A method and system for monitoring a generator circuit breaker tank production line

Abstract

The application 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 the production line and extracting station fusion features; a graph network is constructed, and station state vectors of each station containing the cumulative influence of upstream processes are acquired through the graph network; the station state vectors are input into a prototype network and a reconstruction network trained by normal working condition data in parallel; for any box on the production line, the station state vectors corresponding to the box at each station are concatenated according to the production time sequence, the production quality evolution track of the box is constructed, the evolution track is compared with a reference track generated based on multiple standard workpieces under normal working conditions, and the final quality evaluation index of the box is acquired. The application has the effect of improving the quality monitoring precision.

Description

TECHNICAL FIELD

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

[0002] As a key protection device in the power system, the quality of the generator circuit breaker box directly relates to the safety and reliability of the entire device. The box production usually involves multiple processes such as blanking, bending, welding, spraying, and assembly, and the process flow is complex, with high requirements for dimensional accuracy, welding quality, and geometric tolerance. The traditional quality control method mainly relies on manual inspection and offline sampling inspection, such as using a caliper, a three-coordinate measuring machine, and other tools to measure the finished product or semi-finished product. This method is not only inefficient and real-time, but also the detection 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 found in the final product, it is difficult to quickly and accurately trace the specific workstation and root cause of the problem due to the lack of process data, resulting in long problem positioning and process optimization cycle and high cost.

[0003] With the development of industrial automation and intelligent manufacturing technology, some production lines have begun to introduce machine vision and various sensors for online monitoring. For example, a two-dimensional or three-dimensional vision system is used to detect the key dimensions or surface defects of the box, or the running parameters (such as current, voltage, and speed) of the welding equipment and robots are monitored to indirectly judge the processing quality. However, these methods still have significant limitations. On the one hand, they mostly use single-mode data for analysis, such as relying only on image data or only on equipment sensor data, with single information dimension, making it difficult to fully reflect the subtle changes in complex working conditions and product quality. On the other hand, existing monitoring systems usually treat each workstation as an independent "information island", separately collecting and analyzing data, and fail to effectively establish the internal relationship between workstations. This mode ignores the objective law that the processing deviation of the upstream process will continuously accumulate and pass down to the downstream, eventually affecting the quality of the finished product. In summary, the related art cannot capture the quality evolution process from a systematic perspective of the entire production line, and it is also difficult to identify unknown or sudden production abnormalities caused by the comprehensive influence of multiple processes. SUMMARY

[0004] The present application provides a generator circuit breaker box production line monitoring method and system to solve the problem that the related art cannot capture the quality evolution process from a systematic perspective of the entire production line, and it is also difficult to identify unknown or sudden production abnormalities caused by the comprehensive influence of multiple processes.

[0005] In a first aspect, the present application provides a generator circuit breaker box production line monitoring method, which adopts the following technical solution:

[0006] A generator circuit breaker box production line monitoring method, comprising: acquiring multi-modal data of each station on the production line, the multi-modal data comprising production line two-dimensional images, production line three-dimensional point cloud images and device operation sensor data; identifying a preset key shape and position area of the box in the production line two-dimensional image, and extracting geometric features of the key shape and position area, extracting working condition state features in the sensor data, and acquiring station fusion features based on the geometric features and the working condition state features; constructing a graph network, taking each station of the production line as a node of the graph network, taking the physical flow relationship between the stations as a directed edge between the nodes, and taking the station fusion features as the initial features of each node; acquiring a station state vector of each station containing the cumulative influence of the upstream process through the graph network; inputting the station state vector into a prototype network and a reconstruction network trained by normal working condition data in parallel, and in response to the similarity of the station state vector in the prototype network with all known fault prototypes being lower than a first threshold value and the reconstruction error of the station state vector through the reconstruction network being higher than a second threshold value, determining the current working condition as an unknown production abnormality; for any box on the production line, concatenating the station state vectors corresponding to the box at each station in the production sequence to construct a production quality evolution track of the box, comparing the evolution track with a reference track generated based on a plurality of standard workpieces under normal working conditions, and obtaining a final quality evaluation index of the box.

[0007] By fusing multi-modal data such as production line two-dimensional images, production line three-dimensional point cloud images and device sensors, the real production state of each station can be more comprehensively and accurately described, and the problem of information fragmentation caused by relying on a single data source is overcome. Further, a graph network model is constructed for the entire production line, effectively revealing the internal correlation between the stations and the cumulative influence of the upstream process on the downstream, breaking down the barriers of the "information island" between the stations, 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 abnormalities that have not been preset can be effectively identified, improving the coverage and reliability of fault monitoring. For a single box, a quality evolution track throughout the entire process is constructed, and a quantitative comparison is made with the reference track, which not only gives an accurate final quality evaluation, but also provides clear and complete process data support for tracing the root cause of product quality problems.

[0008] Optionally, the key shape and position area comprises a welding seam area and / or a mounting hole area on the box; the geometric features comprise the height, width and flatness of the welding seam and / or the diameter, roundness and center position deviation of the mounting hole; and the working condition state features comprise statistical features of device operation parameters within a processing period.

[0009] The extracted geometric features such as the height, width and flatness of the weld, the diameter, roundness and center position deviation of the mounting hole are the commonly used size control and shape control parameters in current industrial detection. These parameters can be extracted with high precision through the three-dimensional point cloud image of the production line and image processing, and have clear engineering standards as quality criteria.

[0010] Optionally, the step of obtaining the station fusion feature based on the geometric feature and the working condition state feature comprises: splicing the geometric feature and the working condition state feature to obtain an initial feature vector, using a neural network to fuse and reduce the dimension of the initial feature vector, and obtaining the station fusion feature.

[0011] The splicing method unifies features from different sources into a high-dimensional vector, ensuring that the information retains the maximum original expression ability before encoding; then a neural network is used for dimension reduction to improve feature compactness.

[0012] Optionally, the step of obtaining the station state vector of each station containing the cumulative influence of the upstream process comprises: for any node, aggregating the station state vector of the node in the last iteration and the station state vectors of all upstream adjacent nodes to obtain an intermediate vector, and using a nonlinear activation function to process the intermediate vector to obtain the latest station state vector.

[0013] By considering the feature state of the current node and the upstream node, the "upstream quality transmission effect on downstream quality" can be accurately reflected, such as the possibility of welding error caused by poor stamping accuracy. This aggregation operation preserves the process logic relationship between stations, forming a chain feedback mechanism.

[0014] Optionally, the cosine similarity is used as the similarity between the station state vector and the known fault prototype.

[0015] Optionally, the reconstruction network is an autoencoder network, and the reconstruction error is the mean square error between the input station state vector and the vector reconstructed by the autoencoder network.

[0016] Optionally, the average result of the production quality evolution trajectory generated under normal working conditions of multiple standard workpieces is used as the reference trajectory.

[0017] This reduces the accidental error caused by a single reference sample, and can better represent the stable trend of the standard production process.

[0018] Optionally, the weighted sum of the dynamic time warping distance and the trajectory shape difference between the evolution trajectory and the reference trajectory is used as the final quality evaluation index.

[0019] In a second aspect, the application provides a generator circuit breaker box production line monitoring system, which adopts the following technical scheme:

[0020] The generator circuit breaker box production line monitoring system comprises a processor and a memory, and the memory stores computer program instructions which, when executed by the processor, implement the generator circuit breaker box production line monitoring method according to the above.

[0021] The computer program of the generator circuit breaker box production line monitoring method is generated and stored in the memory to be loaded and executed by the processor, so that the system is convenient to use according to the memory and the processor.

[0022] The present application has the following technical effects:

[0023] The fusion of multi-modal data, graph neural network modeling, dual-path anomaly identification and trajectory comparison evaluation constructs a production line monitoring system with systematization, real-time and accuracy, effectively solving the problems of "information island", "lack of full-chain tracking" and "inability to identify sudden anomalies" in the prior art. BRIEF DESCRIPTION OF DRAWINGS

[0024] Figure 1 It is a method flowchart of the generator circuit breaker box production line monitoring method of the embodiment of the present application. DETAILED DESCRIPTION

[0025] The embodiment of the present application discloses a generator circuit breaker box production line monitoring method, referring to Figure 1 , comprising steps S1-S5.

[0026] S1: Obtain multi-modal data of each station on the production line, and the multi-modal data comprises production line two-dimensional images, production line three-dimensional point cloud images and device running sensor data.

[0027] An industrial camera with high resolution is installed above the corresponding device of each process, which is used to shoot the top image of each box when it passes, forming a production line two-dimensional image, and the image resolution is 4096 by 3072 pixels. These image data can clearly reflect whether there are scratches on the surface of the box or whether the positioning is accurate. At the same time, a laser profile measurement device is integrated at the end of the welding robot, which will scan the weld area after welding to collect three-dimensional point cloud data at a frequency of five hundred points per second, thereby accurately constructing the profile of the weld to obtain a production line three-dimensional point cloud image.

[0028] In addition, during the welding process, the sensor collects current and voltage data of the welding power supply in real time, for example, recording that the average current of a certain welding is 200 amperes and the voltage is 24 volts. At the same time, the torque sensor installed on the wrist of the robot arm records the force change during the execution of the welding task, such as the contact force peak of 50 Newtons. These electrical and mechanical parameters together constitute a data stream reflecting the real-time working condition of the equipment, providing a process basis for subsequent quality analysis, and thus obtaining the equipment operation sensor data.

[0029] S2: Identify the preset key shape and position area of the box in the production line two-dimensional image, extract the geometric features of the key shape and position area, extract the working condition state features in the sensor data, and obtain the station fusion features based on the geometric features and the working condition state features.

[0030] Specifically, a target detection model such as YOLOv5 is used to process the production line two-dimensional image, and the key shape and position areas such as welds, bending angles or mounting holes on the box are automatically located; then, the two-dimensional region coordinates in the production line two-dimensional image are mapped to the three-dimensional point cloud in the production line three-dimensional point cloud image, and the width and height of the weld, the angle of the bending, the flatness and other geometric features are calculated based on the three-dimensional point cloud; for the equipment operation sensor data, a sliding time window is used to extract the mean, variance, kurtosis and other statistical values in each window as the working condition state features; the geometric feature vector corresponding to the geometric features and the working condition state feature vector corresponding to the working condition state features are spliced to obtain an initial feature vector, and the initial feature vector is input into a multi-layer perception MLP for dimension reduction and deep fusion. The output result is the station fusion feature.

[0031] Specifically, when processing a box, the key geometric features of the weld are extracted from the three-dimensional contour data, and the average height of the weld is calculated to be 1.5 mm, the average width is 4 mm, and the flatness is 0.2 mm. At the same time, the features of a key mounting hole are analyzed from the production line two-dimensional image, and the diameter is obtained to be 10.1 mm, the roundness is 0.995, and the center position is offset by 0.15 mm from the design reference point. These numerical values together constitute a geometric feature vector.

[0032] Meanwhile, the equipment operating parameters in the processing cycle of the analysis station are analyzed, and the average value of the welding current is calculated as 198 amperes, the standard deviation is 5 amperes, and the maximum value of the robot end torque is 45 N per meter. These statistical values constitute the working condition state feature vector. The above-mentioned geometric feature vector and the working condition state feature vector are spliced to form an initial feature vector of higher dimension. The initial feature vector is 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, and outputs a low-dimensional station fusion feature containing both product geometric information and equipment state information. The station fusion feature can more compactly represent the comprehensive quality status of the current station.

[0033] S3: Construct a graph network, take each station of the production line as a node of the graph network, take the physical flow relationship between the stations as a directed edge between the nodes, and take the station fusion feature as the initial feature of each node; obtain the station state vector of each station containing the cumulative influence of the upstream processes through the graph network.

[0034] Based on the information in the production line, a graph network is constructed, for example, a production line including three processes of unloading, bending and welding can construct a directed graph including three nodes, in which the unloading node points to the bending node, and the bending node points to the welding node. The station fusion feature obtained in the foregoing step is taken as the initial feature of the corresponding node. The graph attention network GAT model is used, and through the built-in attention mechanism, different weights are assigned to the features from different upstream stations in the message passing process. The aggregated features can reflect the influence of upstream processing quality on the current station. The node feature output after multi-layer graph network calculation is the station state vector of each station containing the cumulative influence.

[0035] In an optional embodiment, the graph network is a graph convolution network.

[0036] The step of obtaining the station state vector of each station containing the cumulative influence of the upstream processes includes: for any node, aggregating the station state vector of the node in the last iteration and the station state vectors of all upstream adjacent station nodes to obtain an intermediate vector, and using a nonlinear activation function to process the intermediate vector to obtain the latest station state vector.

[0037] Specifically, the entire production line is abstracted as a directed graph, in which the stamping, welding, painting and other stations are nodes in the graph, and the direction of material flow defines the connection relationship between the nodes. For example, the welding station is a downstream node of the stamping station. The graph convolution network runs on the production line graph structure, and each node maintains a station state vector representing the quality state of the workpiece at the station.

[0038] When the information is updated, taking the welding station as an example, the calculation of the new station state vector of the welding station depends not only on its own state at the previous moment, but also aggregates the station state vector from the upstream stamping station.

[0039] The step of aggregating the station state vector of the node at the station state vector of all upstream adjacent station nodes in the previous iteration to obtain the intermediate vector can be a weighted sum, such as multiplying the old station state vector of the welding station by a weight of 0.6, and adding the station state vector of the stamping station multiplied by a weight of 0.4 to obtain the intermediate vector.

[0040] The aggregated result is then processed by a nonlinear activation function such as ReLU to generate the updated station state vector of the welding station. The above process simulates how the processing quality of the upstream process affects the quality state of the downstream process, and realizes the cross-station propagation and accumulation of quality information.

[0041] S4: input the station state vector into the prototype network and the reconstruction network trained by normal working condition data in parallel, and determine the current working condition as an unknown production anomaly in response to the similarity of the station state vector to all known fault prototypes in the prototype network being lower than a first threshold value, and the reconstruction error of the station state vector through the reconstruction network being higher than a second threshold value.

[0042] The prototype network calculates the similarity of the station state vector to the known fault prototype using cosine similarity; the reconstruction network is an autoencoder network, and the reconstruction error is the mean square error between the input station state vector and the vector reconstructed by the autoencoder network.

[0043] Specifically, the prototype network pre-stores a plurality of known faults, such as weld-through and size out-of-tolerance, calculated by historical data, and the corresponding feature vector center, i.e. prototype, judges whether it belongs to a known fault by calculating the Euclidean distance between the current station state vector and each fault prototype; the reconstruction network uses a variational autoencoder VAE trained only by station state vectors in normal working conditions, and the network learns the distribution rule of the normal state. When the input station state vector is significantly different from the normal state, the mean square error MSE between the reconstruction output and the original input will significantly increase, thereby identifying the anomaly; only when a station state vector does not match any known fault prototype and its reconstruction error is high, it is determined to be a new, unknown production anomaly.

[0044] For example, in the prototype network, a virtual welding failure prototype vector is included due to too low welding current. After a new workpiece generates its station state vector at the welding station, the cosine similarity between the station state vector and the virtual welding failure prototype vector is calculated. If the calculated similarity score is 0.92, which is a high score very close to 1, it is judged that the current workpiece is most likely to have a virtual welding failure.

[0045] For the reconstruction network, an autoencoder trained only by station state vectors of normal workpieces is used. When a station state vector from a normal production workpiece is input, the autoencoder can well compress and decompress it, and the reconstructed vector is almost identical to the input vector, and the mean square error between the two may be only 0.05. However, if a station state vector representing an unknown abnormality or a new type of failure is input, since the network has not learned this mode, its reconstruction effect will be poor, resulting in a large difference between the reconstructed vector and the input vector, and the calculated mean square error may be as high as 2.8, thereby effectively identifying that it is a new type of abnormality that has never been seen before.

[0046] S5: For any box on the production line, concatenate its station state vectors corresponding to each station in the production time sequence to construct its production quality evolution trajectory, compare the evolution trajectory with a reference trajectory based on a plurality of standard workpieces generated under normal working conditions, and obtain the final quality evaluation index of the box.

[0047] For any box, arrange the station state vectors S1, S2,..., SN generated at station 1, station 2, and the last station N in order to form a multi-dimensional time sequence, which is the quality evolution trajectory of the box; a standard reference trajectory is calculated by, for example, the dynamic time warping centroid average DBA method, by pre-collecting the evolution trajectories of hundreds of qualified standard pieces; the time alignment similarity between the current box trajectory and the reference trajectory is calculated by the dynamic time warping DTW algorithm, i.e. the dynamic time warping distance; the Frechet distance between the two trajectories is also 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.

[0048] In an optional embodiment, the reference trajectory is obtained by averaging the production quality evolution trajectories of a plurality of standard workpieces generated 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.

[0049] For example, 50 standard workpieces that are confirmed to have no defects can be selected and let them pass through the entire production line completely. After each workpiece passes through all stations, a production quality evolution trajectory is generated, which is formed by connecting the state vectors of each station in sequence. Aligning and averaging the 50 trajectories point by point, a reference trajectory representing an ideal production process is obtained.

[0050] When a new workpiece is completed and its own quality evolution trajectory is generated, two calculations are performed. First, the dynamic time warping distance between the new trajectory and the reference trajectory is calculated, resulting in a value such as 4.5, which reflects the matching degree of the two trajectories in time sequence. Second, the shape difference between the two trajectories is calculated, for example using the Frechet distance, resulting in a value such as 3.0. According to the preset weight, 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 evaluation index is calculated as 4.5 multiplied by 0.7 plus 3.0 multiplied by 0.3, resulting in 4.05. The final score quantifies the overall deviation of the current workpiece production process from the ideal process.

[0051] The embodiment of the present application also discloses a generator circuit breaker box production line monitoring system, comprising a processor and a memory, 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 realized.

[0052] The above system also includes a communication bus and a communication interface and other components well known to those skilled in the art, the settings and functions of which are known in the art, and therefore will not be described here.

[0053] The above are preferred embodiments of the present application, and are not intended to limit the protection scope of the present application, therefore: any equivalent changes made on the structure, shape, principle of the present application should be covered within the protection scope of the present application.

Claims

1. A method of monitoring a generator circuit breaker tank production line, characterized by, The multi-modal data of each station on the production line is acquired, and the multi-modal data includes production line two-dimensional images, production line three-dimensional point cloud images, and device running sensor data; a preset key shape and position area of a box body in the production line two-dimensional image is identified, and geometric features of the key shape and position area are extracted, working condition state features in the sensor data are extracted, and station fusion features are acquired based on the geometric features and the working condition state features; a graph network is constructed, each station of the production line is taken as a node of the graph network, a physical flow relationship between the stations is taken as a directed edge between the nodes, and the station fusion features are taken as initial features of each node; station state vectors of each station containing accumulated influences of upstream processes are acquired through the graph network; the station state vectors are input into a prototype network and a reconstruction network trained by normal working condition data in parallel, and in response to similarities of the station state vectors with all known fault prototypes in the prototype network being lower than a first threshold value and reconstruction errors of the station state vectors passing through the reconstruction network being higher than a second threshold value, a current working condition is determined as an unknown production anomaly; for any box body on the production line, station state vectors corresponding to the box body at the stations are concatenated in a production time sequence to construct a production quality evolution track of the box body, and the evolution track is compared with a reference track generated based on a plurality of standard workpieces under a normal working condition to acquire a final quality evaluation index of the box body.

2. The method of claim 1, wherein the method further comprises: The key shape and position area includes a welding seam area and / or a mounting hole area on the box body; the geometric features include a height, a width, and a flatness of the welding seam and / or a diameter, a roundness, and a center position deviation of the mounting hole; and the working condition state features include statistical features of device running parameters within a machining period.

3. The method of claim 1, wherein the method further comprises: The step of acquiring the station fusion features based on the geometric features and the working condition state features includes: splicing the geometric features and the working condition state features to obtain an initial feature vector, and fusing and reducing dimensions of the initial feature vector using a neural network to obtain the station fusion features.

4. The method of claim 1, wherein, The graph network is a graph convolution network.

5. The method of claim 1, wherein, The step of acquiring the station state vectors of each station containing accumulated influences of upstream processes includes: for any node, aggregating a station state vector of the node in a last round of iteration with station state vectors of all upstream adjacent station nodes to obtain an intermediate vector, and processing the intermediate vector using a nonlinear activation function to obtain a latest station state vector.

6. The method of claim 1, wherein, The cosine similarity is taken as the similarity of the station state vector with the known fault prototype.

7. The method of claim 1, wherein, The reconstruction network is an autoencoder network, and the reconstruction error is a mean square error between the input station state vector and a vector reconstructed by the autoencoder network.

8. The method of claim 1, wherein, An average result of production quality evolution tracks of the plurality of standard workpieces generated under the normal working condition is taken as the reference track.

9. The method of claim 1, wherein, A weighted sum of a dynamic time warping distance and a track shape difference between the evolution track and the reference track is taken as the final quality evaluation index.

10. A generator circuit breaker tank production line monitoring system characterized by, The method comprises the following steps: A processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, a generator circuit breaker box production line monitoring method according to any one of claims 1-9 is realized.

Images