An adaptive temperature control and early warning method for pickling process
By combining deep residual auto-coding and differentiable causal discovery networks, the problems of response lag and fuzzy risk positioning in traditional pickling process control systems under complex operating conditions are solved. This enables adaptive temperature control and early warning in the pickling process section, improves control accuracy and equipment safety, and ensures flexible production and efficient energy consumption management.
Patent Information
- Application Number
- CN202611457460.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-18
- Publication Date
- 2026-11-17
AI Technical Summary
Traditional pickling process control systems suffer from problems such as limited state assessment dimensions, delayed response, ambiguous risk identification, and uncoordinated control responses when facing complex and fluctuating operating conditions, leading to increased thermal stress and energy consumption in equipment.
By employing a deep residual autoencoder network and a differentiable causal discovery network, combined with temperature gradient positive correlation constraints and thermal inertia continuity constraints of adjacent tank sections, a low-dimensional process state characterization vector is generated. Through causal contribution analysis and counterfactual reasoning, adaptive temperature control and early warning are achieved, the deep influence paths in the process state evolution are analyzed, and the control strategy is optimized to avoid extreme regulation.
It improves the control precision and equipment operation safety of the pickling process, realizes the intelligent dynamic balance between pickling quality and heat consumption of strip steel surface, avoids equipment damage and energy waste, and enhances the flexibility and robustness of production.
Smart Images

Figure CN123113375A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of industrial metal surface treatment technology, specifically to an adaptive temperature control and early warning method for pickling processes. Background Technology
[0002] The continuous strip pickling production line is a core link in the metallurgical cold rolling process. The operational efficiency of its pickling section directly determines the quality of oxide scale removal from the strip surface and the energy consumption of the entire production line. In traditional control architectures, the underlying control system typically relies on conventional feedback regulation mechanisms or basic steady-state mechanism analysis models to perform isolated single-loop temperature tracking for each independent physical pickling tank. However, the internal physical reactions of the continuous pickling process are highly complex, involving tank liquid flow, continuous strip threading and winding, and high-frequency interactions between various heat exchange components. Existing technologies often treat the massive amounts of multi-source operating variables collected as parallel discrete data, seriously neglecting the physical coordination and correlation of various operating conditions during dynamic evolution. Furthermore, they sever the constraints of temperature gradient transmission and the extremely strong thermal inertia continuity between adjacent tank sections. This results in traditional control models exhibiting significant defects when facing complex fluctuating operating conditions, such as a single dimension of state assessment, severe response lag, and low global coordinated scheduling efficiency.
[0003] Existing pickling process control schemes' quality detection and status early warning algorithms are mostly limited to shallow fitting of surface statistical correlations, failing to penetrate the complex process phenomena to reveal the substantial causal driving forces behind the intertwined and interconnected factors of equipment performance degradation, drastic changes in control parameters, and abnormal incoming material attributes. This black-box mapping approach, lacking the decomposition of marginal partial derivative effects and the quantification of causal contributions, makes the system prone to ambiguous risk tracing when facing quality degradation risks, thus transmitting control weights that deviate from the core to the execution end. Traditional multi-objective optimization control strategies generally lack forward-looking virtual control simulations and thermal load boundary verification barriers before issuing global temperature setpoints. Due to the failure to introduce counterfactual intervention and review mechanisms to proactively determine the chain of risks that local temperature commands may trigger, the system's blindly issued control commands are prone to exceeding the physical minimum power limit or maximum output extreme value of the underlying heating equipment in the process section. This mechanical optimization, lacking safety boundary filtering, not only makes it difficult to maintain a smooth adaptive transition between quality feedback and energy consumption, but also causes destructive sudden changes in absolute temperature difference between adjacent tank sections due to unreasonable gradient distribution. The surge in equipment thermal stress not only exacerbates the mechanical fatigue of the heat exchange structure, but also completely disrupts the safe and stable operation of the entire pickling process section.
[0004] To address the aforementioned issues, there is an urgent need in this field for a novel adaptive temperature control and early warning collaborative solution that can deeply integrate deep physical operational constraints with dynamic tracking and mining mechanisms of causal relationships. This technology urgently needs to be able to penetrate multi-variable coupling barriers under the precise characterization of high-dimensional process states, achieving layer-by-layer stripping of risk contribution and adaptive iteration of control preferences. Furthermore, it should use forward-looking virtual simulation intervention to completely block unreasonable control directions, thereby completely eliminating non-compliant heat load span deviations within a closed-loop multi-objective optimization framework. This will enable a safe, flexible, and highly intelligent balance between the surface cleaning quality of the strip steel and the underlying heat energy consumption. Summary of the Invention
[0005] To address the aforementioned technical problems, an adaptive temperature control and early warning method for the pickling process is provided. This technical solution solves the problems mentioned above.
[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: An adaptive temperature control and early warning method for a pickling process section includes: Step S1: Collect multi-source operation data streams of the pickling process section, input the multi-source operation data streams into the deep residual auto-encoding network, and add temperature gradient positive correlation constraint terms and adjacent tank thermal inertia continuity constraint terms in the encoding and reconstruction stage, and output a low-dimensional process state characterization vector. Step S2: Input the low-dimensional process state characterization vector into the differentiable causal discovery network for directed edge parameter learning, and generate a dynamic causal effect matrix of the influence path of the characterization variable. Step S3: Based on the dynamic causal effect matrix decomposition, the marginal partial derivative influence values of each feature in the low-dimensional process state characterization vector are decomposed to calculate the hierarchical quality risk score set and the causal contribution vector of each corresponding variable node. Step S4: Based on the numerical distribution of each node in the causal contribution vector, generate the initial control preference vector of quality preference weight parameter and energy consumption preference weight parameter; Step S5: Determine whether the maximum score value in the stratified quality risk score set is greater than the preset risk warning limit. When the maximum score value is greater than the preset risk warning limit, enter the counterfactual reasoning intervention loop judgment. Step S6: In the counterfactual reasoning intervention loop judgment, multiple sets of local tank segment temperature offset commands are generated. The local tank segment temperature offset commands of each set are fused with the low-dimensional process state characterization vector and substituted into the dynamic causal effect matrix for forward inference calculation, and the counterfactual reasoning future risk score set is output. Step S7: Extract the minimum risk value from the counterfactual reasoning future risk score set, and correct the initial control preference vector to the target intervention preference vector according to the local tank segment temperature bias command direction that generates the minimum risk value. Step S8: Substitute the target intervention preference vector into the multi-objective optimization control model containing the mass loss objective function and the energy consumption evaluation objective function for iterative solution. Extract the minimum preference space distance solution in the frontier solution set of the multi-objective optimization control model as the target temperature control set value sequence and output it to the bottom controller for execution.
[0007] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention proposes an adaptive temperature control and early warning method for the pickling process, overcoming the limitations of single-loop discrete adjustment and shallow data fitting in traditional pickling control architectures. It deeply integrates physical heat transfer mechanism constraints, dynamic causal tracing, and forward-looking intervention mechanisms, significantly improving the control accuracy and equipment operation safety of the pickling process under multi-variable, strongly coupled conditions. By constructing an end-to-end adaptive temperature control and early warning closed-loop system, it not only completely reverses the passive situation of lagging control response and ambiguous risk positioning in existing technologies, but also achieves intelligent dynamic balance between strip surface pickling quality assessment and overall heat energy consumption while ensuring that the underlying heat exchange equipment is protected from extreme thermal stress damage. This provides optimized control assurance with both flexibility and high robustness for continuous production processes.
[0008] 2. In terms of specific technical approach, this invention effectively overcomes the inherent defects of fragmented multi-source monitoring data in previous methods. By deeply embedding the temperature gradient distribution and the thermal inertia continuity of adjacent tank sections as rigid physical constraints into the dimensionality reduction and reconstruction process, it successfully eliminates invalid data redundancy and extracts a high-fidelity process state characterization that truly reflects the thermodynamic evolution of the continuous production line. Furthermore, it completely abandons the black-box mode of traditional algorithms that rely on surface statistical correlations for risk assessment. Instead, it relies on a continuously differentiable causal optimization link to accurately analyze the deep directional influence paths in the process state evolution process. This progressive breakthrough from mechanism-driven to causal quantitative decomposition enables the control system to peel away layers and accurately locate the original driving sources of quality degradation in a complex environment interwoven with multiple disturbances from processes, equipment, and incoming materials. This lays a highly targeted and physically traceable data foundation for subsequent adaptive adjustment of quality and energy consumption preferences, fundamentally eliminating the problem of control weight allocation deviating from key causal nodes under dangerous operating conditions. 3. This solution addresses the critical weakness of traditional control systems, which lack forward-looking simulations and are prone to triggering extreme control commands. It introduces a dual anti-disturbance defense mechanism combining counterfactual reasoning loops and absolute boundary filtering. When a risk warning is triggered, the system no longer mechanically issues intervention settings directly. Instead, it first projects multiple sets of local temperature bias strategies onto a virtual causal link for scenario-based simulation. By optimizing the future risk projection results, it preemptively corrects intervention preferences, completely preventing cascading control instability that might result from blindly seeking optimization. In the final decision-making stage of multi-objective optimization, this solution transforms the operating power limits and allowable thermal stress spans of the underlying heat exchange equipment into rigid elimination criteria. Within the multi-dimensional solution space, it precisely eliminates all illegal setting sequences that could potentially cause destructive absolute temperature differences between adjacent physical tank sections. This type of optimization decision, which prioritizes equipment safety potential within a non-dominated sorting and screening framework, perfectly balances the continuity and smoothness of the speed regulation process, eliminates mechanical fatigue and breakage damage to heat exchange elements caused by frequent sudden heating and cooling, and ultimately outputs the leading-edge optimal solution that can effectively suppress the spread of quality fluctuations while strictly compressing the energy consumption of the medium to the most ideal state allowed by the process boundary. Attached Figure Description
[0009] Figure 1 This is a flowchart of an adaptive temperature control and early warning method for an acid pickling process.
[0010] Figure 2 A schematic diagram of the data flow for counterfactual reasoning intervention and multi-objective optimized temperature control. Detailed Implementation
[0011] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.
[0012] Reference Figure 1 As shown, an adaptive temperature control and early warning method for a pickling process section includes: This invention can be applied to the pickling process section of a continuous strip steel pickling production line, primarily addressing the issues of delayed bath temperature control and untimely quality risk warnings caused by the combined effects of coupled thermal states of various pickling tanks, strip speed fluctuations, and changes in incoming material properties. The method is achieved collaboratively by an industrial server installed in the pickling main control room, a low-level controller, a field temperature acquisition module, a strip speed acquisition module, and execution interfaces connected to the heating equipment of each physical tank section. The industrial server is responsible for calculating the deep residual autoencoding network, the differentiable causal discovery network, the counterfactual reasoning intervention loop determination, and the multi-objective optimization control model. The low-level controller receives the target temperature setpoint sequence and converts it into adjustment instructions for the heating equipment of each physical tank section.
[0013] Step S1: Collect multi-source operation data streams of the pickling process section, input the multi-source operation data streams into the deep residual auto-encoding network, and add temperature gradient positive correlation constraint terms and adjacent tank thermal inertia continuity constraint terms in the encoding and reconstruction stage, and output a low-dimensional process state characterization vector. First, the on-site temperature acquisition module and strip speed acquisition module collect multi-source operational data streams from the pickling process section. Taking a pickling process section with five acid tanks as an example, at least one tank temperature sensor is installed in each acid tank. The sampling period can be set to one to five seconds. The collected tank temperature time series of each acid tank is written into the real-time data buffer in the form of timestamp, acid tank number, and tank temperature value. The strip speed acquisition module reads the strip speed series from the unit's drive system. The data format includes the strip speed value at the same timestamp. To enable data of different dimensions to enter the same network, the industrial server aligns the tank temperature time series and strip speed series of each acid tank according to the same time window before inputting them into the deep residual autoencoder network. For example, a sample window is formed every sixty seconds, and normalization is performed according to the pre-saved mean and scale parameters. The processed multi-source operational data stream is fed into a deep residual autoencoder network. During the encoding and reconstruction stage, a positive correlation constraint term based on temperature gradient and a thermal inertia continuity constraint term based on adjacent strip segments are added. This ensures that the network retains the fundamental physical relationship between temperature changes and strip movement in the pickling process section when compressing the data. The low-dimensional process state representation vector output by the deep residual autoencoder network serves as the unified state input for subsequent causal learning, risk scoring, and control optimization.
[0014] Step S2: Input the low-dimensional process state characterization vector into the differentiable causal discovery network for directed edge parameter learning, and generate a dynamic causal effect matrix of the influence path of the characterization variable. After the low-dimensional process state representation vector is formed, the industrial server inputs it into a differentiable causal discovery network for directed edge parameter learning. The differentiable causal discovery network does not simply treat each variable as independent; instead, it learns the direction and intensity of influence between different feature nodes within the same computational graph, ultimately generating a dynamic causal effect matrix representing the influence paths of the representation variables. In this dynamic causal effect matrix, each row and column corresponds to a feature node in the low-dimensional process state representation vector, and the non-zero weight elements in the matrix represent a learnable directed connection path between two corresponding feature nodes. The dynamic causal effect matrix is then passed to the risk calculation thread to decompose the marginal partial derivative influence values of each feature within the low-dimensional process state representation vector.
[0015] Step S3: Based on the dynamic causal effect matrix decomposition, the marginal partial derivative influence values of each feature in the low-dimensional process state characterization vector are decomposed to calculate the hierarchical quality risk score set and the causal contribution vector of each corresponding variable node. In the risk calculation thread, the industrial server decomposes the influence of each feature within the low-dimensional process state representation vector based on the dynamic causal effect matrix. The system calculates the marginal partial derivative influence values of each feature node on the endpoint quality assessment feature node along the variable influence path given by the dynamic causal effect matrix, and categorizes these marginal partial derivative influence values into equipment bottleneck-related, process control-related, and incoming material characteristic-related components according to their source attributes. After normalization, the system obtains the equipment bottleneck risk score, process adjustability risk score, and incoming material induced risk score, and merges these three scores into a hierarchical quality risk score set. Simultaneously, the system generates a causal contribution vector based on the proportion of marginal partial derivative influence values for each variable node. Each element in the causal contribution vector corresponds to a physical node in the initial node set of the network, representing the proportion of that physical node's contribution to the current quality risk. This vector directly serves the generation of the subsequent initial control preference vector.
[0016] Step S4: Based on the numerical distribution of each node in the causal contribution vector, generate the initial control preference vector of quality preference weight parameter and energy consumption preference weight parameter; After obtaining the causal contribution vector, the preference generation module generates an initial control preference vector containing quality preference weight parameters and energy consumption preference weight parameters based on the numerical distribution of each node within the causal contribution vector. For example, when the proportion of values corresponding to the tank temperature node in the causal contribution vector is in a high range, it indicates that the current quality risk is more easily affected by tank temperature disturbances. In this case, the preference generation module will increase the quality preference weight parameter and correspondingly decrease the energy consumption preference weight parameter, so that the subsequent multi-objective optimization control model has a clear bias between the quality loss objective function and the energy consumption evaluation objective function. The initial control preference vector is stored in the form of two-dimensional parameter pairs, with the first dimension being the quality preference weight parameter and the second dimension being the energy consumption preference weight parameter, and is written into the shared memory area of the control cycle.
[0017] Step S5: Determine whether the maximum score value in the stratified quality risk score set is greater than the preset risk warning limit. When the maximum score value is greater than the preset risk warning limit, enter the counterfactual reasoning intervention loop judgment. Subsequently, the early warning judgment module determines whether the maximum score value within the hierarchical quality risk score set exceeds the preset risk warning limit. The preset risk warning limit can be set based on historical product surface quality judgment results; for example, when the score is normalized to zero to one, the preset risk warning limit is set to 0.7. When the maximum score value among the equipment bottleneck risk score, process adjustable risk score, and incoming material induced risk score exceeds the preset risk warning limit, it indicates that there is a high quality risk in the current operating condition. The system does not immediately and directly change the temperature setting of the underlying controller, but instead enters a counterfactual reasoning intervention loop judgment, first deducing the future risk changes that may be caused by temperature bias commands for different local tank segments in the dynamic causal effect matrix. This avoids overshoot caused by directly applying a large temperature adjustment based solely on the current risk score.
[0018] Step S6: In the counterfactual reasoning intervention loop judgment, multiple sets of local tank segment temperature offset commands are generated. The local tank segment temperature offset commands of each set are fused with the low-dimensional process state characterization vector and substituted into the dynamic causal effect matrix for forward inference calculation, and the counterfactual reasoning future risk score set is output. In the counterfactual reasoning intervention loop judgment, the industrial server generates multiple sets of local tank segment temperature bias instructions. Each set of local tank segment temperature bias instructions includes the tank segment position, temperature bias direction, and temperature bias magnitude. For example, the first and second acid tanks in the front pickling tank group are adjusted upward by one degree Celsius, and the fourth and fifth acid tanks in the rear pickling tank group are adjusted downward by one degree Celsius. After generation, the system fuses each set of local tank segment temperature bias instructions with the low-dimensional process state representation vector. The fusion process involves finding the feature positions related to the tank liquid temperature in the low-dimensional process state representation vector, converting the temperature bias magnitude to the same normalized scale as the feature position, and writing it into the corresponding position. The remaining equipment state feature nodes, process parameter feature nodes, and incoming material attribute feature nodes remain in their original states. The fused vector is substituted into the dynamic causal effect matrix for forward inference calculation. The industrial server propagates the impact of the temperature bias layer by layer according to the directed connection path in the dynamic causal effect matrix, obtaining the future risk results corresponding to each set of local tank segment temperature bias instructions, and organizing all future risk results into a counterfactual reasoning future risk score set. Specifically, the system first converts the temperature bias amplitude of each group into a corresponding dimensionless bias value using the same normalization parameters as in the training phase. Then, it copies the current low-dimensional process state representation vector and superimposes the dimensionless bias value only at the temperature-related feature positions of the bath liquid to obtain the intervention state vector corresponding to that group of commands; the original values are kept at non-battery liquid temperature feature positions. The system substitutes the intervention state vector into the dynamic causal effect matrix group by group to obtain the future risk results corresponding one-to-one with the local bath segment temperature bias commands of each group.
[0019] Step S7: Extract the minimum risk value from the counterfactual reasoning future risk score set, and correct the initial control preference vector to the target intervention preference vector according to the local tank segment temperature bias command direction that generates the minimum risk value. After the counterfactual reasoning future risk score set is output, the control strategy correction module extracts the minimum risk value. The system finds the direction of the local tank temperature bias command that generates the minimum risk value and corrects the initial control preference vector to the target intervention preference vector based on this direction. If the local tank temperature bias command that generates the minimum risk value is manifested as heating up the front pickling tank group and cooling down the rear pickling tank group, it indicates that the current quality risk is more suitable to be improved by strengthening the front reaction, and the target intervention preference vector will relatively increase the quality preference value. If the deduction results show that the risk of a slight, proportional increase in overall temperature is lower, the system will adjust the relative relationship between the quality preference value and the energy consumption preference value without significantly increasing the energy consumption burden. After the target intervention preference vector is generated, it is fed into the multi-objective optimization control model as the preference benchmark for the final optimization.
[0020] Step S8: Substitute the target intervention preference vector into the multi-objective optimization control model containing the mass loss objective function and the energy consumption evaluation objective function for iterative solution, and extract the minimum preference space distance solution in the frontier solution set of the multi-objective optimization control model as the target temperature control set value sequence and output it to the bottom controller for execution; In one specific embodiment, after the target intervention preference vector is determined, the industrial server substitutes the target intervention preference vector into a multi-objective optimization control model containing both a quality loss objective function and an energy consumption evaluation objective function for iterative solution. The multi-objective optimization control model reads the working boundaries of the heating equipment associated with each physical tank segment from the controller and generates a group of candidate temperature setpoint sequences within these boundaries. Each group of candidate temperature setpoint sequences represents a temperature setting scheme for a complete pickling process segment; for example, five physical tank segments correspond to five temperature setpoints. The multi-objective optimization control model calculates the quantified values of quality loss and energy consumption corresponding to each group of candidate temperature setpoint sequences, and sorts them through non-dominated hierarchical comparison to form the frontier solution set of the multi-objective optimization control model. Finally, the system selects the solution with the minimum preference space distance from the frontier solution set of the multi-objective optimization control model according to the target intervention preference vector, converts this solution into a target temperature control setpoint sequence, and outputs it to the lower-level controller for execution via the industrial communication bus. After receiving the target temperature control setpoint sequence, the lower-level controller distributes it to the temperature control loops of the heating equipment associated with each physical tank segment, enabling the pickling process segment to enter the next control cycle.
[0021] Regarding step S1, the input of multi-source operating data streams into the deep residual autoencoder network, and the addition of temperature gradient positive correlation constraint terms and adjacent tank segment thermal inertia continuity constraint terms in the encoding and reconstruction stage, outputting a low-dimensional process state characterization vector, includes: Collect multi-source operational data streams for the pickling process section; the multi-source operational data streams include: the time series of the bath temperature and the strip speed series of each pickling tank; The multi-source operation data stream of the pickling process section is input into the encoder of the deep residual autoencoder network. The residual convolutional layer inside the encoder module performs high-dimensional feature compression and nonlinear dimensionality reduction on the multi-source operation data stream to obtain the bottleneck feature matrix of the hidden layer. The bottleneck feature matrix of the hidden layer is passed to the decoder module of the deep residual autoencoder network. The decoder module is used to upsample and reconstruct the bottleneck feature matrix of the hidden layer, and output the corresponding multi-source reconstructed data stream. Based on separating the reconstruction bath temperature sequence and the reconstruction strip speed sequence from the multi-source reconstruction data stream, the first-order difference value between adjacent time steps of the reconstruction bath temperature sequence is extracted as the amplitude of the temperature rise trend. The incremental consumption of pickling basic reaction rate caused by strip operation is calculated based on the reconstructed strip linear velocity sequence. The difference between the increase in temperature and the increase in the basic acid pickling reaction rate is used to generate an absolute deviation value by performing an absolute value operation. This absolute deviation value is then used as a positive correlation constraint term for the temperature gradient. Extract the reconstruction temperatures of preceding and subsequent slot segments with the same characteristic time point from the multi-source reconstructed data stream; Calculate the smoothed temperature distribution difference between the reconstruction temperature of the preceding and subsequent tank sections; The overshoot margin within the smooth temperature distribution drop value that exceeds the preset thermal capacity tolerance limit is obtained as the thermal inertia continuity constraint term for adjacent slot segments. Calculate the original mean square error between the multi-source reconstructed data stream and the multi-source running data stream, and then add the original mean square error, the temperature gradient positive correlation constraint term, and the thermal inertia continuity constraint term of adjacent slot segments to obtain the total network reconstruction loss function value. The network gradient backpropagation calculation operation is performed according to the total reconstruction loss function value of the network, and the connection weight parameters inside the deep residual autoencoder network are updated and iterated using the gradient descent instruction; The bottleneck feature matrix of the hidden layer generated by the last forward propagation of the deep residual autoencoder network after updating and iterating the connection weight parameters is saved as a low-dimensional process state representation vector.
[0022] Based on the above scheme, in practical applications, the input of multi-source operational data streams and the generation of low-dimensional process state representation vectors can be executed according to a more detailed network training process. When collecting multi-source operational data streams from the pickling process section, the time series of the tank temperature of each acid tank can come from an acid-resistant temperature probe, and the strip speed series can come from the encoder speed measurement signal. After the two types of data are timestamped in the edge acquisition cabinet, they are sent to the industrial server via industrial Ethernet. After receiving the multi-source operational data streams from the pickling process section, the encoder of the deep residual autoencoder network uses the residual convolutional layer inside the encoder module to perform high-dimensional feature compression and nonlinear dimensionality reduction operations on the multi-source operational data streams. The kernel length of the residual convolutional layer can be set to three to seven sampling points, the number of layers can be set to three to six, and the dimension of the hidden layer can be set to one-quarter to one-sixteenth of the original input dimension. After processing by the encoder, the hidden layer bottleneck feature matrix is obtained. This hidden layer bottleneck feature matrix retains the main information of the temperature changes of each acid tank, the strip speed changes, and the coupling relationship between the two. The network structure consists of the input layer, the encoder residual convolutional layer, the hidden layer bottleneck feature matrix, the decoder upsampling layer and residual connection, and the output layer connected in sequence. The encoder and decoder adopt corresponding compression and restoration structures with respect to the hidden layer bottleneck feature matrix.
[0023] Subsequently, the bottleneck feature matrix of the hidden layer is passed to the decoder module of the deep residual autoencoder network. The decoder module uses an upsampling layer and residual connection corresponding to the encoder to perform upsampling mapping and reconstruction of the bottleneck feature matrix of the hidden layer, and outputs the corresponding multi-source reconstructed data stream. The multi-source reconstructed data stream contains the reconstructed bath temperature sequence and the reconstructed strip speed sequence. The processor extracts the first-order difference value between adjacent time steps from the reconstructed bath temperature sequence as the temperature rise trend amplitude. For example, if the reconstructed bath temperature at the current time point is 78 degrees Celsius and the reconstructed bath temperature at the previous time point is 77.6 degrees Celsius, then the temperature rise trend amplitude is the amplitude corresponding to the difference between the two. Based on the reconstructed strip speed sequence, the system calculates the consumption increment of the pickling basic reaction rate caused by the strip operation in conjunction with the pickling basic reaction rate calibration table. In the embodiments of the present invention, the calibration table can be obtained offline by fitting historical data from the field, with the strip speed as the input and the consumption increment under the same normalized scale as the temperature change as the input. Next, the absolute value of the difference between the temperature rise trend and the incremental consumption of the basic pickling reaction rate is calculated to generate an absolute deviation value, which is then used as a positive correlation constraint term for the temperature gradient. Before calculating the difference, both the temperature rise trend and the incremental consumption of the basic pickling reaction rate are dimensionlessly standardized using the mean of the corresponding variables in the training samples as the center and the corresponding scale parameter as the divisor, and mapped to the same numerical scale. The absolute value is used to characterize the deviation of the relative calibration coupling relationship between the two: a positive or negative difference indicates a deviation, and both should increase the constraint loss; the sign of the difference does not determine the direction of heating or cooling in this constraint term, and the specific intervention direction is determined by the subsequent counterfactual reasoning step.
[0024] Within the same network training cycle, the industrial server also extracts the reconstructed temperatures of adjacent preceding and subsequent tank segments at the same characteristic time point from the multi-source reconstructed data stream. The difference between the reconstructed temperatures of the preceding and subsequent tank segments is smoothed to obtain a smoothed temperature distribution drop value. The smoothing process can use the average of three adjacent sampling points to reduce the impact of single-point noise. The preset thermal tolerance limit can be set according to the acid tank material, acid capacity, and heat exchanger response capability, for example, between two and five degrees Celsius. When the smoothed temperature distribution drop value is greater than the preset thermal tolerance limit, the excess portion is obtained as the overshoot margin, which serves as the thermal inertia continuity constraint term for adjacent tank segments. Subsequently, the system calculates the original mean square error between the multi-source reconstructed data stream and the multi-source running data stream, and then weights and sums the original mean square error, the temperature gradient positive correlation constraint term, and the thermal inertia continuity constraint term for adjacent tank segments to obtain the total network reconstruction loss function value. The training process involves performing backpropagation calculations of the network gradient based on the total reconstruction loss function value. Gradient descent instructions are used to update and iterate the connection weight parameters within the deep residual autoencoder network, with a learning rate ranging from 0.001 to 0.01. After completing a predetermined number of iterations, the bottleneck feature matrix of the hidden layer generated by the final forward propagation of the deep residual autoencoder network after updating the connection weight parameters is obtained and saved as a low-dimensional process state representation vector.
[0025] The total network reconstruction loss function is expressed as:
[0026] In the formula, Let be the total reconstruction loss function value of the network, which is a dimensionless quantity; This is the reconstructed error term after the original mean squared error has been dimensionlessly scaled using the statistical scale of the training samples. This is the constraint quantity of the temperature gradient positive correlation constraint term after being dimensionlessly scaled according to the statistical scale of the training samples; The constraint quantity is the dimensionless constraint quantity after the thermal inertia continuity constraint term of adjacent slot segments is scaled out by the statistical scale of the training samples. These are the weighting coefficients for the reconstructed error term; The weighting coefficients for the positive correlation constraint of the temperature gradient; These are the weighting coefficients for the thermal inertia continuity constraint between adjacent slot sections. All three weighting coefficients are non-negative dimensionless numbers, and their sum is one. They are preset during the system commissioning phase.
[0027] Regarding step S2, the step of inputting the low-dimensional process state representation vector into a differentiable causal discovery network for directed edge parameter learning to generate a dynamic causal effect matrix of the influence path of the representation variable includes: extracting equipment state feature nodes, process parameter feature nodes and incoming material attribute feature nodes from the low-dimensional process state representation vector to construct an initial node set of the network; constructing an initialized fully-connected adjacency matrix as the basic connection architecture of the differentiable causal discovery network, and performing full-connection mapping on the causal connection weights between each feature node in the initial node set of the network; calculating the matrix trace of the continuous adjacency matrix, adding the matrix exponential trace as a structural penalty term for the topology of the acyclic directed graph into the training loss function of the differentiable causal discovery network for optimization iterative calculation, and extracting the continuous adjacency matrix with updated connection weights to obtain a dynamic causal effect matrix.
[0028] In a specific embodiment, on the basis that the aforementioned low-dimensional process state representation vector has been obtained, the directed edge parameter learning of the differentiable causal discovery network is carried out in a node modeling manner. The industrial server extracts equipment state feature nodes, process parameter feature nodes and incoming material attribute feature nodes from the low-dimensional process state representation vector to construct an initial node set of the network. The equipment state feature nodes may correspond to the operating states of the heating equipment supporting each physical tank section, the process parameter feature nodes may correspond to process states such as tank liquid temperature and strip steel line speed, and the incoming material attribute feature nodes may correspond to the attributes that have been determined before entering the pickling process section, such as strip steel specification and oxide scale grade. After the initial node set of the network is formed, the system constructs an initialized fully-connected adjacency matrix as the basic connection architecture of the differentiable causal discovery network, and performs full-connection mapping on the causal connection weights between each feature node in the initial node set of the network. The initialized fully-connected adjacency matrix can be assigned with small random numbers, and the main diagonal is assigned with zero to avoid nodes pointing to themselves.
[0029] Specifically, when training the differentiable causal discovery network, the industrial server converts the fully-connected adjacency matrix into a continuous adjacency matrix, calculates the matrix trace of the continuous adjacency matrix, and adds the matrix exponential trace as a structural penalty term for the topology of the acyclic directed graph into the training loss function of the differentiable causal discovery network for optimization iterative calculation. In addition to the prediction error, the training loss function also includes a structural penalty term, so that the obtained connection structure conforms to the topology of the acyclic directed graph as much as possible. After several rounds of optimization iterative calculation, edges with weak connection weights are compressed to close to zero, and the retained non-zero weight elements reflect more credible variable influence paths. When the training is completed, the system extracts the continuous adjacency matrix with updated connection weights to obtain a dynamic causal effect matrix. This matrix will be attached with a timestamp and stored in the model cache area for the decomposition of edge partial derivative influence values and the cyclic judgment of counterfactual reasoning intervention in the same control cycle.
[0030] The training loss function of the differentiable causal discovery network is expressed as:
[0031] In the formula, The training loss function value for a differentiable causal discovery network; For connection weight prediction error; The non-negative weight coefficients of the acyclic structure penalty term are preset during the system debugging phase; It is a continuous adjacency matrix with the same number of rows and columns; This represents the Hadamard element-wise multiplication operation; Indicates matrix exponentiation; Represents the matrix trace operation; The initial number of nodes in the network is also the number of rows and columns in the continuous adjacency matrix, and both are positive integers. It represents the structural penalty term for the topological structure of an acyclic directed graph.
[0032] Regarding step S3, the marginal partial derivative influence values of each feature within the low-dimensional process state characterization vector based on dynamic causal effect matrix decomposition, and the calculation of the hierarchical quality risk score set and the corresponding causal contribution vector of each variable node, include: Analyze the directed connection paths corresponding to the non-zero weight elements within the dynamic causal effect matrix; Along the directed connection path, according to the chain rule of calculus, calculate the local partial derivative gradient values of each feature node in the low-dimensional process state characterization vector relative to the final quality evaluation feature node. The local partial derivative gradient values of the same feature node on different directed connection paths are summed to generate the edge partial derivative influence values of each feature node. Based on the preset node attribute association mapping table, the edge partial derivative influence values are respectively assigned to the equipment bottleneck partial derivative group, process control partial derivative group and incoming material characteristic partial derivative group; The internal feature values of the equipment bottleneck partial derivative group, process control partial derivative group and incoming material characteristic partial derivative group are extracted, normalized and summed respectively, and the equipment bottleneck risk score, process adjustable risk score and incoming material induced risk score are calculated and merged to generate a hierarchical quality risk score set. Perform absolute value transformation on all marginal partial derivative influence values to generate a global absolute value sequence, and calculate the weight of each absolute value element in the global absolute value sequence relative to the sum of the elements in the sequence; The weight values are combined and arranged into a one-dimensional array format causal contribution vector according to the original physical node arrangement sequence in the initial node set of the network.
[0033] In one specific embodiment, based on the already generated dynamic causal effect matrix, the risk calculation thread further analyzes the directed connection paths corresponding to the non-zero weight elements within the dynamic causal effect matrix. For each directed connection path, the system calculates the local partial derivative gradient values of each feature node in the low-dimensional process state characterization vector relative to the endpoint quality assessment feature node, according to the chain rule of calculus. The endpoint quality assessment feature nodes can be trained using historical labeled data of online surface quality inspection results, pickling under-washing levels, and over-washing levels, and serve as the terminal nodes for risk calculation during runtime. If a feature node is simultaneously located on multiple directed connection paths, the system sums the local partial derivative gradient values of the same feature node on different directed connection paths to generate the marginal partial derivative influence values corresponding to each feature node.
[0034] Subsequently, the industrial server reads the preset node attribute association mapping table. This preset node attribute association mapping table was established during the system debugging phase, and it records the correspondence between each feature node and the equipment bottleneck partial derivative group, process control partial derivative group, and incoming material characteristic partial derivative group. The edge partial derivative influence values are respectively assigned to the equipment bottleneck partial derivative group, process control partial derivative group, and incoming material characteristic partial derivative group. To avoid scoring bias caused by inconsistent dimensions of different feature nodes, the system extracts the feature values within the equipment bottleneck partial derivative group, process control partial derivative group, and incoming material characteristic partial derivative group, normalizes them, and sums them to calculate the equipment bottleneck risk score, process adjustability risk score, and incoming material induced risk score. The three scores are merged to generate a hierarchical quality risk score set. In parallel, the system performs absolute value transformation on all edge partial derivative influence values to generate a global absolute value sequence, calculates the proportion of each absolute value element in the global absolute value sequence relative to the sum of the elements in the sequence, and finally combines and arranges the proportion values into a one-dimensional array format causal contribution vector according to the original physical node arrangement sequence in the initial node set of the network. The sum of all elements in this causal contribution vector is one, which facilitates direct comparison of the importance of different nodes in the future.
[0035] Regarding step S4, the initial control preference vector for generating the quality preference weight parameter and energy consumption preference weight parameter based on the numerical distribution of each node in the causal contribution vector includes: Extract the sum of the absolute values of the values corresponding to the tank temperature nodes in the causal contribution vector as the temperature contribution cumulative scalar. Determine whether the cumulative temperature contribution scalar falls within the preset temperature sensitivity range; When the cumulative scalar of temperature contribution falls within the preset temperature sensitive range, the preset benchmark quality weight is increased by the first preset step as the quality preference weight parameter, and the preset benchmark energy consumption weight is decreased by the second preset step as the energy consumption preference weight parameter. The quality preference weight parameter and the energy consumption preference weight parameter are combined into a two-dimensional parameter pair to form the initial control preference vector.
[0036] In one specific embodiment, after the causal contribution vector has been formed, the generation process of the initial control preference vector can be further refined. The system retrieves the position corresponding to the tank temperature node from the causal contribution vector and extracts the sum of the absolute values of the values corresponding to the tank temperature node in the causal contribution vector as the temperature contribution accumulation scalar. If each of the five acid tanks corresponds to a tank temperature node, the system will add the corresponding values of the five tank temperature nodes to obtain a single temperature contribution accumulation scalar. The preset temperature sensitivity range can be set to 0.3 to 0.6, or it can be adjusted according to historical quality fluctuations. When it is determined that the temperature contribution accumulation scalar falls within the preset temperature sensitivity range, the system increases the preset benchmark quality weight by a first preset step size as the quality preference weight parameter, and decreases the preset benchmark energy consumption weight by a second preset step size as the energy consumption preference weight parameter. For example, if the preset baseline quality weight is 0.5 and the first preset step size is 0.1, then the quality preference weight parameter is 0.6; if the preset baseline energy consumption weight is 0.5 and the second preset step size is 0.1, then the energy consumption preference weight parameter is 0.4. Finally, the quality preference weight parameter and the energy consumption preference weight parameter form a two-dimensional parameter pair and are merged into the initial control preference vector. This process makes the control preference no longer fixed, but adjusted according to the numerical distribution of each node within the causal contribution vector.
[0037] When the cumulative scalar of temperature contribution does not fall into the preset temperature sensitive range, the system first determines whether it is below or above the range: when it is below the lower limit of the range, the quality preference weight parameter and the energy consumption preference weight parameter maintain the preset baseline weight values respectively; when it is above the upper limit of the range, the upward and downward adjustment results when it falls into the range are used, and the two weights are limited to the range of zero to one to form saturation processing and avoid further amplification of weight changes.
[0038] Regarding step S5, determining whether the maximum score value within the stratified quality risk scoring set is greater than the preset risk warning limit includes: Determine whether the maximum score is less than or equal to the preset risk warning limit; When the maximum score is less than or equal to the preset risk warning limit, the steady-state tracking timer at the system's bottom layer is started and the steady-state tolerance timing parameter is automatically accumulated; Determine whether the steady-state tolerance timing parameters have reached the preset hysteresis exit limit; When the steady-state tolerance timing parameter reaches the preset hysteresis exit limit, the historical control preference vector stored in the historical buffer pool is extracted, and the values in the historical control preference vector are smoothly reduced and covered to the value positions corresponding to the initial control preference vector according to the preset decay slope.
[0039] In one specific embodiment, after the hierarchical quality risk score set is calculated, the early warning judgment module also includes a smooth exit process for steady-state scenarios. The system first determines whether the maximum score value is less than or equal to a preset risk warning limit. When the maximum score value is less than or equal to the preset risk warning limit, it indicates that there is no need to enter the counterfactual reasoning intervention loop judgment at this time. At this point, the system's underlying steady-state tracking timer is activated, and the steady-state tolerance timing parameter is automatically accumulated. The steady-state tolerance timing parameter can be accumulated according to the control cycle; for example, if the control cycle is ten seconds, it accumulates ten seconds after each control cycle.
[0040] The system then determines whether the steady-state tolerance timing parameter has reached the preset hysteresis exit limit. This preset hysteresis exit limit can be set to three to ten minutes to avoid frequent control state switching when the risk score drops briefly. When the steady-state tolerance timing parameter reaches the preset hysteresis exit limit, the industrial server retrieves the historical control preference vector stored in the historical buffer pool. Following a preset decay slope, it smoothly reduces the values in the historical control preference vector to the corresponding values in the initial control preference vector. The historical buffer pool can store historical control preference vectors from the most recent control cycles. This smooth reduction gradually brings the initial control preference vector back to the normal control level, reducing unnecessary jumps in the target temperature setpoint sequence received by the underlying controller.
[0041] Regarding step S6, in the counterfactual reasoning-based intervention loop determination, generating multiple sets of local slot temperature bias commands includes: The pickling process section is divided into a front pickling tank group and a rear pickling tank group according to the flow direction of the pickling medium. Read the current total heat load budget limit for each pickling tank; Under the condition that the current total heat load budget limit is not exceeded, the following heat load redistribution strategy sequences are constructed respectively: the first heat load redistribution strategy sequence of heating the front pickling tank group in combination with cooling the rear pickling tank group, the second heat load redistribution strategy sequence of cooling the front pickling tank group in combination with heating the rear pickling tank group, and the third heat load redistribution strategy sequence of overall proportional slight heating. The first heat load redistribution strategy sequence, the second heat load redistribution strategy sequence, and the third heat load redistribution strategy sequence are respectively parsed and converted into a set of specific temperature perturbations that can be recognized by the main control system. The set of specific temperature perturbations is used as multiple sets of local tank segment temperature bias commands.
[0042] In a specific embodiment, after the maximum score value is greater than the preset risk warning limit and enters the counterfactual reasoning intervention loop determination, the generation of multiple sets of local tank section temperature offset instructions can be organized according to the flow direction of the pickling medium. The system divides the pickling process section into a front-section pickling tank group and a rear-section pickling tank group according to the flow direction of the pickling medium. Taking five pickling tanks as an example, the first two pickling tanks along the flow direction of the pickling medium can be classified as the front-section pickling tank group, and the last three pickling tanks can be classified as the rear-section pickling tank group. The division boundary can also be adjusted according to the on-site pickling medium circulation pipeline and process cards.
[0043] Subsequently, the industrial server reads the current total heat load budget upper limit value of each pickling tank. The current total heat load budget upper limit value is derived from the comprehensive conversion result of the underlying controller on the available thermal power of the heating equipment supporting each physical tank section, the steam valve opening margin, and the plant utility energy supply restriction, and the unit can be uniformly kilowatt.
[0044] Specifically, under the condition that the current total heat load budget upper limit value is not exceeded, the system respectively constructs a first heat load reallocation strategy sequence in which the front-section pickling tank group is heated up in coordination with the rear-section pickling tank group being cooled down, a second heat load reallocation strategy sequence in which the front-section pickling tank group is cooled down in coordination with the rear-section pickling tank group being heated up, and a third heat load reallocation strategy sequence for overall proportional slight temperature increase. The first heat load reallocation strategy sequence is suitable for virtual deduction when the front-section reaction is insufficient, the second heat load reallocation strategy sequence is suitable for virtual deduction when the rear-section compensation is insufficient, and the third heat load reallocation strategy sequence is used to evaluate the risk change caused by slight overall heat increase. After the construction is completed, the system respectively parses and converts the first heat load reallocation strategy sequence, the second heat load reallocation strategy sequence and the third heat load reallocation strategy sequence into a specific temperature perturbation set recognizable by the main control system. The specific temperature perturbation set records the temperature offset direction and temperature offset amplitude of each pickling tank, for example: plus one degree Celsius, minus one degree Celsius, and plus zero point five degrees Celsius. Finally, the specific temperature perturbation set is input into the forward deduction calculation link as multiple sets of local tank section temperature offset instructions.
[0045] Regarding step S7, the correcting the initial control preference vector to the target intervention preference vector according to the direction of the local tank section temperature offset instruction that generates the minimum risk value includes: parsing the local tank section temperature offset instruction corresponding to the generated minimum risk value; calculating the theoretical change amount of the front-section heat load and the theoretical change amount of the rear-section heat load corresponding to the local tank section temperature offset instruction; calculating the arithmetic division ratio between the theoretical change amount of the front-section heat load and the theoretical change amount of the rear-section heat load to obtain a heat offset inclination coefficient; using the heat offset inclination coefficient to perform a direct proportional mapping amplification operation on the quality preference weight parameter in the initial control preference vector to obtain an amplified quality preference value; The energy consumption preference weight parameters in the initial control preference vector are inversely proportionally mapped and reduced to obtain the reduced energy consumption preference values. The amplified quality preference value and the reduced energy consumption preference value are concatenated to generate the target intervention preference vector.
[0046] In one specific embodiment, after the counterfactual reasoning future risk score set is generated, the formation of the target intervention preference vector needs to strictly adhere to the local tank segment temperature bias command that generates the minimum risk value. The system parses the local tank segment temperature bias command corresponding to the minimum risk value, reading the temperature bias amplitudes of the front and rear pickling tank groups. Combining the acid volume, specific heat parameters, and temperature changes of each pickling tank, the system calculates the theoretical changes in front and rear heat loads corresponding to the local tank segment temperature bias command. The theoretical changes in front and rear heat loads use the same thermal power scale to ensure the comparability of subsequent arithmetic division ratios. The system calculates the heat bias tilt coefficient by dividing the theoretical changes in front and rear heat loads by arithmetic. When the theoretical change in rear heat load is close to zero, the system can use a minimum protection value in the division to prevent calculation anomalies, and this minimum protection value is pre-written into the control program parameter table.
[0047] Specifically, after obtaining the heat bias tilt coefficient, the system uses this coefficient to perform a proportional mapping amplification operation on the quality preference weight parameters within the initial control preference vector, obtaining the amplified quality preference value. Simultaneously, it performs an inverse proportional mapping amplification operation on the energy consumption preference weight parameters within the initial control preference vector, obtaining the reduced energy consumption preference value. To prevent preference parameters from exceeding limits, both the amplified and reduced energy consumption preference values can be restricted to a normalized range of zero to one.
[0048] Subsequently, the system concatenates the amplified quality preference value with the reduced energy consumption preference value to generate a target intervention preference vector. This target intervention preference vector inherits the results of the counterfactual reasoning intervention loop determination, and therefore is closer to the intervention direction with the lowest current risk than the initial control preference vector.
[0049] Regarding step S8, the adaptive temperature control and early warning method for the pickling process section, its characteristic is that iteratively solving the problem by substituting the target intervention preference vector into a multi-objective optimization control model containing a quality loss objective function and an energy consumption evaluation objective function includes: Extract the current maximum output thermal power extreme value and minimum guaranteed thermal power extreme value of the heating equipment for each physical tank segment of the pickling process section stored in the controller, and determine the upper limit constraint boundary and lower limit constraint boundary of temperature regulation for each physical tank segment; By using the orthogonal combination of the upper limit constraint boundary and the lower limit constraint boundary of temperature regulation for each physical tank segment, a global multidimensional optimization search space covering the entire pickling process segment is constructed. Within the global multidimensional optimization search space, multiple groups of candidate temperature setpoint sequences are generated according to a preset uniform distribution probability. Extract the candidate temperature values corresponding to two physical slot segments that are spatially adjacent within each group of candidate temperature setpoint sequences, and calculate the absolute temperature difference span between adjacent slot segments. Compare the relationship between the absolute temperature difference span between adjacent tank sections and the preset allowable temperature difference limit for equipment thermal stress. Candidate temperature setpoint sequence groups whose absolute temperature difference span between adjacent tank sections is greater than the preset allowable temperature difference limit of equipment thermal stress are directly deleted and destroyed from the global multidimensional optimization search space, while candidate temperature setpoint sequence groups that pass the comparison are retained as valid setpoint sequence groups. In one specific embodiment, after the target intervention preference vector has been determined, the iterative solution of the multi-objective optimization control model still needs to establish a safe search boundary. The industrial server extracts the current maximum output heat power extreme value and minimum guaranteed heat power extreme value of the heating equipment for each physical tank segment in the pickling process stored internally in the controller. Based on the calibration relationship between heat power and tank liquid temperature regulation capability, it determines the upper limit constraint boundary and lower limit constraint boundary of temperature regulation for each physical tank segment. For example, if the allowable tank liquid temperature range for a certain physical tank segment under the current operating conditions is 70 degrees Celsius to 85 degrees Celsius, then 70 degrees Celsius is used as the lower limit constraint boundary of temperature regulation, and 85 degrees Celsius is used as the upper limit constraint boundary of temperature regulation.
[0050] Subsequently, the system constructs a global multidimensional optimization search space covering the entire pickling process segment by orthogonally combining the upper and lower temperature regulation constraints of each physical tank segment. If the pickling process segment includes five physical tank segments, the global multidimensional optimization search space contains five temperature dimensions, each with corresponding upper and lower boundaries. The orthogonal combination referred to in this application means treating the candidate temperature of each physical tank segment as an independent coordinate dimension, and using the lower to upper temperature regulation constraint boundary of that tank segment as the value range of that dimension; the Cartesian product of the value ranges of all dimensions constitutes the global multidimensional optimization search space, rather than implying that the temperature changes of each tank segment are not coupled with each other in the actual pickling process.
[0051] Specifically, after establishing the global multidimensional optimization search space, the system generates multiple candidate temperature setpoint sequence groups within the global multidimensional optimization search space according to a preset uniform distribution probability. The number of candidates can be set to one hundred to one thousand groups depending on the computing power of the industrial server. Each candidate temperature setpoint sequence group contains candidate temperature values for all physical tank segments. To avoid excessive temperature differences between adjacent tank segments, the system extracts the candidate temperature values corresponding to two spatially adjacent physical tank segments within each candidate temperature setpoint sequence group and calculates the absolute temperature difference span between adjacent tank segments. The preset allowable temperature difference limit for equipment thermal stress can be set according to the tank structure, lining material, and heat exchange equipment design requirements, for example, three to eight degrees Celsius. Subsequently, the system compares the absolute temperature difference span between adjacent tank segments with the preset allowable temperature difference limit for equipment thermal stress. Candidate temperature setpoint sequence groups with an absolute temperature difference span greater than the preset allowable temperature difference limit for equipment thermal stress are directly deleted and destroyed from the global multidimensional optimization search space, while the candidate temperature setpoint sequence groups that pass the comparison are retained as valid setpoint sequence groups. After this screening, the subsequent mass loss objective function and energy consumption evaluation objective function will only evaluate the acceptable temperature setting scheme for the equipment.
[0052] Regarding step S8, the adaptive temperature control and early warning method for the pickling process section, its characteristic is that extracting the minimum preference space distance solution within the frontier solution set of the multi-objective optimization control model as the target temperature control setpoint sequence and outputting it to the underlying controller for temperature control includes: The effective setpoint sequence groups are substituted into the mass loss objective function and energy consumption evaluation objective function inside the multi-objective optimization control model in parallel, and the corresponding mass loss adaptive measurement value and energy consumption adaptive measurement value are output for each effective setpoint sequence group. Based on the adaptive metric values of mass loss and energy consumption, all effective setpoint sequence groups are subjected to non-dominated hierarchical comparison and sorting. The inferior dominated groups with two adaptive metric values greater than the reference objects at the same level are eliminated. The top-level effective setpoint sequence group combination with absolute superior and inferior dominance relationship is obtained as the frontier solution set of the multi-objective optimization control model. The mass loss adaptation metric and energy consumption adaptation metric corresponding to each top-level effective setpoint sequence group in the frontier solution set of the multi-objective optimization control model are analyzed to generate a two-dimensional planar mapped coordinate set. Read the quality preference value and energy consumption preference value contained in the target intervention preference vector, and use the quality preference value and energy consumption preference value to calibrate the ideal preference reference origin in the same coordinate system of the two-dimensional planar mapping coordinate set; Calculate the linear Euclidean geometric distance from each coordinate point in the two-dimensional mapped coordinate set to the ideal preference reference origin; Select the top-level valid setpoint sequence group uniquely bound to the coordinate point with the smallest linear Euclidean geometric distance value, convert it into a target temperature control setpoint sequence, and output it to the underlying controller for temperature control execution.
[0053] In a specific embodiment, after the valid setpoint sequence group is formed, the system enters the final Pareto front solution screening and output stage. The industrial server respectively and parallelly substitutes the valid setpoint sequence groups into the quality loss objective function and the energy consumption evaluation objective function inside the multi-objective optimization control model, and outputs the quantified quality loss fitness value and the quantified energy consumption fitness value corresponding to each group of valid setpoint sequence groups. The quality loss objective function can be established based on the historical calibration relationship of the end-point quality assessment feature nodes, the under-pickling risk and the over-pickling risk. The energy consumption evaluation objective function is established based on the estimated thermal power consumption of the supporting heating equipment in each physical tank section, and is weighted and normalized according to the difference between the maximum output thermal power and the minimum guaranteed thermal power of each tank section, so as to retain the difference in the adjustable thermal power range of different tank sections. The outputs of the two types of functions are both normalized to the same evaluation interval, so as to facilitate subsequent non-dominated level comparison and sorting. The system performs non-dominated level comparison and sorting on all valid setpoint sequence groups based on the quantified quality loss fitness value and the quantified energy consumption fitness value, eliminates the inferior dominated groups whose two quantified fitness values are larger than those of the reference objects at the same level, and obtains the top-level valid setpoint sequence groups with absolute advantage-dominance relationship as the Pareto front solution set of the multi-objective optimization control model.
[0054] For any candidate temperature setpoint sequence, the quality loss objective function is expressed as:
[0055] In the formula, is the value of the quality loss objective function corresponding to the candidate temperature setpoint sequence, which is a dimensionless quantity; is the candidate temperature setpoint sequence, and the unit of each temperature setpoint in the sequence is degree Celsius; is the under-pickling risk score corresponding to this candidate temperature setpoint sequence, which is a dimensionless quantity; is the over-pickling risk score corresponding to this candidate temperature setpoint sequence, which is a dimensionless quantity; is the weight coefficient of the under-pickling risk score; is the weight coefficient of the over-pickling risk score. Both weight coefficients are non-negative dimensionless numbers, the sum of the two is 1, and they are preset according to the historical calibration relationship of the end-point quality assessment feature nodes.
[0056] The energy consumption evaluation objective function is expressed as:
[0057] In the formula, The objective function value for energy consumption evaluation corresponding to the candidate temperature setpoint sequence is a dimensionless quantity; The total number of physical slots participating in the optimization process, and is a positive integer; This is the index of the physical slot segment, ranging from one to the total number of physical slot segments participating in the optimization. The candidate temperature setting value for the i-th physical tank segment, in degrees Celsius; Let be the normalized expected thermal power of the i-th physical tank segment under the corresponding candidate temperature setpoint. It is a dimensionless quantity and its value ranges from zero to one. The superscript * indicates that it has been normalized according to the minimum guaranteed thermal power and the maximum output thermal power of the tank segment, and does not represent the optimal solution. This represents the summation over all physical slots participating in the optimization process. The difference between the maximum output heat power and the minimum guaranteed heat power of each slot is used for weighting, and the sum of these differences is used to complete the overall normalization.
[0058] The normalized expected thermal power of each tank section is calculated using the following formula:
[0059] In the formula, The expected absolute thermal power of the i-th physical tank segment at the candidate temperature setpoint; and These are the minimum guaranteed thermal power and maximum output thermal power of the current tank segment stored by the controller, respectively. All three are expressed in the same power unit, such as kilowatt. The tank segment participating in this normalization calculation has a non-zero adjustable thermal power range, i.e. Greater than In the same round of candidate scheme comparison, these two boundary values of each slot segment remain fixed.
[0060] Substituting the above normalization relationship into the energy consumption evaluation objective function, we can obtain its equivalent form:
[0061] Therefore, in the same round of comparisons, the denominator is a fixed positive value, and the total sum of the minimum guaranteed heat power deducted is also a fixed value, thus reducing... This ranking follows the same direction as reducing the sum of the expected absolute heat power of each participating optimization tank segment. Only when the difference between the maximum output heat power and the minimum guaranteed heat power of each tank segment is equal does this weighted expression degenerate into an equal-weighted arithmetic mean. This index is used for comparing expected heat power within the same control cycle; under the condition of consistent evaluation duration and calculation of corresponding heat energy consumption based on this expected power, it can be used to compare expected heat energy consumption, but does not directly represent the measured energy saving rate or economic cost.
[0062] After the front-end solution set of the multi-objective optimization control model is determined, the system parses the quantitative fitness value of quality loss and the quantitative fitness value of energy consumption corresponding to each top-level valid setpoint sequence group in the front-end solution set of the multi-objective optimization control model, and generates a planar two-dimensional mapping coordinate set. The horizontal coordinates in the planar two-dimensional mapping coordinate set may correspond to the quantitative fitness value of quality loss, the vertical coordinates may correspond to the quantitative fitness value of energy consumption, and all coordinate points are uniquely bound to one top-level valid setpoint sequence group.
[0063] Subsequently, the system reads the quality preference value and the energy consumption preference value contained in the target intervention preference vector, and calibrates the ideal preference reference origin in the same coordinate system of the planar two-dimensional mapping coordinate set by using the quality preference value and the energy consumption preference value. The operation thread calculates the straight-line Euclidean geometric distance from each coordinate point in the planar two-dimensional mapping coordinate set to the ideal preference reference origin. A smaller value of the straight-line Euclidean geometric distance indicates that the top-level valid setpoint sequence group corresponding to the coordinate point is closer to the current control intention. Finally, the system selects the top-level valid setpoint sequence group uniquely bound to the coordinate point with the smallest straight-line Euclidean geometric distance value, converts it into a target temperature control setpoint sequence, and outputs the sequence to the underlying controller for temperature control execution. The underlying controller adjusts the output of the heating equipment supporting each physical tank section according to the target temperature control setpoint sequence, so that the pickling process section maintains the energy consumption evaluation result at an acceptable level while the quality risk is controlled.
[0064] The foregoing shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above-mentioned embodiments. The embodiments and descriptions described above only illustrate the principle of the present invention, and various changes and improvements can be made to the present invention without departing from the spirit and scope of the present invention, and these changes and improvements all fall within the scope of the present invention as claimed. The claimed scope of the present invention is defined by the appended claims and their equivalents.
Claims
1. An acid washing process section adaptive temperature control and early warning method, characterized in that, include: Step S1: Collect multi-source operation data streams of the pickling process section, input the multi-source operation data streams into the deep residual auto-encoding network, and add temperature gradient positive correlation constraint terms and adjacent tank thermal inertia continuity constraint terms in the encoding and reconstruction stage, and output a low-dimensional process state characterization vector. Step S2: Input the low-dimensional process state characterization vector into the differentiable causal discovery network for directed edge parameter learning, and generate a dynamic causal effect matrix of the influence path of the characterization variable. Step S3: Based on the dynamic causal effect matrix decomposition, the marginal partial derivative influence values of each feature in the low-dimensional process state characterization vector are decomposed to calculate the hierarchical quality risk score set and the causal contribution vector of each corresponding variable node. Step S4: Based on the numerical distribution of each node in the causal contribution vector, generate the initial control preference vector of quality preference weight parameter and energy consumption preference weight parameter; Step S5: Determine whether the maximum score value in the stratified quality risk score set is greater than the preset risk warning limit. When the maximum score value is greater than the preset risk warning limit, enter the counterfactual reasoning intervention loop judgment. Step S6: In the counterfactual reasoning intervention loop judgment, multiple sets of local tank segment temperature offset commands are generated. The local tank segment temperature offset commands of each set are fused with the low-dimensional process state characterization vector and substituted into the dynamic causal effect matrix for forward inference calculation, and the counterfactual reasoning future risk score set is output. Step S7: Extract the minimum risk value from the counterfactual reasoning future risk score set, and correct the initial control preference vector to the target intervention preference vector according to the local tank segment temperature bias command direction that generates the minimum risk value. Step S8: Substitute the target intervention preference vector into the multi-objective optimization control model containing the mass loss objective function and the energy consumption evaluation objective function for iterative solution. Extract the minimum preference space distance solution in the frontier solution set of the multi-objective optimization control model as the target temperature control set value sequence and output it to the bottom controller for execution.
2. The method according to claim 1, wherein the method is characterized by, The multi-source operational data stream is input into a deep residual autoencoder network, and temperature gradient positive correlation constraints and thermal inertia continuity constraints between adjacent tank segments are added in the encoding and reconstruction stage. The output is a low-dimensional process state characterization vector, which specifically includes: Collect multi-source operational data streams for the pickling process section; the multi-source operational data streams include: the time series of the bath temperature and the strip speed series of each pickling tank; The multi-source operation data stream of the pickling process section is input into the encoder of the deep residual autoencoder network. The residual convolutional layer inside the encoder module performs high-dimensional feature compression and nonlinear dimensionality reduction on the multi-source operation data stream to obtain the bottleneck feature matrix of the hidden layer. The bottleneck feature matrix of the hidden layer is passed to the decoder module of the deep residual autoencoder network. The decoder module is used to upsample and reconstruct the bottleneck feature matrix of the hidden layer, and output the corresponding multi-source reconstructed data stream. Based on separating the reconstruction bath temperature sequence and the reconstruction strip speed sequence from the multi-source reconstruction data stream, the first-order difference value between adjacent time steps of the reconstruction bath temperature sequence is extracted as the amplitude of the temperature rise trend. The incremental consumption of pickling basic reaction rate caused by strip operation is calculated based on the reconstructed strip linear velocity sequence. The difference between the increase in temperature and the increase in the basic acid pickling reaction rate is used to generate an absolute deviation value by performing an absolute value operation. This absolute deviation value is then used as a positive correlation constraint term for the temperature gradient. Extract the reconstruction temperatures of preceding and subsequent slot segments with the same characteristic time point from the multi-source reconstructed data stream; Calculate the smoothed temperature distribution difference between the reconstruction temperature of the preceding and subsequent tank sections; The overshoot margin within the smooth temperature distribution drop value that exceeds the preset thermal capacity tolerance limit is obtained as the thermal inertia continuity constraint term for adjacent slot segments. Calculate the original mean square error between the multi-source reconstructed data stream and the multi-source running data stream, and then add the original mean square error, the temperature gradient positive correlation constraint term, and the thermal inertia continuity constraint term of adjacent slot segments to obtain the total network reconstruction loss function value. The network gradient backpropagation calculation operation is performed according to the total reconstruction loss function value of the network, and the connection weight parameters inside the deep residual autoencoder network are updated and iterated using the gradient descent instruction; The bottleneck feature matrix of the hidden layer generated by the last forward propagation of the deep residual autoencoder network after updating and iterating the connection weight parameters is saved as a low-dimensional process state representation vector.
3. The adaptive temperature control and early warning method for a pickling process section according to claim 2, characterized in that, The low-dimensional process state representation vector is input into a differentiable causal discovery network for directed edge parameter learning, generating a dynamic causal effect matrix of the influence path of the representation variable, specifically including: Extract equipment state feature nodes, process parameter feature nodes, and incoming material attribute feature nodes from the low-dimensional process state representation vector to construct the initial node set of the network; An initial fully connected adjacency matrix is constructed as the basic connection architecture of the differentiable causal discovery network, and the causal connection weights between each feature node in the initial node set of the network are fully mapped. The trace of the continuous adjacency matrix is calculated, and the trace of the matrix exponent is added as a structural penalty term for the acyclic directed graph topology to the training loss function of the differentiable causal discovery network for optimization and iterative calculation. The continuous adjacency matrix with updated connection weights is extracted to obtain the dynamic causal effect matrix.
4. The adaptive temperature control and early warning method for a pickling process section according to claim 3, characterized in that, Based on the marginal partial derivative influence values of each feature within the low-dimensional process state characterization vector through dynamic causal effect matrix decomposition, the hierarchical quality risk score set and the corresponding causal contribution vector of each variable node are calculated, including: Analyze the directed connection paths corresponding to the non-zero weight elements within the dynamic causal effect matrix; Along the directed connection path, according to the chain rule of calculus, calculate the local partial derivative gradient values of each feature node in the low-dimensional process state characterization vector relative to the final quality evaluation feature node. The local partial derivative gradient values of the same feature node on different directed connection paths are summed to generate the edge partial derivative influence values of each feature node. Based on the preset node attribute association mapping table, the edge partial derivative influence values are respectively assigned to the equipment bottleneck partial derivative group, process control partial derivative group and incoming material characteristic partial derivative group; The internal feature values of the equipment bottleneck partial derivative group, process control partial derivative group and incoming material characteristic partial derivative group are extracted, normalized and summed respectively, and the equipment bottleneck risk score, process adjustable risk score and incoming material induced risk score are calculated and merged to generate a hierarchical quality risk score set. Perform absolute value transformation on all marginal partial derivative influence values to generate a global absolute value sequence, and calculate the weight of each absolute value element in the global absolute value sequence relative to the sum of the elements in the sequence; The weight values are combined and arranged into a one-dimensional array format causal contribution vector according to the original physical node arrangement sequence in the initial node set of the network.
5. The adaptive temperature control and early warning method for a pickling process section according to claim 4, characterized in that, Based on the numerical distribution of each node within the causal contribution vector, an initial control preference vector is generated for the quality preference weight parameter and the energy consumption preference weight parameter, including: Extract the sum of the absolute values of the values corresponding to the tank temperature nodes in the causal contribution vector as the temperature contribution cumulative scalar. Determine whether the cumulative temperature contribution scalar falls within the preset temperature sensitivity range; When the cumulative scalar of temperature contribution falls within the preset temperature sensitive range, the preset benchmark quality weight is increased by the first preset step as the quality preference weight parameter, and the preset benchmark energy consumption weight is decreased by the second preset step as the energy consumption preference weight parameter. The quality preference weight parameter and the energy consumption preference weight parameter are combined into a two-dimensional parameter pair to form the initial control preference vector.
6. The adaptive temperature control and early warning method for a pickling process section according to claim 5, characterized in that, Determining whether the maximum score value within the stratified quality risk score set exceeds the preset risk warning limit includes: Determine whether the maximum score is less than or equal to the preset risk warning limit; When the maximum score is less than or equal to the preset risk warning limit, the steady-state tracking timer at the system's bottom layer is started and the steady-state tolerance timing parameter is automatically accumulated; Determine whether the steady-state tolerance timing parameters have reached the preset hysteresis exit limit; When the steady-state tolerance timing parameter reaches the preset hysteresis exit limit, the historical control preference vector stored in the historical buffer pool is extracted, and the values in the historical control preference vector are smoothly reduced and covered to the value positions corresponding to the initial control preference vector according to the preset decay slope.
7. The adaptive temperature control and early warning method for a pickling process section according to claim 6, characterized in that, In the counterfactual reasoning intervention loop decision, multiple sets of local tank segment temperature bias commands are generated, including: The pickling process section is divided into a front pickling tank group and a rear pickling tank group according to the flow direction of the pickling medium. Read the current total heat load budget limit for each pickling tank; Under the condition that the current total heat load budget limit is not exceeded, the following heat load redistribution strategy sequences are constructed respectively: the first heat load redistribution strategy sequence of heating the front pickling tank group in combination with cooling the rear pickling tank group, the second heat load redistribution strategy sequence of cooling the front pickling tank group in combination with heating the rear pickling tank group, and the third heat load redistribution strategy sequence of overall proportional slight heating. The first heat load redistribution strategy sequence, the second heat load redistribution strategy sequence, and the third heat load redistribution strategy sequence are respectively parsed and converted into a set of specific temperature perturbations that can be recognized by the main control system. The set of specific temperature perturbations is used as multiple sets of local tank segment temperature bias commands.
8. The adaptive temperature control and early warning method for a pickling process section according to claim 7, characterized in that, The initial control preference vector is corrected to the target intervention preference vector according to the local tank segment temperature bias command direction that produces the minimum risk value, including: Analyze the local tank segment temperature offset command corresponding to the minimum risk value; Calculate the theoretical changes in front zone heat load and rear zone heat load corresponding to the local tank temperature offset command; The heat offset tilt coefficient is obtained by calculating the arithmetic division ratio between the theoretical change in heat load in the front zone and the theoretical change in heat load in the back zone. The quality preference weight parameters in the initial control preference vector are proportionally mapped and amplified using the heat bias tilt coefficient to obtain the amplified quality preference values. The energy consumption preference weight parameters in the initial control preference vector are inversely proportionally mapped and reduced to obtain the reduced energy consumption preference values. The amplified quality preference value and the reduced energy consumption preference value are concatenated to generate the target intervention preference vector.
9. The adaptive temperature control and early warning method for a pickling process section according to claim 8, characterized in that, The target intervention preference vector is substituted into a multi-objective optimization control model containing both a quality loss objective function and an energy consumption evaluation objective function for iterative solution, including: Extract the current maximum output thermal power extreme value and minimum guaranteed thermal power extreme value of the heating equipment for each physical tank segment of the pickling process section stored in the controller, and determine the upper limit constraint boundary and lower limit constraint boundary of temperature regulation for each physical tank segment; By using the orthogonal combination of the upper limit constraint boundary and the lower limit constraint boundary of temperature regulation for each physical tank segment, a global multidimensional optimization search space covering the entire pickling process segment is constructed. Within the global multidimensional optimization search space, multiple groups of candidate temperature setpoint sequences are generated according to a preset uniform distribution probability. Extract the candidate temperature values corresponding to two physical slot segments that are spatially adjacent within each group of candidate temperature setpoint sequences, and calculate the absolute temperature difference span between adjacent slot segments. Compare the relationship between the absolute temperature difference span between adjacent tank sections and the preset allowable temperature difference limit for equipment thermal stress. Candidate temperature setpoint sequence groups whose absolute temperature difference span between adjacent tank sections exceeds the preset allowable temperature difference limit for equipment thermal stress are directly deleted and destroyed from the global multidimensional optimization search space, while candidate temperature setpoint sequence groups that pass the comparison are retained as valid setpoint sequence groups.
10. The adaptive temperature control and early warning method for a pickling process section according to claim 1, characterized in that, Extracting the minimum preference space distance solution within the frontier solution set of the multi-objective optimization control model as the target temperature control setpoint sequence and outputting it to the underlying controller for temperature control execution includes: The effective setpoint sequence groups are substituted into the mass loss objective function and energy consumption evaluation objective function inside the multi-objective optimization control model in parallel, and the corresponding mass loss adaptive measurement value and energy consumption adaptive measurement value are output for each effective setpoint sequence group. Based on the adaptive metric values of mass loss and energy consumption, all effective setpoint sequence groups are subjected to non-dominated hierarchical comparison and sorting. The inferior dominated groups with two adaptive metric values greater than the reference objects at the same level are eliminated. The top-level effective setpoint sequence group combination with absolute superior and inferior dominance relationship is obtained as the frontier solution set of the multi-objective optimization control model. The mass loss adaptation metric and energy consumption adaptation metric corresponding to each top-level effective setpoint sequence group in the frontier solution set of the multi-objective optimization control model are analyzed to generate a two-dimensional planar mapped coordinate set. Read the quality preference value and energy consumption preference value contained in the target intervention preference vector, and use the quality preference value and energy consumption preference value to calibrate the ideal preference reference origin in the same coordinate system of the two-dimensional planar mapping coordinate set; Calculate the linear Euclidean geometric distance from each coordinate point in the two-dimensional mapped coordinate set to the ideal preference reference origin; The top-level valid setpoint sequence group is selected by uniquely binding the coordinate point with the smallest linear Euclidean geometric distance value, and converted into a target temperature control setpoint sequence for output to the bottom-level controller for temperature control.
