Fault detection and early warning system for belt conveyor

Abstract

The invention, which relates to the technical field of industrial equipment state monitoring, discloses a belt conveyor fault detection and early warning system comprising a multi-mode sensing module, an edge calculation module, a feature fusion and analysis module, a risk level determination module, an intelligent early warning module and a visual feedback module. Multi-source data such as images, sound, tension, temperature and vibration in the running process of a conveyor are collected, features are extracted and preprocessed through edge calculation, and a structured feature vector is constructed; through deep fusion and fault identification model analysis, operation state identification and trend score generation, and in combination with a dynamic threshold algorithm, a risk level is determined, the method has a hierarchical response mechanism, supports user feedback driven model online learning and optimization, and improves the accuracy and adaptive ability of fault identification; the system has the advantages of being comprehensive in detection, timely in response and high in intelligent degree, and is suitable for various complex conveying environments.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial equipment state monitoring, and particularly relates to a belt conveyor fault detection and early warning system. BACKGROUND

[0002] The belt conveyor is a continuous conveying equipment widely used in coal mines, ports, power, metallurgy and other fields, and has the advantages of large conveying capacity, stable operation and strong adaptability. However, in the long-term operation process, the belt conveyor is prone to various faults such as belt deviation, tearing, slipping, drum jamming and bearing overheating, which not only affects the conveying efficiency, but also may cause safety accidents.

[0003] In the prior art, the fault detection of the belt conveyor mainly relies on manual inspection, and there are problems of response lag, incomplete coverage and high misjudgment rate. Although some automatic detection systems can realize the monitoring of part parameters, they are mostly single index detection, such as monitoring only the belt tension, temperature or speed, and cannot realize multi-parameter fusion analysis, lack the comprehensive judgment ability of fault trend, and it is difficult to issue early warning information in time. SUMMARY

[0004] The purpose of the present application is to provide a belt conveyor fault detection and early warning system to solve the problems in the background art.

[0005] In order to achieve the above purpose, the present application provides the following technical scheme: a belt conveyor fault detection and early warning system, comprising:

[0006] A multi-modal perception module is used to collect image, sound, temperature, tension and vibration data generated during the operation of the conveyor;

[0007] An edge computing module is used to extract and preprocess the collected data, and encapsulate the extracted feature data into a structured feature vector;

[0008] A feature fusion and analysis module is used to receive the structured feature vector, and based on a deep fusion model, to weight and combine the multi-modal features to construct a comprehensive feature representation; input the comprehensive feature into a pre-trained multi-fault classification model to identify whether the current conveyor operation state is abnormal and the fault type;

[0009] A risk level determination module is used to calculate a fault trend score based on time series features, and determine a risk level in combination with a dynamic threshold algorithm;

[0010] An intelligent early warning module is used to generate early warning information when the risk level exceeds a preset threshold, and adopt a hierarchical response strategy according to the risk level;

[0011] A visualization module is configured to display real-time status information, fault diagnosis results and warning levels to a user, and obtain user input feedback information, which is used for continuous online learning and updating of the multi-fault classification model.

[0012] Preferably, the structured feature vector includes image features, sound features and sensor features.

[0013] Preferably, the deep fusion model is configured to perform weighted combination on the multi-modal features to construct a comprehensive feature representation, including:

[0014] The structured feature vector is classified, cached and time-sequentially sorted.

[0015] The multi-modal feature vectors in the same time period are normalized and initialized in weight.

[0016] The weighted multi-modal features are combined and mapped by the deep fusion model to generate a comprehensive feature representation, the deep fusion model being composed of multiple nonlinear feature mapping units for extracting inter-modal correlation features.

[0017] Preferably, the comprehensive feature is input into a pre-trained multi-fault classification model to identify whether the current conveyor operating state is abnormal and the fault type, including:

[0018] The comprehensive feature representation is subjected to feature dimension standardization and data integrity verification.

[0019] The verified comprehensive feature is input into a pre-trained multi-fault classification model, the model being constructed based on historical operating samples and having feature distributions corresponding to different fault types as classification boundaries.

[0020] The model calculates matching probabilities of each fault type, and determines whether the current conveyor operating state is abnormal based on the principle of maximum probability or a confidence threshold.

[0021] Based on the identified fault type and corresponding confidence, a risk level evaluation result is generated.

[0022] Preferably, the time-series features are used to calculate a fault trend score, and a dynamic threshold algorithm is used to determine the risk level, specifically including:

[0023] The continuously collected comprehensive feature representations are time-sequentially segmented by a sliding window, each window containing a preset number of sampling periods.

[0024] The change rate and fluctuation amplitude of key feature parameters in each window are calculated to generate a time-series feature curve.

[0025] Calculate a trend score based on the time series characteristic curve, the trend score being obtained by comparing the current rate of change with the deviation degree of the historical stable interval;

[0026] Compare the trend score with a dynamic threshold, when the trend score exceeds the dynamic threshold, determine the risk level and output the corresponding early warning signal.

[0027] Preferably, the early warning information is generated when the risk level exceeds the preset threshold, and a hierarchical response strategy is adopted according to the risk level, comprising:

[0028] Determine the risk level and judge whether it exceeds the set early warning response threshold;

[0029] When the threshold is exceeded, generate early warning information containing fault description, location, risk level and recommended measures in combination with fault type, trend score and duration;

[0030] According to the risk level division rule, the corresponding response strategy is triggered.

[0031] Preferably, the feedback information is used for continuous online learning and updating of the multi-fault classification model, comprising: structuring the user feedback information and associating it with the corresponding comprehensive feature representation for labeling to form new training samples; input the new training samples into the multi-fault classification model to perform online incremental training, and realize continuous optimization and updating of the multi-fault classification model.

[0032] In the above technical solution, the technical effects and advantages provided by the present application are:

[0033] 1、The present application realizes all-round, real-time sensing and intelligent diagnosis of the running state of the belt conveyor by constructing an integrated fault detection and early warning system of multi-modal perception, edge preprocessing, cloud fusion analysis and risk grading response. Compared with the traditional scheme relying on single signal judgment, the present application introduces multi-source feature fusion of image, sound, tension, temperature and vibration, and combines time series trend analysis and dynamic threshold algorithm, which significantly improves the accuracy of fault recognition and the timeliness of early warning.

[0034] 2、The present application also introduces an online learning mechanism driven by user feedback, which continuously optimizes the fault classification model through structured storage and incremental training, so that the system has continuous learning and self-adaptive ability, can adapt to the changing trend under different working conditions, and effectively reduces the false positive rate and the false negative rate. The overall system architecture is flexible and highly expandable, and is suitable for various complex conveying scenes such as mines, ports and power plants. BRIEF DESCRIPTION OF DRAWINGS

[0035] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0036] Figure 1 This is a flowchart of the system modules of the present invention. Detailed Implementation

[0037] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0038] For examples, please refer to Figure 1 As shown in this embodiment, a belt conveyor fault detection and early warning system includes:

[0039] The multimodal sensing module is used to collect image, sound, temperature, tension and vibration data generated during the operation of the conveyor;

[0040] The edge computing module is used to extract and preprocess the features of the collected data, and encapsulate the extracted feature data into structured feature vectors.

[0041] The feature fusion and analysis module is used to receive the structured feature vector, and to perform weighted combination of multimodal features based on a deep fusion model to construct a comprehensive feature representation; the comprehensive features are then input into a pre-trained multi-fault classification model to identify whether there are any abnormalities in the current operating status of the conveyor and the type of fault.

[0042] The risk level determination module is used to calculate the fault trend score based on time series characteristics and combine it with a dynamic threshold algorithm to determine the risk level.

[0043] The intelligent early warning module is used to generate early warning information when the risk level exceeds the preset threshold, and adopt a graded response strategy according to the risk level, including local audible and visual alarms, remote push or triggering shutdown commands;

[0044] The visualization module is used to display real-time status information, fault diagnosis results and warning levels to users, and to obtain user input feedback information. The feedback information is used to continuously learn and update the multi-fault classification model online.

[0045] In this invention, the multimodal sensing module serves as the data acquisition front-end of the entire system. Its main function is to comprehensively, in real-time, and synchronously collect key status information generated during the operation of the belt conveyor. The multimodal sensing module includes an image acquisition unit, a sound acquisition unit, a temperature sensor, a tension sensor, and an acceleration sensor, which are respectively installed in key parts of the conveyor, such as the drive roller area, tensioning device, idler roller group, conveyor belt edge, and guide wheels. Wherein:

[0046] The image acquisition unit uses an industrial camera, preferably installed above or to the side of the conveyor belt, to acquire images of the conveyor belt's surface condition. Through image acquisition, it is possible to visually identify fault signs such as belt surface tears, belt misalignment, material accumulation, and foreign object obstruction.

[0047] The sound acquisition unit uses a condenser or MEMS microphone array, deployed near the moving parts of the equipment, to collect ambient sound signals during the conveyor's operation. By performing spectral analysis on the mechanical operating noise, abnormal vibrations, bearing damage, or roller malfunctions can be identified as abnormal noises.

[0048] Temperature sensors, including infrared temperature sensors or thermocouples, are used to monitor the surface temperature of key components such as drive motors, rollers, and bearings, and to capture abnormal temperature rises in real time, thereby determining whether there are potential hazards such as overload or lubrication failure.

[0049] Tension sensors are installed on tensioning devices or the return section of conveyor belts to measure the tension of the belt during operation, preventing slippage, belt damage, or reduced transmission efficiency caused by abnormal tension.

[0050] Accelerometers are deployed on idler supports or at critical connection points to detect vibration characteristics of the conveyor. Abnormal vibration patterns often indicate potential problems such as loose components, structural fatigue, or eccentric operation of components.

[0051] Each sensor unit is connected to the edge computing module via wired communication or industrial-grade wireless communication (such as CAN bus, RS485, LoRa, etc.) to ensure synchronous acquisition and timestamp alignment of multi-source data.

[0052] In one embodiment of the present invention, the edge computing module is located at the on-site control terminal of the conveyor, serving as a local data processing unit of the system. Its main function is to perform preliminary processing on the raw data such as images, sounds, temperature, tension, and vibration collected by the multimodal sensing module, extract feature information reflecting the operating status, and finally encapsulate this information into a structured feature vector for use by subsequent analysis modules. The edge computing module mainly includes the following processing steps:

[0053] Because different sensors have different data acquisition frequencies and response times, to ensure the consistency of various data in time, the edge computing module first synchronizes the image, sound, and sensor-acquired data according to a unified time base. The system achieves coordinated pairing of different types of data by adding timestamps to each set of acquired data and aligning them according to time sequence.

[0054] The edge computing module performs basic cleaning processing on the synchronized raw data, mainly including outlier removal and signal stability enhancement. For example, it removes blurry frames from images, smooths abrupt changes in sensor data, and performs simple suppression of background noise in sound signals.

[0055] The edge computing module extracts key parameter information that can be used to identify the operating status of the conveyor based on the physical characteristics of various types of data. For example, it extracts the conveyor belt edge contour, the number and location of surface damage areas from image data; it extracts the presence of abnormal knocking sounds, friction sounds, and other characteristic changes during operation from sound data; it extracts the temperature rise trend of key components (such as motors and bearings) from temperature data; and it extracts the stability, fluctuation amplitude, and changes in characteristics during operation from tension and vibration data.

[0056] The extracted image features, sound features, and sensor features are integrated and encapsulated in a structured manner into a structured feature vector. A structured feature vector is a data set composed of several numerical terms representing different state parameters, which can comprehensively reflect the current operating status of the conveyor.

[0057] In this invention, the feature fusion and analysis module is used to perform in-depth feature fusion and fault identification analysis on the structured feature vectors from the edge computing module, thereby determining whether there are any abnormalities in the operating status of the conveyor and its specific fault type.

[0058] This module includes two main functional units: a multimodal feature fusion unit and a fault identification unit. The two units are connected and processed together through comprehensive feature representation.

[0059] The multimodal feature fusion unit processes the structured feature vectors uploaded by the edge computing module to construct a fused comprehensive feature representation. Its processing flow includes the following steps:

[0060] Because the system receives a large amount of feature data from different modalities in real time (such as image features, sound features, tension features, temperature features, and vibration features), and because there are differences in acquisition latency and frequency among the data sources, the received structured feature vectors are first classified and cached according to the data source and timestamp. The system establishes a multi-channel buffer and uses a queue method to sort the features according to the acquisition time, ensuring the temporal alignment and integrity of different modal data within each fusion cycle.

[0061] To eliminate differences in numerical scale, dimensions, and statistical distribution among different modal features and improve the computational stability of subsequent fusion models, the system normalizes the multimodal feature vectors within the same fusion cycle. Specifically, the minimum value of each dimension is subtracted from the feature value of that dimension, and then divided by the difference between the maximum and minimum values, thereby mapping the features to a unified interval of 0 to 1.

[0062] Subsequently, the system assigns initial weights to each modal feature. The initial weights are determined based on the following factors: the modality's stability in historical samples (i.e., a smaller standard deviation indicates higher stability), its correlation with the target fault type (obtained statistically through information gain, mutual information, etc.), and the system's preset strategy factor. The initial weight values ​​are set between 0.1 and 1.

[0063] The preprocessed multimodal features are input into a deep fusion model for combined mapping to construct a comprehensive feature representation. The deep fusion model consists of multiple nonlinear feature mapping units. Each mapping unit consists of a fully connected layer, an activation function (such as ReLU or Tanh), and a normalization layer, used to perform dimensionality transformation, intermodal feature interaction extraction, and nonlinear combination of the input features.

[0064] For example, for a set of input feature vectors X=[x1,x2,...,xn], where n is the total number of features, the model performs a first-level weighted transformation and activation process, which can be formally expressed as: the output value of each fusion unit = activation function (weight matrix × input vector + bias term).

[0065] In practice, the model uses historical sample sets for supervised learning during the training phase to optimize the parameters of each layer with the goal of minimizing fault classification error. In actual operation, the model uses the weighted combined output as a comprehensive feature representation vector, with a length typically set between 64 and 256 dimensions. The specific length can be adjusted according to the system's computing resources and recognition accuracy requirements.

[0066] Fault Identification Unit: After the comprehensive feature representation is generated, the system proceeds to the fault identification unit to identify the fault type and determine the operating status. This process includes the following steps:

[0067] To ensure the consistency of the input format for the pre-trained model, the comprehensive feature representation is first standardized in terms of dimensionality. This involves linearly transforming the current input features based on the mean and standard deviation used during model training to maintain a consistent numerical distribution. Simultaneously, the feature vector dimensions and numerical range are validated for completeness, ensuring no missing or outlier values. Any inputs that do not meet the requirements are automatically discarded or supplemented with default values.

[0068] The comprehensive features, after standardization, are input into a multi-fault classification model for operational status identification. This classification model is built upon historical operational data and trained using supervised learning methods. Its core principle is to map different types of fault samples to specific feature distribution regions.

[0069] During the training phase, the system classifies historical samples according to known fault labels and extracts their corresponding comprehensive feature vectors. Machine learning methods (such as support vector machines, random forests, or lightweight neural networks) are then used to model the distribution boundaries of each fault type. The model training objective is to minimize the difference between the predicted and actual labels while improving generalization ability on unknown samples.

[0070] In actual operation, after receiving the comprehensive feature vector, the model outputs the matching probability value for each type of fault. For example, the output is: [normal operation: 0.85, belt slippage: 0.05, bearing overheating: 0.04, roller abnormal vibration: 0.06].

[0071] The current operating status of the conveyor is determined based on the principle of maximum matching probability. If the probability value corresponding to "normal operation" is lower than a set threshold (e.g., 0.75), the current status is considered abnormal. Furthermore, if the matching probability of a certain fault type exceeds a set confidence threshold (e.g., 0.6), it is identified as a possible fault type. The confidence threshold can be adjusted according to different application scenarios and supports online system updates to improve sensitivity or stability.

[0072] In this invention, the risk level determination module is used to perform trend analysis and risk assessment on the operating status of the belt conveyor, so as to provide early warning and intervention before the fault fully manifests, thereby improving the system's proactive prevention and control capabilities and operational safety. The risk level determination module includes the following processing steps:

[0073] First, the continuously received composite feature representations are processed using a sliding window in chronological order. Specifically, within a set time window length (e.g., each window contains 10 sampling periods, with a 5-second interval between each period), the composite features are segmented, with each window representing a continuous feature observation period. The sliding window slides forward with a set overlap step size (e.g., 50%), forming multiple time series segments to ensure the continuity of feature trends.

[0074] Within each sliding window, key feature parameters (such as the rate of temperature rise, vibration amplitude, and the proportion of change in the cracked area in the image) are selected, and their rate of change and fluctuation amplitude are calculated to construct a time-series feature curve. The specific processing is as follows:

[0075] Rate of change: refers to the speed at which each feature increases or decreases within the current window. It is calculated as the difference between the current value of the feature and the mean of the previous window, divided by the mean of the previous window.

[0076] Fluctuation amplitude: The difference between the maximum and minimum values ​​of this feature within the current window, reflecting short-term stability;

[0077] Arrange the above indicators by time to form a multi-dimensional time series feature set containing multiple key features, which can be used as a basis for judging trends.

[0078] The trend score is used to quantify the degree of deviation between the current comprehensive feature sequence and the historical normal state. The scoring process includes the following steps: extracting historical stable interval data from the initial stage of operation or a set time period as a reference benchmark; comparing the rate of change within the current window with the mean and standard deviation in the historical interval; the trend score is the multiple of the standard deviation of the current rate of change from the historical mean, calculated as: Trend Score = (Current Rate of Change - Historical Mean) ÷ Historical Standard Deviation; the trend score reflects the degree of difference between the current operating state and the long-term stable state; a higher score indicates that the system may be deviating from its normal operating trajectory.

[0079] To adapt to changes in the equipment's operating environment and conditions, a dynamic threshold algorithm is introduced instead of a fixed threshold criterion. This algorithm dynamically calculates the warning threshold based on the system's recent operating status, primarily according to the following rule: Dynamic threshold = , where μ is the mean of recent trend scores, σ is the standard deviation, and k is the sensitivity coefficient, usually set between 1.5 and 2.5; the threshold can be adaptively adjusted according to the actual application scenario, for example: a higher threshold can be used during the start-up phase of the conveyor, and a lower threshold can be used during normal operation;

[0080] The current trend score is compared with a dynamic threshold to determine whether the warning conditions have been met. When the trend score exceeds the threshold, it is considered to pose a potential risk, and the risk level is classified according to the degree to which the score exceeds the limit, as follows:

[0081] Low risk (trend score exceeds the threshold but does not exceed twice the threshold): Record the status change and observe.

[0082] Medium risk (trend score exceeds twice the threshold but is below three times the threshold): The system generates an early warning message and notifies the on-duty personnel;

[0083] High risk (trend score exceeds 3 times the threshold): The system triggers an alarm and can take protective measures such as shutdown and speed limit.

[0084] In this invention, the intelligent early warning module is used to promptly generate early warning information when the system detects that the risk level exceeds a set threshold, and trigger corresponding response measures according to the risk level classification rules. This module has functions such as data judgment, information generation, level linkage, and feedback output, enabling closed-loop early warning control from fault trend identification to response execution. The processing flow of the intelligent early warning module includes the following steps:

[0085] This step is used to receive the risk level result output from the risk level determination module and determine whether it exceeds the preset warning response threshold in order to decide whether to enter the warning processing procedure.

[0086] The warning response threshold set in the system is a configurable, tiered parameter used to distinguish whether an active response should be triggered. Its value can be set based on actual system operating experience and the impact of historical faults, and generally includes the following levels:

[0087] Threshold 1: Low-risk response starting point (e.g., score exceeds the dynamic threshold but is less than twice the threshold);

[0088] Threshold 2: Starting point for medium-risk response (e.g., score exceeding twice the dynamic threshold but less than three times);

[0089] Threshold 3: High-risk response threshold (e.g., score exceeding 3 times the dynamic threshold).

[0090] During operation, the risk level corresponding to each assessment result is compared. If the risk level exceeds the threshold of 1 or above, the process proceeds to the next step.

[0091] After confirming that the risk level exceeds the response threshold, the system activates the early warning information generation module. This module automatically assembles a standardized early warning message based on the following parameters:

[0092] Fault type: Output by a multi-fault classification model, including but not limited to belt misalignment, slippage, tearing, bearing overheating, roller vibration, etc.

[0093] Trend score: derived from the risk level determination module, used to indicate the degree of deviation of the current state from the historical state;

[0094] Fault duration: refers to the cumulative time during which the fault trend score continues to exceed the dynamic threshold, in seconds;

[0095] Risk level: Automatically determined based on the ratio of trend score to dynamic threshold, divided into three levels: low, medium, and high;

[0096] Recommended measures: Based on the preset rule base, the system automatically matches one or more response suggestions according to the fault type and level, such as "check the tensioning device", "suggest reduced load operation", "stop the machine immediately", etc.

[0097] Fault location: The system combines the sensor deployment location or image recognition results to identify the specific location or area number where the fault occurred.

[0098] This information is stored in a structured manner and includes metadata such as a unique event number and generation timestamp, making it easy to track and manage.

[0099] The generated warning information will enter the response processing module, where different response strategies will be automatically executed based on its corresponding risk level. The system has the following built-in hierarchical response mechanisms:

[0100] Low risk (Level 1): Only records to the local event log, displays a yellow warning icon on the interface of the on-duty personnel, and does not output control commands;

[0101] Medium risk (Level 2): ​​Triggers the local audible and visual alarm and sends the warning information to the host computer or DCS system via industrial communication protocols (such as Modbus TCP, RS485), and pushes it to the remote monitoring platform at the same time;

[0102] High risk (Level 3): Based on the execution of all medium risk operations, the linkage control system issues an automatic shutdown or speed limit control command and sends an emergency alarm notification to the designated maintenance personnel's mobile app via wireless means (such as 4G / 5G or LoRa).

[0103] The above response strategy can be extended and configured according to actual project needs. The system provides a strategy configuration interface to support customization of response content under different fault types.

[0104] In this invention, the visualization module is used to provide users with real-time information display of the belt conveyor's operating status and a human-machine interaction interface, and has the dual functions of information visualization presentation and user feedback collection.

[0105] The visualization module connects to the fault diagnosis and early warning modules via data interfaces, receiving the following operational information in real time: operating status parameters of various parts of the conveyor (such as tension, vibration, temperature, and operating speed); fault types automatically identified by the system and their locations; assessed risk levels (such as low, medium, and high); fault trend curves and historical alarm records. This information is displayed graphically on the user terminal device (e.g., dashboards, line graphs, heatmaps, prompt windows, etc.), allowing operators to quickly grasp the current equipment status and potential risks.

[0106] The visual interface includes a feedback window, allowing users to perform the following actions on the system's diagnostic results: confirm the accuracy of the fault type determined by the system; mark false alarms, missed alarms, or misidentifications; and input the actual fault type or additional explanatory information. All user actions are automatically recorded in the system background and bound to a timestamp, user identity information, and corresponding diagnostic context.

[0107] The user feedback information is used to drive the system's multi-fault classification model to continuously learn and optimize online. The specific implementation steps are as follows:

[0108] First, the user-input feedback information is structured. The feedback includes: the actual fault type label (or no fault); the system's original diagnostic result; the label selected or corrected by the user; the time point and duration of the fault occurrence; and the comprehensive feature representation vector corresponding to the diagnostic result (generated by the feature fusion module). This information is then assembled into a complete data sample and stored in the incremental learning sample library. The sample structure includes fields such as feature input, target label, and source method (user annotation).

[0109] To achieve continuous model optimization, an incremental training mechanism is used to update the existing multi-fault classification model. Incremental training refers to targeted adjustments to model parameters based solely on new samples without retraining the entire model. The implementation is as follows: a training frequency is selected (e.g., daily scheduled training, or when the cumulative number of samples exceeds a set limit); newly labeled samples are combined with a portion of the original training samples to form a mini-training set; fine-tuning training is performed based on existing model parameters, updating the weight parameters; incremental training employs a stability constraint strategy to avoid "catastrophic forgetting" of the model caused by new samples; a sliding window approach can be used, retaining only the latest N feedback samples (e.g., N=500) to control the model's generalization ability and training cost.

[0110] After each round of incremental training, the new model will be internally evaluated. Evaluation metrics include: improvement in classification accuracy; change in false positive rate; improvement in consistency with user feedback. When the evaluation metrics meet the set update criteria, the system will replace the current model with the latest version for use in the next round of operation.

[0111] The multi-fault classification model can employ a lightweight neural network structure (such as a multilayer perceptron MLP) or an ensemble decision tree model (such as a random forest) to ensure efficient deployment of the system on edge devices or cloud servers. Input dimensions: Composed of comprehensive feature representation vectors (e.g., 64-dimensional or 128-dimensional); Output categories: The number of corresponding fault types (e.g., normal, slippage, misalignment, bearing overheating, belt tear, etc.); Loss function: The cross-entropy loss function is used to measure the difference between the predicted label and the true label; Optimization method: Adam or SGD optimization algorithms can be used for model training and fine-tuning.

[0112] Example 2: To verify the technical effectiveness of the belt conveyor fault detection and early warning system proposed in this invention, a system test deployment and data testing were conducted in a simulated coal mine conveying environment. The test platform consisted of the following components:

[0113] The conveyor is 50 meters long and is equipped with 4 sets of drive motors, multiple sets of idlers, a tensioning device and guide wheels;

[0114] Install a multimodal sensing module, including: 3 industrial cameras, 2 audio acquisition devices, 4 infrared temperature sensors, 4 acceleration sensors and 2 tension sensors;

[0115] The edge computing module uses an industrial embedded processor (quad-core 2.0GHz, 4GB RAM).

[0116] Model inference and learning are performed using a GPU server (NVIDIA RTX 3060);

[0117] The human-computer interaction terminal is a touch-screen industrial tablet.

[0118] During the experiment, data was collected and tested under the following five typical fault conditions, as shown in Table 1:

[0119] 300 samples were collected for each type of fault, totaling 1500 fault samples and 1000 normal samples as a control group. The sampling period was 5 seconds. After feature extraction, a structured feature vector was formed and input into the system for identification.

[0120] The recognition accuracy was compared between the traditional rule-based single-modal judgment method (comparison scheme A) and the multimodal fusion system described in this invention (scheme B). The results are as follows:

[0121] Table 2 Accuracy Comparison Table

[0122] As shown in Table 2, the average recognition accuracy increased from 82.1% to 95.3%, indicating that the present invention can significantly improve the accuracy of fault identification through multimodal fusion and trend analysis mechanism.

[0123] In the second phase of the experiment, 20 false positive samples confirmed by user feedback were continuously input into the system as training data, triggering an online incremental learning mechanism. After the incremental update, the model's accuracy in recognizing new test samples further improved to 96.8%, and the false positive rate decreased by 15%, verifying that the system has good adaptive optimization capabilities.

[0124] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A fault detection and early warning system for a belt conveyor, characterized in that: include: The multimodal sensing module is used to collect image, sound, temperature, tension and vibration data generated during the operation of the conveyor; The edge computing module is used to extract and preprocess the features of the collected data, and encapsulate the extracted feature data into structured feature vectors. The feature fusion and analysis module is used to receive the structured feature vector, and to perform weighted combination of multimodal features based on a deep fusion model to construct a comprehensive feature representation; the comprehensive features are then input into a pre-trained multi-fault classification model to identify whether there are any abnormalities in the current operating status of the conveyor and the type of fault. The risk level determination module is used to calculate the fault trend score based on time series characteristics and combine it with a dynamic threshold algorithm to determine the risk level. The intelligent early warning module is used to generate early warning information when the risk level exceeds a preset threshold, and to adopt a graded response strategy according to the risk level. The visualization module is used to display real-time status information, fault diagnosis results and warning levels to users, and to obtain user input feedback information. The feedback information is used to continuously learn and update the multi-fault classification model online.

2. The belt conveyor fault detection and early warning system according to claim 1, characterized in that: The structured feature vector includes image features, sound features, and sensor features.

3. The belt conveyor fault detection and early warning system according to claim 1, characterized in that: The method of constructing a comprehensive feature representation by weighting and combining multimodal features based on a deep fusion model includes: Classify, cache, and sort structured feature vectors according to time series. Perform feature normalization and weight initialization on multimodal feature vectors within the same time period; The weighted multimodal features are combined and mapped by a deep fusion model to generate a comprehensive feature representation. The deep fusion model consists of multiple layers of nonlinear feature mapping units, which are used to extract intermodal correlation features.

4. The belt conveyor fault detection and early warning system according to claim 3, characterized in that: The step of inputting comprehensive features into a pre-trained multi-fault classification model to identify whether there are any abnormalities in the current operating status of the conveyor and the type of fault includes: Perform feature dimension standardization and data integrity verification on the comprehensive feature representation; The verified comprehensive features are input into a pre-trained multi-fault classification model, which is built based on historical operating samples and uses the feature distribution corresponding to different fault types as the classification boundary. The model calculates the matching probability of each fault type, and judges whether there is an abnormality in the current operating status of the conveyor based on the principle of maximum probability or confidence threshold. Based on the identified fault types and their corresponding confidence levels, risk level assessment results are generated.

5. The belt conveyor fault detection and early warning system according to claim 4, characterized in that: The process of calculating a fault trend score based on time-series features and determining the risk level using a dynamic threshold algorithm specifically includes: The continuously acquired comprehensive feature representations are segmented into sliding windows according to time sequence, with each window containing a preset number of sampling periods; Calculate the rate of change and fluctuation amplitude of key feature parameters within each window to generate time series feature curves; A trend score is calculated based on the time series characteristic curve, and the trend score is obtained by comparing the deviation of the current rate of change from the historical stable interval. The trend score is compared with a dynamic threshold. When the trend score exceeds the dynamic threshold, the risk level is determined and a corresponding warning signal is output.

6. The belt conveyor fault detection and early warning system according to claim 1, characterized in that: The method for generating early warning information when the risk level exceeds a preset threshold and adopting a graded response strategy based on the risk level includes: The risk level is assessed and verified to determine whether it exceeds the set early warning response threshold. When the threshold is exceeded, an early warning message is generated by combining the fault type, trend score and duration, which includes fault description, location, risk level and recommended measures. Based on the risk level classification rules, the corresponding response strategy is triggered.

7. The belt conveyor fault detection and early warning system according to claim 6, characterized in that: The feedback information is used for continuous online learning and updating of the multi-fault classification model, including: storing user feedback information in a structured manner and associating and labeling it with the corresponding comprehensive feature representation to form new training samples; inputting the new training samples into the multi-fault classification model to perform online incremental training, thereby realizing continuous optimization and updating of the multi-fault classification model.

Images