matrix through a machine learning algorithm, extract the low-dimensional feature matrix of the high-dimensional matrix, calculate the current failure probability at each moment during this period based on the low-dimensional feature matrix, and compare it with the preset failure probability threshold to determine whether there is a potential safety hazard currently; When it is determined that there is a potential safety hazard, decompress the low-dimensional matrix, calculate the contribution value of each dimension in the high-dimensional matrix to the failure probability, and determine the electric vehicle corresponding to the boundary feature with the largest contribution to the failure probability as the single entity prone to safety hazards; Analyze the charging data of the single entity prone to safety hazards, combine it with the deep learning prediction model, calculate the deviation degree between the current state and the predicted state, compare the charging data of the single entity prone to safety hazards with the three-level alarm threshold, and give an alarm according to the corresponding level of the deviation degree.

5. A method for early warning of electric vehicle charging safety based on cloud-edge-end collaboration according to claim 4, characterized in that The failure probability is calculated by computing the Hotelling T 2 statistic and the SPE statistic of the low-dimensional matrix, thereby calculating a comprehensive index φ, and regressing it to a probability distribution to obtain the failure probability.

6. The method for warning of electric vehicle charging safety based on cloud-edge-terminal collaboration according to claim 5, characterized in that Hotelling T 2 The Hotelling T statistic and the SPE statistic are calculated based on the matrix in the dimensionality reduction process of the charging data training set; among them, the Hotelling T 2 statistic is calculated according to the following formula: The SPE statistic is calculated according to the following formula: where X i is the boundary feature matrix of module i at a certain moment, P k is the dimensionality reduction transformation matrix, which is composed of principal eigenvectors, S is the diagonal matrix composed of the principal eigenvalues of the training sample set, I is the identity matrix, and k is the number of principal components.

7. A method for early warning of electric vehicle charging safety based on cloud-edge-terminal collaboration according to claim 6, characterized in that, Based on the Hotelling T 2 and the solution of the SPE statistic, according to the comprehensive index obtain the system failure rate function of the sample at the current moment: wherein, δ 2 are respectively the control limits of Hotelling T 2 and SPE, φ is a symmetric positive definite matrix, and the comprehensive index conforms to a probability distribution of chi-square distribution with degree of freedom h and coefficient g. According to the probability distribution function, the current failure probability function is obtained. The failure probability h is calculated according to the following formula:

8. The method for warning of electric vehicle charging safety based on cloud-edge-terminal collaboration according to claim 4, wherein The deviation degree is quantified using the relative deviation, and the formula is: Wherein, the current measured value is X real , and the deep learning predicted value is X pred .

9. The method for warning of electric vehicle charging safety based on cloud-edge-end collaboration according to claim 8, characterized in that, Giving an alarm according to the corresponding level of the deviation degree is specifically: The preset three-level alarm thresholds are Thres1, Thres2, and Thres3 respectively, where Thres1 < Thres2 < Thres3. The specific alarm rules are as follows: Trigger condition for first-level alarm: ΔX rel ≥Thres1 and ΔX rel <Thres2 Alarm strategy: Send a warning signal to the charging terminal or the edge side and turn on the yellow warning light; Push real-time alarm information to the operation and maintenance personnel through the cloud platform, including deviation parameters, current value, predicted value, and deviation ratio; automatically increase the data collection frequency of this terminal and continuously monitor the trend; Secondary alarm trigger condition: ΔX rel ≥Thres2 and ΔX rel <Thres3 Alarm strategy: Trigger an orange audible and visual alarm, and synchronize the alarm between the local terminal and the cloud; Automatically generate on-site inspection work orders and assign nearby operation and maintenance personnel to check the charging equipment connection, battery appearance, and charging terminal sensor cluster; Three-level alarm Trigger condition: ΔX rel ≥ Thres3 Alarm strategy: Immediately send a power-off command to the edge side through the cloud to cut off the charging circuit; trigger a red emergency alarm to notify the vehicle owner, charging station administrator, and safety supervision department; Automatically save the complete charging data for 10 minutes before and after the current moment as the original basis for fault analysis.

Images