A cable defect determination method and system

Abstract

The application discloses a cable defect judgment method and system, relates to the fields of cable quality supervision and machine vision detection technology, and comprises the following steps: acquiring a cable appearance image; identifying a suspected defect area and extracting a shape evidence by a target detection model; pre-constructing a knowledge rule network composed of process state nodes, risk event nodes and defect type nodes based on cable production process knowledge, and calculating process risk prior of each defect type according to current process parameters; fusing the shape evidence and the process risk prior to obtain knowledge rule support, checking detection confidence within a preset amplitude range by using the knowledge rule support, generating final checking confidence and rule conflict markers; and performing hierarchical judgment according to the final checking confidence and the rule conflict markers; and the application combines supervision process information with visual detection, performs auxiliary checking on the detection result by using a knowledge rule network within a limited amplitude, and improves the accuracy and interpretability of cable defect judgment.

Description

Technical Field

[0001] This invention relates to the fields of cable quality supervision and machine vision inspection technology, specifically to a method and system for determining cable defects. Background Technology

[0002] During the production processes of cables, including insulation extrusion, sheath extrusion, cooling and shaping, traction, and winding, surface defects such as scratches, cracks, breaks, bulges, dents, and roughness may occur. Some defects not only affect the appearance quality of the cable but may also reduce the integrity of the sheath or insulation layer, increasing quality risks during subsequent transportation, laying, and operation.

[0003] Currently, patent CN113947563A discloses a dynamic defect detection method for cable manufacturing quality based on deep learning. This method uses the YOLOv5 model to locate cable regions, uses the UNet segmentation model with MobileNet-V2 as the backbone network to extract cable features, and then judges whether the cable is abnormal based on the cable features. This scheme uses a deep learning model for cable image detection, and outputs the suspected defect areas on the cable surface and their detection confidence through the target detection model. Its judgment process mainly relies on image features and the confidence of the model output to determine defects.

[0004] However, in actual cable manufacturing supervision scenarios, the acquired images may have visual features similar to defects due to reflections, stains, printed textures, and local shadows. Existing methods that only rely on the output confidence of the target detection model to issue alarms may still produce false alarms, resulting in low accuracy of defect judgment results. Summary of the Invention

[0005] To address the shortcomings of existing technologies that rely solely on the output confidence level of target detection models for alarms, which may lead to false alarms and low accuracy in defect determination, this invention proposes a cable defect determination method and system. By constructing a supervision knowledge rule network of "process status - risk event - defect type", supervision experience is transformed into calculable process risk priors. Combined with morphological evidence of suspected defect areas, the confidence level output by the target detection model is dynamically verified within a limited range, thereby solving the problems existing in the prior art.

[0006] A method for determining cable defects includes the following steps: Acquire cable appearance images during the cable manufacturing process, and extract suspected defect areas on the cable surface from the cable appearance images and their detection confidence levels; Based on cable manufacturing process knowledge, supervision experience, and historical defect records, a supervision knowledge rule network is pre-constructed, consisting of process state nodes representing deviations from normal ranges of process parameters, risk event nodes representing risks caused by abnormal process parameters, and defect type nodes representing cable appearance defects. The activation degree of each process state node is calculated based on the process parameters during the current cable supervision process. The activation degree of each process state node is then forward-propagated according to the weights of the rule edges in the supervision knowledge rule network to calculate the activation degree of each risk event node. Based on the activation degree of each risk event node, the corresponding process risk prior for each defect type node is calculated. Morphological evidence of suspected defect areas is extracted and combined with prior process risks of corresponding defect types to dynamically verify the detection confidence level; based on the final verification confidence level, suspected defect areas are classified and judged to generate cable defect judgment results.

[0007] Furthermore, the calculation of the activation level of each process state node based on the process parameters in the current cable manufacturing process specifically includes the following steps: Regarding the first Each process parameter is defined as a production process parameter value. The reference value is The allowable deviation is Then the activation level of the process state node is: This represents a preset constant to prevent the denominator from being zero; Regarding the first For each risk event node, calculate the activation level of the risk event node based on all process status nodes pointing to that risk event node: ; in: Indicates the first Each risk event node at the detection time activation level, This indicates all nodes pointing to risk events. The set of preceding process state nodes, Indicates process status node activation level, Indicates process status node Pointing to risk event nodes The rule edge weights.

[0008] Furthermore, the calculation of the process risk prior corresponding to each defect type node based on the activation degree of each risk event node specifically includes the following steps: Regarding the first For each defect type node, calculate the corresponding process risk prior based on all risk event nodes pointing to that defect type node: ; in: Indicates the first Class defects at the time of detection Prior knowledge of process risks Indicates all nodes pointing to the defect type. The set of preceding risk event nodes, Indicates risk event nodes Pointer to defect type node The rule edge weights.

[0009] Furthermore, the determination of the rule edge weights includes setting initial weights based on supervision experience or, when historical data is available, adjusting the rule edge weights based on the number of occurrences of the preceding state and the number of occurrences of the corresponding subsequent state; wherein, when the amount of historical data is insufficient, the initial weights set based on supervision experience are used; as historical data gradually accumulates, the rule edge weights are adjusted using historical records.

[0010] Furthermore, the morphological evidence of the suspected defective area includes an indication of the area ratio of the suspected defective area relative to the cable body area. The aspect ratio of the detection box is represented by and . ;in, Indicates the area of ​​the main cable section. This represents a preset constant to prevent the denominator from being zero. These represent the width and height of the detection frame, respectively.

[0011] Furthermore, the suspected defect areas on the cable surface and their detection confidence scores are extracted from the cable appearance image. Specifically, the cable appearance image is input into a pre-trained target detection model, which outputs the suspected defect areas and detection confidence scores on the cable surface.

[0012] Furthermore, the extraction of morphological evidence from suspected defect areas, combined with prior knowledge of process risks corresponding to the defect type, and the dynamic verification of the detection confidence level specifically include the following steps: Extract morphological evidence from suspected defect areas, and match each morphological feature in the morphological evidence with the morphological rule reference interval of the corresponding defect type to calculate the degree of morphological feature matching. ; in: Indicates the first The first suspected defect area individual morphological features, This indicates the allowable buffer range when morphological features deviate from the reference range. This indicates the degree of matching between a single morphological feature and the corresponding defect type rule. The range of values ​​is to ; Indicates the first The first class defect corresponding to Lower limit of each morphological feature; Indicates the first The first class defect corresponding to The upper limit of the morphological features; the first Class defects, the first The reference interval for each morphological feature is represented as follows: ; Based on morphological evidence vectors Various characteristics and categories Calculate the morphological evidence matching score based on the degree of matching between the corresponding morphological rule intervals: ; in: Indicates the first Morphological evidence matching score of suspected defective areas, Indicates the number of morphological features, Indicates the first Weights of each morphological feature; The detection confidence level is dynamically verified based on the morphological evidence matching score and the prior process risk of the corresponding defect type.

[0013] Furthermore, the dynamic verification of the detection confidence based on the morphological evidence matching score and the corresponding process risk prior of the defect type specifically includes the following steps: The knowledge rule support is calculated based on the morphological evidence matching score and the corresponding defect type's process risk prior, specifically expressed as follows: ; in: Representing knowledge rules for the first The level of support for each suspected defective area; This indicates the weight of prior process risks. ; Indicates the current process status relative to the category. Risk prior; Indicates the current suspected defect area and category The degree of matching of morphological rules; The detection confidence level is dynamically verified within a preset range using the knowledge rule support.

[0014] Furthermore, the dynamic verification of the detection confidence using knowledge rule support within a preset range specifically includes the following steps: The detection confidence score output by the target detection model According to the support of knowledge rules Detection confidence Perform dynamic verification: ; in: This indicates the initial confidence level after verification by the knowledge rule support. Indicates the strength of the knowledge rule verification; At the same time, knowledge rules provide positive support for visual detection results; At the same time, knowledge rules provide inverse inhibition on visual detection results; At the same time, the original confidence level of the visual detection model is not changed; The range of confidence level adjustments caused by knowledge rules is limited, and the final confidence level after verification is: ; in: Indicates the first i The final verification confidence level for each suspected defective area This represents the initial confidence level calculated based on the support of knowledge rules. This represents the raw confidence score output by the object detection model. and These represent the maximum allowed downward correction range and the maximum allowed upward correction range, respectively. ; The operation indicates that the initial confidence level is limited to a given range; When the confidence level of the visual detection model is higher than the high confidence threshold for visual detection, and the support level of the knowledge rule is lower than the minimum support threshold for the knowledge rule, a rule conflict marker is generated. ; in: Indicates the first Does the suspected defective area have any rule conflicts? This represents the high-confidence threshold for visual detection. This represents the minimum support threshold for knowledge rules.

[0015] The present invention also includes a cable defect determination system, comprising: The acquisition module is used to acquire cable appearance images during the cable manufacturing process and extract suspected defect areas on the cable surface from the cable appearance images and their detection confidence levels. The prior calculation module is used to pre-construct a supervision knowledge rule network based on cable production process knowledge, supervision experience, and historical defect records. This network consists of process state nodes representing deviations from normal ranges of process parameters, risk event nodes representing risks caused by abnormal process parameters, and defect type nodes representing cable appearance defects. The module calculates the activation degree of each process state node based on the process parameters during the current cable supervision process. It then forward-propagates the activation degree of each process state node according to the weights of the rule edges in the supervision knowledge rule network to calculate the activation degree of each risk event node. Finally, it calculates the process risk prior corresponding to each defect type node based on the activation degree of each risk event node. The judgment module is used to extract morphological evidence of suspected defect areas and combine it with the process risk prior of the corresponding defect type to dynamically verify the detection confidence; based on the final verification confidence, the suspected defect areas are classified and judged to generate cable defect judgment results.

[0016] This invention provides a method for determining cable defects, which has the following advantages: This invention constructs a three-layer supervision knowledge rule network consisting of process status nodes, risk event nodes, and defect type nodes. It transforms the process parameters obtainable during cable production into quantifiable process risk priors and combines them with morphological evidence of suspected defect areas to dynamically verify the original detection confidence of suspected defect areas within a limited range. When there are interfering factors such as reflections, stains, printed textures, and local shadows in the image, this method can use process status information to assist in verifying the visual inspection results. This effectively overcomes the technical shortcomings of high false alarm rate and low accuracy caused by relying solely on image features and model output confidence for defect judgment, and improves the accuracy and interpretability of cable defect judgment. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the intelligent cable defect determination method in an embodiment of the present invention. Figure 2 This is a schematic diagram of the three-layer structure of the manufacturing supervision knowledge rule network in this embodiment of the invention: "process status node - risk event node - defect type node". Figure 3 This is a schematic diagram of the process for bounded dynamic verification of visual detection confidence in an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0019] Traditional cable quality supervision mainly relies on manual visual inspection, sampling records, and production parameter records. While manual visual inspection can combine on-site experience to identify defects, its efficiency is affected by personnel experience, workload, and shooting conditions, and different supervisors may have different standards for judging suspected defects. Judging product quality solely based on production parameters also has limitations: abnormal process parameters do not necessarily lead to defects, and even when process parameters are within the normal range, appearance defects may still occur due to equipment wear, localized material abnormalities, or occasional mechanical interference. Currently, patent CN111539935B discloses an online detection method for cable surface defects based on machine vision. This method extracts the contours of cable surface defects and calculates the defect area through image preprocessing, grayscale stretching, illumination compensation, adaptive median filtering, fixed threshold binarization, and morphological processing. This scheme can improve the efficiency of online detection of cable surface defects, but its judgment process mainly relies on image features and does not use the process status information obtained during supervision for defect reliability verification. Patent CN113947563A discloses a deep learning-based dynamic defect detection method for cable manufacturing processes. This method uses a YOLOv5 model to locate cable regions, extracts cable features using a UNet segmentation model with MobileNet-V2 as the backbone, and then determines whether the cable is abnormal based on the segmentation features. This scheme uses a deep learning model for cable image detection, but still relies primarily on image processing and feature comparison for judgment. Patent CN114757949B discloses a computer vision-based method and system for detecting defects in wires and cables. This method obtains defect regions through grayscale features, region growing, image enhancement, and edge detection, and adjusts the process by a production process controller after identifying the defect type. This scheme can link the production process after detection, but it does not disclose how to form a priori defect risk based on the current process state, nor how to use the priori process risk to verify the results of a single visual inspection.

[0020] However, in cable manufacturing supervision scenarios, the acquired images may have visual features similar to defects due to reflections, stains, printed textures, and local shadows. Relying solely on the output confidence of the target detection model to issue alarms may still result in false alarms, leading to low accuracy in defect determination.

[0021] Based on this, this invention addresses the problems of existing cable appearance defect detection methods, which mainly rely on image features and struggle to fully utilize process status information during the manufacturing process, and whose target detection models are easily affected by reflections, stains, printed textures, and local shadows. It proposes an intelligent cable defect judgment method driven by manufacturing supervision knowledge rules. This method uses a target detection model to quickly locate suspected defect areas in cable appearance images. Without requiring a large number of process parameters, it associates cable batches, process types, and a few key process parameters with corresponding images, transforming cable manufacturing supervision experience into a three-layer knowledge rule network of "process status—risk event—defect type." Based on the current process status, it calculates the process risk prior for each defect type. The process risk prior and morphological evidence of suspected defect areas are used for visual inspection confidence verification. High-risk results, conflicting results, and low-confidence results are handled differently, outputting defect judgment records that allow for tracing the source of risk, providing a basis for manual review, production adjustments, and the preparation of manufacturing supervision reports.

[0022] like Figure 1 As shown, the method specifically includes the following steps: S1. Obtain cable manufacturing supervision images and process information.

[0023] S1.1 Obtain cable manufacturing supervision images and related information Acquire cable appearance images collected during cable production, acceptance, or supervision, and obtain the supervision-related information corresponding to the images.

[0024] The cable appearance image can be obtained by a fixed industrial camera, a mobile shooting device, or video frame extraction; the image must include at least the main body area of ​​the cable to be inspected and the surface area to be inspected. Supervision-related information includes the cable batch number, inspection time, and process type; when data is available, it also includes at least one process parameter corresponding to the current process. Process parameters are preferentially selected from extrusion temperature, traction speed, and cooling temperature. If process parameters are temporarily unavailable, the process parameter set is allowed to be empty, and degradation processing is performed according to the method described in step S3.4.

[0025] This invention does not require the simultaneous acquisition of all process parameters, nor does it require the additional collection of a large number of fields such as work order number, specifications, production date, insulation thickness, sheath thickness, and take-up tension. In this step, the default image shooting angle meets the detection requirements, and image rotation, tilt correction, or cable axial direction estimation is not performed.

[0026] S1.2 Obtaining production process parameter values ​​based on time windows Because production line process parameters may fluctuate instantaneously, to avoid anomalies at a single sampling point from excessively affecting subsequent judgments, the time of image acquisition is used as the reference. Set a time window at the center: in, This is the preset time window length.

[0027] Regarding the first For each process parameter, obtain the parameter sample values ​​within a time window and calculate the parameter median: in: Indicates the first j Each process parameter at time... The sampled values; This represents the production process parameter value associated with the current monitoring image; This indicates the median calculation operation.

[0028] Compared to using instantaneous parameter values ​​directly, using the median value within a time window can reduce the impact of sensor noise, occasional jumps, and short-term fluctuations on subsequent risk calculation results.

[0029] S1.3, Constructing the Supervision Data Unit For the time of data collection Cable manufacturing monitoring images, constructing monitoring data tuples : in: Image showing the appearance of the cable; Indicates the cable batch number; Indicates the detection time; Indicates the process type; Represents a set of production process parameter values; m This indicates the number of currently available process parameters, and ,when At that time, set a neutral process risk prior in accordance with the method described in step S3.4.

[0030] Image of cable appearance Used for suspected defect detection in step S2; process type and set of production process parameter values Used for process risk prior calculation in step S3; cable batch numbering Detection time and process type Used for result association and storage in step S5.

[0031] S2. Cable Defect Target Detection and Defect Morphology Evidence: S2.1 Identify suspected defect areas using a target detection model. The cable appearance image obtained in step S1 The system takes a pre-trained target detection model as input and outputs suspected defect areas on the cable surface. The target detection model can employ a YOLO series target detection network or other target detection networks with equivalent target localization and classification capabilities. In this invention, the target detection model is used to output candidate results for suspected defects; the model's detection results are not directly used as the final manufacturing inspection judgment.

[0032] For the The visual inspection results for a suspected defective area are as follows: in: This indicates the suspected defect category output by the target detection model. The location of the detection box indicating the suspected defect area, This represents the original detection confidence score output by the object detection model, and Suspected defects include at least one of the following: scratches, cracks, breaks, bulges, and dents.

[0033] The location of the detection box is represented as follows: in: These represent the x-coordinate and y-coordinate of the center of the detection frame, respectively. These represent the width and height of the detection frame, respectively.

[0034] S2.2 Obtain images of suspected defect areas. Based on the position of the detection frame Images of suspected defective areas are extracted from the original cable appearance image: in: Indicates the first Image of a suspected defect area This indicates that a region capture operation is performed based on the detection box.

[0035] The main cable area can be output synchronously by the target detection model, or a preset shooting area can be used as the main cable area, and its pixel area can be calculated. .

[0036] S2.3 Extracting evidence of suspected defect morphology For each suspected defective area, two morphological indicators, relative area and aspect ratio, are extracted.

[0037] (1) Relative area: Let the area of ​​the cable region be... , No. The area of ​​each suspected defect detection frame is The relative area of ​​the suspected defect region is: in: Indicates the area ratio of the suspected defective area to the main cable area. Indicates the area of ​​the main body of the cable, This represents a preset constant to prevent the denominator from being zero.

[0038] (2) Aspect Ratio: The aspect ratio of the suspected defective area is: in: The aspect ratio indicates the ratio between the longer and shorter sides of the detection frame. The aspect ratio is used to help distinguish different appearances such as slender scratches, local cracks, blocky damage, bulges and dents.

[0039] S2.4 Form a candidate set of suspected defects for use in the visual detection confidence verification in step S4.

[0040] Regarding the first A suspected defect area forms a morphological evidence vector: Visual inspection results With morphological evidence vector By performing correlation, a set of potential defect candidates is obtained: in: This indicates the number of suspected defect areas output by the target detection model. Indicates the first The category, location, and initial confidence level of each suspected defect. Indicates the first A vector of morphological evidence suggesting a defect.

[0041] S3. Prior calculation of process risks driven by supervision knowledge rules: S3.1 Constructing a knowledge rule network for supervision Based on cable manufacturing process knowledge, supervision experience, and historical defect records, a supervision knowledge rule network is constructed: in: Represents a set of rule nodes. Represents the set of directed associations between nodes. This represents the set of weights corresponding to a directed association.

[0042] rule node set Includes three types of nodes: in: Represents the set of process state nodes; Represents a set of risk event nodes; This represents a set of nodes representing defect types.

[0043] Process status nodes are used to indicate whether process parameters deviate from the normal range, including at least one of the following: extrusion temperature too high, extrusion temperature too low, traction speed too high, traction speed too low, and cooling temperature abnormal.

[0044] Risk event nodes are used to represent intermediate risk states that may be caused by abnormal process parameters, including at least one of the following: abnormal plasticizing state, unstable surface forming, abnormal traction matching, abnormal cooling and shaping, and local mechanical friction risk.

[0045] The defect type node is used to indicate the type of cable appearance defect to be determined, including at least one of scratches, cracks, breaks, bulges, and dents.

[0046] A directed association relationship is established according to the three-layer structure of "process status - risk event - defect type".

[0047] Exemplary rules include: (1) Extrusion temperature is too low → abnormal plasticization state → cracks.

[0048] (2) Extrusion temperature is too high → surface molding is unstable → bulging.

[0049] (3) The traction speed is too high → the traction matching is abnormal → the dent is formed.

[0050] (4) Excessive traction speed → Risk of local mechanical friction → Scratches.

[0051] (5) Abnormal cooling temperature → Abnormal cooling and shaping → Cracks.

[0052] (6) Abnormal cooling temperature → Abnormal cooling and shaping → Dent

[0053] The rules do not imply that any abnormal process condition will necessarily lead to a corresponding defect, but rather describe the degree to which the process condition affects the risk of defect occurrence.

[0054] S3.2 Calculate the activation level of process state nodes Regarding the first There are several process parameters, let their production process parameter values ​​be... The reference value is The allowable deviation is The degree of deviation in process state is: in: Indicates the first The activation level of each process parameter corresponds to the state node. The value range is 0 to 1. Indicates the first Reference values ​​for each process parameter Indicates allowable deviation, This represents a preset constant to prevent the denominator from being zero.

[0055] When it is necessary to distinguish between parameters that are too high and too low, according to The positive or negative sign determines the corresponding process state node.

[0056] For example: when When, activate the "parameter too high" node; when At that time, the "parameters are too low" node is activated.

[0057] S3.3 Determine the weights of the rule edges For any directed association Set the edge weights of the rules ,and: The edge weights are used to represent the degree of influence of the preceding node on the succeeding node.

[0058] The weight of a regular edge can be determined using one of the following two methods: (1) Method 1: Set initial weights based on supervision experience: The initial weights are set by cable production personnel, supervision personnel or quality management personnel based on historical experience.

[0059] For example: Extrusion temperature too low → abnormal plasticization state: 0.8; Cooling temperature abnormal → abnormal cooling and shaping: 0.7; Cooling and shaping abnormal → cracks: 0.6.

[0060] (2) Method 2: Adjusting rule weights using historical records: When historical data is available, adjust the rule edge weights based on the number of occurrences of the preceding state and the number of occurrences of the corresponding succeeding state. in: Represents the preceding node With subsequent nodes Number of times they appear together Represents the preceding node Total number of occurrences This represents the smoothing coefficient.

[0061] When the amount of historical data is insufficient, the initial weights set by the supervision experience are used; as historical data gradually accumulates, the weights of the rule edges are adjusted using historical records.

[0062] S3.4 Calculate and output the prior knowledge of process risks. Regarding the first For each risk event node, calculate the activation level of the risk event node based on all process status nodes pointing to that risk event node: in: Indicates the first Each risk event node at the detection time activation level, This indicates all nodes pointing to risk events. The set of preceding process state nodes, Indicates process status node activation level, Indicates process status node Pointing to risk event nodes The rule edge weights.

[0063] Regarding the first For each defect type node, calculate the corresponding process risk prior based on all risk event nodes pointing to that defect type node: in: Indicates the first Class defects at the time of detection Prior knowledge of process risks Indicates all nodes pointing to the defect type. The set of preceding risk event nodes, Indicates risk event nodes Pointer to defect type node The edge weights of the rules. The range of values ​​for the prior process risk is... to The greater the prior process risk, the more likely the current process state is to produce the corresponding type of defect; the smaller the prior process risk, the less likely the current process state is to clearly support the defect type.

[0064] When a certain process parameter cannot be obtained, the process status node corresponding to that parameter is not activated.

[0065] When all process parameters are temporarily unavailable, set the process risk priors corresponding to each defect type to neutral values: This ensures that the visual inspection function can still operate independently, and the entire judgment process will not be interrupted due to the lack of some process data.

[0066] The final output is the prior set of process risks corresponding to the current detection moment: in: Indicates the number of defect types, Indicates the first Prior process risks corresponding to class defects.

[0067] S4, Detection Confidence Verification Driven by Both Knowledge and Vision S4.1 Calculate the morphological evidence matching score For each type of defect, a morphological evidence reference range is set.

[0068] With the first Taking a type of defect as an example, its first The reference interval for each morphological feature is represented as follows: in: Indicates the first The first class defect corresponding to Lower limit of each morphological feature; Indicates the first The first class defect corresponding to The upper limit of each morphological feature; the morphological rules for different defect types can be set based on the statistical results of existing defect samples and the experience of the supervisors.

[0069] For example, scratches are usually characterized by a relatively small area and a large length-to-width ratio; bulges are usually characterized by a relatively small area and a relatively small length-to-width ratio; and localized damage is usually characterized by a relatively large area. Regarding the first The first suspected defect area, the first Class defects and the first Calculate the degree of morphological feature matching for each morphological feature: in: Indicates the first The first suspected defect area individual morphological features, This indicates the allowable buffer range when morphological features deviate from the reference range. This indicates the degree of matching between a single morphological feature and the corresponding defect type rule. The range of values ​​is to .

[0070] Based on morphological evidence vectors Various characteristics and categories Calculate the morphological evidence matching score based on the degree of matching between the corresponding morphological rule intervals: in: Indicates the first Morphological evidence matching score of suspected defective areas, Indicates the number of morphological features, Indicates the first Weights of morphological features .

[0071] In this invention, a setting is provided. The relative area and aspect ratio are used as two morphological features.

[0072] S4.2, Computational Knowledge Rule Support Regarding the first Each suspected defect area is categorized according to its corresponding defect type. Prior to process risks Matching score with morphological evidence Calculate the support degree of knowledge rules: in: Representing knowledge rules for the first The degree of support for each suspected defective area This indicates the weight of prior process risks. , Indicates the current process status relative to the category. Risk prior, Indicates the current suspected defect area and category The degree of matching of morphological rules.

[0073] S4.3 Bounded dynamic verification of visual detection confidence. The raw detection confidence score output by the object detection model First, calculate its logarithmic probability: Then, based on the support of knowledge rules Dynamically verify the original detection confidence level: in: This indicates the initial confidence level after verification by the knowledge rule support. Indicates the strength of knowledge rule verification; At the same time, knowledge rules provide positive support for visual detection results; At the same time, knowledge rules provide inverse inhibition on visual detection results; At the same time, the original confidence level of the visual detection model is not changed.

[0074] Considering the potential for data acquisition errors or instantaneous fluctuations in process parameters, the adjustment range for the confidence level caused by the knowledge rules is limited. The final confidence level after verification is:

[0075] in: Indicates the first i The final verification confidence level for each suspected defective area This represents the initial confidence level calculated based on the support of knowledge rules. This represents the raw confidence score output by the object detection model. and These represent the maximum allowed downward correction range and the maximum allowed upward correction range, respectively. The strength of the knowledge rule verification and the maximum correction range can be determined based on the verification samples or the monitoring and alarm strategy. The operation means limiting the initial confidence level to a given range.

[0076] Through the above processing, the knowledge rules are made to assist in verifying the visual inspection results only within a preset range, avoiding excessive correction of the inspection results due to abnormal process parameters.

[0077] S4.4 Generate rule conflict markers and output verification results When the visual detection model has high confidence but low knowledge rule support, rule conflict markers are generated: in: Indicates the first Does the suspected defective area have any rule conflicts? This represents the high-confidence threshold for visual detection. This represents the minimum support threshold for knowledge rules.

[0078] The rule conflict marker is not used to directly negate the visual inspection result, but to transfer the corresponding area to the manual review process.

[0079] Output of this step: in: Indicates the suspected defect category. Indicates the position of the detection box. This represents the original confidence level of the object detection model. The score represents the matching score of morphological evidence. Indicates the degree of support for knowledge rules. Indicates the final verification confidence level. Indicates a rule conflict flag.

[0080] S5. Intelligent Defect Judgment and Supervision Prompt Generation S5.1. Classify and determine based on the final verification confidence level. Set defect confirmation threshold and threshold for review ,and: .

[0081] Regarding the first One suspected defect area, based on the final verification confidence level. Conflict flags with rules A classification judgment is made.

[0082] (1) Confirming the defect: When the following conditions are met, the defect will be identified. The suspected defect area was confirmed as a defect: (2) Pending manual review: When any of the following conditions are met, the first... i The following suspected defective areas have been identified as areas requiring manual review: or: (3) Low-risk candidate areas When the following conditions are met, the first i One suspected defective region was identified as a low-risk candidate region: Low-risk candidate regions do not immediately generate defect alarms, but can be retained in the system background for subsequent tracing and model optimization.

[0083] S5.2, Identify the main sources of risk For areas with confirmed defects or awaiting manual review, identify the risk propagation path that contributes the most to the current defect category.

[0084] Set process state nodes Risk event nodes and defect category nodes If a risk transmission path is established, then the contribution of that path is: Regarding the first For each suspected defective region, select the path that contributes the most: in: Indicates the first i The main risk source paths corresponding to each suspected defect area Indicates the activation level of process status nodes. This indicates the weight between process status nodes and risk event nodes. This indicates the weight between risk event nodes and defect category nodes.

[0085] For example: excessively high traction speed → risk of localized mechanical friction → scratches.

[0086] S5.3 Generate supervision prompt information For confirmed defects and areas requiring manual review, corresponding supervision prompts are generated based on the main risk source paths.

[0087] Example prompts include: (1) When the main risk path is “high traction speed → local mechanical friction risk → scratches”, the output prompt is: It is recommended to check the changes in traction speed, the contact parts of the traction device and the mechanical friction.

[0088] (2) When the main risk path is “abnormal cooling temperature → abnormal cooling and shaping → crack”, the output prompt is: It is recommended to check the cooling temperature record and cooling and shaping status.

[0089] (3) When a rule conflict is flagged When the visual inspection results conflict with the current process status, it is recommended to manually review the image and check the completeness of the process parameter records.

[0090] The prompts are intended to assist supervisors in making judgments and tracing processes, and do not directly execute production equipment control operations.

[0091] S5.4 Generate intelligent cable defect judgment record For each confirmed defect or area requiring manual review, generate an intelligent cable defect judgment record: in: Indicates the cable batch number; Indicates the detection time; Indicates the process type; This represents the corresponding cable appearance image; Indicates the suspected defect category; Indicates the location of a suspected defect; This represents the original confidence level of the object detection model; Indicates the final verification confidence level; Indicates the degree of support for knowledge rules; Indicates a rule conflict flag; Indicate the main sources and pathways of risk; This indicates a construction supervision notice.

[0092] This invention transforms a small number of process parameters during the manufacturing supervision process into process risk priors corresponding to different defect types. Using these process risk priors and evidence of suspected defect morphology, the visual confidence level output by the target detection model is dynamically verified within a limited range. The process risk priors do not directly replace the visual inspection results, nor are they simply linearly weighted with the visual confidence level. Instead, they serve as an external verification basis, appropriately enhancing, suppressing, or marking conflicts in the visual inspection results. This avoids excessive interference from process parameter recording errors or occasional fluctuations in image judgment. This invention does not require image rotation, tilt correction, or defect direction estimation. During image acquisition, the cable body to be inspected is positioned at a preset shooting angle, or an existing image with an angle that meets the manufacturing supervision requirements is directly used. Subsequent processing assumes the image angle is correct.

[0093] Compared with the prior art, the present invention has the following beneficial effects: (1) It can achieve lightweight integration of manufacturing process information and visual inspection results: This invention only needs to obtain the cable batch number, inspection time, process type and at least one key process parameter to associate the current process status with the cable appearance image. By suppressing instantaneous fluctuations through robust parameter values ​​within the time window, the influence of sensor noise and occasional jumps on the judgment results is reduced.

[0094] (2) Improves the reliability of cable defect determination: This invention constructs a three-layer knowledge rule network of "process status - risk event - defect type", transforms supervision experience into calculable process risk priors, and combines morphological evidence of suspected defect areas to dynamically verify the confidence level output by the target detection model within a limited range. This can reduce false alarms caused by reflections, stains, printed textures and local shadows, while avoiding excessive interference of abnormal process parameters with visual inspection results.

[0095] (3) Differentiated handling of conflict results: When the visual inspection result is significantly inconsistent with the knowledge rule support, this invention does not directly delete the suspected defect, but generates a rule conflict marker and transfers the corresponding area to the manual review process. This method can reduce false alarms while reducing the risk of real defects being incorrectly removed.

[0096] (4) It can generate interpretable and traceable supervision judgment records: This invention associates and saves defect categories, defect locations, original detection confidence levels, final verification confidence levels, main risk source paths, and supervision prompts, providing a basis for subsequent manual review, quality tracking, and supervision report preparation. This invention is mainly implemented through software, does not rely on complex production line hardware, and is easy to deploy and apply in cable supervision scenarios.

[0097] Based on the same inventive concept, this invention also proposes a cable defect determination system, comprising: The acquisition module is used to acquire cable appearance images during the cable manufacturing process and extract suspected defect areas on the cable surface from the cable appearance images and their detection confidence levels.

[0098] The prior calculation module is used to pre-construct a supervision knowledge rule network based on cable production process knowledge, supervision experience, and historical defect records. This network consists of process state nodes representing deviations from normal process parameters, risk event nodes representing risks caused by abnormal process parameters, and defect type nodes representing cable appearance defects. The module calculates the activation degree of each process state node based on the process parameters during the current cable supervision process. It then forward-propagates the activation degree of each process state node according to the weights of the rule edges in the supervision knowledge rule network to calculate the activation degree of each risk event node. Finally, it calculates the prior process risk corresponding to each defect type node based on the activation degree of each risk event node.

[0099] The judgment module is used to extract morphological evidence of suspected defect areas and combine it with the process risk prior of the corresponding defect type to dynamically verify the detection confidence; based on the final verification confidence, the suspected defect areas are classified and judged to generate cable defect judgment results.

[0100] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for determining cable defects, characterized in that, Includes the following steps: Acquire cable appearance images during the cable manufacturing process, and extract suspected defect areas on the cable surface from the cable appearance images and their detection confidence levels; Based on cable production process knowledge, supervision experience and historical defect records, a supervision knowledge rule network is pre-constructed, consisting of process status nodes that characterize process parameters deviating from the normal range, risk event nodes that characterize the risks caused by abnormal process parameters, and defect type nodes that characterize the types of cable appearance defects. The activation degree of each process state node is calculated based on the process parameters in the current cable manufacturing process. The activation degree of each process state node is then forward-propagated according to the weight of the rule edge in the manufacturing knowledge rule network to calculate the activation degree of each risk event node. The process risk priors corresponding to each defect type node are calculated based on the activation degree of each risk event node. The morphological evidence of suspected defect areas is extracted and combined with the prior process risk of the corresponding defect type to dynamically verify the detection confidence. Based on the final verification confidence level, suspected defect areas are classified and judged to generate cable defect judgment results.

2. The cable defect determination method according to claim 1, characterized in that, The calculation of the activation level of each process state node based on the process parameters in the current cable manufacturing process specifically includes the following steps: Regarding the first Each process parameter is defined as a production process parameter value. The reference value is The allowable deviation is Then the activation level of the process state node is: , This represents a preset constant to prevent the denominator from being zero; Regarding the first For each risk event node, calculate the activation level of the risk event node based on all process status nodes pointing to that risk event node: ； in: Indicates the first Each risk event node at the detection time activation level, This indicates all nodes pointing to risk events. The set of preceding process state nodes, Indicates process status node activation level, Indicates process status node Pointing to risk event nodes The rule edge weights.

3. The cable defect determination method according to claim 2, characterized in that, The calculation of the process risk prior for each defect type node based on the activation degree of each risk event node specifically includes the following steps: Regarding the first For each defect type node, calculate the corresponding process risk prior based on all risk event nodes pointing to that defect type node: ； in: Indicates the first Class defects at the time of detection Prior knowledge of process risks Indicates all nodes pointing to the defect type. The set of preceding risk event nodes, Indicates risk event nodes Pointer to defect type node The rule edge weights.

4. The cable defect determination method according to claim 3, characterized in that, The determination of the rule edge weights includes setting initial weights based on supervision experience or, when historical data is available, adjusting the rule edge weights based on the number of occurrences of the preceding state and the number of occurrences of the corresponding subsequent state. Specifically, when the amount of historical data is insufficient, the initial weights set based on supervision experience are used; as historical data gradually accumulates, the rule edge weights are adjusted using historical records.

5. The cable defect determination method according to claim 1, characterized in that, Morphological evidence of the suspected defect area includes an indication of the area ratio of the suspected defect area relative to the main cable body area. The aspect ratio of the detection box is represented by and . ;in, Indicates the area of ​​the main cable section. This represents a preset constant to prevent the denominator from being zero. These represent the width and height of the detection frame, respectively.

6. The cable defect determination method according to claim 1, characterized in that, The suspected defect areas and their detection confidence scores are extracted from the cable appearance image. Specifically, the cable appearance image is input into a pre-trained target detection model, which outputs the suspected defect areas and detection confidence scores of the cable surface.

7. The cable defect determination method according to claim 6, characterized in that, The process of extracting morphological evidence of suspected defect areas and combining it with prior process risks corresponding to the defect type to dynamically verify the detection confidence level specifically includes the following steps: Extract morphological evidence from suspected defect areas, and match each morphological feature in the morphological evidence with the morphological rule reference interval of the corresponding defect type to calculate the degree of morphological feature matching. ； in: Indicates the first The first suspected defect area individual morphological features, This indicates the allowable buffer range when morphological features deviate from the reference range. This indicates the degree of matching between a single morphological feature and the corresponding defect type rule. The range of values ​​is to ; Indicates the first The first class defect corresponding to Lower limit of each morphological feature; Indicates the first The first class defect corresponding to The upper limit of the morphological features; the first Class defects, the first The reference interval for each morphological feature is represented as follows: ; Based on morphological evidence vectors Various characteristics and categories Calculate the morphological evidence matching score based on the degree of matching between the corresponding morphological rule intervals: ； in: Indicates the first Morphological evidence matching score of suspected defective areas, Indicates the number of morphological features, Indicates the first Weights of each morphological feature; The detection confidence level is dynamically verified based on the morphological evidence matching score and the prior process risk of the corresponding defect type.

8. The cable defect determination method according to claim 7, characterized in that, The dynamic verification of the detection confidence based on the morphological evidence matching score and the corresponding process risk prior of the defect type specifically includes the following steps: The knowledge rule support is calculated based on the morphological evidence matching score and the corresponding defect type's process risk prior, specifically expressed as follows: ； in: Representing knowledge rules for the first The level of support for each suspected defective area; This indicates the weight of prior process risks. ; Indicates the current process status relative to the category. Risk prior; Indicates the current suspected defect area and category The degree of matching of morphological rules; The detection confidence level is dynamically verified within a preset range using the knowledge rule support.

9. The cable defect determination method according to claim 8, characterized in that, The step of dynamically verifying the detection confidence level using knowledge rule support within a preset range specifically includes the following steps: The detection confidence score output by the target detection model According to the support of knowledge rules Detection confidence Perform dynamic verification: ； in: This indicates the initial confidence level after verification by the knowledge rule support. Indicates the strength of the knowledge rule verification; At the same time, knowledge rules provide positive support for visual detection results; At the same time, knowledge rules provide inverse inhibition on visual detection results; At the same time, the original confidence level of the visual detection model is not changed; The range of confidence level adjustments caused by knowledge rules is limited, and the final confidence level after verification is: ； in: Indicates the first i The final verification confidence level for each suspected defective area This represents the initial confidence level calculated based on the support of knowledge rules. This represents the raw confidence score output by the object detection model. and These represent the maximum allowed downward correction range and the maximum allowed upward correction range, respectively. ; The operation means limiting the initial confidence level to a given range.

10. A cable defect determination system, characterized in that, include: The acquisition module is used to acquire cable appearance images during the cable manufacturing process and extract suspected defect areas on the cable surface from the cable appearance images and their detection confidence levels. The prior calculation module is used to pre-construct a supervision knowledge rule network based on cable production process knowledge, supervision experience, and historical defect records. This network consists of process state nodes that characterize process parameters deviating from the normal range, risk event nodes that characterize the risks caused by abnormal process parameters, and defect type nodes that characterize the types of cable appearance defects. The activation degree of each process state node is calculated based on the process parameters in the current cable manufacturing process. The activation degree of each process state node is then forward-propagated according to the weight of the rule edge in the manufacturing knowledge rule network to calculate the activation degree of each risk event node. The process risk priors corresponding to each defect type node are calculated based on the activation degree of each risk event node. The determination module is used to extract morphological evidence of suspected defect areas and combine it with the process risk prior of the corresponding defect type to dynamically verify the detection confidence. Based on the final verification confidence level, suspected defect areas are classified and judged to generate cable defect judgment results.

Images