Underwater wet-plugging electric connector insulation characteristic degradation detection method

Abstract

The invention relates to the technical field of insulation degradation detection, in particular to an insulation characteristic degradation detection method for an underwater wet-plugging electric connector, which is characterized in that a multi-angle appearance image of the connector is acquired through a pressure-resistant underwater camera device, and the image quality and the recognition accuracy are improved by adopting a defogging and color correction technology. And further extracting defect characteristics such as cracks, corrosion and aging on the surface of the connector through an image by virtue of a graph neural network model, realizing automatic identification and positioning of defect types, and then executing a local electrical test on the identified defect region to obtain electrical parameters such as insulation resistance, capacitance value and medium absorption ratio. Image recognition information and electrical test data are constructed into a feature matrix, a fusion evaluation model composed of a logistic regression model and a weighted scoring mechanism is input, an insulation degradation degree value is output, and a detection result is generated, so that the accuracy of insulation detection of the underwater wet-plugging electric connector is improved, and the detection time is effectively shortened.

Description

Technical Field

[0001] This invention relates to the field of insulation degradation detection technology, and more particularly to a method for detecting the insulation degradation of underwater wet-plug electrical connectors. Background Technology

[0002] Underwater wet-plug electrical connectors are widely used in critical fields such as marine engineering, deep-sea exploration, underwater robotics, and seabed communications. They are exposed to complex environments with high pressure, moisture, and high salt corrosion, making them highly susceptible to insulation degradation, which in turn affects the safe operation and mission execution capabilities of equipment. Especially in deep-sea operations, connector insulation failure can lead not only to communication interruptions and power transmission failures, but also potentially to short circuits, equipment damage, and serious safety accidents.

[0003] Traditional insulation testing methods rely on manual disassembly followed by electrical testing, or the use of underwater testing instruments for macroscopic assessment of the overall condition. These methods suffer from the inability to assess the degree of degradation at specific defect locations, leading to excessively long testing times. Some methods attempt to manually inspect the connector surface condition based on underwater images, but underwater images generally suffer from low visibility, color distortion, and image blurring, making it difficult to support precise identification and defect localization. Furthermore, the limited range of electrical test results and their susceptibility to environmental disturbances make it difficult to accurately determine the level of insulation degradation risk based solely on electrical test data. Summary of the Invention

[0004] To address the above problems, this invention provides a method for detecting the degradation of insulation characteristics of underwater wet-plug electrical connectors.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for detecting insulation degradation of underwater wet-plug electrical connectors includes the following steps: S1. Acquire appearance images of the target underwater wet-plug electrical connector from several angles using a pressure-resistant camera device installed underwater, and perform dehazing enhancement and color correction processing on the appearance images to generate enhanced image data. S2. Based on enhanced image data, visual feature information of the connector surface area is extracted through an image recognition model, and defect types including insulation material cracks, corrosion spots and seal aging are identified, and defect type prediction results are output. S3. Based on the defect type prediction result, apply a preset test voltage to the defect area through the detection device and collect electrical test data including insulation resistance, capacitance value and dielectric absorption ratio. S4. Based on the defect type prediction results and electrical test data, construct a feature matrix that includes defect location codes, electrical parameter deviation values ​​and historical state vectors. Input the feature matrix into a fusion evaluation model constructed by a logistic regression model and weighted scoring rules. Calculate the insulation characteristic degradation degree value through feature weighting and generate the detection result based on the insulation characteristic degradation degree value.

[0006] Furthermore, the acquisition of appearance images of the target underwater wet-plug electrical connector from several angles using a pressure-resistant camera device installed underwater includes: Using an underwater camera module with a pressure-resistant shell and a ring light source, several shooting angles are set around the connector, and the camera module is controlled to sequentially capture images of the connector's appearance from different angles along a preset path.

[0007] Furthermore, the method of performing dehazing enhancement and color correction processing on the appearance image includes the following steps: Based on the collected original appearance images, the local brightness distribution and color channel histograms of the images are extracted to construct image quality perception feature vectors; Based on the image quality-perceived feature vector, image dehazing based on dark channel prior is performed on the original image to obtain the first enhanced image; Based on the color distribution of the first enhanced image and the preset underwater color compensation curve, color shift correction and white balance adjustment are performed to obtain enhanced image data.

[0008] Furthermore, the image recognition model is trained through the following steps: Based on the appearance images of underwater wet-plug electrical connectors collected from different waters and service periods, an image annotation sample set is constructed, which includes image data and corresponding defect type labels. Based on the image data, image segmentation and scale normalization are performed on the connector region in the image, and graph structure samples are constructed based on the adjacency relationship between image blocks to generate graph structure input data containing graph nodes, edge connections and local feature vectors. Based on graph structure input data, a graph neural network model containing several layers of graph convolution and region attention mechanisms is used to perform node feature aggregation and region difference weighting operations, and output the corresponding defect type prediction results. Based on the classification error loss value between the defect type prediction result and the defect type label, a supervised learning strategy is used to iteratively optimize the weight parameters of the graph neural network model to obtain a converged image recognition model.

[0009] Furthermore, the classification error loss value between the defect type prediction result and the defect type label is calculated using the following loss function: ; in, This represents the classification error loss value. Total number of defect type categories; For the first The true label value of the class defect; The graph neural network model predicts the first... The probability value of a class of defects; For the first Weighting coefficients for class defects.

[0010] Further, S3 includes the following steps: Based on the coordinates of the defect area marked by the image recognition results, the electrical contact port corresponding to the connector under test and its electrical test path relative to the ground wire are determined. According to the electrical test path, a DC test voltage with a preset amplitude and duration is applied to the target contact port through a detection device, and its leakage current change curve is collected during the test to calculate the corresponding insulation resistance value. While applying the test voltage, the capacitance change trend is extracted based on the port charge and discharge response curve, and the corresponding insulation capacitance value and dielectric absorption ratio are calculated.

[0011] Furthermore, the construction of a feature matrix containing defect location codes, electrical parameter deviation values, and historical state vectors based on defect type prediction results and electrical test data includes the following steps: Based on the defect type and corresponding spatial coordinates marked in the defect type prediction results, the defect category label and normalized location index of each defect target are extracted to construct the defect location encoding vector. Based on the insulation resistance and capacitance values ​​collected from the electrical test data, and combined with the preset connector safety reference range, the insulation resistance deviation rate, capacitance deviation rate, and dielectric absorption offset rate are calculated to construct an electrical parameter deviation vector. Based on the historical periodic inspection records of the target connector, the prediction results of historical defect types and the historical electrical parameter change sequence are extracted, the trend change rate and parameter fluctuation range are calculated, and a historical state vector is constructed. The defect location encoding vector, electrical parameter deviation vector, and historical state vector are concatenated according to a preset feature arrangement to generate a feature matrix.

[0012] Further, the feature matrix is ​​input into a fusion evaluation model constructed from a logistic regression model and a weighted scoring rule. The output insulation characteristic degradation value is calculated through feature weighting, including the following steps: A training dataset is constructed based on a series of historical underwater wet-plug electrical connector test samples with known insulation degradation states. The training dataset includes the feature matrix and actual degradation level label for each connector sample. Based on the training dataset, gradient optimization is performed on the weight parameters of the logistic regression model using the maximum likelihood estimation method to minimize the cross-entropy loss between the model output and the actual degenerate label, thereby obtaining a converged set of logistic regression weight parameters. Based on the set of logistic regression weight parameters, a linear combination and Sigmoid mapping process is performed on the feature matrix of the current detection sample to obtain a logistic score value representing the tendency of insulation degradation. According to the preset weighted scoring rules, risk-sensitive weights are assigned to each dimension of the feature matrix, and weighted cumulative calculation is performed to obtain the structural risk score. The degree of insulation degradation is obtained by linearly weighting the logical score and the structural risk score.

[0013] Furthermore, the logistic regression model is a sparse logistic regression model that includes an L1 regularization term.

[0014] Furthermore, the generation of detection results based on the degree of insulation characteristic degradation includes: Based on the insulation degradation value output by the fusion assessment model, a preset multi-level risk threshold table is called to perform interval matching processing on the degradation value to obtain the corresponding risk level label. Based on the risk level label, combined with the defect type prediction results, electrical test data and detection time information, the detection results are generated in a structured manner.

[0015] The beneficial effects of this invention are as follows: This invention acquires images of the connector's appearance from different angles using a pressure-resistant underwater camera device, and combines image dehazing enhancement and color correction technologies to improve image clarity and color reproduction. Based on a graph neural network recognition model, it extracts microscopic defect features from the connector surface, including key degradation signs such as cracks, corrosion, and aging, enabling automatic identification and location of defect types, avoiding the subjectivity and missed detection risks of manual inspection. Through local electrical testing of the defect area, it collects insulation resistance, capacitance, and dielectric absorption ratio, constructs a feature matrix integrating visual information and electrical parameters, and performs a comprehensive judgment using a fusion evaluation model composed of a logistic regression model and weighted scoring rules. This achieves quantitative output of the insulation degradation degree value and generates detection results, effectively improving the accuracy of insulation detection for underwater wet-plug connectors and significantly shortening the detection time. Attached Figure Description

[0016] Figure 1 This is a flowchart of the steps in the underwater wet-plug electrical connector insulation characteristic degradation detection method of the present invention.

[0017] Figure 2 This is a flowchart of the training steps of the image recognition model in this invention. Detailed Implementation

[0018] Please see Figures 1-2 As shown, this invention relates to a method for detecting the insulation degradation of underwater wet-plug electrical connectors, comprising the following steps: S1. Acquire appearance images of the target underwater wet-plug electrical connector from several angles using a pressure-resistant camera device installed underwater, and perform dehazing enhancement and color correction processing on the appearance images to generate enhanced image data. S2. Based on enhanced image data, visual feature information of the connector surface area is extracted through an image recognition model, and defect types including insulation material cracks, corrosion spots and seal aging are identified, and defect type prediction results are output. S3. Based on the defect type prediction result, apply a preset test voltage to the defect area through the detection device and collect electrical test data including insulation resistance, capacitance value and dielectric absorption ratio. S4. Based on the defect type prediction results and electrical test data, construct a feature matrix that includes defect location codes, electrical parameter deviation values ​​and historical state vectors. Input the feature matrix into a fusion evaluation model constructed by a logistic regression model and weighted scoring rules. Calculate the insulation characteristic degradation degree value through feature weighting and generate the detection result based on the insulation characteristic degradation degree value.

[0019] In some embodiments, an underwater camera module with pressure resistance is first used to acquire multi-angle images of the target electrical connector in water. This module is equipped with a ring light source to eliminate local shadow interference in deep-water environments and can be combined with a robotic arm to perform circumferential imaging. The acquired images undergo dehazing processing using an improved dark channel prior algorithm, and an underwater color restoration model is further introduced to complete color correction, resulting in enhanced image data with high fidelity and high recognizability, significantly superior to traditional methods that directly use the original image for identification. In step S2, based on the enhanced image data, a defect identification model for the connector surface is constructed. Unlike existing methods using general convolutional neural networks, this invention employs a structure recognition algorithm based on graph neural networks. By dividing the connector shell image into regions and constructing a graph structure, each image region is treated as a graph node, and edge weights between regions are introduced to capture spatial correlation. A region attention mechanism is then used to achieve high-precision prediction of defect types. The identified defect types include, but are not limited to, insulation layer cracks, metal contact corrosion spots, and aging of the sealing ring, among other typical degradation phenomena affecting insulation performance. In step S3, a preset voltage is applied to the target area and electrical test data is collected, based on the defect location region identified in the image recognition results. The collected parameters not only cover insulation resistance and capacitance values ​​but also specifically introduce the dielectric absorption ratio as an auxiliary indicator to more comprehensively reflect the degree of degradation of the material's dielectric properties. Unlike traditional whole-machine testing, this step achieves a closed loop of identification, location, and testing, effectively improving testing accuracy and efficiency. In step S4, a multimodal feature matrix integrating visual and electrical data is constructed. This matrix not only encodes the spatial location and predicted category of the defect but also quantifies the deviation between electrical parameters and historical data. By introducing a logistic regression model as the basic classifier and combining it with a set of empirical weights to construct a weighted scoring rule, a weighted summation of the contribution of each feature dimension is completed, ultimately outputting a quantitative result of the insulation degradation degree in continuous value form. Compared to the traditional "pass / fail" judgment method, this method enables more granular performance evaluation, helping maintenance personnel to formulate more targeted handling strategies.

[0020] Furthermore, the acquisition of appearance images of the target underwater wet-plug electrical connector from several angles using a pressure-resistant camera device installed underwater includes: Using an underwater camera module with a pressure-resistant shell and a ring light source, several shooting angles are set around the connector, and the camera module is controlled to sequentially capture images of the connector's appearance from different angles along a preset path.

[0021] Specifically, an underwater camera module integrating a pressure-resistant shell and a ring-shaped lighting system is employed. Multiple shooting angles (one shooting point every 30°) are set around its circumference according to the connector's spatial configuration to achieve 360° surface coverage. The camera module is equipped with a CMOS imaging unit, enabling low-light imaging, and a pre-mounted physical filter removes scattering interference caused by suspended particles in the seawater. In terms of control strategy, the camera module moves along a preset trajectory through an integrated structure with an underwater robotic arm or guide rail slider, enabling stable shooting at different angles and ensuring repeatability of image acquisition angles and consistency of image posture. Furthermore, the ring-shaped light source system has an adaptive brightness adjustment function, dynamically adjusting the brightness distribution of the LED array based on the light reflection conditions of the current shooting area to reduce glare and shadows on the metal shell. This significantly improves the signal-to-noise ratio and structural detail representation of underwater images. Compared to existing methods using fixed-angle single-view imaging, this multi-angle dynamic acquisition method offers higher coverage and image robustness, effectively overcoming the problems of local blind spots and glare interference in underwater imaging.

[0022] Furthermore, the method of performing dehazing enhancement and color correction processing on the appearance image includes the following steps: Based on the collected original appearance images, the local brightness distribution and color channel histograms of the images are extracted to construct image quality perception feature vectors; Based on the image quality-perceived feature vector, image dehazing based on dark channel prior is performed on the original image to obtain the first enhanced image; Based on the color distribution of the first enhanced image and the preset underwater color compensation curve, color shift correction and white balance adjustment are performed to obtain enhanced image data.

[0023] In some embodiments, a local brightness scan is performed on the acquired original image. The extreme brightness distribution of each region is statistically analyzed based on a sliding window strategy, and grayscale histogram features of the RGB channels are extracted. The extracted features are used to quantify the local contrast, global brightness uniformity, and color shift trend of the image, forming a one-dimensional image quality perception vector, which serves as the reference input for subsequent image enhancement algorithms. To address the common problems of scattering blur and low contrast in underwater images, an improved dark channel prior dehazing algorithm is selected for image enhancement. Unlike traditional dark channel algorithms, this method introduces a variable receptive field convolution kernel in the dark channel calculation stage to enhance robustness to local color block anomalies and reflective areas. In the transmittance estimation stage, the global atmospheric light value adaptive weight is adjusted through the image quality perception vector, thereby avoiding illumination estimation bias caused by the underwater blue-green background. Finally, guided filtering is used to optimize the dehazing result, outputting the first enhanced image with enhanced contrast. In the color correction stage, this embodiment designs a nonlinear color mapping strategy for underwater scenes. First, the RGB color value distribution of each pixel in the first enhanced image is extracted. Based on a preset underwater fading sample library, the current color distribution is modeled with an offset. Subsequently, a nonlinear color compensation curve based on a LUT (Look-Up Table) is constructed, and the color compensation weights for different regions are dynamically adjusted in conjunction with the image quality perception vector. Simultaneously, adaptive white balance adjustment based on the gray-world assumption is performed to correct warm or cool color shifts caused by chromatic aberration.

[0024] Furthermore, the image recognition model is trained through the following steps: Based on the appearance images of underwater wet-plug electrical connectors collected from different waters and service periods, an image annotation sample set is constructed, which includes image data and corresponding defect type labels. Based on the image data, image segmentation and scale normalization are performed on the connector region in the image, and graph structure samples are constructed based on the adjacency relationship between image blocks to generate graph structure input data containing graph nodes, edge connections and local feature vectors. Based on graph structure input data, a graph neural network model containing several layers of graph convolution and region attention mechanisms is used to perform node feature aggregation and region difference weighting operations, and output the corresponding defect type prediction results. Based on the classification error loss value between the defect type prediction result and the defect type label, a supervised learning strategy is used to iteratively optimize the weight parameters of the graph neural network model to obtain a converged image recognition model.

[0025] In some embodiments, a representative image annotation sample set is first constructed based on underwater wet-plug electrical connector image samples collected from multiple different geographical waters and different service stages. Each image sample is meticulously annotated manually to mark typical defect types appearing on the connector surface, such as microcracks, corrosion points, and material aging areas. The corresponding defect type labels include features such as "surface cracks," "electrode corrosion," and "sealing ring aging." To overcome underwater image blurring and complex background interference, image block segmentation and scale normalization are performed on the connector area in the image during the preprocessing stage to ensure that defect images of different sizes and shooting angles are aligned in a unified feature space. To preserve the spatial structure information between regions, a graph structure sample is constructed, where each image block is treated as a graph node, adjacent image blocks are connected by edges, and feature vectors composed of local brightness, texture, and edge direction information are calculated as the initial input to the graph nodes. During training, the proposed graph neural network model consists of multiple graph convolutional layers and a region attention module. Graph convolution is used to spatially aggregate features of adjacent nodes, while the attention module learns weighted weights based on the contextual differences between nodes, thereby highlighting regions with abnormal textures or color deviations and effectively suppressing interference from redundant regions. The model output is the defect type prediction result corresponding to each graph structure sample. To optimize classification performance, the cross-entropy loss function is used to measure the difference between the prediction result and the manual annotation. The Adam optimizer is used to perform backpropagation and weight updates, and training is iterated until the loss value converges. Compared with traditional global image classification methods, this scheme significantly improves the fine-grained defect recognition capability in low-contrast, unstructured image backgrounds through graph structure modeling and region weighting mechanisms, exhibiting stronger robustness and adaptability.

[0026] Furthermore, the classification error loss value between the defect type prediction result and the defect type label is calculated using the following loss function: ; in, This represents the classification error loss value. Total number of defect type categories; For the first The true label value of the class defect; The graph neural network model predicts the first... The probability value of a class of defects; For the first Weighting coefficients for class defects.

[0027] It's important to note that this loss function takes the defect type probability distribution output by the graph neural network model and the actual defect labels in the image annotation sample set as input. It calculates the error by comparing the deviation between the predicted probability and the actual label for each class, and introduces class weight coefficients to adjust the contribution of different classes to the total loss, thus mitigating training bias that may result from imbalanced sample distribution. Specifically, the loss function is a multi-class weighted cross-entropy, which assigns different importance coefficients to each defect label class, thereby optimizing the identification of easily confused or minority classes. For example, in statistical analysis of the sample set, the number of samples of certain minor corrosion or edge aging defects is much lower than that of common cracks. Therefore, if the model does not weight these defects during training, it is prone to low accuracy in identifying small sample classes. By setting high weight values ​​to enhance the sensitivity of these defects in the loss function, the model can focus more on optimizing the feature representation ability of small sample classes during parameter updates, thereby improving overall classification performance. Furthermore, during the loss calculation stage, the model's learning of node feature aggregation and regional difference attention weights in the graph neural network will be guided by the gradient of the loss function, gradually improving the model's ability to identify surface micro-difference features and achieving high-confidence defect type output.

[0028] Further, S3 includes the following steps: Based on the coordinates of the defect area marked by the image recognition results, the electrical contact port corresponding to the connector under test and its electrical test path relative to the ground wire are determined. According to the electrical test path, a DC test voltage with a preset amplitude and duration is applied to the target contact port through a detection device, and its leakage current change curve is collected during the test to calculate the corresponding insulation resistance value. While applying the test voltage, the capacitance change trend is extracted based on the port charge and discharge response curve, and the corresponding insulation capacitance value and dielectric absorption ratio are calculated.

[0029] In some embodiments, firstly, based on the coordinates of the defect area marked by the image recognition model, the spatial position of the corresponding connector in the structural map is determined, and the electrical contact port number it is connected to is matched. Combining the wiring topology information of the defect area and the connector port, an electrical test path from the target port to the grounding terminal is established, forming a structured path parameter set to ensure accurate positioning and path uniqueness in subsequent tests. Subsequently, the control detection device applies a DC regulated voltage of preset amplitude and test duration to the target contact port, collects the leakage current response curve during the test in real time, and performs noise filtering and baseline calibration on the current fluctuation data. Based on Ohm's law, the collected voltage and current data are calculated to obtain the equivalent insulation resistance value of the insulator in the defect area, and a difference analysis is performed with the resistance reference value of a similar defect-free sample for subsequent feature fusion processing. Furthermore, to obtain the polarization characteristic parameters of the insulating material, the charging and discharging response current data of the port are simultaneously collected during voltage application. By analyzing the recovery trend of the current curve in the early stage of charging and after discharging, the changing trend of the insulation capacitance is extracted, and the dielectric absorption ratio (DAR value) is calculated in conjunction with a standard absorption time window. This ratio can effectively characterize the polarization delay and absorption performance of insulating materials, helping to reveal the degree of material aging. Through the above steps, a local electrical test data vector containing insulation resistance, capacitance, and dielectric absorption ratio is finally formed.

[0030] Furthermore, the construction of a feature matrix containing defect location codes, electrical parameter deviation values, and historical state vectors based on defect type prediction results and electrical test data includes the following steps: Based on the defect type and corresponding spatial coordinates marked in the defect type prediction results, the defect category label and normalized location index of each defect target are extracted to construct the defect location encoding vector. Based on the insulation resistance and capacitance values ​​collected from the electrical test data, and combined with the preset connector safety reference range, the insulation resistance deviation rate, capacitance deviation rate, and dielectric absorption offset rate are calculated to construct an electrical parameter deviation vector. Based on the historical periodic inspection records of the target connector, the prediction results of historical defect types and the historical electrical parameter change sequence are extracted, the trend change rate and parameter fluctuation range are calculated, and a historical state vector is constructed. The defect location encoding vector, electrical parameter deviation vector, and historical state vector are concatenated according to a preset feature arrangement to generate a feature matrix.

[0031] In the specific implementation process, the system first extracts the corresponding defect category label (such as crack, corrosion, aging) and spatial location coordinates for each identified defect target based on the defect type prediction results output by the graph neural network model. Then, it normalizes the location coordinates using the connector structure diagram and maps them to a location index code. The defect type label is further combined with the normalized location index to form a multi-dimensional defect location encoding vector, used to accurately calibrate the distribution of the target defect in three-dimensional space. For modeling electrical performance parameter deviations, starting from the insulation resistance, capacitance, and dielectric absorption ratio data obtained in the previous processing step, and referencing the safety reference range determined by design standards or empirical testing, the deviation rate of the current test data relative to the benchmark value is calculated. For example, the resistance deviation rate = (actual value - reference lower limit) / reference value range, used to reflect the degree of insulation performance degradation. In this way, the insulation resistance deviation rate, capacitance deviation rate, and dielectric absorption offset rate are obtained respectively, combined into an electrical parameter deviation vector, effectively revealing the changes in electrical characteristics caused by the current defect. Regarding historical state modeling, the system calls upon the target connector's previous test records to extract historical defect types and electrical test data sequences. By calculating the switching frequency of defect categories, the trend slope of resistance and capacitance, and the maximum and minimum differences through a sliding window, the degradation trend and stability of the equipment status over time are quantified, and a historical state vector is constructed. Finally, the above three sub-vectors are concatenated according to a preset feature arrangement strategy to form a feature matrix with consistent dimensions and a standardized structure.

[0032] Further, the feature matrix is ​​input into a fusion evaluation model constructed from a logistic regression model and a weighted scoring rule. The output insulation characteristic degradation value is calculated through feature weighting, including the following steps: A training dataset is constructed based on a series of historical underwater wet-plug electrical connector test samples with known insulation degradation states. The training dataset includes the feature matrix and actual degradation level label for each connector sample. Based on the training dataset, gradient optimization is performed on the weight parameters of the logistic regression model using the maximum likelihood estimation method to minimize the cross-entropy loss between the model output and the actual degenerate label, thereby obtaining a converged set of logistic regression weight parameters. Based on the set of logistic regression weight parameters, a linear combination and Sigmoid mapping process is performed on the feature matrix of the current detection sample to obtain a logistic score value representing the tendency of insulation degradation. According to the preset weighted scoring rules, risk-sensitive weights are assigned to each dimension of the feature matrix, and weighted cumulative calculation is performed to obtain the structural risk score. The degree of insulation degradation is obtained by linearly weighting the logical score and the structural risk score.

[0033] In some embodiments, a feature matrix is ​​first constructed for each underwater wet-plug electrical connector sample to be tested. This feature matrix integrates three types of information: defect location encoding, electrical parameter deviation values, and historical state vectors. The defect location encoding vector is generated by the defect type and its corresponding spatial coordinates output by the graph neural network recognition model, and the relative positional features of the connector structure are encoded using a normalized indexing method. The electrical parameter deviation vector is calculated by standardizing the insulation resistance value, capacitance value, and dielectric absorption ratio obtained from the current detection, and comparing them with a preset safety reference threshold range. The historical state vector is based on the connector's previous detection records, extracting the defect type prediction results and the change trajectory of electrical parameters in the time series, and further quantifying their trend slope and fluctuation range. Subsequently, the above feature matrix is ​​input into the fusion evaluation model, which consists of a logistic regression unit and a weighted scoring unit. The logistic regression unit uses maximum likelihood estimation to train the model weight parameters, and iteratively optimizes the cross-entropy loss between the feature matrix and the known degradation level labels to enable the model to have a strong degradation classification boundary learning ability. During the inference phase, the logistic regression model maps the feature matrix to logistic scores, representing the probability level of connector insulation degradation. Simultaneously, a weighted scoring unit, based on preset risk-sensitive weights, weights and accumulates the information from each dimension of the feature matrix to generate a structural risk score, characterizing the structural degradation risk of the connector. Finally, a linear weighted average is performed on the logistic score and the structural risk score to output a comprehensive numerical index representing the current degree of connector insulation performance degradation. This index can serve as a decision support basis for maintenance planning and degradation risk early warning, significantly improving the accuracy and robustness of insulation degradation detection. Furthermore, the logistic regression model is a sparse logistic regression model that includes an L1 regularization term.

[0034] Specifically, the logistic regression model uses a sparse logistic regression model with L1 regularization as the core classifier to improve the ability to identify key degradation factors in the high-dimensional feature matrix. During training, by pairing the feature matrices of multiple historical samples with degradation level labels, the loss function is first constructed using the maximum likelihood estimation method, and the L1 norm is introduced as a regularization term. This allows the model to apply sparse constraints to redundant dimensions in the input features while fitting the prediction task, forcing the weights of unimportant features to approach zero, thereby achieving a unity of feature selection and modeling. This sparsity not only enhances the model's generalization ability but also improves the interpretability of the final degradation determination. During model parameter optimization, gradient descent is used to jointly optimize the logistic regression weights and the L1 regularization term until the loss function converges. In the inference phase, the model can achieve high-precision prediction of insulation degradation based on the input connector detection feature matrix, relying only on a small number of high-weight key feature dimensions. This design is particularly suitable for industrial defect assessment tasks with limited sample size but complex feature dimensions, effectively suppressing the risk of overfitting.

[0035] Furthermore, the generation of detection results based on the degree of insulation characteristic degradation includes: Based on the insulation degradation value output by the fusion assessment model, a preset multi-level risk threshold table is called to perform interval matching processing on the degradation value to obtain the corresponding risk level label. Based on the risk level label, combined with the defect type prediction results, electrical test data and detection time information, the detection results are generated in a structured manner.

[0036] Specifically, the system first obtains the insulation degradation degree value output by the fusion evaluation model. This value is a continuous index between 0 and 1, used to characterize the current degradation trend of underwater wet-plug electrical connectors in terms of insulation performance. To qualitatively determine the risk level of this degradation degree value, the system calls a preset multi-level risk threshold table. This threshold table is constructed based on marine engineering service life standards, statistical data from previous tests, and failure case analysis, setting risk intervals such as [0, 0.3), [0.3, 0.6), [0.6, 0.8), and [0.8, 1.0], corresponding to risk level labels of "safe," "early warning," "mild degradation," and "severe degradation," respectively. The system performs interval matching on the degradation degree value and assigns the corresponding risk level label based on the matching result. Subsequently, the system combines the defect type prediction results output by the image recognition model (such as seal aging, corrosion spots, etc.), the insulation resistance and capacitance deviation values ​​collected during electrical testing, and the execution timestamp of the testing task to construct a structured test result object. The object is organized in JSON format, and the fields include: connector identification code, inspection time, risk level label, defect type details, electrical performance deviation index, and recommended maintenance operation level.

[0037] The above embodiments are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for detecting the insulation characteristic degradation of underwater wet-plug electrical connectors, characterized in that, Includes the following steps: S1. Acquire appearance images of the target underwater wet-plug electrical connector from several angles using a pressure-resistant camera device installed underwater, and perform dehazing enhancement and color correction processing on the appearance images to generate enhanced image data. S2. Based on enhanced image data, visual feature information of the connector surface area is extracted through an image recognition model, and defect types including insulation material cracks, corrosion spots and seal aging are identified, and defect type prediction results are output. S3. Based on the defect type prediction result, apply a preset test voltage to the defect area through the detection device and collect electrical test data including insulation resistance, capacitance value and dielectric absorption ratio. S4. Based on the defect type prediction results and electrical test data, construct a feature matrix that includes defect location codes, electrical parameter deviation values ​​and historical state vectors. Input the feature matrix into a fusion evaluation model constructed by a logistic regression model and weighted scoring rules. Calculate the insulation characteristic degradation degree value through feature weighting and generate the detection result based on the insulation characteristic degradation degree value.

2. The method for detecting insulation degradation of underwater wet-plug electrical connectors according to claim 1, characterized in that, The acquisition of appearance images of the target underwater wet-plug electrical connector from several angles using a pressure-resistant camera device installed underwater includes: Using an underwater camera module with a pressure-resistant shell and a ring light source, several shooting angles are set around the connector, and the camera module is controlled to sequentially capture images of the connector's appearance from different angles along a preset path.

3. The method for detecting insulation degradation of underwater wet-plug electrical connectors according to claim 1, characterized in that, The process of performing dehazing enhancement and color correction on the appearance image includes the following steps: Based on the collected original appearance images, the local brightness distribution and color channel histograms of the images are extracted to construct image quality perception feature vectors; Based on the image quality-perceived feature vector, image dehazing based on dark channel prior is performed on the original image to obtain the first enhanced image; Based on the color distribution of the first enhanced image and the preset underwater color compensation curve, color shift correction and white balance adjustment are performed to obtain enhanced image data.

4. The method for detecting insulation degradation of underwater wet-plug electrical connectors according to claim 1, characterized in that, The image recognition model is trained through the following steps: Based on the appearance images of underwater wet-plug electrical connectors collected from different waters and service periods, an image annotation sample set is constructed, which includes image data and corresponding defect type labels. Based on the image data, image segmentation and scale normalization are performed on the connector region in the image, and graph structure samples are constructed based on the adjacency relationship between image blocks to generate graph structure input data containing graph nodes, edge connections and local feature vectors. Based on graph structure input data, a graph neural network model containing several layers of graph convolution and region attention mechanisms is used to perform node feature aggregation and region difference weighting operations, and output the corresponding defect type prediction results. Based on the classification error loss value between the defect type prediction result and the defect type label, a supervised learning strategy is used to iteratively optimize the weight parameters of the graph neural network model to obtain a converged image recognition model.

5. The method for detecting insulation degradation of underwater wet-plug electrical connectors according to claim 4, characterized in that, The classification error loss value between the defect type prediction result and the defect type label is calculated using the following loss function: ； in, This represents the classification error loss value. Total number of defect type categories; For the first The true label value of the class defect; The graph neural network model predicts the first... The probability value of a class of defects; For the first Weighting coefficients for class defects.

6. The method for detecting insulation degradation of underwater wet-plug electrical connectors according to claim 1, characterized in that, S3 includes the following steps: Based on the coordinates of the defect area marked by the image recognition results, the electrical contact port corresponding to the connector under test and its electrical test path relative to the ground wire are determined. According to the electrical test path, a DC test voltage with a preset amplitude and duration is applied to the target contact port through a detection device, and its leakage current change curve is collected during the test to calculate the corresponding insulation resistance value. While applying the test voltage, the capacitance change trend is extracted based on the port charge and discharge response curve, and the corresponding insulation capacitance value and dielectric absorption ratio are calculated.

7. The method for detecting insulation degradation of underwater wet-plug electrical connectors according to claim 6, characterized in that, The process of constructing a feature matrix containing defect location codes, electrical parameter deviation values, and historical state vectors based on defect type prediction results and electrical test data includes the following steps: Based on the defect type and corresponding spatial coordinates marked in the defect type prediction results, the defect category label and normalized location index of each defect target are extracted to construct the defect location encoding vector. Based on the insulation resistance and capacitance values ​​collected from the electrical test data, and combined with the preset connector safety reference range, the insulation resistance deviation rate, capacitance deviation rate, and dielectric absorption offset rate are calculated to construct an electrical parameter deviation vector. Based on the historical periodic inspection records of the target connector, the prediction results of historical defect types and the historical electrical parameter change sequence are extracted, the trend change rate and parameter fluctuation range are calculated, and a historical state vector is constructed. The defect location encoding vector, electrical parameter deviation vector, and historical state vector are concatenated according to a preset feature arrangement to generate a feature matrix.

8. The method for detecting insulation degradation of underwater wet-plug electrical connectors according to claim 7, characterized in that, The process of inputting the feature matrix into a fusion evaluation model constructed from a logistic regression model and a weighted scoring rule, and calculating the output insulation property degradation value through feature weighting, includes the following steps: A training dataset is constructed based on a series of historical underwater wet-plug electrical connector test samples with known insulation degradation states. The training dataset includes the feature matrix and actual degradation level label for each connector sample. Based on the training dataset, gradient optimization is performed on the weight parameters of the logistic regression model using the maximum likelihood estimation method to minimize the cross-entropy loss between the model output and the actual degenerate label, thereby obtaining a converged set of logistic regression weight parameters. Based on the set of logistic regression weight parameters, a linear combination and Sigmoid mapping process is performed on the feature matrix of the current detection sample to obtain a logistic score value representing the tendency of insulation degradation. According to the preset weighted scoring rules, risk-sensitive weights are assigned to each dimension of the feature matrix, and weighted cumulative calculation is performed to obtain the structural risk score. The degree of insulation degradation is obtained by linearly weighting the logical score and the structural risk score.

9. The method for detecting insulation degradation of underwater wet-plug electrical connectors according to claim 8, characterized in that, The logistic regression model is a sparse logistic regression model that includes an L1 regularization term.

10. The method for detecting insulation degradation of underwater wet-plug electrical connectors according to claim 8, characterized in that, The generation of detection results based on the degree of insulation characteristic degradation includes: Based on the insulation degradation value output by the fusion assessment model, a preset multi-level risk threshold table is called to perform interval matching processing on the degradation value to obtain the corresponding risk level label. Based on the risk level label, combined with the defect type prediction results, electrical test data and detection time information, the detection results are generated in a structured manner.

Images