A fault diagnosis and early warning system and method for electrical equipment based on federated learning and hybrid intelligence

Abstract

This invention relates to the field of power system fault diagnosis and operation and maintenance technology, specifically to an electrical equipment fault diagnosis and early warning system and method based on federated learning and hybrid intelligence. Through a complete closed-loop architecture of "data acquisition—edge computing—federated learning—hybrid intelligent diagnosis—three-level early warning interaction," it achieves full-process coverage from bottom-level multi-source heterogeneous data acquisition to high-level intelligent decision-making and early warning. Specifically, a CNN-LSTM hybrid model is used to simultaneously extract the spatial and temporal features of faults, improving diagnostic accuracy by approximately 15-25% compared to a single model. The federated learning module allows local terminals to transmit encrypted model parameters without sharing raw data, protecting data privacy while enabling cross-regional knowledge sharing. Dynamic weighted aggregation and local model verification mechanisms solve the problem of model performance degradation caused by differences in data distribution across different regions and equipment, and prevent abnormal models from polluting the global model.

Description

Technical Field

[0001] This invention relates to the field of power system fault diagnosis and operation and maintenance technology, specifically to an electrical equipment fault diagnosis and early warning system and method based on federated learning and hybrid intelligence, which is particularly suitable for online monitoring, intelligent diagnosis, hierarchical early warning and operation and maintenance scheduling of electrical line faults and main switch mechanical faults. Background Technology

[0002] Electrical lines (including overhead lines and cable lines) and main switching equipment (such as circuit breakers, disconnectors, and load switches) are the core components of a power system. Electrical lines are exposed to the outdoor environment for extended periods, making them susceptible to short circuits, open circuits, and other faults due to lightning strikes, wind deflection, tree obstructions, and pollution. Main switching equipment, on the other hand, can experience mechanical failures such as jammed operating mechanisms, aging coils, and worn contacts, leading to switching failures and ultimately power outages.

[0003] Traditional fault diagnosis methods rely primarily on manual experience or single-parameter threshold judgments, resulting in the following drawbacks: fault data from different substations and regions cannot be shared, leading to limited model training data and poor generalization ability; centralized uploading of raw data poses a privacy risk. Existing methods mostly rely solely on electrical quantities (voltage, current) or mechanical quantities (vibration, travel) for diagnosis, failing to fully integrate multi-source heterogeneous data, resulting in low recognition rates for complex faults. Static rule bases cannot adapt to equipment aging, changes in operating conditions, and new fault types, leading to a gradual decline in diagnostic performance. They only provide single alarm signals, lacking classification, location information, and handling suggestions, resulting in slow response times and low efficiency for maintenance personnel.

[0004] In recent years, the development of deep learning and federated learning technologies has provided new ideas for solving the above problems. For example, Chinese invention patent CN119398106B proposes a power distribution network fault handling device; CN117290756B proposes a transmission line fault identification method, device, and electronic equipment based on federated learning; CN121304140B proposes a transmission line fault diagnosis and operation and maintenance scheduling method and system based on networked learning. However, existing technologies still have not solved the problems of deeply integrating multi-source mechanical and electrical features, local federated learning and global model verification, digital twin reinforcement learning optimization, and systematically integrating three-level hierarchical early warning with natural language interaction. Summary of the Invention

[0005] In view of this, the purpose of the present invention is to provide an electrical equipment fault diagnosis and early warning system and method based on federated learning and hybrid intelligence, so as to solve the technical problems involved in the above background art.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: An electrical equipment fault diagnosis and early warning system based on federated learning and hybrid intelligence includes a data acquisition module, an edge computing module, a federated learning module, a hybrid intelligence diagnosis module, and a three-level early warning and interaction module. The data acquisition module is configured to synchronously acquire the three-phase voltage and three-phase current of the electrical circuit, as well as the vibration signal, travel signal, and coil current signal of the main switchgear, forming a multi-source heterogeneous monitoring dataset; The edge computing module is deployed on the local terminal and includes a feature extraction unit, a feature library and diagnostic rule library construction unit, and an local training unit. The feature extraction unit adopts a hybrid model of convolutional neural network and long short-term memory network to extract spatiotemporal joint features from the multi-source heterogeneous monitoring dataset. The convolutional neural network is used to extract the spatial distribution features of the fault waveform image, and the long short-term memory network is used to extract the temporal dependence features of the fault evolution. The feature library and diagnostic rule library construction unit is used to map the spatiotemporal joint features to preset fault classification labels, and construct a classification feature library and corresponding diagnostic rule library including electrical circuit faults and main switch mechanical faults. The in-situ training unit is used to train a preset lightweight diagnostic model locally based on the classification feature library and the diagnostic rule library, and generate local model parameters. The federated learning module includes a parameter transmission interface, a dynamic weighted aggregation unit, and a local model verification unit. The parameter transmission interface is configured to encrypt the local model parameters and upload them to the central server, and receive the aggregated parameters sent by the central server. The dynamic weighted aggregation unit, deployed on the central server, is used to calculate the dynamic weight of each local terminal based on the number of samples of each local terminal and the loss function of the local model, and to perform weighted aggregation of the parameters of each local model based on the dynamic weight to generate global model parameters. The local model verification unit is deployed on the local terminal. After receiving the aggregated parameters from the central server, it first performs a performance evaluation on the local verification set. If the evaluation result is better than the current local model, the model parameters are updated; otherwise, the update is rejected and the existing parameters are maintained. The hybrid intelligent diagnostic module includes a Transformer deep inference unit, a digital twin simulation unit, and a reinforcement learning optimization unit; The Transformer deep inference unit is configured to receive the global model parameters or the verified local model parameters, perform contextual understanding on long-sequence power grid state information, and generate preliminary fault diagnosis results. The digital twin simulation unit is used to create a digital twin model of the equipment based on the preliminary fault diagnosis results, and to simulate the fault evolution process and analyze hypothetical scenarios. The reinforcement learning optimization unit is used to perform interactive training in the simulation environment provided by the digital twin simulation unit, dynamically optimize the fault response strategy and repair plan, and generate diagnostic conclusions including fault causes, fault locations, handling suggestions and preventive measures. The three-level early warning and interaction module is configured to divide the early warning into three levels based on the confidence and urgency of the diagnostic conclusion: prompt level, warning level, and emergency level; and output early warning information corresponding to the early warning level through three methods: AI intelligent voice interaction, remote platform early warning push, and mobile APP message prompt. The early warning information includes the cause of the fault, the location of the fault, handling suggestions, and preventive measures.

[0007] Furthermore, in the dynamic weighted aggregation unit, the first... Dynamic weight of each local terminal Calculate using the following formula: ; in, For the first The number of local fault samples possessed by each local terminal. For the first The loss function value of the local diagnostic model for each local terminal on the validation set. The total number of local terminals participating in federated learning.

[0008] Furthermore, the local model verification unit is specifically configured as follows: it loads the aggregated parameters issued by the central server into the local model, calculates the accuracy of the updated model using the locally reserved verification set, and accepts the update if the accuracy is higher than the accuracy of the model before the update by more than a preset threshold; if the accuracy is lower than the accuracy of the model before the update by less than the preset threshold, it rejects the update and sends a rejection event to the central server.

[0009] Furthermore, in the three-level early warning and interaction module, the confidence level of the diagnostic conclusion corresponding to the prompt level is not less than the first threshold and less than the second threshold, which is used to indicate that the device has potential abnormalities but has not yet formed a clear fault. The confidence level of the diagnostic conclusion corresponding to the warning level is not less than the second threshold and less than the third threshold, and is used to indicate that the equipment has a clear fault but has not yet endangered continuous operation; The confidence level of the diagnostic conclusion corresponding to the emergency level is not less than the third threshold, which is used to indicate that the equipment has suffered a serious malfunction and needs to be shut down for repair immediately. Wherein, the first threshold is less than the second threshold, and the second threshold is less than the third threshold.

[0010] Furthermore, the AI ​​intelligent voice interaction supports natural language question and answer. When maintenance personnel issue a voice query command, the system will broadcast the cause of the fault, the location of the fault, the handling suggestions and preventive measures in the diagnostic conclusion in voice form, and adopt different voice prompt frequencies and semantic urgency according to the warning level.

[0011] Furthermore, in the feature library and diagnostic rule library construction unit, the classification features of the electrical circuit faults include three-phase voltage imbalance, harmonic distortion rate, zero-sequence current amplitude, and transient traveling wave polarity; the classification features of the main switch mechanical faults include the time-domain peak value of the vibration signal, frequency-domain energy distribution, inflection point position of the stroke curve, and the start time and duration of the coil current.

[0012] Furthermore, the reinforcement learning optimization unit adopts a near-end policy optimization algorithm. Its state space includes the current health score of the device, load rate, and historical fault frequency. Its action space includes switch opening and closing operations, load transfer instructions, and maintenance priority ranking. Its reward function comprehensively considers fault recovery time, power supply reliability impact, and remaining lifespan loss of the device.

[0013] A fault diagnosis and early warning method for electrical equipment based on federated learning and hybrid intelligence, applied to the aforementioned fault diagnosis and early warning system for electrical equipment based on federated learning and hybrid intelligence, includes the following steps: S1. Synchronously collect the three-phase voltage and three-phase current of the electrical circuit, as well as the vibration signal, travel signal, and coil current signal of the main switchgear, to form a multi-source heterogeneous monitoring dataset; S2. At the local terminal, a hybrid model of convolutional neural network-long short-term memory network is used to extract spatiotemporal joint features from the multi-source heterogeneous monitoring dataset. The convolutional neural network is used to extract the spatial distribution features of the fault waveform image, and the long short-term memory network is used to extract the temporal dependence features of the fault evolution. The spatiotemporal joint features are mapped to preset fault classification labels to construct a classification feature library and a corresponding diagnostic rule library including electrical line faults and main switch mechanical faults. S3. Train the preset lightweight diagnostic model locally based on the classification feature library and diagnostic rule library to generate local model parameters. Upload the local model parameters to the central server after encryption and receive the aggregated parameters issued by the central server. S4. The central server calculates the dynamic weight of each local terminal based on the number of samples of each local terminal and the loss function of the local model, and performs weighted aggregation of the parameters of each local model based on the dynamic weight to generate global model parameters and distribute them to each local terminal. S5. After receiving the global model parameters, the local terminal first performs a performance evaluation on the local validation set. If the evaluation result is better than the current local model, the model parameters are updated; otherwise, the update is rejected and the existing parameters are maintained. S6. Use the Transformer model to perform contextual understanding on long-sequence power grid state information and generate preliminary fault diagnosis results; S7. Based on the preliminary fault diagnosis results, create a digital twin model of the equipment, simulate the fault evolution process and analyze hypothetical scenarios, and dynamically optimize the fault response strategy through reinforcement learning to generate diagnostic conclusions including fault causes, fault locations, handling suggestions and preventive measures. S8. Based on the confidence level and urgency of the diagnostic conclusion, three warning levels are defined: alert level, warning level, and emergency level. Warning information corresponding to the warning level is output through three methods: AI intelligent voice interaction, remote platform warning push, and mobile APP message prompt. The warning information includes the cause of the fault, the location of the fault, handling suggestions, and preventive measures.

[0014] Furthermore, in step S4, the formula for calculating the dynamic weight is: ; in, For the first The number of local fault samples possessed by each local terminal. For the first The loss function value of the local diagnostic model for each local terminal on the validation set. The total number of local terminals participating in federated learning.

[0015] Furthermore, in step S8, the AI ​​intelligent voice interaction supports natural language question and answer. When the maintenance personnel issue a voice query command, the system broadcasts the cause of the fault, the location of the fault, the handling suggestions and preventive measures in the diagnostic conclusion in voice form, and adopts different voice prompt frequencies and semantic urgency according to the warning level.

[0016] The beneficial effects of this invention are: This invention provides an electrical equipment fault diagnosis and early warning system and method based on federated learning and hybrid intelligence. Through a complete closed-loop architecture of "data acquisition—edge computing—federated learning—hybrid intelligent diagnosis—three-level early warning interaction," it achieves full-process coverage from bottom-level multi-source heterogeneous data acquisition to high-level intelligent decision-making and early warning. Specific effects include: using a CNN-LSTM hybrid model to simultaneously extract spatial features (such as waveform images) and temporal features (such as evolution trends) of faults, improving diagnostic accuracy by approximately 15-25% compared to a single model. The federated learning module eliminates the need for local terminals to share raw data, transmitting only encrypted model parameters, protecting data privacy while enabling cross-regional knowledge sharing. Dynamic weighted aggregation and local model verification mechanisms address the performance degradation caused by differences in data distribution across different regions and devices, preventing abnormal models from contaminating the global model. Transformer deep inference + digital twin + reinforcement learning realizes an intelligent closed loop from fault identification to policy optimization, enabling early simulation of fault evolution paths and generation of optimal repair solutions. The three-level early warning system combines three interaction methods: AI voice, platform, and APP, to meet the needs of immediacy, intuitiveness, and operability in different operation and maintenance scenarios, and significantly reduce the false alarm and missed alarm rates. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall architecture of an electrical equipment fault diagnosis and early warning system based on federated learning and hybrid intelligence according to the present invention. Figure 2 This is a block diagram of the edge computing module of the present invention; Figure 3 This is a block diagram of the federated learning module of the present invention; Figure 4 This is a block diagram of the hybrid intelligent diagnostic module of the present invention; Figure 5 This is a flowchart illustrating the steps of an electrical equipment fault diagnosis and early warning method based on federated learning and hybrid intelligence according to the present invention. Explanation of icon numbers: 1-Data acquisition module; 2-Edge computing module; 21-Feature extraction unit; 22-Feature library and diagnostic rule base construction unit; 23-In-situ training unit; 3- Federated Learning Module; 31- Parameter Transmission Interface; 32- Dynamic Weighted Aggregation Unit; 33- Local Model Validation Unit; 4- Hybrid Intelligent Diagnostic Module; 41- Transformer Deep Inference Unit; 42- Digital Twin Simulation Unit; 43- Reinforcement Learning Optimization Unit; 5-Level 3 Early Warning and Interaction Module. Detailed Implementation

[0018] The present invention will be further described below with reference to the accompanying drawings and specific embodiments: like Figure 1 As shown, the present invention provides an electrical equipment fault diagnosis and early warning system based on federated learning and hybrid intelligence, including a data acquisition module 1, an edge computing module 2, a federated learning module 3, a hybrid intelligence diagnosis module 4, and a three-level early warning and interaction module 5. The data acquisition module 1 is configured to synchronously acquire the three-phase voltage and three-phase current of the electrical circuit, as well as the vibration signal, travel signal, and coil current signal of the main switchgear, to form a multi-source heterogeneous monitoring dataset; like Figure 2 As shown, the edge computing module 2 is deployed on the local terminal and includes a feature extraction unit 21, a feature library and diagnostic rule library construction unit 22, and an local training unit 23. The feature extraction unit 21 uses a hybrid model of convolutional neural network and long short-term memory network to extract spatiotemporal joint features from the multi-source heterogeneous monitoring dataset. The convolutional neural network is used to extract the spatial distribution features of the fault waveform image, and the long short-term memory network is used to extract the temporal dependence features of the fault evolution. The feature library and diagnostic rule library construction unit 22 is used to map the spatiotemporal joint features to preset fault classification labels to construct a classification feature library and a corresponding diagnostic rule library including electrical circuit faults and main switch mechanical faults. The in-situ training unit 23 is used to train a preset lightweight diagnostic model locally based on the classification feature library and the diagnostic rule library, and generate local model parameters. like Figure 3 As shown, the federated learning module 3 includes a parameter transmission interface 31, a dynamic weighted aggregation unit 32, and a local model verification unit 33. The parameter transmission interface 31 is configured to upload the local model parameters to the central server after encryption, and to receive the aggregated parameters sent by the central server. The dynamic weighted aggregation unit 32 is deployed on the central server and is used to calculate the dynamic weight of each local terminal based on the number of samples of each local terminal and the loss function of the local model, and to perform weighted aggregation of the parameters of each local model based on the dynamic weight to generate global model parameters. The local model verification unit 33 is deployed on the local terminal. After receiving the aggregated parameters from the central server, it first performs a performance evaluation on the local verification set. If the evaluation result is better than the current local model, the model parameters are updated; otherwise, the update is rejected and the existing parameters are maintained. like Figure 4As shown, the hybrid intelligent diagnostic module 4 includes a Transformer deep inference unit 41, a digital twin simulation unit 42, and a reinforcement learning optimization unit 43; The Transformer deep inference unit 41 is configured to receive the global model parameters or the verified local model parameters, perform contextual understanding on the long sequence of power grid state information, and generate preliminary fault diagnosis results. The digital twin simulation unit 42 is used to create a digital twin model of the equipment based on the preliminary fault diagnosis results, and to simulate the fault evolution process and analyze hypothetical scenarios. The reinforcement learning optimization unit 43 is used to perform interactive training in the simulation environment provided by the digital twin simulation unit, dynamically optimize the fault response strategy and repair plan, and generate diagnostic conclusions including fault causes, fault locations, handling suggestions and preventive measures. The three-level early warning and interaction module 5 is configured to divide the early warning into three levels based on the confidence and urgency of the diagnostic conclusion: prompt level, warning level, and emergency level; and output early warning information corresponding to the early warning level through three methods: AI intelligent voice interaction, remote platform early warning push, and mobile APP message prompt. The early warning information includes the cause of the fault, the location of the fault, handling suggestions, and preventive measures.

[0019] As can be seen from the above description, the present invention has the following beneficial effects: This invention provides an electrical equipment fault diagnosis and early warning system based on federated learning and hybrid intelligence. Through a complete closed-loop architecture of "data acquisition—edge computing—federated learning—hybrid intelligent diagnosis—three-level early warning interaction," it achieves full-process coverage from bottom-level multi-source heterogeneous data acquisition to high-level intelligent decision-making and early warning. Specific effects include: using a CNN-LSTM hybrid model to simultaneously extract spatial features (such as waveform images) and temporal features (such as evolution trends) of faults, improving diagnostic accuracy by approximately 15-25% compared to a single model. The federated learning module allows local terminals to transmit encrypted model parameters without sharing raw data, protecting data privacy while enabling cross-regional knowledge sharing. Dynamic weighted aggregation and local model verification mechanisms solve the problem of model performance degradation caused by differences in data distribution across different regions and devices, preventing abnormal models from polluting the global model. Transformer deep inference + digital twin + reinforcement learning realizes an intelligent closed loop from fault identification to policy optimization, which can simulate fault evolution paths in advance and generate optimal repair solutions. The three-level early warning system combines three interaction methods: AI voice, platform, and APP, to meet the needs of immediacy, intuitiveness, and operability in different operation and maintenance scenarios, and significantly reduce the false alarm and missed alarm rates.

[0020] Furthermore, in the dynamic weighted aggregation unit, the first... Dynamic weight of each local terminal Calculate using the following formula: ; in, For the first The number of local fault samples possessed by each local terminal. For the first The loss function value of the local diagnostic model for each local terminal on the validation set. The total number of local terminals participating in federated learning.

[0021] As can be seen from the above description, the dynamic weights are limited. The specific calculation method is as follows. Compared to the equal-weighted aggregation of the traditional Federal Average (FedAvg), this formula has the following outstanding advantages: it considers the sample size simultaneously. and model performance (loss function value) This gives stakeholders with large datasets and high diagnostic accuracy greater influence in the global model, accelerating model convergence. (Logarithmic function) This approach smooths out weight imbalances caused by large differences in sample size, preventing a single site from dominating the overall model. The formula requires no additional communication overhead; it only needs to include the loss value when uploading parameters, making it easy to deploy in engineering projects.

[0022] Furthermore, the local model verification unit is specifically configured as follows: it loads the aggregated parameters issued by the central server into the local model, calculates the accuracy of the updated model using the locally reserved verification set, and accepts the update if the accuracy is higher than the accuracy of the model before the update by more than a preset threshold; if the accuracy is lower than the accuracy of the model before the update by less than the preset threshold, it rejects the update and sends a rejection event to the central server.

[0023] As described above, the addition of a "local refusal to update" mechanism to traditional federated learning brings the following significant benefits: When the aggregated parameters issued by the central server perform poorly on the local validation set (e.g., due to an abnormal gradient uploaded by a certain site), the local model can refuse to update, ensuring that the online system always uses the optimal model. For sites with extremely unique data distributions (e.g., very few samples from newly built power plants), forced updates may lead to a precipitous drop in performance. This mechanism allows them to continue using the original model, avoiding a "getting worse with each iteration" scenario. Feeding back the refusal event to the central server can trigger retraining of the global model or isolation of abnormal sites, forming a closed-loop optimization.

[0024] Furthermore, in the three-level early warning and interaction module, the confidence level of the diagnostic conclusion corresponding to the prompt level is not less than the first threshold and less than the second threshold, which is used to indicate that the device has potential abnormalities but has not yet formed a clear fault. The confidence level of the diagnostic conclusion corresponding to the warning level is not less than the second threshold and less than the third threshold, and is used to indicate that the equipment has a clear fault but has not yet endangered continuous operation; The confidence level of the diagnostic conclusion corresponding to the emergency level is not less than the third threshold, which is used to indicate that the equipment has suffered a serious malfunction and needs to be shut down for repair immediately. Wherein, the first threshold is less than the second threshold, and the second threshold is less than the third threshold.

[0025] As described above, based on diagnostic confidence thresholds, alerts are categorized into three levels: "Alert," "Warning," and "Emergency," with clearly defined device states for each level. This results in the following effects: Maintenance personnel can adopt differentiated response strategies based on the alert level (e.g., alert levels only require attention, while emergency levels require immediate shutdown), avoiding resource waste caused by applying the same handling method to all anomalies. Tiered alerts reduce the frequent push of low-confidence alerts, allowing maintenance personnel to focus on truly urgent faults and improving work efficiency. Each alert level is directly linked to a confidence interval, facilitating user understanding of the alert's reliability and enabling subsequent automatic system decision-making (e.g., emergency levels automatically trigger work orders).

[0026] Furthermore, the AI ​​intelligent voice interaction supports natural language question and answer. When maintenance personnel issue a voice query command, the system will broadcast the cause of the fault, the location of the fault, the handling suggestions and preventive measures in the diagnostic conclusion in voice form, and adopt different voice prompt frequencies and semantic urgency according to the warning level.

[0027] As described above, dynamically binding AI voice interaction with alert levels achieves the following effects: In scenarios such as inspections and emergency repairs, maintenance personnel can obtain information about the cause, location, handling suggestions, and preventative measures for faults via voice without needing to view the screen, improving on-site operational efficiency. Emergency levels utilize higher frequency and more urgent semantics (such as repeated broadcasts and faster speech), while prompt levels use a more conventional tone, allowing maintenance personnel to judge the severity of faults solely through hearing, shortening emergency response time. The system supports voice queries by maintenance personnel for specific devices or time periods, and can generate customized answers based on a diagnostic conclusion database, lowering the barrier to human-computer interaction.

[0028] Furthermore, in the feature library and diagnostic rule library construction unit, the classification features of the electrical circuit faults include three-phase voltage imbalance, harmonic distortion rate, zero-sequence current amplitude, and transient traveling wave polarity; the classification features of the main switch mechanical faults include the time-domain peak value of the vibration signal, frequency-domain energy distribution, inflection point position of the stroke curve, and the start time and duration of the coil current.

[0029] As described above, the key characteristic parameters for electrical line faults and main switch mechanical faults are specifically defined, resulting in the following effects: The listed characteristics (such as three-phase voltage imbalance, harmonic distortion rate, and vibration signal time-domain peak value) are all recognized physical indicators in the power industry, facilitating engineer verification and optimization, and also enabling the formation of a standardized feature library. Simultaneously, it includes both electrical quantity characteristics (voltage / current / harmonics) and mechanical quantity characteristics (vibration / stroke / coil current), compensating for the insufficient ability of single electrical diagnostics to identify mechanical faults (such as jamming of operating mechanisms). By directly inputting high-quality, manually designed features into the neural network instead of raw data, the model's dependence on massive amounts of data can be reduced, accelerating convergence.

[0030] Furthermore, the reinforcement learning optimization unit adopts a near-end policy optimization algorithm. Its state space includes the current health score of the device, load rate, and historical fault frequency. Its action space includes switch opening and closing operations, load transfer instructions, and maintenance priority ranking. Its reward function comprehensively considers fault recovery time, power supply reliability impact, and remaining lifespan loss of the device.

[0031] As described above, defining the specific algorithm (PPO) and its state / action / reward space design for the reinforcement learning optimizer yields the following results: PPO training in a digital twin environment can simulate thousands of fault scenarios and intervention measures, avoiding the safety risks associated with trial and error on real equipment. The reward function simultaneously considers fault recovery time, power supply reliability, and remaining equipment lifespan depletion, ensuring that the generated repair strategy is not limited to "fastest repair" but rather "comprehensive optimal repair." The state space includes dynamic variables such as equipment health, load rate, and historical fault frequency, enabling the strategy to automatically adjust based on equipment aging and real-time load, achieving personalized operation and maintenance.

[0032] like Figure 5 As shown, this invention also provides a method for fault diagnosis and early warning of electrical equipment based on federated learning and hybrid intelligence, applied to the aforementioned fault diagnosis and early warning system for electrical equipment based on federated learning and hybrid intelligence, comprising the following steps: S1. Synchronously collect the three-phase voltage and three-phase current of the electrical circuit, as well as the vibration signal, travel signal, and coil current signal of the main switchgear, to form a multi-source heterogeneous monitoring dataset; S2. At the local terminal, a hybrid model of convolutional neural network-long short-term memory network is used to extract spatiotemporal joint features from the multi-source heterogeneous monitoring dataset. The convolutional neural network is used to extract the spatial distribution features of the fault waveform image, and the long short-term memory network is used to extract the temporal dependence features of the fault evolution. The spatiotemporal joint features are mapped to preset fault classification labels to construct a classification feature library and a corresponding diagnostic rule library including electrical line faults and main switch mechanical faults. S3. Train the preset lightweight diagnostic model locally based on the classification feature library and diagnostic rule library to generate local model parameters. Upload the local model parameters to the central server after encryption and receive the aggregated parameters issued by the central server. S4. The central server calculates the dynamic weight of each local terminal based on the number of samples of each local terminal and the loss function of the local model, and performs weighted aggregation of the parameters of each local model based on the dynamic weight to generate global model parameters and distribute them to each local terminal. S5. After receiving the global model parameters, the local terminal first performs a performance evaluation on the local validation set. If the evaluation result is better than the current local model, the model parameters are updated; otherwise, the update is rejected and the existing parameters are maintained. S6. Use the Transformer model to perform contextual understanding on long-sequence power grid state information and generate preliminary fault diagnosis results; S7. Based on the preliminary fault diagnosis results, create a digital twin model of the equipment, simulate the fault evolution process and analyze hypothetical scenarios, and dynamically optimize the fault response strategy through reinforcement learning to generate diagnostic conclusions including fault causes, fault locations, handling suggestions and preventive measures. S8. Based on the confidence level and urgency of the diagnostic conclusion, three warning levels are defined: alert level, warning level, and emergency level. Warning information corresponding to the warning level is output through three methods: AI intelligent voice interaction, remote platform warning push, and mobile APP message prompt. The warning information includes the cause of the fault, the location of the fault, handling suggestions, and preventive measures.

[0033] As can be seen from the above description, the present invention has the following beneficial effects: This invention provides a fault diagnosis and early warning method for electrical equipment based on federated learning and hybrid intelligence, corresponding to the aforementioned system architecture. It defines a complete fault diagnosis and early warning method flow (S1-S8), producing a similar effect to the system but with a greater emphasis on the orderly nature of the method steps: from data acquisition → feature extraction → in-situ training → weighted aggregation → local verification → Transformer inference → digital twin simulation → three-level early warning, forming a standardized processing pipeline that facilitates software development and deployment. Each operation step has clearly defined input and output, allowing different engineers to achieve consistent diagnostic results, facilitating cross-team collaboration and patent infringement evidence collection. Edge computing (S2-S3) handles tasks with high real-time requirements, while the cloud or central layer (S4-S7) handles complex inference tasks, achieving a rational allocation of computing resources.

[0034] Furthermore, in step S4, the formula for calculating the dynamic weight is: ; in, For the first The number of local fault samples possessed by each local terminal. For the first The loss function value of the local diagnostic model for each local terminal on the validation set. The total number of local terminals participating in federated learning.

[0035] As can be seen from the above description, the dynamic weights are limited. The specific calculation method is as follows. Compared to the equal-weighted aggregation of the traditional Federal Average (FedAvg), this formula has the following outstanding advantages: it considers the sample size simultaneously. and model performance (loss function value) This gives stakeholders with large datasets and high diagnostic accuracy greater influence in the global model, accelerating model convergence. (Logarithmic function) This approach smooths out weight imbalances caused by large differences in sample size, preventing a single site from dominating the overall model. The formula requires no additional communication overhead; it only needs to include the loss value when uploading parameters, making it easy to deploy in engineering projects.

[0036] Furthermore, in step S8, the AI ​​intelligent voice interaction supports natural language question and answer. When the maintenance personnel issue a voice query command, the system broadcasts the cause of the fault, the location of the fault, the handling suggestions and preventive measures in the diagnostic conclusion in voice form, and adopts different voice prompt frequencies and semantic urgency according to the warning level.

[0037] As described above, dynamically binding AI voice interaction with alert levels achieves the following effects: In scenarios such as inspections and emergency repairs, maintenance personnel can obtain information about the cause, location, handling suggestions, and preventative measures for faults via voice without needing to view the screen, improving on-site operational efficiency. Emergency levels utilize higher frequency and more urgent semantics (such as repeated broadcasts and faster speech), while prompt levels use a more conventional tone, allowing maintenance personnel to judge the severity of faults solely through hearing, shortening emergency response time. The system supports voice queries by maintenance personnel for specific devices or time periods, and can generate customized answers based on a diagnostic conclusion database, lowering the barrier to human-computer interaction.

[0038] The following are several preferred embodiments or application embodiments to help those skilled in the art better understand the technical content of the present invention and the technical contributions made by the present invention compared with the prior art: Example 1 like Figure 1As shown, the electrical equipment fault diagnosis and early warning system based on federated learning and hybrid intelligence provided in this embodiment includes: multiple local terminals deployed in substations or distribution rooms (each terminal corresponds to one line or one switch), a regional central server (which can be located in a municipal dispatch center), and a mobile APP and remote monitoring platform used by maintenance personnel. The local terminals and the central server are connected via a dedicated power network or a 5G network, and the transmitted data is encrypted using the national cryptographic algorithm SM4.

[0039] The local terminal includes a data acquisition module and an edge computing module; the central server includes a central aggregation unit of the federated learning module, a Transformer deep inference unit of the hybrid intelligent diagnostic module, a digital twin simulation unit, and a reinforcement learning optimization unit; the three-level early warning and interaction module is partially deployed on the central server, and partially deployed on the mobile APP and remote platform.

[0040] The electrical equipment fault diagnosis and early warning method based on federated learning and hybrid intelligence provided in this embodiment is applied to the above-mentioned system. The data acquisition and preprocessing are as follows: This embodiment uses a 10kV outgoing line and its associated vacuum circuit breaker of a 110kV substation as an example for illustration.

[0041] Step 1: Synchronous Acquisition of Multi-Source Heterogeneous Data The data acquisition module synchronously acquires the following signals at a sampling rate of 12.8kHz: Electrical circuit signals: instantaneous values ​​of the secondary side voltages of the three-phase voltage transformers A, B, and C; instantaneous values ​​of the secondary side currents of the three-phase current transformers A, B, and C.

[0042] Main switch mechanical signals: Vibration sensor (sampling rate 50kHz) installed on the circuit breaker operating mechanism outputs vibration acceleration signal; travel sensor (linear displacement sensor) outputs moving contact travel signal; Hall current sensor collects closing / opening coil current signal.

[0043] The acquisition window is set to cover 5 cycles before the fault occurs to 10 cycles after the fault occurs. For example, when the sudden change in line current exceeds the set value, waveform recording is triggered, storing data for a total of 15 cycles before and after the fault.

[0044] Step 2: Data Cleaning and Standardization The edge computing module filters the raw data (using a bandpass filter to remove power frequency noise and high-frequency interference), removes outliers (using the 3σ criterion), and normalizes the data. The three-phase current and three-phase voltage data are normalized to the [0,1] interval, while the vibration signal, travel signal, and coil current signal are normalized to their respective maximum amplitudes. After normalization, the three-phase current / voltage waveform data are converted into an RGB three-channel fault transient waveform image (224×224 pixels), and the vibration signal, travel signal, and coil current signal are each retained as a one-dimensional time sequence (2048 points in length).

[0045] Step 3: Extracting spatiotemporal joint features using a CNN-LSTM hybrid model The transient waveform image of the fault is input into the convolutional neural network (CNN) branch. The CNN adopts the ResNet-18 structure, which includes 5 convolutional blocks. Each convolutional block is followed by batch normalization and ReLU activation function. Finally, a 512-dimensional spatial feature vector is output through global average pooling. This vector represents the distribution pattern of the fault waveform in different frequency bands and spatial locations.

[0046] Simultaneously, a one-dimensional time sequence of vibration signal, stroke signal, and coil current signal is input into a Long Short-Term Memory (LSTM) branch. The LSTM consists of two layers, each with 128 hidden units, and a time step of 256. The LSTM outputs a 256-dimensional time-series feature vector, which captures the dynamic evolution of the mechanical fault (e.g., the start time, peak value, and duration of the closing coil current, and the inflection point of the stroke curve).

[0047] By concatenating the 512-dimensional spatial feature vector with the 256-dimensional temporal feature vector, a 768-dimensional spatiotemporal joint feature vector is obtained.

[0048] Step 4: Construct the feature library and diagnostic rule library Based on historical fault samples (including actual field faults and simulated test data), each fault type is labeled. This embodiment defines electrical circuit fault types as follows: A-phase grounding, B-phase grounding, C-phase grounding, AB two-phase short circuit, BC two-phase short circuit, CA two-phase short circuit, and three-phase short circuit; main switch mechanical fault types include: operating mechanism jamming, coil breakage, auxiliary switch poor contact, energy storage spring fatigue, and contact wear.

[0049] For each sample, a 768-dimensional feature vector is associated with its fault label to form a feature library. Based on this, a decision tree-based rule extraction algorithm is used to automatically generate an interpretable diagnostic rule library. For example: Rule 1: If the three-phase voltage imbalance is greater than 15% and the zero-sequence current amplitude is greater than 0.2 times the rated current, it is determined to be a "single-phase ground fault". The fault location is calculated based on traveling wave ranging.

[0050] Rule 2: If the peak value of the vibration signal in the time domain is greater than 10g, the duration of the coil current is greater than 35ms, and the last 10% of the stroke curve takes more than 5ms, then it is determined as "the operating mechanism is stuck".

[0051] The feature library and diagnostic rule library are stored in the local database of the edge computing module and are updated regularly in sync with the central server.

[0052] Step 5: Local Lightweight Model Training Each local terminal trains a lightweight diagnostic model using samples from its local feature library (e.g., 800 fault samples collected from a substation over the past year). This embodiment uses MobileNetV3-Small as the base classifier, with 768-dimensional spatiotemporal joint features as input and probability distributions for 16 fault types (including normal states) as output. Training employs the Adam optimizer with an initial learning rate of 0.001, a batch size of 32, and 20 epochs. After training, the local model parameters are obtained. (t is the training round, k is the site number).

[0053] Step 6: Parameter Upload and Dynamic Weighted Aggregation Each local terminal will transmit the encrypted local model parameters and the loss function value of this training. Uploaded to the central server. The central server processes the samples based on the number of samples from each terminal. Calculate the dynamic weights based on the loss value: ; For example, there are three sites: Site A (800 samples, loss 0.23), Site B (1200 samples, loss 0.18), and Site C (200 samples, loss 0.45). Their calculated weights are 0.35, 0.55, and 0.10, respectively. The central server calculates a weighted average of the parameters from each site according to these weights to obtain the global model parameters. .

[0054] Step 7: Local Model Validation and Update The central server distributes the aggregated global parameters to each site. Each site evaluates its performance on its locally reserved validation set (20% of the total samples). If the accuracy of the updated model is more than 2 percentage points higher than the current model, the update is accepted; otherwise, the update is rejected, and a "reject" event is sent to the central server. For example, if site C's local accuracy is actually lowered (from 85% to 82%) due to a small sample size and high noise, site C will reject the update and continue using the original model. Upon receiving the rejection event, the central server will appropriately reduce the weight of that site or temporarily exclude it in the next round of aggregation.

[0055] Step 8: Transformer Contextual Reasoning For the global or local model optimized by federated learning, when a new fault occurs, the system organizes the multi-dimensional time-series data (including voltage, current, vibration, travel, and coil current) for 15 cycles before and after the fault into a token sequence of length 4096, and inputs it into the Transformer model (12 layers, 8 attention heads). The Transformer captures the long-distance dependencies between various electrical and mechanical quantities through a self-attention mechanism. For example, it notices that the time of current change is highly correlated with the time of vibration impact, thus inferring that the fault is caused by a circuit breaker forced tripping due to a line short circuit. The Transformer outputs preliminary diagnostic results, including the fault category (electrical / mechanical) and confidence level.

[0056] Step 9: Digital Twin Simulation and Reinforcement Learning Optimization The preliminary diagnostic results are input into the digital twin platform. Based on the equipment nameplate parameters, historical operating data, and real-time topology, the digital twin platform establishes a high-fidelity simulation model. For example, for a circuit breaker determined to have "operating mechanism jamming," the digital twin model will simulate the mechanism's operating characteristics under different temperature and lubrication conditions, generating 100 possible jamming evolution paths.

[0057] The reinforcement learning optimizer uses the PPO algorithm and is trained in a digital twin environment. The state space includes: device health (0~1), current load rate (%), ambient temperature, and fault type code. The action space includes: immediate repair, repair with a 4-hour delay, load transfer, and remote reclosing attempt. The reward function is defined as: After 2000 simulation iterations, the reinforcement learning model outputs the optimal strategy. For example, for a circuit breaker that is slightly stuck and has a low load rate, it is recommended to "delay maintenance for 4 hours and transfer some loads in advance"; for a circuit breaker that is severely stuck and is located in an important load area, it is recommended to "immediately shut down the power for maintenance and put the backup power supply into operation".

[0058] The final diagnostic conclusions include: cause of the fault ("closing coil core jamming caused mechanism to fail to operate"), fault location ("10kV 101 circuit breaker operating mechanism box"), handling recommendations ("replace the closing coil, clean the core, add grease"), and preventive measures ("conduct a low voltage opening and closing test every quarter and monitor the coil current waveform").

[0059] Step 10: Classification of Three-Level Early Warning Levels Based on diagnostic confidence The warning system is divided into three levels based on the numbers (0~1) and the severity of the fault consequences: Warning Level (Blue): The equipment exhibits potential anomalies but has not yet developed into a definite fault. For example, the peak coil current is 15% higher than the historical average but is still within acceptable limits.

[0060] Warning level (yellow): The equipment has exhibited a clear malfunction but has not yet jeopardized continuous operation. For example, there are obvious shock pulses in the vibration signal, but it can still continue to operate for 24 hours.

[0061] Emergency Level (Red): The equipment has suffered a serious malfunction and requires immediate shutdown and repair. For example, the three-phase current suddenly drops to zero and the circuit breaker position signal indicates that it has tripped.

[0062] Step 11: Multimodal early warning output AI-powered intelligent voice interaction: A voice module is integrated into the inspection robot or the smart helmet worn by maintenance personnel. When an emergency level is triggered, the system proactively announces: "Emergency warning! The operating mechanism of the 10kV101 circuit breaker is jammed. Please proceed immediately to handle the situation. The fault location is in the third cabinet of the second row in the switch room." Maintenance personnel can also use voice to inquire: "What warnings are currently in effect?" The system replies: "There is one warning: Phase A ground fault on line. Live-line testing is recommended." Remote platform early warning push: On the dispatch center's large screen, the fault point is displayed in the form of a heat map and flashing icons. Clicking the icon will bring up a detailed diagnostic report (including waveform, characteristic parameters, and handling suggestions).

[0063] Mobile App Message Notification: Pushes a message with sound and vibration to the on-duty engineer's mobile phone, including a fault summary, the expected scope of impact, and a list of recommended repair tools.

[0064] Step 12: Feedback and Continuous Learning After troubleshooting, maintenance personnel fill in the actual cause of the fault and the handling result through the app. The system uses this feedback as new labeled samples, adds them to the feature library, and triggers the next round of incremental training. The federated learning framework performs a global model update monthly, enabling the system to continuously adapt to equipment aging and changes in operating conditions.

[0065] Using the above method, a six-month field test was conducted on 10 10kV lines and 30 circuit breakers at a power supply company. The results showed that: The accuracy of fault diagnosis has increased from 78.3% to 96.8% using traditional methods; the average time from the occurrence of a fault to the issuance of a Level 3 warning has been reduced from 15 minutes to 18 seconds; the number of power outages avoided due to timely warnings has increased by 6, reducing economic losses by approximately 1.2 million yuan; and maintenance personnel have saved an average of 8 minutes each time by using AI voice interaction to review drawings and manuals on-site.

[0066] In summary, this invention provides an electrical equipment fault diagnosis and early warning system and method based on federated learning and hybrid intelligence. Through a complete closed-loop architecture of "data acquisition—edge computing—federated learning—hybrid intelligent diagnosis—three-level early warning interaction," it achieves full-process coverage from bottom-level multi-source heterogeneous data acquisition to high-level intelligent decision-making and early warning. Specific effects include: using a CNN-LSTM hybrid model to simultaneously extract spatial features (such as waveform images) and temporal features (such as evolution trends) of faults, improving diagnostic accuracy by approximately 15-25% compared to a single model. The federated learning module eliminates the need for local terminals to share raw data, transmitting only encrypted model parameters, protecting data privacy while enabling cross-regional knowledge sharing. Dynamic weighted aggregation and local model verification mechanisms address the performance degradation caused by differences in data distribution across different regions and devices, preventing abnormal models from contaminating the global model. Transformer deep inference + digital twin + reinforcement learning achieve an intelligent closed loop from fault identification to policy optimization, enabling early simulation of fault evolution paths and generation of optimal repair solutions. The three-level early warning system combines three interaction methods: AI voice, platform, and APP, to meet the needs of immediacy, intuitiveness, and operability in different operation and maintenance scenarios, and significantly reduce the false alarm and missed alarm rates.

[0067] The present invention has been described with reference to the foregoing embodiments and accompanying drawings; however, the foregoing embodiments are merely examples for implementing the present invention. It must be noted that the disclosed embodiments do not limit the scope of the present invention. Conversely, modifications and equivalents included within the spirit and scope of the claims are all included within the scope of the present invention.

Claims

1. A fault diagnosis and early warning system for electrical equipment based on federated learning and hybrid intelligence, characterized in that, It includes a data acquisition module, an edge computing module, a federated learning module, a hybrid intelligent diagnostic module, and a three-level early warning and interaction module; The data acquisition module is configured to synchronously acquire the three-phase voltage and three-phase current of the electrical circuit, as well as the vibration signal, travel signal, and coil current signal of the main switchgear, forming a multi-source heterogeneous monitoring dataset; The edge computing module is deployed on the local terminal and includes a feature extraction unit, a feature library and diagnostic rule library construction unit, and an local training unit. The feature extraction unit adopts a hybrid model of convolutional neural network and long short-term memory network to extract spatiotemporal joint features from the multi-source heterogeneous monitoring dataset. The convolutional neural network is used to extract the spatial distribution features of the fault waveform image, and the long short-term memory network is used to extract the temporal dependence features of the fault evolution. The feature library and diagnostic rule library construction unit is used to map the spatiotemporal joint features to preset fault classification labels, and construct a classification feature library and corresponding diagnostic rule library including electrical circuit faults and main switch mechanical faults. The in-situ training unit is used to train a preset lightweight diagnostic model locally based on the classification feature library and the diagnostic rule library, and generate local model parameters. The federated learning module includes a parameter transmission interface, a dynamic weighted aggregation unit, and a local model verification unit. The parameter transmission interface is configured to encrypt the local model parameters and upload them to the central server, and to receive the aggregated parameters sent by the central server. The dynamic weighted aggregation unit, deployed on the central server, is used to calculate the dynamic weight of each local terminal based on the number of samples of each local terminal and the loss function of the local model, and to perform weighted aggregation of the parameters of each local model based on the dynamic weight to generate global model parameters. The local model verification unit is deployed on the local terminal. After receiving the aggregated parameters from the central server, it first performs a performance evaluation on the local verification set. If the evaluation result is better than the current local model, the model parameters are updated; otherwise, the update is rejected and the existing parameters are maintained. The hybrid intelligent diagnostic module includes a Transformer deep inference unit, a digital twin simulation unit, and a reinforcement learning optimization unit; The Transformer deep inference unit is configured to receive the global model parameters or the verified local model parameters, perform contextual understanding on long-sequence power grid state information, and generate preliminary fault diagnosis results. The digital twin simulation unit is used to create a digital twin model of the equipment based on the preliminary fault diagnosis results, and to simulate the fault evolution process and analyze hypothetical scenarios. The reinforcement learning optimization unit is used to perform interactive training in the simulation environment provided by the digital twin simulation unit, dynamically optimize the fault response strategy and repair plan, and generate diagnostic conclusions including fault causes, fault locations, handling suggestions and preventive measures. The three-level early warning and interaction module is configured to divide the early warning into three levels based on the confidence and urgency of the diagnostic conclusion: prompt level, warning level, and emergency level; The system outputs warning information corresponding to the warning level through three methods: AI intelligent voice interaction, remote platform warning push, and mobile APP message notification. The warning information includes the cause of the fault, the location of the fault, handling suggestions, and preventive measures.

2. The electrical equipment fault diagnosis and early warning system based on federated learning and hybrid intelligence according to claim 1, characterized in that, In the dynamic weighted aggregation unit, the first... Dynamic weight of each local terminal Calculate using the following formula: ； in, For the first The number of local fault samples possessed by each local terminal. For the first The loss function value of the local diagnostic model for each local terminal on the validation set. The total number of local terminals participating in federated learning.

3. The electrical equipment fault diagnosis and early warning system based on federated learning and hybrid intelligence according to claim 1, characterized in that, The local model verification unit is specifically configured as follows: it loads the aggregated parameters issued by the central server into the local model, calculates the accuracy of the updated model using the locally reserved verification set, and accepts the update if the accuracy is improved by more than a preset threshold compared to the accuracy of the model before the update; if the accuracy is lower than the preset threshold compared to the accuracy of the model before the update, it rejects the update and sends a rejection event to the central server.

4. The electrical equipment fault diagnosis and early warning system based on federated learning and hybrid intelligence according to claim 1, characterized in that, In the three-level early warning and interaction module, the confidence level of the diagnostic conclusion corresponding to the prompt level is not less than the first threshold and less than the second threshold, which is used to indicate that the device has potential abnormalities but has not yet formed a clear fault. The confidence level of the diagnostic conclusion corresponding to the warning level is not less than the second threshold and less than the third threshold, and is used to indicate that the equipment has a clear fault but has not yet endangered continuous operation; The confidence level of the diagnostic conclusion corresponding to the emergency level is not less than the third threshold, which is used to indicate that the equipment has suffered a serious malfunction and needs to be shut down for repair immediately. Wherein, the first threshold is less than the second threshold, and the second threshold is less than the third threshold.

5. The electrical equipment fault diagnosis and early warning system based on federated learning and hybrid intelligence according to claim 1, characterized in that, The AI ​​intelligent voice interaction supports natural language question and answer. When maintenance personnel issue a voice query command, the system will broadcast the cause of the fault, the location of the fault, the handling suggestions and preventive measures in the diagnostic conclusion in voice form, and adopt different voice prompt frequencies and semantic urgency according to the warning level.

6. The electrical equipment fault diagnosis and early warning system based on federated learning and hybrid intelligence according to claim 1, characterized in that, In the feature library and diagnostic rule library construction unit, the classification features of the electrical circuit faults include three-phase voltage imbalance, harmonic distortion rate, zero-sequence current amplitude, and transient traveling wave polarity; the classification features of the main switch mechanical faults include the time domain peak value of the vibration signal, frequency domain energy distribution, inflection point position of the stroke curve, and the start time and duration of the coil current.

7. The electrical equipment fault diagnosis and early warning system based on federated learning and hybrid intelligence according to claim 1, characterized in that, The reinforcement learning optimization unit adopts a near-end policy optimization algorithm. Its state space includes the current health score of the device, load rate and historical fault frequency. Its action space includes switch opening and closing operations, load transfer instructions and maintenance priority ranking. Its reward function comprehensively considers fault recovery time, power supply reliability impact and remaining lifespan loss of the device.

8. A method for fault diagnosis and early warning of electrical equipment based on federated learning and hybrid intelligence, applied to the fault diagnosis and early warning system for electrical equipment based on federated learning and hybrid intelligence as described in any one of claims 1 to 7, characterized in that, Includes the following steps: S1. Synchronously collect the three-phase voltage and three-phase current of the electrical circuit, as well as the vibration signal, travel signal, and coil current signal of the main switchgear, to form a multi-source heterogeneous monitoring dataset; S2. At the local terminal, a hybrid model of convolutional neural network-long short-term memory network is used to extract spatiotemporal joint features from the multi-source heterogeneous monitoring dataset. The convolutional neural network is used to extract the spatial distribution features of the fault waveform image, and the long short-term memory network is used to extract the temporal dependence features of the fault evolution. The spatiotemporal joint features are mapped to preset fault classification labels to construct a classification feature library and a corresponding diagnostic rule library including electrical line faults and main switch mechanical faults. S3. Train the preset lightweight diagnostic model locally based on the classification feature library and diagnostic rule library to generate local model parameters. Upload the local model parameters to the central server after encryption and receive the aggregated parameters issued by the central server. S4. The central server calculates the dynamic weight of each local terminal based on the number of samples of each local terminal and the loss function of the local model, and performs weighted aggregation of the parameters of each local model based on the dynamic weight to generate global model parameters and distribute them to each local terminal. S5. After receiving the global model parameters, the local terminal first performs a performance evaluation on the local validation set. If the evaluation result is better than the current local model, the model parameters are updated; otherwise, the update is rejected and the existing parameters are maintained. S6. Use the Transformer model to perform contextual understanding on long-sequence power grid state information and generate preliminary fault diagnosis results; S7. Based on the preliminary fault diagnosis results, create a digital twin model of the equipment, simulate the fault evolution process and analyze hypothetical scenarios, and dynamically optimize the fault response strategy through reinforcement learning to generate diagnostic conclusions including fault causes, fault locations, handling suggestions and preventive measures. S8. Based on the confidence level and urgency of the diagnostic conclusion, three warning levels are defined: alert level, warning level, and emergency level. The system outputs warning information corresponding to the warning level through three methods: AI intelligent voice interaction, remote platform warning push, and mobile APP message notification. The warning information includes the cause of the fault, the location of the fault, handling suggestions, and preventive measures.

9. The electrical equipment fault diagnosis and early warning method based on federated learning and hybrid intelligence according to claim 8, characterized in that, In step S4, the formula for calculating the dynamic weight is: ； in, For the first The number of local fault samples possessed by each local terminal. For the first The loss function value of the local diagnostic model for each local terminal on the validation set. The total number of local terminals participating in federated learning.

10. The electrical equipment fault diagnosis and early warning method based on federated learning and hybrid intelligence according to claim 8, characterized in that, In step S8, the AI ​​intelligent voice interaction supports natural language question and answer. When the operation and maintenance personnel issue a voice query command, the system will broadcast the cause of the fault, the location of the fault, the handling suggestions and preventive measures in the diagnostic conclusion in voice form, and adopt different voice prompt frequencies and semantic urgency according to the warning level.

Images