Distribution network fault rapid positioning method based on artificial intelligence

Abstract

The invention relates to the field of distribution network fault detection, in particular to a distribution network fault rapid positioning method based on artificial intelligence. According to the method, a double-layer star topology is established, a dispatching center is effectively connected with a transformer substation and a power distribution area, first-level and second-level power distribution network data are acquired by using a multi-modal sensor, and fault prediction is performed through a federated learning framework; analyzing topological relevance in combination with a graph convolutional network, generating a candidate fault transformer area set, and determining a fault type through cosine similarity verification and current waveform correlation analysis; a fault distribution network point is accurately positioned by using a traveling wave distance measurement technology. When an unknown fault is detected, the system triggers an incremental learning mechanism to update global model parameters, and the fault diagnosis capability is continuously optimized. Based on the method, rapid and accurate distribution network fault detection and positioning are finally realized.

Description

Technical Field

[0001] The present invention relates to the field of distribution network fault detection, and in particular to a distribution network fault rapid location method based on artificial intelligence. Background Art

[0002] The distribution network is an important part of the power system, responsible for distributing the electric energy transmitted by the substation to each user. Due to the complex structure of the distribution network and the wide distribution of equipment, a fault may cause a large-scale power outage, affecting the operation of the social economy. Therefore, rapid fault detection and accurate positioning are crucial to improving power supply reliability and reducing power outage time. The rapid diagnosis of distribution network faults through intelligent technology can not only improve the automation level of the power system, but also reduce operation and maintenance costs and improve the safety and stability of the power grid.

[0003] However, at this stage, distribution network fault detection still has many drawbacks at the data level. On the one hand, traditional monitoring methods mainly rely on SCADA (Supervisory Control and Data Acquisition) and fault recorders, with low data sampling rates, making it difficult to accurately capture fault characteristics. On the other hand, due to the wide distribution of distribution networks and limited sensor deployment density, fault data is incomplete, affecting the accuracy of fault diagnosis. In addition, the distribution network environment is complex, and external interference factors (such as load fluctuations, weather changes, etc.) easily introduce noise, making it difficult to further improve the fault location accuracy based on traditional methods. Therefore, how to use artificial intelligence and multi-source data fusion technology to overcome data quality problems and achieve efficient and accurate fault location is the focus of current research.

[0004] Therefore, a method for rapid location of distribution network faults based on artificial intelligence is proposed. Summary of the invention

[0005] The present invention provides a method for quickly locating distribution network faults based on artificial intelligence. The method constructs a two-level star topology with a dispatching center and a substation as the core, collects distribution network data through multimodal sensors, and performs fault prediction based on federated learning. Time alignment and graph convolutional networks are used to analyze topological associations, screen candidate fault areas, and accurately identify fault types through cosine similarity, current waveform correlation, and dynamic time regularization. Finally, the fault distribution point is located in combination with traveling wave ranging, and incremental learning is triggered when an unknown fault is detected, dynamically optimizing the global model to improve the accuracy and adaptability of fault detection.

[0006] To achieve the above object, the present invention provides the following technical solutions: A method for quickly locating distribution network faults based on artificial intelligence, comprising: A primary star topology is established with the dispatch center as the primary central node and the substation as the primary edge node. A secondary star topology is established with the substation as the secondary central node and the distribution area as the secondary edge node. The primary distribution network data is collected by multimodal sensors deployed in the substation, a local server of the substation is used as an edge node, and a dispatch center is used as a federation aggregation center to establish a federated learning framework, and the primary distribution network data is predicted using the federated learning framework to obtain the faulty primary edge node; The secondary distribution network data is collected by multimodal sensors deployed in the distribution substation area, and the primary distribution network data and the secondary distribution network data are aligned in time series to construct a cross-level feature matrix; based on the cross-level feature matrix, the topological correlation is analyzed through a pre-trained graph convolutional network to generate a set of candidate fault substation areas; The candidate fault area set is verified by cosine similarity and misjudged data is eliminated, and the distribution network fault type is further determined by combining current waveform correlation analysis and dynamic time warping; The coordinates of the faulty distribution network point are obtained through traveling wave ranging to locate the distribution network fault.

[0007] Furthermore, the federated learning framework includes edge nodes and a federated aggregation center, the edge nodes use the LSTM sub-model to train the primary distribution network data, and the federated aggregation center aggregates the parameters of the LSTM sub-model and generates a global model.

[0008] Furthermore, the training of the federated learning framework includes: The federation aggregation center initializes the global model and sends the global model to the edge nodes; The edge node uses the LSTM sub-model to train the primary distribution network data, and uploads the trained model parameters to the federation aggregation center; The federation aggregation center performs FedAvg aggregation on the model parameters, updates the global model and distributes it to the edge nodes.

[0009] Furthermore, constructing a cross-level feature matrix includes: Obtain primary distribution network data and secondary distribution network data; Unify the timestamp format of primary distribution network data and secondary distribution network data, find out the common timestamp of primary distribution network data and secondary distribution network data, and use linear interpolation to complete the sparse data to align the two data; The aligned data are merged to obtain the cross-level feature matrix.

[0010] Furthermore, the cosine similarity verification process includes: Obtain the secondary distribution network data when the distribution network is operating normally, and calculate its mean data; Calculate the cosine similarity between the secondary distribution network data of each distribution area in the candidate fault area set and the mean data; If the cosine similarity is greater than the similarity threshold, it is determined to be misjudged data and removed from the candidate fault station area set; otherwise, it is determined to be non-misjudged data and retained.

[0011] Furthermore, determining the distribution network fault type includes: Obtain current waveform data before and after the fault and perform preprocessing; Dynamic time warping method is used to compare the waveforms before and after the fault, and the optimal alignment path and DTW distance are calculated; Calculate the Pearson correlation coefficient of the current waveform to determine the waveform change trend; The type of distribution network fault is determined by combining the DTW distance and the Pearson correlation coefficient.

[0012] Furthermore, obtaining the coordinates of the fault distribution point includes: The arrival time difference of the traveling wave between the faulty distribution point and the faulty substation is recorded, and the coordinates of the faulty distribution point are calculated based on the wave speed.

[0013] Furthermore, the method further comprises: When an unknown fault is detected, the incremental learning mechanism is triggered to update the global model parameters.

[0014] Furthermore, triggering the incremental learning mechanism to update the global model parameters includes: Calculate the similarity between the current fault feature and the known fault library. If it is lower than the set threshold, it is determined to be an unknown fault. Cache unknown fault data, perform preliminary training on edge nodes to learn new patterns, and upload trained model parameters to the dispatch center through each edge node to perform global model updates; FedAvg is used for parameter aggregation, and the compatibility of new and old failure modes is ensured through knowledge distillation and playback mechanisms; After confirming the new fault mode, add it to the fault database and deploy the updated model to the edge nodes.

[0015] The beneficial effects of the present invention are: 1. By constructing a two-layer star topology, data flow is efficiently organized under the hierarchical architecture of the dispatching center-substation-distribution station area, making fault detection more accurate and efficient. The substation multimodal sensors are used to collect the primary distribution network data, and the data is pre-processed on the local server based on edge computing to reduce data transmission delays and improve real-time response capabilities. Combined with the federated learning framework, the direct transmission of raw data is avoided, which not only ensures data privacy and security, but also makes full use of the computing power of each edge node to improve the generalization ability of the fault prediction model. Ultimately, this method can quickly locate the faulty primary edge node, provide a basis for subsequent accurate fault location, and improve the intelligent operation and maintenance level of the distribution network.

[0016] 2. By deploying multimodal sensors in distribution substations, collecting secondary distribution network data in a refined manner, and combining the primary distribution network data for time series alignment, a unified cross-level feature matrix is ​​constructed to enhance data integrity and temporal consistency. Using pre-trained graph convolutional networks to mine distribution network topological correlations can effectively model complex power network structures and improve the ability to understand fault propagation paths. Ultimately, this method can accurately screen candidate fault substation sets, reduce misjudgments and missed judgments, lay a solid foundation for subsequent precise fault location and type analysis, and improve the level of intelligent distribution network fault diagnosis.

[0017] 3. Through cosine similarity check, the misjudgment data is screened out to improve the accuracy of fault identification. At the same time, the current waveform correlation analysis and dynamic time regularization are combined to accurately match the fault mode, so as to accurately determine the type of distribution network fault. Furthermore, the traveling wave ranging technology is used to calculate the coordinates of the fault point in combination with the propagation time difference of the fault signal to achieve high-precision fault location. This method integrates a variety of intelligent analysis methods, which can effectively reduce misjudgments and missed judgments, and improve the accuracy of fault identification and location, providing strong support for the intelligent operation and maintenance and rapid repair of the distribution network.

[0018] 4. To achieve timely updating of fault types, this method also detects unknown faults through similarity judgment, triggers incremental learning, optimizes the global model through federated learning after preliminary training on edge nodes, and uses knowledge distillation and playback mechanisms to ensure compatibility between new and old faults. Finally, the model is dynamically updated to improve the accuracy and adaptability of distribution network fault identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings: Figure 1 This is a flow chart of a method for quickly locating distribution network faults based on artificial intelligence provided by the present invention; Figure 2 This is another flow chart of a method for quickly locating distribution network faults based on artificial intelligence provided by the present invention. DETAILED DESCRIPTION

[0020] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0021] Embodiment 1 A method for quickly locating distribution network faults based on artificial intelligence, such as Figure 1 As shown, including: S110: Establish a primary star topology with the dispatching center as the primary central node and the substation as the primary edge node, and establish a secondary star topology with the substation as the secondary central node and the distribution area as the secondary edge node; S120: Collecting primary distribution network data through multimodal sensors deployed in the substation, establishing a federated learning framework with a local server of the substation as an edge node and a dispatching center as a federated aggregation center, and using the federated learning framework to predict the primary distribution network data to obtain a faulty primary edge node; Furthermore, the federated learning framework includes edge nodes and a federated aggregation center, the edge nodes use the LSTM sub-model to train the primary distribution network data, and the federated aggregation center aggregates the parameters of the LSTM sub-model and generates a global model.

[0022] The federation aggregation center is based on distributed LSTM sub-model training, which can integrate the information of each edge node and improve the generalization ability of the global model, making it more suitable for fault prediction in different scenarios, thereby improving the intelligent diagnosis capability and overall operation stability of the distribution network system.

[0023] Furthermore, the training of the federated learning framework includes: The federation aggregation center initializes the global model and sends the global model to the edge nodes; The edge node uses the LSTM sub-model to train the primary distribution network data, and uploads the trained model parameters to the federation aggregation center; The federation aggregation center performs FedAvg aggregation on the model parameters, updates the global model and distributes it to the edge nodes.

[0024] Specifically, the federation aggregation center initializes a global LSTM model parameter , the initial global model parameters Send it to each edge node; each edge node uses the local primary distribution network data to train the LSTM sub-model. After the training is completed, each edge node will update the model parameters It is sent to the federation aggregation center, which performs FedAvg aggregation. The aggregation formula can be expressed as: ; in, represents the global model, Indicates The local model parameters of the edge nodes, Indicates the number of data samples used by the node. Represents the total amount of data of all edge nodes participating in the training; the federation aggregation center updates the global model The data is distributed back to each edge node, and each edge node continues to use local data for the next round of training based on the new global model. The above process is iterated until the global model converges or the preset number of training rounds is reached.

[0025] The federated learning framework improves the intelligence and efficiency of distribution network fault detection through distributed training. The federated aggregation center initializes the global model and sends it to the edge nodes, so that each substation can use local data to independently train the LSTM sub-model, thereby reducing data transmission, protecting privacy, and reducing the communication burden. The trained model parameters are aggregated through the FedAvg method, which effectively integrates the feature information of multiple substations and improves the generalization ability and adaptability of the model. Finally, the optimized global model is distributed to the edge nodes again to achieve continuous optimization and improve the accuracy and response speed of distribution network fault location.

[0026] S130: Collecting secondary distribution network data through multimodal sensors deployed in distribution substations, aligning the primary distribution network data with the secondary distribution network data in time series, and constructing a cross-level feature matrix; based on the cross-level feature matrix, analyzing topological associations through a pre-trained graph convolutional network, and generating a set of candidate fault substations; Furthermore, constructing a cross-level feature matrix includes: Obtain primary distribution network data and secondary distribution network data; Unify the timestamp format of primary distribution network data and secondary distribution network data, find out the common timestamp of primary distribution network data and secondary distribution network data, and use linear interpolation to complete the sparse data to align the two data; The aligned data are merged to obtain the cross-level feature matrix.

[0027] Specifically, the multimodal sensors deployed in the substation collect primary distribution network data, including at least current, voltage, and traveling wave signals; the multimodal sensors deployed in the distribution station area collect secondary distribution network data, including at least load information, current waveform, and short-circuit faults. Both types of data are accompanied by timestamps; all timestamps are converted into a unified format (such as Unix timestamp or ISO8601 format) to extract primary distribution network data. and secondary distribution network data All timestamp sets of , calculate the common timestamp set , and use linear interpolation to complete the sparse data to align the two data. In a feasible embodiment, and is the public timestamp, and is a public timestamp, then For sparse data, and The data between the two are linearly interpolated, and the interpolation formula is: ; in, and Contiguous timestamps indicating the existence of data, and represents the corresponding measured value, Represents the interpolated measurement value; the aligned data are merged to obtain the cross-level feature matrix , where each column corresponds to all the distribution network data at the same timestamp.

[0028] Through time alignment and data completion, we ensure that the primary distribution network data and the secondary distribution network data are analyzed at the same time scale to improve data consistency and integrity. We use a unified timestamp format and linear interpolation method to complete sparse data to avoid information bias caused by missing data, thereby enhancing the accuracy of cross-level data fusion.

[0029] S140: verifying the candidate fault area set by cosine similarity and eliminating misjudged data, and further determining the distribution network fault type by combining current waveform correlation analysis and dynamic time warping; Furthermore, the cosine similarity verification process includes: Obtain the secondary distribution network data when the distribution network is operating normally, and calculate its mean data; Calculate the cosine similarity between the secondary distribution network data of each distribution area in the candidate fault area set and the mean data; If the cosine similarity is greater than the similarity threshold, it is determined to be misjudged data and removed from the candidate fault station area set; otherwise, it is determined to be non-misjudged data and retained.

[0030] Specifically, collect the secondary distribution network data of the distribution station area under normal operation and build a historical normal data set ,set up For the The secondary distribution network characteristic data sampled by times, then the mean data For each distribution area in the candidate fault area set, obtain its current secondary distribution network data , the process of calculating cosine similarity is: ; in, represents the cosine similarity, represents the dot product of two vectors, Represents the Euclidean norm of two vectors; sets the similarity threshold If the cosine similarity is greater than the similarity threshold, it is determined as misjudged data and removed from the candidate fault area set; otherwise, it is determined as non-misjudged data and retained.

[0031] The cosine similarity is used to verify the candidate fault area data, and by comparing it with the historical mean data under normal operating conditions, the misjudged data can be effectively identified and eliminated. By setting the similarity threshold, the normal area and the area where the fault actually occurred can be accurately distinguished, the false alarm rate can be reduced, and the accuracy and reliability of fault location can be improved.

[0032] Furthermore, determining the distribution network fault type includes: Obtain current waveform data before and after the fault and perform preprocessing; Dynamic time warping method is used to compare the waveforms before and after the fault, and the optimal alignment path and DTW distance are calculated; Calculate the Pearson correlation coefficient of the current waveform to determine the waveform change trend; The type of distribution network fault is determined by combining the DTW distance and the Pearson correlation coefficient.

[0033] Specifically, sensors deployed in the distribution area are used to obtain After the failure The current waveform data is obtained and data preprocessing is performed, including normalization, noise reduction and time window selection. The optimal alignment path between the pre-fault waveform and the post-fault waveform is calculated, and nonlinear time deformation is allowed to find the best match between the two waveforms and calculate the DTW distance: ; in, The current waveform data before the fault occurs is shown in The value at a time point, Indicates the current waveform data after the fault occurs. The Pearson correlation coefficient of the current waveform is calculated as: ; in, represents the Pearson correlation coefficient, represents the mean value of the waveform before the fault, Represents the mean of the waveform after the fault; combines the DTW distance and Pearson correlation coefficient, as well as the distance threshold and coefficient threshold , determine the distribution network fault type. In a feasible embodiment, the distribution network fault type determination rule is shown in Table 1.

[0034] Table 1. Rules for determining distribution network fault types

[0035] By obtaining the current waveform data before and after the fault, and using dynamic time warping to calculate the optimal alignment path and DTW distance, the problem of inconsistent waveform time scales before and after the fault is overcome; at the same time, the Pearson correlation coefficient is introduced to analyze the waveform change trend and improve the ability to identify fault characteristics. Finally, combined with DTW distance and correlation analysis, the type of distribution network fault can be accurately determined, improving the accuracy and robustness of fault diagnosis.

[0036] Furthermore, obtaining the coordinates of the fault distribution point includes: The arrival time difference of the traveling wave between the faulty distribution point and the faulty substation is recorded, and the coordinates of the faulty distribution point are calculated based on the wave speed.

[0037] By using traveling wave ranging technology, the coordinates of the faulty distribution network point are calculated by recording the arrival time difference of the traveling waves and combining the wave speed to accurately and quickly locate the fault, thereby improving the timeliness and accuracy of distribution network fault diagnosis.

[0038] S150: Acquire the coordinates of the faulty distribution network point through traveling wave ranging to complete the fault location of the distribution network.

[0039] Embodiment 2 As another optional embodiment of the present application, refer to Figure 2 , is a flow chart of a second embodiment of a method for rapid location of distribution network faults based on artificial intelligence provided by the present application. This embodiment is mainly an extended solution of the method for rapid location of faults described in the above embodiment 1, such as Figure 2 As shown, the method includes: S210: Establish a primary star topology with the dispatching center as the primary central node and the substation as the primary edge node, and establish a secondary star topology with the substation as the secondary central node and the distribution area as the secondary edge node; S220: Collecting primary distribution network data through multimodal sensors deployed in the substation, establishing a federated learning framework with a local server of the substation as an edge node and a dispatching center as a federated aggregation center, and using the federated learning framework to predict the primary distribution network data to obtain a faulty primary edge node; S230: Collecting secondary distribution network data through multimodal sensors deployed in distribution substations, aligning the primary distribution network data with the secondary distribution network data in time series, and constructing a cross-level feature matrix; based on the cross-level feature matrix, analyzing topological associations through a pre-trained graph convolutional network, and generating a set of candidate fault substations; S240: Verify the candidate fault area set by cosine similarity and remove misjudged data, and further determine the distribution network fault type by combining current waveform correlation analysis and dynamic time warping; S250: Acquire the coordinates of the fault distribution network point through traveling wave ranging to complete the fault location of the distribution network; S260: When an unknown fault is detected, an incremental learning mechanism is triggered to update global model parameters.

[0040] The detailed process of steps S210-S250 can refer to the relevant introduction of steps S110-S150 in Example 1, which will not be repeated here.

[0041] Further, see Figure 2 In S260, the incremental learning mechanism is triggered to update the global model parameters including: Calculate the similarity between the current fault feature and the known fault library. If it is lower than the set threshold, it is determined to be an unknown fault. Cache unknown fault data, perform preliminary training on edge nodes to learn new patterns, and upload trained model parameters to the dispatch center through each edge node to perform global model updates; FedAvg is used for parameter aggregation, and the compatibility of new and old failure modes is ensured through knowledge distillation and playback mechanisms; After confirming the new fault mode, add it to the fault database and deploy the updated model to the edge nodes.

[0042] By calculating the similarity between the current fault characteristics and the known fault library, unknown faults are automatically identified and incremental learning is triggered. By performing preliminary training on the edge nodes and uploading model parameters to the dispatch center, FedAvg is used to update the global model to ensure the compatibility of new and old fault modes. The knowledge distillation and playback mechanism is introduced to optimize the stability and accuracy of the model. Finally, the new fault mode is added to the database and updated to the edge nodes to improve the intelligence and adaptability of distribution network fault diagnosis.

[0043] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for quickly locating distribution network faults based on artificial intelligence, characterized in that: include: A primary star topology is established with the dispatch center as the primary central node and the substation as the primary edge node. A secondary star topology is established with the substation as the secondary central node and the distribution area as the secondary edge node. The primary distribution network data is collected by multimodal sensors deployed in the substation, a local server of the substation is used as an edge node, and a dispatch center is used as a federation aggregation center to establish a federated learning framework, and the primary distribution network data is predicted using the federated learning framework to obtain the faulty primary edge node; The secondary distribution network data is collected by multimodal sensors deployed in the distribution station area, the primary distribution network data and the secondary distribution network data are time-series aligned, and a cross-level feature matrix is ​​constructed; Based on the cross-level feature matrix, the topological association is analyzed through a pre-trained graph convolutional network to generate a set of candidate fault areas; The candidate fault area set is verified by cosine similarity and misjudged data is eliminated, and the distribution network fault type is further determined by combining current waveform correlation analysis and dynamic time warping; The coordinates of the faulty distribution network point are obtained through traveling wave ranging to locate the distribution network fault.

2. According to the method of claim 1, the method is characterized in that: The federated learning framework includes edge nodes and a federated aggregation center. The edge nodes use the LSTM sub-model to train the primary distribution network data, and the federated aggregation center aggregates the parameters of the LSTM sub-model and generates a global model.

3. The method for quickly locating distribution network faults based on artificial intelligence according to claim 1 is characterized in that: The training of the federated learning framework includes: The federation aggregation center initializes the global model and sends the global model to the edge nodes; The edge node uses the LSTM sub-model to train the primary distribution network data, and uploads the trained model parameters to the federation aggregation center; The federation aggregation center performs FedAvg aggregation on the model parameters, updates the global model and distributes it to the edge nodes.

4. The method for quickly locating distribution network faults based on artificial intelligence according to claim 1 is characterized in that: Constructing a cross-level feature matrix includes: Obtain primary distribution network data and secondary distribution network data; Unify the timestamp format of primary distribution network data and secondary distribution network data, find out the common timestamp of primary distribution network data and secondary distribution network data, and use linear interpolation to complete the sparse data to align the two data; The aligned data are merged to obtain the cross-level feature matrix.

5. The method for quickly locating distribution network faults based on artificial intelligence according to claim 1 is characterized in that: The cosine similarity verification process includes: Obtain the secondary distribution network data when the distribution network is operating normally, and calculate its mean data; Calculate the cosine similarity between the secondary distribution network data of each distribution area in the candidate fault area set and the mean data; If the cosine similarity is greater than the similarity threshold, it is determined to be misjudged data and removed from the candidate fault station area set; otherwise, it is determined to be non-misjudged data and retained.

6. The method for quickly locating distribution network faults based on artificial intelligence according to claim 1, characterized in that: Determine the distribution network fault type including: Obtain current waveform data before and after the fault and perform preprocessing; Dynamic time warping method is used to compare the waveforms before and after the fault, and the optimal alignment path and DTW distance are calculated; Calculate the Pearson correlation coefficient of the current waveform to determine the waveform change trend; The distribution network fault type is determined by combining the DTW distance and the Pearson correlation coefficient.

7. The method for quickly locating distribution network faults based on artificial intelligence according to claim 1 is characterized in that: Obtaining the coordinates of the fault distribution point includes: The arrival time difference of the traveling wave between the faulty distribution point and the faulty substation is recorded, and the coordinates of the faulty distribution point are calculated based on the wave speed.

8. The method for quickly locating distribution network faults based on artificial intelligence according to claim 1 is characterized in that: The method further comprises: When an unknown fault is detected, the incremental learning mechanism is triggered to update the global model parameters.

9. The method for quickly locating distribution network faults based on artificial intelligence according to claim 1, characterized in that: Triggering the incremental learning mechanism to update global model parameters includes: Calculate the similarity between the current fault feature and the known fault library. If it is lower than the set threshold, it is determined to be an unknown fault. Cache unknown fault data, perform preliminary training on edge nodes to learn new patterns, and upload trained model parameters to the dispatch center through each edge node to perform global model updates; FedAvg is used for parameter aggregation, and the compatibility of new and old failure modes is ensured through knowledge distillation and playback mechanisms; After confirming the new fault mode, add it to the fault database and deploy the updated model to the edge nodes.

Images