A method and system for detecting defects of a chain of a port rubber-tyred gantry crane trolley

Abstract

The application discloses a kind of port tyre gantry crane trolley chain defect detection method and system, it is related to port machinery equipment condition monitoring technical field, the process noise covariance matrix of kalman filter is dynamically self-adaptive adjusted by collecting vibration signal, fundamentally overcome the tracking loss problem under strong vibration working condition.It is also suppressed by the application of front-end adaptive polarized light filtering technology, the strong light interference of metal surface is improved from source image quality.Moreover, the application constructs the bidirectional feedback verification logic of detection and tracking, and combines the risk quantification evaluation model based on weibull distribution, realizes the business closed loop from bottom feature fine perception to upper active early warning, significantly improves the operation safety and automation operation efficiency of equipment.

Description

Technical Field

[0001] This invention belongs to the field of port machinery and equipment condition monitoring technology, and specifically relates to a method and system for detecting defects in the trolley chain of a port tire-mounted gantry crane. Background Technology

[0002] As the main loading and unloading equipment in modern container terminal yard operations, the trolley drive chain of a rubber-tired gantry crane is the core mechanical component for transmitting power, enduring huge alternating loads and high-frequency start-stop impacts over a long period. Due to the open-air environment of ports, the chain is inevitably affected by sea breeze erosion, salt spray corrosion, and dust adhesion, easily leading to changes in its physical properties under prolonged high-intensity operation. To ensure the safe operation of equipment and promote the unmanned construction of smart ports, relying on non-contact computer vision technology to monitor the chain surface condition online in real time, thereby obtaining the chain's running trajectory and surface feature information, and realizing long-term equipment health assessment and predictive maintenance, has become a key development direction for the status perception and operation and maintenance management of automated terminal equipment. However, existing chain visual inspection technologies face many bottlenecks in actual complex working conditions. On the one hand, the strong sunlight in open-air ports and the lubricating oil on the metal chain surface easily produce strong specular reflection, leading to frequent false detections and missed detections in pure visual inspection models. On the other hand, conventional detection algorithms often use horizontal bounding boxes for feature extraction, which introduces a lot of background noise when dealing with tilted or twisted chains, making accurate positioning impossible. Furthermore, the strong mechanical vibrations accompanying the movement of the trolley in a tire-mounted gantry crane can cause the motion prediction models of traditional target tracking algorithms to completely fail, resulting in frequent interruptions to the target trajectory and loss of identification. Moreover, traditional vision solutions often separate the detection and tracking processes, lacking a two-way feedback verification mechanism between detection and tracking, as well as a risk quantification assessment model based on time-series features, making it difficult to truly provide reliable intelligent operation and maintenance decision support on-site. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides a method and system for detecting defects in the trolley chain of a port tire-mounted gantry crane, thereby solving the aforementioned technical problems.

[0004] On the one hand, a method for detecting defects in the trolley chain of a port rubber-tired gantry crane is proposed, including the following steps:

[0005] S1: Real-time acquisition of images of the trolley chain in operation, and preprocessing of the images;

[0006] S2: Input the preprocessed running image into the pre-trained YOLOv11-OBB object detection model to obtain defect targets, wherein the defect targets include oriented bounding boxes and defect types;

[0007] S3: Based on the BoTSORT tracking algorithm, perform cross-frame association of defect targets and output the association results, and adjust the process noise covariance of the association results;

[0008] S4: Perform bidirectional feedback fusion of the defect target and the associated results, perform time-series filtering on the fused output, and obtain the filtered result as a stable defect target;

[0009] S5: Based on the characteristic parameters of the defect target, calculate the defect severity score according to the preset defect risk quantification assessment model, and output the risk level and operation and maintenance early warning information based on the defect severity score.

[0010] Preferably, when acquiring real-time images of the trolley chain and preprocessing the images in step S1, the specific steps include:

[0011] At least two different angles of operation are captured using a linear polarizing filter;

[0012] Calculate the index data of the running image, including the area ratio of reflective areas and the average brightness;

[0013] If the indicator data does not exceed the preset threshold, the acquisition angle corresponding to the running image with the highest indicator data is selected for subsequent running image acquisition.

[0014] Preferably, the specific process of adjusting the process noise covariance in step S3 includes: collecting the vibration signal of the large vehicle;

[0015] The spectral features of the vibration signal are extracted using FFT spectral analysis, and the spectral features include the dominant vibration frequency and amplitude characteristics;

[0016] The extracted spectral features are matched and identified with a preset vibration mode library to obtain the current vibration mode. The vibration mode library includes at least strong vibration modes and steady modes.

[0017] The process noise covariance matrix of the Kalman filter is dynamically adjusted based on the current vibration mode. The dynamic adjustment process is as follows: the process noise covariance matrix is ​​increased in the strong vibration mode and decreased in the steady mode.

[0018] Preferably, step S3, which involves cross-frame association of defect targets based on the BoTSORT tracking algorithm and defect features, specifically includes the following steps:

[0019] Obtain the set of predicted states of defect trajectories in the current frame from the historical frames, and the set of defect targets newly output by the YOLOv11-OBB model in the current frame;

[0020] The appearance feature vectors of the newly output defect targets are extracted respectively, and the motion features of the defect trajectory are extracted using the Kalman filter;

[0021] Calculate the total similarity between each defect trajectory and each newly output defect target;

[0022] Total similarity The specific calculation formula is as follows:

[0023] ,

[0024] in Cosine similarity of appearance features for The preset weight parameters, For IoU similarity of motion features, for The preset weight parameters;

[0025] Obtain the total similarity of all defective targets and construct a defective target similarity matrix;

[0026] Global optimization is performed on the target similarity matrix to find the optimal one-to-one matching relationship between the defect trajectory and the newly output detection target.

[0027] Preferably, step S4, the bidirectional feedback fusion, specifically includes the following steps:

[0028] Forward fusion: Using the defect trajectory of the previous frame, the target position of the current frame is predicted by the Kalman filter, the bounding box around the predicted position is expanded by a set multiple as the region of interest, and the weight of the defect target in the region of interest is increased.

[0029] Backward fusion: For new defect targets that do not match defect trajectories, the defect targets in the historical N frames are traced back to perform defect trajectory consistency verification. If a defect trajectory can be matched in the consecutive historical N frames and the total similarity is higher than the set value, the new target is confirmed to exist and initialized as a new defect trajectory.

[0030] Preferably, step S4 includes the following time-series filtering process:

[0031] Using a sliding window mechanism, the detection confidence of the same defective target over N consecutive frames is exponentially weighted and averaged to calculate the fusion confidence of the defective target. :

[0032] ,

[0033] Among them, weight Calculated based on the following formula: ,

[0034] in, For the first Frame detection confidence, The attenuation coefficient is used to retain defective targets with a fusion confidence level greater than a preset threshold.

[0035] Preferably, the defect risk quantification assessment model in step S5 is as follows:

[0036] ,

[0037] in, Score the severity of the defect; For the preset defect type weights, for Weighting coefficients; To normalize the score for defect size, for Weighting coefficients; The score is normalized to the duration of the defect. for The weighting coefficients.

[0038] Preferably, when outputting risk level and operation and maintenance early warning information, the specific steps include: scoring based on the severity of the defect. Output risk level;

[0039] Implement multi-level alarms or maintain continuous monitoring based on risk levels.

[0040] Preferably, step S5 further includes the following step:

[0041] The remaining service life of the chain is predicted based on the Weibull distribution model, and the specific formula is as follows:

[0042] ,

[0043] in, For reliability; Remaining lifespan; The scale parameter characterizing the characteristic lifetime is obtained by maximum likelihood estimation from historical maintenance data that records the historical running time of the chain; Shape parameters characterizing the failure distribution pattern;

[0044] If the reliability is lower than the set reliability threshold, the corresponding remaining lifetime is calculated.

[0045] On the other hand, a defect detection system for the trolley chain of a port tire-mounted gantry crane is proposed, including the following:

[0046] The image acquisition unit includes: an industrial camera for real-time acquisition of images of the trolley chain running, the industrial camera being equipped with an adjustable-angle linear polarization filter; and an accelerometer for real-time acquisition of trolley vibration signals.

[0047] Data processing unit: It has a pre-trained YOLOv11-OBB model and BoTSORT algorithm module embedded in it, which is used to perform defect detection analysis and risk quantification assessment on the received running images and vibration signals, and calculate the defect severity score.

[0048] Early warning and display unit: Communicatively connected to the data processing unit, used to display risk level and operation and maintenance early warning information based on the severity score of defects.

[0049] The beneficial effects of this invention are as follows: By acquiring vibration signals and dynamically and adaptively adjusting the process noise covariance matrix of the Kalman filter, the tracking loss problem under strong vibration conditions is fundamentally overcome. Furthermore, the application of front-end adaptive polarization filtering technology suppresses strong reflective interference from metal surfaces, improving image quality from the source. Moreover, this invention constructs a bidirectional feedback verification logic for detection and tracking, and combines it with a risk quantification assessment model based on Weibull distribution, realizing a closed-loop business process from fine-grained perception of underlying features to proactive early warning at the upper layer, significantly improving the operational safety and automated maintenance efficiency of the equipment. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the 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.

[0051] Figure 1 A flowchart illustrating the steps of a method for detecting defects in the trolley chain of a port tire-mounted gantry crane provided by the present invention;

[0052] Figure 2 This is a schematic diagram of a defect detection system for the trolley chain of a port tire-mounted gantry crane provided by the present invention. Detailed Implementation

[0053] The following disclosure provides many different embodiments or examples for implementing various embodiments of the invention. To simplify the disclosure, specific embodiments are described below. Of course, these are merely examples and are not intended to limit the scope of the invention.

[0054] The embodiments of the invention will now be described in detail with reference to the accompanying drawings.

[0055] like Figure 1 As shown, a method for detecting defects in the trolley chain of a port rubber-tired gantry crane includes the following steps:

[0056] S1: Real-time acquisition of images of the trolley chain in operation, and preprocessing of the images;

[0057] S2: Input the preprocessed running image into the pre-trained YOLOv11-OBB object detection model to obtain defect targets, wherein the defect targets include oriented bounding boxes and defect types;

[0058] S3: Based on the BoTSORT tracking algorithm, perform cross-frame association of defect targets and output the association results, and adjust the process noise covariance of the association results;

[0059] S4: Perform bidirectional feedback fusion of the defect target and the associated results, perform time-series filtering on the fused output, and obtain the filtered result as a stable defect target;

[0060] S5: Based on the characteristic parameters of the defect target, calculate the defect severity score according to the preset defect risk quantification assessment model, and output the risk level and operation and maintenance early warning information based on the defect severity score.

[0061] In its specific implementation, this invention first acquires real-time images of the trolley chain in operation and performs image preprocessing. These images are then input into a pre-trained YOLOv11-OBB model to extract defect targets. In the anchor frame generation stage of this model, a novel prior constraint on link spacing is introduced. Specifically, the mechanical design dimensions of the trolley chain are obtained and mapped to a baseline aspect ratio and dimension baseline value at the image pixel scale. This serves as the center to limit the aspect ratio search range and rotation angle generation range of the directional anchor frames, thereby eliminating invalid anchor frames that deviate from the physical dimensions of the chain and accurately outputting targets containing directional bounding boxes and defect types. After acquiring the target, the system uses the BoTSORT algorithm to perform cross-frame data association and simultaneously collects equipment hardware vibration signals to dynamically adjust the process noise covariance matrix of the Kalman filter to adapt to mechanical jitter. Then, it performs bidirectional feedback fusion, uses forward trajectory prediction to expand the region of interest to recall ambiguous targets, and combines backward tracing of historical frames to verify newly detected targets to remove transient environmental noise. The fused output is then subjected to time-series filtering through a sliding window to extract stable defect targets. Finally, the extracted feature parameters are substituted into the quantitative evaluation model to calculate the severity score and output the risk level and warning information. The underlying principle of this system design lies in coupling a purely data-driven computer vision model with the inherent physical boundary constraints and real-time stress state of port machinery. It utilizes cross-modal sensing feedback to overcome the limitations of single visual features. Compared to conventional algorithms that rely on horizontal bounding boxes and are susceptible to interference from lighting and vibration, this implementation not only effectively eliminates background feature redundancy under chain tilting and torsional postures, but also ensures the continuity of target identity and the rigor of defect judgment under strong vibration conditions. The overall invention fundamentally reduces the high-frequency false alarm and missed alarm rate in complex environments, and significantly improves the accuracy of predictive maintenance and the level of unmanned operation and maintenance of core port machinery.

[0062] More specifically, the real-time acquisition of the running image of the trolley chain in step S1, and the preprocessing of the running image, specifically includes the following steps:

[0063] At least two different angles of operation are captured using a linear polarizing filter;

[0064] Calculate the index data of the running image, including the area ratio of reflective areas and the average brightness;

[0065] If the indicator data does not exceed the preset threshold, the acquisition angle corresponding to the running image with the highest indicator data is selected for subsequent running image acquisition.

[0066] In the image acquisition stage, an industrial camera with a front-mounted linear polarizing filter is used to visually perceive the chain area. Specifically, an electrically rotating polarizing filter mechanism is deployed in front of the camera lens to sequentially acquire images of the trolley chain at three polarization angles of 0°, 45°, and 90° according to a preset step size. The pixel ratio of the reflective area and the average brightness value of each image are calculated as image quality evaluation indicators. When the calculation results do not exceed the system's preset reflective interference threshold, the polarization angle with the best indicators is selected as the fixed acquisition position, that is, the angle with the smallest reflective area ratio and the most suitable average brightness. If a certain angle index exceeds a threshold, an angle fine-tuning mechanism is triggered. This design principle is based on the highly polarized characteristics of specular reflection light on metal surfaces. By rotating the polarizer, reflected light waves in specific vibration directions can be selectively filtered, while retaining the true texture information in diffuse reflection light. Compared with the traditional fixed light source and single polarization angle acquisition method, this implementation method can significantly reduce the specular highlight interference caused by oil film and water stains on the metal surface of the chain in the strong sunlight environment of the port, improve the signal-to-noise ratio of the defect area, and provide high-quality input images for subsequent directional detection, thereby reducing the false detection rate and improving the detection rate of micro-cracks at the source.

[0067] More specifically, the specific process of adjusting the process noise covariance in step S3 includes: collecting the vibration signal of the large vehicle;

[0068] The spectral features of the vibration signal are extracted using FFT spectral analysis, and the spectral features include the dominant vibration frequency and amplitude characteristics;

[0069] The extracted spectral features are matched and identified with a preset vibration mode library to obtain the current vibration mode. The vibration mode library includes at least strong vibration modes and steady modes.

[0070] The process noise covariance matrix of the Kalman filter is dynamically adjusted based on the current vibration mode. The dynamic adjustment process is as follows: the process noise covariance matrix is ​​increased in the strong vibration mode and decreased in the steady mode.

[0071] In specific implementation, in the trolley chain defect detection system, to address the problem of target tracking failure caused by mechanical vibration during operation, the method involves deploying triaxial accelerometers at key nodes of the trolley's traveling mechanism to collect vibration signals in real time. The FFT spectrum analysis module extracts the dominant frequency components and amplitude envelope features from the vibration signals. These time-frequency features are then matched against a pre-set vibration mode library. This vibration mode library is constructed based on historical operating data and includes at least strong vibration modes during the trolley's start / brake phase and stable modes during uniform speed operation. After identifying the current vibration mode, the system dynamically adjusts the core parameter of the Kalman filter in the BoTSORT tracking algorithm—the process noise covariance matrix Q. In strong vibration modes, the Q value is increased to increase the uncertainty weight of the prediction model, making the filter more inclined to accept new observations from the detection module and avoiding trajectory breakage caused by mechanical vibration. In stable modes, the Q value is decreased to improve the reliability of motion prediction and enhance trajectory continuity. In a specific embodiment, when the trolley starts accelerating on the yard track, the accelerometer detects that the dominant vibration frequency is concentrated in the 15-20Hz range and the amplitude exceeds the threshold. The system determines this as a strong vibration mode and automatically multiplies the diagonal elements of the Q matrix by an amplification factor of 3.0, significantly increasing the weight of the detection results in data association. After the equipment enters the uniform speed operation phase, the vibration characteristics drop below 5Hz, and the system switches to a stable mode, restoring the Q value to the baseline level, allowing Kalman prediction to dominate. This design principle is based on the trade-off mechanism between process noise and observation noise in Kalman filtering theory. By introducing external physical vibration sensing, it achieves closed-loop adaptive adjustment of algorithm parameters, overcoming the limitation of traditional tracking algorithms that cannot adapt to changes in operating conditions under fixed noise parameters. This significantly enhances the robustness and continuous monitoring capability of the system in complex vibration environments.

[0072] More specifically, in step S3, when performing cross-frame association of defect targets based on the BoTSORT tracking algorithm and defect features, the following steps are included:

[0073] Obtain the set of predicted states of defect trajectories in the current frame from the historical frames, and the set of defect targets newly output by the YOLOv11-OBB model in the current frame;

[0074] The appearance feature vectors of the newly output defect targets are extracted respectively, and the motion features of the defect trajectory are extracted using the Kalman filter;

[0075] Calculate the total similarity between each defect trajectory and each newly output defect target;

[0076] Total similarity The specific calculation formula is as follows:

[0077] ,

[0078] in Cosine similarity of appearance features for The preset weight parameters, For IoU similarity of motion features, for The preset weight parameters;

[0079] Obtain the total similarity of all defective targets and construct a defective target similarity matrix;

[0080] Global optimization is performed on the target similarity matrix to find the optimal one-to-one matching relationship between the defect trajectory and the newly output detection target.

[0081] In practical implementation, the depth and appearance feature vectors of all detected targets in the current frame are first extracted. Then, the position and velocity state of historical trajectories in the current frame are predicted using a Kalman filter. Finally, the total similarity of each trajectory-detection combination is calculated. For example, when the system simultaneously detects 3 historical trajectories and 5 defect targets in the current frame, a 5×3 similarity matrix is ​​constructed. The Hungarian algorithm is used to globally optimize this matrix, finding a one-to-one matching scheme that maximizes the total matching score, effectively solving the problem of trajectory ID switching and loss in multi-target scenarios. This design principle is based on multimodal feature fusion and combination optimization theory, using appearance and movement... The weighted fusion of dynamic features overcomes the instability of single features under strong vibration and lighting changes, while the Hungarian algorithm ensures the allocation of trajectory IDs in a globally optimal sense, avoiding suboptimal matching caused by greedy algorithms. Compared with existing technologies that use simple distance thresholds or rely solely on single features for association, this approach maintains the consistency of target identity, especially in complex scenarios with drastic changes in reflection on the chain surface or temporary occlusion. The overall solution ensures the robustness of visual tracking in harsh industrial environments through mathematical optimization, providing a reliable data foundation for the long-term evolution analysis of defects and fundamentally improving the accuracy of predictive maintenance decisions.

[0082] More specifically, in step S4, the bidirectional feedback fusion specifically includes the following steps:

[0083] Forward fusion: Using the defect trajectory of the previous frame, the target position of the current frame is predicted by the Kalman filter, the bounding box around the predicted position is expanded by a set multiple as the region of interest, and the weight of the defect target in the region of interest is increased.

[0084] Backward fusion: For new defect targets that do not match defect trajectories, the defect targets in the historical N frames are traced back to perform defect trajectory consistency verification. If a defect trajectory can be matched in the consecutive historical N frames and the total similarity is higher than the set value, the new target is confirmed to exist and initialized as a new defect trajectory.

[0085] In a specific implementation, when a suspected crack is detected on the chain surface, the forward fusion module first calls the Kalman filter to predict the precise position of the target in the current frame based on historical trajectories. It then expands the bounding box by 1.5 times the original size around the predicted coordinates to form a dynamic region of interest. All detection responses within this region are weighted by 1.8 times to ensure that minor cracks blurred by vibration are not missed. Simultaneously, the backward fusion module initiates reverse verification for newly detected targets not associated with historical trajectories. It traces detection records from four consecutive historical images and calculates the temporal consistency between appearance features and motion trajectories. If a target with a similar shape and motion trend is detected in at least three frames and the cumulative similarity exceeds a threshold of 0.75, it is determined to be a genuine defect and a new trajectory is initialized; otherwise, it is considered transient interference and filtered out. This design principle is based on the spatiotemporal consistency constraint theory. The forward channel utilizes the continuity of physical motion to enhance the credibility of weak-response targets, while the backward channel verifies the authenticity of new targets through historical evidence chains.

[0086] More specifically, in step S4, the time-series filtering process includes the following:

[0087] Using a sliding window mechanism, the detection confidence of the same defective target over N consecutive frames is exponentially weighted and averaged to calculate the fusion confidence of the defective target. :

[0088] ,

[0089] Among them, weight Calculated based on the following formula: ,

[0090] in, For the first Frame detection confidence, The attenuation coefficient is used to retain defective targets with a fusion confidence level greater than a preset threshold.

[0091] In practical implementation, a sliding time window is maintained to store the defect feature parameter sequence of the most recent N frames. Dynamic weights are assigned to each frame within the window, with the current frame having the highest weight and the weights of historical frames decreasing exponentially with time difference. For example, when the system detects a suspected crack on the surface of a link in the chain, the fusion window includes the crack length measurements of the current frame and the previous four frames. Due to the instantaneous vibration of the trolley at the track joint, the measurement value of the current frame is abnormally large. However, through the exponential weighted fusion mechanism, the stable measurement values ​​of historical frames still have a significant impact on the final result, effectively suppressing the interference of single-frame outliers. This design principle is based on the exponential smoothing theory of time series. By assigning a higher weight to the current frame, the system ensures a rapid response to changes in actual defects. At the same time, the progressively decreasing weights of historical frames maintain measurement stability, achieving a balance between dynamic response and noise suppression. Compared with traditional mean filtering or fixed weight fusion methods, this implementation can effectively reduce the standard deviation of defect size measurement in the port measurement environment and maintain the continuity and consistency of defect features when the trolley passes through uneven track areas, avoiding false alarms caused by instantaneous vibration.

[0092] More specifically, the defect risk quantification assessment model in step S5 is shown in the following formula:

[0093] ,

[0094] in, Score the severity of the defect; For the preset defect type weights, for Weighting coefficients; To normalize the score for defect size, for Weighting coefficients; The score is normalized to the duration of the defect. for The weighting coefficients.

[0095] In the port RTG trolley chain defect risk assessment system, the specific implementation of this invention achieves accurate assessment of defect severity through a multi-dimensional quantitative model, with weighting coefficients... , , All parameters are determined through machine learning optimization based on historical fault data. In a specific embodiment, when the system detects a surface crack in a link of the chain, it first calculates the pixel area of ​​the crack region by high-precision OBB bounding box segmentation, divides it by the standard area of ​​a single link of that chain model to obtain a size normalization score, and combines the crack morphology characteristics to classify it as a deep crack type to obtain a type score. The system then confirms that the crack is stably present in 7 consecutive frames of images through a two-way feedback fusion mechanism, and finally merges the three-dimensional scores to output the risk level. This design principle is based on the deep fusion of mechanical failure physical model and data-driven scoring, breaking through the limitations of traditional manual inspection relying on experience judgment or single size threshold alarms, and upgrading defect assessment from qualitative description to quantifiable engineering parameters. Compared with the prior art, this implementation can identify high-risk defects several days in advance in the actual port operation environment, especially significantly improving the accuracy of quantitative assessment of early microcracks.

[0096] More specifically, when outputting risk levels and operational warning information, the following steps are included: scoring based on the severity of the defect. Output risk level;

[0097] Implement multi-level alarms or maintain continuous monitoring based on risk levels.

[0098] In the risk decision-making phase of the trolley chain defect detection system, the specific implementation of this invention is based on a defect severity score. Implement a tiered early warning mechanism: The system matches the real-time calculated defect score with a preset threshold range, when... A value ≥0.8 is considered a Level 1 risk, triggering an emergency stop warning and initiating safety braking in conjunction with the trolley control system; when 0.5 ≤ When the value is less than 0.8, it is classified as a Level 2 risk, and an important maintenance warning work order is generated and pushed to the mobile terminal of the maintenance personnel; when the value is less than or equal to 0.3... When the value is less than 0.5, it is classified as a level three risk, a general maintenance reminder is issued, and it is included in the weekly maintenance plan; when... A value less than 0.3 indicates a level 4 risk, and the system continuously monitors and records the defect evolution trend. In a specific embodiment, when a deep crack is detected on a link of the chain and the surface wear area expands, the system calculates a score based on the crack length, depth, and stress concentration factor. =0.75, triggering a level 2 risk warning, automatically generating an electronic work order containing the defect location coordinates, evolution trend chart, and recommended maintenance procedures, and simultaneously pushing the warning information to the port equipment management platform via the OPC UA interface; this design principle is based on the risk quantification and classification management theory, mapping defect characteristic parameters to discrete risk levels with clear operation and maintenance actions, realizing closed-loop control from perception to decision-making; compared with traditional fixed threshold alarms or manual experience judgment methods, the overall solution significantly improves equipment safety and operation and maintenance economy through precise quantitative threshold division and automated decision execution.

[0099] More specifically, step S5 also includes the following steps:

[0100] The remaining service life of the chain is predicted based on the Weibull distribution model, and the specific formula is as follows:

[0101] ,

[0102] in, For reliability; Remaining lifespan; The scale parameter characterizing the characteristic lifetime is obtained by maximum likelihood estimation from historical maintenance data that records the historical running time of the chain; Shape parameters characterizing the failure distribution pattern;

[0103] If the reliability is lower than the set reliability threshold, the corresponding remaining lifetime is calculated.

[0104] In a specific embodiment, when the system continuously detects deep wear defects in a certain RTG trolley chain with a severity score of 0.78, it calculates the remaining safe operating time to be 216 hours based on the chain material characteristics and port operation intensity parameters. The system automatically generates a level-three warning and pushes it to the equipment management platform, recommending replacement before the next planned shutdown window. This design principle is based on the combination of mechanical fatigue cumulative damage theory and statistical life distribution model, transforming the instantaneous defect characteristics of visual inspection into time-varying reliability indicators. This breaks through the limitations of traditional periodic maintenance or simple threshold alarms, achieving a leap from state perception to life prediction. Compared with existing technologies that rely on manual experience to judge replacement cycles or fixed time interval maintenance, this implementation provides a data-driven decision-making basis for the operation and maintenance of large port equipment by quantifying the mapping relationship between chain health status and remaining life.

[0105] like Figure 2 As shown, a defect detection system for the trolley chain of a port tire-mounted gantry crane includes the following:

[0106] The image acquisition unit includes: an industrial camera for real-time acquisition of images of the trolley chain running, the industrial camera being equipped with an adjustable-angle linear polarization filter; and an accelerometer for real-time acquisition of trolley vibration signals.

[0107] Data processing unit: It has a pre-trained YOLOv11-OBB model and BoTSORT algorithm module embedded in it, which is used to perform defect detection analysis and risk quantification assessment on the received running images and vibration signals, and calculate the defect severity score.

[0108] Early warning and display unit: Communicatively connected to the data processing unit, used to display risk level and operation and maintenance early warning information based on the severity score of defects.

[0109] The port tire-mounted gantry crane trolley chain defect detection system of the present invention has the same operation and effect as the above-mentioned port tire-mounted gantry crane trolley chain defect detection method, and will not be described again here.

[0110] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented using necessary general-purpose hardware platforms, or it can be implemented through a combination of hardware and software. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a computer product. The present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0111] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.

Claims

1. A method for detecting defects in the trolley chain of a port tire-mounted gantry crane, characterized in that, Includes the following steps: S1: Real-time acquisition of images of the trolley chain in operation, and preprocessing of the images; S2: Input the preprocessed running image into the pre-trained YOLOv11-OBB object detection model to obtain defect targets, wherein the defect targets include oriented bounding boxes and defect types; S3: Based on the BoTSORT tracking algorithm, perform cross-frame association of defect targets and output the association results, and adjust the process noise covariance of the association results; S4: Perform bidirectional feedback fusion of the defect target and the associated results, perform time-series filtering on the fused output, and obtain the filtered result as a stable defect target; S5: Based on the characteristic parameters of the defect target, calculate the defect severity score according to the preset defect risk quantification assessment model, and output the risk level and operation and maintenance early warning information based on the defect severity score.

2. The method for detecting defects in the trolley chain of a port tire-mounted gantry crane according to claim 1, characterized in that, When acquiring real-time images of the trolley chain in step S1 and preprocessing the images, the specific steps include: acquiring at least two images from different angles using a linear polarizing filter; calculating the index data of the images, including the area ratio of reflective regions and average brightness; and, provided that the index data does not exceed a preset threshold, selecting the acquisition angle corresponding to the image with the highest index data for subsequent image acquisition.

3. The method for detecting defects in the trolley chain of a port tire-mounted gantry crane according to claim 1, characterized in that, The specific process of adjusting the process noise covariance in step S3 includes: acquiring the vibration signal of the trolley; extracting the spectral features of the vibration signal using FFT spectral analysis, the spectral features including the dominant vibration frequency and amplitude features; matching and identifying the extracted spectral features with a preset vibration mode library to obtain the current vibration mode, the vibration mode library including at least a strong vibration mode and a stable mode; and dynamically adjusting the process noise covariance matrix of the Kalman filter based on the current vibration mode, the dynamic adjustment process being as follows: increasing the process noise covariance matrix in the strong vibration mode and decreasing the process noise covariance matrix in the stable mode.

4. The method for detecting defects in the trolley chain of a port tire-mounted gantry crane according to claim 1, characterized in that, Step S3, based on the BoTSORT tracking algorithm and defect features, involves cross-frame association of defect targets, specifically including the following steps: obtaining the predicted state set of defect trajectories in historical frames for the current frame, and the set of newly output defect targets from the YOLOv11-OBB model in the current frame; extracting the appearance feature vectors of the newly output defect targets, and using the Kalman filter to extract the motion features of the defect trajectories; calculating the total similarity between each defect trajectory and each newly output defect target; and calculating the total similarity. The specific calculation formula is as follows: ,in Cosine similarity of appearance features for The preset weight parameters, For IoU similarity of motion features, for The preset weight parameters are used; the total similarity of all defective targets is obtained, and a defective target similarity matrix is ​​constructed; the target similarity matrix is ​​globally optimized to solve for the optimal one-to-one matching relationship between the defect trajectory and the newly output detection target.

5. The method for detecting defects in the trolley chain of a port tire-mounted gantry crane according to claim 4, characterized in that, In step S4, the bidirectional feedback fusion specifically includes the following steps: Forward fusion: using the defect trajectory of the previous frame, the target position of the current frame is predicted by a Kalman filter, and the bounding box around the predicted position is expanded by a set multiple as the region of interest, and the weight of the defect target in the region of interest is increased; Backward fusion: for new defect targets that have not been matched with defect trajectories, the defect targets in the historical N frames are traced back to perform defect trajectory consistency verification. If a defect trajectory can be matched in the consecutive historical N frames and the total similarity is higher than the set value, then the new target is confirmed to exist and initialized as a new defect trajectory.

6. The method for detecting defects in the trolley chain of a port tire-mounted gantry crane according to claim 5, characterized in that, In step S4, the time-series filtering process specifically includes the following: Using a sliding window mechanism, the detection confidence of the same defective target over N consecutive frames is exponentially weighted to calculate the fusion confidence of the defective target. : Among them, weight Calculated based on the following formula: ,in, For the first Frame detection confidence, The attenuation coefficient is used to retain defective targets with a fusion confidence level greater than a preset threshold.

7. The method for detecting defects in the trolley chain of a port tire-mounted gantry crane according to claim 1, characterized in that, The defect risk quantification assessment model in step S5 is shown in the following formula: ,in, Score the severity of the defect; For the preset defect type weight, for Weighting coefficients; To normalize the score for defect size, for Weighting coefficients; The score is normalized to the duration of the defect. for The weighting coefficients.

8. The method for detecting defects in the trolley chain of a port tire-mounted gantry crane according to claim 7, characterized in that, When outputting risk levels and operational warning information, the specific steps include: scoring based on the severity of the defect. Output risk level; execute multi-level alarms or maintain continuous monitoring based on risk level.

9. The method for detecting defects in the trolley chain of a port tire-mounted gantry crane according to claim 8, characterized in that, Step S5 further includes the following step: predicting the remaining service life of the chain based on the Weibull distribution model, as shown in the following formula: ,in, For reliability; Remaining lifespan; The scale parameter characterizing the characteristic lifetime is obtained by maximum likelihood estimation from historical maintenance data that records the historical running time of the chain; The shape parameter characterizes the failure distribution pattern; if the reliability is lower than the set reliability threshold, the corresponding remaining lifetime is calculated.

10. A defect detection system for the trolley chain of a port tire-mounted gantry crane, characterized in that, Includes the following: The image acquisition unit includes: an industrial camera for real-time acquisition of images of the trolley chain operation, the industrial camera being equipped with an adjustable-angle linear polarization filter; an accelerometer for real-time acquisition of trolley vibration signals; a data processing unit: embedding a pre-trained YOLOv11-OBB model and BoTSORT algorithm module, used to perform defect detection analysis and risk quantification assessment on the received operation images and vibration signals, and calculate the defect severity score; and an early warning and display unit: communicatively connected to the data processing unit, used to display the risk level and maintenance early warning information based on the defect severity score.

Images