Coal conveying belt intelligent inspection system for safe coal conveying

Abstract

The invention discloses an intelligent inspection system for a coal conveying belt for safe coal conveying. The intelligent inspection system comprises a data acquisition module for collecting data from a plurality of data sources; the relay data processing and analyzing module is used for determining and processing the detected abnormal temperature rise data and can transmit the data to equipment where the next sensor is located; the data preprocessing module is used for carrying out standardization processing on the collected data and marking accurate timestamps and spatial position information for each data record; the time-space alignment module is used for accurately aligning the data from different sources processed by the data preprocessing module in the time dimension; identifying data with spatial attributes, and establishing a spatial relationship among different data points; the improved Transform model is used for processing the data processed by the space-time alignment module; and the fault prediction and result display module predicts the fault and visually displays the position of the fault. The method breaks through the traditional single-sensor threshold detection, and achieves the precise positioning of a composite fault.

Description

Technical Field

[0001] The present invention relates to the technical field of safe coal transportation, and in particular to an intelligent inspection system for a coal conveyor belt used for safe coal transportation. Background Art

[0002] In the coal transportation sector, traditional manual inspections suffer from low efficiency, limited coverage, and a high risk of missed inspections. Furthermore, the long distances and multi-operational conditions of coal conveyor belts can easily lead to fires, wire breakage, belt tears, and deviation. Existing monitoring systems often rely on single-point sensors or a single monitoring method, making them incapable of addressing the demands of fusion and analysis of heterogeneous multi-source data in complex environments. Summary of the Invention

[0003] To this end, the present invention provides an intelligent inspection system for a coal conveyor belt for safe coal transportation to solve the problems in the prior art.

[0004] In order to achieve the above object, the present invention provides the following technical solutions:

[0005] An intelligent inspection system for coal conveyor belts for safe coal transportation, comprising a data acquisition module, a data preprocessing module, a spatiotemporal alignment module, an improved Transformer model, a relay data processing and analysis module, and a fault prediction and result display module;

[0006] The data acquisition module collects data from multiple data sources;

[0007] The relay data processing and analysis module is installed on the mobile inspection equipment, the fixed monitoring equipment at the turning point, and the lower belt monitoring equipment; it is used to determine and process the detected abnormal temperature increase data and can transmit the data to the equipment where the next sensor is located;

[0008] The data preprocessing module cleans the collected data to remove noise, outliers, and missing values; standardizes each source data, maps the data to a set numerical range, and annotates each data record with accurate timestamps and spatial location information to ensure subsequent spatiotemporal alignment operations;

[0009] The spatiotemporal alignment module precisely aligns data from different sources processed by the data preprocessing module in the temporal dimension to ensure the consistency of timestamps across all data sources; identifies data with spatial attributes and establishes spatial relationships between different data points;

[0010] The improved Transformer model processes the data after the spatiotemporal alignment module. The improved Transformer model includes an input layer, a multi-head self-attention mechanism layer, a feedforward neural network layer, a residual connection, and an output layer.

[0011] The fault prediction and result display module predicts faults based on the output results of the improved Transformer model and intuitively displays the fault location.

[0012] Furthermore: the data collected by the data acquisition module includes data from various sensors, equipment log data and surrounding environment monitoring data; various sensors are integrated on several mobile inspection devices, fixed monitoring devices at corners and lower belt monitoring devices.

[0013] Furthermore: the mobile inspection equipment is installed above the coal conveyor belt. Since the conveying distance of the coal conveyor belt is very long, each long slope section and horizontal conveying section is divided into several small sections, and each small section is equipped with a reciprocating mobile inspection equipment; a fixed monitoring device is separately installed at each turning point of the coal conveyor belt; and several lower belt monitoring devices are set above the lower belt of the coal conveyor belt to monitor the condition of the empty belt.

[0014] Furthermore: the next sensor re-detects the belt point where the abnormal temperature rise was first detected; when, after comparison, the next sensor still detects an abnormal temperature rise, the device where the next sensor is located transmits the comparison data to the device where the next sensor is located; the next sensor detects the point again, and when it is still confirmed that the temperature has indeed abnormally risen and exceeded the set threshold, the device where the next sensor is located directly determines that the point is already burning with a hidden fire or an open fire, and directly transmits the data to the fault prediction and result display module, and issues a fire warning or alarm; at the same time, the data is transmitted to the data acquisition module;

[0015] When the next comparison is made with the data from the initial sensor of the abnormal point, if it is found that the temperature of the abnormally elevated temperature point is decreasing or the abnormally elevated data at that point disappears, the comparison data will continue to be sent to the next two sensors. After triple comparison, if it is found that the abnormally elevated temperature point has indeed disappeared, the data transmission ends.

[0016] When one of the two subsequent sensors detects an abnormal temperature increase again, the sensor that detected the temperature increase again will be used as the initial sensor of the abnormal point, and the cycle will be restarted, repeating the previous detection and comparison; at the same time, the data will be recorded and transmitted to the fault prediction and result display module as a suspicious fire warning.

[0017] Furthermore: the spatiotemporal alignment module mainly realizes its function through time alignment and space alignment;

[0018] Time alignment: Based on the timestamp information of the data processed by the data preprocessing module, all data are converted to a unified time base. For time series data, resampling technology is used to unify the sampling rate of data with different sampling frequencies. For event-driven data, the data is sorted and aligned according to the order in which events occur. A time alignment matrix is ​​constructed to represent the corresponding relationship between different data sources at various time points.

[0019] Spatial alignment: For data with spatial location attributes, a spatial coordinate system is established and the data of different devices or monitoring points are mapped to this coordinate system; a spatial interpolation algorithm is used to encrypt sparse spatial data to make the data distribution in the entire monitoring area more uniform, facilitating spatial feature extraction and fusion; and the spatial correlation matrix between data is calculated to measure the similarity or correlation between data at different locations, providing a weight basis for subsequent spatial feature fusion.

[0020] Furthermore: the input layer uses the multi-source heterogeneous data after spatiotemporal alignment as the input of the improved Transformer model, and organizes it into a suitable input tensor form according to the feature dimension of the data;

[0021] The multi-head self-attention mechanism layer introduces a multi-head self-attention mechanism, which allows the improved Transformer model to simultaneously focus on different position information and feature information when processing input data. Each head calculates the attention weight separately, and then performs a weighted sum to obtain the output. Finally, the outputs of all heads are spliced ​​together.

[0022] The feedforward neural network layer is connected to the feedforward neural network FFN after the multi-head self-attention mechanism layer. The feedforward neural network FFN consists of two fully connected layers with a ReLU activation function in the middle. The feedforward neural network FFN is used to further perform feature transformation and nonlinear mapping on the output of the self-attention mechanism.

[0023] Residual connections, which add residual connections between the multi-head self-attention mechanism layer and the feedforward neural network layer; residual connections add the input directly to the output;

[0024] The output layer outputs the fault type, confidence, and location coordinates of the fault probability map through a fully connected layer. The number of neurons in the output layer is set according to the number of fault type classifications, the confidence quantization level, and the accuracy requirements of the location coordinates.

[0025] Furthermore: the model training of the improved Transformer model mainly adopts the following method;

[0026] 1) Loss Function Design: A comprehensive loss function is designed, including the cross-entropy loss for fault type prediction, the mean squared error loss for confidence prediction, and the Euclidean loss for location coordinate prediction. By weightedly combining these loss terms, the learning objectives of different tasks are balanced, guiding the model to simultaneously optimize fault diagnosis accuracy, confidence assessment rationality, and location accuracy during training.

[0027] 2) Optimization Algorithm Selection: The Adam optimization algorithm is used to update model parameters. The Adam optimization algorithm combines the advantages of the momentum method and adaptive learning rate. It can automatically adjust the learning rate during training, accelerate convergence, and improve model stability.

[0028] 3) Data partitioning: The preprocessed dataset is divided into training, validation, and test sets according to a certain ratio. The training set accounts for 70%-80% of the total dataset and is used for model training. The validation set accounts for 10%-15% and is used to adjust model hyperparameters and monitor whether the training process is overfitting. The test set accounts for 10%-15% and is used to evaluate the final performance of the model. During the training process, stratified sampling or random sampling methods are used to ensure that the data in each batch can fully represent the distribution characteristics of the entire dataset.

[0029] Furthermore, the temporal and spatial correlation matrices calculated during the spatiotemporal alignment process are used to provide prior knowledge for the multi-head self-attention mechanism, guiding the improved Transformer model to pay more attention to spatiotemporal related data features during feature fusion.

[0030] In the output of the multi-head self-attention mechanism, the feature vectors of different data sources are weighted and summed according to the attention weights to obtain the fused feature representation;

[0031] Features of different scales are processed and fused separately in different layers or modules of the improved Transformer model.

[0032] Furthermore: Fault prediction is based on the fault type probability distribution output by the improved Transformer model, and the most likely fault type is determined using threshold judgment or Top-K selection methods.

[0033] Going a step further: when displaying the prediction results, for each predicted fault point, its location is marked with a different color or symbol on the coal conveyor belt map, and the corresponding fault type and confidence information are displayed. The fault point location is marked on the coal conveyor belt in real time based on the belt transmission speed and displayed.

[0034] The present invention has the following advantages: breaking through the traditional single-sensor threshold detection and realizing the precise positioning of complex faults; effectively integrating multimodal data such as infrared thermal imaging and laser scanning through the spatiotemporal alignment module and the improved Transformer model, and improving the early identification capability of complex faults; the relay data processing module compares abnormal data through multiple sensors step by step, significantly reducing the false alarm rate caused by environmental noise, and reducing the amount of data processed by the data acquisition module.

[0035] Other features and advantages of the present invention will be set forth in the description which follows, and in part will be obvious from the description, or may be learned by practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] To more intuitively illustrate the prior art and the present application, exemplary drawings are provided below. It should be understood that the specific shapes and structures shown in the drawings should not generally be considered as limiting conditions for implementing the present application. For example, based on the technical concepts disclosed in this application and the exemplary drawings, those skilled in the art are capable of easily making routine adjustments or further optimizations to the addition / reduction / attribution division of certain units (components), the specific shapes, positional relationships, connection methods, and dimensional ratios.

[0037] Figure 1 A system block diagram of an intelligent inspection system for coal conveyor belts for safe coal transportation provided in one embodiment of the present application.

[0038] Figure 2 This is a schematic diagram of the position structure of the transmission inspection equipment on the coal conveyor belt in the intelligent inspection system for coal conveyor belt for safe coal transportation of the present invention.

[0039] Figure 3 It is a schematic cross-sectional structural diagram of the installation positions of the mobile inspection equipment and the lower belt fixed monitoring equipment in the intelligent inspection system for coal conveyor belts of the present invention.

[0040] In the picture:

[0041] 1. Coal conveyor belt; 2. Mobile inspection equipment; 3. Fixed monitoring equipment at the turning point; 4. Fixed monitoring equipment for the lower belt; 5. Driving mechanism. DETAILED DESCRIPTION

[0042] The following specific embodiments illustrate the implementation of the present invention. People familiar with this technology can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. It should be understood that these embodiments are only to further illustrate the present invention and cannot be understood as limiting the scope of protection of the present invention. Technical engineers in this field can make some non-essential improvements and adjustments to the present invention based on the content of the above invention; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0043] See also Figure 1-Figure 3 ,An intelligent inspection system for coal conveyor belts for safe coal transportation, including a data acquisition module, a data preprocessing module, a time-space alignment module, an improved Transformer model, a relay data processing and analysis module, and a fault prediction and result display module.

[0044] The data acquisition module collects relevant data from multiple data sources, including but not limited to various sensor data, equipment log data, and surrounding environment monitoring data (such as humidity, electromagnetic interference, etc.).

[0045] Multiple sensors are integrated into several mobile inspection devices 2, fixed monitoring devices 3 at the turning points, and fixed monitoring devices 4 on the lower belt.

[0046] The mobile inspection device 2 is set up above the coal conveyor belt 1. A driving mechanism 5 is provided on both sides of the mobile inspection device 2. The driving mechanism 5 can enable the mobile inspection device 2 to move on the coal conveyor belt 1. Since the conveying distance of the coal conveyor belt 1 is very long, each long slope section and horizontal conveying section is divided into several small sections, and each small section is provided with a reciprocating mobile inspection device 2. A fixed turning monitoring device 3 is separately set up at each turning point of the coal conveyor belt 1. Several lower belt fixed monitoring devices 4 are set above the lower belt of the coal conveyor belt 1 to monitor the condition of the empty belt.

[0047] The sensors on the mobile inspection device 2, the fixed monitoring device 3 at the turning point, and the fixed monitoring device 4 on the lower belt mainly include:

[0048] High frame rate infrared thermal imager (resolution 0.05°C) is used to detect abnormal belt surface temperature;

[0049] Linear array laser scanner (accuracy ±0.1mm) to measure belt deviation in real time;

[0050] Acoustic emission sensor array (frequency range 20kHz-1MHz) to capture the broken wire rope signal inside the belt;

[0051] Multispectral camera (covering visible light + near-infrared bands) can identify surface oil stains, foreign matter and tiny cracks.

[0052] The relay data processing and analysis module is installed on the mobile inspection device 2, the fixed monitoring device 3 at the turning point, and the fixed monitoring device 4 on the lower belt. It is mainly used to determine and process the abnormal temperature increase data detected by the high-frame-rate infrared thermal imager (the initial sensor data at the abnormal point) and transmit the data to the device where the next sensor is located.

[0053] The second sensor re-detects the belt point detected by the initial sensor of the abnormal point; after comparison, if the second sensor still detects an abnormal temperature increase, the device where the second sensor is located will transmit the comparison data to the device where the third sensor is located; the third sensor detects the point again, and when it is still confirmed that the temperature has indeed increased abnormally and exceeded the set threshold, the device where the third sensor is located will directly determine that the point is already burning with a dark flame or an open flame, and transmit the data directly to the fault prediction and result display module, and issue a fire warning or alarm; at the same time, the data will be transmitted to the data acquisition module.

[0054] When the second sensor compares the data of the initial sensor at the abnormal point with that of the abnormal point, and finds that the temperature at the abnormally elevated temperature point is decreasing or the abnormally elevated data at that point disappears, it continues to send the comparison data to the next two sensors. After triple comparison, when it is found that the abnormally elevated temperature point has indeed disappeared, the data transmission ends.

[0055] When one of the two subsequent sensors detects an abnormal temperature increase again, the sensor that detected the temperature increase again will be used as the initial sensor of the abnormal point, and the above process will be repeated; at the same time, the data will be recorded and transmitted to the fault prediction and result display module as a suspicious fire warning.

[0056] For example, if the sensor of mobile inspection device 2 at point A detects an abnormality at point X on the belt, the data from point A is transmitted to the next device at point B (mobile inspection device 2 / fixed monitoring device 3 at the bend / fixed monitoring device 4 on the lower belt). The device at point B then retests point X. If an abnormal temperature rise is detected again, the data from both points A and B are transmitted to the next device at point C, which then retests point X. If a temperature abnormality is confirmed, the device at point C directly transmits the data to the fault prediction and result display module. If no temperature abnormality is detected at point B, the data is reconfirmed at points C and D. Once the temperature is normal, the data transmission ends and the data is recorded. If the temperature detected at point B decreases or the abnormal temperature disappears, and point C detects a temperature increase again, point C is reset to the sensor location where the abnormal temperature was first detected. The previous operation is repeated, and the data from point C is uploaded to the fault prediction and result display module, issuing a suspected fire warning.

[0057] The data preprocessing module cleans the collected data, removes noise, outliers and missing values, and uses appropriate interpolation methods or filling strategies based on statistical laws to handle missing data.

[0058] Since the data dimensions and numerical ranges of different data sources vary greatly, the source data are standardized, such as using Z-score standardization or Min-Max standardization to map the data to a set numerical range to improve the stability and convergence speed of model training.

[0059] In addition, each data record is annotated with an accurate timestamp and spatial location information to ensure subsequent spatiotemporal alignment. For data whose spatial location cannot be directly determined, its relative position can be determined by associating it with the device installation location or other indirect methods. For example, the sensor data at that spatial location can be determined based on the conveyor belt speed, the section where the mobile inspection device 2 is located, and the speed of the mobile inspection device 2 in that section.

[0060] The spatiotemporal alignment module precisely aligns data from different sources (such as equipment operation data and sensor monitoring data) in the time dimension, ensuring that the timestamps of each data source are consistent so that subsequent analysis can be performed based on a unified time reference. For data with spatial attributes (such as sensor data), the spatiotemporal alignment module can identify and establish spatial relationships between different data points.

[0061] The spatiotemporal alignment module mainly realizes its functions through time alignment and space alignment.

[0062] Time alignment: convert all data into a unified time base based on the timestamp information of the data processed by the data preprocessing module; for time series data, resampling technology is used to unify the sampling rate of data with different sampling frequencies; for event-driven data, sort and align according to the order in which events occur.

[0063] A time alignment matrix is ​​constructed to represent the correspondence between different data sources at various time points, so that data at the same time can be accurately combined in subsequent fusion operations.

[0064] Spatial alignment: For data with spatial location attributes, a spatial coordinate system (such as a two-dimensional plane coordinate system or a three-dimensional space coordinate system) is established to map the data of different devices or monitoring points into this coordinate system.

[0065] The spatial interpolation algorithm is used to encrypt the sparse spatial data to make the data distribution in the entire monitoring area more uniform, which is convenient for spatial feature extraction and fusion. The spatial correlation matrix between the data is calculated to measure the similarity or correlation between data at different locations, providing a weight basis for subsequent spatial feature fusion.

[0066] The improved Transformer model processes the data after the spatiotemporal alignment module. The improved Transformer model includes an input layer, a multi-head self-attention mechanism layer, a feedforward neural network layer, a residual connection, and an output layer.

[0067] Input layer: The improved Transformer model uses the temporally and spatially aligned multi-source heterogeneous data as input and organizes it into an appropriate input tensor format based on the data's feature dimensions. For example, for each combination of time point and spatial location, the corresponding multi-source data features are concatenated into a vector as an input element.

[0068] Multi-head self-attention mechanism layer: The introduction of a multi-head self-attention mechanism allows the improved Transformer model to simultaneously focus on different positional information and feature information when processing input data. Each head calculates the attention weight separately, and then performs a weighted sum to obtain the output. Finally, the outputs of all heads are spliced ​​together. Through the multi-head self-attention mechanism, the improved Transformer model can automatically learn complex dependencies between data, including spatiotemporal correlations and feature interactions.

[0069] Feedforward neural network layer: A feedforward neural network (FFN) is connected after the multi-head self-attention mechanism layer. The FFN consists of two fully connected layers with a ReLU activation function in the middle. The FFN is used to further perform feature transformation and nonlinear mapping on the output of the self-attention mechanism to enhance the expressive power of the model.

[0070] Residual connections: To alleviate the vanishing gradient problem and accelerate the training convergence of the improved Transformer model, residual connections are added between the multi-head self-attention mechanism layer and the feedforward neural network layer. Residual connections directly add the input to the output, making it easier for the improved Transformer model to optimize the gradient flow in deep networks.

[0071] Output layer: After the multi-layer stacking of the above structure, a fully connected layer finally outputs the three dimensions of the fault probability map: fault type, confidence level, and location coordinates. The number of neurons in the output layer is set based on the number of fault type classifications, the confidence level quantization, and the required accuracy of the location coordinates. A softmax function is used to predict the probability distribution of the fault type and normalize the confidence level.

[0072] The model training of the improved Transformer model mainly adopts the following methods;

[0073] 1) Loss Function Design: A comprehensive loss function is designed, including the cross-entropy loss for fault type prediction, the mean squared error loss for confidence prediction, and the Euclidean loss for location coordinate prediction. By weightedly combining these loss terms, we balance the learning objectives of different tasks and guide the model to simultaneously optimize fault diagnosis accuracy, confidence assessment rationality, and location accuracy during training.

[0074] 2) Optimization Algorithm Selection: The Adam optimization algorithm is used to update model parameters. Combining the advantages of momentum and adaptive learning rates, the Adam optimization algorithm automatically adjusts the learning rate during training, accelerating convergence and improving model stability. Hyperparameters such as the initial learning rate, learning rate decay strategy, and number of training rounds are set appropriately, and adjustments and optimizations are made based on performance on the validation set.

[0075] 3) Data Partitioning: The preprocessed dataset is divided into training, validation, and test sets according to a specific ratio. Typically, the training set accounts for 70%-80% of the total dataset and is used for model training; the validation set accounts for 10%-15% and is used to adjust model hyperparameters and monitor overfitting during training; and the test set accounts for 10%-15% and is used to evaluate the final model performance. During training, stratified or random sampling methods are used to ensure that each batch of data fully represents the distribution characteristics of the entire dataset.

[0076] In addition, the temporal and spatial correlation matrices calculated during the spatiotemporal alignment process provide prior knowledge for the multi-head self-attention mechanism, guiding the improved Transformer model to place greater emphasis on spatiotemporal-related data features during feature fusion. For example, when calculating attention weights, attention scores can be weighted based on time intervals and spatial distances, giving higher weights to data features at adjacent time points and close spatial locations.

[0077] In the output results of the multi-head self-attention mechanism, the feature vectors of different data sources are weighted and summed according to the attention weights to obtain the fused feature representation; this feature fusion method based on the attention mechanism can dynamically and adaptively fuse according to the intrinsic structure and correlation of the data, highlighting important feature information and suppressing irrelevant noise.

[0078] Considering that multi-source heterogeneous data may have different feature scales (such as time scale, spatial scale and eigenvalue scale), a multi-scale feature fusion strategy is adopted; features of different scales are processed and fused separately in different layers or modules of the improved Transformer model.

[0079] For example, at the bottom layer, short-term features of time series data can be extracted and fused, then at the middle layer, spatial features can be combined for medium-scale fusion, and finally at the high layer, global spatiotemporal information can be comprehensively considered for large-scale feature fusion to generate the final composite fault probability map.

[0080] The fault prediction and result display module predicts faults based on the output of the improved Transformer model. After confidence assessment and calibration, it visually annotates the location coordinates output by the improved Transformer model on a two-dimensional or three-dimensional spatial map. The fault location can be intuitively displayed using geographic information system (GIS) software or professional visualization tool libraries (such as Matplotlib and Plotly).

[0081] Fault prediction is mainly based on the fault type probability distribution output by the improved Transformer model, and the most likely fault type is determined by threshold judgment or Top-K selection method.

[0082] For example, a probability threshold (such as 0.5) is set. When the probability of a certain fault type exceeds the threshold, it is determined that the fault type has occurred; or the top K fault types with the highest probability are selected as the candidate fault set, and further analysis and judgment are performed in combination with domain expert knowledge or actual maintenance experience.

[0083] Confidence assessment and calibration,postprocess the confidence results of the model predictions and use the Platt Scaling calibration method to convert them into confidence estimates that are more in line with the actual situation.

[0084] Platt Scaling fits the relationship between the model prediction probability and the true label on the validation set to obtain a calibration function, which is used to adjust the confidence prediction results on the test set to make it more reliably reflect the uncertainty of the model.

[0085] When displaying the prediction results, for each predicted fault point, its location is marked on the map with a different color or symbol, and the corresponding fault type and confidence information are displayed, so that users can quickly understand the location and possibility of the fault.

[0086] In actual application scenarios, after the model is deployed, its performance changes need to be continuously monitored. As the equipment ages, the operating environment changes, and new failure modes emerge, the performance of the model may gradually decline. Therefore, it is necessary to regularly collect new data and retrain and update the model. By continuously iterating and optimizing the model, it can adapt to the ever-changing actual needs and maintain a high level of fault diagnosis accuracy and reliability.

[0087] This paper uses a spatiotemporal alignment module and an improved Transformer model to effectively integrate multimodal data such as infrared thermal imaging and laser scanning, thereby improving the early identification capability of complex faults (such as hidden fires and wire rope breaks).

[0088] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. An intelligent inspection system for coal conveyor belts for safe coal transportation, characterized in that: include: The data acquisition module collects data from multiple data sources, including data from various sensors, equipment log data, and surrounding environment monitoring data; Various sensors are integrated into several mobile inspection devices, fixed monitoring devices at turns, and lower belt monitoring devices; The relay data processing and analysis module is installed on the mobile inspection equipment, the fixed monitoring equipment at the turning point, and the lower belt monitoring equipment. It is used to determine and process the detected abnormal temperature increase data and can transmit the data to the equipment where the next sensor is located. The relay data processing and analysis module can also transmit the data to the data acquisition module or the fault prediction and result display module according to the data situation; The data preprocessing module cleans the collected data to remove noise, outliers, and missing values; standardizes each source data, maps the data to a set numerical range, and annotates each data record with accurate timestamps and spatial location information; The spatiotemporal alignment module precisely aligns data from different sources processed by the data preprocessing module in the temporal dimension to ensure the consistency of timestamps across all data sources; identifies data with spatial attributes and establishes spatial relationships between different data points; The improved Transformer model processes the data after the spatiotemporal alignment module. The improved Transformer model includes an input layer, a multi-head self-attention mechanism layer, a feedforward neural network layer, a residual connection, and an output layer. The fault prediction and result display module predicts faults based on the output results of the improved Transformer model and intuitively displays the fault location.

2. The intelligent inspection system for coal conveyor belts for safe coal transportation according to claim 1 is characterized in that: The mobile inspection equipment is installed above the coal conveyor belt. Since the conveying distance of the coal conveyor belt is very long, each long slope section and horizontal conveying section is divided into several small sections, and each small section is equipped with a reciprocating mobile inspection equipment; a fixed monitoring device is installed at each turning point of the coal conveyor belt; and several lower belt monitoring devices are set above the lower belt of the coal conveyor belt.

3. The intelligent inspection system for coal conveyor belts for safe coal transportation according to claim 1 is characterized in that: The next sensor re-detects the belt point where the abnormal temperature rise was detected for the first time; when, after comparison, the next sensor still detects an abnormal temperature rise, the device where the next sensor is located transmits the comparison data to the device where the next sensor is located; the next sensor again detects the belt point where the temperature rises abnormally, and when it is still confirmed that the temperature has indeed risen abnormally and exceeds the set threshold, the device where the next sensor is located directly determines that the point is already burning with a hidden flame or an open flame, and transmits the data directly to the fault prediction and result display module, and issues a fire warning or alarm; at the same time, the data is transmitted to the data acquisition module; When the next comparison is made with the data from the initial sensor of the abnormal point, if it is found that the temperature of the abnormally elevated temperature point is decreasing or the abnormally elevated data at that point disappears, the comparison data will continue to be sent to the next two sensors. After triple comparison, when it is confirmed that the abnormally elevated temperature point has disappeared, the data transmission ends. When one of the two subsequent sensors detects an abnormal temperature increase again, the sensor that detected the temperature increase again will be used as the initial sensor of the abnormal point, and the cycle will be restarted, repeating the previous detection and comparison; at the same time, the data will be recorded and transmitted to the fault prediction and result display module as a suspicious fire warning.

4. The intelligent inspection system for coal conveyor belts for safe coal transportation according to claim 1 is characterized in that: The spatiotemporal alignment module mainly realizes its function through time alignment and space alignment; Time alignment: Based on the timestamp information of the data processed by the data preprocessing module, all data are converted to a unified time base. For time series data, resampling technology is used to unify the sampling rate of data with different sampling frequencies. For event-driven data, the data is sorted and aligned according to the order in which the events occur. Construct a time alignment matrix to represent the correspondence between different data sources at different time points; Spatial alignment: For data with spatial location attributes, a spatial coordinate system is established and the data of different devices or monitoring points are mapped to the coordinate system; The spatial interpolation algorithm is used to encrypt the sparse spatial data to make the data distribution in the entire monitoring area more uniform, which is convenient for spatial feature extraction and fusion. The spatial correlation matrix between the data is calculated to measure the similarity or correlation between data at different locations, providing a weight basis for subsequent spatial feature fusion.

5. The intelligent inspection system for coal conveyor belts for safe coal transportation according to claim 1 is characterized in that: The input layer uses the multi-source heterogeneous data after spatiotemporal alignment as the input of the improved Transformer model and organizes it into a suitable input tensor form according to the feature dimension of the data; The multi-head self-attention mechanism layer introduces a multi-head self-attention mechanism, which allows the improved Transformer model to simultaneously focus on different position information and feature information when processing input data. Each head calculates the attention weight separately, and then performs a weighted sum to obtain the output. Finally, the outputs of all heads are spliced ​​together. The feedforward neural network layer is connected to the feedforward neural network FFN after the multi-head self-attention mechanism layer. The feedforward neural network FFN consists of two fully connected layers with a ReLU activation function in the middle. The feedforward neural network FFN is used to further perform feature transformation and nonlinear mapping on the output of the self-attention mechanism. Residual connections, which add residual connections between the multi-head self-attention mechanism layer and the feedforward neural network layer; residual connections add the input directly to the output; The output layer outputs the fault type, confidence, and location coordinates of the fault probability map through a fully connected layer. The number of neurons in the output layer is set according to the number of fault type classifications, the confidence quantization level, and the accuracy requirements of the location coordinates.

6. The intelligent inspection system for coal conveyor belts for safe coal transportation according to claim 1, characterized in that: The model training of the improved Transformer model mainly adopts the following method: 1) Loss Function Design: A comprehensive loss function is designed, including the cross-entropy loss for fault type prediction, the mean squared error loss for confidence prediction, and the Euclidean loss for location coordinate prediction. By weightedly combining these loss terms, the learning objectives of different tasks are balanced, guiding the model to simultaneously optimize fault diagnosis accuracy, confidence assessment rationality, and location accuracy during training. 2) Optimization Algorithm Selection: The Adam optimization algorithm is used to update model parameters. The Adam optimization algorithm combines the advantages of the momentum method and adaptive learning rate. It can automatically adjust the learning rate during training, accelerate convergence, and improve model stability. 3) Data partitioning: The preprocessed dataset is divided into training, validation, and test sets according to a certain ratio. The training set accounts for 70%-80% of the total dataset and is used for model training. The validation set accounts for 10%-15% and is used to adjust model hyperparameters and monitor whether the training process is overfitting. The test set accounts for 10%-15% and is used to evaluate the final performance of the model. During the training process, stratified sampling or random sampling methods are used to ensure that the data in each batch can fully represent the distribution characteristics of the entire dataset.

7. The intelligent inspection system for coal conveyor belts for safe coal transportation according to claim 1, characterized in that: The temporal and spatial correlation matrices calculated during the spatiotemporal alignment process are used to provide prior knowledge for the multi-head self-attention mechanism, guiding the improved Transformer model to pay more attention to spatiotemporal related data features during feature fusion. In the output of the multi-head self-attention mechanism, the feature vectors of different data sources are weighted and summed according to the attention weights to obtain the fused feature representation; Features of different scales are processed and fused separately in different layers or modules of the improved Transformer model.

8. The intelligent inspection system for coal conveyor belts for safe coal transportation according to claim 1, characterized in that: Fault prediction is based on the fault type probability distribution output by the improved Transformer model, and the most likely fault type is determined using threshold judgment or Top-K selection methods.

9. The intelligent inspection system for coal conveyor belts for safe coal transportation according to claim 1, characterized in that: When displaying the prediction results, for each predicted fault point, its location is marked with a different color or symbol on the coal conveyor belt map, and the corresponding fault type and confidence information are displayed. The fault point location is marked on the coal conveyor belt in real time according to the belt transmission speed and displayed.

Images