Deep learning-based belt deviation adjusting method for belt conveyor

Abstract

The invention provides a belt conveyor belt deviation adjusting method based on deep learning, and belongs to the technical field of belt deviation adjusting. The method comprises the steps that non-contact distance measuring sensor arrays are symmetrically arranged along the two sides of a belt conveyor, and belt edge distance data are collected in real time; preprocessing the belt edge distance data to obtain preprocessed data; extracting spatial distribution features and time sequence features of the preprocessed data, and performing feature fusion to obtain fused features; generating a deviation adjustment action instruction through a deep learning model based on the fusion features; and the execution mechanism executes the deviation adjustment action instruction to complete belt deviation adjustment. According to the method, the strategy is optimized through the reward function, meanwhile, the control function of the PID controller is optimized through the preset target function, the deviation adjustment action instruction is further optimized, and the loss of the belt is reduced while precise deviation adjustment of belt deviation adjustment is achieved.

Description

Technical Field

[0001] This invention relates to the field of belt alignment, and in particular to a deep learning-based method for belt alignment of a belt conveyor. Background Technology

[0002] Belt conveyors are crucial continuous material handling equipment in industrial production, widely used in mining, ports, power, metallurgy, and other fields. Their core component, the conveyor belt, is prone to deviation during operation due to factors such as uneven tension, uneven material loading, idler roller malfunction, or installation errors. This can lead to material spillage, belt edge wear, and even safety accidents such as equipment jamming, tearing, or fire. Therefore, real-time and accurate deviation detection and effective adjustment are core technical requirements for ensuring stable equipment operation.

[0003] Currently, belt misalignment detection mainly relies on mechanical limit switches, photoelectric switches, or vision-based inspection methods. Mechanical limit switches trigger signals via contact baffles, offering a simple structure but prone to belt wear and exhibiting lag in response. While photoelectric switches provide non-contact detection, they only offer binary signals, failing to quantify the degree of misalignment and exhibiting poor anti-interference capabilities. Vision-based methods acquire misalignment through image processing, but are limited by ambient lighting, dust, and algorithm complexity, resulting in insufficient real-time performance and robustness. Misalignment control often relies on threshold judgments from the aforementioned detection technologies or is driven by manual experience, employing simple automatic misalignment devices (such as electric push rod self-aligning rollers), which suffer from problems such as misalignment lag, low accuracy, and repetitive oscillations.

[0004] The existing technology has the following shortcomings: (1) Detection level: mechanical and photoelectric switches cannot provide continuous information on belt deviation, and visual detection has weak resistance to environmental interference; (2) Single-point measurement is difficult to reflect the overall belt deviation status and is easily affected by material edge shaking and noise interference; (3) Insufficient dynamic optimization, the causes of deviation are complex, and traditional threshold control is difficult to adaptively adjust the adjustment parameters, resulting in under-adjustment or over-adjustment; (4) Lack of intelligent analysis of the time-series deviation trend, unable to predict the risk of deviation deterioration and make timely decisions on adjustment strategies.

[0005] Therefore, a deep learning-based method for belt alignment of conveyors is needed. Summary of the Invention

[0006] In view of this, the present invention provides a belt misalignment method for belt conveyors based on deep learning. It uses a laser rangefinder or ultrasonic rangefinder to obtain real-time belt edge distance information, and combines deep learning technology to achieve accurate, robust, and real-time online detection of belt misalignment, and intelligently generates the optimal misalignment action command, which significantly improves the accuracy and efficiency of misalignment.

[0007] Therefore, the present invention provides the following technical solution: A deep learning-based method for belt alignment in a belt conveyor includes: A non-contact ranging sensor array is symmetrically deployed on both sides of the belt conveyor to collect real-time data on the distance between the belt edges. The belt edge distance data is preprocessed to obtain preprocessed data; The spatial distribution features and time series features of the preprocessed data are extracted, and feature fusion is performed to obtain fused features; Based on the fusion features, a deep learning model is used to generate adjustment action instructions; The actuator executes the adjustment action command to complete the belt adjustment.

[0008] Furthermore, based on the fused features, a bias adjustment action command is generated using a deep learning model, including: Based on fusion features, a policy network is used to generate bias adjustment action instructions; The control coefficients of the PID controller are optimized with the objective functions of minimizing cumulative offset and minimizing mechanical loss. Update the deviation adjustment command using the optimized control coefficients; The updated bias adjustment action instruction is executed, and the reward function is calculated and fed back to the policy network to optimize the policy network's generation of bias adjustment action instructions.

[0009] Furthermore, the objective function is:

[0010] in, To optimize the cycle, for Belt offset at any moment λ represents the action performed at time t, and λ is the action penalty coefficient, which is dynamically adjusted based on historical wear data. , and These are the control coefficients for the PID controller.

[0011] Furthermore, the reward function:

[0012] in, for Belt offset at any moment I represents the amplitude of the action, and II represents the steady-state indication function. , and These are the weighting coefficients. For dynamic threshold .

[0013] Furthermore, the dynamic threshold:

[0014] in, This represents the average historical offset. The deviation is the standard deviation. These are adaptive coefficients.

[0015] Further, the step of extracting the spatial distribution features and time series features of the preprocessed data, and performing feature fusion to obtain fused features, includes: Spatial distribution features of multiple sensors are extracted based on preprocessed data using a one-dimensional convolutional neural network. Long Short-Term Memory (LSTM) networks are used to extract time-series features of each sensor based on preprocessed data. The fused features are obtained by weighted summation of spatial distribution features and time series features.

[0016] Further, the belt edge distance data is preprocessed to obtain preprocessed data, including: Wavelet transform was used to filter out belt vibration noise; Sensor data is reconstructed using a variational autoencoder, and data points with reconstruction errors exceeding a threshold are discarded.

[0017] Advantages and positive effects of the present invention: This method uses laser or ultrasonic ranging sensors for distance detection, reducing external interference and improving the accuracy of raw data. Preprocessing of the raw data further enhances the accuracy of the detected data. The spatial distribution and time-series features of the preprocessed data are then fused as input to a deep learning model to generate belt alignment commands, and the strategy is optimized using a reward function. Simultaneously, the control function of the PID controller is optimized using a preset objective function, further refining the belt alignment commands and achieving precise belt alignment while reducing belt wear. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart of the belt alignment method for a belt conveyor based on deep learning in Embodiment 1 of the present invention; Figure 2 This is a structural diagram of the belt alignment method for a belt conveyor based on deep learning in Embodiment 1 of the present invention; Figure 3This is a schematic diagram of the sensor installation on the belt conveyor in Embodiment 1 of the present invention. Detailed Implementation

[0020] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0021] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0022] This invention provides a deep learning-based belt alignment method for conveyor belts. The method optimizes the control parameters of a PID controller based on the offset at future time steps using a deep learning model. The optimized control parameters are then used to control the actuator. The execution results of the actuator are fed back to the deep learning model as a reward function to further optimize the policy network.

[0023] Example 1 A deep learning-based method for belt alignment in a belt conveyor, characterized by the following steps: Along both sides of the belt conveyor, a pair of high-precision laser or ultrasonic rangefinder sensors are deployed every 30 meters, with the installation angle perpendicular to the edge section of the belt.

[0024] The actuator is equipped with a hydraulic servo-controlled alignment roller system with a stroke of ±15°, a pressure range of 0-10MPa, and a response time of <100ms.

[0025] A central server is deployed to store historical data and run model training clusters and edge computing devices. Multiple sensors use one edge computing device, and the central server and edge computing devices work together through federated learning.

[0026] S1. Real-time data acquisition of belt edge distance is achieved through a range sensor array, with a synchronous sampling frequency of 120Hz; S2. Preprocess the distance data collected by the laser rangefinder.

[0027] S3. Extract the spatial distribution features and time series features of the preprocessed data, and perform feature fusion to obtain fused features.

[0028] 1) Extract the spatial distribution features of multiple sensors based on the preprocessed distance data sequence using a one-dimensional convolutional neural network.

[0029] 2) Long Short-Term Memory (LSTM) network is used to extract time series features of each sensor based on preprocessed distance data.

[0030] 3) Dynamically assign weights to time series features and spatial distribution features through an attention mechanism, and then sum them up by weight to obtain fused features.

[0031] Input data for a one-dimensional convolutional neural network: That is, 1 sample, 8 sensors, 360 time points of data: data is collected for 3 seconds at a sampling frequency of 120Hz.

[0032] Output data: S= S is a 128-dimensional feature vector, representing spatial distribution characteristics; Input data for Long Short-Term Memory (LSTM) networks: This represents 360 time data points of a sensor.

[0033] Output data: , A 64-dimensional feature vector represents the time-series features of a sensor.

[0034] Spatial distribution features and time series features are fused through an attention mechanism.

[0035] Dynamic weight calculation: Let spatial weight coefficients be set. Time weighting coefficient , .

[0036] Fusion characteristics:

[0037] Trigger conditions are dynamically allocated based on preset non-class settings. and In this embodiment, it includes: The sudden impact of a large coal lump causes the conveyor belt to deviate. Triggering conditions: sensor reading standard deviation greater than 85mm or belt offset greater than 210mm; spatial weighting coefficient. The value is 0.92, which is the time weighting coefficient. The value is 0.18.

[0038] Tape aging and deviation, triggering conditions: offset acceleration greater than 0.05, spatial weighting coefficient. The value is 0.22, which is the time weighting coefficient. The value is 0.78.

[0039] S4. The DDPG (Deep Deterministic Policy Gradient) algorithm is adopted to generate adjustment action instructions through the policy network based on the fusion features, guide the execution mechanism to complete the adjustment, and feed the execution results back to the policy network to optimize the strategy.

[0040] 1) Adjustment action commands, including: adjustment roller angle, pressure and direction.

[0041] 2) The reward function is defined as:

[0042] in, for Belt offset at any moment For the range of motion, , and These are the weighting coefficients. Ⅱ is the dynamic threshold; Ⅱ is the steady-state indication function:

[0043] Dynamic threshold The formula is expressed as:

[0044] in, This represents the average historical offset. The deviation is the standard deviation. The adaptive coefficients are dynamically adjusted based on the predicted offset trend over the next 3 seconds using a long short-term memory network. value.

[0045] initialization Value Dynamically correct according to preset working condition classification ,include: If a large impact occurs suddenly on the belt conveyor, then ; If the belt conveyor maintains a stable tendency to run off-center, then... .

[0046] 3) Calculate the adjustment action amount using a PID controller:

[0047] in, , and The control coefficients of the PID controller. To adjust the amount of motion.

[0048] By optimizing the control coefficients through reinforcement learning, and considering both ensuring the stability of the belt conveyor system and extending equipment life, the optimization objectives are constructed as minimizing cumulative offset and minimizing mechanical loss, expressed by the following formula:

[0049] in, To optimize the cycle, for Belt offset at any moment λ represents the range of motion, and λ is the motion penalty coefficient, which is dynamically adjusted based on historical wear data.

[0050] S5. Simultaneously set early warning conditions to detect whether the transmitter is faulty in real time.

[0051] When the sensor data reconstruction error is greater than 3 within the preset continuous sampling period If so, a fault warning will be triggered; If the rate of change of offset is less than the set threshold after the adjustment action is performed, a fault warning will be triggered.

[0052] Example 2 A deep learning-based method for belt alignment in a belt conveyor includes the following steps: S1. Real-time data acquisition of belt edge distance is achieved by a non-contact ranging sensor array symmetrically deployed on both sides of the belt conveyor, with a sampling frequency ≥100Hz.

[0053] Non-contact ranging sensors are laser rangefinders or ultrasonic rangefinders, which are deployed in pairs along the belt conveyor at intervals of 20-30 meters, with the installation angle perpendicular to the edge cross-section of the belt being measured.

[0054] The alignment mechanism is a servo motor-driven alignment roller device.

[0055] S2. Preprocess the collected data to obtain a preprocessed dataset: Wavelet transform was used to filter out belt vibration noise; Sensor data is reconstructed using a variational autoencoder, and data points with reconstruction errors exceeding a threshold are discarded.

[0056] S3. Extract the spatial distribution features and time series features of the preprocessed data, and perform feature fusion to obtain fused features.

[0057] S4. The DDPG (Deep Deterministic Policy Gradient) algorithm is adopted to generate adjustment action instructions through the policy network based on the fusion features, guide the execution mechanism to complete the adjustment, and feed the execution results back to the policy network to optimize the strategy.

[0058] 1) Generate the deviation adjustment action command based on the fusion features through the policy network: (Actuator 01, +7.8°, 0.5MPa) The optimized PID control parameters are obtained by minimizing the cumulative offset and minimizing the mechanical loss as objective functions; the adjustment action command is determined based on the optimized PID control parameters.

[0059] 2) The actuator executes the adjustment command and calculates the reward function: By collecting sensor data and performing rolling calculations on the data from the most recent hour, a dynamic threshold and reward function are calculated. time =300ms, offset =3.7mm, range of motion =+1.5°, steady state indicator function Ⅱ=0, , and The weighting coefficients are set to 0.7, 0.2, and 0.1 respectively, and the dynamic threshold is... =3.1mm; T is the optimization period, typically 10-60 seconds.

[0060] Historical offset mean =2.1mm, standard deviation of offset =0.8mm, adaptive coefficient =2.0, then the dynamic threshold =3.1mm, in this embodiment, the offset of 3.7mm > 3.1mm triggers the adjustment action, and the reward function... =-2.89.

[0061] 3) Feed the reward function back to the policy network to optimize the strategy.

[0062] Compared with traditional methods, it has more advantages under the same offset sequence.

[0063] Assuming a fixed threshold =5mm, then at time =300ms, offset When the threshold is 3.7mm, traditional designs offer no penalty for movement. Dynamic threshold. When predicting the rate of increase in offset, the threshold is automatically tightened (3.1mm < 5mm) to avoid incorrectly awarding rewards at 3.7mm.

[0064] This method enables precise adjustment of belt conveyors.

[0065] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for belt alignment of a belt conveyor based on deep learning, characterized in that, include: A non-contact ranging sensor array is symmetrically deployed on both sides of the belt conveyor to collect real-time data on the distance between the belt edges. The belt edge distance data is preprocessed to obtain preprocessed data; The spatial distribution features and time series features of the preprocessed data are extracted, and feature fusion is performed to obtain fused features; Based on the fusion features, a deep learning model is used to generate adjustment action instructions; The actuator executes the adjustment action command to complete the belt adjustment.

2. The method according to claim 1, characterized in that, Based on the fused features, a deep learning model is used to generate adjustment action instructions, including: Based on fusion features, a policy network is used to generate bias adjustment action instructions; The control coefficients of the PID controller are optimized with the objective functions of minimizing cumulative offset and minimizing mechanical loss. Update the deviation adjustment command using the optimized control coefficients; The updated bias adjustment action instruction is executed, and the reward function is calculated and fed back to the policy network to optimize the policy network's generation of bias adjustment action instructions.

3. The method according to claim 2, characterized in that, The objective function is: in, To optimize the cycle, for Belt offset at any moment λ represents the action performed at time t, and λ is the action penalty coefficient, which is dynamically adjusted based on historical wear data. , and These are the control coefficients for the PID controller.

4. The method according to claim 2, characterized in that, The reward function: in, for Belt offset at any moment I represents the amplitude of the action, and II represents the steady-state indication function. , and These are the weighting coefficients. For dynamic threshold .

5. The method according to claim 4, characterized in that, The dynamic threshold: in, This represents the average historical offset. The deviation is the standard deviation. These are adaptive coefficients.

6. The method according to claim 1, characterized in that, The step of extracting the spatial distribution features and time series features of the preprocessed data and performing feature fusion to obtain fused features includes: Spatial distribution features of multiple sensors are extracted based on preprocessed data using a one-dimensional convolutional neural network. Long Short-Term Memory (LSTM) networks are used to extract time-series features of each sensor based on preprocessed data. The fused features are obtained by weighted summation of spatial distribution features and time series features.

7. The method according to claim 1, characterized in that, Preprocessing the belt edge distance data to obtain preprocessed data includes: Wavelet transform was used to filter out belt vibration noise; Sensor data is reconstructed using a variational autoencoder, and data points with reconstruction errors exceeding a threshold are discarded.

Images