High-robustness and high-perception-accuracy small sample fault diagnosis method for belt conveyors

Abstract

The present application relates to the technical field of coal mine electromechanical equipment safety monitoring, computer vision and deep learning edge computing, and specifically relates to a high-robustness and high-perception-accuracy small sample fault diagnosis method for a belt conveyor, and the present application constructs a vertical domain robust mask autoencoder VDR-MAE: through multi-scale frequency domain perception dynamic position coding, the anti-interference ability of feature extraction in an industrial high-noise scene is enhanced; through cross-scale fault anchoring reconstruction loss, the problem that small target faults and early faults are submerged by global loss, resulting in low recognition accuracy, is solved; through a mask feature semantic alignment adaptation module, the fine-tuning adaptation effect in a small sample fine-tuning scene is optimized. The accuracy of VDR-MAE in three typical fault diagnosis tasks of surface tearing, edge deviation and surface wear of the coal mine belt conveyor is 90.4%, and the missed detection rate is 4.6%, thereby providing reliable technical support for fault diagnosis of the coal mine belt conveyor.

Description

Technical Field

[0001] This invention relates to the fields of safety monitoring of coal mine electromechanical equipment and the interdisciplinary technologies of computer vision and deep learning edge computing, specifically to a method for small-sample fault diagnosis of belt conveyors with high robustness and high perception accuracy. Background Technology

[0002] Belt conveyors are core transportation equipment in underground coal mining, responsible for transporting raw coal from the working face to the surface storage system. In large underground coal mines in China, the length of main transport roadways with belt conveyors typically exceeds 2km, with operating speeds reaching 4m / s. Under long-term high-load operation, they are prone to surface tearing, edge misalignment, and surface wear. Failure to provide timely warnings can easily lead to major safety accidents such as raw coal spillage, conveyor belt breakage, and fires, with direct economic losses from a single accident reaching millions of yuan. Belt conveyor fault monitoring mainly relies on two types of technologies: contact sensing and visual inspection. While the former (such as tear sensors and misalignment switches) is direct, it has inherent limitations such as high deployment costs, physical vulnerability, and delayed warnings. The latter, with its advantages of non-contact operation, flexible deployment, and early warning, has become a research focus. However, the complex underground environment poses severe challenges to the robustness, real-time performance, and small-sample adaptability of visual inspection.

[0003] With the development of deep learning technology, masked autoencoders (MAEs) based on self-supervised learning have been gradually applied to the field of industrial vision fault diagnosis due to their advantages of not requiring a large amount of labeled data and having strong feature extraction capabilities. In 2021, He Kaiming's team proposed an MAE architecture that uses 75% random masking, the encoder only processes visible patch tokens, and the decoder reconstructs the whole image. It has achieved better results than supervised ViT in general image recognition tasks. However, when directly applied to the fault diagnosis scenario of coal mine belt conveyors, there are three core problems that are difficult to adapt.

[0004] 1) Weak anti-interference capability of location coding: MAE uses fixed sinusoidal location coding or learnable location coding, but underground coal mine images have problems such as uneven lighting, dust obstruction, and periodic texture interference on the surface of the conveyor belt. Fixed location coding is difficult to effectively distinguish the location features of noise disturbances and real faults, and the diagnostic accuracy is less than 80% in low-light and high-noise scenes. 2) Unreasonable weighting of the loss function: MAE uses global pixel mean squared error (MSE) as the reconstruction loss. The features of the normal conveyor belt area, which account for more than 95%, will dominate the loss update. This results in the reconstruction accuracy of minor tears with a width of less than 5mm and early weak manifestation faults with surface wear of less than 2mm being less than 40%, and the early fault identification accuracy being less than 80%. 3) Poor generalization ability of fine-tuning with small samples: There are few fault samples in coal mines, and the number of labeled samples for a single type of fault is generally less than 1,000. There is a feature distribution offset between the pre-trained general features and the fault features. Direct fine-tuning is prone to mode collapse. The generalization accuracy is less than 80% in small sample scenarios.

[0005] Currently, some researchers have made customized improvements to MAE, but these improvements are mostly geared towards general image classification and remote sensing image recognition scenarios. There is no customized improvement plan for the entire process of fault diagnosis of coal mine belt conveyors, nor has an adaptable technical system covering the entire process of pre-training, fine-tuning, and inference been formed, which cannot meet the actual application needs of complex underground scenarios. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a robust and highly accurate method for small-sample fault diagnosis of belt conveyors, specifically designed to solve the issues of poor robustness in feature extraction under high noise conditions, easy loss of fault features in small targets, low accuracy in early fault identification, and weak generalization ability in small-sample scenarios in visual fault diagnosis of coal mine belt conveyors. The technical solution adopted by this invention to solve these problems is as follows: A robust and highly accurate fault diagnosis method for small-sample belt conveyors, characterized by comprising: S1. Using the pre-trained dataset and fine-tuning dataset that have already been built, additional specialized test datasets are constructed. S2. Construct a multi-scale frequency domain sensing dynamic position coding, cross-scale fault anchoring reconstruction loss, and mask feature semantic alignment and adaptation module. S3. Introduce multi-scale frequency domain sensing dynamic position coding, cross-scale fault anchoring reconstruction loss and mask feature semantic alignment and adaptation module into the mask autoencoder MAE to construct the vertical domain robust mask autoencoder VDR-MAE. S4. First, pretrain VDR-MAE using the pretraining dataset, then fine-tune the pretrained VDR-MAE using the fine-tuning dataset, and finally evaluate the performance of the fine-tuned VDR-MAE using the specialized test dataset. S5. Use VDR-MAE to process conveyor belt diagrams and perform small-sample fault diagnosis of belt conveyors.

[0007] Furthermore, in step S1, both the pre-training dataset and the fine-tuning dataset consist of a conveyor belt map, a fault mask map, and a conveyor belt mask map; The conveyor belt diagram consists of a normal conveyor belt diagram and a faulty conveyor belt diagram. The fault categories in the faulty conveyor belt diagram include: surface tearing, edge misalignment, and surface wear. Furthermore, the process of constructing the special test dataset is as follows: 100,000 historical conveyor belt images were collected collaboratively from other mining areas, and 240 faulty conveyor belt images were selected from them to construct the special test dataset, which includes four special scenarios: low illumination, high noise, dust obstruction, and early failure, with 60 faulty conveyor belt images for each scenario.

[0008] Furthermore, the specific implementation of multi-scale frequency domain sensing dynamic position coding in step S2 is as follows: Perform a dimensionality transformation permutation on each Patch Token output by the linear projection layer of the Masked Autoencoder (MAE) to obtain the image patch. ;right A two-dimensional discrete cosine transform (2D-DCT) is performed to obtain the discrete cosine transform matrix. The discrete cosine transform matrix is ​​then split into three frequency domain components based on the discrete cosine transform coefficients: the upper left quarter region is the low-frequency component, corresponding to global illumination and large-scale structures; the middle half region is the mid-frequency component, corresponding to the periodic texture of the conveyor belt; and the lower right quarter region is the high-frequency component, corresponding to fault edges and noise. The sinusoidal position coding of MAE is improved using the low-frequency, mid-frequency, and high-frequency components. The low-frequency, mid-frequency, and high-frequency components are substituted into the sinusoidal function for calculation, and a multi-scale frequency domain sensing dynamic position coding is constructed based on the calculation results.

[0009] Furthermore, in step S2, the specific implementation of the cross-scale fault anchoring reconstruction loss is as follows: At the output of the last Transformer Encoder Block of the Masked Autoencoder (MAE), a lightweight fault region candidate head is added. The fault region candidate head consists of two 1×1 convolutional layers and outputs the probability score of each Patch Token containing a fault. Only retain Patch tokens with a value higher than 0.5 are used as Patch tokens containing fault regions. Based on the reconstructed image output by the decoder of the conveyor belt map and the mask autoencoder (MAE) and the Patch tokens containing fault regions, the global pixel mean square error (MSE) loss and the cross-scale fault anchoring reconstruction loss are calculated. The final loss is constructed based on the global pixel MSE loss and the cross-scale fault anchoring reconstruction loss.

[0010] Furthermore, in step S2, the specific implementation of the mask feature semantic alignment and adaptation module is as follows: pre-train the MAE to generate a pre-trained MAE, and fine-tune the original MAE based on the pre-trained MAE and output VRD-MAE.

[0011] Furthermore, before fine-tuning, the pre-trained dataset from step S1... Input a normal conveyor belt diagram and a faulty conveyor belt diagram into the pre-trained MAE; The output of the last layer Transformer Encoder Block of the pre-trained MAE A list of 196 Patch Tokens is used. The mean vector of each Patch Token list is calculated as the image feature vector for each conveyor belt map. The conveyor belt maps are then grouped according to the state corresponding to the image feature vector. Each image feature vector is grouped together with other image feature vectors of the same state. The average vector of all image feature vectors within each group is calculated to obtain the initial class prototype vector of dimension 768 for each state. , Index of the state; Furthermore, during fine-tuning, firstly, for each conveyor belt image in the current batch of the input pre-trained MAE, an image feature vector is calculated; then, the image feature vector is compared with the true state of the conveyor belt image. The L2 distance is calculated; finally, the L2 distances of each conveyor belt map in the current batch are summed to obtain the semantic alignment loss; the semantic alignment loss is weighted and then the cross-entropy classification loss of MAE is introduced to obtain the fine-tuned total loss.

[0012] Furthermore, during fine-tuning, the initial category prototype vector is updated to the category prototype vector in the 5th generation. Thereafter, the category prototype vector is updated once every 3 iterations. The update method is basically the same as the method for calculating the initial category prototype vector. The difference is that an additional momentum constraint of 0.9 is added during the update to avoid the model from collapsing due to prototype oscillation.

[0013] The beneficial effects of this invention are as follows: Multi-scale frequency domain sensing dynamic position coding improves the robustness of MAE in feature extraction from conveyor belt maps with faults in low-light, high-noise, and dust-obscured environments downhole. In a specialized test dataset, the fault diagnosis accuracy increased from 86.7% to 88.4%, solving the problem of poor anti-interference capability of MAE's sinusoidal position coding in complex production environments. Cross-scale fault anchoring reconstruction loss increases the average recognition accuracy of MAE for early faults such as minor tears (width less than 5mm), early deviation (deviation less than 20mm), and slight wear (wear less than 2mm) from 79.7% to 87.2%, while reducing the false alarm rate from 15.4% to 4.8%. This solves the problem of small-scale, weak-performance fault features being overwhelmed by global loss, achieving accurate perception of early faults and advancing fault warning time by 12-24 hours, significantly reducing the risk of fault escalation.

[0014] The mask feature semantic alignment and adaptation module improved the fault diagnosis accuracy of MAE from 88.7% to 90.4% in small sample fine-tuning scenarios with less than 100 training samples per class. The model's generalization ability was significantly improved, which solved the problem of pattern collapse caused by feature distribution shift under small sample conditions. It also reduced the demand for fault labeling data and reduced the cost of data collection and labeling. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating a method for diagnosing small-sample faults in a belt conveyor with high robustness and high sensing accuracy, as described in this application. Figure 2 This is a schematic diagram of the structure of VDR-MAE in the embodiments of this application. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will now be described in further detail with reference to the accompanying drawings and the following embodiments, so that the public can better understand the implementation method of this invention. The specific implementation scheme of this invention is as follows: First, MAE uses sinusoidal positional encoding or learnable positional encoding, failing to consider the uneven lighting, dust obstruction, and periodic texture interference on the conveyor belt surface in images of underground coal mine belt conveyors. Fixed positional encoding cannot distinguish between noise disturbances and the positional features of actual faults, resulting in poor robustness of the model in identifying fault features in low-light, high-noise scenes. Second, MAE uses global pixel mean squared error (MSE) as the reconstruction loss, failing to consider the distribution characteristic that the fault area of ​​the belt conveyor accounts for less than 5% of the pixels. The global loss is dominated by the features of the extremely high proportion of normal conveyor belt areas, resulting in low reconstruction accuracy of the model for small-scale, weakly represented faults (such as early wear and minor tears), and poor accuracy in identifying early faults in subsequent diagnostic tasks. Third, after the MAE pre-training is completed, during the fine-tuning stage, a multilayer perceptron (MLP) is only connected after the last Transformer Encoder layer as the classification head for fine-tuning. This does not take into account the small sample scenario where belt conveyor fault samples are scarce (less than 100 labeled samples for a single fault class). Furthermore, there is a distributional offset between the general image features learned during the pre-training stage and the semantic features of fault diagnosis. Direct fine-tuning is prone to overfitting, resulting in poor generalization ability for small samples.

[0017] Based on this, this application provides a highly robust and accurate small-sample fault diagnosis method for belt conveyors. The core improvement logic is to "enhance the anti-interference ability of feature extraction, focus on small targets and weak fault (early fault) perception, and reduce semantic distribution offset". The three core modules of MAE, namely "position encoding, loss function and fine-tuning adaptation", are customized and improved in the whole chain to construct VDR-MAE and improve the intelligent level of visual fault diagnosis of coal mine belt conveyors.

[0018] The various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0019] This application provides a method for diagnosing small-sample faults in belt conveyors with high robustness and high sensing accuracy, such as... Figure 1 As shown, the method may include steps S101-S105: S101 uses the pre-trained dataset and fine-tuning dataset that have already been built, and additional specialized test datasets are constructed.

[0020] In this embodiment of the application, the process of constructing the pre-training dataset, the fine-tuning dataset, and the specialized test dataset is as follows: We collected 20,000 images of conveyor belts from the internet, with faulty conveyor belts accounting for about 40%. These images were used as pre-training data. The fault categories of the faulty conveyor belt images included: surface tearing, edge misalignment, and surface wear. An explosion-proof network camera was deployed every 50 meters along the main transport lane belt conveyor, for a total of 60 cameras, to collect data. A total of 80,000 conveyor belt images were collected, covering light intensity of 10 lux to 500 lux, dust concentration of 2 mg / m³ to 12 mg / m³, and no-load to full-load scenarios. Faulty conveyor belt images accounted for approximately 0.3%. From the 80,000 conveyor belt images, 320 images were selected, 80 for normal and 80 for faulty conveyor belts, as fine-tuning data. We collected 100,000 historical conveyor belt images from other mining areas, and selected 240 faulty conveyor belt images to construct a special test dataset. This dataset includes four special scenarios: low illumination (brightness below 50 lux), high noise (signal-to-noise ratio less than 30 dB), dust obstruction (obstruction area greater than 30%), and early failure. There are 60 faulty conveyor belt images for each scenario as the special test dataset.

[0021] Furthermore, three coal mine electromechanical engineers with more than 5 years of experience used the open-source LabelMe tool to perform cross-semantic segmentation annotation on the pre-training data and fine-tuning data, obtaining a fault mask map with the fault area as the foreground and a conveyor belt mask map with the conveyor belt area as the foreground; using the conveyor belt map, fault mask map and conveyor belt mask map, a dataset was constructed to obtain the pre-training dataset and the fine-tuning dataset, respectively.

[0022] In this embodiment, the explosion-proof network camera is installed 1.2m directly above the conveyor belt, covering the full width of the conveyor belt. The model is Hikvision DS-2XE6422FWD-IZHS(D), which meets the explosion-proof requirements of coal mine Ex d IIC T6 Gb, has an IP68 protection rating, a resolution of 1920×1080, a sampling frame rate of 25fps, and a shutter speed of 1 / 1000s.

[0023] S102, constructs a multi-scale frequency domain sensing dynamic position coding, cross-scale fault anchoring reconstruction loss, and mask feature semantic alignment and adaptation module.

[0024] In the embodiments of this application, such as Figure 2 As shown, the implementation of scale-frequency domain-aware dynamic position coding includes: Perform a dimensionality transformation permutation on each Patch Token output by the linear projection layer of MAE to obtain the image patch. ;right A two-dimensional discrete cosine transform (2D-DCT) is performed to obtain the discrete cosine transform matrix. The matrix is ​​then decomposed into three frequency domain components based on the discrete cosine transform coefficients: the upper left quarter is the low-frequency component, corresponding to global illumination and large-scale structures; the middle half is the mid-frequency component, corresponding to the periodic texture of the conveyor belt; and the lower right quarter is the high-frequency component, corresponding to fault edges and noise. The low-frequency, mid-frequency, and high-frequency components are used to improve the sinusoidal position coding of the dynamic position coding system (MAE), constructing a multi-scale frequency domain sensing dynamic position coding system. The calculation formula is as follows: ,in, The spatial coordinates of the patch in the conveyor belt diagram; for The location encoding vector of the Patch Token at that point; It is a sine function; , The frequency and offset parameters for sinusoidal position encoding; The frequency domain feature weights have a value range of [value range missing]. In this invention, 0.35 is used; For the learnable first The weighted coefficients of each frequency domain component are used as parameters of the model and trained together with other model parameters; for place The Each frequency domain component Corresponding to low-frequency components, Corresponding to the intermediate frequency component, Corresponding to high-frequency components.

[0025] This improvement enhances the robustness of MAE in extracting features from conveyor belt maps with low illumination, high noise, and dust obstruction in downhole environments. In the special test dataset, the fault diagnosis accuracy increased from 86.7% to 88.4%, solving the problem of poor anti-interference capability of MAE's sinusoidal position coding in complex production environments.

[0026] In the embodiments of this application, such as Figure 2 As shown, the implementation of cross-scale fault anchoring reconstruction loss includes: At the output of the last Transformer Encoder Block in MAE, a lightweight fault region candidate head is added. The fault region candidate head consists of two 1×1 convolutional layers, outputting the probability score of the patch corresponding to each patch token containing a fault. Only retain Patch tokens with a value higher than 0.5 are used as Patch tokens containing fault regions. Based on the conveyor belt map, the reconstructed image output from the MAE decoder, and the Patch tokens containing fault regions, the global pixel mean square error (MSE) loss and the cross-scale fault anchoring reconstruction loss are calculated. The final loss is constructed based on the global pixel MSE loss and the cross-scale fault anchoring reconstruction loss, and its calculation formula is as follows: ,in, The final loss is determined by the global pixel MSE loss. and cross-scale fault anchoring reconstruction loss Weighted composition; The total number of pixels in the conveyor belt map or reconstructed image; The index of pixels for conveying a tape map or reconstructed image; The first part of the conveyor belt diagram The normalized true value of each pixel, with a value range of . ; For the reconstruction of the image The normalized predicted value of each pixel, with a value range of [value range missing]. ; The weighting coefficients for cross-scale fault anchoring reconstruction loss have a range of values. The optimal value selected by the present invention through grid search is 1.8; The number of Patch Tokens containing fault regions output in the fault region candidate header; An index for the Patch Token containing the faulty region; For the first The probability score of each Patch Token, with a value range of [value missing]. ; For the general VGG16 model Feature extraction functions for each convolutional layer; These are the indices of the three convolutional layers of different scales, conv2_2, conv3_3, and conv4_3, in the general VGG16 model, with index values ​​of 1, 2, and 3 respectively. For the first A Patch Token containing the faulty region; For The corresponding patch position is used as a reference to reconstruct the patch at the same position in the image; For Permut operations; This is an L1 norm operation used to calculate the absolute error between features.

[0027] This improvement increases the average accuracy of MAE in identifying early faults such as minor tears (width <5mm), early deviation (deviation <20mm), and light wear (wear <2mm) from 79.7% to 87.2%, while reducing the false alarm rate from 15.4% to 4.8%. It solves the problem of small-scale, weak-performing fault characteristics being overwhelmed by global losses, and achieves accurate perception of early faults. It can advance the fault warning time by 12-24 hours, significantly reducing the risk of fault expansion.

[0028] In the embodiments of this application, such as Figure 2 As shown, the implementation of the mask feature semantic alignment and adaptation module includes: Before fine-tuning, the pre-trained dataset... Input two normal and faulty conveyor belt images into the pre-trained MAE. Output the last layer of the pre-trained MAE's Transformer Encoder Block. A list of 196 PatchTokens is generated; the mean vector of each PatchToken list is calculated as the image feature vector for each conveyor belt map; the conveyor belt maps are grouped according to the state (including normal state and fault category) corresponding to the image feature vectors. We have several image feature vectors, grouping image feature vectors of the same state together; we calculate the average vector of all image feature vectors within each group to obtain the initial class prototype vector of dimension 768 for each state. , Index of the state; Furthermore, during fine-tuning, firstly, for each conveyor belt image in the current batch of the input pre-trained MAE, an image feature vector is calculated; then, the image feature vector is compared with the true state of the conveyor belt image. The L2 distance is calculated; finally, the L2 distances of each conveyor belt map in the current batch are summed to obtain the semantic alignment loss; the semantic alignment loss is weighted and then the cross-entropy classification loss of MAE is introduced to obtain the fine-tuning total loss, the calculation formula of which is: ,in: To fine-tune the total loss; Cross-entropy is used for classification loss; These are the weight coefficients for the semantic alignment loss, with values ​​ranging from... The optimal value selected by the present invention through hyperparameter search is 0.55; To fine-tune the batch size, the optimal value selected by the present invention through hyperparameter search is 16; For the first Image feature vectors of a conveyor belt diagram; For the first Class State , For the first The actual state of a conveyor belt diagram; This is an L2 norm operation used to calculate the L2 distance between the image feature vector and the initial class prototype vector; Furthermore, during fine-tuning, the initial category prototype vector is updated to the category prototype vector in the 5th generation. Thereafter, the category prototype vector is updated every 3 iterations. The update method is basically the same as the method for calculating the initial category prototype vector, the difference being that an additional momentum constraint of 0.9 is added during the update to avoid model collapse caused by prototype oscillation fluctuations. The update formula is as follows: ,in, The number of iterations to update the category prototype vector; and The first Daihe The first generation Class prototype vector; For the first The first generation The mean vector of all image feature vectors of the class.

[0029] S103 introduces multi-scale frequency domain-aware dynamic position coding, cross-scale fault anchoring reconstruction loss, and mask feature semantic alignment adaptation modules into the mask autoencoder (MAE) to construct a vertically robust mask autoencoder (VDR-MAE).

[0030] S104: First, VDR-MAE is pre-trained using a pre-training dataset. Then, the pre-trained VDR-MAE is fine-tuned using a fine-tuning dataset. Finally, the fine-tuned VDR-MAE is evaluated using a specialized test dataset.

[0031] S105, use VDR-MAE to process conveyor belt diagrams and perform small sample fault diagnosis of belt conveyors.

[0032] Experimental verification and effect analysis All control experiments in this embodiment adhere to the "five same principles": same hardware platform, same dataset partitioning, same data augmentation strategy, same number of training epochs, and same optimizer configuration. Only the model architecture and training strategy are changed to ensure fairness in the comparison.

[0033] Five control experiments were set up to fully verify the performance of VDR-MAE. All experiments used the same dataset and hardware environment. The evaluation metrics included five core indicators: diagnostic accuracy, recall, F1 score, false negative rate, and false positive rate.

[0034] 1. Benchmark performance comparison experiment Four mainstream industrial vision fault diagnosis methods—MAE, YOLOv8n, Faster R-CNN, and supervised ViT-B / 16—were selected as controls. The performance comparison results on the test set of the fine-tuning dataset are shown in the table below:

[0035] Experimental results show that the VDR-MAE proposed in this invention is superior to the control method in all core indicators. The accuracy is 1.7 percentage points higher than that of MAE, and the false alarm rate is 5.1 percentage points lower. It fully meets the technical requirements proposed by the mine for fault identification accuracy ≥90%, false alarm rate ≤5%, and false alarm rate ≤5%.

[0036] 2. Robustness Experiment in Complex Scenarios The diagnostic accuracy of this invention and MAE were compared on fault conveyor belt images and normal images in three special scenarios: low illumination, high noise, and dust obstruction, respectively, in the special test dataset. The results are as follows: In the low illumination scenario, the accuracy of this invention is 89.4%, while the accuracy of MAE is 86.9%, an improvement of 2.5 percentage points; in the high noise scenario, the accuracy of this invention is 88.7%, while the accuracy of MAE is 87.3%, an improvement of 1.4 percentage points; in the dust obstruction scenario, the accuracy of this invention is 87.2%, while the accuracy of MAE is 85.8%, an improvement of 1.4 percentage points; in the normal scenario, the accuracy of this invention is 90.8%, while the accuracy of MAE is 90.4%, an improvement of 0.4 percentage points.

[0037] Experimental results reveal that the performance improvement of multi-scale frequency-domain aware dynamic position coding is significantly higher in three special scenarios than in normal scenarios. This indicates that the gain of this coding mainly comes from its ability to decouple noise and signal in the frequency domain—through learnable weighting coefficients. The model automatically suppresses the contributions of high-frequency noise components (dust, sensor noise) and strongly periodic mid-frequency components (conveyor belt texture) to location encoding, making the attention mechanism more focused on the non-periodic frequency domain perturbations introduced by anomalies. In contrast, fixed sinusoidal encoding treats all frequency bands equally, causing noise location information to contaminate the similarity calculation of Patch Tokens.

[0038] 3. Performance Experiments in Small Sample Scenarios Fine-tuning was performed using 40, 60, and 80 single-type fault conveyor belt diagrams, respectively, and the diagnostic accuracy of the present invention was compared with that of MAE: With 40 single-type fault conveyor belt diagrams, the accuracy of the present invention was 85.4%, while that of MAE was 82.7%, an improvement of 2.7 percentage points; with 60 single-type fault conveyor belt diagrams, the accuracy of the present invention was 87.2%, while that of MAE was 85.8%, an improvement of 1.4 percentage points; with 80 single-type fault state diagrams, the accuracy of the present invention was 90.4%, while that of MAE was 88.7%, an improvement of 1.7 percentage points.

[0039] Experimental results show that the mask feature semantic alignment and adaptation module effectively reduces the distribution offset between pre-trained general features and fault semantic features, significantly reduces the model's requirement for labeled data, and solves the pain point of scarce coal mine fault samples.

[0040] 4. Early Fault Identification Experiment The recognition performance of this invention and MAE was compared on early fault samples in a special test dataset: the average recognition accuracy of this invention was 87.2% and the false alarm rate was 4.8%; while the average recognition accuracy of MAE was only 79.7% and the false alarm rate was 15.4%.

[0041] Experimental results demonstrate that cross-scale anchored weighted loss effectively enhances the model's ability to perceive weak performance and small-scale faults, providing early warnings of faults 12 to 24 hours in advance and preventing the faults from escalating into major safety incidents.

[0042] Practical application cases The VDR-MAE proposed in this invention was officially deployed and put into operation in the main transport roadway of a large state-owned coal mine in June 2025. As of August 2026, it had been running stably for 14 months, monitoring the belt conveyor for a total of approximately 7,900 hours. A total of 12 faults of various types were identified, including: 2 surface tears (minimum identified tear width 2.1 mm), 8 early deviations (minimum identified deviation amount 12 mm), and 2 early wears (minimum identified wear amount 1.2 mm). The actual operating accuracy rate was 96.1%, the missed detection rate was 1.57%, and the false alarm rate was 3.93%. All indicators met the preset requirements.

[0043] On November 12, 2025, the system issued a warning of a surface tear with a width of 3.2 mm at a distance of 1.2 km from the working face, with a fault confidence level of 0.92. The ground dispatch center immediately arranged for maintenance personnel to shut down the machine for inspection and promptly repaired the torn area, preventing a belt breakage accident caused by the tear expanding. The estimated reduction in direct economic losses was approximately 2.8 million yuan. On February 23, 2026, the system issued a warning of an early deviation fault at a distance of 1.8 km from the ground. Maintenance personnel only needed 15 minutes to adjust the position of the idler rollers to restore normal operation, preventing a coal spillage and a shutdown accident of more than 3 hours caused by the deviation worsening, reducing direct economic losses by approximately 350,000 yuan.

[0044] The various embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system and device embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the description of the method embodiments.

[0045] The devices and methods provided in this application are one-to-one correspondences. Therefore, the devices also have similar beneficial technical effects as their corresponding methods. Since the beneficial technical effects of the methods have been described in detail above, the beneficial technical effects of the systems and devices will not be repeated here.

[0046] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0047] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A robust and highly accurate fault diagnosis method for small-sample belt conveyors, characterized in that, include: S1. Using the pre-trained dataset and fine-tuning dataset that have already been built, additional specialized test datasets are constructed. S2. Construct a multi-scale frequency domain sensing dynamic position coding, cross-scale fault anchoring reconstruction loss, and mask feature semantic alignment and adaptation module. S3. Introduce multi-scale frequency domain sensing dynamic position coding, cross-scale fault anchoring reconstruction loss and mask feature semantic alignment and adaptation module into the mask autoencoder MAE to construct the vertical domain robust mask autoencoder VDR-MAE. S4. First, pretrain VDR-MAE using the pretraining dataset, then fine-tune the pretrained VDR-MAE using the fine-tuning dataset, and finally evaluate the performance of the fine-tuned VDR-MAE using the specialized test dataset. S5. Use VDR-MAE to process conveyor belt diagrams and perform small-sample fault diagnosis of belt conveyors.

2. The method for diagnosing small-sample faults in belt conveyors with high robustness and high sensing accuracy according to claim 1, characterized in that: In step S1, both the pre-training dataset and the fine-tuning dataset consist of a conveyor belt map, a fault mask map, and a conveyor belt mask map. The conveyor belt diagram consists of a normal conveyor belt diagram and a faulty conveyor belt diagram. The fault categories in the faulty conveyor belt diagram include: surface tearing, edge misalignment, and surface wear.

3. The method for diagnosing small-sample faults in belt conveyors with high robustness and high sensing accuracy according to claim 2, characterized in that, The process of constructing the special test dataset is as follows: 100,000 historical conveyor belt images were collected collaboratively from other mining areas, and 240 faulty conveyor belt images were selected from them to construct the special test dataset, which includes four special scenarios: low light, high noise, dust obstruction, and early failure, with 60 faulty conveyor belt images for each scenario.

4. The method for diagnosing small-sample faults in belt conveyors with high robustness and high sensing accuracy according to claim 1, characterized in that, The specific implementation of multi-scale frequency domain sensing dynamic position coding in step S2 is as follows: Perform a dimensionality transformation permutation on each Patch Token output by the linear projection layer of the Masked Autoencoder (MAE) to obtain the image patch. ;right Two-dimensional discrete cosine transform (2D-DCT) is performed to obtain the discrete cosine transform matrix. The discrete cosine transform matrix is ​​then split into three frequency domain components according to the discrete cosine transform coefficients: the upper left 1 / 4 region is the low frequency component, corresponding to global illumination and large-scale structures; the middle 1 / 2 region is the mid frequency component, corresponding to the periodic texture of the conveyor belt. The lower right quarter area contains high-frequency components, corresponding to fault edges and noise. The sinusoidal position coding of MAE is improved by using low-frequency, mid-frequency and high-frequency components. The low-frequency, mid-frequency and high-frequency components are substituted into the sinusoidal function for calculation, and a multi-scale frequency domain sensing dynamic position coding is constructed based on the calculation results.

5. The method for diagnosing small-sample faults in belt conveyors with high robustness and high sensing accuracy according to claim 1, characterized in that, In step S2, the specific implementation of the cross-scale fault anchoring reconstruction loss is as follows: At the output of the last Transformer Encoder Block of the Masked Autoencoder (MAE), a lightweight fault region candidate head is added. The fault region candidate head consists of two 1×1 convolutional layers and outputs the probability score of each Patch Token containing a fault. Only retain Patch tokens with a value higher than 0.5 are used as Patch tokens that contain faulty regions; Based on the reconstructed image output by the decoder of the conveyor belt map and the mask autoencoder (MAE) and the PatchToken containing the fault region, the global pixel mean square error (MSE) loss and the cross-scale fault anchoring reconstruction loss are calculated. The final loss is constructed based on the global pixel MSE loss and the cross-scale fault anchoring reconstruction loss.

6. The method for diagnosing small-sample faults in belt conveyors with high robustness and high sensing accuracy according to claim 1, characterized in that, In step S2, the specific implementation of the mask feature semantic alignment and adaptation module is as follows: pre-train the MAE to generate a pre-trained MAE, and fine-tune the original MAE based on the pre-trained MAE and output VRD-MAE.

7. The method for diagnosing small-sample faults in belt conveyors with high robustness and high sensing accuracy according to claim 6, characterized in that, Before fine-tuning, the pre-trained dataset from step S1... Input a normal conveyor belt diagram and a faulty conveyor belt diagram into the pre-trained MAE; The output of the last layer Transformer Encoder Block of the pre-trained MAE A list of 196 PatchTokens is used. The mean vector of each PatchToken list is calculated as the image feature vector for each conveyor belt map. The conveyor belt maps are then grouped according to the state corresponding to the image feature vector. Each image feature vector is grouped together with other image feature vectors of the same state. The average vector of all image feature vectors within each group is calculated to obtain the initial class prototype vector of dimension 768 for each state. , The index of the state.

8. The method for diagnosing small-sample faults in belt conveyors with high robustness and high sensing accuracy according to claim 6, characterized in that, During fine-tuning, firstly, for each conveyor belt image in the current batch of the pre-trained MAE, an image feature vector is calculated; then, the image feature vector is compared with the true state of the conveyor belt image. The L2 distance is calculated; finally, the L2 distances of each conveyor belt map in the current batch are summed to obtain the semantic alignment loss; the semantic alignment loss is weighted and then the cross-entropy classification loss of MAE is introduced to obtain the fine-tuned total loss.

9. The method for diagnosing small-sample faults in belt conveyors with high robustness and high sensing accuracy according to claim 6, characterized in that, During fine-tuning, the initial category prototype vector is updated to the category prototype vector in the 5th generation. After that, the category prototype vector is updated once every 3 iterations. The update method is basically the same as the method for calculating the initial category prototype vector. The difference is that an additional momentum constraint of 0.9 is added during the update to avoid the model from collapsing due to prototype oscillation.

Images