A mine chain feeder chain detection method and detection system

Abstract

The present application relates to a kind of mine chain feeder chain detection method and detection system, and detection system includes computer analysis system, tension detection system, laser radar detection system and video detection system.The method includes the following steps:1 obtains the tension data, texture data and shape deformation data of chain at multiple continuous time points;2 data is normalized to obtain tension feature a, texture feature b and shape feature c;3 construct data set at t time;4 input data set at t time into trained Transform-LSTM model and output the health index h of chain;5 set the health threshold H of chain, if the health index h of chain is less than the health threshold H of chain, then determine that chain is in healthy state.The present application measures tension, texture and shape of chain in real time, and can obtain the safety range and state of chain in real time through comprehensive comparison and judgment.

Description

Technical Field

[0001] This invention relates to the field of mining conveying equipment, and specifically to a method and system for detecting the chain of a mining chain feeder. Background Technology

[0002] Chain feeders occupy a crucial position in the entire conveying process, serving as the starting point for material transport. A chain breakage could have disastrous consequences. Restoring a chain feeder after an accident requires significant manpower, resources, and time, representing a fatal blow to the production system. To prevent chain breakage, traditional conventional protection devices are typically used on-site. However, these conventional devices have the following drawbacks:

[0003] Limited protection scope: Conventional protection devices are typically designed for specific types of faults or hazardous situations, and therefore their protection scope is limited. For other unforeseen faults or abnormal situations, conventional protection devices may not provide effective protection.

[0004] Response time delay: Conventional protection devices require a certain amount of time to detect faults and trigger protection measures, which may cause some delays. This is especially true for high-speed systems, where delays can have serious consequences and make it impossible to predict the chain status.

[0005] False alarms and missed alarms: Conventional protection devices may produce false alarms or missed alarms under certain conditions, leading to unnecessary shutdowns or interventions.

[0006] Complexity and high cost: Some conventional protection devices require complex engineering installation and commissioning, resulting in high costs. Especially in large industrial systems, deploying multiple different types of protection devices increases system complexity and maintenance costs. Summary of the Invention

[0007] To address the problems existing in the above-mentioned technologies, the present invention provides a method for detecting the chain of a mining chain feeder, comprising the following steps:

[0008] S1 obtains the tension data, texture data, and shape deformation data of the chain at multiple consecutive moments;

[0009] S2 normalizes the data obtained in S1 to obtain tension feature a, texture feature b, and shape feature c;

[0010] S3 constructs a dataset at time t, which includes all tension features a, texture features b, and shape features c from the M time points prior to time t;

[0011] S4 inputs the dataset at time t into the trained Transformer-LSTM model and outputs the chain health indicator h;

[0012] The Transformer-LSTM model consists of an LSTM module and a Transformer module.

[0013] S5 sets the health threshold H of the chain. If the chain's health index h is less than the health threshold H, the chain is judged to be in a healthy state.

[0014] This also includes step S41, where the dataset at time t is input into the LSTM module, and the long memory output of the LSTM module is... and short memory The data will be integrated through Flatten and Dense layers, and nonlinear calculations will be performed using the ReLU activation function to transform it into one-dimensional sequence data as an intermediate result.

[0015] The LSTM module consists of an input gate, a forget gate, and an output gate.

[0016] The input gate formula is:

[0017] ;

[0018] in, This is the output of the entry at time t. It is the input matrix at time t. The hidden state at time t-1 This is the weight matrix. For the bias term of the input gate, for Activation function;

[0019] The forgetting gate formula is:

[0020] ;

[0021] in, For the output of the forget gate, Here is the weight matrix for the forget gate. For the bias term of the forget gate, The hidden state at time t-1 It is the input matrix at time t;

[0022] The cell state update formula is:

[0023] ;

[0024] in, To hide the cell states, Wc is the weight matrix of the cell states, and bc is the bias term of the cell states. The hidden state at time t-1 It is the input matrix at time t;

[0025] The formula for the output gate is:

[0026] ;

[0027] in, For the output of the output gate, The hidden state at time t-1 is the input matrix at time t, Wo is the weight matrix of the output gate, and bo is the bias term of the output gate;

[0028] Long memory The formula is:

[0029] ;

[0030] Short memory The formula is:

[0031] ;

[0032] This represents the hidden state at time t+1.

[0033] The Transformer module consists of multiple linearly connected frames, each frame comprising four parts: input, Encoder block, Decoder block, and output.

[0034] Both the Encoder block and the Decoder block are composed of a Multi-Head Attention module and a fully connected neural network.

[0035] This also includes step S42, which adds positional encoding to each data point in the dataset at time t and the intermediate results.

[0036] The formula for Positional Encoding is:

[0037]

[0038] Where pos represents the absolute position of the current value in the time series, pos=1,2,3…, The dimension of the input sequence data is represented by 2i and 2i+1, which represent parity.

[0039] In step S43, the Encoder block transforms the input data into matrix A by the Multi-HeadAttention module, performs residual convolution and standardization on matrix A, performs linear and nonlinear transformations through a fully connected neural network, and outputs matrix A'.

[0040] In the Decoder block, the Multi-Head Attention module receives matrix A', concatenates matrix A' with the intermediate result after position encoding, and then transforms it through Multi-Head Attention and a fully connected neural network to obtain matrix B'.

[0041] In this context, the dataset at time t after position encoding serves as the input to the Encoder block of the first frame, while the Encoder blocks of the remaining frames receive the output of the Decoder block of the previous frame as input.

[0042] The matrix B' output by the last frame is linearly transformed to obtain a one-dimensional vector, and the health index h is obtained by the ReLU activation function.

[0043] The tension data is obtained by a tension sensor installed on the chain.

[0044] The texture data is obtained in the following way:

[0045] A line laser is used to scan the object under test, and a camera captures two-dimensional images of the deformed laser lines on the chain surface. At the same time, a laser sensor emits a laser beam and measures the flight time or phase information of the light from the laser to the chain surface. Then, the deformed laser lines are processed for data denoising, and the three-dimensional point cloud data of the chain is calculated by a line laser scanning algorithm.

[0046] The method for obtaining shape deformation data is as follows:

[0047] The chain image is acquired, and then the YOLOv5 instance segmentation algorithm is used to identify the chain image to obtain the chain's feature points or edge information. The shape features of the chain are extracted and calculated, and the shape features are compared with the reference image. The shape deformation range data is obtained by calculating the difference between the shape features and the reference image.

[0048] The morphological characteristics include the length and / or angle and / or degree of bending of the chain.

[0049] This application also provides a chain detection system for a mining chain feeder, including a computer analysis system, a tension detection system, a lidar detection system, and a video detection system, wherein the computer analysis system is connected to the tension detection system, the lidar detection system, and the video detection system, respectively.

[0050] The beneficial effects of this invention are as follows: This invention uses sensor, video, and lidar detection technologies to measure the tension, texture, and shape of the chain in real time. Through comprehensive comparison and judgment, the safe range and status of the chain are obtained. Real-time monitoring can prevent downtime accidents caused by chain breakage, thereby improving the service life and operating efficiency of the equipment. At the same time, this detection method is characterized by strong real-time performance, high reliability, and ease of implementation. Attached Figure Description

[0051] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0052] Figure 1 This is a schematic diagram of the system composition of the present invention.

[0053] Explanation of reference numerals in the attached figures

[0054] 1. Computer analysis system; 2. Tension detection system; 21. Fixed frame; 22. Tension sensor; 23. Driven wheel tensioning screw; 3. LiDAR detection system; 4. Video detection system. Detailed Implementation

[0055] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art can make similar extensions without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0056] like Figure 1 The image shown is the detection system of this application;

[0057] The system includes: a computer analysis system 1, a tension detection system 2, a lidar detection system 3, and a video detection system 4. The computer analysis system 1 is connected to the tension detection system 2, the lidar detection system 3, and the video detection system 4, respectively.

[0058] Specifically;

[0059] The tension detection system 2 includes: a fixed frame 21, a tension sensor 22, and a driven wheel tensioning screw 23. The fixed frame 21 is composed of a feeder body baffle structure and is used to fix the tension sensor 22. The tension sensor 22 is a key component for measuring tension. It is installed on the fixed frame 21 and is used to measure the deformation or stress of an object caused by tension, and then convert it into an electrical signal. The driven wheel tensioning screw 23 is composed of a screw and an adjusting nut and is used to adjust the mechanical tension so that the tension sensor 22 is subjected to reasonable and uniform force. The signal processor is responsible for receiving the electrical signal from the tension sensor 22 and converting it into a readable digital signal, which is then connected to the computer analysis system 1 via Ethernet.

[0060] Computer analysis system 1 includes: a processor, memory, a graphics processor, a storage device, and a monitor. The processor is a multi-core high-performance processor; the memory is large-capacity, high-speed memory; the graphics processor is a high-performance dedicated graphics card; the storage device uses a fast solid-state drive as the system disk and a large-capacity hard drive for data storage; the monitor is a high-resolution monitor, and the operating system is Windows. Computer analysis system 1 receives and analyzes real-time data and image information collected by tension detection system 2, lidar detection system 3, and video detection system 4.

[0061] The lidar detection system 3 includes a lidar transceiver and a processing module. The lidar transceiver illuminates the target object with a high-intensity laser beam, receives the laser pulses reflected back from the target object, and measures distance and other properties. The lidar transceiver scans the laser beam to cover the detection area. The processing module detects and quantifies the received laser pulse signals from the lidar transceiver, converts them into digital signals for further processing, calculates the distance to the target object by measuring the time from laser pulse emission to reception, and then connects to the computer analysis system 1 via Ethernet.

[0062] The video detection system 4 includes: a video acquisition device and a video analysis and processing module; the video analysis and processing module is connected to the video acquisition device via a serial cable. The video analysis and processing module optimizes, identifies, and classifies the video data acquired by the video acquisition device, uses a deep learning model to associate the target with a specific category or label, extracts information about the target's behavior, features, and attributes, and connects to the computer analysis system 1 via Ethernet.

[0063] The method for detecting the chain of a mining chain feeder provided in this application includes the following steps:

[0064] S1 obtains the tension data, texture data, and shape deformation data of the chain at multiple consecutive moments;

[0065] Tension data is obtained from tension sensors installed on the chain;

[0066] The method for acquiring texture data using the LiDAR detection system 3 is as follows:

[0067] The linear laser scanning algorithm, including data acquisition, point cloud processing, and surface reconstruction, is employed. A linear laser is used to scan the object under test, and a camera acquires two-dimensional images of the deformed laser lines on the chain surface. Simultaneously, a laser sensor emits a laser beam and measures the flight time or phase information of the light from the laser to the chain surface. Then, the deformed laser lines are denoised, and the three-dimensional point cloud data of the chain is calculated using the linear laser scanning algorithm.

[0068] The method for collecting shape deformation data using video detection system 4 is as follows:

[0069] The chain image is acquired, and then the YOLOv5 instance segmentation algorithm is used to identify the chain image to obtain the chain's feature points or edge information. The shape features of the chain are extracted and calculated, and the shape features are compared with the reference image. The shape deformation range data is obtained by calculating the difference between the shape features and the reference image.

[0070] Shape characteristics include the length and / or angle and / or degree of bending of the chain.

[0071] S2 normalizes the data obtained in S1 to obtain tension feature a, texture feature b, and shape feature c;

[0072] S3 constructs a dataset at time t, which includes all tension features a, texture features b, and shape features c from the M time points prior to time t;

[0073] S4 inputs the dataset at time t into the trained Transformer-LSTM model and outputs the chain health indicator h;

[0074] The Transformer-LSTM model consists of an LSTM module and a Transformer module.

[0075] Specifically, in step S41, after inputting the dataset at time t into the LSTM module, the long memory output by the LSTM module is... and short memory The data will be integrated through Flatten and Dense layers, and nonlinear calculations will be performed using the ReLU activation function to transform it into one-dimensional sequence data as an intermediate result.

[0076] An LSTM module consists of an input gate, a forget gate, and an output gate.

[0077] The input gate formula is:

[0078] ;

[0079] in, This is the output of the entry at time t. The hidden state at time t-1 This is the weight matrix. For the bias term of the input gate, for Activation function It is the input matrix at the current time t, that is, the matrix composed of tension feature a, texture feature b and shape feature c from the M times before time t, which is also the matrix composed of the dataset at time t.

[0080] The forgetting gate formula is:

[0081] ;

[0082] in, For the output of the forget gate, Here is the weight matrix for the forget gate. For the bias term of the forget gate, The hidden state at time t-1 It is the input matrix at time t;

[0083] The cell state update formula is:

[0084] ;

[0085] in, To hide the cell states, Wc is the weight matrix of the cell states, and bc is the bias term of the cell states. The hidden state at time t-1 It is the input matrix at time t;

[0086] The formula for the output gate is:

[0087] ;

[0088] in, For the output of the output gate, The hidden state at time t-1 is the input matrix at time t, Wo is the weight matrix of the output gate, and bo is the bias term of the output gate;

[0089] Long memory The formula is:

[0090] ;

[0091] Short memory The formula is:

[0092] ;

[0093] This represents the hidden state at time t+1.

[0094] During detection, taking M=10 as an example, and setting time N as time 11, then during the first detection, i.e. the detection at time N, random initialization is performed to form a matrix with the same format as the input matrix at time xt mentioned earlier. Each value in this matrix is ​​a random number between 0 and 1. This matrix represents the hidden state at time N-1.

[0095] The Transformer module consists of multiple linearly connected frames, which include four parts: input, encoder block, decoder block, and output.

[0096] Both the Encoder block and the Decoder block consist of a Multi-Head Attention module and a fully connected neural network.

[0097] Step S42: Add positional encoding to each data point in the dataset at time t and the intermediate results;

[0098] The formula for Positional Encoding is:

[0099]

[0100] Where pos represents the absolute position of the current value in the time series, pos=1,2,3…, The dimension of the input sequence data is represented by 2i and 2i+1, which represent parity.

[0101] In step S43, the Encoder block transforms the input data into matrix A by the Multi-Head Attention module, performs residual convolution and standardization on matrix A, and performs linear and nonlinear transformations through a fully connected neural network to output matrix A'.

[0102] In the Decoder block, the Multi-Head Attention module receives matrix A', concatenates matrix A' with the intermediate result after position encoding, and then transforms it through Multi-Head Attention and a fully connected neural network to obtain matrix B'.

[0103] In this context, the dataset at time t after position encoding serves as the input to the Encoder block of the first frame, while the Encoder blocks of the remaining frames receive the output of the Decoder block of the previous frame as input.

[0104] The matrix B' of the last frame is linearly transformed to obtain a one-dimensional vector, and the health index h is obtained by the ReLU activation function.

[0105] S5 sets the chain's health threshold H. If the chain's health index h is less than the chain's health threshold H, the chain is considered to be in a healthy state. If the chain is in an unhealthy state, the chain is adjusted.

[0106] For training the model of this application, datasets at multiple time points can be collected in step S3. The datasets at multiple time points are divided into training set and test set in a ratio of 8:2. The above model is trained using the training set. The model structure and parameters are adjusted to achieve the best fitting effect. The model prediction effect is evaluated using the test set after each training. Finally, the model with high prediction accuracy and stable performance is selected as the final application model.

[0107] Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

Claims

1. A method for detecting the chain of a mining chain feeder, characterized in that, Includes the following steps: S1 obtains the tension data, texture data, and shape deformation data of the chain at multiple consecutive moments; S2 normalizes the data obtained in S1 to obtain tension feature a, texture feature b, and shape feature c; S3 constructs a dataset at time t, which includes all tension features a, texture features b, and shape features c from the M time points prior to time t; S4 inputs the dataset at time t into the trained Transformer-LSTM model and outputs the chain health indicator h; The Transformer-LSTM model consists of an LSTM module and a Transformer module. S5 sets the health threshold H of the chain. If the health index h of the chain is less than the health threshold H, the chain is judged to be in a healthy state. The Transformer module consists of multiple linearly connected frames, each frame comprising four parts: input, Encoder block, Decoder block, and output. Both the Encoder block and the Decoder block are composed of a Multi-Head Attention module and a fully connected neural network. In the Encoder block, the input data is transformed into matrix A by the Multi-Head Attention module. Residual convolution and standardization are performed on matrix A, and linear and nonlinear transformations are performed through a fully connected neural network to output matrix A'. In the Decoder block, the Multi-Head Attention module receives matrix A', concatenates matrix A' with the intermediate result after position encoding, and then transforms it through Multi-Head Attention and a fully connected neural network to obtain matrix B'. In this context, the dataset at time t after position encoding serves as the input to the Encoder block of the first frame, while the Encoder blocks of the remaining frames receive the output of the Decoder block of the previous frame as input. The matrix B' output by the last frame is linearly transformed to obtain a one-dimensional vector, and the health index h is obtained by the ReLU activation function.

2. The method for detecting the chain of a mining chain feeder according to claim 1, characterized in that, It also includes step S41, which involves inputting the dataset at time t into the LSTM module and then processing the long memory output of the LSTM module. and short memory The data will be integrated through Flatten and Dense layers, and nonlinear calculations will be performed using the ReLU activation function to transform it into one-dimensional sequence data as an intermediate result.

3. The method for detecting the chain of a mining chain feeder according to claim 2, characterized in that, An LSTM module consists of an input gate, a forget gate, and an output gate. The input gate formula is: ； in, This is the output of the entry at time t. It is the input matrix at time t. The hidden state at time t-1 This is the weight matrix. For the bias term of the input gate, for Activation function; The forgetting gate formula is: ； in, For the output of the forget gate, Here is the weight matrix for the forget gate. For the bias term of the forget gate, The hidden state at time t-1 It is the input matrix at time t; The cell state update formula is: ； in, To hide the cell states, Wc is the weight matrix of the cell states, and bc is the bias term of the cell states. The hidden state at time t-1 It is the input matrix at time t; The formula for the output gate is: ； in, For the output of the output gate, The hidden state at time t-1 is the input matrix at time t, Wo is the weight matrix of the output gate, and bo is the bias term of the output gate; Long memory The formula is: ; Short memory The formula is: ； This represents the hidden state at time t+1.

4. The method for detecting the chain of a mining chain feeder according to claim 2, characterized in that, It also includes step S42, which adds positional encoding to each data point in the dataset at time t and the intermediate results; The formula for Positional Encoding is: ， Where pos represents the absolute position of the current value in the time series, pos=1,2,3…, The dimension of the input sequence data is represented by 2i and 2i+1, which represent parity.

5. The method for detecting the chain of a mining chain feeder according to claim 1, characterized in that, The tension data is obtained by a tension sensor installed on the chain.

6. The method for detecting the chain of a mining chain feeder according to claim 1, characterized in that, The texture data is obtained in the following way: A line laser is used to scan the object under test, and a camera captures two-dimensional images of the deformed laser lines on the chain surface. At the same time, a laser sensor emits a laser beam and measures the flight time or phase information of the light from the laser to the chain surface. Then, the deformed laser lines are processed for data denoising, and the three-dimensional point cloud data of the chain is calculated by a line laser scanning algorithm.

7. The method for detecting the chain of a mining chain feeder according to claim 1, characterized in that, The method for obtaining shape deformation data is as follows: The chain image is acquired, and then the YOLOv5 instance segmentation algorithm is used to identify the chain image to obtain the chain's feature points or edge information. The shape features of the chain are extracted and calculated, and the shape features are compared with the reference image. The shape deformation range data is obtained by calculating the difference between the shape features and the reference image. The shape features include the length and / or angle and / or degree of curvature of the chain.

8. A chain detection system for a mining chain feeder, characterized in that, It includes a computer analysis system, a tension detection system, a lidar detection system, and a video detection system, wherein the computer analysis system is connected to the tension detection system, the lidar detection system, and the video detection system, respectively; The computer analysis system constructs a dataset at time t based on tension data from the tension detection system, texture data from the lidar detection system, and shape deformation data from the video detection system. The dataset at time t is then input into the trained Transformer-LSTM model, which outputs the chain's health index h. The Transformer-LSTM model consists of an LSTM module and a Transformer module. The Transformer module consists of multiple linearly connected frames, each frame comprising four parts: input, Encoder block, Decoder block, and output. Both the Encoder block and the Decoder block are composed of a Multi-Head Attention module and a fully connected neural network. In the Encoder block, the input data is transformed into matrix A by the Multi-Head Attention module. Residual convolution and standardization are performed on matrix A, and linear and nonlinear transformations are performed through a fully connected neural network to output matrix A'. In the Decoder block, the Multi-Head Attention module receives matrix A', concatenates matrix A' with the intermediate result after position encoding, and then transforms it through Multi-Head Attention and a fully connected neural network to obtain matrix B'. In this context, the dataset at time t after position encoding serves as the input to the Encoder block of the first frame, while the Encoder blocks of the remaining frames receive the output of the Decoder block of the previous frame as input. The matrix B' output by the last frame is linearly transformed to obtain a one-dimensional vector, and the health index h is obtained by the ReLU activation function.

Images