Partial discharge identification system algorithm based on ultraviolet sensor array

Abstract

The invention discloses a partial discharge identification system algorithm based on an ultraviolet sensor array, and relates to the technical field of high-voltage electrical equipment monitoring. The method comprises the following steps: step a, nine ultraviolet photoelectric detectors arranged in a 3 * 3 matrix are arranged to form a sensor array, and each sensing unit independently collects an ultraviolet light intensity signal generated by partial discharge; according to the invention, the spatial distribution information of partial discharge is obtained by arranging the ultraviolet sensor array, the limitation of single-point detection is broken through, the anti-interference capability of the system in a complex electromagnetic environment is enhanced, discriminative features are automatically extracted from a spatial intensity relationship by using a self-adaptive optimized feature transformation function, and the accuracy of partial discharge detection is improved. The defect that a traditional method depends on artificial experience is overcome, the effectiveness of features and the generalization recognition capability of modes are improved, and the adaptability of the system to random noise in actual working conditions and the data stability are further improved by combining preprocessing and data enhancement strategies.

Description

Technical Field

[0001] This invention relates to the field of high-voltage electrical equipment monitoring technology, and in particular to an algorithm for a partial discharge identification system based on an ultraviolet sensor array. Background Technology

[0002] Partial discharge, a non-penetrating discharge phenomenon caused by microscopic defects within the insulation system of high-voltage electrical equipment, often begins to appear in the early stages of insulation material aging, moisture absorption, or the presence of air gaps. Its discharge mode, energy intensity, and development law are closely related to the insulation degradation mechanism. During the long-term operation of power equipment, different types of partial discharge correspond to different insulation defect morphologies and severity. By accurately identifying the characteristic differences of these discharge types, potential insulation hazards inside the equipment can be effectively revealed. This provides a key basis for predicting the remaining life of the insulation system, planning preventive maintenance strategies, and avoiding sudden insulation breakdown accidents, thereby ensuring the safe and stable operation of the power grid.

[0003] Traditional detection methods mainly include pulsed current method, ultrasonic method, ultra-high frequency method and optical detection method, etc. However, existing technologies have the following prominent problems: 1. The detection method is limited and the anti-interference ability is weak. The single sensor is susceptible to electromagnetic interference and environmental noise, resulting in a high false alarm rate under complex working conditions. 2. Pattern recognition has poor real-time performance. Traditional methods rely on host computers for offline analysis, which is difficult to meet the real-time diagnostic needs of edge devices. 3. The algorithm has high complexity and is difficult to port to embedded systems. Although deep learning models have high accuracy, they have a large number of parameters and are computationally complex, making them difficult to deploy on resource-constrained microcontroller platforms such as STM32. 4. Feature engineering relies on human experience, and conventional methods require manual feature extraction, which has insufficient generalization ability and is difficult to adapt to new discharge types.

[0004] Ultraviolet detection utilizes ultraviolet photons generated during partial discharge for non-contact monitoring. Its physical properties determine that this method is naturally immune to electromagnetic interference. At the same time, the fast propagation speed of light signals makes the detection response almost without delay. These advantages make it uniquely valuable in the field of partial discharge monitoring of high-voltage equipment. However, current ultraviolet sensing systems generally adopt a single detection point architecture, which can only acquire the intensity information of the discharge source from a single perspective. It cannot capture the spatial distribution characteristics and energy diffusion patterns of the discharge phenomenon. This lack of spatial information makes it difficult for the system to distinguish the characteristic patterns of different types of discharge. At the same time, existing systems mostly transmit the raw signal to the host computer for post-processing, lacking efficient recognition algorithms deployed at the sensor end. They have neither designed simplified models to address the limitations of microcontroller resources nor achieved real-time feature extraction and pattern judgment at the data acquisition source. This results in significant time delays in the entire detection process, which cannot meet the urgent needs of power equipment fault early warning for real-time performance and on-site decision-making.

[0005] To address this issue, we provide an algorithm for a partial discharge identification system based on an ultraviolet sensor array. Summary of the Invention

[0006] The purpose of this invention is to provide an algorithm for a partial discharge identification system based on an ultraviolet sensor array. This algorithm acquires spatial distribution information of discharge by arranging an ultraviolet sensor array, achieves embedded and efficient classification by using adaptive feature extraction and a lightweight multilayer perceptron model, and combines fuzzy comprehensive decision-making with time-series information for comprehensive judgment. This solves the problems of weak anti-interference ability of single-point detection, poor real-time performance of pattern recognition that is difficult to meet edge requirements, difficulty in porting complex algorithms, and insufficient generalization ability due to reliance on human experience in feature engineering in the prior art.

[0007] To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution: This invention relates to an algorithm for a partial discharge identification system based on an ultraviolet sensor array. Step a: Arrange nine ultraviolet photodetectors in a 3×3 matrix to form a sensor array. Each sensing unit independently collects the ultraviolet light intensity signal generated by partial discharge, and after analog-to-digital conversion, forms a 9-dimensional original signal vector. Step b: Perform normalization preprocessing on the original signal vector, calculated using the following formula: ; in The signal after normalization The original response signal, This is the average response value of the array. Standard deviation, To prevent zero constant; Step c: Extract primary feature vectors based on the signals of each sensing unit and their spatial neighborhood relationships. The calculation formula is as follows: ; in For the first Primary features of the unit To differentiate the signal from that of adjacent units, The mean of the neighborhood signal. , , It is an adaptively determinable feature transformation function; Step d: Transform all primary features into an intermediate feature vector using a nonlinear mapping function, expressed as: ; in For the first One intermediate feature component It is a nonlinear mapping operator; Step e: Use the trained lightweight multilayer perceptron model to classify the intermediate feature vectors, calculate the response intensity of each category and normalize it to obtain the confidence level of the current period; Step f: Combining dynamic change indicators and historical recognition stability, a fuzzy comprehensive decision model is used to calculate the comprehensive confidence score for each category. The category with the highest score is selected as the recognition result output. By deploying a 3×3 matrix ultraviolet sensor array, the spatial distribution information of partial discharge is acquired simultaneously, overcoming the limitations of single-point detection and laying the foundation for subsequent feature extraction and pattern recognition based on spatial relationships. The original signal is then normalized and preprocessed to eliminate the influence of sensor differences and environmental fluctuations, improving data stability. Based on this, rich and discriminative spatial-intensity joint primary features are extracted by comprehensively utilizing the unit's own intensity, neighborhood differences, and regional mean. These features are then automatically fused and abstracted using a nonlinear mapping function to achieve dimensionality reduction and form a compact intermediate feature vector. A trained lightweight multilayer perceptron model is used to efficiently classify the intermediate features, generating the confidence score for the current period, achieving real-time computation on the embedded platform. Finally, combining dynamic change indicators and historical recognition stability, a fuzzy comprehensive decision model is used to perform multi-dimensional information fusion, outputting the recognition result with the highest comprehensive confidence score. The entire method systematically integrates the three major advantages of spatial perception, efficient and lightweight model and intelligent temporal decision-making, thereby achieving high real-time performance, high accuracy and robust partial discharge type identification under complex working conditions.

[0008] The present invention further includes step b, which involves augmenting and partitioning the preprocessed data. Specifically, Gaussian white noise with a standard deviation of 0.05 is added to the data using a MATLAB script. 1000 samples are generated for each discharge category, and all samples are randomly divided into a training set (80%) and a test set (20%). Adding Gaussian white noise for data augmentation effectively simulates random interference in actual working conditions, expands the dataset, and improves the model's generalization ability and noise immunity. Standardized data partitioning ensures the scientific rigor and reliability of model training and evaluation.

[0009] The present invention is further configured such that, in step e, the response intensity of each category is calculated and normalized to obtain the confidence level of the current period, specifically including: Calculate the response intensity for each category: ; in For category numbering, Assign feature weights to each category; normalize the response intensity of each category to obtain the confidence level for the current period: ; in and These represent the minimum and maximum response strengths, respectively. As a small constant, the category response intensity is calculated by weighted summation, reflecting the differences in the contribution of different features to each category; further, the response intensity is transformed into a unified confidence metric through normalization, making the confidence scores of different categories and different periods comparable, and providing standardized input for subsequent comprehensive decision-making.

[0010] The present invention is further configured such that, in step f, the comprehensive confidence score of each category is calculated using a fuzzy comprehensive decision model and the identification result is selected, specifically including: The comprehensive confidence score for each category is calculated using a fuzzy comprehensive decision model. ; in As a dynamically changing indicator, For historical stability parameters, , , These are the weighting coefficients, and their sum is 1; The category with the highest overall confidence score is selected as the recognition result output, and a resampling and labeling mechanism is triggered for events with confidence scores below a set threshold. The fuzzy comprehensive decision model organically integrates the current confidence level, dynamic change trend, and historical performance to achieve a more comprehensive evaluation. The mechanism of setting a threshold to trigger resampling can automatically identify low-confidence events, thereby initiating a review process, effectively reducing false alarms and false negatives, and improving the system's ability to respond to boundary situations and abnormal events.

[0011] The present invention is further configured such that step f includes a model deployment step: automatically extracting the network weights, biases, batch normalized parameters, and standardized parameters from the trained lightweight multilayer perceptron model, converting them into C language format files suitable for the STM32 system, and simultaneously generating a C language header file containing constant definitions, parameter arrays, and inference function interfaces. This achieves seamless integration from the training environment to the embedded deployment environment. The automated conversion toolchain avoids errors that may occur during manual porting, and the generated standardized C language file is easy to integrate directly into the STM32 project, greatly simplifying the deployment process and improving development efficiency and system reliability.

[0012] The present invention is further configured such that the feature transformation function in step c is... , , During the model training phase, the neighborhood range is determined by adaptive optimization using the backpropagation algorithm. The neighborhood range is defined as the eight units outside the current unit within a 3×3 region centered on the current unit. The adaptive optimization of the feature transformation function replaces the traditional design that relies on human experience, enabling the extracted features to better serve the final classification goal and improving the effectiveness of the features. The explicit 3×3 neighborhood definition makes full use of the compact structure of the array, ensuring the consistency and integrity of spatial relationship feature calculation.

[0013] The present invention is further configured such that the nonlinear mapping operator in step d... It is implemented using fully connected layers based on Sigmoid or ReLU activation functions, and the output dimension of the mapping layer is... Less than the total number of primary features input To achieve feature dimensionality reduction, mature and efficient sigmoid or ReLU activation functions are used to ensure the expressive power and training stability of nonlinear mapping. The forced dimensionality reduction design effectively compresses the feature dimension, significantly reducing subsequent computation and storage overhead while retaining key information, making it particularly suitable for resource-constrained embedded scenarios.

[0014] The present invention is further configured such that the lightweight multilayer perceptron model in step e includes an input layer, at least one hidden layer and an output layer, the number of neurons in the hidden layer does not exceed 32, and the overall parameters are optimized to adapt to the storage and computing power constraints of the STM32 series microcontrollers. The network structure has been strictly designed to be lightweight to ensure that the model parameters and computational complexity match the limited memory and computing power of microcontrollers such as STM32, thereby truly realizing real-time, online recognition on edge devices and avoiding the latency and bandwidth consumption caused by data uploading.

[0015] The present invention is further configured such that the dynamic change index The historical stability parameter is obtained by calculating the relative rate of change between the current response intensity and the response intensity of the previous period. Based on statistics of consistency and confidence fluctuations in the most recent N identifications of this category, a dynamic change index was derived. The ability to detect abrupt changes or trends in discharge activity helps in identifying early faults or rapidly developing defects. Historical stability parameters This reflects the consistency of the identification results and gives higher credibility to categories that appear consistently.

[0016] The present invention is further configured such that the spatial distribution information provided by the ultraviolet sensor array, the computational efficiency of the lightweight multilayer perceptron model, and the comprehensive utilization of dynamic and historical information by the fuzzy comprehensive decision model enable the system to achieve high real-time and high-accuracy partial discharge type identification under complex working conditions. It systematically integrates three core advantages: spatial sensing array, efficient lightweight model, and intelligent time-series decision. The spatial array solves the problem of single information source, the lightweight model solves the problem of embedded real-time computing, and the fuzzy decision solves the problem of uncertainty in single identification. The three work together to enable the entire system to maintain reliable identification performance with high real-time and high accuracy even in complex industrial environments with strong interference and limited resources.

[0017] The present invention has the following beneficial effects: 1. This invention acquires spatial distribution information of partial discharge by arranging an ultraviolet sensor array, overcoming the limitations of single-point detection and enhancing the system's anti-interference capability in complex electromagnetic environments. It automatically extracts discriminative features from spatial intensity relationships using an adaptively optimized feature transformation function, overcoming the shortcomings of traditional methods that rely on human experience, improving the effectiveness of features and the generalization recognition capability of patterns. Combined with preprocessing and data augmentation strategies, it further enhances the system's adaptability to random noise and data stability in actual working conditions.

[0018] 2. This invention employs a lightweight multilayer perceptron model, whose network structure has been simplified and parameter optimized to adapt to the storage and computing power constraints of microcontrollers such as STM32. This enables the direct deployment and efficient operation of the algorithm on resource-constrained embedded platforms, meeting the real-time diagnostic needs of edge devices and avoiding delays caused by data uploads. By integrating current confidence, temporal trends, and historical recognition stability information through a fuzzy comprehensive decision model, the invention performs multi-dimensional verification and comprehensive judgment on single recognition results, improving the system's ability to cope with boundary conditions and abnormal events, thereby achieving reliable recognition under complex working conditions. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below.

[0020] Figure 1 This is a flowchart illustrating an algorithm for a partial discharge identification system based on an ultraviolet sensor array. Figure 2 This is a schematic diagram of the data augmentation and partitioning process in an algorithm for a partial discharge identification system based on an ultraviolet sensor array; Figure 3 This is a schematic diagram of the confidence calculation process in an algorithm for a partial discharge identification system based on an ultraviolet sensor array; Figure 4 This is a schematic diagram of the fuzzy comprehensive decision-making process in an algorithm for a partial discharge identification system based on an ultraviolet sensor array; Figure 5 This is a schematic diagram of the model deployment process in an algorithm for a partial discharge identification system based on an ultraviolet sensor array. Detailed Implementation

[0021] The technical solutions of the present invention will be described below with reference to the accompanying drawings. The described embodiments are only some embodiments of the present invention, and not all embodiments.

[0022] Please see Figures 1-5 This invention relates to an algorithm for a partial discharge identification system based on an ultraviolet sensor array, comprising the following steps: Step a: Arrange nine ultraviolet photodetectors in a 3×3 matrix to form a sensor array. Each sensing unit independently collects the ultraviolet light intensity signal generated by partial discharge, and after analog-to-digital conversion, forms a 9-dimensional original signal vector. Step b: Perform normalization preprocessing on the original signal vector. The calculation formula is as follows: ; in The signal after normalization The original response signal, This is the average response value of the array. Standard deviation, To prevent zero constant; Step c: Extract primary feature vectors based on the signals of each sensing unit and their spatial neighborhood relationships. The calculation formula is as follows: ; in For the first Primary features of the unit To differentiate the signal from that of adjacent units, The mean of the neighborhood signal. , , It is an adaptively determinable feature transformation function; Step d: Transform all primary features into an intermediate feature vector using a nonlinear mapping function, expressed as: ; in For the first One intermediate feature component It is a nonlinear mapping operator; Step e: Use the trained lightweight multilayer perceptron model to classify the intermediate feature vectors, calculate the response intensity of each category and normalize it to obtain the confidence level of the current period; Step f: Combining dynamic change indicators with historical recognition stability, calculate the comprehensive confidence score of each category using a fuzzy comprehensive decision model, and select the category with the highest score as the recognition result output.

[0023] Specifically, by deploying a 3×3 matrix ultraviolet sensor array, the spatial distribution information of partial discharge is acquired synchronously, overcoming the limitations of single-point detection and laying the foundation for subsequent feature extraction and pattern recognition based on spatial relationships. Furthermore, the original signal is preprocessed by normalization to eliminate the influence of sensor differences and environmental fluctuations, improving data stability. Based on this, rich and discriminative spatial-intensity joint primary features are extracted by comprehensively utilizing the unit's own intensity, neighborhood differences, and regional mean. Subsequently, these features are automatically fused and abstracted using a nonlinear mapping function to achieve dimensionality reduction and form a compact intermediate feature vector. A trained lightweight multilayer perceptron model is used to efficiently classify the intermediate features, generating the confidence score for the current period, enabling real-time computation on an embedded platform. Finally, combining dynamic change indicators and historical recognition stability, a fuzzy comprehensive decision model is used to fuse multi-dimensional information, outputting the recognition result with the highest comprehensive confidence. This entire method systematically integrates the three major advantages of spatial perception, efficient lightweight models, and intelligent temporal decision-making, thereby achieving high real-time performance, high accuracy, and robust partial discharge type recognition under complex working conditions.

[0024] Step b also includes augmenting and partitioning the preprocessed data. Specifically, Gaussian white noise with a standard deviation of 0.05 is added to the data using a MATLAB script. 1000 samples are generated for each discharge category, and all samples are randomly divided into a training set (80%) and a test set (20%). Step e calculates the response intensity of each category and normalizes it to obtain the confidence level for the current cycle. Specifically, this includes: Calculate the response intensity for each category: ; in For category numbering, Assign feature weights to each category; normalize the response intensity of each category to obtain the confidence level for the current period: ; in and These represent the minimum and maximum response strengths, respectively. For a small constant, step f calculates the comprehensive confidence score for each category using a fuzzy comprehensive decision model and selects the recognition result, specifically including: The comprehensive confidence score for each category is calculated using a fuzzy comprehensive decision model. ; in As a dynamically changing indicator, For historical stability parameters, , , These are the weighting coefficients, and their sum is 1; The category with the highest overall confidence score is selected as the recognition result output, and a resampling and labeling mechanism is triggered for events with confidence scores below a set threshold. Step f includes the model deployment step: automatically extracting the network weights, biases, batch normalized parameters, and standardized parameters from the trained lightweight multilayer perceptron model, converting them into C language format files suitable for the STM32 system, and simultaneously generating a C language header file containing constant definitions, parameter arrays, and inference function interfaces. The feature transformation function in step c... , , The neighborhood is determined adaptively through backpropagation during the model training phase, and is defined as the 8 units excluding the current unit within a 3×3 region centered on the current unit. The nonlinear mapping operator in step d... It is implemented using fully connected layers based on Sigmoid or ReLU activation functions, and the output dimension of the mapping layer is... Less than the total number of primary features input To achieve feature dimensionality reduction, the lightweight multilayer perceptron model in step e includes an input layer, at least one hidden layer, and an output layer. The number of neurons in the hidden layer does not exceed 32. The overall parameters have been optimized to adapt to the storage and computing power constraints of the STM32 series microcontroller, and the dynamic changing indicators... The historical stability parameter is obtained by calculating the relative rate of change between the current response intensity and the response intensity of the previous period. Based on the statistical analysis of the consistency and confidence fluctuation of this category in the most recent N identifications, it is concluded that the spatial distribution information provided by the ultraviolet sensor array, the computational efficiency of the lightweight multilayer perceptron model, and the comprehensive utilization of dynamic and historical information by the fuzzy comprehensive decision model enable the system to achieve high real-time performance and high accuracy in identifying partial discharge types under complex working conditions.

[0025] Specifically: By adding Gaussian white noise for data augmentation, random disturbances in actual working conditions are effectively simulated, expanding the dataset and improving the model's generalization ability and noise resistance. Standardized data partitioning ensures the scientific rigor and reliability of model training and evaluation. Weighted summation of category response intensity reflects the differences in the contribution of different features to each category. Furthermore, normalization transforms the response intensity into a unified confidence metric, making confidence levels comparable across different categories and periods, providing standardized input for subsequent comprehensive decision-making. The fuzzy comprehensive decision model organically integrates current confidence, dynamic trends, and historical performance, achieving a more comprehensive evaluation. A threshold-triggered resampling mechanism automatically identifies low-confidence events, initiating a review process, effectively reducing false positives and false negatives, and improving the system's ability to handle boundary conditions and abnormal events. Seamless integration from the training environment to the embedded deployment environment is achieved. An automated conversion toolchain avoids errors that may occur during manual porting. The generated standardized C language files facilitate direct integration into STM32 projects, greatly simplifying the deployment process and improving efficiency. To improve development efficiency and system reliability, adaptive optimization of the feature transformation function replaces the traditional design relying on manual experience, enabling extracted features to better serve the final classification goal and improving feature effectiveness. A clear 3×3 neighborhood definition fully utilizes the compact structure of the array, ensuring consistency and integrity in spatial relationship feature calculation. The use of mature and efficient Sigmoid or ReLU activation functions guarantees the expressive power and training stability of nonlinear mappings. Forced dimensionality reduction effectively compresses feature dimensions, significantly reducing subsequent computation and storage overhead while retaining key information, making it particularly suitable for resource-constrained embedded scenarios. A rigorous lightweight design of the network structure ensures that the number of model parameters and computational complexity match the limited memory and computing power of microcontrollers such as STM32, thus truly achieving real-time, online recognition on edge devices, avoiding latency and bandwidth consumption caused by data uploads, and dynamically changing indicators. The ability to detect abrupt changes or trends in discharge activity helps in identifying early faults or rapidly developing defects. Historical stability parameters It reflects the consistency of the recognition results, gives higher credibility to categories that appear continuously and stably, and systematically integrates three core advantages: spatial sensing array, efficient lightweight model and intelligent temporal decision. The spatial array solves the problem of single information source, the lightweight model solves the problem of embedded real-time computing, and fuzzy decision solves the problem of uncertainty in single recognition. The three work together to enable the whole system to maintain reliable recognition performance with high real-time performance and high accuracy even in complex industrial environments with strong interference and tight resources.

[0026] Please see Figures 1-5 Based on Specific Implementation Example 1, a high-voltage switchgear online intelligent monitoring system is developed.

[0027] Specifically, this embodiment is applied to online partial discharge intelligent monitoring of high-voltage switchgear. The system deploys a 3×3 ultraviolet sensor array on the inner side of the switchgear's insulation panel. Each sensing unit in the array is distributed in an equally spaced grid to synchronously capture the ultraviolet light signal emitted by partial discharge within the cabinet. The raw signal is converted from analog to digital to form a 9-dimensional vector. Normalization preprocessing is used to eliminate deviations caused by differences in sensor sensitivity and environmental temperature drift. In the feature extraction stage, the system constructs a primary feature vector based on the signal strength of each sensing unit, the signal gradient difference between it and its eight neighboring units, and the average intensity of its 3×3 neighborhood. The feature transformation function is adaptively optimized through backpropagation during the training stage. Subsequently, a fully connected layer based on the ReLU activation function is used to perform nonlinear fusion and dimensionality reduction on the primary features to obtain intermediate features with stronger discriminative power. A lightweight multilayer perceptron model is deployed on an embedded processor with no more than 32 hidden layer neurons to achieve rapid classification of intermediate features and output the current confidence level of each discharge category. The system also integrates a fuzzy comprehensive decision module to dynamically fuse the current confidence level, the rate of change of signal intensity in adjacent detection cycles, and the stability performance of the category in the past identifications. Finally, it outputs the discharge type with the highest comprehensive confidence level and can report to the monitoring center in real time through the communication module to realize continuous assessment and early warning of the insulation status of the switchgear.

[0028] Please see Figures 1-5 Based on Specific Embodiment 1, a portable cable terminal inspection instrument.

[0029] Specifically, this embodiment addresses the partial discharge inspection needs of power cable terminals by designing a portable identification instrument. The instrument integrates a flexibly mountable 3×3 ultraviolet sensor array at its front end. The array utilizes a flexible circuit board to adapt to installation on different curved surfaces. The acquired ultraviolet signals are normalized and enhanced with Gaussian white noise injection before being input into an embedded identification algorithm. Primary feature extraction fully leverages the spatial topology of the array, processing unit-independent signals, neighborhood difference signals, and regional average signals through pre-trained feature transformation functions. Intermediate feature mapping uses a Sigmoid activation function to ensure stable output values. Dimensionality reduction design decreases computational burden. The lightweight multilayer perceptron model undergoes pruning and quantization, significantly compressing the overall parameter count and enabling millisecond-level inference on a low-power microcontroller. The instrument incorporates fuzzy decision logic, performing a comprehensive score based on real-time signal dynamic changes and historical identification consistency. It supports on-site identification of typical discharge modes such as corona discharge and surface discharge. Identification results and confidence levels are displayed intuitively on the instrument screen and can be transmitted to mobile terminals via Bluetooth, facilitating rapid diagnosis and recording by inspection personnel.

[0030] Please see Figures 1-5 Based on Specific Embodiment 1, a transformer multi-physics field coupled discharge identification system.

[0031] Specifically, this embodiment is applicable to the composite monitoring scenario of partial discharge inside oil-immersed transformers. The system combines an ultraviolet array with vibration and ultrasonic sensors, but uses the ultraviolet array as the primary identification method. The ultraviolet sensor array is arranged near the transformer's observation window to capture ultraviolet photons generated by discharge in the oil. In the signal preprocessing stage, sliding window normalization is used to adapt to background light fluctuations caused by changes in transformer load. In the feature extraction stage, in addition to extracting the spatial features of each unit signal, the system also introduces the signal change rate in the time dimension as a supplementary feature to form a spatiotemporal joint primary feature vector. The intermediate feature mapping layer adopts a multi-branch fully connected structure to extract feature representations at different abstract levels. The final intermediate features are obtained through feature splicing and compression. The lightweight perceptron model classifies the features based on this and outputs the confidence level of each discharge type. The fuzzy decision module introduces the environmental noise level as an adjustment factor to dynamically adjust the weight coefficients of the current confidence level and historical stability, thereby improving the robustness of identification in the complex electromagnetic and vibration interference environment of the transformer and effectively distinguishing various insulation defect modes such as oil bubble discharge and floating discharge.

[0032] Please see Figures 1-5 Based on Specific Embodiment 1, an edge device for detecting electric arcs in a distributed photovoltaic power station.

[0033] Specifically, this embodiment is applicable to arc fault detection on the DC side of distributed photovoltaic power stations. The system adapts the ultraviolet array partial discharge identification algorithm to the detection and classification of arc light. The sensor array uses a high-sensitivity ultraviolet photodetector, which is placed in the photovoltaic combiner box or key nodes of the inverter to capture the ultraviolet spectral signal emitted by potential arcs. To adapt to outdoor light fluctuations, the system introduces a background light dynamic compensation mechanism in the preprocessing stage to enhance the effective components of the signal. Feature extraction focuses on the spatial diffusion mode and transient intensity distribution of arc light on the array. The feature transformation function is trained through an arc sample library. The intermediate feature mapping adopts a binary weight design, which greatly reduces the computational complexity and storage requirements. The lightweight perceptron model is extremely pruned and can run on an ultra-low power edge processor, meeting the strict energy consumption requirements of photovoltaic scenarios. The fuzzy decision module combines the arc duration, intensity change rate and consistency with historical arc events to make a comprehensive judgment, realizing the identification and alarm of different fault types such as series arcs and parallel arcs. The device supports wireless networking, and the identification results can be uploaded to the power station monitoring platform in real time, providing key information for the safe operation of the photovoltaic system.

[0034] The preferred embodiments of the present invention disclosed above are only for the purpose of illustrating the present invention. The preferred embodiments do not describe all the details in detail, nor do they limit the invention to the specific implementation described herein. This specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the present invention, so that those skilled in the art can better understand and utilize the present invention.

Claims

1. An algorithm for partial discharge identification system based on ultraviolet sensor array, characterized in that: Includes the following steps: Step a: Arrange nine ultraviolet photodetectors in a 3×3 matrix to form a sensor array. Each sensing unit independently collects the ultraviolet light intensity signal generated by partial discharge, and after analog-to-digital conversion, forms a 9-dimensional original signal vector. Step b: Perform normalization preprocessing on the original signal vector, calculated using the following formula: ； in The signal after normalization The original response signal, This is the average response value of the array. Standard deviation To prevent zero constant; Step c: Extract primary feature vectors based on the signals of each sensing unit and their spatial neighborhood relationships. The calculation formula is as follows: ； in For the first Primary features of the unit To differentiate the signal from that of adjacent units, The mean of the neighborhood signal. , , It is an adaptively determinable feature transformation function; Step d: Transform all primary features into an intermediate feature vector using a nonlinear mapping function, expressed as follows: ； in For the first One intermediate feature component It is a nonlinear mapping operator; Step e: Use the trained lightweight multilayer perceptron model to classify the intermediate feature vectors, calculate the response intensity of each category and normalize it to obtain the confidence level of the current period; Step f: Combining dynamic change indicators with historical recognition stability, calculate the comprehensive confidence score of each category using a fuzzy comprehensive decision model, and select the category with the highest score as the recognition result output.

2. The algorithm for a partial discharge identification system based on an ultraviolet sensor array according to claim 1, characterized in that: Step b also includes enhancing and dividing the preprocessed data. Specifically, Gaussian white noise with a standard deviation of 0.05 is added to the data using a MATLAB script. 1,000 samples are generated for each discharge category, and all samples are randomly divided into a training set accounting for 80% and a test set accounting for 20%.

3. The algorithm for a partial discharge identification system based on an ultraviolet sensor array according to claim 1, characterized in that: Step e involves calculating and normalizing the response intensity of each category to obtain the confidence level for the current period, specifically including: Calculate the response intensity of each category ； in For category numbering, The feature weights corresponding to this category are assigned; the response intensities of each category are normalized to obtain the confidence level for the current period. ； in and These represent the minimum and maximum response strengths, respectively. It is a tiny constant.

4. The algorithm for a partial discharge identification system based on an ultraviolet sensor array according to claim 1, characterized in that: Step f, which involves calculating the comprehensive confidence score for each category using a fuzzy comprehensive decision model and selecting the identification results, specifically includes: The comprehensive confidence score for each category is calculated using a fuzzy comprehensive decision model. ； in As a dynamically changing indicator, For historical stability parameters, , , These are the weighting coefficients, and their sum is 1; The category with the highest overall confidence score is selected as the recognition result output, and a resampling and labeling mechanism is triggered for events with confidence scores below a set threshold.

5. The algorithm for a partial discharge identification system based on an ultraviolet sensor array according to claim 1, characterized in that: Step f includes a model deployment step: automatically extracting the network weights, biases, batch normalized parameters, and standardized parameters from the trained lightweight multilayer perceptron model, converting them into C language format files suitable for the STM32 system, and simultaneously generating a C language header file containing constant definitions, parameter arrays, and inference function interfaces.

6. The algorithm for a partial discharge identification system based on an ultraviolet sensor array according to claim 1, characterized in that: Feature transformation function in step c , , The neighborhood range is determined adaptively through backpropagation during the model training phase, and is defined as the eight units other than the current unit within a 3×3 region centered on the current unit.

7. The algorithm for a partial discharge identification system based on an ultraviolet sensor array according to claim 1, characterized in that: The nonlinear mapping operator in step d It is implemented using fully connected layers based on Sigmoid or ReLU activation functions, and the output dimension of the mapping layer is... Less than the total number of primary features input This is to achieve feature dimensionality reduction.

8. The algorithm for a partial discharge identification system based on an ultraviolet sensor array according to claim 1, characterized in that: The lightweight multilayer perceptron model in step e includes an input layer, at least one hidden layer, and an output layer. The number of neurons in the hidden layer does not exceed 32, and the overall parameters are optimized to adapt to the storage and computing power constraints of the STM32 series microcontroller.

9. The algorithm for a partial discharge identification system based on an ultraviolet sensor array according to claim 4, characterized in that: The dynamic change index The historical stability parameter is obtained by calculating the relative rate of change between the current response intensity and the response intensity of the previous period. This is based on statistics of the consistency and confidence fluctuation of this category in the most recent N identifications.

10. The algorithm for a partial discharge identification system based on an ultraviolet sensor array according to claim 1, characterized in that: The system, with its spatial distribution information provided by the ultraviolet sensor array, computational efficiency of the lightweight multilayer perceptron model, and comprehensive utilization of dynamic and historical information by the fuzzy comprehensive decision model, can achieve high real-time performance and high accuracy in identifying partial discharge types under complex operating conditions.

Images