Intelligent detection system and method for tearing of lower-layer belt of mining long-distance conveyor

Abstract

The invention provides an intelligent detection system and method for tearing of a lower-layer belt of a mining long-distance conveyor, and belongs to the technical field of mining machinery and intelligent monitoring control. According to the technical scheme, the intelligent detection system for the tearing of the lower-layer belt of the mining long-distance conveyor comprises a sensing measurement module, a data transmission module, a tearing detection module, an alarm and control module and a data storage and analysis module, the invention further provides an intelligent detection method for tearing of the lower-layer belt of the conveyor. The intelligent detection method comprises the following steps that S1, sensing data are collected; s2, constructing a tear sensing database; s3, constructing a tearing state monitoring model; s4, optimizing the tearing state monitoring model; and S5, tearing state monitoring. The system has the advantages that the tension sensors are installed on the lower layer of the conveying belt, real-time collection of multi-point tension data along the path of the belt is achieved, and the system can stably receive sensor data underground through a transmission network constructed through the LoRa wireless communication protocol.

Description

Technical Field

[0001] The present invention relates to the technical field of mining machinery and intelligent monitoring and control, and in particular to an intelligent detection system and method for tearing of the lower belt of a long-distance conveyor for mining. Background Art

[0002] In the mining industry, long-distance, bidirectional belt conveyors are critical equipment for ore transportation. Their efficient transport capabilities and flexible bidirectional transport capabilities make them crucial in mining production processes and widely used in various mining operations. However, the extremely harsh operating environment underground poses significant challenges to the conveyor belts of these belt conveyors.

[0003] In terms of load, ore production in mines fluctuates significantly, and conveyor belts often have to carry weights far exceeding their designed loads. This subjects the belts to prolonged periods of high tension and friction, significantly accelerating aging and damage. Regarding humidity, groundwater seepage within mines and moisture introduced by ventilation systems expose conveyor belts to prolonged humidity, which not only degrades the belt's physical properties but can also cause corrosion and weaken its structural strength. The abundant dust permeating the underground mines not only coats the conveyor belt surface, affecting its transmission efficiency, but can also enter its interior, exacerbating wear. Combined with aging caused by long-term equipment operation, this significantly increases the probability of conveyor belt failures such as tearing and wear. Currently, mining companies rely primarily on traditional manual inspections and simple sensor monitoring methods for conveyor belt inspection. Manual inspections are limited by human physiological and psychological factors. In the high-load, high-risk environment of underground mines, it's difficult for inspectors to maintain full concentration for extended periods. Furthermore, the dim lighting and cramped conditions underground make it difficult to thoroughly and carefully inspect every detail of the conveyor belt, making even minor damage easily overlooked. At the same time, manual inspection is extremely inefficient and cannot meet the needs of continuous production in mines. Frequent shutdown inspections will also affect production progress. Summary of the Invention

[0004] The purpose of the present invention is to provide an intelligent detection system and method for tearing of the lower belt of a long-distance conveyor for mining. The detection system and method can solve the technical problems existing in the prior art, such as difficulty in detecting tearing of the lower conveyor belt, delayed response, high false alarm rate and poor environmental adaptability, and realize high-precision, real-time and intelligent monitoring of the operating status of the lower layer of the conveyor belt and tear fault warning, thereby improving the safety and operation and maintenance efficiency of the underground transportation system in mines.

[0005] In order to achieve the above-mentioned purpose of the invention, the present invention adopts a technical solution specifically as follows: a method for intelligently detecting tearing of the lower belt of a conveyor, comprising the following steps:

[0006] S1. Sensor data acquisition: Several tension sensors, vibration sensors, temperature sensors, and optical sensors are installed on the lower belt of a long-distance, bidirectional belt conveyor used in mining. Strain data, vibration data, temperature data, and crack monitoring data are collected using a wireless sensor network to obtain real-time status data of the belt.

[0007] S2. Build a tear perception database: By analyzing the working conditions of long-distance, bidirectional conveyor belts used in mines, a tear perception database is built, which includes belt vibration, tension, temperature, and crack data.

[0008] S3. Constructing a tearing state monitoring model: using the database constructed in step S2, a deep learning algorithm is used for model training to construct a conveyor belt tearing state monitoring model;

[0009] S4. Optimization of the tearing state monitoring model: Network pruning and model compression technologies are used to lightweight the monitoring model, enabling the model to achieve real-time and fast calculations on a micro-industrial host;

[0010] S5. Tearing status monitoring: Based on the real-time sensor data collected in step S1 as input, the tearing status monitoring model optimized in step S4 is used to calculate and output the tearing status of the lower belt of the long-distance bidirectional transport belt conveyor for mining in real time, thereby realizing real-time monitoring of the belt.

[0011] Furthermore, in step S2, the construction of the tearing perception database includes the following steps:

[0012] S2.1. The sensor collects data once per second and transmits the data to the data acquisition terminal through the wireless module. The data of each sensor is x(t), and the time series data is:

[0013] X={x(1),x(2),…,x(t)}

[0014] Among them, t represents the timestamp, which means the data collected by the sensor at the tth second;

[0015] S2.2. Each sensor transmits data to the coordinator node. After receiving the data, the coordinator node stores the information in the database and adds a timestamp and sensor location mark to each data;

[0016] S2.3. After data collection, wavelet transform is first used to perform denoising to remove signal fluctuations caused by external interference, and a low-pass filter is used to filter out high-frequency noise. The transfer function is:

[0017]

[0018] Where f is the frequency of the signal, fc is the cutoff frequency of the filter, and the signal x~(t) obtained after filtering is the denoised data;

[0019] S2.4. The processed data are classified and stored according to time series and sensor location to form a structured database.

[0020] Furthermore, in step S3, the construction of the tearing state monitoring model includes the following steps:

[0021] S3.1. Extract useful feature data from the established tear perception database, focusing on belt tension, vibration frequency, temperature fluctuation, and crack development.

[0022] S3.2. Use principal component analysis to extract representative features from large-scale data and reduce the data dimension;

[0023] S3.3. Use the LSTM neural network model to train the feature data;

[0024] S3.4. During the training process, the labeled normal state and tearing state datasets are used to optimize the model parameters so that it can accurately distinguish between different states;

[0025] S3.5. After model training is completed, use the validation set to test the model to check the accuracy of the model.

[0026] Furthermore, in the deep learning algorithm described in step S3, the size of the convolution kernel k is selected to be 3×3, and the default value of the learning rate η is 0.01;

[0027] The network pruning and model compression techniques described in step S3.2 determine the pruning ratio p% based on the change in the accuracy of the monitoring validation set, increasing layer by layer from 20% to 50%, and the quantization value xquant ranges from -127 to 127;

[0028] Furthermore, in step S4, the tearing state monitoring model is optimized, including the following steps:

[0029] S4.1. Prune the trained LSTM model to remove neurons and connections that have little impact on the model output, thereby reducing the model's complexity and computational effort.

[0030] S4.2. Use weight pruning to prune redundant filters in the convolutional layer.

[0031] S4.3. Quantize the pruned LSTM model to reduce the storage space of the model.

[0032] Furthermore, in the principal component analysis method described in step S4, characteristic data with eigenvalues ​​greater than 1 are selected as principal components according to the Kaiser criterion.

[0033] Furthermore, in step S5, the tearing state monitoring includes the following steps:

[0034] S5.1. The wireless transmission system continues to receive real-time data from each sensor via the LoRa protocol. Each data packet from the sensor contains information such as timestamp, sensor location, and data value.

[0035] S5.2. The sensor data is sent to the coordinator node through the terminal node, and the coordinator node transmits the data to the computing module;

[0036] S5.3. The calculation module uses the optimized LSTM model to analyze the received real-time data, calculate the current state of the belt, and determine whether there is a tearing risk;

[0037] S5.4. If a tear risk is detected, the calculation module will trigger an alarm, output specific alarm information, and display the tear location and risk level on the monitoring screen;

[0038] S5.5. When the system detects signs of tearing, it will immediately notify the operator through sound and light alarms or through the system to remind them to take maintenance measures in a timely manner.

[0039] An intelligent detection system for lower belt tearing of a long-distance conveyor for mining, comprising:

[0040] Sensing and measurement module: includes tension sensors installed on the sides and top of the conveyor belt, a vibration sensor for detecting belt vibration, a temperature sensor, and an optical sensor installed on the belt surface;

[0041] Data transmission module: including sensor signal acquisition and transmission interface, data transmission network;

[0042] Tear detection module: This module uses data analysis algorithms to deeply process the collected sensor data and determine whether tearing has occurred.

[0043] Alarm and control module: used to respond promptly and take emergency measures;

[0044] Data storage and analysis module: It has data storage functions, provides cloud storage and local storage options, and can perform historical data analysis;

[0045] Self-learning optimization module: It can optimize the system's tear detection algorithm based on historical data and real-time data to improve detection accuracy.

[0046] Furthermore, the tear detection module includes an anomaly detection algorithm, an image processing algorithm and a deep learning algorithm.

[0047] Furthermore, the self-learning optimization module is capable of algorithm self-optimization and adaptive adjustment;

[0048] The algorithm self-optimization is to accumulate historical data so that the system can continuously train and optimize the tear detection algorithm;

[0049] The adaptive adjustment means that the system can automatically adjust the parameters and alarm thresholds of the detection algorithm according to environmental changes and working conditions.

[0050] Compared with the prior art, the present invention has the following beneficial effects:

[0051] 1. By installing a tension sensor beneath the conveyor belt and integrating it with a high-definition camera for image processing, this system enables continuous and stable operation unaffected by light, dust, and obstructions, effectively addressing the difficulty of acquiring sensor information in harsh underground environments. Furthermore, the low-power, highly reliable transmission network constructed using the LoRa wireless communication protocol enables the system to stably receive sensor data underground, avoiding the complex deployment and maintenance challenges of traditional wiring communication methods.

[0052] 2. In terms of data analysis and processing, this invention innovatively combines a one-dimensional convolutional neural network (1D-CNN) with a long short-term memory network (LSTM) to construct a tear state recognition model suitable for tension time series data. The convolutional network is used to extract the local variation characteristics of the tension signal, while the LSTM is used to capture the dynamic evolution patterns within the long time series. The two work together to effectively improve the accuracy of tear state recognition. Compared with traditional static threshold judgment or simple logical judgment methods, this deep learning model has stronger adaptability and can meet the needs of tear detection in different transportation environments and load conditions.

[0053] 3. To achieve efficient operation of the model at the edge, the present invention also introduces network pruning and model compression technologies, and performs lightweight optimization on the tear detection model, enabling it to run quickly on a micro-industrial host, meeting the needs of real-time computing and low-power equipment operation underground, and significantly improving the system's response speed and deployment flexibility.

[0054] 4. This invention establishes a clearly structured and fully documented tear status database, storing and managing each sensor data point, determination results, tear location, and alarm information. This database also continuously optimizes model weights through a closed-loop data feedback mechanism. The system, equipped with intelligent analysis and trend prediction capabilities, can proactively identify potential tear risks, enabling true predictive maintenance and effectively reducing equipment failure rates and operational costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0056] Figure 1 The present invention is a flowchart of an intelligent detection method for tearing of the lower belt of a long-distance conveyor for mining.

[0057] Figure 2 This is a schematic diagram of the sensor layout of the intelligent detection system for lower belt tearing of a long-distance mining conveyor according to an embodiment of the present invention. DETAILED DESCRIPTION

[0058] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. Of course, the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0059] like Figure 1-2 As shown, the present invention provides an intelligent detection system for lower belt tearing of a long-distance conveyor for mining, including a sensing and measurement module, a data transmission module, a tearing detection module, an alarm and control module, a data storage and analysis module, and a self-learning optimization module.

[0060] (1) The perception measurement module includes:

[0061] Tension sensor: High-precision tension sensor with a sensitivity better than ±0.5% and an operating temperature range of -40°C to +85°C. It is installed on the side and top of the conveyor belt to monitor belt tension changes in real time.

[0062] Vibration sensor: A three-axis vibration sensor with a frequency range of 0.1Hz-10kHz and a resolution of 0.005g is used to detect the vibration generated by the belt during operation. In particular, when the belt is torn or loose, the vibration pattern changes significantly. The sensor can capture these changes in time and provide early warning signals of tearing;

[0063] Temperature sensor: Choose a thermocouple temperature sensor with a measurement range of -50°C to +150°C, a temperature resolution of 0.1°C, and an accuracy of ±0.5°C. Install it on the belt surface to monitor belt temperature changes in real time, thus preventing equipment failures caused by overheating.

[0064] Optical sensor: Choose a high-resolution industrial camera with a resolution of no less than 7680x4320 pixels, a frame rate of no less than 30fps, a CCD sensor type, and a GigE Vision interface type. Install it on the belt surface or on the side bracket. Use image recognition technology to monitor abnormal conditions such as cracks and tears on the belt surface in real time, and use image processing algorithms to accurately locate the damage.

[0065] (2) Data transmission module:

[0066] The module includes a sensor signal acquisition and transmission interface, responsible for transmitting the data collected by each sensor to the central processing unit through the industrial communication protocol LoRa;

[0067] LoRa terminal node device:

[0068] Installed near each sensor collection point, it is responsible for encoding and packaging the data collected by tension sensors, vibration sensors, etc., and sending it to the coordinator node through the LoRa protocol;

[0069] The data transmission network uses wireless transmission to quickly transmit the collected raw data to the tear detection module.

[0070] (3) Tear detection module: The tear detection module includes anomaly detection algorithm, image processing algorithm and deep learning algorithm;

[0071] Anomaly detection algorithm: Through real-time analysis of tension, vibration and temperature data, it identifies abnormal fluctuations that exceed safety thresholds and determines the possible risk of tearing.

[0072] Image processing algorithms: When the system uses optical sensors or video monitoring, image processing algorithms can use image recognition technology to initially locate cracks, tears, or damage on the belt surface. Edge detection and feature extraction methods are used to further improve detection accuracy.

[0073] Deep Learning Algorithm: A deep learning model, Convolutional Neural Network (CNN), is used to optimize the accuracy of tear detection and automatically identify more complex damage patterns.

[0074] (4) Alarm and control module: used to respond promptly and take emergency measures, including sound and light alarms, industrial control relays, and human-machine interface (HMI) industrial display screens.

[0075] (5) Data storage and analysis module:

[0076] This module contains the embedded database SQLite;

[0077] Data storage eMMC, which stores the operating system, model runtime library, and deployment scripts;

[0078] The analysis and calculation adopts the industrial computer of tear detection module.

[0079] (6) Self-learning optimization module:

[0080] By accumulating historical data, the system can continuously train and optimize the tear detection algorithm, improve its ability to identify complex damage, and automatically adjust the detection algorithm parameters and alarm thresholds according to environmental changes and working conditions to ensure stability and accuracy under various operating conditions.

[0081] In order to better achieve the above-mentioned invention effects, this embodiment provides an intelligent detection method for tearing of the lower belt of a long-distance conveyor for mining, comprising the following steps:

[0082] Step 1: Sensor data acquisition: Install at least eight sensors on the lower belt of a long-distance, bidirectional belt conveyor used in mining. Use a wireless sensor network to collect strain data, vibration data, temperature data, and crack monitoring data to obtain real-time status data of the belt.

[0083] The following takes the collection of tension data as an example:

[0084] The sensing module completes the real-time collection of multi-point tension data of the lower belt: Several tension sensors and high-definition cameras are installed at key locations along the running direction of the lower belt to form a sensing module, which is used to sense the stress state of the lower belt during the conveying process in real time. Each tension sensor collects the longitudinal tension changes during the operation of the belt to form tension time series data Ti(t), where i is the sensor number and t is the timestamp;

[0085] First, the camera is used to initially identify the tearing location and locate the cracks, tears or damage on the belt surface. Edge detection and feature extraction methods are used to further improve the accuracy of detection. The specific steps are as follows:

[0086] Get an image or video stream of the belt, grayscale it, and Gaussian filter it to reduce noise:

[0087] Igray=rgb2gray(I)

[0088] Ismooth=gaussian_filter(Igray,σ)

[0089] Where I is the original image, σ is the standard deviation of the Gaussian filter;

[0090] Use the Canny edge detection algorithm to detect cracks or damaged areas on the belt surface:

[0091] E=Canny(Ismooth,low_threshold,high_threshold)

[0092] Where E is the edge image, and the low threshold and high threshold are used to control the sensitivity of edge detection;

[0093] Perform contour extraction on the detected edges to extract the crack area on the belt surface:

[0094] Contours = findContours(E)

[0095] Analyze the area and aspect ratio of the crack profile and compare it with the predetermined standards to determine whether it is a tear or damage;

[0096] The severity of the damage can be further determined by the location and size of the cracks;

[0097] Damage assessment formula:

[0098] D=Area(C)×AspectRatio(C)

[0099] Where D is the damage assessment value, C is the crack profile, and the area and aspect ratio are the key parameters for evaluation;

[0100] Each sensing unit then collects information using Fiber Bragg Grating (FBG) sensing technology. Its working principle is that when the belt is stressed, the FBG attached to the belt or at the tensioning part produces a slight deformation, causing its reflection wavelength λB to shift. The amount of this shift can be used to infer the stress on the belt.

[0101] The calculation formula for the change in fiber Bragg grating reflection wavelength is as follows:

[0102] ΔλB=λB·((1-pe)·ε)

[0103] in:

[0104] ΔλB: shift of reflected wavelength;

[0105] λB: original reflection wavelength;

[0106] pe: effective photoelastic coefficient of optical fiber;

[0107] ε: belt strain (caused by tension);

[0108] The relationship between tension F and strain ε is established by Hooke's law:

[0109] F=E·A·ε1F

[0110] in:

[0111] E: elastic modulus of belt material;

[0112] A: effective cross-sectional area of ​​the sensor adhesion area;

[0113] ε1: strain calculated by grating wavelength shift;

[0114] By combining the above two formulas, the grating wavelength change can be accurately converted into belt tension value. The sensing module outputs these tension data in the form of time series as follows:

[0115] Ti(t)=f(ΔλB)

[0116] Where i represents the sensor number and t is the timestamp;

[0117] After data collection, wavelet transform is used to remove noise and remove signal fluctuations caused by external interference. A low-pass filter is used to filter out high-frequency noise. The transfer function is:

[0118]

[0119] Wherein, f is the frequency of the signal, fc is the cutoff frequency of the filter, and the selection range is 5Hz-50Hz. The signal x~(t) obtained after filtering is the denoised data.

[0120] Step 2: Build a tear perception database: By analyzing the operating conditions of long-distance, bidirectional conveyor belts used in mines, a tear perception database is constructed. This database contains belt vibration, tension, temperature, and crack data for tear state analysis and model training. Taking tension data as an example:

[0121] (1) Data acquisition completes the orderly acquisition and synchronization marking of tension data: the tension signal output by the sensing module is uniformly scheduled by the data acquisition module after A / D conversion. The sampling frequency is set to f = 1000 Hz, that is, 1000 data points are collected per second, and a sampling timestamp is added to achieve time synchronization of data collected by all sensors. The tension data within each sampling period Δt is organized as a tensor:

[0122] T={T1(t),T2(t),...,Tn(t)}

[0123] Where n is the total number of sensors;

[0124] (2) Constructing a tension data database: Based on the processed multi-point data, a structured tension data database is constructed to facilitate subsequent storage, query, and intelligent analysis. The database construction method includes:

[0125] Database type: Uses lightweight embedded relational database SQLite, which is convenient for deployment in mine field environments and real-time data writing;

[0126] Database table structure design:

[0127] Table name: BeltTensionData

[0128] Field Description:

[0129] id: auto-increment primary key, uniquely identifies the data entry;

[0130] timestamp: timestamp, records the data collection time, accurate to milliseconds;

[0131] sensor_id: sensor number, identifying the source of tension data;

[0132] tension_value: normalized tension value;

[0133] status_flag: data status flag (0 means normal, 1 means abnormal rejection);

[0134] The data writing strategy batch-inserts pre-processed tension data in sampling time order, and checks data integrity before each write to ensure no omissions. A maximum database capacity threshold is set, and a cyclic overwrite writing strategy is adopted when it exceeds the threshold to achieve dynamic data updates and historical record maintenance. Indexes are established for the timestamp and sensor_id fields to improve data query efficiency.

[0135] Step 3: Construct a tearing state monitoring model: Using the database constructed in step 2, a deep learning algorithm is used for model training to construct a belt tearing state monitoring model for long-distance bidirectional transport belt conveyors used in mines.

[0136] For the four characteristic data of tension value, vibration frequency, temperature fluctuation and crack development, the principal component analysis method is used:

[0137] The first step is to calculate the covariance matrix:

[0138]

[0139] Where G is the matrix composed of feature data

[0140] Then the eigenvalue decomposition:

[0141] C=NΛN T

[0142] Where Λ is the eigenvalue diagonal matrix, n is the number of samples, and N is the decomposition matrix.

[0143] According to the Kaiser criterion, the feature data with eigenvalues ​​greater than 1 are selected as the principal components to find the projection direction and form the projection matrix W, so as to maximize the variance of the projected data and reduce the data dimension.

[0144] The data analysis module then inputs the feature vector of each sensor into a pre-trained convolutional neural network model for tear recognition. The model is a three-layer one-dimensional convolutional neural network (1D-CNN) with the following structure:

[0145] First layer: Conv1D (kernel = 3, filters = 32) + ReLU + MaxPooling

[0146] Second layer: Conv1D (kernel = 3, filters = 64) + ReLU + MaxPooling

[0147] The third layer: fully connected layer (Dense) + Softmax output classification

[0148] The input is:

[0149] Fi=[Ti(t),Di(t),σi(t)]

[0150] The output is:

[0151] Pi=CNN(Fi)∈{0,1}

[0152] Output 1 indicates that the area is torn;

[0153] Convolutional neural networks are used to extract spatial features from the raw tension data. These features are further fed into the long short-term memory (LSTM) model as time series input to analyze the evolution trend of the tension sequence in the time dimension. The input layer receives the tension feature sequence extracted from the CNN, with a dimension of X∈RT×F, where T is the number of time steps and F is the feature dimension of each step.

[0154] LSTM layer:

[0155] Single-layer LSTM network, the number of hidden units is set to 128;

[0156] Use bidirectional LSTM (Bi-LSTM) to improve time series modeling capabilities;

[0157] The activation function uses tanh;

[0158] Output layer: fully connected layer + Sigmoid output;

[0159] Output tearing probability pt∈[0,1].

[0160] The workflow of the LSTM unit is as follows:

[0161] ft = σ(Wf [ht-1, xt] + bf) (forget gate)

[0162] it=σ(Wi·[ht-1,xt]+bi)(input gate)

[0163] C~t=tanhf0WC·[ht-1,xt]+bC)(candidate state)

[0164] Ct=ft*Ct-1+it*C~t(unit state)

[0165] ot=σ(Wo·[ht-1,xt]+bo)(output gate)

[0166] ht=ot*tanhf0(Ct)(hidden state)

[0167] Among them, xt is the CNN feature input at the tth time step, ht is the hidden layer state, and Ct is the unit state;

[0168] Discrimination output mechanism:

[0169] Input the hidden state hT of the last time step into the fully connected layer;

[0170] Use the sigmoid activation function to output the tearing probability:

[0171] ptear=σ(Wfc·hT+bfc)

[0172] When ptear>0.8, the system determines that the current state is at high risk of tearing.

[0173] The cross entropy loss function is used for model training. The loss function is as follows:

[0174]

[0175] Where yi is the true label, y^ is the model prediction value, and n is the number of samples;

[0176] The CNN network is trained using the backpropagation algorithm, and the weights and biases are adjusted using the optimization algorithm:

[0177]

[0178] Among them, wt is the current weight, η is the learning rate, is the gradient of the loss function.

[0179] Step 4: Optimize the tear state monitoring model: Use network pruning and model compression technology to lightweight the monitoring model, so that the model can achieve real-time and fast calculation on the micro industrial host, improving the system's response speed and accuracy. Specifically:

[0180] Tear state monitoring model optimization: To improve the real-time operation capability and resource efficiency of the tear state discrimination model on the mine site's industrial control computer, this step lightweights the original three-layer one-dimensional convolutional neural network (1D-CNN). The optimization method includes network pruning and model compression.

[0181] Network Pruning

[0182] Pruning objects: Prune redundant filters in the convolutional layer to reduce the number of unimportant channels and reduce the model size;

[0183] Pruning strategy: Use a pruning method based on weight sparsity to calculate the L1 norm of each convolution kernel Wi∈Rk×c×l:

[0184] |||Wi|||1=∑j=1k∑m=1c∑n=1l||wjmn||

[0185] The pruning ratio p% is determined by monitoring the changes in the accuracy of the validation set, and is increased layer by layer from 20% to 50%. The first p% channels with the largest L1 norm are retained, and the remaining filters are set to zero and the weights are frozen during training. In order to prevent the destruction of existing features, the fine-tuning learning rate is 1 / 10 of the initial learning rate.

[0186] Retraining: After pruning, fine-tune the model using the training set and the Adam optimizer to maintain convergence stability and accuracy.

[0187] Model Quantization

[0188] Target platform adaptation: To enable the model to be deployed on low-power industrial control devices, 8-bit fixed-point quantization (INT8 quantization) is performed.

[0189] Quantization method: Symmetric linear quantization is used, and the following transformations are performed on both the convolutional layer weights and activation values:

[0190] xquant=round(xfloat / s)

[0191] The quantization value xquant range is (-127, 127), s is the scale factor, and the maximum value is set by channel. During the quantization process, the TensorFlow Lite quantization calibration tool is used to perform static quantization on the trained model.

[0192] Post-quantization calibration: Use the training set to perform post-quantization calibration to ensure that the accuracy drop is controlled within 1%;

[0193] Optimization results

[0194] Original model size: 2.3MB → Optimized model size: 680KB

[0195] Inference latency: Single prediction time reduced from approximately 45ms to 13ms (based on ARM Cortex-A7 testing)

[0196] The accuracy remains above 97% with no noticeable degradation.

[0197] Step 5: Tear status monitoring: Using the real-time sensor data collected in step 1 as input, combined with the tear status monitoring model optimized through network pruning and model compression in step 4, real-time monitoring of the tear status of the lower belt of the long-distance bidirectional transport belt conveyor for mining is achieved.

[0198] Real-time tension change data collected by sensors deployed on the lower layer of the conveyor belt is transmitted via the LoRa wireless communication protocol. Each sensor node, acting as an end device, sends data to the nearest coordinator node via the LoRa link. Through real-time analysis of tension, vibration, and temperature data, abnormal fluctuations exceeding safety thresholds are identified and potential tear risks are determined. The specific steps are as follows:

[0199] The time series data of tension T(t), vibration V(t) and temperature θ(t) are collected through sensors;

[0200] The data collected each time is {T(t), V(t), θ(t)} and recorded at timestamp t;

[0201] For each moment t, make the following judgment:

[0202] The data of each sensor is Z-score standardized using the following formula:

[0203]

[0204] Among them, U t is the data at the current moment, μ X and σX are the mean and standard deviation of the data respectively;

[0205] Calculate the U value for each data (tension, vibration, temperature): U T (t) corresponds to tension data, U V (t) Corresponding vibration data U θ (t) corresponds to temperature data;

[0206] Set the threshold α as the criterion for anomaly detection;

[0207] If the Z value of any data exceeds the threshold α, it is determined to be abnormal data;

[0208] Detection formula:

[0209] if ZU(t)>α then the data is anomalous.

[0210] Where U∈{T,V,θ};

[0211] The coordinator node aggregates data from multiple terminal nodes and forwards the data to the industrial computer at the center of the system via high-speed 5LoRa. Once the LSTM model determines that the current state is at high tearing risk (pt>θ), the system immediately triggers an audible and visual alarm (buzzer + LE light). The system automatically saves the current abnormal data, the contextual data of the previous and subsequent time periods, and the model judgment results to the local database and the cloud database. The system then uses this historical data for model retraining or reinforcement learning optimization.

[0212] 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, improvements, etc. 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 detection method for tearing of the lower belt of a conveyor, characterized in that: The steps include: S1. Sensor data acquisition: Several tension sensors, vibration sensors, temperature sensors, and optical sensors are installed on the lower belt of a long-distance, bidirectional belt conveyor used in mining. Strain data, vibration data, temperature data, and crack monitoring data are collected using a wireless sensor network to obtain real-time status data of the belt. S2. Build a tear perception database: By analyzing the working conditions of long-distance, bidirectional conveyor belts used in mines, a tear perception database is built, which includes belt vibration, tension, temperature, and crack data. S3. Constructing a tearing state monitoring model: using the database constructed in step S2, a deep learning algorithm is used for model training to construct a conveyor belt tearing state monitoring model; S4. Optimization of the tearing state monitoring model: Network pruning and model compression technologies are used to lightweight the monitoring model, enabling the model to achieve real-time and fast calculations on a micro-industrial host; S5. Tearing status monitoring: Based on the real-time sensor data collected in step S1 as input, the tearing status monitoring model optimized in step S4 is used to calculate and output the tearing status of the lower belt of the long-distance bidirectional transport belt conveyor for mining in real time, thereby realizing real-time monitoring of the belt.

2. The intelligent detection method for tearing of the lower belt of a conveyor according to claim 1 is characterized in that: In step S2, the construction of the tearing perception database includes the following steps: S2.

1. The sensor collects data once per second and transmits the data to the data acquisition terminal through the wireless module. The data of each sensor is x(t), and the time series data is: X={x(1),x(2),…,x(t)} Among them, t represents the timestamp, which means the data collected by the sensor at the tth second; S2.

2. Each sensor transmits data to the coordinator node. After receiving the data, the coordinator node stores the information in the database and adds a timestamp and sensor location mark to each data; S2.

3. After data collection, wavelet transform is first used to perform denoising to remove signal fluctuations caused by external interference, and a low-pass filter is used to filter out high-frequency noise. The transfer function is: Where f is the frequency of the signal, fc is the cutoff frequency of the filter, and the signal x~(t) obtained after filtering is the denoised data; S2.

4. The processed data are classified and stored according to time series and sensor location to form a structured database.

3. The intelligent detection method for tearing of the lower belt of a conveyor according to claim 1 is characterized in that: In step S3, the tearing state monitoring model is constructed, including the following steps: S3.

1. Extract useful feature data from the established tear perception database, focusing on belt tension, vibration frequency, temperature fluctuation, and crack development. S3.

2. Use principal component analysis to extract representative features from large-scale data and reduce the data dimension; S3.

3. Use the LSTM neural network model to train the feature data; S3.

4. During the training process, the labeled normal state and tearing state datasets are used to optimize the model parameters so that it can accurately distinguish between different states; S3.

5. After model training is completed, use the validation set to test the model to check the accuracy of the model.

4. The intelligent detection method for tearing of the lower belt of a conveyor according to claim 3 is characterized in that: In the deep learning algorithm described in step S3, the convolution kernel size k is selected as 3×3, and the learning rate η is set to 0.01 by default; The network pruning and model compression technology described in step S3.2 determines the pruning ratio p% based on the change in the accuracy of the monitoring validation set, and increases layer by layer from 20%-50%. The quantization value xquant ranges from -127 to 127.

5. The intelligent detection method for tearing of the lower belt of a conveyor according to claim 1 is characterized in that: In step S4, the tearing state monitoring model is optimized, including the following steps: S4.

1. Prune the trained LSTM model to remove neurons and connections that have little impact on the model output, thereby reducing the model's complexity and computational effort. S4.

2. Use weight pruning to prune redundant filters in the convolutional layer. S4.

3. Quantize the pruned LSTM model to reduce the storage space of the model.

6. The intelligent detection method for tearing of the lower belt of a conveyor according to claim 5, characterized in that: In the principal component analysis method described in step S4, characteristic data with eigenvalues ​​greater than 1 are selected as principal components according to the Kaiser criterion.

7. The intelligent detection method for tearing of the lower belt of a conveyor according to claim 1 is characterized in that: In step S5, the tearing state monitoring includes the following steps: S5.

1. The wireless transmission system continues to receive real-time data from each sensor via the LoRa protocol. Each data packet from the sensor contains information such as timestamp, sensor location, and data value. S5.

2. The sensor data is sent to the coordinator node through the terminal node, and the coordinator node transmits the data to the computing module; S5.

3. The calculation module uses the optimized LSTM model to analyze the received real-time data, calculate the current state of the belt, and determine whether there is a tearing risk; S5.

4. If a tear risk is detected, the calculation module will trigger an alarm, output specific alarm information, and display the tear location and risk level on the monitoring screen; S5.

5. When the system detects signs of tearing, it will immediately notify the operator through sound and light alarms or through the system to remind them to take maintenance measures in a timely manner.

8. An intelligent detection system for tearing of the lower belt of a long-distance conveyor for mining, characterized in that: include: Sensing and measurement module: includes tension sensors installed on the sides and top of the conveyor belt, a vibration sensor for detecting belt vibration, a temperature sensor, and an optical sensor installed on the belt surface; Data transmission module: including sensor signal acquisition and transmission interface, data transmission network; Tear detection module: This module uses data analysis algorithms to deeply process the collected sensor data and determine whether tearing has occurred. Alarm and control module: used to respond promptly and take emergency measures; Data storage and analysis module: It has data storage functions, provides cloud storage and local storage options, and can perform historical data analysis; Self-learning optimization module: It can optimize the system's tear detection algorithm based on historical data and real-time data to improve detection accuracy.

9. The intelligent detection system for tearing of the lower belt of a long-distance conveyor for mining according to claim 8 is characterized in that: The tear detection module includes an anomaly detection algorithm, an image processing algorithm and a deep learning algorithm.

10. The intelligent detection system for tearing of the lower belt of a long-distance conveyor for mining according to claim 8, characterized in that: The self-learning optimization module is capable of algorithm self-optimization and adaptive adjustment; The algorithm self-optimization is to accumulate historical data so that the system can continuously train and optimize the tear detection algorithm; The adaptive adjustment means that the system can automatically adjust the parameters and alarm thresholds of the detection algorithm according to environmental changes and working conditions.

Images