Power distribution cabinet switch intelligent identification system and method

Abstract

The invention relates to a power distribution cabinet switch intelligent identification system and method, and the system comprises a data collection unit, a data processing unit, a feature extraction unit, a state classification unit, a test optimization unit, a multi-modal data fusion and analysis unit, and a system control unit. Generating a sub-image containing complete features in combination with a cutting extension algorithm; the feature extraction unit adopts a multi-scale convolution module and a Vision Transform to extract local and global features, and optimizes feature expression through an SE attention mechanism; the state classification unit is combined with a support vector machine (SVM) classifier and Focal Loss to accurately classify normal, closed and tripping states; the test optimization unit is used for enhancing the generalization ability of the model by utilizing Test Time Training (TTT) and self-supervised learning; and the multi-modal data fusion unit is used for modeling an interaction relationship between the image and the sensor data through a cross-modal Transform.

Description

Technical Field

[0001] The present invention relates to an intelligent switch identification system for a power distribution cabinet based on target detection and an improved triplet selection strategy, and in particular to an intelligent switch identification system and method for a power distribution cabinet applied in the technical field of intelligent power distribution switches. Background Art

[0002] With the rapid development of intelligent power systems, real-time monitoring and intelligent management of the switch status of distribution cabinets, as the core equipment for power distribution, have become an important part of ensuring the safe operation of the power grid. Traditional methods for identifying the switch status of distribution cabinets mainly rely on manual inspections and simple sensor monitoring, which have the following problems: low detection efficiency: manual inspection cycles are long and inefficient, and it is difficult to detect abnormal states of distribution cabinets in a timely manner; insufficient intelligence level: simple sensor monitoring methods are easily affected by environmental interference and are difficult to cope with state recognition needs in complex scenarios.

[0003] In recent years, intelligent monitoring methods based on image recognition have been gradually applied to switch status monitoring of distribution cabinets. The switch images are collected in real time by cameras, and deep learning models (such as the YOLO series) are used for status recognition. However, the existing technology still faces the following major problems in practical applications: Complex multi-switch monitoring: The distribution cabinet usually contains multiple switches, and the status of each switch needs to be monitored at the same time, which increases the complexity of recognition and the computational burden. Existing methods, such as directly using YOLO for multi-category detection, result in high model complexity and difficulty in meeting real-time requirements. Data imbalance problem: In the switch status of the distribution cabinet, the number of samples in the normal state is far greater than the number of samples in the abnormal state (such as tripping), resulting in insufficient recognition ability of the model for a minority class during training. High labeling cost: Switch status recognition requires a large amount of labeled data, and there are fewer samples in the abnormal state. The labeling workload is large and the cost is high. Directly using YOLO for multi-category detection requires detailed labeling of each switch state, which increases the difficulty and cost of labeling.

[0004] Difficulty in feature extraction: The differences between switch states are small, especially under different lighting and angles, the appearance of the switch varies greatly, which increases the difficulty of feature extraction and recognition. Existing methods perform poorly in fine-grained classification and are prone to misjudgment.

[0005] In the prior art, although there are some switch state recognition methods based on image processing and sensor data fusion, there are still some limitations, such as:

[0006] Chinese invention patent CN116740472B discloses a method for identifying the switch state of an electric switch cabinet. The method realizes the identification of the target switch state by fusing image data and current data, thereby improving the accuracy of state monitoring. However, the technology has the following shortcomings: Simple fusion method: Only a preset fusion algorithm is used, and the correlation characteristics between image and sensor data are not fully utilized, resulting in limited data fusion effect; Lack of deep learning modeling: Deep learning models are not used to jointly model multimodal data, making it difficult to handle interference and abnormal situations in complex environments.

[0007] Chinese invention patent CN113537224B proposes a vision-based method for automatic recognition of the switch status of a distribution cabinet. It uses boundary feature points and perspective projection transformation to correct the image, and combines morphological filtering to extract the geometric features of the switch to achieve recognition of the switch status. However, this method still has the following limitations: Reliance on image features: Switch status recognition is only based on two-dimensional visual features, and other sensor data is not combined to analyze abnormal conditions, which limits its applicability; Lack of intelligent prediction capabilities: It is impossible to predict potential abnormal switch conditions based on historical data, and it is difficult to meet the needs of intelligent preventive maintenance.

[0008] Although the image recognition method based on deep learning has made certain progress in the switch status monitoring of distribution cabinets, it still has the following shortcomings: Data imbalance: Most existing methods fail to effectively deal with the data imbalance problem during the training process, resulting in low accuracy of the model in recognizing abnormal states; High computational complexity: In order to achieve high accuracy, the existing models are usually designed to be more complex, resulting in large consumption of computing resources and difficulty in meeting the needs of real-time monitoring; Poor generalization ability: The model performs well on the training data set, but in actual applications, due to environmental changes (such as lighting, angle, etc.), the generalization ability of the model is limited, the recognition effect is unstable, and the labeling requirements are high: Existing methods rely on a large amount of labeled data for supervised learning, especially when abnormal state samples are scarce, the labeling cost is high and data acquisition is difficult; The contradiction between real-time and recognition accuracy: In the existing technology, some methods, such as directly using YOLO for switch state classification, can achieve a certain degree of real-time performance, but there are deficiencies in the accuracy of switch state recognition, especially when the state changes are subtle, misjudgment is prone to occur. Summary of the invention

[0009] In view of the above-mentioned prior art, the technical problem to be solved by the present invention is how to achieve efficient detection and feature extraction of distribution cabinet switches to effectively deal with data imbalance, intelligent multi-modal fusion precise classification and generalization optimization, predictive maintenance and intelligent early warning.

[0010] To solve the above problems, the present invention provides a switch intelligent identification system and method for a power distribution cabinet, comprising a data acquisition unit, a data processing unit, a feature extraction unit, a state classification unit, a test optimization unit, a data fusion and analysis unit, and a system control unit;

[0011] The data acquisition unit includes at least one high-resolution camera for collecting real-time image data of the switch of the power distribution cabinet, and the image covers a variety of lighting conditions and angles; at least one multimodal sensor for collecting environmental parameters of the power distribution cabinet, including temperature, current and humidity;

[0012] The data processing unit calls the pre-trained YOLOv5 target detection model and configures it to detect the position of the switch of the power distribution cabinet in real time; according to the detection result, the switch area is expanded and cropped to generate a sub-image containing the switch image features;

[0013] The feature extraction unit includes a multi-scale convolution module, which is used to extract local detail features by combining 3×3, 5×5 and 7×7 convolution kernels; Vision Transformer is called to extract global context features of the switch image through position encoding and multi-head self-attention mechanism; SE attention mechanism is used to weight feature channels, strengthen the expression of important features, and suppress redundant features;

[0014] The state classification unit is configured with a support vector machine (SVM) classifier to classify the extracted feature vectors; combined with a category balancing mechanism (such as Focal Loss), a higher weight is given to abnormal state samples; the state of the output switch includes normal, closed, and tripped;

[0015] The test optimization unit supports dynamic optimization based on Test Time Training (TTT) and adjusts model parameters in real time. It also combines self-supervised learning tasks (such as image reconstruction) to enhance the model's adaptability to test data distribution.

[0016] The data fusion and analysis unit calls the pre-trained cross-modal Transformer model to fuse image data and environmental sensor data; analyzes historical time series data and calls the LSTM or time Transformer model to predict the probability of future abnormal states;

[0017] The system control unit is used to coordinate data transmission and functional operations between units; and provide system task allocation and status monitoring functions.

[0018] In the above-mentioned intelligent distribution cabinet switch recognition system and method based on target detection and improved triplet selection strategy, a pre-trained YOLOv5 target detection model is used, combined with a dynamic reasoning optimization algorithm, to achieve rapid detection of switch positions under multi-lighting and complex environments; the detection boundary is dynamically adjusted through a cropping and expansion algorithm to ensure that the cropped sub-image contains complete switch features, improve the stability of detection and extraction, introduce a cross-modal Transformer model, and use a multi-head attention mechanism to dynamically model image features and sensor data (such as temperature, current, humidity, etc.) to enhance the robustness of data fusion; the weight is dynamically adjusted according to the correlation between the modal features and the abnormal state to improve the sensitivity of abnormal state recognition, and the state classification unit is combined with a support vector machine (SVM) classifier and a category balancing mechanism (Focal Loss), solve the problem of data imbalance, and significantly improve the classification accuracy of minority class states (such as tripping); adjust the model parameters in real time through the test optimization unit (TTT), combine self-supervised learning tasks (such as image reconstruction) to enhance the generalization ability of the model in a dynamic environment, use the time series analysis module (LSTM or time Transformer) to perform dynamic time window modeling on the historical data of the switch state, and combine the real-time environmental data to predict possible abnormal states (such as tripping risks) in the future; the system control unit supports real-time alarm and remote feedback to improve operation and maintenance efficiency. The present invention can be widely used in distribution cabinet monitoring, industrial control cabinet status recognition and other scenarios in the power system. Through the innovative application of deep learning models and multimodal data fusion technology, the present invention significantly improves the recognition accuracy and operation reliability in complex environments, reduces the cost of manual inspections, and provides effective support for the intelligent operation and maintenance of power systems.

[0019] As a further improvement of the present application, the data acquisition unit includes: a high-resolution camera with a resolution of 1080p or above, supporting low-light environment imaging; a temperature sensor, a current sensor and a humidity sensor for collecting environmental data.

[0020] As a further improvement of the present application, the YOLOv5 model called by the data processing unit has the following configuration: using a dynamic inference optimization algorithm to detect the switch position in real time; the detection results are processed by a cropping and expansion algorithm to ensure that the cropped area contains complete switch image features; the cropped image is normalized to improve the accuracy of subsequent feature extraction.

[0021] As a further improvement of the present application, the feature extraction unit includes: a multi-scale convolution module, which extracts fine-grained features of the switch image through convolution kernels of different sizes; a Vision Transformer module, which combines position encoding and multi-head self-attention mechanism for global modeling; and an SE attention mechanism, which is used to re-weight feature channels and improve the expression of important features.

[0022] As another improvement of the present application, the state classification unit combines the category balancing mechanism and adopts the following strategy: giving higher weights to abnormal state samples (such as tripping state) based on Focal Loss; calling the support vector machine (SVM) classifier to classify the features and output normal, closed and tripping states.

[0023] As another improved supplement to the present application, the test optimization unit supports functions based on Test Time Training (TTT), including: dynamically optimizing model parameters according to the real-time input data distribution; combining self-supervised tasks (such as feature reconstruction) to enhance the model generalization ability and maintain high accuracy under lighting changes or switch position offsets.

[0024] A method for intelligently identifying switches in a power distribution cabinet comprises the following steps:

[0025] S1. Data collection: Use a high-resolution camera to collect images of distribution cabinet switches; use multimodal sensors to collect environmental data, including temperature, current, and humidity;

[0026] S2, object detection and image cropping: Call the YOLOv5 model to detect the switch position; crop the detected switch area to generate a switch sub-image;

[0027] S3, feature extraction: use multi-scale convolution module to extract local features; call Vision Transformer to extract global features, and combine SE attention mechanism to optimize feature expression;

[0028] S4, state classification: call the support vector machine (SVM) classifier to classify the extracted features; output the state of the switch, including normal, closed and tripped;

[0029] S5, test optimization: adjust model parameters in real time based on TTT, and enhance model adaptability by combining self-supervision tasks;

[0030] S6. Multimodal data fusion and time series analysis: Fusion of image data and environmental data, calling cross-modal Transformer to model multimodal features; calling LSTM or time Transformer to analyze historical data and predict future abnormal conditions.

[0031] As another improved supplement of the present application, the target detection and image cropping steps include: a YOLOv5 model combined with a dynamic reasoning algorithm to detect switches in real time; and a cropping extension algorithm is used to ensure that the sub-image contains complete features.

[0032] As another improvement of the present application, in the multimodal data fusion and time series analysis steps: the cross-modal Transformer performs multi-head attention modeling on image features and environmental data; and the LSTM or time Transformer predicts the probability of occurrence of future abnormal states.

[0033] In summary, this application has the following beneficial effects:

[0034] 1. By introducing the pre-trained YOLOv5 model and dynamic inference optimization algorithm, efficient positioning of the distribution cabinet switch is achieved. YOLOv5 can quickly and accurately detect the switch position in multiple lighting, complex background and multi-angle environments. The cropping and expansion algorithm dynamically adjusts the cropping boundary to ensure that the generated sub-image fully contains the switch features while suppressing background noise.

[0035] In the feature extraction stage, the multi-scale convolution module captures the fine-grained features (such as edges, textures, etc.) in the switch image and enhances the representation of local features. The Vision Transformer effectively models the global features of the image through position encoding and multi-head self-attention mechanism, improving the model's recognition ability in complex scenes. The SE attention mechanism strengthens the expression of important features and reduces redundant information interference.

[0036] 2. Support vector machine (SVM) classifier combined with category balancing mechanism (Focal Loss): It effectively solves the problem of data imbalance, gives higher classification weights to minority classes (such as tripping status), significantly improves classification accuracy, and achieves efficient distinction between the three states of normal, closed and tripping, ensuring the accuracy and reliability of the recognition results.

[0037] 3. Through the test optimization unit (Test Time Training, TTT), dynamic adjustment of model parameters is achieved, so that the system can adapt to the distribution changes of real-time input data. Combined with self-supervised learning tasks (such as image reconstruction), the robustness of the system in complex environments such as lighting changes, switch position offsets and sensor noise is enhanced. The dynamic optimization and adaptability greatly improve the stability of the system and meet the needs of long-term operation of the distribution cabinet in actual applications.

[0038] 4. The present invention integrates image features with environmental sensor data (such as temperature, current, humidity, etc.). The cross-modal Transformer uses a multi-head attention mechanism to model the interaction between multimodal features and dynamically adjusts the weights of different modalities to ensure the rationality and efficiency of feature fusion. The dynamic weight allocation rule of modal features further enhances the sensitivity of the system in abnormal state recognition.

[0039] This multimodal fusion method enables deep perception and accurate judgment of potential abnormal conditions in complex environments.

[0040] 5. Through the time series analysis module (LSTM or time Transformer), the historical data of the switch status is dynamically modeled in a time window, and the abnormal status that may occur in the future (such as tripping risk) is predicted in combination with the real-time environmental data. The dynamic adjustment mechanism of the time series analysis can accurately capture the fluctuation characteristics of key data, greatly improving the prediction accuracy. The predictive maintenance function provides strong support for the active monitoring and preventive maintenance of the power system.

[0041] 6. The introduction of vibration sensors and sound sensors enables the system to have stronger fault detection capabilities. It can identify abnormal signals such as mechanical vibration or arc discharge noise, and realize real-time alarm of abnormal conditions through alarm lights, buzzers or SMS notifications. Combined with the remote monitoring function, it improves the overall management efficiency of the system, improves the flexibility and adaptability of the system, and meets the actual application needs in complex scenarios.

[0042] 7. Through the optimized cropping and expansion algorithm, feature extraction module and multimodal fusion strategy, redundant calculations are significantly reduced, resource usage and system energy consumption are reduced, and the use of pre-trained large models (YOLOv5, Vision Transformer, cross-modal Transformer, etc.) avoids model training from scratch, shortens the development cycle and reduces deployment costs. The system's automated monitoring and abnormal state recognition capabilities reduce the frequency of manual inspection and maintenance, and significantly reduce operation and maintenance costs.

[0043] 8. In the field of switch recognition in distribution cabinets, a cross-modal Transformer is used to achieve deep fusion of image and sensor data. The dynamically adjusted cropping and expansion algorithm and triplet selection strategy optimize the performance of detection and classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 System interaction for this application Figure 1 ;

[0045] Figure 2 System interaction for this application Figure 2 ;

[0046] Figure 3 System interaction for this application Figure 3 . DETAILED DESCRIPTION

[0047] Three embodiments of the present application are described in detail below with reference to the accompanying drawings.

[0048] Embodiment 1:

[0049] like Figures 1 to 3 As shown, this embodiment provides a system and method for intelligent identification of switches in a power distribution cabinet, including the following functional modules:

[0050] Hardware configuration of the data acquisition unit:

[0051] Use at least one high-resolution camera (resolution not less than 1080p) to support imaging in multiple lighting environments and low-light environments, and collect image data of switches inside the distribution cabinet in real time to ensure coverage of multi-angle views.

[0052] Configure temperature sensors, current sensors, and humidity sensors to synchronously collect environmental data.

[0053] Acquisition scope: The image data covers multiple states of the switch (such as normal, closed, and tripped), and captures the possible impact of external environmental changes (such as high temperature or high humidity) on the switch state.

[0054] In addition to temperature sensors, current sensors and humidity sensors, the data acquisition unit can also be equipped with vibration sensors and sound monitoring equipment:

[0055] Vibration sensor: used to detect abnormal vibration of mechanical parts or switches inside the distribution cabinet and identify possible mechanical failures.

[0056] Acoustic Sensors: Used to monitor audio characteristics of switch operation, such as abnormal noise or arcing sounds, to provide additional diagnostic information.

[0057] Extended scenarios: When multiple environmental parameters are analyzed in combination, complex faults that cannot be identified by a single mode can be discovered, such as the abnormal audio characteristics of arc discharge in a high-humidity environment.

[0058] Data processing unit

[0059] YOLOv5 input preprocessing, add the following preprocessing steps before image input:

[0060] Brightness adjustment: Use histogram equalization technology to balance image brightness and improve switch recognition accuracy in scenes with uneven lighting.

[0061] Contrast correction: The automatic enhancement algorithm optimizes contrast and reduces the impact of low-contrast images on target detection.

[0062] Data enhancement: including random rotation, scaling, and flipping methods to expand the input data distribution of the YOLOv5 model and enhance its robustness.

[0063] YOLOv5 target detection model: Call the pre-trained YOLOv5 model and combine it with the dynamic reasoning optimization algorithm to achieve real-time detection of switch positions.

[0064] The detection results are processed through a cropping and expansion algorithm to ensure that the switch sub-image completely contains the switch features and the surrounding key areas.

[0065] Image normalization: Normalize the pixel values ​​of the cropped switch sub-image to eliminate the impact of brightness or contrast differences on subsequent feature extraction.

[0066] Feature Extraction Unit

[0067] Multi-scale convolution module: Use 3×3, 5×5 and 7×7 convolution kernels to extract fine-grained features of different receptive fields respectively. Vision Transformer: Based on the pre-trained Vision Transformer model, position encoding is used to represent the spatial distribution of switch images, and the multi-head self-attention mechanism is combined to capture the global context information of the switch state.

[0068] SE attention mechanism: weights the feature channels extracted by convolution and Transformer to strengthen key features and suppress background interference.

[0069] State Classification Unit

[0070] Classifier configuration: Call the pre-trained support vector machine (SVM) classifier to classify the extracted feature vectors and output three states of the switch: normal, closed, and tripped.

[0071] In addition to the support vector machine (SVM) classifier, other classification algorithms (such as random forest, k-nearest neighbor, etc.) can also be used to classify the switch state. The most appropriate classifier is selected according to the specific application scenario and data distribution to improve the classification performance.

[0072] Category balancing mechanism: Focal Loss is introduced to optimize the classifier, giving higher weights to minority class samples (such as tripping status) to improve classification accuracy.

[0073] Test Optimization Unit

[0074] Dynamic optimization: Use the Test Time Training (TTT) method to dynamically adjust model parameters in combination with real-time input data distribution to improve the adaptability of the model.

[0075] Self-supervised learning: Use image reconstruction tasks to enhance the model's generalization ability and ensure high recognition rates under conditions such as lighting changes and switch position shifts.

[0076] Data fusion and analysis unit

[0077] Multimodal fusion: Use a cross-modal Transformer to fuse image data with environmental sensor (such as temperature, current) data, and model the relationship between the two through a multi-head attention mechanism.

[0078] Time series analysis: Call the LSTM or time Transformer model to analyze the historical data of switch status and predict possible abnormal status in the future.

[0079] System control unit

[0080] The system control unit is responsible for coordinating the interaction of modules such as the data acquisition unit, data processing unit, and feature extraction unit, and provides real-time task allocation, status monitoring, and abnormality reporting functions.

[0081] The system control unit includes abnormal status alarm and feedback functions:

[0082] Alarm mechanism: When the status classification unit identifies an abnormal state (such as tripping), the system control unit issues a real-time alarm through an alarm light, buzzer or SMS notification.

[0083] Feedback signal: The system supports docking with the power distribution management platform to upload classification results and sensor data in real time for remote monitoring and maintenance decision-making

[0084] Working process

[0085] Data collection phase:

[0086] The high-resolution camera collects real-time images of the distribution cabinet switches, the sensors synchronously record the temperature, current and humidity parameters, and the data acquisition unit transmits the images and sensor data to the data processing unit.

[0087] Object detection and cropping stage:

[0088] The data processing unit calls the YOLOv5 model to detect the switch position in real time, and the cropping and expansion algorithm generates a sub-image containing the switch features and performs normalization.

[0089] Feature extraction stage:

[0090] The feature extraction unit uses a multi-scale convolution module to extract local features, calls Vision Transformer to obtain global features, and uses the SE attention mechanism to optimize feature expression.

[0091] Status classification stage:

[0092] The state classification unit calls the SVM classifier to classify the switch state according to the feature vector.

[0093] The Focal Loss mechanism is used to improve the classification priority of abnormal status (tripping).

[0094] Test optimization phase:

[0095] The test optimization unit dynamically adjusts the model through TTT and combines self-supervision tasks to enhance generalization capabilities.

[0096] Data fusion and prediction stage:

[0097] Cross-modal Transformer to fuse image and sensor data.

[0098] The time series analysis model predicts future abnormal conditions based on historical data.

[0099] System control and task allocation phase:

[0100] The system control unit coordinates the data interaction between modules in real time, monitors the system operation status, and outputs recognition results or abnormal alarms.

[0101] Based on the object detection and dynamic reasoning optimization of the YOLOv5 model, fast detection and accurate cropping of switch positions are achieved.

[0102] The entire process of data collection and processing is automated to support real-time monitoring.

[0103] Combining multi-scale convolution, Vision Transformer and SE attention mechanism, the recognition accuracy of switch states in complex scenarios is significantly improved.

[0104] The Focal Loss mechanism enhances the detection capability of abnormal conditions (such as tripping).

[0105] Through TTT dynamic optimization and self-supervised learning, the system's robustness under various conditions such as lighting, angle, and environmental changes is improved.

[0106] Data fusion and time series analysis support predictive maintenance in complex application scenarios.

[0107] Multimodal perception and intelligent decision-making:

[0108] The cross-modal Transformer fuses images and environmental data to improve the system's understanding of potential anomalies.

[0109] Combined with LSTM or temporal Transformer, it can predict abnormal conditions and provide decision support.

[0110] Through the above embodiments, the system realizes efficient, accurate and robust identification of the switch status of the distribution cabinet, and provides technical support for the intelligent monitoring and management of the power system.

[0111] Embodiment 2:

[0112] like Figures 1 to 3 As shown, this embodiment provides a method for intelligently identifying switches in a distribution cabinet, which specifically includes the following steps:

[0113] Step S1: Data collection

[0114] Hardware configuration:

[0115] Use a high-resolution camera (resolution 1080p and above) to collect image data of switches inside the distribution cabinet. The camera supports multi-angle imaging, covering different switch states (normal, closed and tripped) and complex lighting conditions.

[0116] Equipped with temperature sensors, current sensors and humidity sensors to synchronously collect environmental data of the distribution cabinet.

[0117] Data content:

[0118] The image data contains the appearance features of the switch and the surrounding area, and the sensor data reflects the dynamic changes of environmental parameters.

[0119] Step S2: Object detection and image cropping

[0120] YOLOv5 model call:

[0121] Call the pre-trained YOLOv5 target detection model and detect the switch position in real time through the dynamic inference algorithm.

[0122] During the detection process, the model's multi-scale feature extraction capability is used to accurately locate the various states of the switch (such as normal, closed, and tripped).

[0123] Cropping and expansion: Use the cropping and expansion algorithm to appropriately expand the boundary area of ​​the detection result to ensure that the generated sub-image contains the complete switch features and their surrounding areas.

[0124] Normalization: Normalize the cropped sub-image to unify the brightness and contrast, and improve the stability of subsequent feature extraction.

[0125] Step S3: Feature extraction

[0126] Local feature extraction: A multi-scale convolution module is used to extract fine-grained local features in sub-images through 3×3, 5×5, and 7×7 convolution kernels to capture the texture, shape, and edge information of the switch.

[0127] Global feature modeling: Call the pre-trained Vision Transformer model to model the spatial distribution information of sub-images through position encoding.

[0128] The multi-head self-attention mechanism is combined to extract the global context information of switch features, improving the model's ability to distinguish the switch state in complex backgrounds.

[0129] Feature channel optimization: Use the SE attention mechanism to weight the extracted feature channels, strengthen the expression of key features, and suppress redundant or irrelevant features.

[0130] Input sequence generation: The cropped switch sub-image is split into fixed-size image blocks, and each image block is linearly expanded into a feature vector.

[0131] A position code is added to each image block to indicate its spatial position information in the sub-image.

[0132] Multi-head self-attention mechanism: Calculate the query, key and value matrices for each image block to capture the global contextual relationship between blocks; optimize global feature modeling through weighted normalized attention distribution.

[0133] Step S4: Status classification

[0134] Classifier call: call the support vector machine (SVM) classifier to classify the state of the switch according to the extracted feature vector.

[0135] Category balancing mechanism: Use the Focal Loss mechanism to give higher weights to minority class samples (such as tripping status) to solve the problem of data imbalance.

[0136] Output results: The classification results include three states: normal, closed and tripped.

[0137] Step S5: Test optimization

[0138] Real-time adjustment: Based on the Test Time Training (TTT) method, model parameters are adjusted in real time to adapt to changes in the distribution of input data.

[0139] Self-supervised tasks: Optimize the model’s generalization ability for unknown environmental data through image reconstruction tasks, and enhance the model’s robustness under conditions of lighting and angle changes.

[0140] Step S6: Multimodal data fusion and time series analysis

[0141] Multimodal modeling: Call the pre-trained cross-modal Transformer to fuse image features and environmental sensor data through a multi-head attention mechanism to establish associations between features.

[0142] The fusion process of multimodal features is as follows:

[0143] Input encoding: Image features are extracted through Vision Transformer; environmental sensor data (such as temperature, current) are encoded into vectors through linear layers.

[0144] Multi-head attention modeling: A multi-head attention mechanism is used to capture the interactive relationship between image features and sensor features; the correlation matrix is ​​used to dynamically adjust the weight of each modality feature.

[0145] Time series prediction: Use LSTM or time transformer to model historical data and predict the future abnormal state of the switch (such as possible tripping risk). Optimize the time series analysis process of LSTM or time transformer:

[0146] Dynamic window rules: Adaptively adjust the time window size according to the volatility of historical data (such as temperature and current change rate); use a larger time window for stable data and a smaller window for fluctuating data to capture key dynamic features.

[0147] Negative sample difficulty score: The difficulty score of the negative sample is determined by calculating the feature Euclidean distance between the negative sample and the anchor sample;

[0148] Use the following formula:

[0149]

[0150] Among them, the category sample weight is determined by the inverse value of the number of category samples in the data set.

[0151] Constructing balanced sample pairs

[0152] Selection of anchor samples: In each training batch, we ensure that anchor samples are evenly selected from all categories, including samples of normal states and abnormal states (such as tripping), so that samples of each category can be fully learned by the model.

[0153] Select positive samples

[0154] Selection of positive samples: For each anchor sample, a positive sample is randomly selected from the sample set of the same category as the anchor sample. Since there are fewer samples of abnormal states, we use data augmentation (such as rotation, flipping, brightness adjustment, etc.) to increase the number of samples of abnormal states and enrich the diversity of positive samples.

[0155] Selecting difficult negative samples

[0156] Difficulty score calculation: A difficulty score is calculated for each negative sample, which comprehensively considers the distance between the negative sample and the anchor sample in the feature space and the number of samples in the category to which the negative sample belongs.

[0157] Distance factor: The characteristic distance between the negative sample and the anchor sample. The closer the distance, the higher the difficulty.

[0158] Class balancing factor: Give higher weight to negative samples of minority classes (such as the “trip” state) to improve the model’s ability to distinguish minority classes.

[0159] Comprehensive difficulty score: Difficulty score = distance factor × category balance factor.

[0160] Distance factor: Calculate the feature distance between all negative samples and anchor samples to get the distance value d i .

[0161] Category balancing factor: For each negative sample, the number of samples N of the category to which it belongs is i , calculate the category balance factor

[0162] Difficulty Rating: The smaller the distance and the fewer the category samples, the higher the difficulty score.

[0163] Dynamic Difficulty Threshold Adjustment:

[0164] In the early stage of training, easier negative samples are selected; the difficulty threshold is gradually increased with the training rounds, and finally difficult negative samples are selected to enhance the distinguishing ability of the feature space.

[0165] Set the total number of training rounds to T.

[0166] Initial stage: number of rounds

[0167] Mid-term: Rounds

[0168] Late Stage: Rounds

[0169] Difficulty threshold S t It is used to filter negative samples that meet the requirements of the current training stage. As the training progresses, S t Dynamically adjust to gradually increase the difficulty of negative samples.

[0170] Setting method:

[0171] Initial stage:

[0172] S t =S min

[0173] Choose a difficulty rating higher than S min Negative samples.

[0174] Mid-term stage:

[0175]

[0176] As the number of training rounds t increases, S tFrom S min Increase linearly to S mid .

[0177] Later stages:

[0178]

[0179] S t From S mid Increase linearly to S max .

[0180] S min : Initial difficulty threshold, set to the minimum value of the difficulty score, here the default is 0.

[0181] S mid : Intermediate difficulty threshold, which can be set to the middle value of the difficulty score.

[0182] S max : Maximum difficulty threshold, close to the maximum value of the difficulty score.

[0183] T start 、T mid 、T end : Corresponding to the starting and ending round numbers of the initial, middle and late stages respectively.

[0184] Random selection of negative samples: In the set of negative samples that meet the current difficulty threshold, a negative sample is randomly selected according to the probability distribution of the difficulty score.

[0185] Constructing triples

[0186] Through the above steps, we construct a triplet (anchor sample, positive sample, negative sample) for each anchor sample for model training.

[0187] Introducing the loss function with difficulty weight

[0188] Weighted triplet loss function:

[0189] For each triple, calculate its corresponding difficulty weight. The higher the difficulty, the greater the weight.

[0190] Modify the traditional triple loss function and introduce the difficulty weight w:

[0191] L=w×max{d(f a ,f p )-d(f a ,f n )+α,0}

[0192] in:

[0193] f a ,f p,f n are the feature vectors of anchor samples, positive samples, and negative samples respectively.

[0194] d(·,·) is the distance between the feature vectors (usually the Euclidean distance).

[0195] α is the preset margin.

[0196] w is the difficulty weight, which is determined according to the difficulty score of the negative sample.

[0197] Dynamic adjustment of weights:

[0198] As training progresses, the calculation method of difficulty weights is adjusted to gradually emphasize the impact of high-difficulty negative samples on the model.

[0199] Working process

[0200] Data acquisition: High-resolution cameras collect images of the distribution cabinet switches in real time, and sensors simultaneously collect temperature, current, and humidity data. Images and sensor data are transmitted to the data processing module through the system control unit.

[0201] Marker detection and cropping: The YOLOv5 model is used to detect the switch position, and the cropping and expansion algorithm generates a sub-image containing the switch features. The cropped image is normalized to unify the brightness and contrast.

[0202] In addition to YOLOv5, other efficient target detection models (such as YOLOv7, SSD, Faster R-CNN, etc.) can be used to detect the switch position. The most suitable model can be selected according to actual needs and computing resources to further optimize the detection performance and real-time performance.

[0203] Feature extraction:

[0204] Multi-scale convolution extracts local features, Vision Transformer extracts global features, and SE attention mechanism optimizes feature expression.

[0205] Status classification:

[0206] The SVM classifier classifies the switch state according to the feature vector and combines Focal Loss to optimize the classification results of minority class samples.

[0207] Test optimization:

[0208] TTT adjusts model parameters in real time and combines self-supervision tasks to enhance the model's adaptability to complex environments.

[0209] Multimodal fusion and prediction:

[0210] The cross-modal Transformer models the interaction between images and sensor data.

[0211] The timing analysis module predicts abnormal conditions that may occur in the future.

[0212] YOLOv5's dynamic inference algorithm supports real-time detection of switch positions, and combines it with the cropping and expansion algorithm to generate sub-images containing complete features, improving the accuracy of target positioning and classification.

[0213] The multi-scale convolution module captures local details, the Vision Transformer models global information, and the SE attention mechanism strengthens the expression of key features, significantly improving the recognition accuracy in complex scenes.

[0214] The cross-modal Transformer unifies the modeling of images and sensor data, effectively improving the system's ability to analyze potential abnormal conditions.

[0215] The timing analysis module uses historical data to predict future abnormal conditions and achieve predictive maintenance of switch status.

[0216] TTT and self-supervised learning tasks enhance the adaptability and robustness of the model in dynamic environments.

[0217] Through the above embodiments, the intelligent distribution cabinet switch identification method realizes efficient, accurate and intelligent switch status identification, and provides reliable technical support for real-time monitoring and abnormal warning of the power system.

[0218] The following is a detailed description of Example 3, including specific implementation methods, working processes and corresponding beneficial effects. This example focuses on supplementing the optimization and expansion functions of the system and method, covering trimming and expansion algorithms, multimodal data fusion optimization, and generalization capability enhancement measures, demonstrating the performance improvement and practical application scalability of the technical solution.

[0219] Embodiment 3:

[0220] like Figures 1 to 3 As shown, based on Example 1 and Example 2, this embodiment optimizes and expands the key modules of the system, specifically including:

[0221] Optimization of data processing units

[0222] Crop expansion algorithm improvements:

[0223] After the YOLOv5 detection model locates the switch position, the cropping and expansion algorithm dynamically adjusts the cropping boundaries to ensure that the switch area features are complete.

[0224] Cropping boundary adjustment rules: dynamically expand the boundary according to the confidence of the detection box, moderately reduce the boundary when the confidence is high, and increase the boundary range to capture potential key features when the confidence is low.

[0225] The cropped images are uniformly normalized to eliminate the influence of brightness and contrast differences on feature extraction.

[0226] The specific rules of the cropping expansion algorithm to dynamically adjust the cropping boundaries are:

[0227] Confidence and boundary ratio mapping:

[0228] The boundary extension ratio is inversely proportional to the detection confidence;

[0229] When the confidence level is high, the expansion ratio is reduced (such as increasing by 10%), and when the confidence level is low, the expansion ratio is increased (such as increasing by 30%).

[0230] Background noise suppression:

[0231] The expanded area detects the background area through the convolution kernel, sets the noise area features to zero, and reduces the interference to the subsequent extraction module.

[0232] Optimization of multimodal data fusion

[0233] Cross-modal Transformer improvements:

[0234] The pre-trained cross-modal Transformer model is called to perform multimodal modeling of image features and sensor data (such as temperature, current, and humidity) through a multi-head attention mechanism.

[0235] Fusion rules: Assign weights to each modal feature, and the weights are dynamically adjusted based on the correlation between the feature and the abnormal state.

[0236] Increase the attention interaction between features and improve the ability to recognize abnormal states in complex environments.

[0237] The following dynamic weight allocation rules are introduced in the cross-modal Transformer:

[0238] Relevance weight calculation:

[0239] The correlation between each modal feature and the abnormal state is scored, and the weight is assigned using a weighted average strategy;

[0240] The formula is:

[0241]

[0242] Weight normalization:

[0243] The assigned weights are normalized by the Softmax function to ensure numerical stability during feature fusion.

[0244] Timing analysis prediction optimization:

[0245] Use LSTM or temporal Transformer to perform time series modeling on the historical data of switch states.

[0246] Optimize prediction rules: Introduce dynamic time windows, adjust the sampling frequency of time series according to the real-time volatility of data, and enhance the prediction accuracy of future abnormal conditions (such as tripping).

[0247] The generalization ability of the test optimization unit is enhanced

[0248] Dynamic optimization mechanism:

[0249] Test Time Training (TTT) is used to optimize model parameters in real time, and the weight distribution of the classifier and feature extractor is dynamically adjusted according to the input data distribution.

[0250] The model's receptive field size and attention distribution are dynamically adjusted to account for the effects of lighting, switch position offset, etc. in different environments.

[0251] Self-supervised tasks:

[0252] The image reconstruction task is introduced as an auxiliary objective, and unlabeled data is used for self-supervised learning to enhance the model's generalization ability for unknown data distribution.

[0253] Reconstruction goal: Reconstruct the feature distribution of the input cropped image to ensure the consistency of feature extraction.

[0254] In addition to the image reconstruction task, the self-supervised learning task adds the following contrastive learning task:

[0255] Mission objectives:

[0256] Strengthen the feature consistency of samples in the same state, while maximizing the feature differences of samples in different states.

[0257] Task Implementation:

[0258] The feature extraction module is optimized by contrasting the loss function to make the extracted features more robust and discriminative.

[0259] Working process

[0260] Data processing optimization:

[0261] A high-resolution camera captures images of the distribution cabinet switches. After the YOLOv5 model detects the switch positions, the cropping and expansion algorithm adjusts the cropping boundaries based on the confidence level.

[0262] The cropped sub-image is normalized to eliminate the deviation of image brightness and contrast.

[0263] Multimodal data fusion and time series prediction:

[0264] Image features and environmental data (temperature, current, humidity) are uniformly modeled through a cross-modal Transformer, and multimodal data weights are dynamically allocated.

[0265] Call LSTM or time transformer to analyze historical data and combine it with real-time data to predict possible future abnormal states of the switch.

[0266] Optimization during the testing phase:

[0267] Use TTT to dynamically adjust the classification parameters of the model based on the real-time input data distribution.

[0268] Combined with the image reconstruction task, the feature extraction module is optimized by using unlabeled images to enhance the model adaptability.

[0269] The improved cropping and expansion algorithm dynamically adjusts the cropping boundary to ensure that the cropped sub-image completely contains the key features of the switch and the surrounding environment information.

[0270] Image normalization eliminates differences in brightness and contrast, ensuring stability and consistency in feature extraction.

[0271] The cross-modal Transformer dynamically adjusts the fusion weights of image and sensor data through a multi-head attention mechanism, improving the system's ability to adapt to complex environments.

[0272] LSTM and temporal Transformer combined with dynamic time windows can accurately capture the key features of data fluctuations and significantly improve the prediction accuracy of future abnormal conditions.

[0273] The TTT method optimizes model parameters in real time and improves the adaptability of the model in complex scenarios such as lighting changes and angle deviations.

[0274] The self-supervised task optimizes the model’s feature extraction capabilities through unlabeled images and enhances the learning effect on unknown data distribution.

[0275] The optimization content of Example 3 can be widely applied to scenarios with dynamic environmental characteristics, such as switch status monitoring in high temperature and high humidity environments.

[0276] Cross-modal data fusion supports collaborative analysis of multi-source information, providing a more comprehensive basis for intelligent decision-making in abnormal conditions.

[0277] Combined with TTT's dynamic optimization capabilities, the system can continuously adapt to environmental changes and ensure long-term stable operation.

[0278] Through the optimization and expansion of this embodiment, the performance and prediction ability of the intelligent distribution cabinet switch identification system in complex environments are further enhanced, reflecting the wide applicability and scalability of the technical solution.

[0279] In view of current practical needs, the above-mentioned implementation mode adopted in this application is not limited to the scope of protection. Various changes made within the knowledge scope of technical personnel in this field without departing from the concept of this application still fall within the scope of protection of the present invention.

Claims

1. An intelligent identification system for switches in a power distribution cabinet, characterized in that: It includes data acquisition unit, data processing unit, feature extraction unit, state classification unit, test optimization unit, data fusion and analysis unit, and system control unit; The data acquisition unit includes at least one high-resolution camera for collecting real-time image data of the switch of the power distribution cabinet, and the image covers a variety of lighting conditions and angles; at least one multimodal sensor for collecting environmental parameters of the power distribution cabinet, including temperature, current and humidity; The data processing unit calls a pre-trained YOLOv5 target detection model and is configured to detect the position of the switch of the power distribution cabinet in real time; according to the detection result, the switch area is expanded and cropped to generate a sub-image containing the switch image features; The feature extraction unit includes a multi-scale convolution module for extracting local detail features by combining 3×3, 5×5 and 7×7 convolution kernels; calling Vision Transformer to extract global context features of the switch image through position encoding and multi-head self-attention mechanism; Use SE attention mechanism to weight feature channels, strengthen important feature expressions, and suppress redundant features; The state classification unit is configured with a support vector machine (SVM) classifier to classify the extracted feature vectors; combined with a category balancing mechanism (such as Focal Loss), a higher weight is given to abnormal state samples; and the state of the output switch, including normal, closed and tripped; The test optimization unit supports dynamic optimization based on Test Time Training (TTT) and adjusts model parameters in real time; it combines self-supervised learning tasks (such as image reconstruction) to enhance the model's adaptability to test data distribution; The data fusion and analysis unit calls a pre-trained cross-modal Transformer model to fuse image data and environmental sensor data; analyzes historical time series data and calls an LSTM or temporal Transformer model to predict the probability of occurrence of future abnormal states; The system control unit is used to coordinate data transmission and functional operations between various units; and provide system task allocation and status monitoring functions.

2. The intelligent identification system for switches in a power distribution cabinet according to claim 1 is characterized in that: The data acquisition unit includes: a high-resolution camera with a resolution of 1080p or above, supporting low-light environment imaging; a temperature sensor, a current sensor and a humidity sensor, which are used to collect environmental data.

3. The intelligent identification system for switches in a power distribution cabinet according to claim 1 is characterized in that: The YOLOv5 model called by the data processing unit has the following configuration: using a dynamic reasoning optimization algorithm to detect the switch position in real time; the detection result is processed by a cropping and expansion algorithm to ensure that the cropped area contains complete switch image features; Normalize the cropped image to improve the accuracy of subsequent feature extraction.

4. The intelligent identification system for switches in a power distribution cabinet according to claim 1 is characterized in that: The feature extraction unit includes: a multi-scale convolution module, which extracts fine-grained features of the switch image through convolution kernels of different sizes; a VisionTransformer module, which combines position encoding and multi-head self-attention mechanism for global modeling; and an SE attention mechanism, which is used to re-weight feature channels and improve the expression of important features.

5. The intelligent identification system for switches in a power distribution cabinet according to claim 1 is characterized in that: The state classification unit combines the category balancing mechanism and adopts the following strategy: giving higher weight to abnormal state samples (such as tripping state) based on Focal Loss; The support vector machine (SVM) classifier is called to classify the features and output normal, closed and tripped states.

6. The intelligent identification system for switches in a power distribution cabinet according to claim 1 is characterized in that: The test optimization unit supports functions based on Test Time Training (TTT), including: dynamically optimizing model parameters according to the real-time input data distribution; combining self-supervised tasks (such as feature reconstruction) to enhance the model generalization ability and maintain high accuracy under lighting changes or switch position offsets.

7. A method for intelligent identification of switches in a power distribution cabinet, characterized in that: The following steps are involved: S1, data collection: use a high-resolution camera to collect images of distribution cabinet switches; Use multimodal sensors to collect environmental data, including temperature, current, and humidity; S2, object detection and image cropping: Call the YOLOv5 model to detect the switch position; crop the detected switch area to generate a switch sub-image; S3, feature extraction: use multi-scale convolution module to extract local features; Call Vision Transformer to extract global features and combine SE attention mechanism to optimize feature expression; S4, state classification: call the support vector machine (SVM) classifier to classify the extracted features; output the state of the switch, including normal, closed and tripped; S5, test optimization: adjust model parameters in real time based on TTT, and enhance model adaptability by combining self-supervision tasks; S6, Multimodal data fusion and time series analysis: Fusion of image data and environmental data, calling cross-modal Transformer to model multimodal features; Call LSTM or time transformer to analyze historical data and predict future abnormal conditions.

8. The method for intelligent identification of switches in a power distribution cabinet according to claim 7, characterized in that: The target detection and image cropping steps include: using a YOLOv5 model in combination with a dynamic inference algorithm to detect switches in real time; and using a cropping extension algorithm to ensure that the sub-image contains complete features.

9. The intelligent identification method of a switch in a power distribution cabinet according to claim 7, characterized in that: In the multimodal data fusion and time series analysis steps: the cross-modal Transformer performs multi-head attention modeling on image features and environmental data; the LSTM or time Transformer predicts the probability of occurrence of future abnormal states.

Images