GIS equipment expansion joint hidden danger intelligent detection system and equipment based on deep learning, and storage medium

Abstract

The invention discloses a GIS equipment expansion joint hidden danger intelligent detection system based on deep learning, and relates to the field of intelligent monitoring, and the system comprises an acquisition preprocessing module, a multi-modal data fusion module, a defect detection and classification module, an intelligent hidden danger evaluation module and an optimization feedback module. According to the invention, data of a plurality of sensors are combined, and the health condition of the equipment is comprehensively evaluated through a data fusion technology. The multi-source data fusion can effectively improve the accuracy of a fault detection model, avoids misjudgment caused by limitation of a single data source, and provides more comprehensive and accurate equipment health state evaluation. Through real-time monitoring and early warning, long-term health assessment can be carried out on the equipment, and serious faults of the equipment in the using process are avoided.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent monitoring, and in particular to an intelligent detection system for hidden dangers of expansion joints of GIS devices based on deep learning. Background Art

[0002] With the wide application of industrial automation and intelligent manufacturing, the health status monitoring of equipment has become particularly important. Traditional equipment monitoring methods usually rely on regular maintenance or manual inspection, which have certain lags and blind spots and cannot detect equipment hidden dangers in real time. With the increase in equipment complexity, traditional methods can no longer meet the requirements of real-time monitoring of equipment status and fault warning. Therefore, a more intelligent and real-time equipment health detection system is needed. The faults that may occur during equipment operation need to be identified as early as possible, and traditional regular inspections cannot effectively meet the real-time requirements of equipment health management. Improve the accuracy of fault prediction and avoid downtime or losses caused by equipment failures. Reduce the need for manual inspection and maintenance through intelligent means and lower labor costs.

[0003] In recent years, deep learning and reinforcement learning technologies have made remarkable progress in many fields, especially showing powerful capabilities in computer vision, object detection, fault diagnosis, anomaly detection, etc. Reinforcement learning can optimize the decision-making process through interaction with the environment, has strong adaptability, and is suitable for dealing with complex and dynamically changing detection problems. With the popularization of Internet of Things devices and the explosive growth of sensor data, edge computing, as an emerging computing architecture, is becoming a key technology for processing and analyzing device-side data. Edge computing can transfer data processing and analysis tasks from the cloud to the device side, thereby greatly reducing latency, making decisions in real time, and reducing network bandwidth pressure. Summary of the Invention

[0004] In view of the above or existing problems of the intelligent detection system for hidden dangers of expansion joints of GIS devices based on deep learning, the present invention is proposed.

[0005] To solve the above technical problems, the present invention provides the following technical solutions:

[0006] In a first aspect, an embodiment of the present invention provides an intelligent detection system for hidden dangers of expansion joints of GIS devices based on deep learning, which includes an acquisition and preprocessing module for acquiring infrared, ultrasonic, and visual image data of the expansion joints of GIS devices, enhancing features by image denoising, annotating the data, and creating defect categories; a multi-modal data fusion module for extracting features from the data and then fusing the data;

[0007] The defect detection and classification module is used to perform object detection using YOLOv8, automatically identify the defect location, classify the defects through the ResNet+Attention mechanism, and improve the discrimination ability of similar defects by combining contrastive learning; the intelligent hidden danger assessment module is used to build a GIS device health status assessment model based on GNN, analyze the stress, temperature and vibration conditions of different parts of the expansion joint, calculate the hidden danger level through the fuzzy logic algorithm, and output a prediction report; the optimization feedback module is used to optimize the detection model by combining reinforcement learning with historical data, dynamically adjust the weights, and realize real-time detection at the device end by combining edge computing.

[0008] As a preferred solution of the GIS device expansion joint hidden danger intelligent detection system based on deep learning according to the present invention, wherein: image denoising is used to enhance features, the data is labeled, and defect categories are created, including:

[0009] DnCNN is used for end-to-end denoising, and the loss function is defined: ,

[0010] where y i is the noisy image, x i is the noiseless image, F θ (y i ) is the denoising output of the CNN network, and θ is the network parameter;

[0011] Equalization is performed on the local area to avoid over-enhancement. The local sub-block histogram is calculated, the Canny operator is used for edge extraction, the image gradient is calculated, and then the directional features are extracted through the Gabor filter;

[0012] If a fracture line appears on the surface of the expansion joint, the defect is a crack; if oxidation and rust patches appear on the metal surface, the defect is corrosion; if the surface material peels off or becomes thinner, the defect is wear; if the insulating material discolors and cracks, the defect is insulation aging; if the relative displacement of the connection part is abnormal, the defect is the abnormal relative displacement of the connection part.

[0013] As a preferred solution of the GIS device expansion joint hidden danger intelligent detection system based on deep learning according to the present invention, wherein: feature extraction is performed on the data, and then the data is fused, including:

[0014] The color distribution of the GIS device expansion joint area is characterized by using the color histogram, and the color distribution histogram H c is defined as: ,

[0015] where I(x, y) is the pixel value, δ is the function, and M×N is the image size;

[0016] The gray-level co-occurrence matrix is used to extract the texture features of the image: ,

[0017] where d x , d y is the offset in the direction θ, d is the offset distance, and N d is the normalization factor;

[0018] Then, the features from different sources are directly concatenated: ,

[0019] where F image is the image feature extracted by the CNN, and F sensor is the temperature and current sensor feature of the GIS device.

[0020] As a preferred solution of the intelligent detection system for hidden dangers of the expansion joint of the GIS device based on deep learning according to the present invention, wherein: YOLOv8 is used for object detection to automatically identify the defect location, including:

[0021] Input the preprocessed GIS device image into the detection system, use FPN to detect objects of different sizes, identify specific categories according to the training data, calculate the confidence of each detection box, and screen out high-confidence objects;

[0022] After YOLOv8 object detection, the detection system outputs the defect category, defect location coordinates, and defect confidence score.

[0023] As a preferred solution of the intelligent detection system for hidden dangers of the expansion joint of the GIS device based on deep learning according to the present invention, wherein: calculating the confidence of each detection box and screening out high-confidence objects includes:

[0024] YOLOv8 generates multiple prediction boxes, each box containing the center point coordinates (x, y), width and height (w, h), the probability of target existence P(O), and the probability of each category P(C i ∣O);

[0025] Calculate the final confidence of the detection box: Confidence Score = P(O) × P(C∣O)

[0026] where P(O) is the probability of target existence predicted by YOLOv8, ranging from [0, 1], and P(C∣O) is the probability that the target belongs to a certain defect category, ranging from [0, 1];

[0027] Set the confidence threshold Tc. If Confidence Score > Tc, then retain the detection box, otherwise discard it;

[0028] After screening out high-confidence objects, non-maximum suppression further removes redundant boxes and calculates the intersection over union: ,

[0029] Set the IoU threshold T i , if the IoU of two detection boxes > Ti, keep the box with higher confidence and remove the box with lower confidence;

[0030] Sort the detection boxes in descending order of confidence, calculate the IoU between the current box and the remaining boxes, and if the IoU exceeds the threshold, delete the box with lower confidence.

[0031] As a preferred solution of the intelligent detection system for hidden dangers of expansion joints of GIS devices based on deep learning according to the present invention, wherein: defect classification is performed through the ResNet+Attention mechanism, and contrast learning is combined to improve the discrimination ability of similar defects, including:

[0032] Calculate the importance of each channel through global average pooling and global max pooling: ,

[0033] wherein, σ is the Sigmoid activation function for normalization, and W0, W1, W2 are learning parameters;

[0034] Adopt the SimCLR structure to construct positive and negative samples for contrast learning to improve the discrimination ability of similar defects.

[0035] As a preferred solution of the intelligent detection system for hidden dangers of expansion joints of GIS devices based on deep learning according to the present invention, wherein: a health state evaluation model of GIS devices is constructed based on GNN, the stress, temperature and vibration conditions of different parts of the expansion joint are analyzed, the hidden danger level is calculated through a fuzzy logic algorithm, and a prediction report is output, including:

[0036] Model different parts of the device as nodes in the graph structure, and the interaction between components is represented by the edges in the graph. Each node contains the feature data of this part, and the elements of the adjacency matrix A represent the relationship between node i and node j. If there is an edge between node i and node j, then A ij =1, otherwise A ij =0A. The initial features of the nodes include the stress, temperature and vibration data of the device parts, and the node feature vector is expressed as: ,

[0037] wherein, h i is the feature vector of node i, F i is the stress, T i is the temperature, V i is the vibration;

[0038] The representation of each node is updated by aggregating the features of neighboring nodes. By stacking multiple layers of GCNs, the representation of the node will gradually incorporate information from neighboring nodes. The health state vectors of each part obtained through training the graph neural network model are input into the fuzzy logic algorithm for calculating the hidden danger level. The fuzzy logic system converts the node features into fuzzy sets and infers the hidden danger level through reasoning. The feature data of each node is converted into a fuzzy set through the fuzzification process. During the fuzzy reasoning process, the system calculates the hidden danger level according to the rule base.

[0039] Based on the output of the GNN model and the result of fuzzy logic reasoning, a health state assessment report for each device component is generated.

[0040] As a preferred solution of the GIS device expansion joint hidden danger intelligent detection system based on deep learning according to the present invention, the following steps are included: optimizing the detection model by combining reinforcement learning with historical data, dynamically adjusting the weights, and realizing real-time detection at the device end by combining edge computing.

[0041] The Q-learning algorithm is used to optimize the strategy, and the Q-value update formula is as follows: ,

[0042] where Q(s t , a t ) represents the value of executing action a t in state s t , α is the learning rate, γ is the discount factor, r t is the reward value, and max a Q(s t+1 , a) represents the maximum Q-value in the next state.

[0043] During the training process, the weights of each feature in Q-learning are adjusted according to the contribution degree of each feature. The temperature data has little impact on fault detection, while the vibration data has a greater impact on detection. Higher weights are assigned to the vibration data through the weighting factor.

[0044] In a second aspect, an embodiment of the present invention provides a computer device, including a memory and a processor. The memory stores a computer program, and the computer program instructions, when executed by the processor, implement the steps of the GIS device expansion joint hidden danger intelligent detection system based on deep learning as described in the first aspect of the present invention.

[0045] In a third aspect, an embodiment of the present invention provides a computer-readable storage medium, on which a computer program is stored. The computer program instructions, when executed by the processor, implement the steps of the GIS device expansion joint hidden danger intelligent detection system based on deep learning as described in the first aspect of the present invention.

[0046] The beneficial effects of the present invention are as follows: The present invention combines the data of multiple sensors and comprehensively evaluates the health status of the device through data fusion technology. This multi-source data fusion can effectively improve the accuracy of the fault detection model, avoid misjudgment caused by the limitations of a single data source, and provide a more comprehensive and accurate evaluation of the device's health status. Through real-time monitoring and early warning, the present invention can conduct a long-term health assessment of the device and avoid serious faults during the use of the device. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0048] Figure 1 It is a schematic structural diagram of an intelligent detection system for hidden dangers of expansion joints of GIS devices based on deep learning.

[0049] Figure 2 It is a computer equipment diagram of an intelligent detection system for hidden dangers of expansion joints of GIS devices based on deep learning. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0050] In order to make the above objects, features, and advantages of the present invention more obvious and understandable, the following will make a detailed description of the specific embodiments of the present invention with reference to the drawings in the specification.

[0051] In the following description, many specific details are set forth in order to fully understand the present invention. However, the present invention can also be implemented in other ways different from those described herein. Those skilled in the art can make similar extensions without departing from the connotation of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below.

[0052] Secondly, the so-called "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that can be included in at least one implementation manner of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment that is mutually exclusive with other embodiments.

[0053] Embodiment 1

[0054] Referring to Figures 1 to 2 , which is the first embodiment of the present invention. This embodiment provides an intelligent detection system for hidden dangers of expansion joints of GIS devices based on deep learning, including

[0055] S1: Collect the infrared, ultrasonic, and visual image data of the expansion joint of the GIS device, enhance the features by image denoising, label the data, and create defect categories.

[0056] Preferably, use DnCNN for end-to-end denoising and define the loss function: ,

[0057] where y i is the noisy image, x i is the noise-free image, F θ (y i ) is the denoising output of the CNN network, and θ is the network parameter;

[0058] Perform equalization on the local area to avoid over-enhancement, calculate the local sub-block histogram, use the Canny operator for edge extraction, calculate the image gradient, and then extract the directional features through the Gabor filter;

[0059] If there are fracture lines on the surface of the expansion joint, the defect is a crack; if there are oxidation and rust patches on the metal surface, the defect is corrosion; if the surface material peels off or becomes thinner, the defect is wear; if the insulating material changes color and has cracks, the defect is insulation aging; if the relative displacement of the connection part is abnormal, the defect is the abnormal relative displacement of the connection part.

[0060] In Example 1, the infrared, ultrasonic, and visual image data of the expansion joint of the GIS device are specifically collected through the corresponding infrared imager, ultrasonic sensor, and camera, and then the data are input into the intelligent detection system.

[0061] Furthermore, assume that noisy image data are obtained from the surface of the expansion joint of the GIS device through an infrared sensor. The input image yi is a noisy image with a size of 256x256 pixels and a noise ratio of 10% in the image. The denoised image F θ (yi) is cut into multiple 8x8 small blocks, the histogram of each small block is calculated, and histogram equalization processing is performed. Through local equalization, the local area details of the image become clearer, while avoiding over-enhancement. Use the Canny operator to perform edge extraction on the equalized image. The Canny operator identifies the edge parts in the image by calculating the image gradient. For example, when detecting cracks, the Canny operator can highlight the fracture lines or tiny cracks on the surface of the expansion joint. Calculate the image gradient to help extract the strong edge regions, especially in the crack, wear, and corrosion regions.

[0062] Use the Gabor filter to extract the directional features in the image. The Gabor filter analyzes the image at multiple scales and directions, effectively extracting the detailed information, especially for texture-like defects. For the image Fθ (i) Convolve using Gabor filters with multiple directions and frequencies to obtain filtered results of different scales and directions.

[0063] Use a convolutional neural network for defect classification. The input to the CNN model is the feature map of the denoised image, and the output is the category of the defect. Five defect categories are set: crack, corrosion, wear, insulation aging, and abnormal displacement of the connection part. For each region in the image, the CNN model classifies according to the extracted features (including texture, edge, and directional information). For example, if a significant crack morphology is detected in the image, the CNN will classify that region as a crack. For a corrosion region, the CNN will identify oxidation spots through texture features and classify it as corrosion.

[0064] Locate the position of the defect through edge extraction and directional analysis in the image, and generate the position coordinates of each defect. Output the specific position and classification result of each defect, such as a crack located at coordinates (120, 130) to (200, 230), with high severity, and it is recommended to repair as soon as possible. Corrosion is located at coordinates (50, 60) to (150, 160), with medium severity, and it is recommended to clean and check the coating integrity.

[0065] S2: Extract features from the data and then fuse the data.

[0066] Preferably, use the color histogram to characterize the color distribution of the expansion joint area of the GIS device, and define the color distribution histogram H c : ,

[0067] where I(x, y) is the pixel value, δ is the function, and M×N is the image size;

[0068] Extract the texture features of the image using the gray-level co-occurrence matrix: ,

[0069] where d x , d y is the offset under the direction θ, d is the offset distance, and N d is the normalization factor;

[0070] Then directly splice the features from different sources: ,

[0071] where F image is the image feature extracted by the CNN, and F sensor is the temperature and current sensor feature of the GIS device.

[0072] Furthermore, collect the temperature, pressure, and current sensor data of the expansion joint. Assume the data we collected includes:

[0073] Temperature data (unit: °C), e.g., T = [45.3, 46.1, 47.0, 48.2, 49.3]

[0074] Current data (unit: A), e.g., I = [0.22, 0.24, 0.25, 0.23, 0.26]

[0075] Pressure data (unit: Pa), e.g., P = [2.1, 2.3, 2.2, 2.4, 2.5]

[0076] Assume that an image of a GIS expansion joint is extracted, and the size of the image is 2560×1440 pixels.

[0077] Convert the image to the RGB color space and extract the color value of each pixel. For each pixel (x, y) in the image, calculate its color value in the RGB space and map it to the corresponding color histogram bin.

[0078] For example, assume that we divide the colors into 10 color bins (from red to blue), then each pixel value I(x, y) will fall into one of the bins, and finally form the color histogram Hc of the image. By statistically analyzing the color information of each pixel, we obtain the color distribution of the image, thereby reflecting the color characteristics of different regions in the image.

[0079] Convert the image to a grayscale image to obtain the grayscale value of each pixel in the image. Assume that the offset (dx = 1, dy = 0) is selected to calculate the gray-level co-occurrence matrix in the horizontal direction. For each pair of adjacent pixels (i, j) in the image, calculate the gray-level co-occurrence value between them. For each pair of pixels (I(x, y), I(x + dx, y + dy)) in the image, calculate its gray-level correlation and statistically form the gray-level co-occurrence matrix P(i, j, dx, dy).

[0080] Extract texture features such as contrast, entropy, and homogeneity from the gray-level co-occurrence matrix. Normalize the temperature, current, and pressure sensor data so that the data ranges of different features are consistent.

[0081] Assume that we have normalized the temperature, current, and pressure data respectively, and obtained the new data vectors: T′=[0.1,0.2,0.3,0.4,0.5],I′=[0.05,0.1,0.15,0.1,0.2],P′=[0.1,0.15,0.12,0.18,0.2], P' = [0.1, 0.15, 0.12, 0.18, 0.2].

[0082] The normalized sensor data T', I', and P' are concatenated into a feature vector Fsensor: Fsensor = [0.1, 0.2, 0.3, 0.4, 0.5, 0.05, 0.1, 0.15, 0.1, 0.2, 0.1, 0.15, 0.12, 0.18, 0.2]. The color histogram features Hc extracted from the image, the texture features of the gray-level co-occurrence matrix, and the sensor data feature Fsensor are concatenated together to form a comprehensive feature vector Ftotal.

[0083] S3: Use YOLOv8 for object detection, automatically identify the defect location, classify the defects through the ResNet + Attention mechanism, and improve the discrimination ability of similar defects by combining contrastive learning.

[0084] Preferably, the preprocessed GIS device image is input into the detection system, FPN is used to detect objects of different sizes, specific categories are identified according to the training data, the confidence of each detection box is calculated, and high-confidence objects are screened;

[0085] After YOLOv8 object detection, the detection system outputs the defect category, defect location coordinates, and defect confidence score.

[0086] Preferably, YOLOv8 generates multiple prediction boxes, each box containing the center point coordinates (x, y), width and height (w, h), the probability of object existence P(O), and the probability of each category P(C i ∣O);

[0087] Calculate the final confidence of the detection box: Confidence Score = P(O) × P(C∣O)

[0088] where P(O) is the probability of object existence predicted by YOLOv8, ranging from [0, 1], and P(C∣O) is the probability that the object belongs to a certain defect category, ranging from [0, 1];

[0089] Set the confidence threshold Tc. If Confidence Score > Tc, then retain the detection box, otherwise discard it;

[0090] After screening out high-confidence objects, non-maximum suppression further removes redundant boxes and calculates the intersection over union: ,

[0091] Set the IoU threshold T i , if the IoU of two detection boxes > Ti, retain the box with higher confidence and remove the box with lower confidence;

[0092] Sort the detection boxes in descending order of confidence, calculate the IoU between the current box and the remaining boxes, and if the IoU exceeds the threshold, delete the box with lower confidence.

[0093] Preferably, calculate the importance of each channel through global average pooling and global maximum pooling: ,

[0094] where σ is the Sigmoid activation function for normalization, and W0, W1, W2 are learning parameters;

[0095] Adopt the SimCLR structure, construct positive and negative samples for contrast learning, and improve the ability to distinguish similar defects.

[0096] Furthermore, set the confidence threshold Tc. If the Confidence Score of a certain detection box is greater than Tc, then retain the detection box; otherwise discard it. Assume Tc = 0.5, then only the detection boxes with a confidence score greater than 0.5 will be retained. Assume that multiple detection boxes have been screened out, and their confidence scores and IoUs have been calculated. For each pair of detection boxes with an IoU greater than the threshold Ti, retain the box with higher confidence and discard the box with lower confidence. Set the IoU threshold Ti = 0.5. If the IoU of two boxes is greater than 0.5, and the confidence of box A is higher than that of B, then delete box B. For all boxes, sort them in descending order of confidence, calculate the IoU of each box with the remaining boxes in turn, and remove the redundant boxes.

[0097] After non-maximum suppression, the final output is the category, position coordinates (x, y), width and height (w, h), and confidence score of each defect.

[0098] Assume the detection results are as follows: Detection box 1: Category: crack, position (x, y) = (120, 200), width and height (w, h)=(30, 50), confidence: 0.87;

[0099] Detection box 2: Category: corrosion, position (x, y)=(500, 600), width and height (w, h) = (40, 60), confidence: 0.92;

[0100] Detection box 3: Category: wear, position (x, y)=(700, 800), width and height (w, h) = (25, 40), confidence: 0.55;

[0101] Assume that during the screening process at this step, some redundant boxes are removed through non-maximum suppression, and the final output result has only two boxes: Detection box 1: Category: crack, Location (120, 200), Width and height (30, 50), Confidence: 0.87; Detection box 2: Category: corrosion, Location (500, 600), Width and height (40, 60), Confidence: 0.92.

[0102] S4: Build a GIS equipment health status assessment model based on GNN, analyze the stress, temperature, and vibration conditions of different parts of the expansion joint, calculate the hidden danger level through the fuzzy logic algorithm, and output a prediction report.

[0103] Preferably, different parts of the equipment are modeled as nodes in the graph structure, and the interactions between components are represented by edges in the graph. Each node contains the characteristic data of that part. The elements of the adjacency matrix A represent the relationship between node i and node j. If there is an edge between node i and node j, then A ij = 1, otherwise A ij = 0A. The initial characteristics of the nodes include the stress, temperature, and vibration data of the equipment parts. The node feature vector is expressed as: ,

[0104] where, h i is the feature vector of node i, F i is the stress, T i is the temperature, V i is the vibration;

[0105] Update the representation of each node by aggregating the features of neighboring nodes. By stacking multiple layers of GCN, the representation of the nodes will gradually integrate the information from neighboring nodes. The health status vectors of each part obtained through training the graph neural network model are input into the fuzzy logic algorithm for calculating the hidden danger level. The fuzzy logic system converts the node features into fuzzy sets and conducts reasoning to obtain the hidden danger level; the characteristic data of each node is converted into a fuzzy set through the fuzzification process. During the fuzzy reasoning process, the system calculates the hidden danger level according to the rule base;

[0106] Generate a health status assessment report for each equipment component based on the output of the GNN model and the results of the fuzzy logic reasoning.

[0107] Furthermore, represent the graph structure with the adjacency matrix A, where:

[0108] ,

[0109] h1 = [1200, 80, 0.3], h2 = [1100, 85, 0.4];

[0110] Convert the numerical data into fuzzy sets:

[0111] Force (F): Low (0 - 800 N), Medium (800 - 1500 N), High (1500+ N);

[0112] Temperature (T): Low (0 - 60 °C), Medium (60 - 90 °C), High (90+ °C);

[0113] Vibration (V): Low (0 - 0.2 g), Medium (0.2 - 0.5 g), High (0.5+ g);

[0114] Rule 1, if F is High and T is High, and V is High → Hazard level is Severe;

[0115] Rule 2, if F is Medium and T is High, and V is Low → Hazard level is Medium;

[0116] Rule 3, if F is Low and T is Low, and V is Low → Hazard level is Minor;

[0117] Suppose a component feature, h3 = [1600, 95, 0.6], F = 1600 → High (0.8), T = 95 → High (0.9)

[0118] V = 0.6 → High (0.7), according to Rule 1, Hazard level = 0.8 × 0.9 × 0.7 = 0.504 (Severe).

[0119] S5: Use reinforcement learning combined with historical data to optimize the detection model, dynamically adjust the weights, and combine edge computing to achieve real-time detection at the device end.

[0120] Preferably, use the Q-learning algorithm to optimize the policy, and the Q-value update formula is as follows: ,

[0121] where, Q(s t ,a t ) represents the value of executing action a t in state s t , α is the learning rate, γ is the discount factor, r t is the reward value, max a Q(s t+1 ,a) represents the maximum Q-value in the next state;

[0122] During the training process, adjust the weights of each feature in Q-learning according to their contribution degrees. Temperature data has little impact on fault detection, and vibration data has a greater impact on detection. Assign higher weights to vibration data through weighting factors.

[0123] Furthermore, the reward value r t is designed as:

[0124] ,

[0125] If the device actually has a fault and is successfully detected, r t = +10. If the device has no fault but is misreported, r t = -2. If the fault is not detected, r t = -5;

[0126] Different sensors have different effects on faults. The temperature T has a smaller effect, and the weight is set to 0.2. The vibration V has a larger effect, and the weight is set to 0.6. The force F has a medium effect, and the weight is set to 0.4;

[0127] where w F = 0.4, w T = 0.2w, w V = 0.6; Test the training effect of Q - learning. Test 500 groups of GIS device data:

[0128]

[0129] Result analysis: The fault detection accuracy rate is 94.5%, the false alarm rate is 3.2%, and the undetected rate is 2.3%. After optimization, compared with the fixed strategy, the Q - learning optimization strategy improves the detection accuracy rate by 8.7%.

[0130] It should be noted that this patent includes functions such as image acquisition, data extraction, defect classification, and hidden danger assessment based on deep learning. Deep learning requires a corresponding database, which is also the basis for artificial intelligence training. Specifically, the scheme for the database includes the following construction process:

[0131] 1. Build a basic database

[0132] Utilize the existing data set: Take the common defects of GIS expansion joints that have been made public as the basic database.

[0133] 2. Data acquisition, supplement and correction to complete the database

[0134] 2.1 Self - acquisition: Obtain data of expansion joints in different states through existing or temporarily installed sensors.

[0135] 2.2 Cooperative acquisition: Cooperate with GIS expansion joint manufacturers and manufacturers that monitor the data of expansion joints to obtain the accumulated data.

[0136] 3. Set up a cloud server

[0137] Store the database through the cloud server and update the database regularly.

[0138] 4. Data annotation

[0139] Mark general image recognition defect types and specific image recognition defect types. Enrich and correct specific fault types.

[0140] 5. Data cleaning

[0141] Duplicate removal: Delete duplicate data to avoid overfitting.

[0142] Anomaly detection: Eliminate noise or abnormal data to ensure data quality.

[0143] Format unification: Unify the data format for subsequent processing.

[0144] This embodiment also provides a computer device, which is applicable to the situation of the intelligent detection system for hidden dangers of GIS equipment expansion joints based on deep learning, including a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the intelligent detection system for hidden dangers of GIS equipment expansion joints based on deep learning as proposed in the above embodiment.

[0145] This computer device can be a terminal. This computer device includes a processor, a memory, a communication interface, a display screen, and an input device connected through a system bus. Among them, the processor of this computer device is used to provide computing and control capabilities. The memory of this computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of this computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be achieved through WIFI, a carrier network, NFC (Near Field Communication), or other technologies. The display screen of this computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of this computer device can be a touch layer covering the display screen, or a button, a trackball, or a touchpad provided on the housing of the computer device, or an external keyboard, touchpad, or mouse, etc.

[0146] This embodiment also provides a storage medium, on which a computer program is stored, and when the program is executed by a processor, it implements the intelligent detection system for hidden dangers of GIS equipment expansion joints based on deep learning as proposed in the above embodiment.

[0147] In summary, the present invention combines multiple sensor data and comprehensively evaluates the health status of the equipment through data fusion technology. This multi-source data fusion can effectively improve the accuracy of the fault detection model, avoid misjudgment caused by the limitations of a single data source, and provide a more comprehensive and accurate evaluation of the equipment health status. Through real-time monitoring and early warning, the present invention can conduct a long-term health assessment of the equipment and avoid serious faults during the use of the equipment.

[0148] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered within the scope of the claims of the present invention.

Claims

1. An intelligent detection system for hidden dangers of expansion joints of GIS devices based on deep learning, characterized in that, Including: A data acquisition and preprocessing module, which is used to collect infrared, ultrasonic, and visual image data of the expansion joint of GIS equipment, enhance features by image denoising, label the data, and create defect categories; A multi-modal data fusion module, which is used to extract features from the data and then fuse the data; A defect detection and classification module, which is used to perform object detection using YOLOv8, automatically identify the defect location, classify the defects through the ResNet+Attention mechanism, and improve the discrimination ability of similar defects by combining contrastive learning; An intelligent hidden danger assessment module, which is used to build a health status assessment model of GIS equipment based on GNN, analyze the stress, temperature, and vibration conditions of different parts of the expansion joint, calculate the hidden danger level through a fuzzy logic algorithm, and output a prediction report; An optimization feedback module, which is used to optimize the detection model by combining reinforcement learning with historical data, dynamically adjust the weights, and achieve real-time detection at the device end by combining edge computing.

2. The intelligent detection system for hidden dangers of expansion joints of GIS devices based on deep learning according to claim 1, characterized in that, The enhancing features by image denoising, labeling the data, and creating defect categories includes: End-to-end denoising is performed using DnCNN, and the loss function is defined as: , Among them, y i is the noisy image, x i is the noise-free image, F θ (y i ) is the denoising output of the CNN network, and θ is the network parameter; Equalizing the local area to avoid over-enhancement, calculating the local sub-block histogram, extracting edges using the Canny operator, calculating the image gradient, and then extracting directional features through the Gabor filter; If there is a fracture line on the surface of the expansion joint, the defect is a crack; if there are oxidation and rust patches on the metal surface, the defect is corrosion; if the surface material peels off or becomes thinner, the defect is wear; if the insulating material changes color and has cracks, the defect is insulation aging; if the relative displacement of the connection part is abnormal, the defect is abnormal relative displacement of the connection part.

3. The intelligent detection system for hidden dangers of expansion joints of GIS devices based on deep learning according to claim 1, characterized in that: The extracting features from the data and then fusing the data includes: Characterize the color distribution in the expansion joint area of GIS equipment using a color histogram, and define the color distribution histogram H c : , Where I(x,y) is the pixel value, δ is the function, and M×N is the image size; Extract the texture features of the image using the gray-level co-occurrence matrix: , where d x , d y is the offset in the direction θ, d is the offset distance, and N d is the normalization factor; Then directly splice the features from different sources: , Among them, F image is the image feature extracted by the CNN, and F sensor is the temperature and current sensor feature of the GIS device.

4. The intelligent detection system for hidden dangers of expansion joints of GIS devices based on deep learning according to claim 1, characterized in that: The performing object detection using YOLOv8 and automatically identifying the defect location includes: Inputting the preprocessed GIS equipment image into the detection system, using FPN to detect targets of different sizes, identifying specific categories according to the training data, calculating the confidence of each detection box, and screening high-confidence targets; After YOLOv8 object detection, the detection system outputs the defect category, defect location coordinates, and defect confidence score.

5. The intelligent detection system for hidden dangers of expansion joints of GIS devices based on deep learning according to claim 4, characterized in that, The calculating the confidence of each detection box and screening high-confidence targets includes: YOLOv8 generates multiple prediction boxes, each box containing the center point coordinates (x, y), width and height (w, h), the probability of object existence P(O), and the probability of each class P(C i |O); Calculating the final confidence of the detection box: Confidence Score = P(O)×P(C∣O) Where P(O) is the probability of target existence predicted by YOLOv8, ranging from [0,1], and P(C∣O) is the probability that the target belongs to a certain defect category, ranging from [0,1]; Setting the confidence threshold Tc, if Confidence Score > Tc, then keep the detection box, otherwise discard it; After screening out high-confidence targets, non-maximum suppression further removes redundant bounding boxes and calculates the intersection over union (IoU): , Set the IoU threshold T i , if the IoU of two detection boxes > Ti, keep the box with higher confidence and remove the box with lower confidence; Sorting the detection boxes in descending order of confidence, calculating the IoU between the current box and the remaining boxes, and deleting the boxes with lower confidence if the IoU exceeds the threshold.

6. The intelligent detection system for hidden dangers of expansion joints of GIS devices based on deep learning according to claim 1, characterized in that: The classifying the defects through the ResNet+Attention mechanism and improving the discrimination ability of similar defects by combining contrastive learning includes: Calculate the importance of each channel through global average pooling and global max pooling: , Among them, σ is the Sigmoid activation function for normalization, and W0, W1, and W2 are learning parameters; The SimCLR structure is adopted to construct positive and negative samples for contrast learning to improve the discrimination ability of similar defects.

7. The intelligent detection system for hidden dangers of expansion joints of GIS devices based on deep learning according to claim 1, characterized in that: The GIS device health status evaluation model is constructed based on the GNN to analyze the stress, temperature, and vibration conditions of different parts of the expansion joint, calculate the hidden danger level through the fuzzy logic algorithm, and output a prediction report, including: Model different parts of the device as nodes in a graph structure. The interactions between components are represented by edges in the graph. Each node contains the feature data of that part. The elements of the adjacency matrix A represent the relationship between node i and node j. If there is an edge between node i and node j, then A ij = 1; otherwise, A ij = 0. The initial features of the nodes include the force, temperature, and vibration data of the device parts. The node feature vector is represented as: , Among them, h i is the feature vector of node i, F i is the force, T i is the temperature, V i is the vibration; The representation of each node is updated by aggregating the features of neighbor nodes. By stacking multiple layers of GCN, the representation of the node will gradually integrate the information from neighbor nodes. The health status vectors of each part obtained through the training of the graph neural network model are input into the fuzzy logic algorithm for hidden danger level calculation. The fuzzy logic system transforms the node features into fuzzy sets and infers the hidden danger level; the feature data of each node is transformed into a fuzzy set through the fuzzification process. In the fuzzy inference process, the system calculates the hidden danger level according to the rule base; Based on the output of the GNN model and the result of fuzzy logic inference, a health status evaluation report for each device component is generated.

8. The intelligent hidden danger detection system for the expansion joint of the GIS device based on deep learning according to claim 1, characterized in that, The detection model is optimized by combining reinforcement learning with historical data, the weights are dynamically adjusted, and edge computing is combined to achieve real-time detection at the device end, including: Use the Q-learning algorithm to optimize the strategy, and the Q-value update formula is as follows: , where Q(s t , a t ) represents the value of executing action a t in state s t , α is the learning rate, γ is the discount factor, r t is the reward value, and max a Q(s t+1 , a) represents the maximum Q-value in the next state; During the training process, the weights of each feature in Q-learning are adjusted according to the contribution degree of each feature. The temperature data has little impact on fault detection, and the vibration data has a greater impact on detection. A higher weight is assigned to the vibration data through a weighting factor.

9. A computer device, comprising a memory and a processor, the memory storing a computer program, characterized in that: When the processor executes the computer program, the steps of the intelligent detection system for hidden dangers of the expansion joint of the GIS device based on deep learning according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, the steps of the intelligent detection system for hidden dangers of the expansion joint of the GIS device based on deep learning according to any one of claims 1 to 7 are implemented.

Images