Belt conveyor dynamic tension intelligent regulation and control method and system based on reinforcement learning

Abstract

The invention discloses an intelligent regulation and control method and system for dynamic tension of a belt conveyor based on reinforcement learning. The method comprises the following steps that S1, multi-source operation signals of the belt conveyor are obtained and preprocessed; s2, inputting the preprocessed multi-source operation signal into a feature fusion network, realizing time sequence and frequency domain feature fusion through sensing fusion and modal decoupling, and outputting a multi-modal fusion feature; and S3, constructing a reinforcement learning module, carrying out strategy evolution based on the dynamic change of the weight of the multi-modal fusion feature, and dynamically optimizing a reward function. The invention relates to the technical field of belt conveyor control. According to the intelligent regulation and control method and system for the dynamic tension of the belt conveyor based on reinforcement learning, through organic combination of multi-source signal fusion, reinforcement learning optimization and an online self-adaptive mechanism, the fixity and hysteresis of a traditional tension control method are overcome, and intelligent, precise and robust regulation and control of the dynamic tension of the belt conveyor are achieved.

Description

Technical Field

[0001] This invention relates to the field of belt conveyor control, and more specifically, to a method and system for intelligent dynamic tension control of belt conveyors based on reinforcement learning. Background Technology

[0002] Belt conveyors are key equipment for the continuous transport of materials in industries such as coal mining, ports, power, and metallurgy. Their operational stability is closely related to the control of conveyor belt tension. Excessive belt tension can easily lead to belt deformation, joint damage, and shortened service life; insufficient tension may cause slippage, misalignment, or even conveying interruption. Therefore, achieving precise tension control of belt conveyors under dynamic operating conditions is a core issue affecting their safe and efficient operation.

[0003] Currently, traditional belt conveyor tension control mainly relies on empirical settings or PID-based fixed-parameter control methods. These methods are ill-suited to complex and variable operating conditions. For example, when the conveyor belt is in a start-stop, heavy-load, or material-impact state, tension fluctuates drastically, and fixed-parameter control cannot reflect the system's nonlinear and time-varying characteristics in real time, leading to control lag or over-adjustment. Furthermore, in actual operation, multi-source signals suffer from noise, drift, or sensor malfunctions, further affecting the accuracy and reliability of tension control.

[0004] With the development of artificial intelligence technology, signal modeling methods based on deep learning and adaptive decision-making mechanisms based on reinforcement learning have been gradually introduced into the field of industrial control. Their advantage lies in achieving dynamic optimal control of nonlinear systems through multi-source data fusion and policy self-evolution. However, existing research mostly focuses on single-signal modeling or static control optimization, lacking an integrated intelligent method for multi-modal signal feature fusion and dynamic tension adaptive regulation of belt conveyors. This makes it difficult to maintain system stability and safety under sensor anomalies or complex operating conditions. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for intelligent dynamic tension control of belt conveyors based on reinforcement learning. This invention addresses the problem that existing research focuses on single signal modeling or static control optimization, and lacks an integrated intelligent method for multimodal signal feature fusion and dynamic tension adaptive control of belt conveyors. This makes it difficult to maintain the stability and safety of the system under sensor malfunctions or complex operating conditions, and thus fails to meet the application requirements.

[0006] This invention achieves the above objective through the following technical solution: a method for intelligent dynamic tension control of a belt conveyor based on reinforcement learning, the method comprising the following steps: S1. Acquire and preprocess the multi-source operating signals of the belt conveyor; S2. Input the preprocessed multi-source operating signal into the feature fusion network, and realize the fusion of time-series and frequency-domain features through sensor fusion and mode decoupling, and output multi-modal fusion features; S3. Construct a reinforcement learning module, perform policy evolution based on the dynamic weight changes of the multimodal fusion features, dynamically optimize the reward function, and generate a tension control strategy that adapts to the complex working conditions of the belt conveyor. S4. Based on the tension control strategy, the dynamic tension of the belt conveyor is controlled in real time, and the parameters are reconfigured through an online self-learning mechanism. When the sensor is abnormal or the signal drifts, the control strategy is automatically corrected to ensure that the conveyor belt always runs within the safe tension range.

[0007] Furthermore, in step S1, acquiring and preprocessing the multi-source operating signals of the belt conveyor includes: Multi-source operating signals are collected synchronously by sensor arrays deployed at key parts of the belt conveyor; The multi-source operating signals include at least the conveyor belt longitudinal tension signal, equipment vibration signal, conveyor belt running speed signal, and motor current signal; Each sensor synchronously collects data at a preset sampling frequency to form a raw signal set; The original signals of each dimension are subjected to adaptive preprocessing, including removing interference data by using outlier removal methods and eliminating the dimensional differences between different signals by using standardization methods.

[0008] Furthermore, the preset sampling frequency ranges from 100Hz to 1000Hz and is dynamically adjusted according to the operating speed of the belt conveyor. The signal acquisition duration of the original signal set is dynamically adjusted according to the operating conditions of the belt conveyor. When the load is stable, it is 10s-30s, and when starting, stopping or material impact occurs, it is extended to 30s-60s. The outlier removal method uses the 3σ criterion, and the standardization method uses Z-score standardization.

[0009] Furthermore, in step S2, the preprocessed multi-source operating signal is input into the feature fusion network. Through sensor fusion and modal decoupling, time-series and frequency-domain feature fusion is achieved, outputting multi-modal fused features, including: The standardized multi-source operating signals are segmented according to a preset time window to obtain a set of signal segments; The modal decoupling module performs modal separation on each signal segment to obtain the unique characteristics of each mode; Each modality-specific feature is synchronously input into a joint feature fusion network for cross-modal feature fusion, and time-dependent features and frequency-domain dynamic features are obtained respectively. Multimodal fusion features of each signal segment are obtained by feature concatenation. An adaptive weight allocation mechanism is used to dynamically determine the weight coefficients of each segment feature, and the global multimodal fusion feature is obtained by weighted summation.

[0010] Furthermore, the preset time window length ranges from 50 to 200 time steps to ensure that each segment can fully cover the running time corresponding to 1-2 roller circumferences of the conveyor belt; The modal decoupling module aims to minimize the mutual information between features of different modalities by learning modal-specific extraction parameters through a neural network. The joint feature fusion network is a Transformer-LSTM joint network, where the Transformer layer captures long-term temporal dependencies and the LSTM layer captures the frequency domain dynamic changes of local signals.

[0011] Furthermore, in S3, a reinforcement learning module is constructed to perform policy evolution based on the dynamic changes in the weights of the multimodal fusion features, dynamically optimize the reward function, and generate a tension control policy, including: Construct a reinforcement learning agent that adapts to the control requirements of a belt conveyor, and define the agent's state space, action space, and reward function; Set a threshold for the weight change rate, and design a policy evolution mechanism based on the weight change rate of multimodal fusion features. When the weight change rate exceeds the threshold, the policy network parameters are updated. The reward function is dynamically optimized, and the three core objectives of balancing tension control accuracy, operational stability and strategy adaptability are achieved by adopting adaptive weight coefficients. The agent is trained using reinforcement learning algorithms to generate the optimal tension control strategy.

[0012] Furthermore, the state space is based on global multimodal fusion features, and also incorporates the current operating condition parameters of the belt conveyor. These operating condition parameters include at least: Conveyor belt target tension value, actual operating speed, material load level, and drum temperature; The action space is defined as the operable amount of the belt conveyor tension adjustment actuator, which includes at least the pressure adjustment amount of the hydraulic tensioning device and the torque adjustment amount of the drive motor. The weight change rate threshold is dynamically adjusted through exponential decay, and the reinforcement learning algorithm employs a proximal policy optimization algorithm.

[0013] Furthermore, the adaptive weight coefficient of the reward function is dynamically adjusted according to the current operating status of the belt conveyor. When the tension deviation exceeds the set threshold, the weight coefficient corresponding to the tension control accuracy is increased. When signal fluctuations exceed a set threshold, the weighting coefficient corresponding to operational stability is increased. When weights change frequently, increase the weight coefficient corresponding to the policy adaptability.

[0014] Furthermore, in step S4, the dynamic tension of the belt conveyor is dynamically controlled in real time based on the tension control strategy, and parameter reconfiguration is achieved through an online self-learning mechanism, including: The optimal tension control strategy is converted into control commands that can be recognized by the belt conveyor actuator, and transmitted to the actuator via industrial Ethernet to form a closed-loop control. Set sensor anomaly detection thresholds and signal drift detection thresholds to monitor the integrity and stability of multi-source operating signals in real time; When a sensor malfunction or signal drift is detected, an online self-learning mechanism is triggered. The abnormal signal is processed through data completion or substitution, and the modal decoupling parameters, policy network parameters, and reward function weight coefficients are updated to achieve adaptive correction of the control strategy.

[0015] A reinforcement learning-based intelligent dynamic tension control system for belt conveyors, applied to the aforementioned intelligent dynamic tension control method for belt conveyors, the system comprising: Signal acquisition module, signal preprocessing module, feature fusion module, reinforcement learning decision-making module, real-time control module, and online self-learning module; The signal acquisition module is used to synchronously acquire multi-source operating signals of the belt conveyor through a sensor array; The signal preprocessing module is used to remove abnormal data and standardize the acquired multi-source operating signals; The feature fusion module is used to output multimodal fusion features through modal decoupling and joint network fusion. The reinforcement learning decision-making module is used to construct reinforcement learning agents based on multimodal fusion features, dynamically optimize policies and reward functions, and generate optimal tension control policies. The real-time control module is used to convert the control strategy into control commands and transmit them to the actuator to realize real-time dynamic tension control; The online self-learning module is used to monitor the sensor status and signal stability, and automatically corrects the control strategy when the sensor is abnormal or the signal drifts, so as to ensure that the belt conveyor always operates within the safe tension range.

[0016] The beneficial effects of this invention are as follows: 1. By combining modal decoupling with a Transformer-LSTM joint network to achieve deep fusion of time-series and frequency-domain features, the nonlinear correlations between signals such as conveyor belt tension, vibration, current, and speed can be fully extracted, improving the robustness and accuracy of feature representation.

[0017] 2. Based on the dynamic changes in weights according to multimodal fusion characteristics, strategy evolution and reward function optimization are carried out to achieve a multi-objective balance of tension control accuracy, operational stability and strategy adaptability, which significantly improves the self-learning and decision-making capabilities of the control system in complex dynamic environments.

[0018] 3. By monitoring the integrity and stability of sensor signals in real time, an online self-learning mechanism is automatically triggered when a sensor malfunctions or a signal drifts, completing data completion, parameter updates, and strategy corrections to ensure that the conveyor belt always operates within a safe tension range.

[0019] 4. The system can automatically adjust the sampling frequency and feature weight allocation strategy according to different loads, speeds and operating conditions of the belt conveyor, and has good versatility and scalability, enabling stable operation in different types of conveying systems.

[0020] 5. By organically combining multi-source signal fusion, reinforcement learning optimization, and online adaptive mechanisms, the fixed nature and lag of traditional tension control methods are overcome, and intelligent, precise, and robust control of the dynamic tension of belt conveyors is achieved. Attached Figure Description

[0021] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart illustrating the overall method of the present invention; Figure 2 This is a schematic diagram of the structure of the present invention; Figure 3 This is a system block diagram of the present invention. Detailed Implementation

[0022] The present application will now be described in further detail with reference to the accompanying drawings. It should be noted that the following specific embodiments are only used to further illustrate the present application and should not be construed as limiting the scope of protection of the present application. Those skilled in the art can make some non-essential improvements and adjustments to the present application based on the above application content.

[0023] Example 1: Please see Figure 1-2 This invention provides a technical solution: a method for intelligent dynamic tension control of a belt conveyor based on reinforcement learning, the method comprising: S1. Acquire and preprocess the multi-source operating signals of the belt conveyor; Among them, belt conveyors are continuous transportation equipment that use drive rollers to drive conveyor belts to transport various bulk or packaged goods. They are widely used in industries such as mining, ports, power, and building materials. Multi-source operating signals refer to various signals collected from different parts and types of sensors on the belt conveyor. These signals reflect various state information during the operation of the conveyor, such as speed signals reflecting the speed of the conveyor belt, tension signals reflecting the tension on the conveyor belt, vibration signals reflecting the vibration of conveyor components, and motor current signals reflecting the working status of the motor. Preprocessing involves a series of processing operations on the collected raw multi-source operating signals. The purpose is to remove noise, interference, and other undesirable components from the signals, improve signal quality, and provide accurate and reliable data for subsequent analysis and processing. Common preprocessing methods include filtering and normalization. S2. Input the preprocessed multi-source operating signal into the Transformer-LSTM joint network, and realize the fusion of time-series and frequency-domain features through tension sensing fusion and modal decoupling to output multi-modal fusion features; In the Transformer-LSTM joint network, Transformer is a deep learning model architecture based on self-attention mechanism, originally used in natural language processing. It can capture long-range dependencies between elements at different positions in sequential data. When processing belt conveyor operation signals, it can effectively analyze the correlation between different time points in the signal and mine global features. LSTM is a special type of recurrent neural network (RNN) specifically designed for processing sequential data. By introducing gating mechanisms (input gate, forget gate, output gate), it solves the gradient vanishing and gradient exploding problems that traditional RNNs encounter when processing long sequences, and can better capture long-term dependencies in sequential data. In belt conveyor signal processing, LSTM can learn the pattern of signal changes over time. The joint network combines Transformer and LSTM to fully leverage their advantages. Transformer captures global features, while LSTM captures local temporal features, working together to process the input multi-source operation signals. Tension sensing fusion fuses signals from different tension sensors. Since different sensors may be installed at different locations on the conveyor, the collected tension information has different characteristics and... The key focus is on obtaining more comprehensive and accurate tension information by fusing these signals, thus improving the ability to perceive the tension state of the conveyor. Modal decoupling addresses the issue that signals generated by different physical processes or factors may be coupled together in signal processing. Modal decoupling separates these coupled signals into different modes, each representing a specific physical process or characteristic. For example, it separates the mechanical vibration mode and electrical signal mode from the operating signal of a belt conveyor to more deeply analyze the impact of different factors on the conveyor's operation. The fusion of time-series and frequency-domain features refers to the characteristics of a signal changing over time, reflecting the signal's performance over time. The dynamic changes along the time axis; frequency domain features are obtained by converting time-domain signals to the frequency domain using methods such as Fourier transform, reflecting the distribution of different frequency components in the signal. By fusing time-series and frequency domain features, information from both time and frequency dimensions of the signal can be utilized simultaneously to more comprehensively describe the operating status of the belt conveyor; multimodal fusion features are features obtained after the above processing that integrate multiple modal information, such as different sensor signal modes, time-series and frequency domain modes, etc. These features can more comprehensively and accurately reflect the operating status and characteristics of the belt conveyor, providing more effective input for subsequent reinforcement learning modules; S3. Construct a reinforcement learning module, perform strategy evolution based on the dynamic weight changes of the multimodal fusion features, dynamically optimize the reward function, and generate a tension control strategy that adapts to complex working conditions such as load fluctuations, start-stop transitions, and material impacts of belt conveyors. Among them, reinforcement learning is a machine learning framework that learns the optimal strategy through interaction between the agent and the environment. The agent takes actions based on the current environmental state, and the environment rewards or punishes the agent accordingly. The agent maximizes the long-term cumulative reward by continuously trying and adjusting its action strategy. In the dynamic tension control of belt conveyors, the agent takes corresponding tension control actions based on the operating state of the conveyor. The dynamic change of the weights of multimodal fusion features means that the importance of the fusion features of different modalities in reflecting the operating state of the conveyor may change with time, operating conditions and other factors. By dynamically adjusting the weights of these features, the influence of each feature on tension control in the current state can be more accurately reflected, enabling the reinforcement learning module to make more reasonable decisions based on the actual situation. Policy evolution is the process by which the agent in the reinforcement learning module continuously adjusts and optimizes its action strategy based on environmental feedback and changes in feature weights. Through multiple iterations and learning, the agent's strategy gradually evolves to adapt to different operating conditions and needs, improving the effectiveness of tension control. The reward function, in reinforcement learning, is a function used to measure the reward gained by the agent after taking action. The design of the reward function directly affects the agent's learning direction and policy optimization goals. In the dynamic tension control of belt conveyors, the reward function can be based on... The design considers multiple factors such as the tension state, operational stability, and energy consumption of the conveyor belt to guide the agent in generating an effective tension control strategy. The dynamically optimized reward function adjusts its form and parameters in real time based on the actual operating conditions and needs of the belt conveyor. For example, under different load fluctuations, the weight of tension stability in the reward function is adjusted so that the agent can control tension with the optimal strategy under different operating conditions. The tension control strategy is an action plan generated by the reinforcement learning module to control the tension of the belt conveyor. This strategy considers complex operating conditions such as load fluctuations, start-stop transitions, and material impacts, and can dynamically adjust the tension according to the real-time operating status to ensure the safe and stable operation of the conveyor. S4. Based on the tension control strategy, the dynamic tension of the belt conveyor is controlled in real time, and the parameters are reconfigured through an online self-learning mechanism. When the sensor is abnormal or the signal drifts, the control strategy is automatically corrected to ensure that the conveyor belt always runs in the safe tension range and avoids conveyor belt slippage, tearing or excessive slack. Among these, real-time control immediately adjusts the tension of the belt conveyor based on the tension control strategy generated by the reinforcement learning module. By monitoring the conveyor's operating status in real time, it promptly responds to load changes, start-stop operations, and other conditions, ensuring the tension remains within a suitable range. The online self-learning mechanism continuously collects new operating data and feedback information during system operation, using this information to automatically adjust and optimize the parameters of the reinforcement learning module. Through online self-learning, the system can continuously adapt to changes in the conveyor's operating environment and new operating conditions, improving the accuracy and effectiveness of the control strategy. Parameter reconfiguration resets and adjusts various parameters in the reinforcement learning module based on the results of the online self-learning mechanism. This parameter reconfiguration allows the system to better adapt to different operating conditions, improving the performance of tension control. Sensor anomalies or signal drift are addressed by... During long-term operation, the sensor may malfunction, leading to inaccurate signal acquisition. Alternatively, due to environmental factors or sensor characteristics, the signal value may change slowly over time. These situations can affect the system's accurate judgment of the conveyor's operating status, thus impacting the effectiveness of the tension control strategy. The automatic correction control strategy automatically identifies and takes measures when sensor abnormalities or signal drift are detected. It adjusts and corrects the tension control strategy based on other reliable information or preset correction rules, ensuring that the control strategy still effectively guarantees the conveyor belt operates within the safe tension range. The safe tension range is a reasonable range of conveyor belt tension set to ensure the normal operation of the belt conveyor and prevent belt slippage, tearing, or excessive slack. Within this range, the conveyor belt can stably and efficiently transport materials while extending its service life.

[0024] It should be noted that during use, acquiring and preprocessing multi-source operating signals can comprehensively and accurately reflect the conveyor status, providing a reliable data foundation for subsequent analysis. The Transformer-LSTM joint network achieves the fusion of time-series and frequency-domain features through tension sensing fusion and modal decoupling, outputting multi-modal fusion features. This allows for in-depth mining of hidden signal information, more accurate understanding of operating patterns, and the construction of a reinforcement learning module. Based on the dynamic evolution strategy of multi-modal fusion feature weight changes and the optimization of the reward function, it can generate tension control strategies adapted to complex working conditions, improving the adaptability and effectiveness of control. Real-time control and parameter reconfiguration through an online self-learning mechanism automatically correct the strategy when sensors malfunction or signals drift, ensuring that the conveyor belt is always within a safe tension range, effectively avoiding problems such as slippage, tearing, or excessive slack, ensuring stable operation of the conveyor, and reducing failure rate and maintenance costs.

[0025] In one embodiment, acquiring and preprocessing the multi-source operating signals of the belt conveyor includes: Multi-source operating signals are synchronously collected by sensor arrays deployed at key parts of the belt conveyor. These include: longitudinal tension signals of the conveyor belt collected by tension sensors installed on the carrying and return sections of the conveyor belt, reflecting the stretching state of the conveyor belt and directly related to the core target of tension control; equipment vibration signals collected by vibration sensors fixed to the bearing seats of the drive roller and the idler roller, indirectly reflecting the smoothness of roller operation and the fit of the conveyor belt; conveyor belt running speed signals collected by speed sensors installed on the output shaft of the drive motor, which are related to the load transmission efficiency and tension matching relationship; and motor current signals collected by current sensors connected in series in the power supply circuit of the drive motor, reflecting the size of the motor load and indirectly mapping the changes in conveyor belt tension. Each sensor operates at a preset sampling frequency, set according to the belt conveyor's running speed, with a range of 100Hz-1000Hz. Higher sampling rates are used during high-speed operation. Data is collected synchronously to form the original signal set.

[0026] in, This indicates the signal dimension, corresponding to 4 core signal types, with a value of 4. This indicates the signal acquisition duration, which is dynamically adjusted according to the operating conditions of the belt conveyor. When the load is stable, it is taken as 10s-30s, and extended to 30s-60s during start-up, shutdown, or material impact. The raw signals of each dimension are preprocessed for belt conveyor compatibility. The 3σ criterion is used to eliminate abnormal data caused by material impact or transient sensor interference, such as extreme values ​​where tension signal mutations exceed three times the standard deviation of the normal range. Then, the Z-score standardization method is used to eliminate dimensional differences between different signals, such as tension (N), current (A), and velocity (m / s). The standardized expression is:

[0027] in, Indicates the first dimensional signal in the first The raw data at that moment Indicates the first The mean of the dimensional signal, Indicates the first The standard deviation of the signal. The data is standardized signal data; preprocessing ensures that the data conforms to the operating characteristics of the belt conveyor, providing a unified and reliable data foundation for subsequent feature fusion.

[0028] This design deploys sensor arrays at key locations, synchronously collecting multi-source signals at preset frequencies to form a raw signal set. Then, abnormal data is removed from the signals in each dimension using the 3σ criterion, and Z-score standardization is used to eliminate dimensional differences. The multi-source signals comprehensively reflect the operating status of the belt conveyor. The preset frequency acquisition can adapt to different operating speeds, ensuring data integrity. Abnormal data removal and standardization eliminate errors caused by external interference and dimensional differences, making the data unified and reliable. This provides a high-quality foundation for subsequent feature fusion, more accurately reflects the equipment's operating characteristics, and improves the accuracy of subsequent control.

[0029] In one embodiment, the step of inputting the preprocessed multi-source operating signal into the Transformer-LSTM joint network, achieving temporal and frequency domain feature fusion through tension sensing fusion and modal decoupling, and outputting multi-modal fused features includes: The standardized multi-source operating signals are segmented according to a fixed time window set by the operating cycle characteristics of the belt conveyor, with the time window length being... The value range is 50-200 time steps, ensuring that each segment can completely cover the running time corresponding to 1-2 roller circumferences of the conveyor belt, reflecting the tension fluctuation pattern within a short cycle. Each segment contains continuous... The complete signal data at each time step is processed to obtain a set of signal segments:

[0030] in, Number of segments; Each signal segment is decoupled using a mode decoupling module based on mutual information constraints. Modal separation is performed. This module addresses the coupling characteristics of multi-source signals in belt conveyors, such as how tension changes synchronously cause fluctuations in motor current and vibration signals. Aiming to minimize the mutual information between different modal features, it uses a neural network to learn specific extraction parameters for each mode, separating the unique features of tension, vibration, velocity, and motor current modes. To avoid feature distortion caused by intermodal interference, the decoupling process satisfies:

[0031] in, For the first The decoupling parameters of the three modes are pre-trained using historical operating data of the belt conveyor, with the optimization objective being to maximize the correlation between the characteristics of each mode and the tension control target; The modality-specific features are synchronously input into the Transformer-LSTM joint network for cross-modal feature fusion, adapting to the time-series operation characteristics of belt conveyors: The Transformer layer employs a multi-head self-attention mechanism, calculating attention weights between different modal features to capture the long-term time-series dependencies between tension signals and motor current and operating speed, such as the hysteretic correlation between current increase and tension rise under heavy load. The LSTM layer utilizes a gating mechanism to capture the frequency domain dynamic changes of local signals, such as instantaneous tension fluctuations caused by material impact and the frequency characteristics of drum vibration, adapting to dynamic operating conditions such as belt conveyor start-stop transitions and material impacts. The specific fusion process is as follows: ; ; ; in, The time-dependent features output by the Transformer layer reflect the tension variation trend of the belt conveyor during long-cycle operation. The frequency domain dynamic characteristics of the LSTM layer output reflect the tension fluctuations under instantaneous operating conditions. For the first Multimodal fusion features of individual signal segments are used to achieve complete preservation of temporal and frequency domain information through feature concatenation; An adaptive weight allocation mechanism adapted to the operational stability of the belt conveyor is adopted to dynamically determine the weight coefficients of each segment feature. During the operation of a belt conveyor, smaller signal fluctuations indicate more stable operating conditions and more reliable effective information contained in the features. For example, under uniform speed and heavy load, smaller tension signal fluctuations indicate higher feature confidence and a larger corresponding weight. The specific weight determination rules are as follows:

[0032] in, As a feature confidence index,

[0033] For the first The standard deviation of the fused features of individual segments; for example, when materials are conveyed uniformly, the fluctuation of segment features is small. Small value, Its large proportion and the characteristic of ensuring stable operating conditions play a dominant role in the global integration; The global multimodal fusion features are obtained by weighted summation. This feature comprehensively integrates the temporal correlation of multi-source signals from belt conveyors, such as the long-term matching relationship between tension and speed and frequency domain dynamics, such as the tension response information of instantaneous impacts, accurately characterizing the equipment's operating status and providing inputs that fit the actual working conditions of belt conveyors for subsequent reinforcement learning strategy decisions.

[0034] This design segments the standardized signal into fixed time windows, separates the features specific to each mode through modal decoupling, and then inputs them into a Transformer-LSTM joint network to fuse the features. An adaptive weight allocation mechanism is used to obtain global fused features. The time window segmentation can fully cover the short-cycle operation time and reflect the tension fluctuation law. Modal decoupling avoids feature distortion. The Transformer-LSTM joint network can simultaneously capture long-term time dependence and local frequency domain dynamic changes. The adaptive weight allocation highlights stable operating condition features. The global fused features comprehensively integrate information and accurately characterize the equipment status, providing a realistic input for reinforcement learning strategy decision-making.

[0035] In one embodiment, the reinforcement learning module, based on the dynamic weight changes of the multimodal fusion features, performs policy evolution, dynamically optimizes the reward function, and generates a tension control policy, including: Construct a reinforcement learning agent that adapts to the control requirements of belt conveyors and define the agent's state space. Action space With reward function : state space Global multimodal fusion features The core is the current operating parameters of the belt conveyor, including the target tension value of the conveyor belt. Actual operating speed, material load level, and drum temperature; Among them, the target tension value of the conveyor belt The tension range is set according to the strength, bandwidth, operating speed, and load level of the conveyor belt material, such as 1500N-3000N for nylon conveyor belts. Material load level is determined by combining motor current and speed signals; The drum temperature is supplemented by a temperature sensor, reflecting the equipment's heat dissipation status; Action space Defined as the operable quantity of the tension adjustment actuator of a belt conveyor, specifically including the pressure adjustment quantity of the hydraulic tensioning device, corresponding to the change range of the conveyor belt tension, and the torque adjustment quantity of the drive motor. The tension is indirectly adjusted by changing the output power of the motor. The range of action values ​​is limited according to the hardware performance of the equipment, such as the hydraulic pressure adjustment range of 0.5MPa-3MPa and the torque adjustment range of 50%-120% of the rated torque. Set a threshold for the rate of change of weights to adapt to the operating characteristics of the belt conveyor. This threshold is used to determine whether to trigger policy evolution: During the initial startup of a belt conveyor and under conditions such as material impact, signal characteristics fluctuate greatly and weights change frequently. If the threshold is too small, it will lead to frequent strategy updates and affect the stability of control. Therefore, the initial threshold is set relatively large. As the system enters a stable operating phase, such as uniform material conveying, the threshold dynamically decays to improve the sensitivity of strategy evolution and adapt to slight changes in operating conditions in a timely manner. The specific threshold determination rules are as follows:

[0036] in, The initial threshold is set based on the initial operational stability of the belt conveyor, with a value ranging from 0.1 to 0.3. During the startup phase, the value is set to 0.2 to 0.3, and after stable operation, it is set to 0.1 to 0.2. The continuous running time of the system. The threshold decay period is set according to the operational stability characteristics of the belt conveyor, with a value range of 100-500 time steps. During heavy-load stable operation, 300-500 time steps are used, while during light-load or frequent start-stop operations, 100-200 time steps are used. The threshold is smoothly adjusted through exponential decay to avoid sudden changes in the strategy. Weight change rate based on multimodal fusion features:

[0037] Design strategy evolution mechanism: The rate of change of weights reflects the dynamic changes in the operating conditions of the belt conveyor, such as... An increase may correspond to changes in signal characteristics caused by sudden changes in material load or conveyor belt wear. When this occurs, it indicates that the original strategy is no longer suitable for the current operating conditions, triggering a strategy network parameter update. The update formula is:

[0038] in, To update the learning rate for the strategy, an adaptive learning rate rule is used for dynamic adjustment, adapting to the variable operating conditions of belt conveyors. , The initial learning rate is 0.001-0.01. After new equipment or major repairs, it is 0.005-0.01. After stable operation, it is 0.001-0.005. This is an indicator function that takes the value 1 when the condition is met, and 0 otherwise. As the number of strategy updates increases, the learning rate gradually decreases to avoid parameter oscillations in the later stages and ensure the stability of the control strategy. The policy objective function is defined as the expected long-term cumulative reward of the agent. The core objective is to maintain the tension of the belt conveyor within the target range while minimizing equipment wear and failure risk. Dynamically optimize reward function The reward function needs to be tailored to the control requirements of the belt conveyor, comprehensively considering three core objectives: tension control accuracy (avoiding slippage and tearing), operational stability (reducing equipment vibration and wear), and strategy adaptability (adapting to load fluctuations, sensor malfunctions, and other operating conditions). An adaptive weighting coefficient is adopted. To balance the priorities of each objective, the weighting coefficients are dynamically adjusted based on the current operating status of the belt conveyor. The specific determination rules are as follows:

[0039]

[0040]

[0041] in, The target tension value for a belt conveyor is determined based on the conveyor belt material, width, design tension, and current load level. For example, the target tension for a polyester conveyor belt under heavy load is taken as 60%-80% of the design tension. To detect the actual tension value in real time, The standard deviation of the multimodal fusion characteristics reflects the operational stability of the belt conveyor. , This is a weighting adjustment coefficient, ranging from 5 to 10, used to control the sensitivity of weight changes, especially in tension-sensitive operating conditions such as high-speed operation. Take 8-10, , For threshold parameters, The setting is based on the allowable deviation range of the belt conveyor tension, typically 5%-10% of the target tension. For example, if the target tension is 2000N... Select 100N-200N; Based on the characteristic stability threshold set during normal operation of the belt conveyor, and obtained from historical data statistics, it meets the following requirements. ; The dynamic adaptation logic of the weighting coefficients is as follows: when the tension deviation... Exceed hour, Automatically increases tension, up to 0.7-0.8, prioritizing tension control accuracy to prevent conveyor belt slippage or tearing; when signal fluctuations occur... Exceed For example, material impact can cause increased vibration. Automatically increases, up to 0.6-0.7, prioritizing system stability and reducing equipment wear; When the weights change frequently The absolute value is relatively large. Automatically increases, up to 0.5-0.6, prioritizing the improvement of the strategy's adaptability to changes in operating conditions; Based on the aforementioned weighting coefficients, the specific expression for the reward function is as follows:

[0042] in, The average of the weight change rates of each segment is used. Negative values ​​are used to penalize situations where excessive tension deviation may lead to failure, or where severe signal fluctuations may exacerbate equipment wear. Positive values ​​are used to reward situations where the strategy is highly adaptable and can quickly adapt to changes in operating conditions. A near-end policy optimization algorithm is used to train the reinforcement learning agent. This algorithm limits the policy update magnitude through clip operations to avoid excessive parameter updates due to sudden changes in operating conditions, thus ensuring training stability. During training, the agent interacts with the belt conveyor's operating environment in real time (simulating different loads, start-stop, material impact, and other conditions), continuously adjusting the policy network parameters, and ultimately generating an optimal tension control strategy adapted to the complex operating conditions of the belt conveyor.

[0043] in, For the policy network, output the probability distribution of different actions.

[0044] This design constructs a reinforcement learning agent, clearly defining the state, action space, and reward function. A weight change rate threshold is set to trigger policy evolution. An evolutionary mechanism is designed based on the weight change rate, dynamically optimizing the reward function. An optimal policy is generated through training using a proximal policy optimization algorithm. The clearly defined space aligns with equipment control requirements, the weight change rate threshold dynamically adjusts the policy evolution sensitivity according to operating conditions, the dynamically optimized reward function balances the priorities of multiple objectives, and the proximal policy optimization algorithm ensures stable training. The resulting optimal policy adapts to complex operating conditions, effectively improving tension control accuracy and reducing equipment wear and failure risks.

[0045] In one embodiment, the real-time control of the dynamic tension of the belt conveyor based on the tension control strategy, and the parameter reconfiguration through an online self-learning mechanism, automatically correcting the control strategy when the sensor malfunctions or the signal drifts, includes: The optimal tension control strategy for the output of the reinforcement learning agent. Convert into control commands that can be recognized by the belt conveyor actuator: For hydraulically tensioned belt conveyors, the tension adjustment amount is converted into a pressure control signal of the hydraulic system, such as a 4-20mA standard current signal. The thrust of the tensioning cylinder is adjusted through the proportional relief valve, thereby changing the tension of the conveyor belt. For counterweight tensioned belt conveyors, the tension can be stabilized by adjusting the output power of the drive motor in conjunction with the motor torque adjustment signal. Control commands are transmitted to the actuators via industrial Ethernet. The actuators adjust their operating parameters in real time according to the commands, forming a closed-loop control of "signal acquisition - feature fusion - strategy decision-making - execution control" to ensure that the dynamic tension of the belt conveyor always tracks the target value. Set anomaly detection threshold to adapt to the sensor deployment characteristics of the belt conveyor. Sensor anomaly detection threshold, and Signal drift detection threshold is used to accurately identify sensor faults and signal distortion, preventing control failure due to sensor problems. Sensor anomaly detection threshold A malfunction in a key sensor of a belt conveyor, such as a tension sensor, can directly affect the control effect, and its weighting coefficient... For larger values, the corresponding anomaly detection time threshold should be more lenient to avoid false positives leading to unnecessary policy adjustments. The specific rule is as follows:

[0046] in, The basic anomaly duration threshold is set, ranging from 3 to 5 sampling periods, depending on the sensor's sampling frequency. For example, at 100Hz sampling, one sampling period is 0.01s. Take a time interval of 0.03s-0.05s. The weighting coefficients for each modal signal are the corresponding values ​​for the tension sensor. It is the largest, therefore its anomaly detection threshold is the most lenient; Signal drift determination threshold Long-term operation of belt conveyors may cause sensor aging, such as zero-point drift of tension sensors. The signal drift threshold needs to be dynamically adjusted in conjunction with operational stability. The specific rules are as follows:

[0047] in, The base drift ratio threshold, ranging from 0.2 to 0.3, is set according to the sensor's accuracy class; for example, 0.25 is used for a 0.5-class sensor. The maximum standard deviation of the multimodal fusion characteristics during the historical operation of the belt conveyor reflects the signal fluctuation limit during normal operation. Real-time monitoring of the integrity and stability of multi-source operating signals of the belt conveyor: Determining sensor malfunctions by analyzing signal reception status, such as continuous signal loss from the tension sensor exceeding [a certain duration]. The system was determined to be a sensor malfunction, and redundant sensor data replacement was initiated. The distribution range of the standardized signal is used to determine whether signal drift exists, such as when the motor current signal deviates from the normal range. The proportion exceeds The signal is identified as drift, which may be caused by sensor aging or loose wiring. When a sensor malfunction or signal drift is detected, the online self-learning mechanism is immediately triggered, and the parameter correction process can be started without stopping the machine, ensuring the continuous operation of the belt conveyor. Based on the latest acquired valid signal data, abnormal signals are supplemented by interpolation of redundant sensor data at the same location, or replaced by other related signals. For example, when the tension signal drifts, the tension is indirectly estimated using motor current and speed signals, and the modal decoupling parameters are updated using the gradient descent method. and policy network parameters At the same time, reconfigure the weight coefficients of the reward function. This enables adaptive correction of the control strategy, with the specific update formula as follows:

[0048]

[0049] in, For the online learning rate, a dynamic learning rate rule is adopted to adapt to the continuous operation requirements of the belt conveyor:

[0050]

[0051] The initial online learning rate is set to 0.0001-0.001 to ensure smooth parameter updates and avoid abrupt changes in parameter regulation. This is an anomaly indicator function. It takes a value of 1 when an anomaly occurs, which reduces the learning rate and avoids parameter oscillation. It takes a value of 0 when the condition is normal. We assign weighting coefficients to each modal signal to ensure more accurate updates of the decoupling parameters corresponding to the core signal. The modal decoupling loss function is defined as the sum of mutual information between modal features, and the optimization objective is to improve the separation accuracy of multi-source signals in the belt conveyor. The objective function for online strategy optimization is defined as the expected short-term cumulative reward under the current operating conditions, ensuring that the strategy can quickly adapt to abnormal operating conditions. Through the online self-learning mechanism, the belt conveyor can continuously optimize the control strategy during long-term operation, adapting to equipment aging, load changes, sensor performance degradation, etc., without the need for downtime for retraining, significantly improving operational reliability and maintenance efficiency.

[0052] This design converts the optimal strategy into control commands and transmits them to the actuator to form a closed-loop control. It sets thresholds for anomaly and drift detection, and triggers an online self-learning mechanism when a problem is detected. This mechanism updates parameters and reconfigures the reward function using valid signals. The closed-loop control ensures dynamic tension tracks the target value, and the anomaly and drift detection thresholds accurately identify problems, preventing control failures. The online self-learning mechanism can correct the strategy without downtime, and the dynamic learning rate rule ensures smooth parameter updates, continuously optimizing the control strategy, adapting to equipment changes, significantly improving operational reliability and maintenance efficiency, and reducing downtime losses.

[0053] Example 2: Please see Figure 3 A reinforcement learning-based intelligent dynamic tension control system for belt conveyors is applied to the aforementioned intelligent dynamic tension control method for belt conveyors. The system includes: Signal acquisition module, signal preprocessing module, feature fusion module, reinforcement learning decision-making module, real-time control module, and online self-learning module; The signal acquisition module is used to synchronously acquire multi-source operating signals of the belt conveyor through a sensor array; The signal preprocessing module is used to remove abnormal data and standardize the acquired multi-source operating signals; The feature fusion module is used to output multimodal fusion features through modal decoupling and joint network fusion. The reinforcement learning decision-making module is used to construct reinforcement learning agents based on multimodal fusion features, dynamically optimize policies and reward functions, and generate optimal tension control policies. The real-time control module is used to convert the control strategy into control commands and transmit them to the actuator to realize real-time dynamic tension control; The online self-learning module is used to monitor the sensor status and signal stability, and automatically corrects the control strategy when the sensor is abnormal or the signal drifts, so as to ensure that the belt conveyor always operates within the safe tension range.

[0054] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0055] The above embodiments provide a detailed description of the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for intelligent dynamic tension control of a belt conveyor based on reinforcement learning, characterized in that, The method includes the following steps: S1. Acquire and preprocess the multi-source operating signals of the belt conveyor; S2. Input the preprocessed multi-source operating signal into the feature fusion network, and realize the fusion of time-series and frequency-domain features through sensor fusion and mode decoupling, and output multi-modal fusion features; S3. Construct a reinforcement learning module, perform policy evolution based on the dynamic weight changes of the multimodal fusion features, dynamically optimize the reward function, and generate a tension control strategy that adapts to the complex working conditions of the belt conveyor. S4. Based on the tension control strategy, the dynamic tension of the belt conveyor is controlled in real time, and the parameters are reconfigured through an online self-learning mechanism. When the sensor is abnormal or the signal drifts, the control strategy is automatically corrected to ensure that the conveyor belt always runs within the safe tension range.

2. The method for intelligent dynamic tension control of belt conveyors based on reinforcement learning according to claim 1, characterized in that, The process of acquiring and preprocessing the multi-source operating signals of the belt conveyor in step S1 includes: Multi-source operating signals are collected synchronously by sensor arrays deployed at key parts of the belt conveyor; The multi-source operating signals include at least the conveyor belt longitudinal tension signal, equipment vibration signal, conveyor belt running speed signal, and motor current signal; Each sensor synchronously collects data at a preset sampling frequency to form a raw signal set; The original signals of each dimension are subjected to adaptive preprocessing, including removing interference data by using outlier removal methods and eliminating the dimensional differences between different signals by using standardization methods.

3. The intelligent dynamic tension control method for belt conveyors based on reinforcement learning according to claim 2, characterized in that: The preset sampling frequency ranges from 100Hz to 1000Hz and is dynamically adjusted according to the operating speed of the belt conveyor. The signal acquisition duration of the original signal set is dynamically adjusted according to the operating conditions of the belt conveyor. When the load is stable, it is 10s-30s, and when starting, stopping or material impact occurs, it is extended to 30s-60s. The outlier removal method uses the 3σ criterion, and the standardization method uses Z-score standardization.

4. The intelligent dynamic tension control method for belt conveyors based on reinforcement learning according to claim 1, characterized in that, In step S2, the preprocessed multi-source operating signal is input into the feature fusion network. Through sensor fusion and modal decoupling, time-series and frequency-domain feature fusion is achieved, outputting multi-modal fused features, including: The standardized multi-source operating signals are segmented according to a preset time window to obtain a set of signal segments; The modal decoupling module performs modal separation on each signal segment to obtain the unique characteristics of each mode; Each modality-specific feature is synchronously input into a joint feature fusion network for cross-modal feature fusion, and time-dependent features and frequency-domain dynamic features are obtained respectively. Multimodal fusion features of each signal segment are obtained by feature concatenation. An adaptive weight allocation mechanism is used to dynamically determine the weight coefficients of each segment feature, and the global multimodal fusion feature is obtained by weighted summation.

5. The intelligent dynamic tension control method for belt conveyors based on reinforcement learning according to claim 4, characterized in that: The preset time window length ranges from 50 to 200 time steps, ensuring that each segment can completely cover the running time corresponding to 1-2 roller circumferences of the conveyor belt; The modal decoupling module aims to minimize the mutual information between features of different modalities by learning modal-specific extraction parameters through a neural network. The joint feature fusion network is a Transformer-LSTM joint network, where the Transformer layer captures long-term temporal dependencies and the LSTM layer captures the frequency domain dynamic changes of local signals.

6. The intelligent dynamic tension control method for belt conveyors based on reinforcement learning according to claim 1, characterized in that, The reinforcement learning module constructed in S3 performs policy evolution based on the dynamic weight changes of the multimodal fusion features, dynamically optimizes the reward function, and generates a tension control policy, including: Construct a reinforcement learning agent that adapts to the control requirements of a belt conveyor, and define the agent's state space, action space, and reward function; Set a threshold for the weight change rate, and design a policy evolution mechanism based on the weight change rate of multimodal fusion features. When the weight change rate exceeds the threshold, the policy network parameters are updated. The reward function is dynamically optimized, and the three core objectives of balancing tension control accuracy, operational stability and strategy adaptability are achieved by adopting adaptive weight coefficients. The agent is trained using reinforcement learning algorithms to generate the optimal tension control strategy.

7. The intelligent dynamic tension control method for belt conveyors based on reinforcement learning according to claim 6, characterized in that, The state space is based on global multimodal fusion features and incorporates the current operating condition parameters of the belt conveyor. These operating condition parameters include at least the following: Conveyor belt target tension value, actual operating speed, material load level, and drum temperature; The action space is defined as the operable amount of the belt conveyor tension adjustment actuator, which includes at least the pressure adjustment amount of the hydraulic tensioning device and the torque adjustment amount of the drive motor. The weight change rate threshold is dynamically adjusted through exponential decay, and the reinforcement learning algorithm employs a proximal policy optimization algorithm.

8. The intelligent dynamic tension control method for belt conveyors based on reinforcement learning according to claim 6, characterized in that: The adaptive weighting coefficient of the reward function is dynamically adjusted according to the current operating status of the belt conveyor. When the tension deviation exceeds the set threshold, the weighting coefficient corresponding to the tension control accuracy is increased. When signal fluctuations exceed a set threshold, the weighting coefficient corresponding to operational stability is increased. When weights change frequently, increase the weight coefficient corresponding to the policy adaptability.

9. The intelligent dynamic tension control method for belt conveyors based on reinforcement learning according to claim 1, characterized in that, In step S4, the dynamic tension of the belt conveyor is dynamically controlled in real time based on the tension control strategy, and parameter reconfiguration is achieved through an online self-learning mechanism, including: The optimal tension control strategy is converted into control commands that can be recognized by the belt conveyor actuator, and transmitted to the actuator via industrial Ethernet to form a closed-loop control. Set sensor anomaly detection thresholds and signal drift detection thresholds to monitor the integrity and stability of multi-source operating signals in real time; When a sensor malfunction or signal drift is detected, an online self-learning mechanism is triggered. The abnormal signal is processed through data completion or substitution, and the modal decoupling parameters, policy network parameters, and reward function weight coefficients are updated to achieve adaptive correction of the control strategy.

10. A dynamic tension intelligent control system for belt conveyors based on reinforcement learning, characterized in that, The system, applied to the method for intelligent dynamic tension control of a belt conveyor according to any one of claims 1-9, comprises: Signal acquisition module, signal preprocessing module, feature fusion module, reinforcement learning decision-making module, real-time control module, and online self-learning module; The signal acquisition module is used to synchronously acquire multi-source operating signals of the belt conveyor through a sensor array; The signal preprocessing module is used to remove abnormal data and standardize the acquired multi-source operating signals; The feature fusion module is used to output multimodal fusion features through modal decoupling and joint network fusion. The reinforcement learning decision-making module is used to construct reinforcement learning agents based on multimodal fusion features, dynamically optimize policies and reward functions, and generate optimal tension control policies. The real-time control module is used to convert the control strategy into control commands and transmit them to the actuator to realize real-time dynamic tension control; The online self-learning module is used to monitor the sensor status and signal stability, and automatically corrects the control strategy when the sensor is abnormal or the signal drifts, so as to ensure that the belt conveyor always operates within the safe tension range.

Images