Intelligent wire harness test system based on self-learning algorithm

Abstract

The invention discloses a wire harness intelligent test system based on a self-learning algorithm, and particularly relates to the field of automobile wire harness testing, the system comprises a data acquisition module, a data preprocessing module, an image segmentation module and a test module; the data acquisition module is used for acquiring image data of the automobile wire harness; the data preprocessing module is used for preprocessing the acquired image; the image segmentation module performs instance segmentation on the wire harness in the image; and the test module evaluates the quality of the wire harness by analyzing the segmentation result. According to the invention, through automatic and intelligent technical means, the accuracy, safety and reliability of automobile wire harness testing are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of automobile wiring harness testing, and more particularly to a wiring harness intelligent testing system based on a self-learning algorithm. Background Art

[0002] The automotive wiring harness is a critical component responsible for information and energy transmission within the vehicle. Its role is to connect sensors, actuators, and electronic control units, ensuring the smooth flow of energy and information within the vehicle. With the rapid development of autonomous driving technology and vehicle electrification, the importance of wiring harnesses is becoming increasingly prominent. In modern vehicles, wiring harnesses must not only support complex electrical functions but also meet higher safety and reliability requirements.

[0003] The existing public document 1 (Design and Implementation of Automobile Wiring Harness Detection Device Based on STM32, 2023) proposes a detection device based on STM32, such as Figure 2 As shown, the detection device uses an STM32F4 as the main control chip and integrates peripheral circuits including a power supply circuit, a clock circuit, a serial communication circuit, a USB circuit, a Flash storage circuit, a relay output interface, and an LCD touch screen interface. The core detection circuit consists of a voltage-controlled constant current source circuit, an analog switch circuit, an address decoder circuit, a test channel switching circuit, a capacitance measurement circuit, and an A / D short-circuit protection circuit, supplemented by an LED status indicator. However, the detection device is tested at the end of production, and the cost of rework or scrapping after problems are found is high.

[0004] The existing open literature 2 (Design and Research of New Energy Vehicle Wiring Harness Assembly Detection System, 2023) proposes a new energy vehicle wiring harness assembly detection system based on machine vision, such as Figure 3 As shown, the new energy vehicle wiring harness assembly inspection system utilizes industrial cameras, lenses, light sources, and other hardware, along with PLC and pneumatic control systems, combined with OpenCV and Qt software to achieve automated inspection of wiring harness brackets and covers. The system uses contour detection to detect bracket features and an improved Faster-RCNN model to enhance cover detection accuracy. Experiments have shown that its accuracy exceeds 99% under varying lighting conditions, with a single inspection taking only 250 milliseconds. However, the use of the improved Faster-RCNN model in the new energy vehicle wiring harness assembly inspection system requires a large amount of training data and computing resources, resulting in high deployment and maintenance costs.

[0005] Therefore, there is an urgent need for an intelligent wire harness testing system that can monitor and evaluate wire harness quality in real time during the manufacturing process and has low deployment and maintenance costs. Summary of the Invention

[0006] In order to overcome the above-mentioned defects of the prior art, the present application provides a wire harness intelligent testing system based on a self-learning algorithm, which is used to collect images of automobile wire harnesses through deploying RGB cameras at each workstation of an assembly line, and to pre-process the collected images, to perform instance segmentation on the wire harnesses in the pre-processed images based on a proposed wire harness instance segmentation algorithm, to identify the wire harnesses, and to extract quality features through contour detection and connected component analysis and match them with design parameters to evaluate the quality of the wire harnesses, so as to solve the problems proposed in the above background art.

[0007] To achieve the above object, the present application provides the following technical solutions:

[0008] A wire harness intelligent testing system based on a self-learning algorithm, comprising a data acquisition module and a data preprocessing module, the data acquisition module being used to collect image data of automobile wire harnesses; the data preprocessing module being used to pre-process the collected images, characterized in that it further comprises an image segmentation module and a testing module, the image segmentation module being used to perform instance segmentation on the wire harnesses in the images; and the testing module being used to evaluate the quality of the wire harnesses by analyzing the segmentation results.

[0009] The image segmentation module performs instance segmentation on the wire harnesses in the images through a wire harness instance segmentation algorithm; the wire harness instance segmentation algorithm uses a DeeplabV3+ model to perform semantic segmentation on the images to generate a binary mask M b ; then, a MaskSlic algorithm is used to perform superpixel segmentation in the wire harness area to extract the centroid coordinates of each superpixel; then, the wire harness instance segmentation algorithm constructs a region adjacency graph, each superpixel being a node, adjacent superpixels being connected by edges, and the nodes being classified into endpoints, line segments and intersection points according to their degrees; then, the graph is clustered by a connected component method to divide the graph into clusters without intersection points and clusters with intersection points; for the case with intersection points, TripletNet combines curvature information to calculate the similarity score of neighbor pairs to generate a prediction matrix W to guide path finding, to expand the path from the endpoints and connect complete wire segments; in addition, CrossNet is used to classify the intersection regions to infer the up-down relationship of the wire harnesses to generate a correct instance mask; finally, the path node sequence is fitted by a cubic B-spline curve and discretized into 2D points to provide a lightweight wire harness model.

[0010] As a further scheme of the present application, the binary mask M bThe wiring harness area is marked as white (foreground) and the other areas are marked as black (background); in order to train the DeeplabV3+ model, 407 car door wiring harness images were taken in a simulated industrial environment. After preprocessing, they were expanded into a dataset of 30,000 synthetic images. The dataset covers wiring harnesses of different shapes and colors, ensuring that the DeeplabV3+ model can learn diverse wiring harness features. During the training process, the DeeplabV3+ model used ResNet-101 as the backbone network and was trained for 200 rounds with a batch size of 10. The optimizer used Adam and the learning rate was set from 10 -6 Gradually adjust to 10 -9 .

[0011] As a further solution of the present invention, the MaskSlic algorithm is combined with the binary mask M generated by semantic segmentation. b , limiting the superpixel clustering to the line bundle area to avoid meaningless segmentation of the entire image. After the segmentation is completed, extract each superpixel R i The centroid coordinates of serve as a reference point for subsequent processing steps (e.g., B-spline interpolation).

[0012] As a further solution of the present invention, the line bundle instance segmentation algorithm constructs a regional adjacency graph (RAG) from the superpixel segmentation results to manage the neighborhood relationship between superpixels and identify endpoints. This process is achieved by creating an undirected unweighted graph G = (V, E). If two superpixels are adjacent, an edge is added between their corresponding nodes, where G represents the graph itself, which is a mathematical structure used to describe the relationship between a set of objects; V represents the vertex set, that is, the set of all superpixel vertices in the graph, and each vertex v i Corresponds to a superpixel in the image; E represents the edge set, which is the set of all edges in the graph. After the graph is constructed, nodes are classified according to their degree (i.e., the number of neighbors): outliers are removed, endpoints mark the start or end of a bundle, line segments form part of the bundle's path, and intersections indicate where bundles cross or branch.

[0013] As a further aspect of the present invention, the wire bundle instance segmentation algorithm then calculates the degree of each node to identify nodes in the graph with a degree greater than 2, i.e., intersection nodes. The algorithm then replaces each intersection group consisting of one or more intersection nodes with an equivalent single node. The equivalent node preserves all neighborhood relationships of the original group, ensuring that critical connectivity information is not lost during the simplification process.

[0014] As a further solution of the present invention, the TripletNet takes as input a 32×32 pixel image patch surrounding a node's neighboring nodes. It generates an embedded representation using a ResNet8 feature extractor and a fully connected layer. This is then optimized using a triplet loss to calculate the Euclidean distance d between the two image patches. Simultaneously, a curvature score is generated using a Von Mises distribution based on the angular information of the node sequence. The final similarity score for each pair of neighboring nodes is the product of the TripletNet score and the curvature score, which is then entered into the corresponding position of the prediction matrix W. If there is no intersection in the graph, the prediction matrix W is empty.

[0015] As a further aspect of the present invention, the CrossNet classifies 64×64 pixel image patches in the intersection region to determine which bundle is on top. The CrossNet predicts the top-bottom relationship at the intersection and outputs a probability value for each possible configuration. By selecting the path with the highest probability as the top bundle, no threshold is required.

[0016] As a further aspect of the present invention, a data acquisition module is used to collect image data of automotive wiring harnesses. Specifically, a high-resolution RGB camera with a resolution of 4586 x 2663 pixels is deployed at each workstation on the assembly line, enabling real-time capture of the wiring harness's assembly status, ensuring image clarity and detail. The high-resolution RGB camera is deployed behind each workstation on the assembly line, capturing the wiring harness's assembly status in real time. During the data acquisition process, a dataset is constructed from the 407 captured images, including simulations of different wiring harness configurations and deformation states.

[0017] As a further embodiment of the present invention, the data preprocessing module preprocesses the captured images, including the following specific contents: the preprocessing includes image scaling, horizontal flipping, scaling enhancement, rotation enhancement, four-point perspective transformation, padding, contrast adjustment, brightness adjustment, gamma correction, blurring, and hue saturation adjustment. The image scaling adjusts the original high-resolution image (4586×2663 pixels) to a uniform size of 480×480 pixels. The purpose of image scaling is to reduce the computational complexity of the image data; secondly, it facilitates the image segmentation module to perform batch processing and feature extraction on the image data to ensure data consistency. The horizontal flipping simulates the presentation of the wire harness in different directions by generating a mirrored version of the image. The scaling enhancement simulates the appearance of an object at different distances or viewing angles by adjusting the size of the wire harness in the image. The rotation enhancement simulates the angular changes that may occur in the wire harness during the assembly process by rotating the image. The four-point perspective transformation simulates the perspective changes of the image. The padding adds additional pixels to the image boundary to avoid detection errors caused by boundary effects. Contrast adjustment, brightness adjustment, and gamma correction simulate image differences caused by varying lighting conditions in industrial environments by altering the image's brightness and contrast. Blur processing introduces slight image noise or out-of-focus effects, helping the wire harness instance segmentation algorithm adapt to scenes with lower image quality and improving its robustness in practical applications. Hue and saturation adjustment simulates wire harness components of different materials or colors by changing the image's color properties, enabling the wire harness instance segmentation algorithm to adapt to diverse wire harness appearances.

[0018] Through preprocessing operations, the data preprocessing module expands the scale and diversity of the training dataset, enabling the wire harness instance segmentation algorithm to learn more representative features.

[0019] As a further solution of the present invention, the test module evaluates the quality of the wire harness by analyzing the segmentation results, including the following specific contents: the test module uses contour detection and connected component analysis to extract quality features from instance masks and spline points, and the quality features include the number of components in the wire harness, the integrity of the wiring path, the accuracy of the color, the accuracy of the position, and the correctness of the direction. Then, the test module matches the extracted quality features with the target parameters preset in the design software Capital one by one. If the actual extracted quality features are completely consistent with the target parameters, the test module will determine that the assembly quality of the wire harness is "OK", indicating that the assembly task has been completed correctly and the wire harness can enter the next production link. However, if there is any mismatch, the test module will be evaluated as "nOK", indicating that the quality is unqualified, and the problem will be visualized to guide rework.

[0020] The technical effects and advantages of the wire harness intelligent testing system based on a self-learning algorithm of the present invention: The present invention realizes the real-time capture and segmentation of the wire harness assembly status by deploying a high-resolution RGB camera at each workstation of the assembly line, combined with data preprocessing and image segmentation modules, and can timely monitor and evaluate the quality of the wire harness during the manufacturing process, reducing the high rework costs caused by delayed problem discovery. The present invention adopts a self-learning algorithm (DeeplabV3+ combined with MaskSlic algorithm) to expand the diversity of the training data set through data preprocessing and optimize the calculation process, eliminating the need to manually specify endpoints, shortening the execution time, and reducing computing resource requirements and deployment and maintenance costs. The present invention improves the accuracy, safety and reliability of automotive wire harness testing through automated and intelligent technical means. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 The figure is a structural diagram of a wiring harness intelligent testing system based on a self-learning algorithm according to the present invention.

[0022] Figure 2 The figure is a schematic diagram of the structure of a detection device based on STM32 in the prior art.

[0023] Figure 3 This is a structural diagram of the existing machine vision-based new energy vehicle wiring harness assembly detection system.

[0024] Figure 4 This is a schematic diagram of the structure of the wire harness instance segmentation algorithm proposed in this invention.

[0025] Figure 5 Schematic diagram of the superpixel processing result of the input image of the present invention. DETAILED DESCRIPTION

[0026] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0027] Example 1

[0028] See Figure 1 As shown in the structural schematic diagram, the present invention provides an intelligent wire harness testing system based on a self-learning algorithm, which includes a data acquisition module, a data preprocessing module, an image segmentation module and a testing module; the data acquisition module is used to acquire image data of automobile wire harnesses; the data preprocessing module preprocesses the acquired images; the image segmentation module performs instance segmentation on the wire harnesses in the image; and the testing module evaluates the quality of the wire harnesses by analyzing the segmentation results.

[0029] Furthermore, the data acquisition module is used to collect image data of automotive wiring harnesses, including: a high-resolution RGB camera is deployed at each workstation on the assembly line. The camera has a resolution of up to 4586×2663 pixels, which can capture the assembly status of the wiring harness in real time, ensuring the clarity and richness of the image. This high-resolution configuration enables the camera to capture every tiny component in the wiring harness, such as connectors, clips, and different types of wiring harnesses (such as fully taped wiring harnesses, spiral taped wiring harnesses, and non-taped wiring harnesses). The high-resolution RGB camera is deployed behind each workstation on the assembly line to capture the assembly status of the wiring harness in real time. During the data acquisition process, the 407 collected images are constructed into a data set, which includes simulations of different wiring harness configurations and deformation states.

[0030] Furthermore, the data preprocessing module preprocesses the captured images, including: image scaling, horizontal flipping, scaling enhancement, rotation enhancement, four-point perspective transformation, padding, contrast adjustment, brightness adjustment, gamma correction, blurring, and hue saturation adjustment. Image scaling adjusts the original high-resolution image (4586×2663 pixels) to a uniform size of 480×480 pixels. The purpose of image scaling is to reduce the computational complexity of image data; secondly, it facilitates the image segmentation module to batch process and extract features from image data, ensuring data consistency. Horizontal flipping simulates the appearance of wire harnesses in different directions by generating mirrored versions of the image. Scaling enhancement simulates the appearance of objects at different distances or viewing angles by adjusting the size of the wire harness in the image. Rotation enhancement simulates the angular changes that may occur during the assembly process of the wire harness by rotating the image. The four-point perspective transformation simulates the perspective changes of the image. Padding adds additional pixels to the image boundaries to avoid detection errors caused by boundary effects. The contrast adjustment, brightness adjustment and gamma correction simulate the image differences caused by changes in lighting conditions (such as time, location or equipment differences) in industrial environments by changing the brightness and contrast characteristics of the image. The blur processing helps the wire harness instance segmentation algorithm adapt to scenes with low image quality by introducing slight image noise or defocus effects, thereby improving its robustness in practical applications. The hue and saturation adjustment simulates wire harness components of different materials or colors by changing the color properties of the image, so that the wire harness instance segmentation algorithm can adapt to diverse wire harness appearances. Through preprocessing operations, the data preprocessing module expands the scale and diversity of the training data set, enabling the wire harness instance segmentation algorithm to learn more representative features.

[0031] Furthermore, the image segmentation module performs instance segmentation on the wire harness in the image, including: the image segmentation module performs instance segmentation on the wire harness in the image by using a wire harness instance segmentation algorithm to identify the wire harness. The wire harness instance segmentation algorithm combines deep learning and computer vision technology, does not require manual endpoint specification, shortens execution time, and aims to achieve efficient and reliable wire harness segmentation. The wire harness instance segmentation algorithm is as follows: Figure 4 As shown in the figure, we first use the deep convolutional neural network (DCNN) to perform semantic segmentation on the image and generate a binary mask M b , the deep convolutional neural network is specifically the DeeplabV3+ model. The binary mask marks the wiring harness area as white (foreground) and other areas as black (background). In order to train the DeeplabV3+ model, 407 car door wiring harness images were taken in a simulated industrial environment. After preprocessing operations, they were expanded to a data set of 30,000 synthetic images. The data set covers wiring harnesses of different shapes and colors to ensure that the DeeplabV3+ model can learn diverse wiring harness features. During the training process, the DeeplabV3+ model used ResNet-101 as the backbone network and was trained for 200 rounds with a batch size of 10. The optimizer used Adam and the learning rate was increased from 10 -6 Gradually adjust to 10 -9 .

[0032] like Figure 5 As shown in Figure 2, the line instance segmentation algorithm then uses the MaskSlic algorithm to perform superpixel segmentation on the image, with the goal of segmenting the image into sub-regions (i.e., superpixels) to simplify image representation and accelerate subsequent processing. The MaskSlic algorithm combines the binary mask M generated by semantic segmentation with the b , limiting the superpixel clustering to the line bundle area to avoid meaningless segmentation of the entire image. After the segmentation is completed, extract each superpixel R i The centroid coordinates of serve as a reference point for subsequent processing steps (e.g., B-spline interpolation).

[0033] The wire harness instance segmentation algorithm then constructs a region adjacency graph (RAG) from the superpixel segmentation results to manage the neighborhood relationships between superpixels and identify endpoints. This process is achieved by creating an undirected unweighted graph G = (V, E). If two superpixels are adjacent, an edge is added between their corresponding nodes. Where G represents the graph itself, which is a mathematical structure used to describe the relationship between a set of objects; V represents the vertex set, that is, the set of all superpixel vertices in the graph, and each vertex v iCorresponding to a superpixel in the image; E represents the edge set, i.e., the set of all edges in the graph. Neighborhood relationships are determined by applying morphological gradient operations to the superpixel mask to extract the boundary pixels of adjacent superpixels. After the graph is constructed, nodes are classified according to their degree (i.e., the number of neighbors): outliers are removed, endpoints mark the start or end of a bundle, line segments form part of the bundle's path, and intersections indicate where bundles cross or branch.

[0034] Subsequently, by calculating the degree of each node, nodes in the graph with a degree greater than 2, i.e., intersection nodes, are identified. Then, the intersection group consisting of one or more intersection nodes is replaced with an equivalent single node. The equivalent node can retain all the neighborhood relationships of the original group, thereby ensuring that key connection information is not lost during the simplification process. The simplified graph reduces the complexity of the wire harness instance segmentation algorithm, making path discovery more direct and efficient. In order to handle different wire harness segments, the connected component method is used to segment the graph into several clusters, each cluster being a subset of the graph. The clusters are divided into two categories: clusters without intersections and clusters with intersections. The clusters without intersections (shown in blue in the visualization) do not contain any intersection nodes and consist only of two endpoints, representing simple segments of the wire harness; while clusters with intersections (shown in yellow in the visualization) contain intersection nodes (marked in red in the visualization).

[0035] To assess the similarity of intersection neighbor pairs and guide subsequent path connections, the wire harness instance segmentation algorithm uses a deep convolutional neural network called TripletNet combined with curvature information to calculate similarity scores for intersection neighbor pairs and construct a prediction matrix W. TripletNet takes 32×32 pixel image patches surrounding the intersection neighbor nodes as input, generates an embedding representation using a ResNet8 feature extractor and a fully connected layer, and then optimizes using a triplet loss to calculate the Euclidean distance d between the two image patches. Simultaneously, a curvature score is generated using a Von Mises distribution based on the angular information of the node sequence. The final similarity score for each intersection neighbor pair is the product of the TripletNet score and the curvature score, which is then entered into the corresponding position of the prediction matrix W. If there is no intersection in the graph, the prediction matrix W is empty.

[0036] Next, the wire bundle instance segmentation algorithm begins path finding, aiming to discover the path for each wire bundle and maximize graph coverage. This approach involves starting from an endpoint and extending the path through neighbor relationships until it reaches another endpoint or an intersection neighbor. For clusters without intersections, the two endpoints are directly connected to form a complete path, which is then added to the path set. For clusters with intersections, partial paths are first created from the endpoints until an intersection neighbor node is encountered. These partial paths are stored in the path set and then connected using the highest-scoring pairs in the prediction matrix W to form a complete path.

[0037] Next, the wire harness instance segmentation algorithm uses CrossNet to perform path layout inference, predict the upper and lower relationships of the wire harness in the intersection area, and generate the correct instance mask M c . The CrossNet classifies the 64×64 pixel image blocks of the intersection area to determine which harness is on top. The CrossNet is used to predict the upper and lower relationships at the intersection and output a probability value for each possible configuration. By selecting the path with the highest probability as the upper harness, there is no need to set a threshold. At the same time, the harness instance segmentation algorithm establishes B-spline modeling, which approximates the path of the harness through a B-spline curve to generate a lightweight spline point. The B-spline modeling uses the centroid coordinates of superpixels to convert the node sequence defining the path into a set of 2D point sequences, and then fits a cubic B-spline curve to obtain a smooth representation of the harness trajectory. Subsequently, the B-spline curve is discretized into a fixed number of 2D points and represented in pixel coordinates.

[0038] Furthermore, the testing module evaluates the quality of the wiring harness by analyzing the segmentation results, including: the testing module uses contour detection and connected component analysis to extract quality features from instance masks and spline points. The quality features include the number of components in the wiring harness, the integrity of the wiring path, the accuracy of color, the accuracy of position, and the correctness of direction. The testing module then matches the extracted quality features one by one with the target parameters preset in the design software Capital. If the actual extracted quality features completely match the target parameters, the testing module will determine that the assembly quality of the wiring harness is "OK", indicating that the assembly task has been completed correctly and the wiring harness can enter the next production link. However, if there is any mismatch, such as insufficient component quantity, wiring path deviation from the design, color error or position deviation, the testing module will assess it as "nOK", indicating that the quality is unqualified, and the problem will be visualized to guide rework.

[0039] This invention deploys high-resolution RGB cameras at each workstation on the assembly line, combined with data preprocessing and image segmentation modules, to achieve real-time capture and segmentation of the wiring harness assembly state. This allows for timely monitoring and assessment of wiring harness quality during the manufacturing process, reducing the high rework costs caused by delayed problem discovery. The invention utilizes a self-learning algorithm (DeeplabV3+ combined with the MaskSlic algorithm) to expand the diversity of the training dataset through data preprocessing and optimize the computational process, eliminating the need for manual endpoint specification, shortening execution time, and reducing computing resource requirements and deployment and maintenance costs. Through automated and intelligent technical means, the invention improves the accuracy, safety, and reliability of automotive wiring harness testing.

[0040] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

[0041] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A wiring harness intelligent testing system based on a self-learning algorithm, comprising a data acquisition module and a data preprocessing module, wherein the data acquisition module is used to acquire image data of automobile wiring harnesses; the data preprocessing module preprocesses the acquired images, characterized in that: The invention also includes an image segmentation module and a testing module, wherein the image segmentation module performs instance segmentation on the wire harness in the image; and the testing module evaluates the quality of the wire harness by analyzing the segmentation results. The image segmentation module uses a wire harness instance segmentation algorithm to perform instance segmentation on the wire harness in the image and identify the wire harness; the wire harness instance segmentation algorithm uses the DeeplabV3+ model to perform semantic segmentation on the image and generate a binary mask M b ; Subsequently, the MaskSlic algorithm is used to perform superpixel segmentation in the harness area and extract the centroid coordinates of each superpixel; then, the harness instance segmentation algorithm constructs a regional adjacency graph, with each superpixel as a node and edges connecting adjacent superpixels, and is classified into endpoints, line segments and intersections according to the node degree; then, the connected component method is used to cluster the graph, dividing the graph into clusters without intersections and clusters with intersections. For the case of intersections, TripletNet combines curvature information to calculate the similarity scores of neighbor pairs and generates a prediction matrix W to guide path finding, extending the path from the endpoints and connecting complete wire segments; in addition, the harness instance segmentation algorithm uses CrossNet to classify the intersection area and infer the upper and lower relationships of the harness to generate the correct instance mask; finally, the path node sequence is fitted with a cubic B-spline curve and discretized into 2D points to provide a lightweight harness model.

2. The wire harness intelligent testing system based on self-learning algorithm according to claim 1 is characterized in that , the binary mask M b The harness area is marked as white and the other areas are marked as black. When training the DeeplabV3+ model, ResNet-101 was used as the backbone network, and 200 rounds of training were performed with a batch size of 10, an Adam optimizer, and a learning rate of 10. -6 Gradually adjust to 10 -9 .

3. The wire harness intelligent testing system based on self-learning algorithm according to claim 1 is characterized in that , the MaskSlic algorithm combines the binary mask M generated by semantic segmentation b , limiting the superpixel clustering to the line bundle region.

4. The wire harness intelligent testing system based on self-learning algorithm according to claim 1, characterized in that: The data acquisition module deploys RGB cameras at each workstation on the assembly line to capture the wiring harness assembly status in real time and construct a dataset containing 407 images covering different wiring harness configurations and deformation states.

5. The wire harness intelligent testing system based on self-learning algorithm according to claim 1, characterized in that: The data preprocessing module scales the image to 480×480 pixels, performs horizontal flipping, zoom enhancement, rotation enhancement, four-point perspective transformation, padding, contrast adjustment, brightness adjustment, gamma correction, blurring, and hue saturation adjustment, expanding the dataset to 30,000 synthetic images.

6. The wire harness intelligent testing system based on self-learning algorithm according to claim 1, characterized in that: After constructing the region adjacency graph, the line bundle instance segmentation algorithm extracts adjacent superpixel boundary pixels through morphological gradient operation to determine the neighborhood relationship, classifies nodes according to node degree, removes outliers, and marks endpoints, line segments and intersections.

7. The wire harness intelligent testing system based on self-learning algorithm according to claim 1, characterized in that: The proposed wire bundle instance segmentation algorithm replaces intersection groups consisting of intersection nodes with degree greater than 2 with equivalent single nodes, preserves neighborhood relationships, and segments the graph into clusters without intersections and clusters with intersections using a connected component method.

8. The wire harness intelligent testing system based on self-learning algorithm according to claim 1, characterized in that: The testing module uses contour detection and connected component analysis to extract quality features, including component count, wiring path integrity, color accuracy, positional accuracy, and directional correctness, and matches the quality features with target parameters in the design software Capital.

9. The wire harness intelligent testing system based on self-learning algorithm according to claim 1, characterized in that: The matching result of the test module is "OK" or "nOK". If the matching result is "nOK", it means that the quality is unqualified, and rework is guided by visualizing the problem.

Images