A multi-parameter real-time monitoring and control system for steel and non-ferrous metal cold rolling process

Abstract

The present application relates to the technical field of multi-parameter real-time monitoring and control system, and specifically discloses a multi-parameter real-time monitoring and control system for steel and non-ferrous metal cold rolling process. The system comprises a multi-source data acquisition module, a signal preprocessing and feature extraction module, a multi-scale deep feature fusion module, a device state evaluation module based on a physical information neural network, and a self-adaptive decision and control output module. Through deep fusion of multi-source heterogeneous signal features and embedding of process physical constraints for state evaluation, the system can realize accurate identification and early warning of early weak faults of the cold rolling equipment, and output graded warning and self-adaptive control suggestions according to the evaluation results, thereby improving the reliability and production efficiency of the equipment.

Description

Technical Field

[0001] This invention belongs to the technical field of multi-parameter real-time monitoring and control systems, specifically relating to a multi-parameter real-time monitoring and control system for the cold rolling process of steel and non-ferrous metals. Background Technology

[0002] In the field of industrial automation and intelligent manufacturing, real-time monitoring and control of the operating status of production equipment is a key link in achieving efficient, stable, and high-quality production. Among them, the cold rolling process of steel and non-ferrous metals, as an important material forming process, directly affects the dimensional accuracy, surface quality, and overall efficiency of the production line through precise management of its equipment status and process parameters.

[0003] The multi-parameter real-time monitoring and control system for the cold rolling process aims to achieve closed-loop control and fault early warning of the production process by collecting and analyzing multi-dimensional signals from key equipment such as rolling mills, rolls, and motors. This system needs to comprehensively process heterogeneous data from various sensors, including those for vibration, temperature, pressure, and current, to build a comprehensive understanding of equipment health status and process stability.

[0004] Existing technologies typically rely on data from a single or a few sensors for threshold alarms, making it difficult to achieve deep insights into the status of complex equipment systems. For example, minor wear or localized defects on the surface of rolling mills have weak early characteristic signals that are easily masked in the conventional vibration monitoring spectrum, causing early warning systems based on traditional vibration analysis to fail to detect such progressive faults in a timely manner. This monitoring blind spot forces equipment maintenance to rely on reactive repairs or periodic overhauls, not only increasing unplanned downtime and affecting production continuity but also potentially leading to more serious equipment damage and safety accidents due to the expansion of defects.

[0005] In addition, existing systems lack the ability to integrate and analyze multi-source parameters. Each parameter is often judged independently, and there is a lack of intelligent diagnostic models based on process mechanisms and data correlation. As a result, they cannot achieve accurate status assessment and risk prediction under changing production conditions. Summary of the Invention

[0006] The purpose of this invention is to provide a multi-parameter real-time monitoring and control system for the cold rolling process of steel and non-ferrous metals, so as to solve the technical contradictions in the prior art caused by insufficient ability to fuse and analyze multi-source heterogeneous data and insensitivity to early weak fault characteristic signals, resulting in delayed early warning and inaccurate condition assessment.

[0007] The technical solution of this invention is a multi-parameter real-time monitoring and control system for the cold rolling process of steel and non-ferrous metals. This system constructs a closed-loop intelligent monitoring and control architecture from data acquisition, feature fusion, state assessment to decision output. The system includes a multi-source heterogeneous data acquisition module, a signal preprocessing and feature extraction module, a multi-scale deep feature fusion module, a physical information neural network-based equipment state assessment module, and an adaptive decision and control output module.

[0008] A multi-source heterogeneous data acquisition module is used to acquire multi-dimensional real-time operating signals of key equipment during the cold rolling process in parallel. This module specifically includes vibration sensing units, temperature sensing units, pressure sensing units, current and voltage sensing units, and process parameter acquisition units. The vibration sensing units are deployed at key measuring points on the mill stand, roll bearing housings, and main drive gearbox, using accelerometers to acquire three-dimensional vibration signals at a sampling frequency of no less than 20 kHz.

[0009] The temperature sensing unit uses a combination of infrared thermal imager and embedded thermocouples to synchronously monitor the working temperature of the roll surface, bearing temperature rise, and motor winding temperature, with a sampling frequency of no less than 100 Hz. The pressure sensing unit is integrated into the mill's hydraulic reduction system and bending roll system, acquiring rolling force and bending roll force signals in real time. The current and voltage sensing unit uses high-precision Hall effect sensors to acquire the three-phase current and voltage waveforms of the main motor and coiling motor. The process parameter acquisition unit obtains rolling speed, reduction rate, tension setpoint, and strip material and specification information in real time from the basic automation system of the rolling line.

[0010] The signal preprocessing and feature extraction module is used to perform noise reduction, alignment, and initial feature extraction on the raw signals acquired by the multi-source heterogeneous data acquisition module. This module first performs wavelet threshold denoising preprocessing on each raw signal to suppress electromagnetic interference and mechanical background noise. Further, the module timestamps the preprocessed signals to ensure strict synchronization of the multi-source data in the time dimension. After alignment, the module performs time-domain feature extraction, frequency-domain feature extraction, and time-frequency-domain feature extraction in parallel. Time-domain feature extraction calculates the root mean square value, peak factor, kurtosis index, and waveform index of the signal. Frequency-domain feature extraction extracts the frequency band including the meshing frequency octave band, bearing characteristic frequency band, and the energy proportion within the range of 0.5 to 3 times the rotational frequency by performing a Fast Fourier Transform on the signal. Time-frequency-domain feature extraction uses continuous wavelet transform to generate the time-spectrum graph of the signal and extracts wavelet energy entropy and singular spectral entropy as features characterizing the non-stationarity and complexity of the signal.

[0011] The multi-scale deep feature fusion module receives and deeply fuses multi-level features output from the signal preprocessing and feature extraction modules to construct a high-dimensional fused feature vector that can characterize the overall state of the device. The core of this module is a cascaded deep neural network architecture. The first level of this architecture consists of multiple parallel feature encoding sub-networks, each specializing in processing feature sets extracted from a single signal source, such as a vibration feature sub-network or a temperature feature sub-network.

[0012] Each feature encoding subnetwork consists of three fully connected layers, mapping the input features to a 128-dimensional latent space representation through a non-linear activation function. The second stage of this architecture is a feature fusion network, whose input is the concatenation of the 128-dimensional latent space vectors output from all feature encoding subnetworks. The feature fusion network comprises two long short-term memory (LSM) layers and one self-attention mechanism layer. The LSM layers are used to capture the dynamic correlation and evolution patterns of multi-source feature sequences over time. The self-attention mechanism layer dynamically calculates the importance weights of features from different signal sources at different time steps for the current device state assessment, and performs weighted fusion of features based on these weights, ultimately outputting a 256-dimensional deep fused feature vector.

[0013] The equipment condition assessment module based on a physical information neural network is used to accurately quantify and assess the health status and failure modes of equipment based on the deep fusion feature vector output by the multi-scale deep feature fusion module and the physical mechanism constraints of the cold rolling process. This module consists of a physical information constraint layer and a condition assessment network. The physical information constraint layer incorporates state-related equations derived from the mechanical and thermodynamic models of the cold rolling process; these equations are embedded in the neural network training process as loss functions.

[0014] The state assessment network takes a deeply fused feature vector as input, and its output layer contains two parallel evaluation branches. The first evaluation branch is the health measurement branch, which outputs a continuous value between 0 and 1 to characterize the overall health status of the equipment, where 1 represents absolute health and 0 represents complete failure. The second evaluation branch is the fault mode recognition and localization branch, which outputs a multi-dimensional vector. Each dimension of the vector corresponds to a preset typical fault mode, including roll surface spalling, bearing inner ring wear, gear tooth breakage, motor rotor eccentricity, and process parameter mismatch. The value represents the probability of the fault mode occurring and the suspected fault location information inferred based on feature contribution.

[0015] The adaptive decision and control output module executes tiered early warning and control strategies based on the health status value and fault probability vector output by the equipment status assessment module. This module has three built-in progressive decision thresholds: a monitoring threshold, an early warning threshold, and a shutdown threshold. When the health status value is higher than the monitoring threshold, the module does not generate any active output, but only archives the status data. When the health status value is lower than the monitoring threshold but higher than the early warning threshold, the module generates a first-level early warning signal. This signal triggers key data monitoring and trend analysis of the relevant equipment components, and pushes the analysis report to the maintenance personnel's terminal.

[0016] When the health status value is below the warning threshold but above the shutdown threshold, the module generates a secondary warning signal and a preliminary control suggestion signal. The secondary warning signal triggers an audible and visual alarm. The preliminary control suggestion signal, based on the identified fault mode probability, sends a fine-tuning instruction to the rolling mill process control system, such as suggesting a 2% to 5% reduction in rolling speed or adjusting the bending roll force setting to alleviate equipment stress. When the health status value is below the shutdown threshold, or the probability of any fault mode exceeds its corresponding safety threshold, the module generates a tertiary emergency shutdown signal and a fault location report. The tertiary emergency shutdown signal is sent directly to the mill main drive control system with the highest priority to execute a safe shutdown sequence. The fault location report details the fault mode, probability, and suspected location, and pushes it to the maintenance terminal and production scheduling system.

[0017] In one embodiment of the present invention, the wavelet thresholding denoising process in the signal preprocessing and feature extraction module employs an improved threshold function and a hierarchical thresholding strategy. The improved threshold function is a continuously differentiable function between hard and soft thresholding, whose expression maximizes the preservation of useful abrupt changes in the signal while ensuring denoising effectiveness. The hierarchical thresholding strategy adaptively adjusts the threshold value according to the different scales of the wavelet decomposition coefficients; for the high-frequency coefficient layer, a threshold derived based on a general thresholding criterion is used. For the low-frequency coefficient layer, a threshold derived from a robust estimation method is used. ;

[0018] The The calculation formula is ,in The standard deviation of noise. The number of signal sampling points; The calculation formula is: ,in The preset adjustment coefficient, This represents the absolute deviation of the median of the wavelet coefficients in the low-frequency coefficient layer.

[0019] As one embodiment of the present invention, the computation process of the self-attention mechanism layer in the multi-scale deep feature fusion module is as follows. First, the feature sequence output by the Long Short-Term Memory network layer is used as the input source for the query vector, key vector, and value vector. Next, the dot product of the query vector and all key vectors is calculated, and an attention score is obtained through scaling. Then, a normalized exponential function is applied to the attention score to transform it into a weight distribution. Finally, the weight distribution and the value vector are weighted and summed to obtain the context-aware fusion feature at the current time step. This mechanism enables the model to dynamically focus on the feature sources and historical moments most relevant to the current abnormal state of the device.

[0020] As one embodiment of the present invention, the method for constructing the physical information constraint layer in the equipment condition assessment module based on a physical information neural network is as follows: Based on the cold rolling process mechanism, a set of correlation equations is established between key process parameters such as rolling force, rolling torque, and roll thermal crown, and vibration spectrum characteristics and temperature distribution characteristics. These correlation equations are then discretized and added as additional regularization loss terms to the total loss function of the condition assessment network.

[0021] For a batch of training samples, in addition to calculating the cross-entropy loss or mean squared error loss between the network's predicted output and the true label, the difference between certain feature representations of the intermediate layer of the network and the theoretical values ​​calculated from the process parameters according to the physical equation is also calculated as the physical information loss term.

[0022] During model training, this physical information loss term, together with the traditional cross-entropy loss and mean squared error loss, undergoes backpropagation and parameter optimization, thereby deeply embedding domain knowledge into the neural network, constraining its output results to conform to physical laws, and improving the model's generalization ability and the physical interpretability of the evaluation results under data-scarce conditions.

[0023] In one embodiment of the present invention, the generation logic of the preliminary control suggestion signal in the adaptive decision and control output module is based on a lightweight reinforcement learning agent. This agent takes the current equipment status assessment result, real-time process parameters, and production quality indicators as state inputs. The agent's action space is defined as a series of discrete process parameter fine-tuning actions, such as slight decreases in rolling speed, slight adjustments in tension, and slight increases in coolant flow rate. The agent's objective is to maximize the comprehensive reward function, which simultaneously considers the degree of improvement in equipment health, the stability of production quality, and the minimization of production losses. Through online learning and simulation training, the agent can learn the optimal process adjustment strategy under specific failure modes, thereby ensuring that the control suggestions not only aim to protect the equipment but also consider the continuity and economy of the production process.

[0024] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0025] 1. This invention constructs a multi-scale deep feature fusion module and employs a cascaded deep neural network and self-attention mechanism to achieve deep and adaptive fusion of time-domain, frequency-domain, and time-frequency-domain features extracted from multi-source heterogeneous signals such as vibration, temperature, pressure, and current. This technique overcomes the limitations of existing systems that rely on independent multi-parameter analysis and lack of correlation. It can effectively extract comprehensive and highly discriminative features characterizing early, subtle faults in equipment from high-dimensional noise backgrounds, significantly improving the system's detection sensitivity and accuracy for progressive faults, such as initial wear of rolls and micro-pitting of bearings, thus enabling proactive fault warning.

[0026] 2. This invention innovatively introduces a device condition assessment module based on a physical information neural network, embedding the mechanism model of the cold rolling process into the data-driven neural network training process in the form of constraint loss. This technical solution effectively integrates the adaptive capability of the data-driven method with the interpretability and extrapolation capability of the physical model, making the condition assessment model not only dependent on historical data but also constrained by physical laws. This significantly improves the robustness and reliability of the system's condition assessment under new operating conditions and new material conditions not covered by the training data, avoiding the physically unreliable output that may occur with purely data-driven models, and providing a solid technical foundation for intelligent diagnosis from perception to cognition.

[0027] 3. This invention designs an adaptive decision-making and control output module with multi-level thresholds and intelligent decision-making capabilities. This module not only achieves a closed loop from state monitoring to control execution, but also introduces reinforcement learning-based control suggestion generation logic, enabling the system to provide optimized adjustment schemes that balance equipment safety and production efficiency while issuing early warnings. This evolution from passive alarm to proactive intervention effectively avoids unplanned downtime, delays fault development through small-scale adaptive adjustments to process parameters, and buys time for planned maintenance. This maximizes the overall operational efficiency and economic benefits of the production line while ensuring long-term reliable equipment operation. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the overall technical solution architecture of a multi-parameter real-time monitoring and control system for the cold rolling process of steel and non-ferrous metals proposed in this invention.

[0029] Figure 2 This is a schematic diagram of the core principle framework of the multi-scale deep feature fusion module in this invention;

[0030] Figure 3 This is a logical flowchart of the device status assessment module based on physical information neural network in this invention;

[0031] Figure 4 This is a schematic diagram of the multi-level interaction relationship and data flow of the adaptive decision-making and control output module in this invention. Detailed Implementation

[0032] This invention provides a multi-parameter real-time monitoring and control system for the cold rolling process of steel and non-ferrous metals. Please refer to the appendix. Figures 1 to 4 This system constructs a closed-loop intelligent monitoring and control architecture, encompassing data acquisition, feature fusion, status assessment, and decision output. The system includes a multi-source heterogeneous data acquisition module, a signal preprocessing and feature extraction module, a multi-scale deep feature fusion module, a device status assessment module based on physical information neural networks, and an adaptive decision and control output module. These modules interact with a real-time database via high-speed industrial Ethernet, forming a collaborative and organic whole.

[0033] The multi-source heterogeneous data acquisition module is used to acquire multi-dimensional real-time operating signals of key equipment during the cold rolling process in parallel. This module specifically includes vibration sensing units, temperature sensing units, pressure sensing units, current and voltage sensing units, and process parameter acquisition units. The vibration sensing units are deployed at key measuring points on the mill stand, roll bearing housings, and main drive gearbox. They employ triaxial integrated circuit piezoelectric accelerometers with a range of ±50g, synchronously acquiring three-dimensional vibration signals at a sampling frequency of no less than 20 kHz. The sensor output signal from each measuring point is transmitted to a distributed data acquisition station located near the mill via a 4-20 mA current loop or a direct digital output interface.

[0034] The temperature sensing unit combines a non-contact infrared thermal imager with embedded K-type armored thermocouples. The infrared thermal imager performs a full-field temperature scan of the work roll and support roll surface at a rate of 100 frames per second, generating a temperature distribution matrix. The embedded thermocouples are precisely installed inside the stator winding of the main drive motor, the outer ring of the roll bearing, and the gearbox lubrication circuit, collecting point temperature data at a sampling frequency of no less than 100 Hz. The pressure sensing unit is integrated into the pressure transmitters of the mill's hydraulic pressing system and bending roll system. These transmitters are directly connected in series on the hydraulic lines, acquiring analog signals of rolling force and bending roll force in real time, with an accuracy of ±0.1% of full scale. The current and voltage sensing unit uses open-type high-precision Hall effect sensors to acquire the instantaneous waveforms of the three-phase current and voltage of the main motor and winding motor at a sampling frequency of 10 kHz, and calculates their effective values, harmonic content, and other electrical parameters.

[0035] The process parameter acquisition unit acquires rolling speed, reduction rate, front and rear tension setpoints, coolant flow rate setpoints, and strip material grade, width, and thickness specifications in real time from the programmable logic controller and process control computer of the rolling line's basic automation system via open industry standard protocols such as OPCUA or ModbusTCP. All sensing units are equipped with independent high-precision temperature-controlled crystal oscillators and synchronize with the master clock server at the nanosecond level via the IEEE 1588 precision time protocol, laying the foundation for subsequent multi-source data alignment.

[0036] The signal preprocessing and feature extraction module is used to perform denoising, alignment, and initial feature extraction on the raw signals acquired by the multi-source heterogeneous data acquisition module. This module is deployed in an industrial edge computing server and receives the raw data streams from the data acquisition modules. The module first performs wavelet thresholding preprocessing on each raw signal. The preprocessing process employs an improved threshold function and a hierarchical thresholding strategy. The improved threshold function is a continuously differentiable function between hard and soft thresholding, and its mathematical expression is:

[0037]

[0038] in, Indicates the first Layer scale, first Wavelet coefficients at each position, These are the denoised coefficients. For the threshold, For shape parameters. This function is applied at a threshold. The wavelet decomposition algorithm is continuous and has derivatives, ensuring denoising while maximizing the preservation of useful abrupt changes in the signal, especially for transient pulses characterizing impact faults. The hierarchical thresholding strategy adaptively adjusts the threshold size based on the scale of the wavelet decomposition coefficients: for high-frequency coefficient layers containing more noise details (i.e., smaller-scale decomposition layers in wavelet decomposition), a threshold derived based on a general thresholding criterion is used. The The calculation formula is ,in The noise standard deviation is estimated from the highest frequency layer coefficients of wavelet decomposition using "median absolute deviation / 0.6745". The number of signal sampling points; for the low-frequency coefficient layer containing the main signal components (i.e., the decomposition layer with a larger scale in wavelet decomposition), a threshold obtained based on a robust estimation method (median absolute deviation method) is used. The The calculation formula is: ,in The preset adjustment coefficient (with a value range of 1.2 to 1.8, preferably 1.5) is used. This represents the absolute deviation of the median of the wavelet coefficients in this low-frequency coefficient layer; and is calculated using the above formula. Always greater than This ensures that high-frequency layer noise is sufficiently suppressed, while the main signal components of the low-frequency layer are not excessively suppressed.

[0039] Furthermore, this module performs timestamp alignment on the preprocessed signals. The alignment algorithm uses the rolling speed signal provided by the process parameter acquisition unit as the reference time axis, as this signal is directly related to the physical position of the strip. The algorithm first checks the timestamp sequences of all data channels, identifying and removing abnormal timestamps caused by network latency or packet loss. Subsequently, using cubic spline interpolation, the sampling sequences of vibration, temperature, pressure, current, and other signals are resampled onto a unified time grid that is strictly synchronized with the reference time axis, with a minimum interval of 1 millisecond. After alignment, the module executes three sub-processes in parallel: time-domain feature extraction, frequency-domain feature extraction, and time-frequency-domain feature extraction.

[0040] The time-domain feature extraction sub-process calculates the root mean square (RMS), peak factor, kurtosis index, and waveform index of the signal within a sliding time window of 1024 sampling points. The RMS reflects the average energy of the signal; the peak factor is the ratio of the peak value to the RMS and is sensitive to impact; the kurtosis index is the ratio of the fourth central moment to the square of the variance, used to detect the sharpness of the signal distribution; and the waveform index is the ratio of the RMS to the absolute mean. The frequency-domain feature extraction sub-process first performs a Hanning windowed Fast Fourier Transform on the signal within the time window to calculate its power spectral density. Then, it extracts the frequency bands including the meshing frequency octave band, the bearing characteristic frequency band, and the energy percentage within the range of 0.5 to 3 times the rotational frequency; the energy sum within the characteristic frequencies of rolling bearing outer ring faults, inner ring faults, rolling element faults, and their sidebands; and the percentage of energy within the range of 0.5 to 3 times the rotational frequency.

[0041] The time-frequency domain feature extraction subprocess employs continuous wavelet transform, using the Morlet wavelet as the mother wavelet to generate a high-resolution time-frequency spectrum of the signal. From the time-frequency spectrum, wavelet energy entropy and singular spectral entropy are extracted. Wavelet energy entropy calculates the uncertainty of the energy distribution of wavelet coefficients at different scales; singular spectral entropy calculates the uncertainty of the singular value distribution by performing singular value decomposition on the time-frequency spectrum matrix. Both together characterize the non-stationarity and complexity of the signal. All extracted feature values ​​are organized into structured feature vectors, tagged with a unified timestamp and device identifier, and sent to the downstream buffer.

[0042] The multi-scale deep feature fusion module receives and deeply fuses multi-level features output by the signal preprocessing and feature extraction modules to construct a high-dimensional fused feature vector that can characterize the overall state of the device. Please refer to the appendix. Figure 2 The core of this module is a cascaded deep neural network architecture, which is deployed on an industrial server with graphics processor acceleration.

[0043] The first level of this architecture consists of multiple parallel feature encoding subnetworks, each specializing in processing a set of features extracted from a single signal source. Specifically, these include vibration feature subnetworks, temperature feature subnetworks, pressure feature subnetworks, current feature subnetworks, and process parameter feature subnetworks. Each feature encoding subnetwork comprises three fully connected layers, with modified linear units used as activation functions between layers. Batch normalization layers are also introduced to accelerate training and improve stability.

[0044] Taking the vibration feature subnetwork as an example, its input dimension is the total number of time-domain, frequency-domain, and time-frequency-domain features extracted from the vibration signal, assumed to be 45 dimensions. The first fully connected layer maps the 45-dimensional input to 256 dimensions; the second layer maps to 128 dimensions; and the third layer outputs a 128-dimensional latent space representation vector. Through a nonlinear activation function, each subnetwork maps its input feature set to a unified, highly expressive 128-dimensional latent space representation, which captures the deep nonlinear relationships of the internal features of the signal source.

[0045] The second level of this architecture is a feature fusion network, whose input is the concatenation of 128-dimensional latent space vectors output from all feature encoding subnetworks. Assuming there are 5 signal sources, the concatenated vector has a dimension of 640. The feature fusion network first contains two layers of Long Short-Term Memory (LSTM) networks. The LSM layers process the concatenated feature vectors at consecutive time steps in a time-series manner, with their hidden state dimension set to 256.

[0046] The cell states and hidden states of the Long Short-Term Memory (LSTM) network layer can capture the dynamic correlation and evolution patterns of multi-source feature sequences over time, such as the gradual accumulation process or periodic change patterns of fault features. Next, the feature fusion network includes one self-attention mechanism layer. The computation process of the self-attention mechanism layer is as follows: First, the 256-dimensional feature sequence output from the last time step of the LSM network layer is transformed into a query vector, a key vector, and a value vector through three independent linear transformation matrices.

[0047] Next, the dot product of the query vector and all key vectors is calculated, and then scaled by dividing by the square root of the key vector dimension to obtain the original attention score matrix. Then, a normalized exponential function is applied to each row of the attention score matrix to transform it into a weight probability distribution, which reflects the importance of features from different historical time steps and different signal sources in the current device state.

[0048] Finally, the weighted probability distributions are summed with their corresponding value vectors to obtain a 256-dimensional context-aware fusion feature vector. This mechanism allows the model to dynamically focus on the feature sources and historical moments most relevant to the current abnormal state of the equipment. For example, in the case of early bearing wear, specific frequency bands in the vibration signal are automatically assigned higher weights. Ultimately, the output of this self-attention layer is further integrated by a fully connected layer to output a 256-dimensional deep fusion feature vector that integrates multi-source, multi-scale, and spatiotemporal equipment state information.

[0049] The equipment condition assessment module based on a physical information neural network is used to accurately quantify and assess the health status and failure modes of equipment based on the deep fusion feature vector output by the multi-scale deep feature fusion module and the physical mechanism constraints of the cold rolling process. Please refer to the appendix. Figure 3 This module consists of a physical information constraint layer and a state assessment network. The physical information constraint layer is not an executable computational layer, but rather a set of state correlation equations derived from the cold rolling process mechanism, existing in the form of loss functions. These equations are constructed as follows: Based on the elastoplastic deformation theory, tribology, and heat conduction theory of the cold rolling process, a set of correlation equations is established between key process parameters such as rolling force, rolling torque, and roll thermal crown, and observable vibration spectrum characteristics and temperature distribution characteristics. For example, there is a definite functional relationship between rolling force and strip deformation resistance, friction coefficient, and reduction rate; while abnormal rolling force fluctuations will excite specific modal vibrations of the mill structure, reflected in the vibration spectrum including the meshing frequency octave band, bearing characteristic frequency band, and the energy increase in the frequency band proportion within the range of 0.5 to 3 times the rotational frequency.

[0050] For example, the surface temperature distribution of a roll is related to rolling speed, cooling conditions, and heat flux density in the contact arc area; abnormal hot spots may indicate wear or uneven cooling. These continuous sets of physical correlation equations are discretized and linearized to approximate a series of equality or inequality constraints. During the training phase of the state evaluation network, these constraints are constructed as additional regularization loss terms and added to the network's total loss function. Specifically, for a batch of training samples, in addition to calculating the cross-entropy loss or mean squared error loss between the network's predicted output and the true label, the difference between certain feature representations in the intermediate layers of the network and the theoretical values ​​calculated from the process parameters according to the physical equations is also calculated as a physical information loss term. This loss term forces the neural network, during the learning process, to ensure that its internal representation and output prediction conform to known physical laws to a certain extent.

[0051] The state evaluation network takes a 256-dimensional deep fusion feature vector as input. Its main structure consists of four fully connected layers, each followed by a modified linear unit activation and dropout layer to prevent overfitting. The network's output layer contains two parallel evaluation branches. The first evaluation branch is a health measurement branch, which outputs a continuous value between 0 and 1 through a single-neuron output layer with a sigmoid activation function, representing the overall health status of the equipment. A value of 1 represents an ideal state where the equipment is absolutely healthy and shows no signs of degradation; a value of 0 represents complete failure and inoperability. This value is continuous and reflects the gradual change in health. The second evaluation branch is a fault mode recognition and localization branch, which outputs a multi-dimensional vector. The dimension of the vector corresponds to the number of preset typical fault mode types, for example, set to 5 dimensions, corresponding to roller surface spalling, bearing inner ring wear, gear tooth breakage, motor rotor eccentricity, and process parameter mismatch.

[0052] Each dimension's value is calculated using the Softmax function or an independent Sigmoid function, representing the probability of that specific fault mode occurring, with values ​​ranging from 0 to 1. Furthermore, this branch contains a substructure based on gradient-based activation mapping, which can inversely deduce the original signal source and feature type that contributes most to the current fault probability based on the network's sensitivity to input features. This, combined with the equipment's 3D model, provides suspected fault location information, such as "the inner ring of the bearing on the drive side of the work roller under frame 3". During model training, the physical information loss term, along with traditional cross-entropy loss and mean squared error loss, undergoes backpropagation and parameter optimization, deeply embedding domain knowledge into the neural network and constraining its output to conform to physical laws. This significantly improves the robustness and reliability of the model's state assessment under new operating conditions and material conditions not covered by the training data.

[0053] The adaptive decision-making and control output module is used to execute tiered early warning and control strategies based on the health value and fault probability vector output by the equipment status assessment module. Please refer to the appendix. Figure 4 This module incorporates three progressive decision thresholds: an attention threshold, an early warning threshold, and a shutdown threshold. These thresholds are not fixed values ​​but are dynamically set and periodically adjusted based on the equipment model, historical operating data, and the importance of the current production task, using an offline optimization algorithm. For example, for the main drive gearbox of a critical rack, the attention threshold might be set to 0.92, the early warning threshold to 0.85, and the shutdown threshold to 0.70. When the health status value exceeds the attention threshold, the module does not generate active control output; instead, it packages the current status assessment results, raw feature data, and process parameters in a compressed format and stores them in a long-term historical database for subsequent model retraining and performance analysis.

[0054] When the health status value is below the attention threshold but above the warning threshold, the module generates a level one warning signal. This signal triggers a monitoring subroutine within the system. This subroutine doubles the data sampling frequency of relevant equipment components, such as the specific bearing housing that issued the warning, and starts an independent trend analysis thread. The trend analysis thread uses an exponentially weighted moving average algorithm and linear regression to predict the health status trend of the component over the next 8 hours. Simultaneously, an analysis report containing the current status, historical curves, and trend predictions is generated and pushed non-blockingly via the enterprise message bus to the mobile terminals and fixed workstations of maintenance personnel responsible for the area, prompting them to strengthen their inspection and monitoring.

[0055] When the health status value is below the warning threshold but above the shutdown threshold, the module simultaneously generates a secondary warning signal and a preliminary control suggestion signal. The secondary warning signal triggers multimedia audio-visual alarms installed in the control room and at the equipment site, emitting a specific combination of sound and light to remind operators that the equipment status has entered the warning zone. The generation logic of the preliminary control suggestion signal is based on a lightweight reinforcement learning agent. This agent takes the current equipment status assessment result, real-time process parameters, and the thickness and shape quality indicators of the most recent 100 meters of strip steel as status inputs. The agent's action space is defined as a series of discrete process parameter fine-tuning actions, such as a slight decrease of 2% in rolling speed, a slight decrease of 5% in rolling speed, a slight increase of 1% in front tension, a slight decrease of 1% in back tension, a slight positive adjustment of the bending roll force of 3%, and a slight increase of 10% in coolant flow rate. The agent's goal is to maximize the comprehensive reward function, which simultaneously considers the degree of improvement in equipment health, the stability of production quality, and the minimization of output loss. Reward Function The specific form is as follows:

[0056]

[0057] in, It is the predicted change in equipment health after an action is taken. It is the predicted change in the standard deviation of production quality indicators. It is the predicted change in output per unit time. , , These are weighting coefficients. The reinforcement learning agent is trained extensively in an offline simulation environment consisting of a high-fidelity rolling process model and an equipment degradation model. During online operation, the agent outputs optimal fine-tuning actions from its policy network based on the current state, generating preliminary control suggestion signals. These signals are sent to the rolling line process control system as suggestion commands via a secure, authenticated application programming interface. Upon receiving the suggestion, the process control system requires operator confirmation or must operate in automatic mode before executing actions such as reducing the rolling speed by 3% or adjusting the bending roll force setpoint to proactively alleviate equipment stress and slow the progression of faults.

[0058] When the health status value falls below the shutdown threshold, or the probability of any fault mode exceeds its corresponding safety threshold (e.g., the probability of bearing inner ring wear exceeds 0.95), the module immediately generates a Level 3 emergency shutdown signal and a detailed fault location report. The Level 3 emergency shutdown signal is sent directly to the safety programmable logic controller (PLC) of the mill's main drive control system with the highest priority, via a hard-wired safety loop and a high-speed real-time Ethernet dual-channel. Upon receiving the signal, the PLC ignores any other process instructions and immediately executes a predefined safety shutdown sequence, including sequentially cutting off the main motor power, engaging mechanical braking, and releasing rolling force and tension, ensuring a smooth and rapid stop for the equipment. The fault location report details the triggered fault modes, their specific probability values, the suspected locations calculated based on feature inversion, and associated sensor reading snapshots. This report is simultaneously pushed to the maintenance terminal, production scheduling system, and factory manufacturing execution system via redundant network paths, providing accurate information for rapid response and spare parts preparation in emergency maintenance.

[0059] The entire system's software architecture adopts a microservice design, with each core module running as an independent microservice on a containerized platform. Modules communicate loosely coupled via publish-subscribe message queues. The system is equipped with a unified management and monitoring interface that displays the real-time operating status of each module, data flow graphs, equipment health curves, warning logs, and control command history. All algorithm models support online incremental learning and periodic full updates, with model versions and parameter configurations managed by a version control system. The system is deeply integrated with the factory's existing equipment management system, maintenance management system, and production execution system, achieving a closed-loop data process encompassing monitoring, diagnosis, maintenance, and production decision-making.

Claims

1. A multi-parameter real-time monitoring and control system for cold rolling processes of steel and non-ferrous metals, characterized in that, include: A multi-source heterogeneous data acquisition module is used to acquire multi-dimensional real-time operating signals of key equipment in the cold rolling process in parallel. The signal preprocessing and feature extraction module is used to perform noise reduction, alignment and feature extraction on the raw signal acquired by the multi-source heterogeneous data acquisition module; The multi-scale deep feature fusion module is used to receive and deeply fuse multi-level features output by the signal preprocessing and feature extraction module to construct a high-dimensional fusion feature vector characterizing the overall state of the device. The equipment condition assessment module based on the physical information neural network is used to quantitatively assess the health status and fault modes of equipment based on the deep fusion feature vector output by the multi-scale deep feature fusion module and in combination with the physical mechanism constraints of the cold rolling process. The equipment condition assessment module based on the physical information neural network includes a physical information constraint layer and a condition assessment network. The condition assessment network includes a health quantification branch and a fault mode recognition and localization branch. The physical information constraint layer is constructed as follows: based on the cold rolling process mechanism, a set of correlation equations is established between key process parameters and vibration spectrum characteristics and temperature distribution characteristics; these correlation equations are discretized and added as an additional regularization loss term to the total loss function of the condition assessment network. During model training, the physical regularization loss term, together with the traditional cross-entropy loss and mean squared error loss, undergoes backpropagation and parameter optimization. An adaptive decision-making and control output module is used to execute a graded early warning and control strategy based on the health value and fault probability vector output by the equipment status assessment module based on the physical information neural network. The adaptive decision-making and control output module has built-in attention thresholds, early warning thresholds, and shutdown thresholds. The generation logic of the initial control suggestion signal in the adaptive decision-making and control output module is based on a lightweight reinforcement learning agent. The reinforcement learning agent uses the current equipment status assessment result, real-time process parameters, and production quality indicators as status inputs. The action space of the reinforcement learning agent is defined as a series of discrete process parameter fine-tuning actions. The goal of the reinforcement learning agent is to maximize the comprehensive reward function, which simultaneously considers the degree of improvement in equipment health, the stability of production quality, and the minimization of output loss. The multi-source heterogeneous data acquisition module includes a vibration sensing unit, a temperature sensing unit, a pressure sensing unit, a current and voltage sensing unit, and a process parameter acquisition unit. The vibration sensing unit is deployed at key measuring points of the mill stand, roll bearing housing and main drive gearbox, and uses an accelerometer to collect three-dimensional vibration signals at a sampling frequency of not less than 20 kHz. The temperature sensing unit uses a combination of an infrared thermal imager and an embedded thermocouple, with a sampling frequency of not less than 100 Hz; the pressure sensing unit is integrated into the rolling mill hydraulic pressing system and the bending roll system. The current and voltage sensing unit acquires the three-phase current and voltage waveforms of the main motor and the winding motor through a high-precision Hall sensor. The process parameter acquisition unit obtains rolling speed, reduction rate, tension set value, and strip material and specification information from the basic automation system of the rolling line in real time. The signal preprocessing and feature extraction module includes a time-domain feature extraction unit, a frequency-domain feature extraction unit, and a time-frequency-domain feature extraction unit that are executed in parallel. The time-domain feature extraction unit calculates the root mean square value, peak factor, kurtosis index, and waveform index of the signal. The frequency domain feature extraction unit extracts frequency bands including the meshing frequency octave band, the bearing characteristic frequency band, and the energy proportion in the range of 0.5 to 3 times the rotation frequency by performing a fast Fourier transform on the signal. The time-frequency domain feature extraction unit uses continuous wavelet transform to generate the time-frequency spectrum of the signal and extracts wavelet energy entropy and singular spectrum entropy from it.

2. The multi-parameter real-time monitoring and control system for cold rolling processes of steel and non-ferrous metals according to claim 1, characterized in that, The denoising process in the signal preprocessing and feature extraction module adopts wavelet threshold denoising preprocessing, and the wavelet threshold denoising process in the signal preprocessing and feature extraction module adopts an improved threshold function and a hierarchical threshold strategy. The improved threshold function is a continuously differentiable function that lies between a hard threshold and a soft threshold. The hierarchical thresholding strategy adaptively adjusts the threshold size according to the different scales of the wavelet decomposition coefficients. For the high-frequency coefficient layer, a threshold derived based on a general thresholding criterion is used. For the low-frequency coefficient layer, a threshold derived from a robust estimation method is used. ; The The calculation formula is ,in The standard deviation of noise. The number of signal sampling points; The calculation formula is: ,in The preset adjustment coefficient, This represents the absolute deviation of the median of the wavelet coefficients in the low-frequency coefficient layer.

3. The multi-parameter real-time monitoring and control system for cold rolling processes of steel and non-ferrous metals according to claim 1, characterized in that, The calculation process of the self-attention mechanism layer in the multi-scale deep feature fusion module is as follows: First, the feature sequence output by the Long Short-Term Memory network layer is used as the input source for the query vector, key vector, and value vector; Next, the dot product of the query vector and all key vectors is calculated, and the attention score is obtained through scaling. Then, a normalized exponential function is applied to the attention scores to transform them into a weight distribution; Finally, the weight distribution and the value vector are weighted and summed to obtain the context-aware fusion feature of the current time step.

4. A multi-parameter real-time monitoring and control system for cold rolling processes of steel and non-ferrous metals according to claim 1, characterized in that, The multi-scale deep feature fusion module includes a cascaded deep neural network architecture. The first level of this architecture consists of multiple parallel feature encoding sub-networks, and the second level consists of a feature fusion network containing a long short-term memory network layer and a self-attention mechanism layer. The feature encoding subnetwork consists of three fully connected layers, which use a non-linear activation function to map the input features to a 128-dimensional latent space representation. The input to the feature fusion network is the concatenation of 128-dimensional latent space vectors output by all feature encoding subnetworks, and the final output is a 256-dimensional deep fusion feature vector.

5. A multi-parameter real-time monitoring and control system for cold rolling processes of steel and non-ferrous metals according to claim 1, characterized in that, The health quantification branch outputs a continuous value between 0 and 1, which is used to characterize the overall health status of the device. The fault mode recognition and localization branch outputs a multi-dimensional vector. Each dimension of the vector corresponds to a preset typical fault mode, and its value represents the probability of the fault mode occurring and the suspected fault location information inferred based on the feature contribution.

6. A multi-parameter real-time monitoring and control system for cold rolling processes of steel and non-ferrous metals according to claim 1, characterized in that, The adaptive decision-making and control output module executes a tiered strategy based on the comparison result between the health score and the threshold: When the health score is below the attention threshold but above the warning threshold, a level one warning signal is generated and key data monitoring and trend analysis are triggered. When the health status value is lower than the warning threshold but higher than the shutdown threshold, a secondary warning signal and a preliminary control suggestion signal are generated. When the health value is lower than the shutdown threshold, or the probability of any fault mode exceeds its corresponding safety threshold, a level three emergency shutdown signal and fault location report are generated.

Images