Conveyor belt damage identification method and system based on X-ray detection

Abstract

The invention discloses a conveyor belt damage identification method and system based on X-ray detection, and relates to the technical field of image identification detection, and the method comprises the following steps: preparing three groups of orthogonally distributed X-ray equipment, starting three groups of orthogonally distributed X-ray sources to synchronously carry out penetration scanning on a running conveyor belt, and starting the X-ray equipment; three-dimensional perspective images of the top, the left side and the right side of the conveying belt are acquired through a linear array detector, meanwhile, an MEMS three-axis acceleration sensor is started to acquire vibration data of the conveying belt in real time, the vibration displacement is calculated, sub-pixel-level geometric correction is carried out based on the vibration displacement, and image offset errors caused by vibration are eliminated. The three groups of orthogonally distributed X-ray sources are arranged for synchronous scanning, and the vibration data are combined to perform offset correction on the image, so that the effect of eliminating the vibration interference during the high-speed operation of the conveying belt is achieved, the stable and clear image is obtained under the dynamic scene of the high-speed operation of the conveying belt, and the detection error caused by poor image quality is reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image recognition detection, in particular to a kind of based on X ray detection conveying belt damage identification method and system. BACKGROUND

[0002] By X ray penetration conveying belt material, in combination with image processing technology to identify internal damage, X ray penetration conveying belt, by detector receives the image formed by the difference in intensity of radiation, in combination with the physical characteristics of steel wire rope core or fabric core carries out damage analysis, which can detect the subtle damage (such as steel wire fracture, adhesive layer peeling, deviation wear, etc.) inside the conveying belt, suitable for high-risk scenarios such as mines, ports, chemical industry, etc., can find potential fault hidden danger in advance.

[0003] In the detection of industrial conveying belt, X ray detection relies on single-view imaging, when the conveying belt is due to vibration image offset, surface attached oil stain or coal dust, it is difficult to cope with vibration interference, stain shielding and micro-damage identification requirements under the high-speed running scene of conveying belt, is easy to cause detection missed detection rate to increase and for micro-damage identification precision is insufficient, influence conveying belt damage detection accuracy and efficiency, and then appear missed detection to lead to safety accident hidden danger or non-planned shutdown loss. SUMMARY

[0004] In view of the deficiencies of the prior art, the present application provides a kind of based on X ray detection conveying belt damage identification method and system, solve the problems mentioned in the background art.

[0005] To achieve the above object, the present application is realized by the following technical scheme: a kind of based on X ray detection conveying belt damage identification method, comprising the following steps: A1, prepare three groups of X ray equipment in orthogonal distribution, start three groups of X ray source in orthogonal distribution and penetrate scanning to the conveying belt in operation synchronously, acquire the three-dimensional perspective image of the top, left side and right side of the conveying belt by line array detector acquisition, simultaneously start MEMS three-axis acceleration sensor and collect conveying belt vibration data in real time, and calculate vibration displacement, carry out sub-pixel level geometric correction based on vibration displacement, eliminate the image offset error caused by vibration; A2, the image corrected after step A1 is input into the pre-trained semantic segmentation network, the stain area is identified by analyzing the pixel gray value difference in the image, the stain area mask is generated and the pixel level shielding treatment is carried out on the area; A3, the non-stain interference image processed in step A2 is input into the backbone network, the shallow texture features and deep semantic features of the three perspective images are extracted respectively, the cross-perspective fusion is carried out on different level features by feature pyramid network, and multi-scale damage feature spectrum is formed; A4. Input the feature map generated in step A3 into the Anchor-Free detection head of the lightweight feature extraction network, combine it with the non-maximum suppression algorithm to perform damage detection, output the specific location coordinates of the conveyor belt damage, damage type and severity quantification score, confirm the damage and damage degree of the conveyor belt, and issue the corresponding audible and visual alarm.

[0006] Preferably, the calculation of the vibration displacement in step A1 uses the Kalman filter algorithm, and the specific calculation formula is as follows: Define the vibration state vector as:

[0007] in, For displacement, For velocity, the predicted value is updated using the state equation:

[0008] in, Here is the state transition matrix. For the control matrix, The acceleration input vector acquired by the MEMS sensor; Covariance update:

[0009] in, To predict the covariance matrix, The process noise covariance matrix; Finally, a measurement update is performed, using the Kalman algorithm for gain, specifically:

[0010] in, For the measurement matrix, To measure the noise covariance, the final corrected displacement is:

[0011] in, For the corrected first Constant conveyor belt vibration displacement For the first The first time data prediction Displacement at any given time For the first The displacement value measured by the MEMS accelerometer at any given time.

[0012] Preferably, the training sample set of the semantic segmentation network in step A2 contains more than 5,000 X-ray images of conveyor belts containing corresponding types of stains, including oil stains, dust, water stains, and debris. The semantic segmentation network algorithm includes: Firstly, the weighted combination of cross-entropy and Dice coefficient is adopted by the loss function, specifically:

[0013] wherein, is the total loss value of the network, is the weighted coefficient, taking the value 0.7, is the cross-entropy loss, is the weighted coefficient of the Dice coefficient loss, taking the value 0.3, is the Dice coefficient loss; Subsequently, the semantic segmentation network extracts image features through the encoder, and the decoder fuses shallow texture features through the skip connection, and outputs the stain area mask with the same size as the input.

[0014] Preferably, in the step A3, the backbone network introduces a cross-view attention mechanism to assign dynamic weights to the feature maps of the three views, enhance the damage feature expression of the occluded area, and the weight coefficient of the dynamic weight ranges from 0.1 to 0.9. The cross-view attention mechanism algorithm is specifically calculated as follows: Let the feature maps of the three views be ; wherein, is the height of the feature map, is the width of the feature map, is the number of feature channels; Then calculate the inter-view feature similarity:

[0015] wherein, is the similarity matrix, is the normalization function, is the scaling coefficient, and are the feature maps of the first and the second view, respectively, is the matrix inner product operation of the feature map; Generate attention weights after feature similarity:

[0016] wherein, is the attention weight coefficient of the first view, is the natural exponential function, is the overall correlation degree of the first view and other views, is the sum of the exponential correlation degrees of the three views; After the attention weight adjustment is completed, the fused feature map is obtained, and the specific process is as follows:

[0017] wherein, is the fused feature map, are the attention weight coefficients of the first, second and third views respectively, are the original feature maps of the three orthogonal views, and the feature response value of the 0.5mm micro damage is improved through the cross-view attention mechanism algorithm.

[0018] Preferably, the lightweight feature extraction network in step A4 adopts a lightweight convolution kernel, and GPU hardware acceleration is used to reduce detection delay and improve micro damage recognition rate.

[0019] A damage identification system based on X-ray detection of a conveyor belt, the damage identification closed loop system comprising: A multi-view scanning module for multi-dimensional X-ray imaging of the conveyor belt, composed of three sets of orthogonally distributed X-ray sources, linear array detector arrays and synchronous trigger modules; A vibration compensation module for collecting vibration data and correcting image offset, which outputs displacement by algorithm processing after collecting conveyor belt vibration parameters, and drives image correction to realize sub-pixel level geometric correction; A deep learning processing module for image feature extraction and damage identification, which performs stain shielding, feature extraction and damage classification on image data; An early warning module for real-time triggering of damage alarm, dynamically adjusting the sensitivity threshold based on the running speed of the conveyor belt, and outputting early warning information through sound and light signals and industrial bus when the damage severity score exceeds the threshold; A prediction module for damage trend prediction, which inputs damage data of the past 30 days based on the LSTM time series network, calculates the damage expansion rate and risk level in the next 7 days, and outputs a 1-10 level maintenance priority ranking; A hardware monitoring module for monitoring the running state of the equipment, which collects X-ray source anode temperature, tube voltage, detector dark current and signal gain in real time, and generates a state abnormal signal when the parameters exceed the preset threshold; A redundancy module for continuous operation guarantee in case of equipment failure, which contains 8-way redundant detector channels and automatic switching circuit, and switches to the redundant channel when the hardware monitoring module detects channel failure; The multi-view scanning module, vibration compensation module, deep learning processing module, early warning and prediction module, and hardware monitoring and redundancy module realize data interaction through the industrial bus to form the damage identification closed loop system.

[0020] Preferably, the deep learning processing module comprises an image preprocessing submodule, a feature extraction submodule, and a damage classification submodule. The image preprocessing submodule: performs noise suppression and dynamic range adjustment on the image, and receives device image data through a PCIe4.0 interface; The feature extraction submodule: integrates a pre-trained semantic segmentation subnetwork and a feature extraction subnetwork, and outputs a 16-layer multi-scale feature map for extracting features of the conveyor belt image; The damage classification submodule: is used for optimizing an Anchor-Free detection head and non-maximum suppression based on a lightweight feature extraction network architecture, and outputs conveyor belt damage coordinates, types, and scores.

[0021] Preferably, the early warning module comprises a threshold response submodule and an early warning output submodule. The threshold response submodule is used for responding to a running speed threshold of the conveyor belt detection process, and calling the conveyor belt to the corresponding threshold in real time through a speed sensor. The early warning output submodule comprises an audible and visual alarm, which synchronously outputs corresponding audible and visual alarm information when the damage score exceeds the threshold.

[0022] Preferably, the prediction module comprises a data storage submodule and a timing calculation submodule. The data storage submodule uses an SSD hard disk to build a damage database, and stores damage records for nearly 30 days, wherein the damage records comprise position coordinates, type labels, size parameters, and environmental temperatures, and support timestamp indexing; The timing calculation submodule: integrates an LSTM neural network chip, inputs a damage data sequence for nearly 30 days, calculates an extended rate for the next 7 days through a trend fitting subunit, and outputs a 1-10 level maintenance priority, wherein the 1 level is the highest level.

[0023] Preferably, the redundancy module comprises a channel detection submodule and a switching control submodule. The channel detection submodule is used for monitoring the signal integrity of the main channel of the 8-way detector in real time, and outputs a channel state value once every 10 ms; The switching control submodule: when receiving an abnormal signal of the hardware monitoring module, cuts off the fault channel and connects the redundant channel through a relay matrix and a redundant channel gating circuit, and simultaneously outputs a switching state feedback signal.

[0024] The application provides an X-ray-based conveyor belt damage identification method and system, which has the following beneficial effects: (1) The application achieves the effect of eliminating the vibration interference when the conveying belt runs at high speed by setting three groups of orthogonally distributed X-ray sources for synchronous scanning and combining vibration data to correct the image offset, realizes the acquisition of stable and clear multi-view images in the dynamic scene of the high-speed running of the conveying belt, breaks through the limitation of single-view imaging through the multi-dimensional imaging mode, can fully capture the damage information of different directions of the conveying belt, avoids the damage shielding problem caused by single view angle, provides a complete and accurate image basis for subsequent damage identification, reduces the detection error caused by poor image quality from the source, and guarantees the continuity and reliability of damage detection.

[0025] (2) The stain area in the image is identified and shielded through the semantic segmentation network, the targeted processing method can effectively distinguish stains from real damage, avoids the case that stains are misjudged as damage or cover damage features, allows the damage identification algorithm to focus on valuable image areas, reduces invalid calculation and false judgment, achieves the effect of excluding interference factors such as oil stains and coal dust, realizes the accurate extraction of the damage features of the conveying belt, thereby improving the accuracy of damage detection and ensuring that each potential damage can be accurately identified, providing a reliable basis for the safety evaluation of the conveying belt.

[0026] (3) The cross-view fusion of the shallow texture features and deep semantic features of the multi-view images is performed, the multi-feature fusion method can integrate damage information under different view angles, make up for the defect that the small damage features are not obvious under single view angle, highlight the subtle damage that is difficult to detect, avoid the case that the small damage is not found and gradually develops into a serious fault, and the damage detection algorithm analysis enhances the effect of expressing small damage features, realizes the effective identification of small damage, thereby reducing the risk of safety accidents and non-scheduled downtime caused by equipment sudden failure, and guarantees the continuous and stable operation of the conveying belt. BRIEF DESCRIPTION OF DRAWINGS

[0027] Figure 1 A method flowchart of the X-ray detection conveying belt damage identification method and system; Figure 2 A method step diagram of the X-ray detection conveying belt damage identification method and system; Figure 3 A system block diagram of the X-ray detection conveying belt damage identification method and system; Figure 4 A sub-module block diagram of the X-ray detection conveying belt damage identification method and system. DETAILED DESCRIPTION

[0028] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0029] Embodiment 1 Please refer to Figures 1-2 The present application provides a kind of based on X-ray detection conveying belt damage identification method and system, to realize above-mentioned purpose, the present application is realized by the following technical solutions: including the following steps: A1, preparation three groups of X-ray equipment that are orthogonal distribution, start three groups of X-ray source that are orthogonal distribution and penetrate scanning to conveying belt in operation synchronously, three-dimensional perspective image of top, left side and right side of conveying belt is acquired by line array detector acquisition, simultaneously start MEMS three-axis acceleration sensor real-time acquisition conveying belt vibration data, and calculate vibration displacement, sub-pixel level geometric correction is carried out based on vibration displacement, eliminate the image offset error caused by vibration; A2, the image after correction in step A1 is input into pre-trained semantic segmentation network, the stain area is identified by analyzing the pixel gray value difference in image, generates stain area mask and carries out pixel level shielding treatment to this area; A3, the image without stain interference after step A2 processing is input into main network, shallow texture feature and deep semantic feature of three visual angle images are extracted respectively, cross-view fusion is carried out to different level features by feature pyramid network, and multi-scale damage feature spectrum is formed; A4, the feature spectrum generated in step A3 is input into Anchor-Free detection head of light weight feature extraction network, damage detection is carried out in combination with non-maximum suppression algorithm, and the specific position coordinates of conveying belt damage, damage type and severity quantization score are output, the damage of conveying belt and damage degree are confirmed, and corresponding sound and light alarm is sent.

[0030] In the embodiment, through multi-view scanning and vibration correction, the limitations of single-view imaging and the image offset problem under high-speed operation are solved, the integrity and stability of image data are ensured, the stain shielding and multi-feature fusion mode effectively exclude the interference factors such as surface dirt of conveying belt under complex working conditions, make damage feature more prominent, light weight detection algorithm is combined with dynamic threshold early warning, realize the closed loop from damage identification to risk early warning, both ensure the detection accuracy, and adapt to the scene demand under different running speeds, provide whole process guarantee for the safe operation of conveying belt.

[0031] Embodiment 2 Specifically, refer to Figure 1 -Figure 2 In step A1, the vibration displacement is calculated using the Kalman filter algorithm, and the specific calculation formula is as follows: Define the vibration state vector as:

[0032] in, For displacement, For velocity, the predicted value is updated using the state equation:

[0033] in, Here is the state transition matrix. For the control matrix, The acceleration input vector acquired by the MEMS sensor; Covariance update:

[0034] in, To predict the covariance matrix, The process noise covariance matrix; Finally, a measurement update is performed, using the Kalman algorithm for gain, specifically:

[0035] in, For the measurement matrix, To measure the noise covariance, the final corrected displacement is:

[0036] in, For the corrected first Constant conveyor belt vibration displacement For the first The first time data prediction Displacement at any given time For the first The displacement value measured by the MEMS accelerometer at any given time.

[0037] In step A2, the training sample set for the semantic segmentation network contains over 5000 X-ray images of conveyor belts containing corresponding types of stains. These stain types include oil stains, dust, water stains, and debris. The semantic segmentation network algorithm includes: First, the loss function employs a weighted combination of cross-entropy and the Dice coefficient, specifically:

[0038] in, This represents the total network loss value. This is the weighting factor, with a value of 0.7. For cross-entropy loss, This is the weighting coefficient for the Dice coefficient loss, with a value of 0.3. For Dice coefficient loss; Subsequently, the semantic segmentation network extracts image features through the encoder, and the decoder fuses shallow texture features through skip connections, outputting a smudge region mask of the same size as the input.

[0039] In step A3, the backbone network introduces a cross-view attention mechanism to assign dynamic weights to the feature maps of the three views, thereby enhancing the representation of damage features in the occluded region. The weight coefficients of the dynamic weights range from 0.1 to 0.9. The specific calculation of the cross-view attention mechanism algorithm is as follows: Let the feature maps from the three perspectives be... ; in, For feature map height, The width of the feature map. Number of feature channels; Then, the feature similarity between viewpoints is calculated:

[0040] in, This is a similarity matrix. For normalization function, This is the scaling factor. and The first The and the first Feature maps from each perspective The inner product operation is performed on the feature maps. Attention weights are generated based on feature similarity.

[0041] in, For the first Attention weight coefficients for each perspective It is a natural exponential function. For the first The overall correlation between one perspective and other perspectives This is the sum of the indexed correlations among the three perspectives; After adjusting the attention weights, the feature maps are fused, as follows:

[0042] in, The fused feature map These are the attention weight coefficients for the 1st, 2nd, and 3rd viewpoints, respectively. The original feature maps of three orthogonal views are respectively, and the feature response value of the 0.5mm micro-damage is improved through the cross-view attention mechanism algorithm.

[0043] In step A4, the lightweight feature extraction network adopts a lightweight convolution kernel, and GPU hardware acceleration is used to reduce detection delay and improve micro-damage recognition rate. In this embodiment, the state transition matrix for updating the predicted value of the state equation in the Kalman filter algorithm is:

[0044] wherein, is the sampling interval; The control matrix for updating the predicted value of the state equation is:

[0045] In the covariance update, the process noise covariance is:

[0046] The measurement noise covariance is:

[0047] The Kalman filter algorithm fuses dynamic prediction and measurement to accurately correct the image offset caused by vibration and provides high-quality data for subsequent processing. The semantic segmentation network with cross-entropy and Dice coefficient weighted loss considers both the classification accuracy of large-area stains and the detail recognition of small-area stains, reduces interference factors, and effectively solves the single-view occlusion problem through dynamic weight allocation, enhances the feature expression of micro-damage, and combines lightweight network with hardware acceleration to ensure accuracy while improving detection speed, making the method suitable for real-time detection of high-speed conveyor belts.

[0048] Embodiment 3 Specifically, refer to Figure 3 Figure 4 A damage identification system based on X-ray detection of a conveyor belt, the damage identification closed-loop system comprising: A multi-view scanning module for multi-dimensional X-ray imaging of the conveyor belt, composed of three sets of orthogonally distributed X-ray sources, linear array detector arrays and synchronous trigger modules; A vibration compensation module for collecting vibration data and correcting image offset, which outputs displacement by algorithm processing after collecting vibration parameters of the conveyor belt, and drives image correction to realize sub-pixel level geometric correction; A deep learning processing module for image feature extraction and damage identification, which performs stain shielding, feature extraction and damage classification on image data; ​The early warning module is used for triggering a damage alarm in real time, dynamically adjusting a sensitivity threshold based on a conveying belt running speed, and outputting early warning information through an audible and visual signal and an industrial bus when a damage severity score exceeds the threshold. The prediction module is used for damage trend prediction, inputs near 30-day damage data based on an LSTM time sequence network, calculates a damage expansion rate and a risk level in the next 7 days, and outputs a 1-10 level maintenance priority ranking. The hardware monitoring module is used for monitoring a device running state, collecting an X-ray source anode temperature, a tube voltage, a detector dark current and a signal gain in real time, and generating a state abnormal signal when a parameter exceeds a preset threshold. The redundancy module is used for continuous running guarantee when a device fails, and comprises 8-way redundant detector channels and an automatic switching circuit. The multi-view scanning module, the vibration compensation module, the deep learning processing module, the early warning and prediction module and the hardware monitoring and redundancy module realize data interaction through an industrial bus, and form a damage identification closed loop system. In the embodiment, the three groups of X-ray sources of the multi-view scanning module are powered by independent adjustable voltage high-voltage power supplies, the linear array detector array is connected with a synchronous trigger through a gigabit Ethernet, and a trigger signal delay is less than 1 ms, so that the three-view image frames are synchronized. The vibration compensation module comprises a MEMS sensor, a signal conditioning circuit and a FPGA image correction chip, analog signals output by the sensor are converted by 16 bits, a displacement amount calculated by a Kalman filtering algorithm run by a processor is transmitted to the FPGA through an SPI interface, and the FPGA corrects image coordinates in real time through sub-pixel interpolation, so that the accuracy of the image is ensured.

[0049] The deep learning processing module adopts an NVIDIA chip, connects multi-view scanning module image data, and adopts quantization acceleration for built-in semantic segmentation, feature fusion and detection network.

[0050] The prediction module is deployed on an edge server, stores near 30-day damage data through a database, and has a small damage expansion prediction error of a conveying belt after 50 rounds of network training, so that the output maintenance priority can be directly connected to an enterprise system. The 8-way redundant detector channels of the redundancy module are arranged in parallel with the main channel, and a switching control circuit relay can ensure the stability of the circuit. The multi-view scanning module realizes full-section imaging through high-precision synchronous triggering, avoids the detection blind area of single view, the hardware and algorithm of the vibration compensation module are cooperatively designed, the image offset is corrected in real time, the reliability of the data is ensured, the high computing power and acceleration technology of the deep learning processing module meet the real-time detection needs of the high-speed conveying belt, the early warning and prediction module is combined, the instant risk alarm is realized, and the prospective maintenance suggestion is provided, and the hardware monitoring and redundancy module improves the stability and fault resistance of the system, so that the whole system can run reliably for a long time in a complex industrial environment.

[0051] Embodiment 4 Specifically, referring to Figure 4 , the deep learning processing module includes an image preprocessing submodule, a feature extraction submodule and a damage classification submodule; The image preprocessing submodule: noise suppression and dynamic range adjustment are performed on the image, and device image data is received through a PCIe4.0 interface; The feature extraction submodule: a pre-trained semantic segmentation subnetwork and a feature extraction subnetwork are integrated, and 16-layer multi-scale feature maps are outputted, which are used for extracting the features of the conveying belt image; The damage classification submodule: an Anchor-Free detection head and a non-maximum suppression based on a lightweight feature extraction network architecture are optimized, and the conveying belt damage coordinates, types and scores are outputted.

[0052] The early warning module includes a threshold response submodule and an early warning output submodule; The threshold response submodule is used for responding to the running speed threshold of the conveying belt detection process, and the conveying belt is called to the corresponding threshold in real time through a speed sensor; The early warning output submodule includes an audible and visual alarm, and the corresponding audible and visual alarm information is outputted when the damage score exceeds the threshold.

[0053] The prediction module includes a data storage submodule and a time sequence calculation submodule: The data storage submodule: an SSD hard disk is used to build a damage database, and damage records in the past 30 days are stored, the damage records include position coordinates, type labels, size parameters and environmental temperature, and support indexing by time stamp; The time sequence calculation submodule: an LSTM neural network chip is integrated, the damage data sequence in the past 30 days is inputted, the future 7-day expansion rate is calculated through a trend fitting subunit, and 1-10 level maintenance priorities are outputted, wherein the 1st level is the highest level.

[0054] The redundancy module includes a channel detection submodule and a switching control submodule: The channel detection submodule is used for monitoring the signal integrity of the main channel of the 8-way detector in real time, and outputs a channel state value once every 10ms; The switching control submodule: when receiving the abnormal signal of the hardware monitoring module, the switching control submodule cuts off the fault channel and connects the redundant channel through the relay matrix and the redundant channel gating circuit, and outputs a switching state feedback signal; In this embodiment, the image preprocessing submodule of the deep learning processing module uses Gaussian filtering to suppress X-ray quantum noise, and the dynamic range adjustment maps the 12-bit image to 8-bit through linear stretching to enhance the contrast between the damage and the background. The semantic segmentation subnetwork of the feature extraction submodule outputs a stain mask, and the stain mask and the original image are operated at the pixel level to obtain a stain-free image. The feature extraction subnetwork outputs 16 layers of feature maps through 16 layers of convolution, wherein the shallow layer retains edge details, and the deep layer extracts semantic information to provide rich features for multi-scale fusion. The Anchor-Free detection head of the damage classification submodule predicts the center coordinates, width and height, and class probability of the damage, and the non-maximum suppression removes repeated predictions by calculating the IOU of the detection box. Finally, the damage score is output. The data storage submodule of the prediction module uses multiple 1TB SSDs to support the writing and querying of 100,000 damage records per day. Through timestamp indexing, data of any time period can be retrieved within 1 second. The LSTM network of the time series calculation submodule contains 3 layers of hidden layers. The trend fitting subunit uses the least squares method to fit the curve of the change of the damage size over time, and outputs the maintenance priority level 1-10.

[0055] The channel detection submodule of the redundancy module sends a CRC check code to the detector every 10ms. When the receiving end fails to verify, the state word has 1 at the xth position, indicating that the xth channel is faulty. The switching control submodule then controls the relay matrix through PLC. After receiving the fault signal, the main channel relay is disconnected, the redundant channel relay is connected, and the current channel state is displayed through the LED indicator. The submodules of the deep learning processing module work together to gradually improve the feature quality and recognition accuracy from image preprocessing to damage classification. The submodules of the early warning module achieve dynamic matching of speed and threshold and reliable output of alarm signals, ensuring timely risk transmission. The submodules of the prediction module provide scientific basis for maintenance decision-making through data storage and time series analysis. The submodules of the redundancy module minimize the impact of equipment failure on detection through real-time detection and rapid switching. The fine design of each module forms a closed-loop optimization in accuracy, efficiency, and reliability, significantly improving the comprehensive performance of the conveyor belt damage detection system.

[0056] Although embodiments of the present application have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and alterations can be made without departing from the principles and spirit of the present application.

Claims

1. A method for identifying damage of a conveyor belt based on X-ray detection, characterized by, The method comprises the following steps: A1, prepare three sets of X-ray equipment in orthogonal distribution, start the three sets of X-ray source in orthogonal distribution to synchronously scan the running conveying belt, acquire the three-dimensional perspective image of the top, left side and right side of the conveying belt through the linear array detector, simultaneously start the MEMS three-axis acceleration sensor to collect the vibration data of the conveying belt in real time, calculate the vibration displacement, carry out sub-pixel level geometric correction based on the vibration displacement, and eliminate the image offset error caused by vibration; A2, input the image corrected in step A1 into the pre-trained semantic segmentation network, identify the stain area by analyzing the pixel gray value difference in the image, generate a stain area mask and perform pixel-level shielding processing on the area; A3, input the stain-free image processed in step A2 into the backbone network, extract the shallow texture features and deep semantic features of the three perspective images respectively, perform cross-perspective fusion on the features at different levels through the feature pyramid network, and form a multi-scale damage feature map; A4, input the feature map generated in step A3 into the Anchor-Free detection head of the lightweight feature extraction network, combine the non-maximum suppression algorithm to perform damage detection, output the specific position coordinates, damage type and severity quantitative score of the conveying belt damage, confirm the damage and damage degree of the conveying belt, and issue corresponding sound and light alarms.

2. The method for identifying damage of a conveyor belt based on X-ray detection according to claim 1, characterized in that, The calculation of the vibration displacement in step A1 adopts the Kalman filtering algorithm, and the specific calculation formula is as follows: The vibration state vector is defined as: ； wherein, is the displacement amount, is the velocity, the predicted value is updated by the state equation: ； wherein, is a state transition matrix, is a control matrix, is an acceleration input vector collected by the MEMS sensor; Covariance update: ； wherein, is a predicted covariance matrix, is a process noise covariance matrix; Finally, the measurement update is performed through the Kalman algorithm gain, which is specifically: ； wherein, is the measurement matrix, is the measurement noise covariance, and finally correcting the displacement amount: ； in, For the corrected first The amount of vibration displacement of the conveyor belt at any given time. For the first The first time data prediction Displacement at any given time For the first The displacement value measured by the MEMS accelerometer at any given time.

3. The method for identifying damage of a conveyor belt based on X-ray detection according to claim 1, characterized in that, The training sample set of the semantic segmentation network in step A2 contains more than 5000 conveying belt X-ray images containing corresponding types of stains, and the stain types include oil stains, dust, water stains and debris. The semantic segmentation network algorithm comprises: Firstly, the weighted combination of cross entropy and Dice coefficient is adopted through the loss function, which is specifically: ； wherein, is a total network loss value, is a weighting coefficient, and takes a value of 0.7, is a cross-entropy loss, is a weighting coefficient of the Dice coefficient loss, and takes a value of 0.3, is a Dice coefficient loss; Subsequently, the semantic segmentation network extracts image features through the encoder, and the decoder fuses shallow texture features through the jump connection to output a stain area mask with the same size as the input.

4. The method for identifying damage of a conveyor belt based on X-ray detection according to claim 1, characterized in that, In step A3, the backbone network introduces a cross-perspective attention mechanism to assign dynamic weights to the feature maps of the three perspectives, enhance the damage feature expression of the occluded area, and the weight coefficient of the dynamic weight ranges from 0.1 to 0.

9. The cross-perspective attention mechanism algorithm is specifically calculated as follows: Let the feature maps of three views be ; wherein, is the feature map height, is the feature map width, is the number of feature channels; Then, the feature similarity between perspectives is calculated: ； wherein, is a similarity matrix, is a normalization function, is a scaling coefficient, and are the feature maps of the th and th view, respectively, is a matrix inner product operation on the feature maps; The attention weight is generated after the feature similarity: ； wherein, is the attention weight coefficient for the th view angle, is the natural exponential function, is the overall correlation between the th view angle and other view angles, is the sum of the exponentialized correlation for the three view angles; After the attention weight adjustment is completed, the fused feature map is obtained, which is specifically as follows: ； wherein, is the fused feature map, are attention weight coefficients of the first, second and third views, respectively, are original feature maps of the three orthogonal views, and the feature response value of the 0.5 mm micro-damage is improved by the cross-view attention mechanism algorithm.

5. The method for identifying damage of a conveyor belt based on X-ray detection according to claim 1, characterized in that, In step A4, the lightweight feature extraction network adopts a lightweight convolution kernel and realizes the reduction of detection delay through GPU hardware acceleration to improve the recognition rate of micro-damage.

6. An X-ray based conveyor belt damage identification system for use in an X-ray based conveyor belt damage identification method according to any one of claims 1 to 5, characterized in that, The damage recognition closed-loop system comprises: A multi-perspective scanning module for multi-dimensional X-ray imaging of the conveying belt, composed of three sets of orthogonal distributed X-ray sources, linear array detector arrays and synchronous trigger modules; A vibration compensation module is configured to collect vibration data and correct image deviation, output displacement through algorithm processing after collecting vibration parameters of the conveying belt, and drive image correction to achieve sub-pixel level geometric correction; A deep learning processing module is configured to extract image features and identify damage, shield stains, extract features, and classify damage from image data; An early warning module is configured to trigger damage alarms in real time, dynamically adjust sensitivity thresholds based on the running speed of the conveying belt, and output early warning information through acoustic and optical signals and an industrial bus when the damage severity score exceeds the threshold; A prediction module is configured to predict damage trends, input damage data of the past 30 days based on an LSTM time series network, calculate damage expansion rates and risk levels for the next 7 days, and output 1-10 level maintenance priority rankings; A hardware monitoring module is configured to monitor the running state of the device, collect X-ray source anode temperature, tube voltage, detector dark current, and signal gain in real time, and generate state anomaly signals when parameters exceed preset thresholds; A redundancy module is configured to ensure continuous operation in the event of device failure, including 8 redundant detector channels and automatic switching circuits that switch to redundant channels when the hardware monitoring module detects channel failure; The multi-view scanning module, vibration compensation module, deep learning processing module, early warning and prediction module, and hardware monitoring and redundancy module interact through an industrial bus to form the damage identification closed-loop system.

7. The X-ray based system for detecting and identifying damage to a conveyor belt of claim 6, wherein, The deep learning processing module includes an image preprocessing submodule, a feature extraction submodule, and a damage classification submodule; The image preprocessing submodule: performs noise suppression and dynamic range adjustment on images, and receives device image data through a PCIe4.0 interface; The feature extraction submodule: integrates a pre-trained semantic segmentation subnetwork and a feature extraction subnetwork, outputs 16-layer multi-scale feature maps, and extracts features from the conveying belt image; The damage classification submodule: optimizes the Anchor-Free detection head and non-maximum suppression based on a lightweight feature extraction network architecture, and outputs the conveying belt damage coordinates, type, and score.

8. The X-ray based system for detecting and identifying damage to a conveyor belt of claim 6, wherein, The early warning module includes a threshold response submodule and an early warning output submodule; The threshold response submodule is configured to respond to the running speed threshold of the conveying belt detection process by calling the conveying belt to the corresponding threshold in real time through a speed sensor; The early warning output submodule includes an audible and visual alarm that outputs corresponding audible and visual alarm information when the damage score exceeds the threshold.

9. The X-ray based system for detecting and identifying damage to a conveyor belt of claim 6, wherein, The prediction module includes a data storage submodule and a time series calculation submodule: The data storage submodule uses an SSD hard drive to build a damage database and store damage records for the past 30 days, which include position coordinates, type labels, size parameters, and environmental temperature, and supports timestamp indexing; The time series calculation submodule: integrates an LSTM neural network chip, inputs a damage data sequence for the past 30 days, calculates the expansion rate for the next 7 days through a trend fitting subunit, and outputs 1-10 level maintenance priority, with level 1 being the highest level.

10. The X-ray based system for detecting and identifying damage to a conveyor belt of claim 6, wherein, The redundancy module includes a channel detection submodule and a switching control submodule: The channel detection sub-module is used for monitoring signal integrity of 8-way probe main channels in real time and outputting a channel state value once every 10 ms; The switching control sub-module is used for cutting off a fault channel and connecting a redundant channel and outputting a switching state feedback signal when receiving an abnormal signal of the hardware monitoring module through a relay matrix and a redundant channel gating circuit.

Images