Light-weight AI-based phase modifier edge diagnosis system and method

Abstract

The invention discloses a phase modifier edge diagnosis system and method based on lightweight AI, relates to the technical field of power management, and aims to solve the technical problem of large automatic defect of current phase modifier diagnosis. Comprising a multi-mode sensing module, a lightweight AI edge deployment module, a network communication module, a data processing and storage module and an autonomous decision-making and fault diagnosis module, and the lightweight AI edge deployment module comprises an adaptive unit to avoid total model transmission. According to the system, significant technical gain is realized through the lightweight AI edge deployment module, the adaptive unit is embedded into the working condition feature mapping layer, model weight is dynamically adjusted in combination with working condition parameters such as the load rate of the phase modifier and the voltage of the power grid, misinformation caused by dirty data input is reduced, and the reliability of the system is improved. And the edge increment updating unit only transmits the deviation sample to perform local parameter updating, so that the diagnosis accuracy, the working condition adaptability and the operation efficiency are comprehensively improved, and efficient and reliable technical support is provided for the edge diagnosis of the phase modifier.

Description

Technical Field

[0001] This invention relates to the field of power management technology, and more specifically, to a synchronous condenser edge diagnosis system and method based on lightweight AI. Background Technology

[0002] In the operation and management of the power system, DC synchronous condensers, as important reactive power compensation devices, play a crucial role in maintaining grid voltage stability and improving transmission capacity. The operating status of key components such as carbon brushes and slip rings directly affects grid stability.

[0003] Existing diagnostic technologies mostly rely on a single sensor to collect data, which is susceptible to electromagnetic interference leading to noise superposition. They also suffer from problems such as temporary data loss and invalid jumps. Furthermore, they lack the ability to identify complex operating conditions such as carbon brush overheating and compound faults, making it difficult to accurately locate the fault location and extent.

[0004] Traditional AI diagnostic models have a large number of parameters, making them difficult to deploy at the edge. Full model transmission and updates increase network load and costs. Furthermore, existing systems lack adaptive adjustment mechanisms based on operating conditions, have fixed detection cycles, cannot dynamically optimize based on equipment health status, suffer from delayed fault warnings, and have short fault-free periods in some systems, making it difficult to cover a major overhaul cycle of a synchronous condenser. Consequently, they fail to meet the power system's requirements for real-time, accurate, and long-term reliable diagnostics. Therefore, we propose a lightweight AI-based edge diagnostic system and method for synchronous condensers. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the existing technology, adapt to the needs of reality, and provide a camera edge diagnosis system and method based on lightweight AI to solve the technical problem of large defects in the current camera diagnosis automation.

[0006] To address the aforementioned technical problems, this invention provides the following technical solution: a lightweight AI-based edge diagnostic system for synchronous condensers, comprising a multimodal perception module, a lightweight AI edge deployment module, a network communication module, a data processing and storage module, and an autonomous decision-making and fault diagnosis module. The lightweight AI edge deployment module includes an adaptive unit that embeds a working condition feature mapping layer into the pruning model, collects synchronous condenser load rate and grid voltage working condition parameters in real time, and dynamically adjusts model weights; a multi-fault generalization training unit that adds a composite fault sample library of the synchronous condenser before model pruning and adjusts the model's ability to identify composite faults through transfer learning; a data missing compensation unit that, based on a temporal difference algorithm, completes temporary data missingness within 1-5 seconds, with a completion error ≤±1% and ≤±0.5℃; and an outlier filtering unit that adds... + Dual filtering mechanism based on operating condition thresholds, firstly... The principle is to identify statistical outliers, and then combine them with the camera condenser operating condition threshold for secondary screening to remove invalid jump data. The edge incremental update unit is designed with a lightweight incremental training module. When the autonomous decision-making unit reports a diagnostic deviation, only the deviation sample is transmitted to the edge side to update the local parameters of the pruning model, avoiding the transmission of the entire model.

[0007] Preferably, the lightweight AI edge deployment module further includes a key feature extraction unit, which extracts fault-sensitive features in advance during the preprocessing stage, including carbon brush temperature gradient and current harmonic distortion rate, to adjust the computational load of subsequent model inference. The key feature extraction unit, together with the data missing compensation unit and the outlier filtering unit, adjusts the clarity of the data preprocessing function in the lightweight AI edge deployment module to optimize the data quality on the edge side.

[0008] Preferably, the lightweight AI edge deployment module further includes a data hierarchical collaboration unit, which divides the preliminary diagnostic structure into three levels of status and outputs key features to the autonomous decision-making and fault diagnosis module based on the severity of the latter two levels of status.

[0009] Preferably, the multimodal sensing module collects the output detection data status of the camera converter, which includes a thermal imaging sensing unit for generating thermal images in real time and extracting the temperature of carbon brush, slip ring and demagnetizing cabinet components, a current sensing unit for collecting carbon brush operating current and odd harmonic components, an arc sensing unit, and a wear state sensing unit for establishing a wear rate model based on recording laser round-trip time by pulse method and converting carbon brush thickness change.

[0010] Preferably, the autonomous decision-making and fault diagnosis module includes a fusion AI model unit, a fault identification unit, a dynamic detection unit, and an alarm unit. The fusion AI model unit includes a random forest-LSTM fusion model, a spatiotemporal graph convolutional network (ST-GCN) model, and a Transformer time series model. The random forest-LSTM fusion model is used to analyze time series data to identify progressive degradation of carbon brushes. The spatiotemporal graph convolutional network model fuses device topology and time series data to predict fault probability. The Transformer model can provide early warning of carbon brush current over-limit and high temperature problems one month in advance.

[0011] Preferably, the fault identification unit classifies the fault types such as carbon brush overheating, sparking, current imbalance, and excessive wear, assesses the degree of fault, and locates the fault location. The dynamic detection unit adjusts the detection cycle according to the health status of the equipment. Under normal conditions, the detection interval is 5 minutes / time, and under abnormal conditions, it switches to 1 second / time. The alarm unit outputs an alarm signal based on the detection status of the dynamic detection unit.

[0012] A lightweight AI-based edge diagnosis method for camera adjustment systems, applied to any of the above-described lightweight AI-based edge diagnosis systems for camera adjustment systems, comprises the following diagnostic steps: Step 1: Based on the multimodal sensing module, collect the core operating parameters of the carbon brush of the DC synchronous condenser in a comprehensive and high-precision manner; Step 2: The collected data is purified and transmitted using the network communication module and the data processing and storage module. Noise caused by electromagnetic interference is removed using the FFT filtering algorithm. The protocol adaptation module adaptively adapts to the corresponding protocol and standardizes the heterogeneous data. The clock synchronization module corrects the spatiotemporal deviation of the multi-source data. Step 3: The output data is used to perform data feature extraction and AI diagnostic inference at the edge based on the lightweight AI edge deployment module and the autonomous decision-making and fault diagnosis module; Step 4: Based on the autonomous decision-making and fault diagnosis module, combined with the AI ​​diagnostic data from Step 3, output fault conclusions and trigger the alarm mechanism, including type identification, periodic adjustment, and alarm mode based on the fault identification unit; Step 5: Based on the alarm data and output handling detection from Step 4, verify the fault status of the camera and create a labeled sample; Step Six: Import the labeled samples output from Step Five into the edge AI model, and periodically update the Random Forest-LSTM and Spatiotemporal Graph Convolutional Network models. At the same time, optimize the sensor acquisition accuracy based on the fault status of the updated data, and perform reliability verification on the optimized data and sensors.

[0013] Preferably, the data feature extraction in step three specifically involves using tensor decomposition technology to jointly analyze multi-dimensional data such as temperature, current, and ignition, screening key fault-related features, and simultaneously using self-supervised learning technology to identify baseline drift and sensor failure using unlabeled data, triggering real-time calibration.

[0014] Preferably, the AI ​​diagnostic inference in step three specifically involves running a pruned AI model based on an edge-side RISC-V chip, i.e., a BERT pruned model with less than 1MB of parameters, to process thermal imaging data and identify abnormal temperature regions using a MobileNet / TinyML model. Furthermore, the above model is used in conjunction with a fusion AI model unit for fusion computation.

[0015] Preferably, the reliability verification in step six specifically involves monitoring the mean time between failures (MTBF) of the system. By optimizing the emergency structure on the synchronous condenser, the MTBF is guaranteed to be ≥8000 hours, covering one major overhaul cycle of the synchronous condenser.

[0016] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention achieves significant technical gains through a lightweight AI edge deployment module: the adaptive unit embeds a working condition feature mapping layer, dynamically adjusting model weights based on working condition parameters such as synchronous condenser load rate and grid voltage, enabling the AI ​​model to adapt to changes in working conditions in real time and improving diagnostic adaptability; the multi-fault generalization training unit introduces a composite fault sample library and leverages transfer learning to enhance the model's ability to identify and generalize composite faults, overcoming the limitations of single fault diagnosis; the data missing compensation unit accurately completes 1-5 seconds of temporary data missing based on a time-series difference algorithm, with the completion error controlled within ≤±1% and ≤±0.5℃, ensuring data integrity; the outlier filtering unit... The dual mechanism of thresholding based on operating conditions eliminates invalid jump data, reducing false alarms caused by "dirty data" input. The edge incremental update unit only transmits deviation samples for local parameter updates, avoiding full model transmission and reducing network load and update costs. In summary, this module ensures lightweight deployment of AI models while comprehensively improving diagnostic accuracy, operating condition adaptability, and operational efficiency, providing efficient and reliable technical support for edge diagnosis of synchronous condensers.

[0017] 2. This invention also utilizes a feature extraction unit to pre-capture fault-sensitive features such as carbon brush temperature gradient and current harmonic distortion rate, which can accurately reduce the computational load of subsequent model inference and improve edge-side computational efficiency. Its collaborative optimization of data preprocessing with data missing compensation and outlier filtering units refines data purification accuracy, reduces interference from invalid data on the model, and further solidifies the diagnostic data foundation. The data hierarchical collaboration unit, through a three-level state division, makes the transmission of preliminary diagnostic results more hierarchical. The latter two levels, with their output of key features, allow the autonomous decision-making module to quickly focus on core information, improving the targeting and response efficiency of fault judgment. Together, these two elements enhance the accuracy of system data processing and the synergy of the decision-making chain.

[0018] 3. This invention also utilizes the system's diagnostic method, which relies on a multimodal sensing module to achieve high-precision, multi-dimensional acquisition of core carbon brush parameters. Combined with FFT filtering, protocol adaptation, and clock synchronization technologies, data purification is achieved, effectively avoiding electromagnetic interference and spatiotemporal deviation issues. Through preprocessing optimization, key feature extraction, and lightweight model inference using a lightweight AI edge deployment module, diagnostic calculations are efficiently completed at the edge, reducing network transmission pressure and resource consumption. The integrated AI model unit accurately identifies fault types, predicts fault probabilities, and provides early warnings. The dynamic detection unit adaptively adjusts the detection cycle based on health status, and the alarm mechanism responds promptly to risks. Labeled samples enable incremental model updates and sensor accuracy optimization. Reliability verification ensures the system covers the overhaul cycle, comprehensively improving the accuracy, real-time performance, and long-term stability of synchronous condenser fault diagnosis. Attached Figure Description

[0019] Figure 1This is a schematic diagram of the system structure of the present invention; Figure 2 This is a schematic diagram of the process structure of the present invention. Detailed Implementation

[0020] like Figures 1 to 2 As shown, this invention relates to a lightweight AI-based edge diagnostic system for synchronous condensers, comprising a multimodal perception module, a lightweight AI edge deployment module, a network communication module, a data processing and storage module, and an autonomous decision-making and fault diagnosis module. The lightweight AI edge deployment module includes an adaptive unit that embeds a condition feature mapping layer into the pruning model, real-time acquisition of synchronous condenser load rate and grid voltage condition parameters, and dynamic adjustment of model weights. A multi-fault generalization training unit adds a composite fault sample library of the synchronous condenser before model pruning, adjusting the model's ability to identify composite faults through transfer learning. A data missing compensation unit, based on a temporal difference algorithm, completes temporary data missing data within 1-5 seconds, with a completion error ≤±1% and ≤±0.5℃. An outlier filtering unit adds... + Dual filtering mechanism based on operating condition thresholds, firstly... The principle is to identify statistical outliers, and then combine them with the camera condenser operating condition threshold for secondary screening to remove invalid jump data. The edge incremental update unit is designed with a lightweight incremental training module. When the autonomous decision-making unit reports a diagnostic deviation, only the deviation sample is transmitted to the edge side to update the local parameters of the pruning model, avoiding the transmission of the entire model.

[0021] This invention achieves significant technical gains through a lightweight AI edge deployment module: the adaptive unit embeds a working condition feature mapping layer, dynamically adjusting model weights based on working condition parameters such as synchronous condenser load rate and grid voltage, enabling the AI ​​model to adapt to changes in working conditions in real time and improving diagnostic adaptability; the multi-fault generalization training unit introduces a composite fault sample library and leverages transfer learning to enhance the model's ability to identify and generalize composite faults, overcoming the limitations of single fault diagnosis; the data missing compensation unit accurately completes 1-5 second temporary data missingness based on a time-series difference algorithm, with the completion error controlled within ≤±1% and ≤±0.5℃, ensuring data integrity; the outlier filtering unit... The dual mechanism of thresholding based on operating conditions eliminates invalid jump data, reducing false alarms caused by "dirty data" input. The edge incremental update unit only transmits deviation samples for local parameter updates, avoiding full model transmission and reducing network load and update costs. In summary, this module ensures lightweight deployment of AI models while comprehensively improving diagnostic accuracy, operating condition adaptability, and operational efficiency, providing efficient and reliable technical support for edge diagnosis of synchronous condensers.

[0022] Specifically, the lightweight AI edge deployment module also includes a key feature extraction unit, which extracts fault-sensitive features in advance during the preprocessing stage, including carbon brush temperature gradient and current harmonic distortion rate, to adjust the computational load of subsequent model inference. The key feature extraction unit, together with the data missing compensation unit and outlier filtering unit, adjusts the clarity of the data preprocessing function in the lightweight AI edge deployment module to optimize the data quality on the edge side. The key feature extraction unit accurately extracts fault-sensitive features such as carbon brush temperature gradient and current harmonic distortion rate during the preprocessing stage, which can effectively reduce the computational complexity of subsequent model inference and improve the efficiency of edge-side computation. It works in conjunction with the data missing compensation unit and the outlier filtering unit to optimize the data preprocessing process, further refine the data purification accuracy, reduce the interference of invalid data on the model, and ensure the high quality of edge-side diagnostic input data.

[0023] Furthermore, the lightweight AI edge deployment module also includes a data hierarchical collaboration unit, which divides the preliminary diagnostic structure into three levels of status and outputs key features to the autonomous decision-making and fault diagnosis module based on the severity of the latter two levels of status. The data hierarchical collaboration unit divides the preliminary diagnostic results into three levels, enabling hierarchical transmission of diagnostic information. For the latter two high-severity levels, key feature outputs are provided to offer targeted information support for autonomous decision-making and fault diagnosis modules, significantly improving the pertinence of fault judgment and the efficiency of decision response.

[0024] This invention also utilizes a feature extraction unit to pre-capture fault-sensitive features such as carbon brush temperature gradient and current harmonic distortion rate, which can accurately reduce the computational load of subsequent model inference and improve edge-side computational efficiency. Its collaborative optimization of data preprocessing with data missing compensation and outlier filtering units refines data purification accuracy, reduces interference from invalid data on the model, and further solidifies the diagnostic data foundation. The data hierarchical collaboration unit, through a three-level state division, makes the transmission of preliminary diagnostic results more hierarchical. The latter two levels, with their key feature outputs, allow the autonomous decision-making module to quickly focus on core information, improving the targeting and response efficiency of fault judgment. Together, these two elements enhance the accuracy of system data processing and the synergy of the decision-making chain.

[0025] It is worth noting that the multimodal sensing module collects the output detection data status of the camera condenser. It includes a thermal imaging sensing unit for generating thermal images in real time and extracting the temperature of carbon brush, slip ring and demagnetizing cabinet components; a current sensing unit for collecting carbon brush operating current and odd harmonic components; an arc sensing unit; and a wear state sensing unit for calculating the wear rate model based on recording the laser round-trip time by pulse method and converting the carbon brush thickness change. The multimodal sensing module uses sensing units such as thermal imaging, current, arcing, and wear status to achieve multi-dimensional and high-precision acquisition of operating parameters of core components such as synchronous condenser carbon brushes, slip rings, and demagnetizing cabinets. This avoids the one-sidedness of data from a single sensing dimension and provides a comprehensive and reliable data source for subsequent diagnostic reasoning.

[0026] It is worth mentioning that the autonomous decision-making and fault diagnosis module includes an AI model fusion unit, a fault identification unit, a dynamic detection unit, and an alarm unit. The AI ​​model fusion unit includes a random forest-LSTM fusion model, a spatiotemporal graph convolutional network (ST-GCN) model, and a Transformer time series model. The random forest-LSTM fusion model is used to analyze time series data to identify the progressive degradation of carbon brushes. The spatiotemporal graph convolutional network model fuses equipment topology and time series data to predict the probability of failure. The Transformer model can provide early warning of carbon brush current exceeding limits and high temperature problems up to one month in advance. In the AI ​​model unit that integrates autonomous decision-making and fault diagnosis modules, the random forest-LSTM fusion model accurately analyzes time-series data to identify progressive degradation of carbon brushes, the spatiotemporal graph convolutional network model integrates equipment topology and time-series data to predict fault probability, and the Transformer model can provide early warning of carbon brush current over-limit and temperature over-limit issues up to one month in advance, significantly improving the accuracy and lead time of fault diagnosis.

[0027] It is worth noting that the fault identification unit classifies the fault types such as carbon brush overheating, sparking, current imbalance, and excessive wear, assesses the degree of fault and locates the fault location. The dynamic detection unit adjusts the detection cycle according to the health status of the equipment. Under normal conditions, the detection interval is 5 minutes / time, and under abnormal conditions, it switches to 1 second / time. The alarm unit outputs an alarm signal based on the detection status of the dynamic detection unit. The fault identification unit can accurately classify fault types such as carbon brush overheating and sparking, and simultaneously complete fault severity assessment and location positioning; the dynamic detection unit adaptively switches the detection cycle according to the equipment health status (normal 5 minutes / time, abnormal 1 second / time), and the alarm unit outputs alarm signals in conjunction, realizing the synergy of fault diagnosis accuracy, detection flexibility and risk response time.

[0028] A lightweight AI-based edge diagnosis method for camera adjustment systems, based on any one of the above, comprises the following diagnostic steps: Step 1: Based on the multimodal sensing module, collect the core operating parameters of the carbon brush of the DC synchronous condenser in a comprehensive and high-precision manner; Step 2: The collected data is purified and transmitted using the network communication module and the data processing and storage module. Noise caused by electromagnetic interference is removed using the FFT filtering algorithm. The protocol adaptation module adaptively adapts to the corresponding protocol and standardizes the heterogeneous data. The clock synchronization module corrects the spatiotemporal deviation of the multi-source data. Step 3: The output data is used to perform data feature extraction and AI diagnostic inference at the edge based on the lightweight AI edge deployment module and the autonomous decision-making and fault diagnosis module; Step 4: Based on the autonomous decision-making and fault diagnosis module, combined with the AI ​​diagnostic data from Step 3, output fault conclusions and trigger the alarm mechanism, including type identification, periodic adjustment, and alarm mode based on the fault identification unit; Step 5: Based on the alarm data and output handling detection from Step 4, verify the fault status of the camera and create a labeled sample; Step Six: Import the labeled samples output from Step Five into the edge AI model, and periodically update the Random Forest-LSTM and Spatiotemporal Graph Convolutional Network models. At the same time, optimize the sensor acquisition accuracy based on the fault status of the updated data, and perform reliability verification on the optimized data and sensors.

[0029] This invention also utilizes a diagnostic method within the system. This method relies on a multimodal sensing module to achieve high-precision, multi-dimensional acquisition of core carbon brush parameters. Combined with FFT filtering, protocol adaptation, and clock synchronization technologies, data purification is achieved, effectively avoiding electromagnetic interference and spatiotemporal deviation issues. Through preprocessing optimization, key feature extraction, and lightweight model inference using a lightweight AI edge deployment module, diagnostic calculations are efficiently completed at the edge, reducing network transmission pressure and resource consumption. The integrated AI model unit accurately identifies fault types, predicts fault probabilities, and provides early warnings. The dynamic detection unit adaptively adjusts the detection cycle based on health status, and the alarm mechanism responds promptly to risks. Labeled samples enable incremental model updates and sensor accuracy optimization. Reliability verification ensures the system covers the overhaul cycle, comprehensively improving the accuracy, real-time performance, and long-term stability of synchronous condenser fault diagnosis.

[0030] Specifically, the data feature extraction in step three involves using tensor decomposition technology to jointly analyze multi-dimensional data such as temperature, current, and ignition, screening key fault-related features, and using self-supervised learning technology to identify baseline drift and sensor failure using unlabeled data, triggering real-time calibration. By combining tensor decomposition technology with multi-dimensional data analysis, key features associated with faults can be accurately screened, improving diagnostic targeting. Meanwhile, self-supervised learning can use unlabeled data to identify baseline drift and sensor failure, triggering real-time calibration and avoiding data bias from affecting diagnostic accuracy.

[0031] Furthermore, the AI ​​diagnostic inference in step three specifically involves running a pruned AI model based on the edge RISC-V chip, namely a BERT pruned model with less than 1MB of parameters, to process thermal imaging data and identify abnormal temperature regions using a MobileNet / TinyML model. The model is then used in conjunction with a fusion AI model unit to perform fusion calculations. Based on a pruning model with less than 1MB of parameters running on the edge RISC-V chip, it adapts to edge resource constraints and efficiently processes thermal imaging data and identifies temperature anomalies; combined with fusion AI model computation, it ensures diagnostic accuracy while controlling resource consumption.

[0032] It is worth noting that the reliability verification in step six specifically involves monitoring the mean time between failures (MTBF) of the system. By optimizing the emergency structure on the synchronous condenser, the MTBF is guaranteed to be ≥8000 hours, covering one major overhaul cycle of the synchronous condenser. By monitoring mean time between failures (MTBF) and optimizing the emergency structure, the system's MTBF is ensured to be ≥8000 hours, which can fully cover the overhaul cycle of synchronous condensers, reduce maintenance interruptions, and improve the long-term operational reliability of the diagnostic system and the stability of power supply.

[0033] The embodiments disclosed in this invention are preferred embodiments, but are not limited thereto. Those skilled in the art can easily understand the spirit of this invention based on the above embodiments and make different extensions and variations, but as long as they do not depart from the spirit of this invention, they are all within the protection scope of this invention.

Claims

1. A camera edge diagnostic system based on lightweight AI, characterized in that, It includes a multimodal perception module, a lightweight AI edge deployment module, a network communication module, a data processing and storage module, and an autonomous decision-making and fault diagnosis module; The lightweight AI edge deployment module includes an adaptive unit, which embeds a working condition feature mapping layer in the pruning model to collect synchronous condenser load rate and grid voltage working condition parameters in real time and dynamically adjust the model weights. The multi-fault generalization training unit incorporates a composite fault sample library from the camera adjustment device before model pruning, and adjusts the model's ability to identify composite faults through transfer learning. The data missing compensation unit, based on the time-series difference algorithm, fills in temporary data missing within 1-5 seconds, with a completion error of ≤±1% and ≤±0.5℃. Outlier filtering unit: Add + Dual filtering mechanism based on operating condition thresholds, firstly... The principle is to identify statistical outliers, and then combine this with the synchronous condenser operating condition threshold for secondary screening to remove invalid jump data; The edge incremental update unit is designed with a lightweight incremental training module. When the autonomous decision-making unit reports a diagnostic deviation, it only transmits the deviation sample to the edge side to update the local parameters of the pruned model, thus avoiding the transmission of the entire model.

2. The camera edge diagnosis system based on lightweight AI according to claim 1, characterized in that, The lightweight AI edge deployment module also includes a key feature extraction unit, which extracts fault-sensitive features in advance during the preprocessing stage, including carbon brush temperature gradient and current harmonic distortion rate, to adjust the computational load of subsequent model inference. The key feature extraction unit, together with the data missing compensation unit and the outlier filtering unit, adjusts the clarity of the data preprocessing function in the lightweight AI edge deployment module to optimize the data quality on the edge side.

3. The camera edge diagnosis system based on lightweight AI according to claim 2, characterized in that, The lightweight AI edge deployment module also includes a data hierarchical collaboration unit, which divides the preliminary diagnostic structure into three levels of status and outputs key features to the autonomous decision-making and fault diagnosis module based on the severity of the latter two levels of status.

4. The camera edge diagnosis system based on lightweight AI according to claim 3, characterized in that, The multimodal sensing module collects the output detection data status of the camera converter. It includes a thermal imaging sensing unit for generating thermal images in real time and extracting the temperature of carbon brush, slip ring and demagnetizing cabinet components, a current sensing unit for collecting carbon brush operating current and odd harmonic components, an arc sensing unit, and a wear state sensing unit for establishing a wear rate model based on recording the laser round-trip time by pulse method and converting the carbon brush thickness change.

5. The camera edge diagnosis system based on lightweight AI according to claim 1, characterized in that, The autonomous decision-making and fault diagnosis module includes a fusion AI model unit, a fault identification unit, a dynamic detection unit, and an alarm unit. The fusion AI model unit includes a random forest-LSTM fusion model, a spatiotemporal graph convolutional network model, and a Transformer time series model. The random forest-LSTM fusion model is used to analyze time series data to identify progressive degradation of carbon brushes. The spatiotemporal graph convolutional network model fuses device topology and time series data to predict the probability of failure. The Transformer model can provide early warning of carbon brush current exceeding limits and high temperature problems one month in advance.

6. The camera edge diagnosis system based on lightweight AI according to claim 5, characterized in that, The fault identification unit classifies fault types such as carbon brush overheating, sparking, current imbalance, and excessive wear, assesses the degree of fault, and locates the fault location. The dynamic detection unit adjusts the detection cycle according to the health status of the equipment. Under normal conditions, the detection interval is 5 minutes / time, and under abnormal conditions, it switches to 1 second / time. The alarm unit outputs an alarm signal based on the detection status of the dynamic detection unit.

7. A lightweight AI-based edge diagnosis method for camera adjustment systems, applied to the lightweight AI-based edge diagnosis system for camera adjustment systems described in any one of claims 1-6, characterized in that, The diagnostic steps of this method are as follows: Step 1: Based on the multimodal sensing module, collect the core operating parameters of the carbon brush of the DC synchronous condenser in a comprehensive and high-precision manner; Step 2: The collected data is purified and transmitted using the network communication module and the data processing and storage module, and noise caused by electromagnetic interference is removed using the FFT filtering algorithm; The protocol adaptation module adaptively adapts to the corresponding protocol and standardizes heterogeneous data, while the clock synchronization module corrects the spatiotemporal deviation of multi-source data. Step 3: The output data is used to perform data feature extraction and AI diagnostic inference at the edge based on the lightweight AI edge deployment module and the autonomous decision-making and fault diagnosis module; Step 4: Based on the autonomous decision-making and fault diagnosis module, combined with the AI ​​diagnostic data from Step 3, output fault conclusions and trigger the alarm mechanism, including type identification, periodic adjustment, and alarm mode based on the fault identification unit; Step 5: Based on the alarm data and output handling detection from Step 4, verify the fault status of the camera and create a labeled sample; Step Six: Import the labeled samples output from Step Five into the edge AI model, and periodically update the Random Forest-LSTM and Spatiotemporal Graph Convolutional Network models. At the same time, optimize the sensor acquisition accuracy based on the fault status of the updated data, and perform reliability verification on the optimized data and sensors.

8. The camera edge diagnosis method based on lightweight AI according to claim 7, characterized in that, The data feature extraction in step three specifically involves using tensor decomposition technology to jointly analyze multi-dimensional data such as temperature, current, and ignition, screening key fault-related features, and using self-supervised learning technology to identify baseline drift and sensor failure using unlabeled data, triggering real-time calibration.

9. The camera edge diagnosis method based on lightweight AI according to claim 7, characterized in that, The AI ​​diagnostic inference in step three specifically involves running a pruned AI model on the edge RISC-V chip, namely a BERT pruned model with less than 1MB of parameters, to process thermal imaging data and identify abnormal temperature regions using a MobileNet / TinyML model. Furthermore, the above models are combined with a fusion AI model unit for fusion computation.

10. A camera edge diagnosis method based on lightweight AI according to claim 7, characterized in that, The reliability verification in step six specifically involves monitoring the mean time between failures (MTBF) of the system. By optimizing the emergency structure on the synchronous condenser, the MTBF is guaranteed to be ≥8000 hours, covering one major overhaul cycle of the synchronous condenser.

Images