Generator circuit breaker box production line monitoring method and system

Abstract

The invention relates to the technical field of data processing, in particular to a generator circuit breaker box production line monitoring method and system. The method comprises the following steps: acquiring multi-modal data of each station on a production line, and extracting station fusion features; constructing a graph network, and obtaining a station state vector of each station containing the accumulated influence of the upstream process through the graph network; inputting the station state vectors into a prototype network and a reconstruction network trained by normal working condition data in parallel; and for any box body on the production line, the station state vectors corresponding to the stations are connected in series according to a production time sequence, a production quality evolution track of the box body is constructed, the evolution track is compared with a reference track generated based on a plurality of standard workpieces under a normal working condition, and a final quality evaluation index of the box body is obtained. The method has the effect of improving the quality monitoring precision.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to a method and system for monitoring a generator circuit breaker box production line. Background Art

[0002] Generator circuit breaker boxes are key protective equipment in power systems, and the manufacturing quality of their boxes is directly related to the safety and reliability of the entire equipment. Box production usually involves multiple processes such as cutting, bending, welding, spraying, and assembly. The process flow is complex and has extremely high requirements for dimensional accuracy, welding quality, and form and position tolerances. Traditional quality control methods mainly rely on manual inspections and offline spot checks, such as using calipers, three-dimensional coordinate measuring machines and other tools to measure finished or semi-finished products. This method is not only inefficient and has poor real-time performance, but the test results are easily affected by human factors, making it difficult to achieve full coverage and real-time monitoring of the production process. In addition, when defects are discovered in the final product, due to the lack of process data support, it is difficult to quickly and accurately trace the specific workstation and root cause of the problem, resulting in long problem location and process optimization cycles and high costs.

[0003] With the advancement of industrial automation and intelligent manufacturing technologies, some production lines have begun to incorporate machine vision and various sensors for online monitoring. For example, two-dimensional or three-dimensional vision systems are used to inspect critical dimensions or surface defects on boxes, or to indirectly assess processing quality by monitoring operating parameters (such as current, voltage, and speed) of welding equipment and robots. However, these methods still have significant limitations. First, they mostly rely on single-modal data for analysis, such as image data or equipment sensor data. This single dimension of information makes it difficult to fully reflect complex working conditions and subtle variations in product quality. Second, existing monitoring systems typically treat each workstation as an independent "information island," collecting and analyzing data separately, failing to effectively establish the inherent connections between workstations. This approach ignores the fact that processing deviations in upstream processes accumulate and propagate downstream, ultimately affecting the quality of the finished product. In summary, these technologies fail to capture the quality evolution process from a systematic perspective across the entire production line, nor do they easily identify unknown or sudden production anomalies caused by the combined impact of multiple processes. Summary of the Invention

[0004] The present application provides a method and system for monitoring a generator circuit breaker box production line to address the problems in related technologies of being unable to capture the quality evolution process from a systematic perspective of the entire production line, and also having difficulty identifying unknown or sudden production anomalies caused by the combined influence of multiple processes.

[0005] In a first aspect, the present application provides a method for monitoring a generator circuit breaker box production line, which adopts the following technical solution: A method for monitoring a generator circuit breaker box production line includes: acquiring multimodal data of each workstation on the production line, the multimodal data including a two-dimensional image of the production line, a three-dimensional point cloud image of the production line, and equipment operation sensor data; identifying preset key shape and position areas of the box in the two-dimensional image of the production line, extracting geometric features of the key shape and position areas, extracting working state features from the sensor data, and acquiring workstation fusion features based on the geometric features and working state features; constructing a graph network, taking each workstation of the production line as a node of the graph network, taking the physical flow relationship between the workstations as the directed edge between the nodes, and taking the workstation fusion features as the initial features of each node; acquiring the features including the key shape and position areas of the box in the two-dimensional image of the production line through the graph network; extracting the geometric features of the key shape and position areas of the box in the two-dimensional image of the production line, and extracting the geometric features of the key shape and position areas; extracting the working state features from the sensor data, and acquiring the workstation fusion features based on the geometric features and the working state features; constructing a graph network, taking each workstation of the production line as a node of the graph network, taking the physical flow relationship between the workstations as the directed edge between the nodes, and taking the workstation fusion features as the initial features of each node; obtaining the features including the key shape and position areas of the box in the two-dimensional image of the production line, and extracting the geometric features of ... The workstation state vectors of each workstation are cumulatively affected by the upstream processes; the workstation state vectors are input in parallel into the prototype network and the reconstruction network trained by normal working condition data. In response to the similarity of the workstation state vector with all known fault prototypes in the prototype network being lower than the first threshold, and its reconstruction error through the reconstruction network being higher than the second threshold, the current working condition is determined to be an unknown production abnormality; for any box on the production line, the workstation state vectors corresponding to each workstation are connected in series according to the production sequence, and its production quality evolution trajectory is constructed. The evolution trajectory is compared with the benchmark trajectory generated under normal working conditions based on multiple standard workpieces to obtain the final quality evaluation index of the box.

[0006] By integrating multimodal data such as two-dimensional images of the production line, three-dimensional point cloud images of the production line, and equipment sensors, the true production status of each workstation can be more comprehensively and accurately portrayed, overcoming the problem of one-sided information caused by relying on a single data source. Furthermore, by using the entire production line to construct a graph network model, the inherent correlation between the various workstations and the cumulative impact of upstream processes on downstream processes are effectively revealed, breaking down the barriers of "information islands" between workstations and achieving overall control of the production process from a holistic perspective. In addition, through the collaborative work of the prototype network and the reconstruction network, unknown production anomalies that have not been preset can be effectively identified, improving the coverage and reliability of fault monitoring. Constructing a quality evolution trajectory throughout the entire process for a single box and performing quantitative comparison with the baseline trajectory can not only provide an accurate final quality assessment, but also provide clear and complete process data support for tracing the root causes of product quality problems.

[0007] Optionally, the key shape and position areas include the welding seam area and / or the mounting hole area on the box; the geometric features include the height, width, flatness of the weld and / or the diameter, roundness, and center position deviation of the mounting hole; the working condition characteristics include the statistical characteristics of the equipment operating parameters during the processing cycle.

[0008] Extracted geometric features, such as weld height, width, and flatness, and mounting hole diameter, roundness, and center deviation, are common dimensional and shape control parameters used in current industrial inspections. These parameters can be extracted with high precision through production line 3D point cloud imaging and image processing, and also have clear engineering standards as quality criteria.

[0009] Optionally, the step of obtaining the workstation fusion feature based on the geometric feature and the working condition feature includes: splicing the geometric feature and the working condition feature to obtain an initial feature vector, and using a neural network to fuse and reduce the dimension of the initial feature vector to obtain the workstation fusion feature.

[0010] The splicing method unifies features from different sources into a high-dimensional vector, ensuring that the information retains the maximum original expressiveness before encoding; then a neural network is used to reduce the dimension and improve the compactness of the features.

[0011] Optionally, the step of obtaining the workstation state vector of each workstation that includes the cumulative impact of upstream processes includes: for any node, aggregating the workstation state vector of the node in the previous iteration with the workstation state vectors of all upstream adjacent workstation nodes to obtain an intermediate vector, and using a nonlinear activation function to process the intermediate vector to obtain the latest workstation state vector.

[0012] By considering the characteristic states of this node and upstream nodes, we can accurately reflect the "transmission effect of upstream quality on downstream quality," such as the potential for welding errors caused by poor stamping accuracy. This aggregation operation preserves the process logic relationship between each station, forming a chain feedback mechanism.

[0013] Optionally, the cosine similarity is used as the similarity between the workstation state vector and the known fault prototype.

[0014] Optionally, the reconstruction network is an autoencoder network, and the reconstruction error is the mean square error between the input workstation state vector and the vector reconstructed by the autoencoder network.

[0015] Optionally, an average result of production quality evolution trajectories generated for multiple standard workpieces under normal working conditions is used as the reference trajectory.

[0016] It reduces the accidental error caused by a single reference sample and can better represent the stable trend of the standard production process.

[0017] Optionally, a weighted sum of the dynamic time warping distance and trajectory shape difference between the evolved trajectory and the reference trajectory is used as the final quality assessment metric.

[0018] In a second aspect, the present application provides a generator circuit breaker box production line monitoring system, which adopts the following technical solutions: A generator circuit breaker box production line monitoring system includes: a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the generator circuit breaker box production line monitoring method described above is implemented.

[0019] The above-mentioned method for monitoring a production line of a generator circuit breaker box is generated into a computer program and stored in a memory so as to be loaded and executed by a processor. Thus, a system is manufactured based on the memory and the processor for easy use.

[0020] This application has the following technical effects: By integrating multimodal data, graph neural network modeling, dual-path anomaly recognition and trajectory comparison evaluation, a production line monitoring system with systematic, real-time and accurate characteristics has been constructed, effectively solving problems such as "information islands", "lack of full-chain tracking" and "inability to identify sudden anomalies" in existing technologies. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 This is a method flow chart of a generator circuit breaker box production line monitoring method according to an embodiment of the present application. DETAILED DESCRIPTION

[0022] The present application embodiment discloses a method for monitoring a generator circuit breaker box production line, referring to Figure 1 , including steps S1 to S5.

[0023] S1: Acquire multimodal data of each workstation on the production line. The multimodal data includes two-dimensional images of the production line, three-dimensional point cloud images of the production line, and equipment operation sensor data.

[0024] A high-resolution industrial camera is installed above the equipment corresponding to each process step. It captures the top surface of each box as it passes, creating a 2D image of the production line with a resolution of 4096 x 3072 pixels. This image data clearly indicates whether there are scratches on the box surface and whether the positioning is accurate. Simultaneously, a laser profile measurement device is integrated into the end of the welding robot. It scans the weld area after completion, collecting 3D point cloud data at a rate of 500 points per second. This accurately constructs the weld contour and creates a 3D point cloud image of the production line.

[0025] Furthermore, during the welding process, sensors collect real-time data on the welding power supply's current and voltage. For example, they record an average current of 200 amps and a voltage of 24 volts for a particular weld. Simultaneously, a torque sensor mounted on the robot's wrist records changes in force during the welding task, such as a peak contact force of 50 Newtons. These electrical and mechanical parameters together form a data stream reflecting the equipment's immediate operating status, providing a process basis for subsequent quality analysis, ultimately resulting in operational sensor data.

[0026] S2: Identify the preset key shape and position areas of the box in the two-dimensional image of the production line, extract the geometric features of the key shape and position areas, extract the working condition features in the sensor data, and obtain the work station fusion features based on the geometric features and working condition features.

[0027] Specifically, a target detection model such as YOLOv5 is used to process the two-dimensional image of the production line to automatically locate key shape and position areas such as welds, bending corners or mounting holes on the box; then, the two-dimensional area coordinates in the two-dimensional image of the production line are mapped to the three-dimensional point cloud of the three-dimensional point cloud image of the production line, and geometric features such as the width and height of the weld, the angle of the bend, and the flatness are calculated based on the three-dimensional point cloud; for the equipment operation sensor data, a sliding time window is used to extract statistical values ​​such as the mean, variance, and kurtosis in each window as working condition status features; the geometric feature vector corresponding to the geometric feature and the working condition status feature vector corresponding to the working condition status feature are spliced ​​to obtain the initial feature vector, and the initial feature vector is input into a multi-layer perceptron MLP for dimensionality reduction and deep fusion, and the output result is the work station fusion feature.

[0028] Specifically, when processing a box, the key geometric features of the weld were extracted from the 3D contour data, and the average weld height was calculated to be 1.5 mm, the average width was 4 mm, and the flatness was 0.2 mm. Simultaneously, the characteristics of a key mounting hole were analyzed from the 2D image of the production line, revealing its diameter to be 10.1 mm, its roundness to be 0.995, and its center position to be offset by 0.15 mm from the design reference point. These values ​​together constitute a geometric feature vector.

[0029] At the same time, the equipment operating parameters within the workstation processing cycle were analyzed, and the average welding current was calculated to be 198 amperes, with a standard deviation of 5 amperes, and the maximum end-of-line torque of the robot was 45 Newtons per meter. These statistical values ​​constitute the working condition state feature vector. The geometric feature vector and the working condition state feature vector are concatenated to form a higher-dimensional initial feature vector. This initial feature vector is then input into a pre-trained three-layer fully connected neural network for compression. For example, the network compresses it from 10 dimensions to 4 dimensions, outputting a low-dimensional workstation fusion feature that contains both product geometry and equipment state information. This fusion feature more compactly represents the overall quality status of the current workstation.

[0030] S3: Construct a graph network, with each workstation on the production line as a node of the graph network, the physical flow relationship between the workstations as the directed edges between the nodes, and the workstation fusion features as the initial features of each node; obtain the workstation state vector of each workstation including the cumulative impact of the upstream process through the graph network.

[0031] A graph network is constructed based on various information in the production line. For example, a production line including three processes: cutting, bending, and welding can construct a directed graph containing three nodes, where the cutting node points to the bending node, and the bending node points to the welding node; the station fusion features obtained in the previous steps of each station are used as the initial features of the corresponding nodes; the graph attention network GAT model is used, and through its built-in attention mechanism, different weights are assigned to features from different upstream stations during the message transmission process. The aggregated features can reflect the impact of upstream processing quality on the current station. The node features output after calculation by the multi-layer graph network are the station state vectors of each station including the cumulative impact.

[0032] In an optional embodiment, the graph network is a graph convolutional network; The steps of obtaining the workstation state vector of each workstation including the cumulative impact of upstream processes include: for any node, aggregating the workstation state vector of the node in the previous iteration with the workstation state vectors of all upstream adjacent workstation nodes to obtain an intermediate vector, and using a nonlinear activation function to process the intermediate vector to obtain the latest workstation state vector.

[0033] Specifically, the entire production line is abstracted as a directed graph, where workstations such as stamping, welding, and painting are nodes in the graph. The direction of material flow defines the connections between nodes. For example, the welding station is a downstream node from the stamping station. A graph convolutional network operates on the production line graph structure, with each node maintaining a station state vector representing the current quality status of the workpiece at that station.

[0034] When updating information, taking the welding station as an example, the calculation of its new station state vector not only depends on its own state at the previous moment, but also aggregates the station state vector from its upstream stamping station.

[0035] The step of aggregating the station state vector of the node in the previous iteration with the station state vectors of all upstream adjacent station nodes to obtain the intermediate vector can be: it can be a weighted sum, for example, multiplying the old station state vector of the welding station itself by a weight of 0.6, and then multiplying the station state vector of the stamping station by a weight of 0.4 to obtain the intermediate vector.

[0036] The aggregated results are then processed through a nonlinear activation function, such as a ReLU, to generate the updated welding station state vector. This process simulates how the processing quality of upstream processes affects the quality status of downstream processes, enabling the cross-station propagation and accumulation of quality information.

[0037] S4: The workstation state vector is input in parallel into the prototype network and the reconstruction network trained by normal working condition data. In response to the similarity of the workstation state vector in the prototype network with all known fault prototypes being lower than the first threshold, and its reconstruction error through the reconstruction network being higher than the second threshold, the current working condition is determined to be an unknown production anomaly.

[0038] The prototype network uses cosine similarity to calculate the similarity between the workstation state vector and the known fault prototype; the reconstruction network is an autoencoder network, and the reconstruction error is the mean square error between the input workstation state vector and the vector reconstructed by the autoencoder network.

[0039] Specifically, the prototype network pre-stores a variety of known faults calculated through historical data, such as weld penetration and dimensional deviation. The corresponding feature vector center is the prototype. The Euclidean distance between the current workstation state vector and each fault prototype is calculated to determine whether it is a known fault; the reconstruction network uses a variational autoencoder (VAE) trained only by the workstation state vector under normal working conditions. The network learns the distribution law of the normal state. When the input workstation state vector is significantly different from the normal state, the mean square error (MSE) between its reconstructed output and the original input will increase significantly, thereby identifying the anomaly; only when a workstation state vector does not match any known fault prototype and its reconstruction error is very high, it will be judged as a new, unknown production anomaly.

[0040] For example, the prototype network includes a prototype vector for a cold weld fault caused by low welding current. When a new workpiece generates its station state vector at a welding station, the cosine similarity between the station state vector and the cold weld fault prototype vector is calculated. If the calculated similarity score is 0.92, which is very close to 1, it is determined that the current workpiece is highly likely to have a cold weld fault.

[0041] The reconstruction network uses an autoencoder trained only on workstation state vectors from normal workpieces. When fed a workstation state vector from a normal production workpiece, the autoencoder effectively compresses and decompresses it, reconstructing a vector that is nearly identical to the input vector, with a mean squared error (MSE) of only 0.05. However, if fed a workstation state vector representing an unknown anomaly or a new type of fault, the network, having not learned this pattern, will perform poorly, resulting in a significant discrepancy between the reconstructed vector and the input vector. The calculated MSE can reach as high as 2.8, effectively identifying this as a new, previously unseen anomaly.

[0042] S5: For any box on the production line, the state vectors corresponding to each workstation are connected in series according to the production sequence to construct its production quality evolution trajectory. The evolution trajectory is compared with the baseline trajectory generated under normal working conditions based on multiple standard workpieces to obtain the final quality evaluation index of the box.

[0043] For any box, the station state vectors S1, S2, ..., SN generated at station 1, station 2 to the last station N are arranged in sequence to form a multidimensional time series, which is the quality evolution trajectory of the box; the evolution trajectories of hundreds of qualified standard parts are collected in advance, and a standard reference trajectory is calculated using methods such as dynamic time warping barycenter average (DBA); the time alignment similarity between the current box trajectory and the reference trajectory, i.e., the dynamic time warping distance, is calculated using the dynamic time warping (DTW) algorithm; the Fréchet distance of the two trajectories is simultaneously calculated as the trajectory shape difference; the weight coefficients W1 and W2 are set, and the final quality evaluation index is defined as W1 multiplied by the dynamic time warping distance plus W2 multiplied by the trajectory shape difference.

[0044] In an optional embodiment, the reference trajectory is obtained by averaging the production quality evolution trajectories generated by multiple standard workpieces under normal working conditions; the final quality evaluation index is calculated based on the weighted sum of the dynamic time warping distance and the trajectory shape difference.

[0045] For example, 50 standard workpieces, confirmed to be free of defects, can be selected and run through the entire production line. Each workpiece, after passing through all workstations, generates a production quality evolution trajectory consisting of the state vectors of each workstation. These 50 trajectories are aligned and averaged point by point to produce a baseline trajectory that represents the ideal production process.

[0046] When a new workpiece is completed and generates its own quality evolution trajectory, two calculations are performed. First, the dynamic time warping distance between the new trajectory and the baseline trajectory is calculated to obtain a value such as 4.5, which reflects the degree of matching between the two trajectories in the time series. Second, the shape difference between the two trajectories is calculated, for example, using the Fréchet distance, to obtain a value such as 3.0. According to the preset weights, for example, the weight of the dynamic time warping distance is 0.7, and the weight of the shape difference is 0.3, the final quality assessment index is calculated as 4.5 times 0.7 plus 3.0 times 0.3, which is 4.05. The final score comprehensively quantifies the overall deviation of the current workpiece production process from the ideal process.

[0047] An embodiment of the present application further discloses a generator circuit breaker box production line monitoring system, comprising a processor and a memory, wherein the memory stores computer program instructions. When the computer program instructions are executed by the processor, a generator circuit breaker box production line monitoring method according to the present application is implemented.

[0048] The above system also includes other components well known to those skilled in the art, such as a communication bus and a communication interface. The configuration and functions of these components are known in the art and will not be described in detail here.

[0049] The above are all preferred embodiments of the present application, and are not intended to limit the scope of protection of the present application. Therefore, any equivalent changes made based on the structure, shape, and principle of the present application should be included in the scope of protection of the present application.

Claims

1. A method for monitoring a generator circuit breaker box production line, characterized in that: Acquire multimodal data of each workstation on the production line, including two-dimensional images of the production line, three-dimensional point cloud images of the production line, and equipment operation sensor data; identify the preset key shape and position areas of the box in the two-dimensional image of the production line, and extract the geometric features of the key shape and position areas, extract the working condition features in the sensor data, and obtain the workstation fusion features based on the geometric features and working condition features; construct a graph network, use each workstation of the production line as a node of the graph network, use the physical flow relationship between the workstations as the directed edge between the nodes, and use the workstation fusion features as the initial features of each node; obtain the cumulative impact of each workstation on the upstream process through the graph network. The workstation state vector of each position is input into the prototype network and the reconstruction network trained by normal working condition data in parallel. In response to the similarity of the workstation state vector with all known fault prototypes in the prototype network being lower than the first threshold, and the reconstruction error of the workstation state vector through the reconstruction network being higher than the second threshold, the current working condition is determined to be an unknown production abnormality. For any box on the production line, the workstation state vectors corresponding to each workstation are connected in series according to the production sequence to construct its production quality evolution trajectory. The evolution trajectory is compared with the benchmark trajectory generated under normal working conditions based on multiple standard workpieces to obtain the final quality evaluation index of the box.

2. A method for monitoring a generator circuit breaker box production line according to claim 1, characterized in that: The key geometric areas include the welding seam area and / or the mounting hole area on the box; the geometric features include the height, width, flatness of the weld and / or the diameter, roundness, and center position deviation of the mounting hole; the working condition characteristics include the statistical characteristics of the equipment operating parameters during the processing cycle.

3. A method for monitoring a generator circuit breaker box production line according to claim 1, characterized in that: The steps of obtaining workstation fusion features based on geometric features and working condition features include: splicing the geometric features and working condition features to obtain an initial feature vector, and using a neural network to fuse and reduce the dimension of the initial feature vector to obtain the workstation fusion features.

4. A method for monitoring a generator circuit breaker box production line according to claim 1, characterized in that: The graph network is a graph convolutional network.

5. A method for monitoring a generator circuit breaker box production line according to claim 1, characterized in that: The steps of obtaining the workstation state vector of each workstation including the cumulative impact of upstream processes include: for any node, aggregating the workstation state vector of the node in the previous iteration with the workstation state vectors of all upstream adjacent workstation nodes to obtain an intermediate vector, and using a nonlinear activation function to process the intermediate vector to obtain the latest workstation state vector.

6. A method for monitoring a generator circuit breaker box production line according to claim 1, characterized in that: The cosine similarity is used as the similarity between the workstation state vector and the known fault prototype.

7. A method for monitoring a generator circuit breaker box production line according to claim 1, characterized in that: The reconstruction network is an autoencoder network, and the reconstruction error is the mean square error between the input workstation state vector and the vector reconstructed by the autoencoder network.

8. A method for monitoring a generator circuit breaker box production line according to claim 1, characterized in that: The average result of the production quality evolution trajectories generated for multiple standard workpieces under normal working conditions is used as the benchmark trajectory.

9. A method for monitoring a generator circuit breaker box production line according to claim 1, characterized in that: The weighted sum of the dynamic time warping distance and trajectory shape difference between the evolved trajectory and the benchmark trajectory is used as the final quality evaluation indicator.

10. A generator circuit breaker box production line monitoring system, characterized in that: include: A processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, a method for monitoring a generator circuit breaker box production line according to any one of claims 1 to 9 is implemented.

Images