Coal mine conveying belt tearing defect identification method based on machine vision

Abstract

The invention discloses a coal mine conveying belt tearing defect identification method based on machine vision, which realizes high-definition image acquisition in a complex environment through a linear array camera and multispectral light supplement, suppresses dust noise by adopting a dark channel prior algorithm, and realizes defect high-precision detection in combination with an improved YOLOv6 model (including a GhostNetv3 trunk and edge feature enhancement module). By means of an innovative morphological filtering and time sequence verification mechanism, the omission ratio is reduced to 1.2%, the false alarm rate is reduced to 1.2%, the detection speed is larger than or equal to 30 FPS, the method is suitable for the coal mine underground environment with low illumination and high dust, and reliable guarantee is provided for safe operation of a belt.

Description

Technical Field

[0001] The present invention relates to the fields of coal mine safety monitoring and machine vision technology, and specifically to a real-time detection method for conveyor belt tear defects suitable for the complex environment of underground coal mines. The method can achieve high-precision identification (detection accuracy ≥ 98%) of defects such as longitudinal tearing, edge wear, and holes in the belt, and is suitable for conveyor systems with belt speeds of 0.5-6 m / s. Background Art

[0002] Current status of conveyor belt tearing defects in complex environments of coal mines in the existing technology:

[0003] Disadvantages of manual inspection: Relying on visual inspection by patrol inspectors, the inspection of a single kilometer of belt takes ≥ 1 hour, and the missed detection rate is as high as 15% (GB / T34028-2017 standard requires a missed detection rate of ≤ 5%); it is unable to detect defects on the inside of the belt, and there are safety risks such as people falling and dust inhalation.

[0004] Limitations of traditional detection technology: Contact detection (such as pressure sensors): can only detect surface defects and create additional resistance to belt operation (increasing energy consumption by 2%-5%);

[0005] Ordinary machine vision: using area array camera (resolution ≤ 1920 × 1080), in low light (<10 lux), high dust (concentration > 1000 mg / m 3 ) environment, the defect recognition accuracy is less than 85%, which cannot meet the complex working conditions underground.

[0006] Therefore, in order to address the problems of "low detection efficiency, poor environmental adaptability, and fuzzy defect classification" in coal mine conveyor belts, it is urgent to develop an intelligent detection method based on machine vision to achieve real-time defect detection (response time < 200ms) and accurate classification in complex environments, and reduce downtime accidents caused by belt tearing (reducing downtime by more than 40 hours per year on average). Summary of the Invention

[0007] The purpose of the present invention is to solve the technical problems existing in the prior art and provide a method for identifying tear defects in coal mine conveyor belts based on machine vision, comprising the following steps:

[0008] a. Use a linear array camera to scan the belt at a frequency of ≥5000 Hz, obtain line images covering the full width of the belt, and stitch them into a complete surface image;

[0009] b. Adaptively pre-process the captured images, including dust and haze correction, low-light enhancement, and Gaussian noise filtering;

[0010] c. Detect defects using a modified YOLOv6 model that includes a GhostNetv3 backbone network and an edge feature enhancement module;

[0011] d. Perform morphological filtering and time series verification on the detection results, and output the final defect type and location.

[0012] Furthermore, the linear array camera has a spectral response range of 400-1000 nm, an installation height of 0.8-1.2 m, and an adaptable belt speed of 0.5-6 m / s.

[0013] Furthermore, the dust haze correction adopts a dark channel priori algorithm, and the specific formula is:

[0014]

[0015] Where A is the atmospheric light value, t(x) is the transmittance, and the transmittance estimation is optimized by guided filtering.

[0016] Furthermore, the edge feature enhancement module extracts the belt edge contour through the Sobel operator and performs channel-level fusion with the feature map output by the backbone network.

[0017] Furthermore, the loss function of the improved YOLOv6 model includes classification loss, bounding box regression loss and defect morphology loss, and the expression is: L = L cls +1.5L bor +2.0L shape

[0018] Among them L shape The defect contour similarity is calculated based on Hu moment features.

[0019] Furthermore, the morphological filtering rules include: longitudinal tear: length / width>5 and edge curvature change>3 places / mm; edge wear: edge contour offset>3mm (relative to the standard edge model).

[0020] Furthermore, the time series verification triggers an alarm when the same type of defects are detected in three consecutive frames, thereby avoiding accidental noise interference.

[0021] Furthermore, it includes: an image acquisition unit: including a linear array camera, a multi-spectral fill light and an air curtain dust protection device; a data processing unit: integrating an adaptive preprocessing module, an improved YOLOv6 detection model and a defect verification module; an alarm unit: supporting sound and light alarms and remote signal transmission (transmission delay <100ms).

[0022] Furthermore, the multi-spectral fill light includes an 850nm infrared lamp and a 450nm blue light, and automatically switches the lighting mode via a photosensor to ensure that the illumination in the detection area is ≥200lux. Furthermore, when executed by a processor, the program implements the belt defect identification method described in any one of claims 1-7, including image preprocessing, improved YOLOv6 detection, and defect verification steps.

[0023] Compared with the existing technology, the present invention integrates multi-spectral supplementary light and dust correction technology, and can 3 In dusty environments, image quality is improved by more than 50%, and detection accuracy is increased by 15% compared to traditional solutions, meeting the environmental adaptability requirements for belt detection in Article 374 of the Coal Mine Safety Regulations.

[0024] The defect detection accuracy of the present invention is greatly improved, and the minimum identification defect size is: longitudinal tear width 2mm, hole diameter 5mm, edge wear offset 1mm, meeting the "first-level detection accuracy" requirements of MT / T907-2023 "Technical Conditions for Coal Mine Belt Conveyor Protection Devices";

[0025] In addition, the classification accuracy of the present invention is ≥98%, which enables accurate determination of defect types (traditional solutions can only detect the presence or absence of defects but cannot classify them); real-time performance and reliability are improved, with a detection speed of 32FPS (detection interval of 31ms at a belt speed of 6m / s), ensuring that each meter of belt is inspected ≥32 times and no dynamic defects are missed; time series verification and morphological filtering reduce the false alarm rate from 8% to 1.2%, reducing invalid alarms by more than 70%, and reducing the workload of operation and maintenance personnel. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] The present invention will be further described below with reference to the accompanying drawings and embodiments;

[0027] Figure 1 : Flowchart of the processing algorithm of the present invention;

[0028] Figure 2 : Schematic diagram of hardware deployment of the detection system of the present invention. DETAILED DESCRIPTION

[0029] This section will describe in detail the specific embodiments of the present invention. The preferred embodiments of the present invention are shown in the accompanying drawings. The purpose of the accompanying drawings is to supplement the description of the text part of the specification with graphics, so that people can intuitively and vividly understand each technical feature and the overall technical solution of the present invention, but it should not be understood as a limitation on the scope of protection of the present invention.

[0030] Reference Figure 1-2 The method for identifying tear defects in coal mine conveyor belts based on machine vision according to an embodiment of the present invention includes:

[0031] 1. Hardware system architecture:

[0032] The image acquisition module includes a linear array camera: a TCD1304DG linear array CCD camera (resolution 5340×1, spectral response 400-1000nm), installed 0.8-1.2m above the belt, with a scanning frequency of 5000Hz, adapting to belt speeds ≤6m / s; multi-spectral fill light: equipped with 850nm infrared LED lights (illuminance ≥200lux) and 450nm blue light LED lights (to suppress dust reflections), automatically switching lighting modes through photosensors; dust-proof device: the camera lens is integrated with an air curtain purge system (air pressure 0.3MPa), which automatically cleans the lens every 5 minutes to ensure that the lens transmittance is ≥95%.

[0033] The data processing platform includes: industrial computer: configured with NVIDIA Jetson AGX Orin (computing power 200TOPS), running Ubuntu 20.04 system; storage module: using SSD hard disk (capacity 2TB), real-time storage of detection images with timestamps (storage period 30 days).

[0034] 2. Image Processing Algorithms

[0035] Adaptive image preprocessing

[0036] Dust noise suppression:

[0037] Gaussian filtering (kernel size 5×5, σ=1.5) was used to remove random noise;

[0038] Dust haze correction based on dark channel prior restores image contrast (contrast improvement of 30%-50%):

[0039] Where: I(x) is the input image, A is the atmospheric light value, t(x) is the transmittance, and t0=0.1 is the minimum transmittance threshold.

[0040] Low-light enhancement: adaptive histogram equalization (CLAHE) expands the image grayscale range to [20,230] and improves the contrast between defects and background.

[0041] 2. Improved YOLOv6 detection model

[0042] Backbone network: GhostNetv3 is used to replace CSPDarknet, which reduces the number of parameters by 40% and increases the detection speed by 25% (FPS ≥ 30);

[0043] Multi-scale feature fusion:

[0044] Added an edge feature enhancement module (EFM) to extract the belt edge contour using the Sobel operator and fuse it with deep semantic features to improve the accuracy of edge wear detection;

[0045] The feature pyramid (FPN) outputs three scales (80×80, 40×40, and 20×20), corresponding to small target (hole), medium target (edge ​​wear), and large target (longitudinal tear) detection respectively;

[0046] Loss function optimization: L = L cls +λ1L box +λ2L shape

[0047] Among them: Lcls is the classification loss (FocalLoss), Lbox is the DIoU loss, L_shape is the contour similarity loss of longitudinal tearing (based on Hu moment features), λ1=1.5,λ2=2.0.

[0048] 3. Defect classification and false detection elimination

[0049] Morphological filtering: longitudinal tear: length / width>5, and the edge is irregularly serrated (curvature change>3 places / mm);

[0050] Edge wear: edge profile deviation > 3mm (relative to the standard belt edge model);

[0051] Time series verification: An alarm is triggered when the same defect is detected in three consecutive frames, avoiding accidental noise interference (the false detection rate is reduced from 8% to 1.2%).

[0052] 3. System Workflow

[0053] 1. Image acquisition: A linear array camera scans the belt at 5000 Hz, generating line images with a width of 5340 pixels and a height of 1 pixel, which are then stitched together to form a complete belt surface image (resolution 5340 × 2000, corresponding to a belt width of 1.2 m).

[0054] 2. Preprocessing: Dust correction, illumination enhancement, and noise filtering are performed in sequence to output a clear belt surface image;

[0055] 3. Defect detection: Improve the real-time inference of the YOLOv6 model to output defect type, location (pixel-level accuracy) and confidence level (threshold ≥ 0.9);

[0056] Result processing: Defects that meet the morphological characteristics trigger audible and visual alarms (response time < 200ms) and are uploaded to the ground monitoring system via industrial Ethernet (transmission delay < 100ms). Specific embodiment:

[0058] The main transport belt of a coal mine has a belt width of 1.2m, a belt speed of 4m / s, an ambient light of 5-15lux, and a dust concentration of 800-1500mg / m 3 , deploy 2 sets of detection devices (1 set each at the nose and tail).

[0059] Hardware parameters (as shown in Table 1)

[0060]

[0061] Table 1

[0062] The specific implementation steps are as follows:

[0063] Dataset construction: 5,000 belt images were collected (including 3,000 defect images: 1,200 longitudinal tears, 1,000 edge wear images, and 800 holes). The defect outlines were marked using the LabelMe annotation tool (accuracy ±1 pixel). Data augmentation: The data was expanded to 20,000 images using rotation (±10°), scaling (0.8-1.2 times), and Gaussian noise (σ=0.05).

[0064] Model training: Training parameters: batch size 32, learning rate 1e-4, training period 50 epochs; performance indicators: mAP@0.5 reached 98.2%, longitudinal tear AP = 97.8%, edge wear AP = 98.5%, hole AP = 98.9%.

[0065] Field tests: Low-light scenarios: With infrared fill light enabled, image contrast increased from 15% to 45%, and defect edge clarity improved by 60%; high-dust scenarios: The air curtain purge system reduced lens contamination from 30% to 5%, and increased detection accuracy from 82% to 98.3%; dynamic detection: At a belt speed of 4m / s, the detection delay was 180ms, meeting real-time requirements (industry standard ≤200ms).

[0066] Comparison of experimental data

[0067] index Solution of the present invention Traditional solution Improvement Detection accuracy 98.3％ 85.2％ 15.4％ Missed detection rate 1.2％ 12.5％ 90.4％ Average detection speed 32FPS 18FPS 77.8％ Adaptability to dusty environments 98.3％ 82.0％ 20.0％

[0068] Table 2

[0069] The embodiments of the present invention are described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Various changes can be made within the scope of knowledge possessed by ordinary technicians in the technical field without departing from the scope of the present invention.

Claims

1. A method for identifying tear defects in coal mine conveyor belts based on machine vision, characterized in that: The following steps are involved: a. Use a linear array camera to scan the belt at a frequency of ≥5000 Hz, obtain line images covering the full width of the belt, and stitch them into a complete surface image; b. Adaptively pre-process the captured images, including dust and haze correction, low-light enhancement, and Gaussian noise filtering; c. Detect defects using a modified YOLOv6 model that includes a GhostNetv3 backbone network and an edge feature enhancement module; d. Perform morphological filtering and time series verification on the detection results, and output the final defect type and location.

2. The method according to claim 1, characterized in that The linear array camera has a spectral response range of 400-1000 nm, an installation height of 0.8-1.2 m, and an adaptable belt speed of 0.5-6 m / s.

3. The method according to claim 1, characterized in that The dust haze correction adopts the dark channel prior algorithm, and the specific formula is: Where A is the atmospheric light value, t(x) is the transmittance, and the transmittance estimation is optimized by guided filtering.

4. The method according to claim 1, wherein The edge feature enhancement module extracts the belt edge contour through the Sobel operator and performs channel-level fusion with the feature map output by the backbone network.

5. The method according to claim 1, wherein The loss function of the improved YOLOv6 model includes classification loss, bounding box regression loss and defect morphology loss, and the expression is: L = L cls +1.5L box +2.0L shape Among them L shape The defect contour similarity is calculated based on Hu moment features.

6. The method according to claim 1, wherein The morphological filtering rules include: longitudinal tear: length / width>5 and edge curvature change>3 places / mm; edge wear: edge contour offset>3mm (relative to the standard edge model).

7. The method according to claim 1, characterized in that The time series verification is to trigger an alarm when the same type of defects are detected in three consecutive frames to avoid accidental noise interference.

8. A detection system for implementing the method according to claims 1-7, characterized in that: include: Image acquisition unit: includes a linear array camera, a multi-spectral fill light, and an air curtain dust protection device; data processing unit: integrates an adaptive preprocessing module, an improved YOLOv6 detection model, and a defect verification module; alarm unit: supports sound and light alarms and remote signal transmission (transmission delay <100ms).

9. The system according to claim 8, characterized in that The multi-spectral fill light includes an 850nm infrared lamp and a 450nm blue light lamp, and automatically switches the lighting mode through a photosensor to ensure that the illumination in the detection area is ≥200lux.

10. A computer-readable storage medium storing an executable program, characterized in that: When the program is executed by a processor, the belt defect recognition method according to any one of claims 1 to 7 is implemented, including image preprocessing, improved YOLOv6 detection and defect verification steps.

Images