A strip steel surface defect tracing method, system, device and medium

Abstract

The application relates to a strip steel surface defect tracing method, system, device and medium. The method comprises the following steps: performing defect recognition on a strip steel surface defect image to obtain a defect type, a geometric feature vector and a texture feature vector, acquiring spatial coordinates and a timestamp corresponding to the image, and constructing a defect feature matrix; constructing a process knowledge graph based on production line equipment parameters, a theoretical speed curve, equipment topological relations, real-time working condition data and historical event records; performing candidate equipment positioning in the process knowledge graph by adopting a periodic matching strategy or a non-periodic matching strategy to generate a matching result set; acquiring an equipment health index from the process knowledge graph, calculating a final confidence and a comprehensive score based on matching parameters in the matching result set, and generating a tracing report. The method integrates multi-dimensional defect features and constructs a process knowledge graph, optimizes a defect and equipment matching strategy, and improves the accuracy and reliability of strip steel surface defect tracing.

Description

Technical Field

[0001] This invention belongs to the field of computer technology, and in particular relates to a method, system, device and medium for tracing the source of surface defects in steel strips. Background Technology

[0002] As a key basic material in the automotive, shipbuilding, home appliance, and pipeline transportation industries, the surface quality of strip steel directly determines the performance and service life of end products. Therefore, accurate tracing of surface defects during strip steel production is of great significance for improving product quality and reducing production costs. The core requirement of surface defect tracing is to pinpoint the specific equipment or process step that causes the defect through correlation analysis between defect characteristics and production processes, providing a technical basis for timely adjustment of equipment parameters and troubleshooting of potential faults.

[0003] However, existing strip steel surface defect processing technologies mostly focus on defect identification and classification. They typically employ deep learning models such as convolutional neural networks to process defect images acquired by vision systems to determine the defect type. However, these technologies only output basic defect classification results, failing to integrate the geometric and textural features of defects with spatiotemporal information such as spatial coordinates and timestamps at the time of defect occurrence. This results in a single dimension of defect feature representation, making it difficult to form a complete data foundation to support source tracing analysis. Furthermore, in terms of process data management, these technologies often use static databases to store production line equipment parameters and theoretical speed curves. They cannot integrate real-time operating condition data such as actual speed and vibration spectrum collected by sensor networks, nor can they establish a dynamic correlation between equipment status, process events, and equipment topology. This results in fragmented process data, failing to provide accurate knowledge support for matching defects with equipment.

[0004] Furthermore, existing strip steel surface defect treatment technologies have significant limitations in the defect-equipment matching process. For example, for periodic defects such as roll marks and vibration marks, the impact of the deviation between the actual and theoretical speeds during production on the actual effect of the roll diameter is not considered. Only fixed roll diameter parameters are used for periodic matching, resulting in large matching errors. For non-periodic defects such as indentation and scratches, there is a lack of quantitative analysis on the consistency of defect direction and spatial distribution patterns. Relying solely on manual experience or single equipment status parameters to screen candidate equipment makes it difficult to achieve accurate positioning. Summary of the Invention

[0005] Therefore, it is necessary to provide a method, system, equipment, and medium for tracing the source of surface defects in strip steel to address the above-mentioned technical problems, aiming to enhance the integration and dynamism of process data and improve the comprehensiveness and accuracy of defect feature characterization.

[0006] Firstly, this application provides a method for tracing the source of surface defects in strip steel, including:

[0007] Defect recognition processing is performed on the strip surface defect images acquired by the vision system to obtain the defect type, geometric feature vector and texture feature vector. The spatial coordinates and timestamps corresponding to the strip surface defect images are obtained. The defect type, spatial coordinates, timestamps, geometric feature vectors and texture feature vectors are integrated to generate a defect feature matrix.

[0008] Obtain production line equipment parameters, theoretical speed curves, and equipment topology relationships; collect real-time operating condition data; and construct and process a process knowledge graph based on production line equipment parameters, theoretical speed curves, equipment topology relationships, real-time operating condition data, and historical event records to generate a process knowledge graph.

[0009] Based on the defect type in the defect feature matrix, either a periodic matching strategy or an aperiodic matching strategy is used to locate candidate equipment in the process knowledge graph, generating a matching result set containing matching parameters.

[0010] The equipment health index is obtained from the process knowledge graph. The confidence level is calculated based on the equipment health index and the matching parameters in the matching result set to obtain the final confidence level. A comprehensive score is calculated based on the final confidence level and the equipment health index to generate a traceability report.

[0011] In one embodiment, defect identification processing is performed on the strip surface defect image acquired based on the vision system to obtain the defect type, geometric feature vector, and texture feature vector, and to obtain the spatial coordinates and timestamp corresponding to the strip surface defect image, including:

[0012] A vision system is used to acquire and process images of the strip surface to obtain images of defects on the strip surface;

[0013] The defect images on the surface of the strip steel are classified using a convolutional neural network based on ResNet50 and spatial pyramid pooling to obtain the defect types.

[0014] The surface defect images of strip steel are processed by a preset multi-scale topology network to extract features, resulting in geometric feature vectors and texture feature vectors.

[0015] A position sensor is used to detect and process the acquisition location and acquisition time of the strip surface defect image to obtain the spatial coordinates and timestamp corresponding to the strip surface defect image.

[0016] In one embodiment, production line equipment parameters, theoretical speed curves, and equipment topology relationships are acquired; real-time operating condition data is collected; and a process knowledge graph is constructed based on the production line equipment parameters, theoretical speed curves, equipment topology relationships, real-time operating condition data, and historical event records to generate a process knowledge graph, including:

[0017] Obtain production line equipment parameters, theoretical speed curves, and equipment topology from the manufacturing execution system. The production line equipment parameters include the set of roll diameters.

[0018] Sensor networks are used to collect and process data on the production line's operating status to obtain real-time operating data, which includes actual speed data and vibration spectrum data.

[0019] Based on vibration spectrum data, the equipment health status is assessed and calculated to obtain the equipment health index.

[0020] Temperature sensors are used to collect and process the temperature of production line equipment to obtain temperature data. The temperature data is then processed to calculate the change value. When the temperature change value exceeds the preset temperature threshold, a temperature disturbance event node is generated.

[0021] The steel throughput of the production line is detected and processed. When the steel throughput exceeds the preset steel throughput threshold, a steel throughput over-limit event node is generated.

[0022] Based on the set of roll diameters and the equipment topology, equipment nodes are constructed to obtain equipment nodes;

[0023] Historical event records are acquired, and correlation processing is performed on equipment nodes, temperature disturbance event nodes, steel overload event nodes, equipment topology relationships, and historical event records to generate a process knowledge graph.

[0024] In one embodiment, based on the defect type in the defect feature matrix, a periodic matching strategy is used to perform candidate equipment location processing in the process knowledge graph, generating a matching result set containing matching parameters, including:

[0025] Based on the defect type, defect data of the type of roller printing or vibration pattern are selected from the defect feature matrix to obtain a periodic defect subset;

[0026] The spatial coordinates of adjacent defects in a periodic defect subset are processed to calculate the spacing, thus obtaining the actual spacing.

[0027] Based on the timestamps of each defect in the periodic defect subset, the actual speed data for the corresponding time period is extracted from the process knowledge graph;

[0028] The actual speed integral value is obtained by performing integral calculations based on the actual speed data for the corresponding time period.

[0029] The theoretical speed curve is obtained from the process knowledge graph, and the theoretical speed curve is processed by integration to obtain the theoretical speed integral value;

[0030] The nominal diameter of the roll is obtained from the process knowledge graph. Based on the actual speed integral value, the theoretical speed integral value, and the nominal diameter of the roll, roll diameter compensation calculation is performed to obtain the effective roll diameter.

[0031] Based on the actual spacing and effective roll diameter, equipment retrieval is performed in the process knowledge graph. Equipment that meets the preset periodic matching conditions is retrieved as candidate equipment. The candidate equipment and effective roll diameter are integrated to obtain matching parameters, and a matching result set containing the matching parameters is generated.

[0032] In one embodiment, based on the defect type in the defect feature matrix, a non-periodic matching strategy is used to perform candidate equipment location processing in the process knowledge graph, generating a matching result set containing matching parameters, including:

[0033] Based on the defect type, defect data of the type of indentation or scratch are selected from the defect feature matrix to obtain a non-periodic defect subset;

[0034] Extract the geometric feature vectors corresponding to each defect from the aperiodic defect subset, and extract the orientation angle data from the geometric feature vectors;

[0035] The standard deviation of the orientation angle data is calculated to obtain the orientation consistency index;

[0036] The strip surface is divided into multiple grid cells based on the spatial coordinates of each defect in the non-periodic defect subset;

[0037] The number of defects in each grid cell is statistically processed to obtain the defect density. Based on the defect density, the distribution entropy is calculated to obtain the distribution entropy.

[0038] Equipment with a health index lower than a preset health index threshold and located upstream of the defect location is selected from the process knowledge graph as candidate equipment.

[0039] The candidate devices, directional consistency index, and distribution entropy are processed to generate a matching result set containing matching parameters, including directional consistency index and distribution entropy.

[0040] In one embodiment, an equipment health index is obtained from a process knowledge graph. A confidence level is calculated based on the equipment health index and matching parameters in the matching result set to obtain a final confidence level. A comprehensive score is then calculated based on the final confidence level and the equipment health index, generating a traceability report, including:

[0041] When the matching result set is the matching result set of the corresponding periodic matching strategy, the effective roller diameter and actual spacing are extracted from the matching result set, the spacing error is calculated on the effective roller diameter and actual spacing to obtain the spacing error value, and the confidence is calculated based on the spacing error value to obtain the periodic confidence.

[0042] When the matching result set is the matching result set of the corresponding non-periodic matching strategy, the distribution entropy and directional consistency index are extracted from the matching result set, and the confidence score is calculated based on the distribution entropy and directional consistency index to obtain the non-periodic confidence score.

[0043] The event nodes associated with candidate equipment in the process knowledge graph are retrieved. When there is a temperature disturbance event node, the periodic confidence level or the non-periodic confidence level is enhanced by a preset first enhancement coefficient to obtain the final confidence level. When there is an event node with excessive steel content, the periodic confidence level or the non-periodic confidence level is enhanced by a preset second enhancement coefficient to obtain the final confidence level.

[0044] The equipment health index of candidate equipment is obtained from the process knowledge graph, and a comprehensive score is obtained by weighting the final confidence level and the equipment health index.

[0045] Candidate devices are sorted according to their comprehensive scores, and the final confidence level, comprehensive score, and corresponding matching parameters are used as core evidence to generate a source tracing report.

[0046] In one embodiment, the periodic confidence level is calculated using the following formula:

[0047]

[0048] in, The confidence level is periodic, and its value ranges from [0,1]. The absolute deviation between the effective roll diameter and the nominal roll diameter, i.e. , The nominal diameter of the roll. For the effective roller diameter, This is the spacing matching penalty coefficient, used to amplify the impact of spacing difference on confidence level. This represents the actual spacing between adjacent periodic defects. The theoretical defect spacing is based on the effective roll diameter. Pi, with a value of 3.1416. This is the vibration stability correction factor. The vibration stability index of the equipment has a value range of [0,1] and is calculated from the power spectral density of the vibration spectrum data.

[0049] Secondly, this application also provides a strip steel surface defect tracing system, including:

[0050] The feature extraction and matrix construction module is used to perform defect recognition processing on the strip surface defect images acquired by the vision system, to obtain the defect type, geometric feature vector and texture feature vector, to obtain the spatial coordinates and timestamps corresponding to the strip surface defect images, and to integrate the defect type, spatial coordinates, timestamps, geometric feature vectors and texture feature vectors to generate a defect feature matrix.

[0051] The process knowledge graph construction module is used to acquire production line equipment parameters, theoretical speed curves, and equipment topology relationships, collect real-time operating condition data, and perform process knowledge graph construction processing based on production line equipment parameters, theoretical speed curves, equipment topology relationships, real-time operating condition data, and historical event records to generate a process knowledge graph.

[0052] The candidate equipment location module is used to perform candidate equipment location processing in the process knowledge graph based on the defect type in the defect feature matrix, using either a periodic matching strategy or an aperiodic matching strategy, and generate a matching result set containing matching parameters.

[0053] The traceability report generation module is used to obtain the equipment health index from the process knowledge graph, calculate the confidence level based on the equipment health index and the matching parameters in the matching result set, obtain the final confidence level, calculate the comprehensive score based on the final confidence level and the equipment health index, and generate the traceability report.

[0054] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the first aspect.

[0055] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the first aspect.

[0056] The aforementioned method, system, equipment, and medium for tracing the source of surface defects in strip steel firstly identifies and processes defect images acquired by a vision system, integrating spatiotemporal features to generate a defect feature matrix, laying the data foundation for subsequent defect-equipment matching. Secondly, it acquires multi-source data from the production line and constructs a process knowledge graph, overcoming the shortcomings of traditional methods' static and fragmented process data, and providing dynamic and correlated knowledge support for matching analysis. Furthermore, it employs a differentiated matching strategy based on defect type to locate candidate equipment, improving the accuracy of equipment location. Finally, it combines equipment health index calculations with confidence scores and comprehensive ratings to generate reports, further enhancing the credibility of the tracing results and the guidance for subsequent production optimization. Attached Figure Description

[0057] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0058] Figure 1 A flowchart of a method for tracing the source of surface defects in strip steel, provided as an exemplary embodiment of the present invention;

[0059] Figure 2 A schematic flowchart of a method for defect identification processing of strip steel surface defect images is provided as an exemplary embodiment of the present invention;

[0060] Figure 3 This is a schematic diagram of a strip steel surface defect tracing system provided as an exemplary embodiment of the present invention. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0062] In one embodiment, such as Figure 1 As shown, a method for tracing the source of surface defects in strip steel is provided. This embodiment illustrates the application of this method to a terminal. It is understood that this method can also be applied to a server, and to a system including both a terminal and a server, and implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0063] S101: Perform defect recognition processing on the strip surface defect image acquired by the vision system to obtain the defect type, geometric feature vector and texture feature vector, obtain the spatial coordinates and timestamp corresponding to the strip surface defect image, and integrate the defect type, spatial coordinates, timestamp, geometric feature vector and texture feature vector to generate a defect feature matrix.

[0064] Specifically, the type of defect determines the possible generation mechanism, such as roll marks related to roll rotation, scratches related to equipment friction, etc. Spatial coordinates can pinpoint the specific location of the defect in the length and width of the strip, thus associating it with the equipment in the corresponding area of ​​the production line. Timestamps can pinpoint the precise production period in which the defect occurred, providing a basis for retrieving process parameters and equipment status during that period. Geometric feature vectors (such as contours and dimensions) and texture feature vectors (such as surface texture distribution) can reflect the essential attributes of the defect, providing a basis for distinguishing defects generated by different equipment. For example, during implementation, the vision system can adopt a combination of linear scan cameras and light sources. The linear scan cameras are deployed along the width of the strip, and the acquisition frame rate is adaptively matched with the strip running speed to ensure that the defect images are not stretched or blurred. Subsequently, defect recognition can be achieved through feature extraction and classification logic. That is, the acquired defect images are first preprocessed, such as grayscale normalization and noise filtering, and then the feature extraction module extracts the defect type identifier, geometric morphological features, and surface texture features to ensure the distinguishability and stability of the features. Spatial coordinates can be acquired through position sensors that are synchronously triggered with the vision system. These sensors can be installed beside the strip conveyor rollers and use laser ranging principles for measurement. Timestamps are synchronized with a unified system clock, ensuring that the spatiotemporal information of each defect image is perfectly aligned with the actual production process. During integrated processing, the data in each dimension is first standardized, normalizing geometric and texture feature vectors to the same numerical range. Then, a defect feature matrix is ​​constructed according to a fixed dimension of defect type-spatial coordinates-timestamp-geometric feature vector-texture feature vector. This matrix is ​​stored in a structured data format, supporting rapid retrieval and vector operations.

[0065] S102: Obtain production line equipment parameters, theoretical speed curves, and equipment topology relationships; collect real-time operating condition data; and perform process knowledge graph construction based on production line equipment parameters, theoretical speed curves, equipment topology relationships, real-time operating condition data, and historical event records to generate a process knowledge graph.

[0066] Specifically, production line equipment parameters determine the physical properties of the strip steel affected by the equipment, while real-time operating data reflects the actual operating status of the equipment. Equipment topology clarifies the positional relationship of each piece of equipment in the production process, and historical event records contain information on abnormal operating conditions that may lead to defects. By combining the above data to construct a process knowledge graph, the dynamic characteristics of the production process can be comprehensively depicted. For example, the process knowledge graph can be constructed following the logic of entity-relationship-attribute. Entity nodes can include equipment nodes (defined as equipment in the production line that directly contacts or affects the surface quality of the strip steel, such as rolls, guide devices, etc.) and event nodes (defined as abnormal operating conditions that may affect the equipment status or strip steel quality, such as sudden temperature changes, exceeding the steel throughput limit, etc.). Relationships can include topological relationships between equipment (based on upstream and downstream relationships in the production process) and association relationships between equipment and events (based on the range of equipment involved when the event occurs). Attributes can include equipment parameter attributes (such as roll diameter), status attributes (such as equipment health index), and event characteristic attributes (such as temperature change value, exceeding limit value). Furthermore, during the construction process, real-time operating data can drive the dynamic updating of equipment status attributes. For example, based on vibration spectrum data, the equipment health index can be calculated in real time to reflect the real-time operating status of the equipment. When abnormal operating conditions are detected, corresponding event nodes are automatically generated and associated with related equipment, ensuring that the knowledge graph can reflect the dynamic changes in the production process in real time. The generation of this process knowledge graph realizes the structured integration of multi-source process data, providing comprehensive and dynamic process knowledge support for the subsequent positioning of candidate equipment, avoiding the shortcomings of traditional static databases that cannot adapt to real-time changes in production.

[0067] S103: Based on the defect type in the defect feature matrix, perform candidate equipment location processing in the process knowledge graph using either a periodic matching strategy or an aperiodic matching strategy, and generate a matching result set containing matching parameters.

[0068] Specifically, due to the fundamental differences in the generation mechanisms of different types of defects—namely, periodic defects such as roll marks and vibration marks are directly related to the periodic movement of equipment (e.g., the rotation of rolls) and their distribution on the strip surface exhibits obvious periodic characteristics—while non-periodic defects such as indentation and scratches are mostly related to non-periodic events such as abnormal contact or localized damage to equipment, and their distribution has no fixed period but may show specific directional or spatial distribution patterns. Therefore, by adopting corresponding matching strategies for different types of defects, the accuracy of equipment positioning can be significantly improved. For example, strategy triggering judgment is first performed based on the defect type identifier in the defect feature matrix. When the defect type is roll mark or vibration mark, a periodic matching strategy is triggered; when the defect type is indentation or scratch, a non-periodic matching strategy is triggered. Subsequently, when locating candidate equipment in the process knowledge graph, the spatial coordinates and timestamps in the defect feature matrix are used as constraints to first lock the range of equipment upstream and downstream of the corresponding spatial location during the defect generation period (based on the equipment topology relationship in the knowledge graph), and then further filtering is performed based on the core parameters of the matching strategy. The core of the periodic matching strategy is to correlate the periodic characteristics of defects with the periodic parameters of the equipment, while the core of the non-periodic matching strategy is to correlate the direction and distribution characteristics of defects with the abnormal states of the equipment. By extracting key quantitative indicators that reflect the degree of correlation between defects and equipment during the matching process, matching parameters can be obtained, providing a data foundation for subsequent confidence level calculations.

[0069] S104: Obtain the equipment health index from the process knowledge graph, calculate the confidence level based on the equipment health index and the matching parameters in the matching result set, obtain the final confidence level, calculate the comprehensive score based on the final confidence level and the equipment health index, and generate a traceability report.

[0070] Specifically, matching parameters reflect the degree of fit between defect characteristics and equipment parameters, while the equipment health index reflects the probability of equipment generating defects. Combining these two allows for a comprehensive assessment of the correlation between candidate equipment and defects. For example, the equipment health index can be retrieved from the corresponding equipment node attributes in the process knowledge graph. This index is a quantitative indicator dynamically calculated based on real-time operating data, reflecting the actual operating status of the equipment during the defect occurrence period. The core of confidence calculation is to transform matching parameters into quantified correlation probabilities. For instance, in periodic matching, the higher the fit between the matching parameters and the equipment's periodic parameters, the higher the confidence level; in non-periodic matching, the higher the fit between defect characteristics and abnormal equipment states, the higher the confidence level. By comprehensively considering the correlation degree of matching parameters and the influence of equipment health status, the final confidence level can be obtained, avoiding misjudgments caused by a single factor. The calculation of the comprehensive score further optimizes the final confidence level. For example, by setting reasonable weight allocations, since matching parameters directly reflect the correlation between defects and equipment, while the equipment health index serves as an auxiliary judgment, the confidence weight corresponding to the matching parameters can be higher than that of the equipment health index, thus obtaining a quantitative indicator that reflects the correlation priority of candidate equipment. By presenting the candidate equipment after ranking by comprehensive scores in a structured manner, a traceability report is generated. This report can include the candidate equipment's identifier, comprehensive score, confidence level, and core matching evidence such as the specific values ​​of matching parameters and equipment health index. It provides production personnel with clear and reliable traceability results, which can directly guide production personnel to conduct equipment inspection and process optimization, effectively shortening fault handling time and reducing production costs.

[0071] The aforementioned method, by acquiring and processing defect images, integrates defect type, geometric features, texture features, spatial coordinates, and timestamps to generate a defect feature matrix, thus addressing the problems of single-dimensional defect feature representation and lack of spatiotemporal information integration in traditional methods. Secondly, it acquires production line equipment parameters, theoretical speed curves, equipment topology relationships, and real-time operating data to construct a process knowledge graph, overcoming the shortcomings of fragmented process data and lack of dynamic correlation in traditional methods. Furthermore, based on defect type, a periodic or non-periodic matching strategy is used to locate candidate equipment in the process knowledge graph, generating a matching result set, effectively solving the strategic limitations of traditional methods in the defect-equipment matching stage. Finally, confidence is calculated by combining equipment health index and matching parameters to generate a comprehensive score and traceability report, enhancing the reliability and credibility of the traceability results and improving the accuracy and efficiency of strip steel surface defect traceability in complex production scenarios.

[0072] In one embodiment, such as Figure 2 As shown, defect recognition processing is performed on strip surface defect images acquired based on a vision system to obtain defect type, geometric feature vector, and texture feature vector, and to obtain the spatial coordinates and timestamps corresponding to the strip surface defect images, including:

[0073] S201: Use a vision system to acquire and process images of the strip surface to obtain images of defects on the strip surface;

[0074] S202: The defect images of the strip steel surface are classified using a convolutional neural network based on ResNet50 and spatial pyramid pooling to obtain the defect type;

[0075] S203: Feature extraction processing is performed on the surface defect image of the strip steel using a preset multi-scale topology network to obtain geometric feature vectors and texture feature vectors;

[0076] S204: The acquisition location and acquisition time of the strip surface defect image are detected and processed using a position sensor to obtain the spatial coordinates and timestamp corresponding to the strip surface defect image.

[0077] Specifically, the vision system can consist of a line scan camera, a high-brightness strip LED light source, an image acquisition card, and an industrial control computer. The line scan camera can be horizontally mounted above the strip conveyor rollers along the width of the strip, with the lens optical axis perpendicular to the strip surface. The mounting height can be set according to the lens focal length to ensure the camera's field of view covers the entire width of the strip. The high-brightness strip LED light source uses cool white light and is symmetrically arranged on both sides of the camera. The light source forms a preset angle with the strip surface, and the light intensity is adjusted by a light source controller to avoid loss of defect details caused by reflections on the strip surface. The image acquisition card can be a high-speed card, supporting a preset range of acquisition rates. The acquisition rate and the strip running speed are linked and controlled by the industrial control computer. During acquisition, image acquisition can be triggered by a photoelectric sensor installed at the entrance of the conveyor rollers. When the strip head reaches the photoelectric sensor position, the sensor sends a trigger signal to the industrial control computer to start image acquisition. Since the original image is a grayscale image, it can be transmitted to an industrial control computer via a data acquisition card for preprocessing. The preprocessing process involves removing image noise using a median filtering algorithm, normalizing the image grayscale values ​​to a preset range using a grayscale stretching algorithm, and cropping out invalid areas at the edges of the strip. Finally, an image of the strip surface defects with effective dimensions can be obtained, which can clearly show minute defects.

[0078] Furthermore, images of surface defects in steel strips can be classified using a convolutional neural network based on ResNet50 and spatial pyramid pooling. This network model uses ResNet50 as its backbone, removing the fully connected layers and global average pooling layers of the original ResNet50. A spatial pyramid pooling module is added at the end of the network. This module sets multiple pooling kernels of different scales to perform multi-scale pooling on the feature maps output by the backbone network. The pooling results of different scales are concatenated into a fixed-dimensional feature vector, which is then connected to a fully connected layer and a Softmax activation function to form a complete classification network—that is, a convolutional neural network based on ResNet50 and spatial pyramid pooling. Before training this network, a sample dataset containing multiple types of defects can be constructed, with the number of samples for each type meeting the training requirements. The sample images undergo the same preprocessing as the acquired images. During training, the input image size is uniformly adjusted to a preset size, and data augmentation methods such as random flipping, rotation, and brightness perturbation are used to improve the model's generalization ability. The Adam optimizer is used as the optimizer, and the cross-entropy loss function is adopted. Training continues until the validation set accuracy reaches the preset requirement. During classification, the pre-processed images of surface defects on the strip are input into the trained network model. The model outputs probability vectors corresponding to the number of defect categories, where the dimension corresponding to the maximum probability value is the defect category. These vectors are then converted into corresponding defect type codes based on a pre-defined category-code mapping relationship, enabling accurate defect type determination. This mapping relationship can cover the correspondence between categories such as scratches, wrinkles, holes, cracks, roll marks, surface dirt, color difference, surface roughness, vibration marks, surface spots, surface scratches, surface indentations, surface yellowing, and surface peeling, and their corresponding codes.

[0079] Specifically, when using a pre-defined multi-scale topology network for feature extraction of strip surface defect images, the multi-scale topology network can adopt a "multi-scale feature fusion-feature separation extraction" structure design to obtain more comprehensive geometric and texture features. The network input is the pre-processed strip surface defect image. It first extracts multi-scale features through three parallel convolutional branches, each using convolutional kernels of different sizes. During the convolution process of each branch, corresponding stride and padding parameters are set, and the number of output channels for the three branches remains consistent. Then, the number of feature map channels in the three branches is unified through the convolutional kernels, and finally, pixel-by-pixel addition is performed to achieve multi-scale feature fusion. After feature fusion, geometric feature vectors and texture feature vectors can be extracted separately through two parallel sub-networks. For example, the geometric feature extraction sub-network can first use the Canny edge detection algorithm to extract the defect contour. This algorithm filters edge pixels by setting high and low thresholds to obtain a clear contour, and then uses a contour tracking algorithm to obtain the contour point coordinates. Based on the contour point coordinates, the contour curvature, aspect ratio, defect area, and orientation angle are calculated. The contour curvature can be calculated by multi-point method to obtain the feature value of each contour point. The aspect ratio is the ratio of the length to the width of the minimum bounding rectangle of the defect. The defect area can be obtained by combining the number of pixels in the contour with pixel equivalent conversion. The orientation angle is the angle between the long side of the minimum bounding rectangle of the defect and the strip length direction. After quantizing and normalizing the above four features to a preset range, a geometric feature vector can be formed. The texture feature extraction sub-network can first convert the fused feature map into a grayscale image, and then calculate the gray-level co-occurrence matrix energy, HOG features and wavelet energy coefficients in sequence. The gray-level co-occurrence matrix energy is obtained by calculating the matrix and obtaining the energy value statistics after setting distance and angle. The HOG features are obtained by dividing into cell units and blocks and calculating the orientation angle features. The wavelet energy coefficients are obtained by taking the energy ratio of each layer component after wavelet basis decomposition. After concatenating the above three types of features, the dimensionality is reduced by principal component analysis and normalized to a preset range to form a texture feature vector. The final output geometric feature vector and texture feature vector can also be associated with feature identifiers to ensure that they correspond to the same defect image.

[0080] Specifically, when using a position sensor to detect and process the acquisition location and time of images of surface defects on strip steel, a position-time dual synchronization mechanism can be constructed to ensure accurate matching of spatiotemporal information with the defect images. The position sensor is a laser rangefinder, installed on the side of the strip steel conveyor rollers. The laser emission direction is perpendicular to the width direction of the strip steel, and the laser spot falls on the positioning marks on the edge of the strip steel. These positioning marks are pre-made, equally spaced scale lines during strip steel production. Time synchronization is achieved using a network time protocol. The industrial control computer, laser rangefinder, and image acquisition card are all connected to the same industrial Ethernet, and the clocks of each device are synchronized to a preset accuracy through a network time protocol server. When the photoelectric sensor triggers image acquisition, the industrial control computer simultaneously records the trigger time (timestamp) of the acquisition card and sends a position detection command to the laser rangefinder. After receiving the command, the laser rangefinder measures the distance between the current positioning mark on the strip steel edge and the sensor, and calculates the X-coordinate of the strip steel surface defect image acquisition location based on the pre-made intervals of the positioning marks. This coordinate is set with the roller inlet as the origin along the length of the strip steel. Simultaneously, a linear displacement sensor installed along the width of the strip measures the position of the camera's field of view along the width of the strip, which is then used as the Y-coordinate. Finally, the X and Y coordinates are integrated with a timestamp to form the spatial coordinates and timestamp corresponding to the surface defect image of the strip. The coordinate data retains a preset number of decimal places, and the timestamp adopts a standardized time format. Data association identifiers are bound to the corresponding defect type, geometric feature vector, and texture feature vector, providing a precise spatiotemporal reference for subsequent integration processing.

[0081] After the above process is completed, the data of each dimension can be standardized first, and the geometric feature vector and texture feature vector can be normalized to the same numerical range. Then, a defect feature matrix can be constructed according to a preset fixed dimension order. This matrix is ​​stored in a structured data format to support fast retrieval and vector operations, providing a complete and standardized data source for the selection of subsequent matching strategies and device positioning.

[0082] In one embodiment, production line equipment parameters, theoretical speed curves, and equipment topology relationships are acquired; real-time operating condition data is collected; and a process knowledge graph is constructed based on the production line equipment parameters, theoretical speed curves, equipment topology relationships, real-time operating condition data, and historical event records to generate a process knowledge graph, including:

[0083] Obtain production line equipment parameters, theoretical speed curves, and equipment topology from the manufacturing execution system. The production line equipment parameters include the set of roll diameters.

[0084] Sensor networks are used to collect and process data on the production line's operating status to obtain real-time operating data, which includes actual speed data and vibration spectrum data.

[0085] Based on vibration spectrum data, the equipment health status is assessed and calculated to obtain the equipment health index.

[0086] Temperature sensors are used to collect and process the temperature of production line equipment to obtain temperature data. The temperature data is then processed to calculate the change value. When the temperature change value exceeds the preset temperature threshold, a temperature disturbance event node is generated.

[0087] The steel throughput of the production line is detected and processed. When the steel throughput exceeds the preset steel throughput threshold, a steel throughput over-limit event node is generated.

[0088] Based on the set of roll diameters and the equipment topology, equipment nodes are constructed to obtain equipment nodes;

[0089] Historical event records are acquired, and correlation processing is performed on equipment nodes, temperature disturbance event nodes, steel overload event nodes, equipment topology relationships, and historical event records to generate a process knowledge graph.

[0090] Specifically, production line equipment parameters can include core parameters of key equipment directly affecting the surface quality of strip steel. Roll diameter sets, as an important component, record the basic physical parameters of each roll, providing data support for subsequent periodic matching. The theoretical speed curve reflects the change in strip steel running speed set in the production plan over time. Equipment topology clarifies the spatial layout and upstream / downstream connection sequence of each piece of equipment in the production process, defining the range of equipment potentially involved when defects occur, and providing spatial constraints for subsequent equipment retrieval. The above data can be standardized and stored in a temporary database to ensure a unified data structure for easy subsequent retrieval and processing. Subsequently, a sensor network can be used to collect and process data on the production line's operating status. The sensor network consists of various sensors deployed at key locations on the production line. Speed ​​sensors are installed at the main drive roller conveyor of the strip steel to collect the actual running speed of the strip steel in real time; vibration sensors are installed at the bearing seats or machine bodies of key equipment such as rolls and guide devices to collect vibration signals generated during equipment operation. This sensor network employs a distributed deployment approach. Each sensor collects data at a preset frequency, which is set based on the equipment's operating characteristics and data requirements to ensure the capture of dynamic changes in the equipment's operating status. After preprocessing, the collected raw data undergoes filtering algorithms to remove noise signals caused by environmental interference. Then, signal amplification and analog-to-digital conversion convert the analog signals into digital signals, ultimately generating standardized actual velocity data and vibration spectrum data, which are transmitted to the data processing center in real time.

[0091] Specifically, based on the principle of vibration spectrum analysis, the health status of the equipment can be quantified and a health index calculated by comparing the difference between the actual vibration spectrum of the equipment during operation and the reference vibration spectrum. The calculation formula is as follows:

[0092]

[0093] in, Indicates the device at time Health index, Indicates the device at time Vibration spectrum data, The vibration frequency, This represents the reference vibration spectrum data under normal operating conditions of the equipment. and These represent the lower and upper limits of the set vibration frequency analysis range, respectively. This formula, by calculating the integral value of the difference between the actual vibration spectrum and the reference vibration spectrum within the set frequency range, reflects the degree of deviation between the equipment's vibration state and its normal state. The smaller the integral value, the closer the equipment health index is to 1, indicating a better equipment health state; conversely, the larger the integral value, the worse the equipment health state, and the more likely it is to produce defects affecting the surface quality of the strip steel. This calculation method achieves a quantitative assessment of the equipment's health state, providing a condition constraint basis for subsequent equipment selection.

[0094] Specifically, temperature sensors can be deployed at key heat-generating components of production line equipment to collect temperature data in real time during equipment operation and store the collected temperature data in a time series. Temperature change values ​​can be calculated from the difference between temperature data at two adjacent collection points, reflecting the rate and magnitude of temperature change. Preset temperature thresholds can be determined based on equipment material characteristics, operating conditions, and statistical analysis of historical fault data. When the calculated temperature change value exceeds this preset threshold, it indicates an abnormal temperature change, which may lead to changes in equipment performance or damage, thereby affecting the surface quality of the strip steel. In this case, a temperature disturbance event node can be automatically generated. This event node includes information such as the time of the event, the identification of the involved equipment, and the temperature change value, providing event evidence for subsequent confidence enhancement. Furthermore, the amount of strip steel passing through the production line can be detected and processed. This amount of strip steel can be detected in real time using a flow detection device installed at a specific location on the production line. This device, based on weighing or volume measurement principles, can count the total amount of strip steel passing through that location in real time. The preset strip steel threshold can be set according to the equipment's design load capacity, service life, and production process requirements, serving as a standard for judging whether the equipment is operating under overload conditions. The real-time steel throughput is compared with a preset steel throughput threshold. When the real-time steel throughput exceeds the preset threshold, it indicates that the equipment is operating under overload conditions. Long-term overload operation may lead to accelerated equipment wear, performance degradation, and ultimately, surface defects in the strip. In this case, an overload event node can be automatically generated. This event node records key information such as the time of the event, the identification of the equipment involved, and the overload value. Together with the temperature disturbance event node, it constitutes key event data affecting the equipment status.

[0095] Specifically, following the principles of entity modeling, equipment nodes can be constructed based on the set of roll diameters and equipment topology. This means using key equipment in the production line as the modeling object, with each equipment node containing basic and status attributes. Basic attributes can include core parameters such as equipment identification, roll diameter, and installation location information, while status attributes can include real-time status data such as equipment health indices. During construction, the equipment topology can be used as a framework, associating each roll diameter parameter in the roll diameter set with its corresponding equipment identification, ensuring the completeness and accuracy of the attribute information for each equipment node. Equipment nodes can be stored in a structured data format, distinguished by a unique equipment identifier, facilitating subsequent association with event nodes and supporting dynamic updates of attribute information to ensure that equipment nodes reflect the actual status of the equipment in real time. Subsequently, historical event records can be retrieved, and equipment nodes, temperature disturbance event nodes, steel throughput exceeding limit event nodes, equipment topology relationships, and historical event records can be correlated to generate a process knowledge graph. The historical event records include various abnormal events that have occurred during production line operation, such as temperature disturbances and steel throughput exceeding limits, along with their handling results, providing a reference for current event analysis and equipment status assessment. During the association processing, a process knowledge graph can be constructed using a triplet model of entity-relationship-attribute based on graph database technology. For example, equipment nodes, temperature disturbance event nodes, and steel throughput exceeding limits event nodes are first designated as entity nodes in the graph. Then, upstream and downstream relationships between equipment nodes are established based on equipment topology, and relationships between event nodes and equipment nodes are established based on the equipment scope involved when an event occurs. Simultaneously, historical event records are associated with their corresponding equipment nodes and event nodes, forming a complete relationship network. This process knowledge graph supports dynamic updates. When new real-time operating data and event nodes are generated, the status attributes and relationships of equipment nodes can be updated promptly, ensuring that the graph can reflect the production line's operating status and equipment associations in real time.

[0096] In one embodiment, based on the defect type in the defect feature matrix, a periodic matching strategy is used to perform candidate equipment location processing in the process knowledge graph, generating a matching result set containing matching parameters, including:

[0097] Based on the defect type, defect data of the type of roller printing or vibration pattern are selected from the defect feature matrix to obtain a periodic defect subset;

[0098] The spatial coordinates of adjacent defects in a periodic defect subset are processed to calculate the spacing, thus obtaining the actual spacing.

[0099] Based on the timestamps of each defect in the periodic defect subset, the actual speed data for the corresponding time period is extracted from the process knowledge graph;

[0100] The actual speed integral value is obtained by performing integral calculations based on the actual speed data for the corresponding time period.

[0101] The theoretical speed curve is obtained from the process knowledge graph, and the theoretical speed curve is processed by integration to obtain the theoretical speed integral value;

[0102] The nominal diameter of the roll is obtained from the process knowledge graph. Based on the actual speed integral value, the theoretical speed integral value, and the nominal diameter of the roll, roll diameter compensation calculation is performed to obtain the effective roll diameter.

[0103] Based on the actual spacing and effective roll diameter, equipment retrieval is performed in the process knowledge graph. Equipment that meets the preset periodic matching conditions is retrieved as candidate equipment. The candidate equipment and effective roll diameter are integrated to obtain matching parameters, and a matching result set containing the matching parameters is generated.

[0104] Specifically, the core characteristics of roll mark and vibration mark defects are caused by the periodic movement of equipment (such as uniform rotation of rolls or periodic vibration of equipment). Their distribution on the strip surface exhibits a periodic pattern with fixed spacing. Therefore, this type of defect data can be separated from the defect feature matrix first. Then, the defect type identifiers stored in the defect feature matrix can be called, and the defect records corresponding to the identifiers (roll mark or vibration mark) can be filtered out using an identifier matching algorithm. The spatial coordinates, timestamps, and other core information of each record are extracted, sorted by timestamp order, and integrated into a subset of periodic defects. This sorting operation ensures that the spacing calculation of adjacent defects conforms to the temporal logic of strip production, avoiding spacing calculation errors caused by data disorder. Furthermore, since the periodicity of periodic defects is mainly reflected in the distribution pattern along the length of the strip, the spatial coordinates can be preprocessed to extract the coordinate values ​​of each defect along the length of the strip, ignoring the interference of the width direction coordinates on the periodicity. Subsequently, the spatial coordinates of adjacent defects within the periodic defect subset are processed to calculate the actual spacing. Adjacent defects are determined by the order of their timestamps; two defect records with consecutive timestamps are considered adjacent. The actual spacing is obtained by calculating the absolute value of the difference between the length coordinates of adjacent defects. If multiple feature points exist in the width direction of the same defect, the average length coordinate of each feature point is taken as the length coordinate of the defect to eliminate the influence of defect width on the spacing calculation. After calculation, outlier removal is performed on the obtained actual spacing data, eliminating spacing values ​​that deviate from the mean by more than a preset multiple to ensure the accuracy of subsequent matching.

[0105] Specifically, the periodic spacing of defects is directly related to the strip running speed and equipment rotation speed. Therefore, the actual strip running speed during the defect occurrence period can be used as the basic data for roll diameter compensation. In implementation, the timestamps of the first and last defects in the periodic defect subset can be extracted first to determine the time interval of defect occurrence. Then, the actual strip speed data within this time interval can be extracted through the time retrieval interface of the process knowledge graph. Furthermore, the extracted speed data must correspond one-to-one with the timestamps. If data is missing, linear interpolation is used to supplement the missing data. If there are outliers with sudden speed changes, a moving average method is used for smoothing to ensure that the speed data accurately reflects the strip running state during the defect occurrence period. Subsequently, integration can be performed based on the actual speed data for the corresponding time period to obtain the actual speed integral value. The core purpose of actual speed integration is to obtain the actual running distance of the strip during the defect occurrence period, which is a key parameter for roll diameter compensation calculation. The integration operation can use the trapezoidal integral method in numerical integration, with the integration interval being the time interval of defect occurrence, the integration variable being time, and the integrand being the actual speed data. The calculation logic is as follows: the time interval is divided into several small time intervals, and the speed within each interval is considered a constant value. The product of the speed within each interval and the time interval (i.e., the running distance within that interval) is calculated. Then, the running distances of all intervals are summed to obtain the actual speed integral value. During the integration process, the precision of the time interval division must match the speed data acquisition frequency to ensure the accuracy of the integration result.

[0106] Specifically, the theoretical speed curve is the benchmark for strip steel running speed set in the production plan. Its integral result is the theoretical running distance of the strip steel during the defect occurrence period, used to compare with the actual running distance to achieve roll diameter compensation. For example, the theoretical speed curve corresponding to the production plan can be extracted from the equipment parameter nodes of the process knowledge graph. This curve is plotted with time on the x-axis and speed on the y-axis. Then, the same trapezoidal integration method as the actual speed integration is used to perform integration within the same time interval as the actual speed integration to obtain the theoretical speed integral value. Furthermore, the integration interval and calculation method of the above theoretical speed integral and actual speed integral can be completely consistent to ensure the effectiveness of subsequent compensation coefficient calculations.

[0107] Specifically, the nominal diameter of the rolls is a design parameter of the equipment. However, in actual production, the actual running speed of the strip deviates from the theoretical speed, causing the actual effective roll diameter to differ from the nominal diameter. Therefore, the effective roll diameter can be obtained through compensation calculation. The calculation formula is as follows:

[0108]

[0109] in, For the effective roller diameter, The nominal diameter of the roll. This is the integral value of the actual speed. This is the integral value of the theoretical speed. This is the timestamp of the first defect in the periodic defect subset. This is the timestamp of the last defect. The formula constructs a compensation coefficient using the ratio of the actual speed to the theoretical speed integral, correcting the nominal roll diameter to reflect the actual effect of the equipment. When the actual speed is greater than the theoretical speed, the compensation coefficient is greater than 1, and the effective roll diameter is greater than the nominal diameter; conversely, the effective roll diameter is smaller than the nominal diameter. Based on the actual spacing and the calculated effective roll diameter, equipment can be retrieved from the process knowledge graph. Equipment meeting preset periodic matching conditions is selected as candidate equipment. The candidate equipment and the effective roll diameter are then integrated to obtain matching parameters. The core logic of the periodic matching condition is that the actual spacing of the periodic defect should be consistent with the periodic characteristics corresponding to the effective roll diameter; that is, the integer multiple deviation between the actual spacing and the effective roll diameter is within a preset allowable range (this range is set according to the defect detection accuracy). For example, node information of all equipment with periodic motion characteristics (such as rolls, drive rolls, etc.) can be extracted from the process knowledge graph first. Combined with the equipment topology, equipment upstream of the defect location (i.e., equipment that may act on the defect area) can be selected. Then, the ratio of the effective roll diameter to the actual spacing of each piece of equipment can be calculated, and equipment with an integer ratio and a deviation within the preset range can be selected as candidate equipment. Finally, the candidate equipment's identifier, model, and calculated effective roller diameter are integrated into matching parameters. After sorting the matching deviations from smallest to largest, a matching result set is constructed, providing accurate candidate equipment and core matching basis for subsequent confidence calculations.

[0110] In one embodiment, based on the defect type in the defect feature matrix, a non-periodic matching strategy is used to perform candidate equipment location processing in the process knowledge graph, generating a matching result set containing matching parameters, including:

[0111] Based on the defect type, defect data of the type of indentation or scratch are selected from the defect feature matrix to obtain a non-periodic defect subset; geometric feature vectors corresponding to each defect are extracted from the non-periodic defect subset, and orientation angle data are extracted from the geometric feature vectors.

[0112] The standard deviation of the orientation angle data is calculated to obtain the orientation consistency index; the strip surface is divided into multiple grid units based on the spatial coordinates of each defect in the non-periodic defect subset;

[0113] The number of defects in each grid cell is statistically processed to obtain the defect density. Based on the defect density, the distribution entropy is calculated to obtain the distribution entropy.

[0114] Equipment with a health index lower than a preset health index threshold and located upstream of the defect location is selected from the process knowledge graph as candidate equipment.

[0115] The candidate devices, directional consistency index, and distribution entropy are processed to generate a matching result set containing matching parameters, including directional consistency index and distribution entropy.

[0116] Specifically, indentation defects are mostly caused by foreign objects on the production line or local bulges in the equipment pressing against the strip, while scratch defects are mostly caused by wear on the cutting edge of the equipment or friction between surface burrs and the strip. Both types of defects are related to non-periodic equipment anomalies (such as localized damage or random entry of foreign objects) and do not exhibit a fixed-interval periodic distribution pattern. Therefore, they can be analyzed separately. For example, the defect type codes stored in the defect feature matrix can be called, and defect records corresponding to the indentation or scratch categories can be filtered out using code matching rules. Core information such as geometric feature vectors, spatial coordinates, and defect identifiers can be extracted from each record and sorted according to the spatial coordinates along the strip length direction to form a subset of non-periodic defects. Since the direction of indentation or scratch defects is directly related to the direction of equipment action (e.g., the scratch direction is consistent with the equipment friction direction, and the indentation direction is related to the foreign object impact direction), and the direction angle is a core feature linking defects and equipment, this parameter can be pre-stored in a specified dimension of the geometric feature vector. Therefore, we can first associate each record in the aperiodic defect subset with its corresponding geometric feature vector through defect identification, extract the dimensional data representing the defect direction from the vector as the original direction angle, and then standardize the original direction angle. That is, taking the strip length direction as 0° as the reference, the original direction angle is uniformly converted into an angle value of 0°-360°. If the defect is linearly extended, the angle of the extension direction is taken as the direction angle; if the defect is an irregular block indentation, the extension direction of the longest axis of the indentation mark is taken as the direction angle. This standardization process can eliminate the difference in angle references between different feature extraction stages and ensure the accuracy of subsequent consistency analysis.

[0117] Specifically, if defects are caused by the same equipment malfunction, their orientation angles will exhibit strong consistency. For example, scratches caused by the same worn part will generally have the same orientation. This consistency can be quantified by calculating the standard deviation; the smaller the standard deviation, the stronger the orientation consistency, indicating a higher probability that the defects originate from the same equipment. For instance, before calculation, outliers in the orientation angle data are removed using the Grubbs criterion to eliminate orientation angle values ​​that deviate from the dataset mean by more than a preset confidence level, thus avoiding interference from individual defects from different sources. Then, the orientation consistency index is calculated using the standard deviation formula:

[0118]

[0119] in, As an indicator of directional consistency, The number of orientation angle data after removing outliers. For the first One direction angle data, for The average value of each orientation angle data. After calculation, As a quantitative result of directional consistency, this indicator can be used to determine whether defects may originate from the same device during subsequent matching.

[0120] Specifically, the spatial distribution of indentation or scratch defects is directly related to the equipment's effective range. For example, the defect density within the effective range of a certain equipment will be significantly higher than in other areas. Mesh generation can convert continuous spatial coordinates into discrete regional units, enabling quantitative statistical analysis of defect distribution density. For instance, the width range is determined by using the two ends of the strip width direction as boundaries, and the length range is determined by using the length coordinates of the first and last defects in the non-periodic defect subset as boundaries, forming a rectangular statistical area encompassing all target defects. This area is then divided into several equally sized mesh units along both the length and width directions. The size of the mesh units can be set according to the typical effective range of the production line equipment, ensuring that the effective range of a single equipment can cover 1-3 mesh units. This avoids overlapping effective areas of different equipment due to excessively large meshes, or scattered defect distribution due to excessively small meshes that cannot be statistically analyzed. After partitioning, a unique identifier can be assigned to each mesh unit, and the spatial coordinate range of each unit can be recorded, establishing a relationship between defects and mesh units (i.e., determining which mesh unit each defect's spatial coordinates fall into and binding the defect identifier to the corresponding mesh identifier).

[0121] The number of defects within each grid cell can then be statistically analyzed to obtain the defect density. Based on this density, distribution entropy can be calculated to obtain the distribution entropy. Defect density reflects the number of defects per unit area, while distribution entropy quantifies the spatial uniformity of the defect density distribution. If defects are concentrated in a few grid cells (low entropy), it indicates that the defects originate from equipment malfunctions in a local area. If they are uniformly distributed (high entropy), they may originate from multiple devices or random factors. For example, all grid cells can be traversed first, and the number of defects bound to each cell can be counted. The defect density can then be calculated based on the area of ​​the grid cell (defect density = number of defects in the cell / cell area). The probability distribution of the defect density for each grid cell can then be calculated. Defect density per grid cell Divide by the sum of the defect densities of all mesh elements The density probability of the cell is obtained. Finally, the distribution entropy is calculated using the information entropy formula, which is:

[0122]

[0123] in, For distribution entropy, For the first Defect density probability of each grid cell This represents the total number of grid cells. The smaller the distribution entropy, the more concentrated the defects are in a specific area, and the higher the likelihood that they are associated with the corresponding equipment in that area.

[0124] Specifically, the occurrence of non-periodic defects is directly related to the health status of equipment. For example, a low equipment health index indicates abnormalities such as wear and loosening, making it more prone to scratches or indentation defects. Furthermore, the equipment must be located upstream of the defect location (in the strip steel production process, abnormalities in upstream equipment will form defects on the downstream strip steel surface, and downstream equipment cannot affect the upstream strip steel area that has already been produced). For instance, the spatial area where defects occur can be first determined (i.e., the rectangular statistical area corresponding to the subset of non-periodic defects). Through the equipment topology relationship interface of the process knowledge graph, all equipment nodes in the opposite direction (i.e., upstream) of this area along the strip steel running direction can be retrieved. Then, the equipment health index is extracted from each equipment node and compared with a preset health index threshold. The preset health index threshold can be calibrated using historical fault data, i.e., statistically analyzing the critical values ​​of the equipment health index when non-periodic defects occur in history. This threshold is then used to filter out equipment with health indices below the threshold, integrating them into a candidate equipment list. The list must include core information such as equipment identification, installation location, and health index value. Furthermore, the directional consistency index and distribution entropy together constitute the quantitative basis for the correlation between defects and candidate devices. That is, for a candidate device, if its direction of action matches the defect directional consistency index (e.g., the angle between the device's friction direction and the defect direction is small) and its area of ​​action coincides with the concentrated area of ​​defect distribution (corresponding to a small distribution entropy), then the correlation is high. Therefore, the candidate device list can be bound to the directional consistency index and distribution entropy data. For each candidate device, auxiliary identifiers of "directional matching degree" and "distribution matching degree" can be added (the directional matching degree is the reciprocal of the directional consistency index, and the distribution matching degree is the reciprocal of the distribution entropy). The devices are then sorted according to the weighted sum of the directional matching degree and the distribution matching degree, with the weights determined based on historical data. Finally, the sorted candidate device list, directional consistency index, distribution entropy, and matching degree identifiers are integrated into a matching result set, which can provide core quantitative parameters for the correlation between defects and devices in subsequent confidence calculations.

[0125] In one embodiment, an equipment health index is obtained from a process knowledge graph; a confidence score is calculated based on the equipment health index and matching parameters in the matching result set to obtain a final confidence score; a comprehensive score is calculated based on the final confidence score and the equipment health index; and a traceability report is generated, including:

[0126] When the matching result set is the matching result set of the corresponding periodic matching strategy, the effective roller diameter and actual spacing are extracted from the matching result set, the spacing error is calculated on the effective roller diameter and actual spacing to obtain the spacing error value, and the confidence is calculated based on the spacing error value to obtain the periodic confidence.

[0127] When the matching result set is the matching result set of the corresponding non-periodic matching strategy, the distribution entropy and directional consistency index are extracted from the matching result set, and the confidence score is calculated based on the distribution entropy and directional consistency index to obtain the non-periodic confidence score.

[0128] The event nodes associated with candidate equipment in the process knowledge graph are retrieved. When there is a temperature disturbance event node, the periodic confidence level or the non-periodic confidence level is enhanced by a preset first enhancement coefficient to obtain the final confidence level. When there is an event node with excessive steel content, the periodic confidence level or the non-periodic confidence level is enhanced by a preset second enhancement coefficient to obtain the final confidence level.

[0129] The equipment health index of candidate equipment is obtained from the process knowledge graph, and a comprehensive score is obtained by weighting the final confidence level and the equipment health index.

[0130] Candidate devices are sorted according to their comprehensive scores, and the final confidence level, comprehensive score, and corresponding matching parameters are used as core evidence to generate a source tracing report.

[0131] Specifically, the effective roll diameter and actual spacing data can be retrieved from the matching result set first, and the nominal roll diameter of the corresponding equipment can be extracted from the equipment nodes of the process knowledge graph. Then, the absolute deviation between the effective roll diameter and the nominal roll diameter is calculated and denoted as... ,in For the effective roller diameter, The nominal diameter of the roll is used; simultaneously, the deviation between the actual gap and the theoretical gap based on the effective roll diameter is calculated, i.e. ,in This is the actual spacing. This refers to the strip length corresponding to one revolution of the equipment during strip operation (i.e., the theoretical defect spacing). The formula for calculating the periodicity confidence level is:

[0132]

[0133] in, The confidence level is periodic, and its value ranges from [0,1]. The spacing matching penalty coefficient, determined by fitting historical defect-equipment matching data, is used to amplify the impact of the deviation between the actual and theoretical spacing on the confidence level. Pi, with a value of 3.1416. This is the vibration stability correction factor. The vibration stability index is calculated from the power spectral density of the vibration spectrum data. Its value is the normalized result of the integral of the deviation between the actual vibration spectrum and the reference spectrum, ranging from [0,1]. A value closer to 1 indicates more stable vibration and a smaller corresponding correction magnitude. Furthermore, if a parameter is missing during the calculation, the historical mean can be used as a substitute to ensure the completeness of the confidence level calculation.

[0134] Specifically, the confidence level of non-periodic defects is positively correlated with the clustering and directional consistency of the defect distribution. Therefore, the maximum reference value of the distribution entropy can be determined first. Maximum reference value for the orientation consistency index (standard deviation of orientation angle) .in The distribution entropy is the value of the defect when it is uniformly distributed across all grid cells. This represents the standard deviation when the orientation angle is randomly distributed within the range of 0°–360°. The non-periodic confidence level can then be calculated using the following formula:

[0135]

[0136] in, For distribution entropy, The standard deviation of the orientation angle is denoted by . This formula achieves the fusion and quantification of two types of indicators by multiplying the distribution entropy and the orientation standard deviation after normalization: the smaller the distribution entropy (the more concentrated the defects) and the smaller the orientation standard deviation (the more consistent the orientation), the closer the non-periodic confidence level is to 1, indicating that the defect is more likely to be associated with the candidate device.

[0137] Furthermore, the event association links in the process knowledge graph can be retrieved using the unique identifier of the candidate equipment to determine whether the equipment is associated with temperature disturbances or excessive steel output during the defect occurrence period. The preset first and second enhancement coefficients are determined through statistical analysis of historical failure cases. For example, if temperature disturbances have a more significant impact on equipment status, the first enhancement coefficient is set higher than the second enhancement coefficient. The calculation logic for the enhancement processing is as follows: ,in To enhance the confidence level before, This represents the enhancement coefficient for the corresponding event. Furthermore, after enhancement, a range constraint needs to be applied to the final confidence level. If >1, then take 1. If the value is less than 0, the value is 0, ensuring that the confidence level is always within the range of [0,1].

[0138] Specifically, the equipment health index of candidate equipment can be obtained from the process knowledge graph. A comprehensive score is obtained by weighting the final confidence level and the equipment health index. A lower equipment health index indicates a worse operating condition and a higher susceptibility to defects; therefore, the comprehensive score must reflect the influence of both confidence level and equipment condition. The weighted summation formula is used in the calculation:

[0139]

[0140] in, The weighting coefficients are determined based on verification data from historical source tracing results. For equipment health index, This indicates the degree of equipment abnormality. After calculation, the comprehensive score is normalized to a range of [0,1], facilitating subsequent sorting and result presentation. Candidate equipment can be sorted from highest to lowest comprehensive score, reflecting the likelihood that a candidate equipment is a defect source. A source tracing report can be generated using the final confidence level, comprehensive score, and corresponding matching parameters as core evidence. The source tracing report must include the candidate equipment's unique identifier, installation location, comprehensive score, final confidence level, and corresponding matching parameters (periodic matching corresponds to effective roller diameter and actual spacing; non-periodic matching corresponds to distribution entropy and directional consistency index), and related event node information (if any). It should be stored in a structured format and support visual output, enabling production personnel to quickly locate high-priority candidate equipment and guide equipment troubleshooting and process optimization.

[0141] Based on the same inventive concept, this application also provides a strip surface defect tracing system for implementing the above-described method for tracing surface defects in strip steel. The solution provided by this system is similar to the implementation described in the above method. Therefore, the specific limitations of one or more embodiments of the strip surface defect tracing system provided below can be found in the limitations of the strip surface defect tracing method described above, and will not be repeated here.

[0142] In one exemplary embodiment, such as Figure 3 As shown, a strip steel surface defect tracing system 300 is provided, comprising:

[0143] The feature extraction and matrix construction module 301 is used to perform defect recognition processing on the strip surface defect image acquired by the vision system, to obtain the defect type, geometric feature vector and texture feature vector, to obtain the spatial coordinates and timestamp corresponding to the strip surface defect image, and to integrate the defect type, spatial coordinates, timestamp, geometric feature vector and texture feature vector to generate a defect feature matrix.

[0144] The process knowledge graph construction module 302 is used to obtain production line equipment parameters, theoretical speed curves and equipment topology relationships, collect real-time operating condition data, and perform process knowledge graph construction processing based on production line equipment parameters, theoretical speed curves, equipment topology relationships, real-time operating condition data and historical event records to generate a process knowledge graph.

[0145] The candidate equipment location module 303 is used to perform candidate equipment location processing in the process knowledge graph according to the defect type in the defect feature matrix, using either a periodic matching strategy or an aperiodic matching strategy, and generate a matching result set containing matching parameters.

[0146] The traceability report generation module 304 is used to obtain the equipment health index from the process knowledge graph, calculate the confidence level based on the equipment health index and the matching parameters in the matching result set, obtain the final confidence level, calculate the comprehensive score based on the final confidence level and the equipment health index, and generate a traceability report.

[0147] In one exemplary embodiment, the present invention also provides a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the strip steel surface defect tracing method of this application. A multi-core processor is preferred to improve the system's parallel processing capability. The memory provides sufficient temporary storage space to support program execution and data processing. The memory capacity should be large enough to accommodate large amounts of data and computational tasks.

[0148] In one exemplary embodiment, the present invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of a method for tracing the source of surface defects in strip steel according to the present application. The computer-readable storage medium may include: a read-only memory, a random access memory, a solid-state drive, or an optical disk, etc.

[0149] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A method for tracing the source of surface defects in strip steel, characterized in that, The method includes: Defect recognition processing is performed on the strip surface defect image acquired by the vision system to obtain the defect type, geometric feature vector and texture feature vector. The spatial coordinates and timestamp corresponding to the strip surface defect image are obtained. The defect type, spatial coordinates, timestamp, geometric feature vector and texture feature vector are integrated to generate a defect feature matrix. Obtain production line equipment parameters, theoretical speed curves, and equipment topology relationships; collect real-time operating condition data; and perform process knowledge graph construction processing based on the production line equipment parameters, theoretical speed curves, equipment topology relationships, real-time operating condition data, and historical event records to generate a process knowledge graph. Based on the defect type in the defect feature matrix, a periodic matching strategy or an aperiodic matching strategy is used to perform candidate equipment location processing in the process knowledge graph to generate a matching result set containing matching parameters. The equipment health index is obtained from the process knowledge graph. Confidence is calculated based on the equipment health index and the matching parameters in the matching result set to obtain the final confidence score. A comprehensive score is calculated based on the final confidence score and the equipment health index to generate a traceability report.

2. The method according to claim 1, characterized in that, The process of performing defect identification processing on the strip surface defect image acquired based on the vision system to obtain the defect type, geometric feature vector, and texture feature vector, and obtaining the spatial coordinates and timestamp corresponding to the strip surface defect image includes: The vision system is used to acquire and process images of the strip surface to obtain defect images of the strip surface; The defect images on the surface of the strip steel are classified using a convolutional neural network based on ResNet50 and spatial pyramid pooling to obtain the defect types. The surface defect image of the strip steel is processed by a preset multi-scale topology network for feature extraction to obtain the geometric feature vector and the texture feature vector; The acquisition location and acquisition time of the strip surface defect image are detected and processed using a position sensor to obtain the spatial coordinates and timestamp corresponding to the strip surface defect image.

3. The method according to claim 1, characterized in that, The process involves acquiring production line equipment parameters, theoretical speed curves, and equipment topology relationships; collecting real-time operating data; and constructing a process knowledge graph based on the production line equipment parameters, theoretical speed curves, equipment topology relationships, real-time operating data, and historical event records. This process knowledge graph includes: The production line equipment parameters, the theoretical speed curve, and the equipment topology are obtained from the manufacturing execution system. The production line equipment parameters include a set of roll diameters. A sensor network is used to collect and process data on the production line's operating status to obtain the real-time operating data, which includes actual speed data and vibration spectrum data. Based on the vibration spectrum data, an equipment health status assessment calculation is performed to obtain the equipment health index; Temperature sensors are used to collect and process the temperature of production line equipment to obtain temperature data. The temperature data is then processed to calculate the change value. When the temperature change value exceeds a preset temperature threshold, a temperature disturbance event node is generated. The steel throughput of the production line is detected and processed. When the steel throughput exceeds the preset steel throughput threshold, a steel throughput over-limit event node is generated. Based on the set of roll diameters and the equipment topology, equipment nodes are constructed to obtain equipment nodes; Historical event records are obtained, and the equipment nodes, temperature disturbance event nodes, steel overload event nodes, equipment topology relationships, and historical event records are associated to generate the process knowledge graph.

4. The method according to claim 1, characterized in that, The step involves using a periodic matching strategy to locate candidate equipment in the process knowledge graph based on the defect types in the defect feature matrix, generating a matching result set containing matching parameters, including: Based on the defect type, defect data of the type of roller printing or vibration pattern are selected from the defect feature matrix to obtain a periodic defect subset; The spatial coordinates of adjacent defects in the periodic defect subset are processed to calculate the spacing, thus obtaining the actual spacing. Based on the timestamps of each defect in the periodic defect subset, the actual speed data for the corresponding time period is extracted from the process knowledge graph; The actual speed integral value is obtained by performing integral calculation based on the actual speed data of the corresponding time period. The theoretical speed curve is obtained from the process knowledge graph, and the theoretical speed curve is integrated to obtain the theoretical speed integral value. The nominal diameter of the roll is obtained from the process knowledge graph. Based on the actual speed integral value, the theoretical speed integral value, and the nominal diameter of the roll, roll diameter compensation calculation is performed to obtain the effective roll diameter. Based on the actual spacing and the effective roll diameter, equipment retrieval processing is performed in the process knowledge graph to retrieve equipment that meets the preset periodic matching conditions as candidate equipment. The candidate equipment and the effective roll diameter are integrated to obtain the matching parameters, and a matching result set containing the matching parameters is generated.

5. The method according to claim 1, characterized in that, The step involves using a non-periodic matching strategy to locate candidate equipment in the process knowledge graph based on the defect type in the defect feature matrix, generating a matching result set containing matching parameters, including: Based on the defect type, defect data of the type of indentation or scratch are selected from the defect feature matrix to obtain a non-periodic defect subset; Extract the geometric feature vector corresponding to each defect from the aperiodic defect subset, and extract the orientation angle data from the geometric feature vector; The standard deviation of the orientation angle data is calculated to obtain the orientation consistency index; The strip surface is divided into multiple grid cells based on the spatial coordinates of each defect in the aforementioned aperiodic defect subset; The number of defects in each grid cell is statistically processed to obtain the defect density, and the distribution entropy is calculated based on the defect density to obtain the distribution entropy; Equipment whose health index is lower than a preset health index threshold and is located upstream of the defect location is retrieved from the process knowledge graph as candidate equipment. The candidate devices, the directional consistency index, and the distribution entropy are processed to generate a matching result set containing the matching parameters, including the directional consistency index and the distribution entropy.

6. The method according to claim 1, characterized in that, The process involves obtaining an equipment health index from the process knowledge graph, calculating a confidence level based on the equipment health index and matching parameters in the matching result set to obtain a final confidence level, calculating a comprehensive score based on the final confidence level and the equipment health index, and generating a traceability report, including: When the matching result set is the matching result set corresponding to the periodic matching strategy, the effective roller diameter and actual spacing are extracted from the matching result set, the spacing error is calculated on the effective roller diameter and the actual spacing to obtain the spacing error value, and the confidence is calculated based on the spacing error value to obtain the periodic confidence. When the matching result set is the matching result set corresponding to the non-periodic matching strategy, the distribution entropy and direction consistency index are extracted from the matching result set, and confidence is calculated based on the distribution entropy and the direction consistency index to obtain the non-periodic confidence. The event nodes associated with the candidate equipment in the process knowledge graph are retrieved. When a temperature disturbance event node exists, the periodic confidence level or the non-periodic confidence level is enhanced by a preset first enhancement coefficient to obtain the final confidence level. When an over-steel quantity event node exists, the periodic confidence level or the non-periodic confidence level is enhanced by a preset second enhancement coefficient to obtain the final confidence level. The equipment health index of the candidate equipment is obtained from the process knowledge graph, and a weighted calculation is performed based on the final confidence level and the equipment health index to obtain the comprehensive score. The candidate devices are sorted according to the comprehensive score, and the final confidence level, the comprehensive score, and the corresponding matching parameters are used as core evidence to generate the source tracing report.

7. The method according to claim 6, characterized in that, The periodic confidence level is calculated using the following formula: in, The periodic confidence level has a value range of [0,1]. The absolute deviation between the effective roll diameter and the nominal roll diameter is, i.e. , The nominal diameter of the roll is... The effective roller diameter, This is the spacing matching penalty coefficient, used to amplify the impact of spacing difference on confidence level. The actual spacing between adjacent periodic defects. The theoretical defect spacing is based on the effective roll diameter. Pi, with a value of 3.1416. This is the vibration stability correction factor. The vibration stability index of the equipment has a value range of [0,1] and is calculated from the power spectral density of the vibration spectrum data.

8. A strip steel surface defect tracing system, characterized in that, The system includes: The feature extraction and matrix construction module is used to perform defect recognition processing on the strip surface defect image acquired by the vision system to obtain the defect type, geometric feature vector and texture feature vector, obtain the spatial coordinates and timestamp corresponding to the strip surface defect image, and integrate the defect type, spatial coordinates, timestamp, geometric feature vector and texture feature vector to generate a defect feature matrix. The process knowledge graph construction module is used to acquire production line equipment parameters, theoretical speed curves, and equipment topology relationships, collect real-time operating condition data, and perform process knowledge graph construction processing based on the production line equipment parameters, theoretical speed curves, equipment topology relationships, real-time operating condition data, and historical event records to generate a process knowledge graph. The candidate equipment location module is used to perform candidate equipment location processing in the process knowledge graph according to the defect type in the defect feature matrix, using either a periodic matching strategy or an aperiodic matching strategy, and generate a matching result set containing matching parameters. The traceability report generation module is used to obtain the equipment health index from the process knowledge graph, calculate the confidence level based on the equipment health index and the matching parameters in the matching result set, obtain the final confidence level, calculate the comprehensive score based on the final confidence level and the equipment health index, and generate a traceability report.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images