A coiled roll safety state prediction control method based on multi-sensor cooperation

Abstract

This invention relates to the field of industrial control technology, specifically to a multi-sensor collaborative method for predictive control of the safety status of a take-up roller. This method involves real-time acquisition of multimodal data from multiple sensors deployed on the take-up roller, fusing this data with digital twin model data to generate a state feature vector. The state prediction model is then used to calculate the predicted future state value and uncertainties. Furthermore, the model predictive controller integrates the actual state, predicted state, and uncertainties to generate an optimal control sequence online. This sequence is received and executed by a programmable logic controller (PLC) to perform closed-loop adjustments to the take-up roller's control parameters. Simultaneously, the system calculates the probability of operational anomalies in real time and issues over-limit alarms. This invention deeply integrates prediction, control, and monitoring, achieving proactive prevention of equipment failures and significantly improving the intelligence level, operational safety, and production efficiency of the programmable control system.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of industrial control technology, in particular to a winding roll safety state prediction control method based on multi-sensor cooperation. BACKGROUND

[0002] In the field of industrial automation, the operation control of winding rolls generally adopts program control systems. Such systems usually take programmable logic controllers (PLCs) as the hardware core, with a set of pre-set, static logic programs running inside; the programs strictly execute a series of determined instruction sequences according to process requirements, such as controlling winding tension through PID loops, setting line speed synchronization of multi-stage drive units, and managing roll diameter calculation, to achieve precise control over the production process.

[0003] However, the existing program control technology has the following limitations in the application scenario of winding rolls:

[0004] (1) In the traditional control architecture, the process control program executed by the PLC and the device state monitoring system running in parallel with it are independent in terms of data and logic, so the traditional program control system cannot perceive the dynamic changes in the state characteristics of the winding roll, causing the control program to continue executing the original instruction sequence when the device health state deteriorates;

[0005] (2) The state monitoring results cannot directly and dynamically guide the program control system to adjust operating parameters to avoid risks, lacking closed-loop active control capabilities.

[0006] Therefore, a winding roll safety state prediction control method based on multi-sensor cooperation is proposed. SUMMARY

[0007] The present application relates to the field of industrial control technology, in particular to a winding roll safety state prediction control method based on multi-sensor cooperation.

[0008] To achieve the above-mentioned purpose, the present application provides the following technical solutions:

[0009] A winding roll safety state prediction control method based on multi-sensor cooperation, comprising:

[0010] Real-time acquisition of multi-modal data through multi-sensors deployed on the winding roll and the conveyor belt;

[0011] Preprocessing of the multi-modal data, and fusion of the preprocessed data with the preset winding roll digital twin model data to generate a feature vector representing the current state of the winding roll;

[0012] Based on the feature vector, in combination with a pre-established state prediction model and physical constraint conditions, the current state of the winding roll is analyzed to obtain its predicted state value and state uncertainty distribution;

[0013] According to the current actual state value, the predicted state value and the state uncertainty distribution, applying a model predictive controller to optimize the control sequence of the coiling roller online to generate an optimal control sequence; receiving and executing the optimal control sequence by a programmable logic controller to adjust the control parameters in a closed loop;

[0014] The running abnormality probability of the coiling roller is calculated by the monitoring and alarming module in real time, and when the probability exceeds a preset safety threshold, an alarm is triggered to prompt an operator to intervene.

[0015] Further, the multi-modal data includes vibration data, temperature data, visual data and operation parameter data; wherein the operation parameter data includes motor torque, coiling tension and conveyor belt speed.

[0016] Further, the process of pre-processing the multi-modal data to generate pre-processed data includes:

[0017] The adaptive filtering technique is applied to the multi-modal data for denoising to obtain first multi-modal data;

[0018] The timestamp synchronization and interpolation method is used to align the data of the first multi-modal data to obtain second multi-modal data;

[0019] The second multi-modal data is normalized to obtain pre-processed data.

[0020] Further, the process of fusing the pre-processed data and the preset coiling roller digital twin model data to generate a feature vector representing the current state of the coiling roller includes:

[0021] Based on the geometric parameters, material properties and operating conditions of the coiling roller, a digital twin model is constructed;

[0022] The time domain and frequency domain analysis method is used to extract features from the pre-processed data to obtain first features;

[0023] The real-time operation parameter data is input into the digital twin model to output virtual sensor data under the same working condition, and the features are extracted to obtain second features;

[0024] The first features and the second features are compared to obtain residual features; the first features, the second features and the residual features are spliced to obtain a feature vector representing the current state of the coiling roller.

[0025] Further, based on the feature vector, the current state of the coiling roller is analyzed in combination with a pre-established state prediction model and physical constraint conditions to output the predicted state value and the state uncertainty distribution.

[0026] inputting the feature vector into a pre-trained long short-term memory network model to obtain a predicted state value;

[0027] inputting the feature vector into a pre-trained physical information neural network model to output a probability vector of the coiling roller under different state categories;

[0028] calculating the highest state probability by using a Shannon entropy calculation formula to obtain a state entropy for representing state uncertainty distribution.

[0029] Further, the model predictive controller is used to optimize the control sequence of the coiling roller, and the process of generating the optimal control sequence comprises:

[0030] a cost function of the model predictive controller is constructed, and the optimization objective is to minimize the deviation between the predicted state value and the current actual state value, reduce state uncertainty, and suppress the sharp change of the control amount;

[0031] a nonlinear programming algorithm is used to solve the optimization problem to obtain the control sequence in the future period of time;

[0032] in each control period, the first instruction of the control sequence is issued to the programmable logic controller, and the new measurement data is used for rolling optimization to generate the optimal control sequence.

[0033] Further, the process of calculating the running abnormal probability of the coiling roller in real time by the monitoring and alarm module comprises:

[0034] the predicted state value is continuously obtained by the edge controller and compared with the actual measurement value of the sensor;

[0035] the difference between the predicted state value and the actual measurement value is combined with the state uncertainty distribution to form an abnormal score;

[0036] the abnormal score is mapped to a running abnormal probability value between 0 and 1 by a normalization function.

[0037] Compared with the prior art, the beneficial effects of the present application are:

[0038] 1. By combining the geometric parameters, material properties and operating conditions of the coiling roller to construct a digital twin model, and by extracting features through time-frequency domain analysis, combining the residual features of real-time and virtual data, and generating a comprehensive feature vector representing the current state of the coiling roller, the accuracy of state representation is improved; by simulating the actual working condition through the digital twin model, the state change of the coiling roller is dynamically captured, and the limitation that the traditional control system cannot perceive the decline of the equipment health state is overcome; the introduction of residual features can highlight the deviation between the actual and expected states, effectively improve the sensitivity of the feature vector to abnormal states, and provide a more reliable basis for subsequent prediction and control.

[0039] 2. By using a pre-trained long short-term memory network model to capture temporal features, high-precision predicted state values ​​are generated. Combined with the output state probability vector of the physical information neural network, the prediction results are ensured to conform to physical laws. The state uncertainty distribution is calculated by Shannon entropy, providing a quantitative confidence index for state prediction and improving the reliability of the prediction results. The state probability vector and uncertainty distribution provide a basis for anomaly detection, which helps to identify potential risks in a timely manner and make up for the shortcomings of the disconnect between traditional system state monitoring and control.

[0040] 3. By constructing a cost function that includes state deviation, uncertainty, and control smoothness, the system can comprehensively consider prediction accuracy and control stability to generate a smooth and efficient optimal control sequence, reducing drastic fluctuations in the control process. The system uses a nonlinear programming algorithm to efficiently solve the optimization problem and combines a rolling optimization strategy to continuously update the control sequence, ensuring that the control commands can dynamically adapt to the state changes of the take-up roll, thereby improving the adaptability and reliability of the take-up roll safety state prediction control. Attached Figure Description

[0041] Fig. 1 This is a flowchart illustrating a multi-sensor collaborative method for predicting and controlling the safety status of a take-up roller according to the present invention.

[0042] Fig. 2 This is a schematic diagram illustrating the process of generating feature vectors in this invention.

[0043] Fig. 3 This is a schematic diagram of the process for generating the optimal control sequence in this invention. Detailed Implementation

[0044] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, 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 are within the scope of protection of the present invention.

[0045] Please see Figs. 1 to 3 This invention provides a method for predictive control of the safety status of a take-up roller based on multi-sensor collaboration, the technical solution of which is as follows:

[0046] Example 1:

[0047] To improve the predictive control effect of the safety status of the take-up roll, a company used a multi-sensor collaborative predictive control method for the safety status of the take-up roll proposed in this invention. The flowchart of the method is as follows: Fig. 1 As shown, it specifically includes:

[0048] Multimodal data is collected in real time by deploying multiple sensors on the take-up roller and conveyor belt;

[0049] Furthermore, the multimodal data includes vibration data, temperature data, visual data, and operational parameter data; among which, the operational parameter data includes motor torque, winding tension, and conveyor belt speed.

[0050] Furthermore, vibration and temperature data are collected by acceleration and temperature sensors mounted on the bearing housing of the take-up roller, respectively; visual data is captured by an industrial camera mounted directly above the take-up roller, equipped with a coaxial ring LED light source to eliminate the visual impact of surface reflection; and operating parameter data are collected by torque sensors mounted on the motor shaft, tension sensors on the production line, and speed encoders on the conveyor belt.

[0051] By collecting multimodal data such as vibration, temperature, vision, and operating parameters, the operating status of the take-up roll is comprehensively captured, making up for the shortcomings of traditional control systems that rely on only a single or a few data sources and cannot perceive dynamic state changes; multidimensional sensor data provides a richer source of information for predictive control of the safe state of the take-up roll.

[0052] The multimodal data is preprocessed and then fused with the preprocessed data and the preset digital twin model data of the take-up roll to generate a feature vector representing the current state of the take-up roll.

[0053] Furthermore, the process of preprocessing multimodal data to generate preprocessed data includes:

[0054] Adaptive filtering techniques are applied to the multimodal data to remove noise, resulting in the first multimodal data.

[0055] The first multimodal data is aligned using timestamp synchronization and interpolation methods to obtain the second multimodal data;

[0056] The second multimodal data is normalized to obtain preprocessed data.

[0057] Furthermore, the visual data in the multimodal data is denoised using Wiener filtering, while the remaining data is denoised using the least mean square algorithm for adaptive filtering.

[0058] Furthermore, during the timestamp synchronization process, for sensors with inconsistent sampling rates, linear interpolation is used to align the low sampling rate data to the highest sampling rate reference, forming a second multimodal data that is fully synchronized in time.

[0059] Furthermore, the aligned modal data are scaled to the [0, 1] interval using the min-max normalization method.

[0060] By using adaptive filtering to remove noise, timestamp synchronization and interpolation methods to align multi-source data, and normalization to unify data scale, the quality and consistency of multimodal data are improved, laying the foundation for subsequent feature extraction and state prediction.

[0061] Furthermore, the process of generating feature vectors is as follows: Fig. 2 As shown, it includes:

[0062] A digital twin model is constructed based on the geometric parameters, material properties, and operating conditions of the take-up roller;

[0063] The first feature is obtained by using time-domain and frequency-domain analysis methods to extract features from the preprocessed data.

[0064] Real-time operational parameter data is input into the digital twin model, and virtual sensor data under the same working conditions is output. Features are then extracted from the data to obtain the second feature.

[0065] The first feature is compared with the second feature to obtain the residual feature; the first feature, the second feature and the residual feature are concatenated to obtain the feature vector representing the current state of the take-up roller.

[0066] Furthermore, a three-dimensional model of the take-up roller was drawn using SolidWorks software and imported into multiphysics simulation software. At the same time, material properties such as Young's modulus, Poisson's ratio, and density were defined. Then, parametric scanning simulation was performed in the simulation software to conduct steady-state dynamics and thermal analysis on each combination of motor speed and take-up tension parameters, and to simulate virtual sensor data under ideal conditions.

[0067] Furthermore, time-domain and frequency-domain analysis includes the calculation of statistics such as mean, variance, and peak value, as well as spectral information obtained through fast Fourier transform;

[0068] Furthermore, before performing time-domain and frequency-domain analysis on the preprocessed data and virtual sensor data, the preprocessed data is decomposed using Daubechies wavelet decomposition to extract low-frequency and high-frequency components. Then, time-domain and frequency-domain analysis is performed on the decomposition results of each layer to generate multi-scale features. By introducing multi-scale decomposition technology, the fine-grained capture of the dynamic changes of the take-up roller can be improved, and the representation ability of the feature vector can be enhanced.

[0069] By constructing a digital twin model that combines the geometric parameters, material properties, and operating conditions of the take-up roll, and extracting features through time-frequency domain analysis, a comprehensive feature vector representing the current state of the take-up roll is generated by combining the residual features of real-time and virtual data, thereby improving the accuracy of state representation. By simulating actual working conditions using the digital twin model, changes in the state of the take-up roll can be dynamically captured, overcoming the limitation of traditional control systems that cannot detect the decline in the health status of the equipment. At the same time, the introduction of residual features can highlight the deviation between the actual and expected states, effectively improving the sensitivity of the feature vector to abnormal states, and providing a reliable basis for state prediction and control.

[0070] Based on the feature vector, combined with the pre-established state prediction model and physical constraints, the current state of the take-up roller is analyzed to obtain its predicted state value and state uncertainty distribution.

[0071] Furthermore, based on the feature vectors, combined with the pre-established state prediction model and physical constraints, the process of analyzing the current state of the take-up roller and outputting its predicted state value and state uncertainty distribution includes:

[0072] The feature vector is input into a pre-trained long short-term memory network model to obtain the predicted state value;

[0073] The feature vector is input into a pre-trained physical information neural network model, which outputs the probability vector of the take-up roller under different state categories.

[0074] The probability of the highest state is calculated using the Shannon entropy formula to obtain the state entropy, which is used to characterize the distribution of state uncertainty.

[0075] Furthermore, the Long Short-Term Memory (LSTM) network model consists of two LSTM layers, one Dropout layer, and one fully connected output layer; each LSTM layer has 128 neurons; the model takes feature vectors as input and outputs future sensor predictions; the training process uses the Adam optimizer with a learning rate of 0.001, a loss function of mean squared error, and trains for 100 epochs with a batch size of 64.

[0076] Furthermore, the Physical Information Neural Network (PINN) model is a fully connected neural network consisting of four hidden layers, each with 256 neurons. The activation function used is ReLU, and the output layer uses the Softmax function for probability generation. This model takes feature vectors as input and outputs the probabilities of various winding roller states. The loss function is a weighted sum of classification cross-entropy loss and physical constraint loss.

[0077] Furthermore, in actual production lines, the probability of abnormal samples occurring is very low, resulting in a small number of usable abnormal samples, thus reducing the reliability of the abnormal state probabilities output by the physical information neural network model. A generative adversarial network (GAN) is introduced to generate a latent state distribution to assist the GAN model in generating more robust state probability vectors. Feature vectors are input into the generator to produce simulated state probability vectors, and a discriminator distinguishes between the probability vectors from the PINN and the generated probability vectors. The probability vectors generated by the GAN are then weighted and averaged with the PINN output to generate a more robust state probability vector, improving...

[0078] The accuracy of PINN in predicting anomaly probabilities.

[0079] By using a pre-trained long short-term memory network model to capture temporal features, high-precision predicted state values ​​are generated. Combined with the output state probability vector of the physical information neural network, and using Shannon entropy to calculate the state uncertainty distribution, a quantitative confidence index is provided for state prediction. Using the state probability vector and uncertainty distribution, potential risks can be identified in a timely manner, making up for the shortcomings of the disconnect between state monitoring and control in traditional systems, thereby improving the reliability of the safe state prediction and control of the take-up roll.

[0080] Based on the current actual state value, predicted state value, and state uncertainty distribution, a model predictive controller is applied to optimize the control sequence of the take-up roller online and generate the optimal control sequence; the optimal control sequence is received and executed by a programmable logic controller to perform closed-loop adjustment of the control parameters;

[0081] Furthermore, the process of using a model predictive controller to optimize the control sequence of the take-up roller and generating the optimal control sequence is as follows: Fig. 3 As shown, it includes:

[0082] The cost function of the model predictive controller is constructed with the optimization objective of minimizing the deviation between the predicted state value and the current actual state value, reducing state uncertainty, and suppressing drastic changes in the control quantity.

[0083] The optimization problem is solved using a nonlinear programming algorithm to obtain the control sequence for a future period of time.

[0084] In each control cycle, the first instruction of the control sequence is sent to the programmable logic controller, and rolling optimization is performed using new measurement data to generate the optimal control sequence;

[0085] Furthermore, the control cycle, prediction time domain, and control time domain of the model predictive controller are set to 100ms, 20 steps, and 10 steps, respectively; the state variables are defined as actual state values, predicted state values, and state entropy; the control variables are set as three-dimensional vectors, including: motor output torque, winding tension setpoint, and conveyor belt speed; the constraints include control quantity constraints and control quantity change rate constraints.

[0086] Furthermore, the cost function includes a state tracking term, an uncertainty suppression term, and a control suppression term; among which, the state tracking term is used to penalize the difference between the predicted state value and the actual state value, the uncertainty suppression term directly penalizes the magnitude of the predicted state uncertainty, and the control suppression term is used to penalize the magnitude of the control action, that is, the magnitude of the increment of the control variable;

[0087] Furthermore, using a nonlinear programming solver, such as IPOPT, the optimization problem can be solved quickly on the edge controller to obtain a control sequence consisting of control variables that minimizes the cost function;

[0088] Furthermore, a DRL model is constructed, with state variables and control variables as the state space and cost function as the reward function. The action space of the DRL is represented as the adjustment of the weights of each item in the cost function. The DRL model is pre-trained offline using historical operating data. In each control cycle, the DRL adjusts the weights and optimizes the cost function according to the real-time operating conditions, so that the model predictive controller can adapt to the optimization requirements under different operating conditions.

[0089] By constructing a cost function that includes state deviation, uncertainty, and control smoothness, the system can comprehensively consider prediction accuracy and control stability to generate a smooth and efficient optimal control sequence, reducing drastic fluctuations in the control process. The system uses a nonlinear programming algorithm to efficiently solve the optimization problem and combines a rolling optimization strategy to continuously update the control sequence, ensuring that the control commands can dynamically adapt to the state changes of the take-up roll, thereby improving the adaptability and reliability of the take-up roll safety state prediction control.

[0090] The monitoring and alarm module calculates the probability of abnormal operation of the take-up roller in real time, and triggers an alarm when the probability exceeds a preset safety threshold to prompt the operator to intervene.

[0091] Furthermore, the process of calculating the probability of abnormal operation of the take-up roller in real time through the monitoring and alarm module includes:

[0092] The edge controller continuously acquires and compares the predicted state values ​​with the actual measurement values ​​from the sensors;

[0093] The difference between the predicted state value and the actual measured value is combined with the state uncertainty distribution to form an anomaly score;

[0094] The anomaly score is mapped to a runtime anomaly probability value between 0 and 1 using a normalization function;

[0095] Furthermore, the deviation between the predicted state value and the actual measurement value of the sensor is standardized, and the standardized result is weighted and summed with the state entropy to obtain the anomaly score;

[0096] Furthermore, the sigmoid function is used to map the anomaly score to a probability value of operational anomaly between 0 and 1. When this value is greater than 0.85, an alarm is triggered to prompt operator intervention.

[0097] By continuously comparing predicted state values ​​with actual measured values ​​using an edge controller, and calculating anomaly scores based on state uncertainty distribution, the probability of operational anomalies is generated in real time, improving the timeliness and accuracy of anomaly detection and compensating for the shortcomings of traditional system state monitoring results that cannot directly guide control.

[0098] This embodiment proposes a multi-sensor collaborative method for predictive control of the safety status of a take-up roll. This method involves real-time acquisition of multimodal data from multiple sensors deployed on the take-up roll, fusing this data with digital twin model data to generate a state feature vector. The state prediction model is then used to calculate the future predicted state value and uncertainties. Subsequently, the model predictive controller integrates the actual state, predicted state, and uncertainties to generate an optimal control sequence online. This sequence is received and executed by a programmable logic controller (PLC) to perform closed-loop adjustments to the take-up roll's control parameters. Simultaneously, the system calculates the probability of operational anomalies in real time and issues over-limit alarms. This invention deeply integrates prediction, control, and monitoring, achieving proactive prevention of equipment failures and significantly improving the intelligence level, operational safety, and production efficiency of the programmable control system.

[0099] Example 2:

[0100] This embodiment takes the optimization of the safety status prediction and control of the take-up roll in a steel strip manufacturing enterprise as an example. The specific process is as follows:

[0101] Multimodal data is collected in real time by deploying multiple sensors on the take-up roller and conveyor belt;

[0102] The company uses an accelerometer, a thermocouple sensor, and an industrial camera mounted on the take-up roller bearing housing to collect vibration data, temperature data, and visual data, respectively. Then, it uses a torque sensor, a tension sensor, and a speed encoder to collect operational parameter data. The reference values ​​are 2000 N·m motor torque, 50 kN take-up tension, and 1.5 m / s conveyor belt speed.

[0103] The multimodal data is preprocessed and then fused with the preprocessed data and the preset digital twin model data of the take-up roll to generate a feature vector representing the current state of the take-up roll.

[0104] The temperature and visual data with low sampling rates are aligned to the 1000Hz sampling rate benchmark of vibration data by linear interpolation, generating time-synchronized preprocessed data.

[0105] The company used SolidWorks to create a 3D model of the take-up roller, defining material properties such as Young's modulus of 200 GPa, Poisson's ratio of 0.3, and density of 7850 kg / m³. The model was then imported into COMSOL Multiphysics software for multiphysics simulation, simulating virtual sensor data under the combination of motor speed and take-up tension.

[0106] Based on the characteristics of the preprocessed data and the simulation data under time-frequency domain analysis, a feature vector is generated;

[0107] Based on the feature vector, combined with the pre-established state prediction model and physical constraints, the current state of the take-up roller is analyzed to obtain its predicted state value and state uncertainty distribution.

[0108] The enterprise inputs the feature vectors into the pre-trained long short-term memory network model and the physical information neural network model respectively to obtain the sensor prediction value for the next 100ms and the probability vectors of the winding roller in normal, slightly abnormal and severely abnormal states. The state entropy is calculated using the Shannon entropy formula to characterize the distribution of state uncertainty.

[0109] The sensor prediction values, i.e. the predicted state values, are as follows: the vibration amplitude is 0.02 m / s², the temperature is 85℃, and the tension deviation is 0.5 kN; the probability vector is [0.85, 0.12, 0.03]. Therefore, the take-up roller is biased towards the normal state.

[0110] Based on the current actual state value, predicted state value, and state uncertainty distribution, a model predictive controller is applied to optimize the control sequence of the take-up roller online and generate the optimal control sequence; the optimal control sequence is received and executed by a programmable logic controller to perform closed-loop adjustment of the control parameters;

[0111] The model predictive controller is set as follows: the control cycle is 100ms, the prediction time domain is 20 steps, and the control time domain is 10 steps; the state variables are defined as the actual state value, the predicted state value, and the state entropy; the constraint ranges for the motor torque, winding tension, and conveyor belt speed in the control variables are 2000-2500 N·m, 45-55 kN, and 1.4-2.0 m / s, respectively.

[0112] The model predictive controller uses the IPOPT nonlinear programming solver to solve the cost function minimization problem and optimizes the take-up roller control sequence. The generated optimal control sequence is as follows: motor torque is adjusted to 2100 N·m, take-up tension is adjusted to 48 kN, and conveyor belt speed is adjusted to 1.6 m / s.

[0113] The monitoring and alarm module calculates the probability of abnormal operation of the take-up roller in real time, and triggers an alarm when the probability exceeds a preset safety threshold to prompt the operator to intervene.

[0114] The edge controller continuously compares the predicted state values ​​with the actual measured values ​​to calculate the deviation value. The standardized result of the deviation value is weighted and summed with the state entropy to generate an anomaly score, which is then mapped to an anomaly probability value of 0-1 using the Sigmoid function. The deviation values ​​of vibration, temperature and tension data are 0.015 m / s², 0.4℃ and 0.3 kN, respectively, and the calculated anomaly scores and anomaly probabilities are 0.62 and 0.65, respectively, which do not exceed the alarm threshold.

[0115] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for predictive control of the safety status of a take-up roller based on multi-sensor collaboration, characterized in that, include: Multimodal data is collected in real time by deploying multiple sensors on the take-up roller and conveyor belt; The multimodal data is preprocessed and fused with the preprocessed data and the preset digital twin model data of the take-up roller to generate a feature vector characterizing the current state of the take-up roller. A digital twin model is constructed based on the geometric parameters, material properties, and operating conditions of the take-up roller; The first feature is obtained by using time-domain and frequency-domain analysis methods to extract features from the preprocessed data. Real-time operational parameter data is input into the digital twin model, and virtual sensor data under the same working conditions is output. Features are then extracted from the data to obtain the second feature. The first feature is compared with the second feature to obtain the residual feature; the first feature, the second feature and the residual feature are concatenated to obtain the feature vector representing the current state of the take-up roller. Based on the feature vector, combined with the pre-established state prediction model and physical constraints, the current state of the take-up roller is analyzed to obtain its predicted state value and state uncertainty distribution. The feature vector is input into a pre-trained long short-term memory network model to obtain the predicted state value; The feature vector is input into a pre-trained physical information neural network model, which outputs the probability vector of the take-up roller under different state categories. The probability of the highest state is calculated using the Shannon entropy formula to obtain the state entropy, which is used to characterize the distribution of state uncertainty. Based on the current actual state value, the predicted state value, and the state uncertainty distribution, a model predictive controller is applied to optimize the control sequence of the take-up roller online and generate the optimal control sequence. The optimal control sequence is received and executed by a programmable logic controller to perform closed-loop adjustment of the control parameters. The cost function of the model predictive controller is constructed with the optimization objective of minimizing the deviation between the predicted state value and the current actual state value, reducing state uncertainty, and suppressing drastic changes in the control quantity. The optimization problem is solved using a nonlinear programming algorithm to obtain the control sequence for a future period of time. In each control cycle, the first instruction of the control sequence is sent to the programmable logic controller, and rolling optimization is performed using new measurement data to generate the optimal control sequence; The monitoring and alarm module calculates the probability of abnormal operation of the take-up roller in real time, and triggers an alarm when the preset safety threshold is exceeded to prompt the operator to intervene.

2. The method for predictive control of the safety status of a take-up roller based on multi-sensor collaboration according to claim 1, characterized in that, The multimodal data includes vibration data, temperature data, visual data, and operational parameter data; among which, the operational parameter data includes motor torque, winding tension, and conveyor belt speed.

3. The method for predictive control of the safety status of a take-up roller based on multi-sensor collaboration according to claim 1, characterized in that, The process of preprocessing multimodal data to generate preprocessed data includes: Adaptive filtering techniques are applied to the multimodal data to remove noise, resulting in the first multimodal data. The first multimodal data is aligned using timestamp synchronization and interpolation methods to obtain the second multimodal data.

4. The method for predictive control of the safety status of a take-up roller based on multi-sensor collaboration according to claim 1, characterized in that, The process of calculating the probability of abnormal operation of the take-up roller in real time through the monitoring and alarm module includes: The edge controller continuously acquires and compares the predicted state values ​​with the actual measurement values ​​from the sensors; The difference between the predicted state value and the actual measured value is combined with the state uncertainty distribution to form an anomaly score; The anomaly score is mapped to a runtime anomaly probability value between 0 and 1 using a normalization function.

Images