A deep learning-based intelligent diagnosis and prediction system for mine equipment failure

Abstract

The application discloses a kind of based on deep learning's mine equipment fault intelligent diagnosis prediction system, it is related to intelligent diagnosis technical field, including, data acquisition and pretreatment module, for collecting the operating data of equipment and pretreatment obtains the data after standardization;Time-frequency chart generation module, for the data after standardization generates preliminary time-frequency chart, and is formed by space attention mechanism combined joint mask enhancement time-frequency chart, based on enhancement time-frequency chart through TextureCrop method optimization NetA model and NetB model, respectively generate scanning image and combine, target time-frequency chart is generated by clipping;High-dimensional feature extraction module, for target time-frequency chart uses deep convolutional neural network and carries out feature extraction and obtains high-dimensional feature vector, uses graph convolution network and updates high-dimensional feature vector.The application improves the accuracy of mine equipment fault diagnosis.

Description

Technical Field

[0001] This invention relates to the field of intelligent diagnostic technology, and in particular to an intelligent diagnostic and prediction system for mine equipment faults based on deep learning. Background Technology

[0002] With the increasing complexity of mining equipment, equipment failure prediction and diagnosis have become crucial tasks in mine safety management. Traditional methods for detecting mining equipment failures mainly rely on manual inspections and rule-based pre-set models. However, these methods have limitations, such as the inability to monitor equipment status in real time, the inability to detect latent faults promptly, and the susceptibility to misdiagnosis in the face of complex and ever-changing mining environments. In recent years, with the development of sensing technology, data processing technology, and deep learning algorithms, data-driven equipment failure prediction and diagnosis methods have gradually become a research hotspot. By utilizing data such as vibration, temperature, and pressure generated during equipment operation, combined with deep learning algorithms, real-time monitoring, diagnosis, and prediction of equipment failures can be achieved, greatly improving the safety and reliability of mining equipment. However, there are still shortcomings in the field of mining equipment failure diagnosis. Existing methods mostly rely on single-type sensor data, neglecting the fusion of multi-sensor data, resulting in a lack of comprehensiveness in failure diagnosis. Traditional deep learning models often use general time-frequency analysis methods, failing to incorporate spatial attention mechanisms and joint masking techniques for feature enhancement, leading to the underutilization of effective information in the time-frequency graph and thus affecting the accuracy of failure prediction. Summary of the Invention

[0003] In view of the aforementioned existing problems, the present invention is proposed.

[0004] Therefore, this invention provides a deep learning-based intelligent fault diagnosis and prediction system for mining equipment, which solves the problems of low fault diagnosis accuracy and insufficient feature extraction in the prior art.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0006] In a first aspect, the present invention provides a deep learning-based intelligent fault diagnosis and prediction system for mining equipment, comprising,

[0007] The data acquisition and preprocessing module is used to collect the equipment's operating data and preprocess it to obtain standardized data.

[0008] The time-frequency map generation module is used to generate a preliminary time-frequency map from the standardized data, and to form an enhanced time-frequency map by combining a joint mask with a spatial attention mechanism. Based on the enhanced time-frequency map, the NetA model and NetB model are optimized by the TextureCrop method, and scanned images are generated and combined. The target time-frequency map is then cropped to generate the target time-frequency map.

[0009] The high-dimensional feature extraction module is used to extract high-dimensional feature vectors from the target time-frequency map using a deep convolutional neural network, and then update the high-dimensional feature vectors using a graph convolutional network.

[0010] The fault classification and prediction module is used to perform fault prediction using a multilayer perceptron based on the updated high-dimensional feature vectors, and uses a metaheuristic optimization algorithm to optimize the hyperparameters to obtain the optimal hyperparameters.

[0011] The fault early warning and judgment module is used to judge the abnormal operating state of the equipment based on the fault prediction results, and to obtain the complete trained model weight file based on the optimal hyperparameters.

[0012] The visualization report generation module is used to perform online enhancement and prediction using the pre-trained complete model weight file, and generate a visualization report by combining the fault prediction judgment results.

[0013] As a preferred embodiment of the intelligent diagnosis and prediction system for mine equipment faults based on deep learning described in this invention, the step of collecting and preprocessing the equipment's operating data to obtain standardized data refers to collecting the equipment's operating data in real time through multiple sensors installed on the mine equipment, and marking the timestamp and equipment number of each collection point; using Discrete Wavelet Transform (DWT) to denoise the collected vibration signals to remove background noise; using a low-pass filter to remove high-frequency noise; and resampling by calculating the equipment's rotation frequency to obtain the denoised and standardized signal and data with a unified sampling frequency.

[0014] As a preferred embodiment of the intelligent diagnosis and prediction system for mine equipment faults based on deep learning described in this invention, the following steps are taken: A preliminary time-frequency map is generated based on standardized data, and an enhanced time-frequency map is formed by combining a joint mask with a spatial attention mechanism. Based on the enhanced time-frequency map, the NetA and NetB models are optimized using the TextureCrop method to generate scanned images, which are then combined. The target time-frequency map is generated by cropping the image. This involves performing continuous wavelet transform on the denoised vibration signal to convert the signal from the time domain to the time-frequency domain. Multi-scale wavelet transform is used to decompose the signal, obtaining preliminary time-frequency maps of the signal in different frequency bands and time windows. A final spatial attention map is generated based on the preliminary time-frequency map. The theoretical fault characteristic frequency is calculated. Based on the calculated theoretical fault characteristic frequency, a joint mask is formed. The joint mask is then weighted and fused with the spatial attention map to generate the enhanced time-frequency map.

[0015] Construct the NetA model network architecture, obtain the global texture energy map E through the TextureCrop method, obtain the attention mask based on the global texture energy map E, multiply the loss map with the attention mask to obtain the weighted loss value, and calculate the total loss function;

[0016] The NetB network architecture is constructed. A pre-trained lightweight CNN model is used as input to historical high-resolution B-scan images. The TextureCrop method is applied to the B-scan images to obtain the texture energy values ​​of the patches. The texture energy value of each patch is calculated to generate the importance score of the patch. The importance score is used as the supervision label to train the lightweight CNN model. The high-resolution B-scan images to be processed are also divided into non-overlapping patches of fixed size. The pre-trained lightweight CNN model outputs the importance score of each patch. The patches are sorted according to the importance score. The top-K high-scoring patches are selected and fed into the NetB model. The enhanced patches are re-stitched with the unselected low-scoring patches according to their spatial positions in the original image to form a complete enhanced B-scan image. The total loss function is calculated. The Adam optimizer is used to backpropagate the total loss function to automatically update all network parameters of the NetB model. The new low-quality B-scan image is input into the converged NetB model. The NetB model outputs the enhanced B-scan image.

[0017] Using B-scan images, defect areas are identified using a defect detection algorithm and correlated with the time period of A-scan images. The depth information of B-scan images is then mapped back to the time point of A-scan images. Time-frequency transformation is performed on A-scan images within the corresponding time period to generate a time-frequency map of the time period. The time-frequency map is then cropped to generate the target time-frequency map.

[0018] As a preferred embodiment of the intelligent diagnosis and prediction system for mine equipment faults based on deep learning described in this invention, the step of obtaining high-dimensional feature vectors by using a deep convolutional neural network to extract features from the target time-frequency map refers to using a deep convolutional neural network for feature extraction, including applying a convolution operation to the input target time-frequency map to obtain local and global feature representations, and then using a pooling layer after the convolutional layer to reduce the dimensionality and obtain the high-dimensional features extracted by the deep convolutional neural network.

[0019] As a preferred embodiment of the intelligent diagnosis and prediction system for mine equipment faults based on deep learning described in this invention, the step of using a graph convolutional network to update high-dimensional feature vectors refers to constructing a graph structure, with each device as a node in the graph. The node features include the sensor data and status information of the device. The relationships between devices are represented by edges in the graph. The graph convolutional network is used to model the relationships between devices. The graph convolutional network transmits node information and updates the features of nodes through multiple convolutional operations.

[0020] As a preferred embodiment of the intelligent fault diagnosis and prediction system for mining equipment based on deep learning described in this invention, the step of using a multilayer perceptron (MLP) for fault prediction based on updated high-dimensional feature vectors refers to constructing an MLP and combining it with a GCN model to form an end-to-end model, calculating the cross-entropy loss between the predicted probability output by the MLP and the true label, and then using the backpropagation algorithm to update all learnable parameters of the GCN and MLP models. During the training process, the Adam optimizer is used to obtain the trained model, and the node features output by the graph convolutional network are combined with the trained MLP to perform fault prediction, thereby obtaining the probability distribution vector.

[0021] As a preferred embodiment of the intelligent diagnosis and prediction system for mine equipment faults based on deep learning described in this invention, the step of using a metaheuristic optimization algorithm to optimize hyperparameters and obtain the optimal hyperparameters refers to using a metaheuristic optimization algorithm to optimize hyperparameters, calculating the updated position of each individual, mapping the new position back to the legal hyperparameter space, and obtaining the optimal hyperparameters.

[0022] As a preferred embodiment of the intelligent diagnosis and prediction system for mine equipment faults based on deep learning described in this invention, the judgment of abnormal operating conditions of equipment based on fault prediction results refers to the maximum value in the fault prediction results output by the multilayer sensor, i.e., the confidence level of the model on the prediction results. If the confidence level is greater than a preset threshold τ and the prediction category is not "normal", an early warning is triggered. The early warning generation includes generating a structured early warning message based on the prediction category.

[0023] As a preferred embodiment of the intelligent diagnosis and prediction system for mine equipment faults based on deep learning described in this invention, the step of obtaining a trained complete model weight file by training the model based on optimal hyperparameters refers to using a labeled equipment dataset, employing optimal hyperparameters and a NetB model and a deep convolutional neural network structure to construct a model including: a NetB model, a GNN model, and a multilayer perceptron. Using the Adam optimizer, the accuracy and F1-score are calculated on the validation set every epoch. If the validation loss does not decrease continuously within a set number of epochs, early stopping is triggered. The weights of the GNN model and the multilayer perceptron (MLP) with the best performance on the validation set, as well as the weights of the NetA / NetB model and the IRV2 feature extractor, are saved to obtain a trained complete model weight file.

[0024] As a preferred embodiment of the deep learning-based intelligent diagnosis and prediction system for mine equipment faults described in this invention, the following steps are taken: Online enhancement and prediction using a pre-trained complete model weight file, combined with the fault prediction results to generate a visual report, refer to loading the pre-trained complete model weight file and deploying it to the inference engine. Multimodal feature fusion and graph construction are then performed again to obtain new node features, which are then fed back into the updated MLP model to obtain the fault prediction probability. When the confidence level maxes out... Only when the predicted category is not "normal" will the corresponding fault area of ​​the device be highlighted in the visualization report and the MLP prediction result be overlaid; otherwise, the status diagram will only be displayed in the normal mode.

[0025] The beneficial effects of this invention are as follows: By employing enhanced time-frequency maps and combining physical priors (such as fault feature frequencies) with data-driven spatial attention maps, key features related to faults can be extracted more accurately, improving the feature representation capability of the time-frequency maps and enhancing the accuracy of fault diagnosis. The Time-Frequency Map is optimized using the TextureCrop method, with the NetA and NetB models focusing on the reconstruction of key regions during training, improving the fine-grained enhancement capability of low-quality time-frequency maps and thus increasing the accuracy of fault prediction. The joint model combining Graph Convolutional Networks (GCN) and Multilayer Perceptrons (MLP) achieves deep fusion and spatiotemporal correlation modeling based on multi-sensor data, solving the problem of insufficient multimodal data processing in traditional methods. A more comprehensive diagnosis is achieved by integrating multidimensional information from the equipment, reducing false alarm and false negative rates. Metaheuristic optimization algorithms are used for hyperparameter optimization to improve the model's generalization ability, ensuring adaptability and robustness in different mining environments, providing a more efficient and accurate solution for intelligent fault diagnosis of mining equipment. Attached Figure Description

[0026] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0027] Figure 1 This is a flowchart of a deep learning-based intelligent diagnosis and prediction system for mine equipment faults, as shown in Example 1.

[0028] Figure 2 This is a structural diagram of a deep learning-based intelligent diagnosis and prediction system for mine equipment faults in Example 1. Detailed Implementation

[0029] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0030] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0031] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0032] Example 1, referring to Figure 1 and Figure 2 This is the first embodiment of the present invention, which provides a deep learning-based intelligent fault diagnosis and prediction system for mining equipment, comprising the following steps:

[0033] S1, Data Acquisition and Preprocessing Module, is used to collect the equipment's operating data and preprocess it to obtain standardized data;

[0034] The time-frequency map generation module is used to generate a preliminary time-frequency map from the standardized data, and to form an enhanced time-frequency map by combining a joint mask with a spatial attention mechanism. Based on the enhanced time-frequency map, the NetA model and NetB model are optimized by the TextureCrop method, and scanned images are generated and combined. The target time-frequency map is then cropped to generate the target time-frequency map.

[0035] Specifically, the collection of equipment operation data and the preprocessing to obtain standardized data refers to the real-time collection of equipment operation data through various sensors (such as vibration, temperature, pressure, sound, etc.) installed on the mining equipment. The collected data includes vibration signals, temperature sensor data, and pressure data, and each collection point should be marked with a timestamp and equipment number. Discrete wavelet transform (DWT) is used to denoise the collected vibration signals to remove background noise and improve signal quality. A low-pass filter is used to remove high-frequency noise and retain effective vibration signals. The data from each sensor are standardized to unify the outputs of different sensors to the same dimension. The mean-variance standardization method is used to adjust the mean of each feature to 0 and the standard deviation to 1.

[0036] By calculating the rotational frequency of the equipment (conveyor drive motor, elevator main shaft) Resampling is performed, where n is the device's rotational speed, and 60 represents converting the device speed n from "revolutions per minute (RPM)" to "revolutions per second (Hz)". 1 minute = 60 seconds. The virtual resampling frequency is calculated based on the device's rotational speed. , where q represents the desired number of sampling points in each rotation cycle (i.e., the number of sampling points per revolution), ensuring that each data point obtains a denoised and standardized signal at the same sampling frequency, and unifying the sampling frequency of the data.

[0037] By introducing multi-sensor data acquisition and preprocessing technology, this method addresses the problem that single-sensor data cannot fully reflect the equipment status in traditional methods. Discrete wavelet transform (DWT) is used to denoise the vibration signal, removing background noise, improving signal quality, and ensuring data accuracy. A low-pass filter is used to remove high-frequency noise, further optimizing the effective portion of the vibration signal and ensuring higher-quality input data. In terms of standardization, mean-variance standardization is employed to unify data from different sensors to the same dimension, eliminating data inconsistencies caused by different sensor types. Resampling based on the equipment rotation frequency ensures a uniform sampling frequency for each data point, resolving the sampling inconsistency problem caused by different equipment rotation speeds in traditional methods, thus improving data quality and consistency.

[0038] Furthermore, the process involves generating a preliminary time-frequency map based on the standardized data, forming an enhanced time-frequency map by combining a joint mask with a spatial attention mechanism, optimizing the NetA and NetB models using the TextureCrop method based on the enhanced time-frequency map, generating scanned images and combining them, and cropping to generate the target time-frequency map. This process involves performing a continuous wavelet transform (CWT) on the denoised vibration signal to convert the signal from the time domain to the time-frequency domain, and using multi-scale wavelet transforms (such as Morlet wavelets) to decompose the signal to obtain preliminary time-frequency maps of the signal in different frequency bands and time windows.

[0039] Based on the initial time-frequency map, a single-layer convolution operation is performed using a fixed-size (e.g., 3×3) convolution kernel, with each pixel position... Extracting local feature vectors from the center , local feature vectors of all locations The input spatial attention mechanism (SAM) module includes a learnable weight vector W obtained from the training dataset through backpropagation and optimization algorithms (such as gradient descent) for each local feature vector. Calculate the inner product ( The attention score is obtained, and the scores of all positions on the entire time-frequency map are concatenated into a two-dimensional response map. Global normalization, such as using the Softmax function, is then applied to the scores of all positions on the entire map to ensure that the weight sum is 1, generating the final spatial attention map. ,in Indicates the position in the time-frequency graph, where It is a time coordinate. These are frequency coordinates;

[0040] Calculate the theoretical fault characteristic frequencies, including the inner race fault frequency and the outer race fault frequency:

[0041] The inner ring failure frequency for:

[0042] ,

[0043] Where N is the number of balls, d is the ball diameter, D is the pitch diameter, and α is the contact angle. To convert RPM (revolutions per minute) to Hz (revolutions per second), It is a constant factor in the geometric relationship between the number of balls and the pitch diameter, where n is the rotational speed;

[0044] The outer ring failure frequency for

[0045] ,

[0046] Based on the calculated inner ring fault frequency and outer ring failure frequency On the frequency axis of the time-frequency plot (such as the result of short-time Fourier transform or wavelet transform), a frequency band window (e.g., a Gaussian window or rectangular window with a width of ±5Hz) is set with each fault frequency as the center. At all time-frequency points (t,f) covered by the window, the mask value is set to 1 (or a high response value), and the remaining area is set to 0 (or a low attenuation value such as 0.2). If multiple faults are considered simultaneously, multiple windows are superimposed to form a joint mask. ;

[0047] Joint mask Spatial attention map Perform weighted fusion to generate an enhanced time-frequency map. The formula is:

[0048] ,

[0049] Wherein, λ is the fusion weight hyperparameter λ∈[0,1], which is used to balance the contribution of data-driven attention and physical prior. By performing a grid search on the validation set for preset candidate values ​​(such as λ∈{0.2,0.5,0.8}), the weight hyperparameter that makes the fault diagnosis accuracy or clustering index (such as NMI) the highest is selected.

[0050] The NetA model (A-scan enhancement) network architecture is constructed, including a U-Net-style encoder-decoder structure. The encoder consists of 4 layers of convolution + batch normalization + ReLU with progressive downsampling. The decoder consists of 4 layers of transposed convolution + skip connections (from the corresponding layer of the encoder). The output layer is Sigmoid activation to generate an enhanced A-scan time-frequency map (two-dimensional image).

[0051] The enhanced time-frequency map is processed using the TextureCrop method with a set of Gabor filters of different directions and scales (e.g., 4 directions × 3 scales, a total of 12 filters). Perform convolution, take the absolute value of each filter output and square it to obtain the local energy response, and sum the energy responses of all filters to obtain the global texture energy map E;

[0052] The attention mask is obtained by applying the normalized Softmax function to the global texture energy map E. The loss map is close to 1 in high-texture regions and close to 0 in low-texture regions. When training the NetA model, the loss map is multiplied by the attention mask to obtain the weighted loss value. The TextureCrop method guides the NetA model to focus more on improving the reconstruction accuracy of key regions (such as information-rich time-frequency locations like fault frequency bands) during the training phase. The formula is as follows:

[0053] ,

[0054] in, Indicates the total number of pixels. This represents the pixel value at position (i,j) of the enhanced A-scan time-frequency plot output by the NetA model (the pixel represents the enhanced signal strength at the corresponding time-frequency point). This indicates the location of a true high-quality A-scan time-frequency diagram. The pixel value (ideal value) is obtained by using a high-precision sensor (such as an ultrasonic probe or laser rangefinder) to detect the target equipment (such as a conveyor, hoist or coal mining machine in a coal mine), and by collecting the original echo signal and performing preprocessing operations on the received signal to obtain the actual equipment.

[0055] Calculate the total loss function :

[0056] ,

[0057] in, These are the weighting coefficients;

[0058] Total loss function All learnable parameters (including convolutional kernel weights in the encoder and decoder) are passed to the NetA model through the backpropagation algorithm. The Adam optimizer iteratively updates the network parameters based on the gradient of the loss. After training, the new low-quality A-scan time-frequency map is input into the converged NetA model. The model forward inference directly outputs the enhanced high-fidelity time-frequency map, which is the final enhanced A-scan image.

[0059] The NetB network architecture is constructed, including the use of the Swing Transformer structure, multi-level residual blocks to extract local and global features, the introduction of channel / spatial attention modules to enhance the response of key regions, and the output of high-resolution B-scan images;

[0060] The device uses high-resolution real B-scan images acquired by a high-precision imaging system (such as a high-sampling-rate ultrasound device or high-quality images calibrated by experts) under the same operating conditions as tags. ;

[0061] Using a pre-trained lightweight CNN model with historical high-resolution B-scan images as input, the TextureCrop method is applied to the B-scan images. This includes convolving each B-scan patch with multi-directional, multi-scale Gabor filters, summing the squared absolute values ​​of the responses (or calculating the variance) to obtain the texture energy value of the patch. Based on the responses of all filters, the texture energy value of each patch is calculated and normalized to generate an importance score for the patch. The importance score is used as the supervision label, and the mean squared error loss function is used for optimization to train the lightweight CNN model.

[0062] The high-resolution B-scan image to be processed is also divided into non-overlapping patches of fixed size. Using a pre-trained lightweight CNN model, all patches of the current output high-resolution B-scan image are inferred separately, and the importance score of each patch is output. The patches are sorted according to the importance score, and the top-K (e.g., the top 20%) high-scoring patches are selected and fed into the backbone network of the NetB model for fine enhancement. This includes extracting and fusing local and global features layer by layer through multi-level residual blocks in the backbone network of the NetB model, using internal channel / spatial attention mechanisms (e.g., CBAM) to enhance key texture and structural information, and outputting enhanced patches of the same size but with clearer details, lower noise, and higher resolution after upsampling or reconstruction layers. The remaining low-scoring patches are skipped or only lightly processed.

[0063] The enhanced tiles will be reassembled with the unselected low-resolution tiles according to their spatial positions in the original image to form a complete enhanced B-scan image. Calculate the total loss function The formula is:

[0064]

[0065] in, and The model is obtained by performing a grid search on the validation set. Specifically, based on the training dataset, combinations of preset candidate values ​​(e.g., β∈{0.1,0.5,1.0}, γ∈{0.1,0.3,0.5}) are tried. For each set of parameters, the PSNR, SSIM, and fault diagnosis accuracy metrics of the model on the validation set are calculated, and the optimal combination is selected. and The value is used as the final hyperparameter. To measure the loss term in terms of structural similarity between the augmented image and the real label, The mathematical symbol for gradient;

[0066] Using the Adam optimizer to optimize the total loss function Backpropagation is performed to automatically update all network parameters of the NetB model (including convolutional layer weights and attention module parameters) to output the scanned image. Gradually approaching the label After training is completed, during the inference phase, the new low-quality B-scan image is input into the converged NetB model, and the model directly outputs a high-resolution, denoised and structurally clear enhanced B-scan image through forward computation.

[0067] By using B-scan images, defect regions are identified using defect detection algorithms (such as edge detection and outlier detection) to determine the location or extent of defects. The depth range corresponding to the defect location in the B-scan image is determined. B-scan images are formed by scanning vibration echo signals in space, and the depth axis directly reflects the propagation distance of the signal in the medium. Enhanced A-scan images are two-dimensional time-frequency representations obtained by time-frequency transformation of echo signals from the same physical source and enhancement using the NetA model. The horizontal axis represents the signal propagation time. According to the physical relationship "depth = speed of sound × time / 2", each depth position in the B-scan image uniquely corresponds to a time point in the enhanced A-scan image, establishing a strict one-to-one mapping between the depth axis and the time axis. Based on this mapping, the depth range of the defect in the B-scan image is converted into the corresponding time period in the enhanced A-scan image, and the time-frequency content of the region is cropped, retaining only the time-frequency part related to the defect location, generating a target time-frequency map aligned with the defect region in the B-scan image.

[0068] By introducing spatial attention mechanism, joint masking, and multi-scale wavelet transform, the accuracy and robustness of mine equipment fault diagnosis are improved. Continuous wavelet transform (CWT) is used to convert the signal from the time domain to the time-frequency domain, and multi-scale wavelets (such as Morlet wavelets) are combined to decompose the signal, capturing the frequency characteristics of equipment faults more meticulously. This improves the feature extraction capability of the signal in different frequency bands and time windows, thus solving the shortcomings of traditional methods in time-frequency map optimization. The combination of spatial attention mechanism (SAM) and joint masking more accurately focuses on key areas in the time-frequency map, enhancing the representation of fault features and avoiding interference from background noise. The combined optimization of etA and NetB models utilizes Gabor filters for texture energy extraction and enhancement, addressing the shortcomings of traditional methods in processing complex time-frequency images. By finely enhancing time-frequency images, the system improves the accuracy of fault area identification and enhances its adaptability to complex fault modes. The A-scan and B-scan image technologies, combined with high-precision sensors, enable precise alignment of fault location and depth, further improving the accuracy of multimodal data fusion. This enhances the performance of the intelligent fault diagnosis and prediction system for mine equipment, enabling real-time and accurate prediction of equipment faults in dynamic real-world environments, thereby improving mine safety and operational efficiency.

[0069] S2, High-dimensional feature extraction module, is used to extract high-dimensional feature vectors from the target time-frequency map using a deep convolutional neural network, and then update the high-dimensional feature vectors using a graph convolutional network.

[0070] Specifically, the step of using a deep convolutional neural network to extract high-dimensional feature vectors based on the target time-frequency map refers to using a deep convolutional neural network (specifically, the Inception-ResNet-v2 model) for feature extraction. This includes applying convolution operations to the input target time-frequency map to obtain local and global feature representations. The convolution formula is:

[0071] ,

[0072] in, For the first Feature map of the layer Let the convolution kernel be from channel a to channel b. For bias terms, For activation functions, such as ReLU, This represents the total number of channels a.

[0073] After the convolutional layers, pooling layers are used for dimensionality reduction to decrease computation and obtain the high-dimensional features extracted by the deep convolutional neural network. The pooling operation formula is:

[0074] ,

[0075] in, This indicates a pooling operation. These are learnable scalar parameters, automatically learned during model training through backpropagation and gradient descent optimization algorithms. The scalar parameters are initialized to 1 or random small values.

[0076] By employing the Inception-ResNet-v2 (IRV2) model for feature extraction, this invention addresses the shortcomings of traditional methods in extracting features from time-frequency maps and capturing both local and global features. Deep convolutional neural networks (DCNNs) extract rich local and global feature representations from the input target time-frequency map through multiple convolutional operations, enhancing the ability to identify complex signal patterns in fault diagnosis. Pooling operations after the convolutional layers further reduce computational load, and by learning optimizable scalar parameters, the impact of the pooling process is automatically adjusted during training, improving the accuracy and efficiency of feature extraction. The deep structure and residual learning mechanism of DCNNs avoid overfitting or information loss problems caused by shallow networks in traditional methods, enabling this invention to maintain high feature extraction capabilities in complex environments and enhancing the accuracy and robustness of mine equipment fault prediction.

[0077] Furthermore, the method of using graph convolutional networks to update high-dimensional feature vectors refers to constructing a graph structure, where each device is treated as a node in the graph. Node features include sensor data and status information of the device (such as vibration, temperature, etc.), and the relationships between devices (such as physical connections and electrical connections) are represented by edges in the graph.

[0078] Graph Convolutional Networks (GCNs) are used to model the relationships between devices. Through multiple layers of convolutional operations, GCNs transmit node information and update node features, ensuring that each device node's features reflect its own state and its spatiotemporal relationships with other device nodes. The formula is as follows:

[0079] ,

[0080] in, It is the update feature of node i. It is a normalization constant. It is a weight matrix. This indicates that in the k-th layer of graph convolution, the feature vector of node j is obtained after the graph convolution operation in the previous layer (or the initial input). It is the activation function, usually ReLU or Tanh, and N(i) is the set of neighboring nodes of node i, that is, all other device nodes that have direct connections (edges) with node i, reflecting the physical or logical associations between devices (such as electrical connections, shared drive shafts).

[0081] By introducing Graph Convolutional Networks (GCNs), the limitations of traditional methods in handling complex relationships between devices are addressed. Traditional equipment fault diagnosis methods often treat devices as independent entities, ignoring the interrelationships between them (such as physical and electrical connections), resulting in the inability to fully utilize the spatiotemporal correlation information between devices. Through the multi-layer convolutional operations of GCNs, node features are updated, allowing the state of each device node to more comprehensively reflect the spatiotemporal correlation information between the device and its surrounding environment and other devices. This not only enhances the accuracy of equipment fault diagnosis and improves the fault mode recognition capability of complex equipment networks, but also allows GCNs to adaptively adjust the weights of each device node, thereby improving the model's generalization ability and reducing misdiagnosis caused by complex dependencies between devices.

[0082] S3, Fault Classification and Prediction Module, is used to perform fault prediction using a multilayer perceptron based on the updated high-dimensional feature vector, and to optimize the hyperparameters using a metaheuristic optimization algorithm to obtain the optimal hyperparameters;

[0083] Specifically, the fault prediction using a multilayer perceptron based on updated high-dimensional feature vectors refers to using historical equipment operation data with real fault labels (where each sample contains node feature vectors extracted by a graph convolutional network and corresponding one-hot encoded fault category labels), constructing a multilayer perceptron (MLP) and jointly forming an end-to-end model with a GCN model, calculating the cross-entropy loss between the predicted probability output by the MLP and the real label, and then using the backpropagation algorithm to update all learnable parameters of the GCN and MLP models. During training, the Adam optimizer is used, combined with an early stopping mechanism and a learning rate decay strategy to improve generalization performance and prevent overfitting, resulting in a trained model.

[0084] Combining node features output by graph convolutional networks Fault prediction is performed using a trained multilayer perceptron (MLP), yielding a probability distribution vector. In other words, each element in the vector represents the probability that a specific type of fault (such as inner ring wear of a bearing, broken gear teeth, motor imbalance, etc.) will occur in the device within a certain period of time (such as the next maintenance cycle or the next 24 hours). For example, an output of [0.1, 0.05, 0.82] can be interpreted as: there is an 82% probability that an "outer ring fault" will occur in the near future. The probability distribution vector formula is:

[0085] ,

[0086] in, It is a multilayer perceptron.

[0087] By combining an end-to-end model of Graph Convolutional Network (GCN) and Multilayer Perceptron (MLP), the accuracy and generalization ability of fault prediction are improved. Existing fault diagnosis methods usually rely on simple feature extraction or traditional classification models, which are difficult to fully capture the complex spatiotemporal dependencies between devices. By introducing the GCN model, the complex relationships between devices can be effectively handled, and high-dimensional feature vectors of devices can be extracted through graph convolution operations, which can more comprehensively reflect the device status. Combined with multilayer perceptron (MLP) for fault prediction, the MLP outputs the probability distribution of different fault types of devices by processing these high-dimensional feature vectors. Based on the historical operating data of the devices, more detailed and accurate fault prediction is performed, which improves the system's ability to identify device faults, especially when multiple fault types exist at the same time, and can provide higher prediction accuracy.

[0088] Furthermore, the use of a metaheuristic optimization algorithm for hyperparameter optimization to obtain the optimal hyperparameters refers to using a metaheuristic optimization algorithm (SFO) for hyperparameter optimization, including determining the range of hyperparameters to be optimized, such as the convolution kernel size and the number of filters, initializing the population, and defining the search space (e.g., randomly generating N=30 sets of hyperparameter configurations, each set being regarded as a "Sailfish individual", denoted as...). (i=1,…,30)), for each individual Train the NetB model with hyperparameter configuration (with only 1 epoch or small sample validation), and calculate the accuracy as the fitness value on the validation set. Find the current optimal individual The formula for updating the position of each individual is:

[0089] ,

[0090] in, This represents the new position (i.e., the new hyperparameter combination) of the i-th Sailfish individual in the next iteration. It is the position of the individual with the highest fitness in the current group (the optimal solution). It is a randomly selected suboptimal solution's position (i.e., a candidate solution with low fitness but still in the population). It is a coefficient that decays with the number of iterations, calculated according to a preset rule: it is usually defined as... ,in The initial step size is 0.1, iter is the current iteration number, and max iter is the maximum number of iterations (e.g., 50). For the current position of the individual, The weight parameters are obtained by performing grid search or metaheuristic optimization (such as SFO, Bayesian optimization) on the validation set.

[0091] new location Mapping back to the valid hyperparameter space (e.g., rounding to integers), reaching the set number of training iterations (e.g., 20 generations), yields the optimal hyperparameters. .

[0092] By using the SFO algorithm to intelligently search within a broader hyperparameter space, the model automatically adjusts hyperparameters such as convolution kernel size and the number of filters, enabling it to adapt to different equipment and fault modes, thereby improving the accuracy of fault prediction. The SFO algorithm initializes a population and optimizes the hyperparameter combination in each iteration using fitness values ​​(such as accuracy on the validation set), ensuring that the selection of hyperparameters is both exploratory and targeted. By dynamically updating the position of each individual and gradually narrowing the search range, the algorithm effectively avoids local optima while reducing computational resource consumption and approaches the global optimum. Compared with traditional methods, the introduction of the SFO algorithm improves the efficiency of model hyperparameter optimization, enabling the model to automatically adjust and obtain optimal performance when facing complex and variable equipment operating data, significantly improving the accuracy and reliability of mine equipment fault prediction. By optimizing hyperparameters through metaheuristic optimization algorithms, the algorithm solves the hyperparameter selection problem existing in traditional methods, improving the intelligence level and adaptability of the fault prediction system.

[0093] S4, Fault Early Warning and Judgment Module, is used to judge the abnormal operating state of the equipment based on the fault prediction results, and to obtain the trained complete model weight file based on the optimal hyperparameters;

[0094] The visualization report generation module is used to perform online enhancement and prediction using the pre-trained complete model weight file and combine the fault prediction judgment results to generate a visualization report.

[0095] Specifically, judging the abnormal operating state of equipment based on fault prediction results refers to using the probability distribution vector output by a multilayer perceptron (MLP). The maximum value in the value is the model's confidence in the prediction result. The accuracy, false alarm rate, and false negative rate of the warning are calculated under different candidate thresholds (e.g., from 0.7 to 0.95, with a step size of 0.01). Combined with business requirements (e.g., tolerating low false alarms but accepting a small number of false negatives), the threshold that optimizes the F1-score or cost-sensitive indicator is selected as the final threshold τ. If the confidence > threshold τ and the prediction category is not "normal", a warning is triggered. Warning generation includes generating a structured warning message based on the prediction category, such as: "Equipment #123: High probability of outer ring failure (confidence 0.92), maintenance is recommended within 72 hours".

[0096] By leveraging the fault prediction results and confidence values ​​of fault category probability distributions based on the output of a multilayer perceptron (MLP), this invention addresses the insufficient accuracy of fault warnings in traditional methods. Traditional fault warning systems often rely solely on fixed thresholds to determine equipment status, easily leading to false alarms or missed alarms, and cannot be finely adjusted according to actual needs. This invention calculates the warning accuracy, false alarm rate, and missed alarm rate under different candidate thresholds, and combines this with business requirements (such as a low false alarm rate or an acceptable missed alarm rate) to adaptively select the optimal warning threshold. By combining confidence values ​​with dynamic thresholds, it achieves both accuracy and flexibility in fault prediction. When the confidence exceeds the set threshold and the predicted category is not "normal," a targeted warning is triggered, generating structured warning information to help maintenance personnel detect potential faults early and handle them promptly. This solves the problem of unstable warning effects caused by uniform threshold settings in traditional methods, providing more reliable fault warnings and maintenance decision support for mining equipment.

[0097] Furthermore, the step of obtaining the trained complete model weight file based on optimal hyperparameters refers to using a labeled device dataset, divided into a training set (80%) and a validation set (20%), and employing optimal hyperparameters. The model was constructed using the NetB model and deep convolutional neural network architecture, including: the NetB model according to... Set the number of convolutional layers and filters. Use GraphSAGE or GAT as the GNN model, with K=2 layers and a multilayer perceptron (MLP). Use the Adam optimizer (learning rate determined by hyperparameters). (Decision), batch size = .batch_size: Calculate the accuracy and F1-score on the validation set in each epoch. If the validation loss does not decrease continuously for a set number of epochs (e.g., 5 epochs), trigger early stopping. Save the GNN model and MLP weights with the best performance on the validation set, as well as the NetA / NetB model and IRV2 feature extractor weights, to obtain the complete trained model weight file (.pt or .h5).

[0098] By introducing optimal hyperparameters for model training, this invention addresses the problems of inaccurate hyperparameter settings and unstable model performance in traditional methods. Existing technologies typically rely on fixed hyperparameter configurations, which can easily lead to overfitting or underfitting during training and make it difficult to achieve optimal performance under different equipment and operating conditions. By employing optimal hyperparameters and dynamic adjustment, this invention can automatically select the best hyperparameter combination during model training, improving the accuracy and generalization ability of fault diagnosis. The trained model can better adapt to different equipment and operating environments in practical applications, improving the reliability and practicality of the fault diagnosis model. By saving the model weight file with the best performance on the validation set, the optimal parameter combination during training is ensured, improving the accuracy and operating efficiency of the intelligent fault diagnosis system for mine equipment.

[0099] Furthermore, online enhancement and prediction are performed using the pre-trained complete model weight file, and a visual report is generated based on the fault prediction results. This involves loading the pre-trained complete model weight file (.pt or .h5) and deploying it to the inference engine. Initialization of the NetA, NetB, and GNN models and the multilayer perceptron (MLP) inference module is completed. Real-time data from the device's vibration, ultrasound, and other sensors is received via Kafka or MQTT. The data is segmented using a sliding window (e.g., every 2 seconds), denoised, and standardized, and then input into the loaded NetA and NetB models respectively. Enhanced A-scan time-frequency maps and B-scan images are generated. Multimodal feature fusion and graph construction are performed again to obtain new node features, which are then fed back into the updated MLP model to obtain the fault prediction probability. When the confidence level maxes out... Only when the predicted category is not "normal" will the corresponding fault area of ​​the device be highlighted in the visualization report, and the MLP prediction results will be superimposed (such as marking high-risk areas with red boxes and displaying fault probabilities with color bars). Otherwise, the status diagram will only be displayed in the normal mode, without triggering alarm indicators. The operation and maintenance personnel can intuitively view the current status and risk location of the device through the report, and realize real-time monitoring and dynamic fault adjustment based on the visualization report.

[0100] By using the pre-trained complete model weight file for online enhancement and prediction, this invention addresses the problems of poor real-time performance and inability to dynamically adjust and optimize in traditional fault diagnosis methods. Existing technologies often rely on offline model training, lacking real-time feedback and adjustment for changes in equipment status. This invention deploys NetA, NetB, and GNN models along with a multilayer perceptron (MLP) inference module into the inference engine, and combines this with real-time sensor data to generate enhanced A-scan time-frequency maps and B-scan images, improving image and feature quality and making fault diagnosis results more accurate. Multimodal feature fusion and graph construction methods are used to update node features, which are then re-input into the MLP model for fault prediction. A set threshold is used to determine whether a fault warning is triggered. Only when the fault prediction confidence exceeds the set threshold and the fault category is "abnormal" will the system generate a structured visualization report, intuitively highlighting the fault area. This helps maintenance personnel understand the equipment status and potential fault risks in real time, avoiding the lack of flexibility and lag in traditional methods, and improving the response speed and accuracy of equipment fault diagnosis.

[0101] This embodiment also provides a computer device applicable to a deep learning-based intelligent diagnosis and prediction system for mine equipment faults, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the deep learning-based intelligent diagnosis and prediction system for mine equipment faults as proposed in the above embodiment.

[0102] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0103] This embodiment also provides a storage medium on which a computer program is stored. When executed by a processor, the program implements a deep learning-based intelligent diagnosis and prediction system for mine equipment faults as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0104] In summary, this invention employs enhanced time-frequency maps, combining physical priors (such as fault feature frequencies) with data-driven spatial attention maps to more accurately extract key features related to faults, improving the feature representation capability of the time-frequency maps and enhancing the accuracy of fault diagnosis. The TextureCrop method optimizes the time-frequency maps, with the NetA and NetB models focusing on key region reconstruction during training, improving the fine-grained enhancement capability of low-quality time-frequency maps and thus increasing the accuracy of fault prediction. A joint model combining Graph Convolutional Networks (GCN) and Multilayer Perceptrons (MLP) enables deep fusion and spatiotemporal correlation modeling based on multi-sensor data, addressing the shortcomings of traditional methods in multimodal data processing. By integrating multidimensional information from the equipment, a more comprehensive diagnosis is achieved, reducing false alarm and false negative rates. A metaheuristic optimization algorithm is used for hyperparameter optimization to improve the model's generalization ability, ensuring adaptability and robustness in different mining environments, providing a more efficient and accurate solution for intelligent fault diagnosis of mining equipment.

[0105] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A deep learning-based intelligent fault diagnosis and prediction system for mining equipment, characterized in that: include, The data acquisition and preprocessing module is used to collect the equipment's operating data and preprocess it to obtain standardized data. The time-frequency map generation module generates a preliminary time-frequency map based on the standardized data, and forms an enhanced time-frequency map by combining a joint mask with a spatial attention mechanism. Based on the enhanced time-frequency map, the NetA model and NetB model are optimized by the TextureCrop method, and scanned images are generated and combined. The target time-frequency map is then cropped to generate the target time-frequency map. The process involves optimizing the NetA and NetB models using the TextureCrop method based on the enhanced time-frequency map, generating scanned images and combining them, cropping to generate the target time-frequency map, constructing the NetA model network architecture, obtaining the global texture energy map E using the TextureCrop method, obtaining the attention mask based on the global texture energy map E, multiplying the loss map with the mask to obtain the weighted loss value, and calculating the total loss function. The NetB network architecture is constructed. A pre-trained lightweight CNN model is used to take historical high-resolution B-scan images as input. The TextureCrop method is applied to the B-scan images to obtain the texture energy values ​​of the patches. The texture energy value of each patch is calculated to generate the importance score of the patch. The importance score is used as the supervision label to train the lightweight CNN model. The high-resolution B-scan images to be processed are also divided into non-overlapping patches of fixed size. The pre-trained lightweight CNN model outputs the importance score of each patch. The patches are sorted according to the importance score. The top-K high-scoring patches are selected and fed into the NetB model. The enhanced patches are re-stitched with the unselected low-scoring patches according to their spatial positions in the original image to form a complete enhanced B-scan image. The total loss function is calculated. The Adam optimizer is used to backpropagate the total loss function to automatically update all network parameters of the NetB model. The new low-quality B-scan image is input into the converged NetB model, and the model outputs the enhanced B-scan image. Using B-scan images, defect regions are identified using defect detection algorithms. The enhanced high-fidelity time-frequency map is directly output as the A-scan image through forward inference of the NetA model. The time periods of the B-scan images and A-scan images are associated. Based on the depth information of the B-scan images, the time points of the A-scan images are mapped back to the time points of the A-scan images. The A-scan images are then subjected to time-frequency transformation within the corresponding time periods to generate time-frequency maps for the time periods. The time-frequency maps are then cropped to generate the target time-frequency map. The high-dimensional feature extraction module uses a deep convolutional neural network to extract high-dimensional feature vectors based on the target time-frequency map, and then uses a graph convolutional network to update the high-dimensional feature vectors. The fault classification and prediction module uses a multilayer perceptron to predict faults based on the updated high-dimensional feature vectors and employs a metaheuristic optimization algorithm to optimize hyperparameters and obtain the optimal hyperparameters. The fault early warning and judgment module judges the abnormal operating state of the equipment based on the fault prediction results, and obtains the trained complete model weight file based on the optimal hyperparameters. The visualization report generation module uses the pre-trained complete model weight file for online enhancement and prediction, and combines the fault prediction judgment results to generate a visualization report.

2. The intelligent fault diagnosis and prediction system for mine equipment based on deep learning as described in claim 1, characterized in that: The process of collecting and preprocessing the operating data of the equipment to obtain standardized data refers to collecting the operating data of the equipment in real time through multiple sensors installed on the mining equipment. Each collection point should be marked with a timestamp and equipment number. The collected vibration signal is denoised using discrete wavelet transform to remove background noise, and a low-pass filter is used to remove high-frequency noise. The equipment's rotation frequency is calculated and resampled to obtain the denoised and standardized signal and data with a unified sampling frequency.

3. The intelligent fault diagnosis and prediction system for mine equipment based on deep learning as described in claim 2, characterized in that: The process of generating a preliminary time-frequency map based on standardized data and forming an enhanced time-frequency map by combining a joint mask with a spatial attention mechanism involves performing continuous wavelet transform on the denoised vibration signal to convert the signal from the time domain to the time-frequency domain, decomposing the signal using multi-scale wavelet transform to obtain preliminary time-frequency maps of the signal in different frequency bands and time windows, generating a final spatial attention map based on the preliminary time-frequency map, calculating the theoretical fault characteristic frequency, forming a joint mask based on the calculated theoretical fault characteristic frequency, and weighted fusing the joint mask with the spatial attention map to generate the enhanced time-frequency map.

4. The intelligent fault diagnosis and prediction system for mine equipment based on deep learning as described in claim 3, characterized in that: The process of extracting high-dimensional feature vectors from target time-frequency maps using deep convolutional neural networks refers to the use of deep convolutional neural networks for feature extraction. This includes applying convolution operations to the input target time-frequency map to obtain local and global feature representations. After the convolutional layers, pooling layers are used for dimensionality reduction to obtain high-dimensional features extracted by the deep convolutional neural network.

5. The intelligent fault diagnosis and prediction system for mine equipment based on deep learning as described in claim 4, characterized in that: The use of graph convolutional networks to update high-dimensional feature vectors refers to constructing a graph structure, treating each device as a node in the graph. Node features include the device's sensor data and status information. Relationships between devices are represented by edges in the graph. Graph convolutional networks are used to model the relationships between devices. Through multiple layers of convolutional operations, the graph convolutional network transmits node information and updates node features.

6. The intelligent fault diagnosis and prediction system for mine equipment based on deep learning as described in claim 5, characterized in that: The method of using a multilayer perceptron for fault prediction based on updated high-dimensional feature vectors refers to... By constructing a multilayer perceptron and combining it with a graph convolutional network model to form an end-to-end model, the cross-entropy loss between the predicted probability output by the multilayer perceptron model and the true label is calculated. Then, the backpropagation algorithm is used to update all learnable parameters of the graph convolutional network model and the multilayer perceptron model. During the training process, the Adam optimizer is used to obtain the trained model. Combining the node features output by the graph convolutional network, the trained multilayer perceptron is used to perform fault prediction to obtain the probability distribution vector.

7. The intelligent fault diagnosis and prediction system for mine equipment based on deep learning as described in claim 6, characterized in that: The process of using a metaheuristic optimization algorithm to optimize hyperparameters and obtain the optimal hyperparameters involves calculating the updated position of each individual, mapping the new position back to the valid hyperparameter space, and thus obtaining the optimal hyperparameters.

8. The intelligent fault diagnosis and prediction system for mine equipment based on deep learning as described in claim 7, characterized in that: The judgment of abnormal operating status of equipment based on fault prediction results refers to the maximum value of the fault prediction result output by the multilayer sensor and the probability distribution of the fault category, that is, the confidence of the model in the prediction result. If the confidence is greater than the final threshold and the prediction category is abnormal, an early warning is triggered. The early warning generation includes generating a structured early warning message based on the prediction category.

9. The intelligent fault diagnosis and prediction system for mine equipment based on deep learning as described in claim 8, characterized in that: The process of using the trained complete model weight file for online enhancement and prediction, and combining the fault prediction judgment results to generate a visualization report, refers to loading the trained complete model weight file and deploying it to the inference engine. Multimodal feature fusion and graph construction are then performed again to obtain new node features, which are then fed back into the updated MLP model to obtain the fault prediction probability. Only when the confidence level is greater than the final threshold and the prediction category is abnormal, is the corresponding fault area of ​​the device highlighted in the visualization report, and the MLP prediction results are superimposed. Otherwise, the status diagram is displayed in the normal mode.

Images