Deep learning based hot-rolled strip steel mechanical property process parameter reverse optimization method

Abstract

The application is suitable for the technical field of process parameter optimization, and provides a hot-rolled strip steel mechanical property process parameter reverse optimization method based on deep learning, comprising the following steps: collecting process parameters and corresponding strip steel mechanical property data of a hot continuous rolling process, and constructing a sample set composed of a process parameter vector and a mechanical property vector; training a deep performance prediction field, a deep credibility field and a deep feasibility potential field, and constructing a collaborative energy function; taking current benchmark process parameters and a plurality of process parameter clustering centers corresponding to historical similar target properties as starting points, performing a multi-startpoint constraint descent search along a gradient direction of the collaborative energy function, and obtaining a candidate process parameter set; and determining a final output from the candidate process parameter set. The application has the ability to perceive and avoid prediction uncertainty, and the optimization is away from process regions unfamiliar to the model, and only searches in a region with reliable prediction, thereby improving the reliability of the reverse optimization result.

Description

Technical Field

[0001] This invention relates to the field of process parameter optimization technology, specifically to a method for reverse optimization of process parameters for the mechanical properties of hot-rolled strip steel based on deep learning. Background Technology

[0002] Hot-rolled strip steel is an important industrial raw material, and its mechanical properties must meet stringent customer order requirements. During the production process, these properties are determined by a series of process parameters such as finishing rolling temperature, coiling temperature, and reduction rate. When customers propose new performance targets, the core technical challenge facing steel companies is how to quickly and accurately reverse-engineer the optimal combination of process parameters.

[0003] Existing deep learning-based reverse optimization methods suffer from the following bottlenecks: First, the forward process-performance prediction model is treated as an absolutely accurate black box, ignoring the high uncertainty of predictions in unfamiliar regions, leading to incorrect guidance to unreliable solutions during reverse optimization. Second, the feasible domain of actual production processes is highly complex and has fuzzy boundaries, making it difficult for traditional methods to define precisely, and the optimized process parameters are often infeasible in practice. Therefore, a deep learning-based reverse optimization method for the mechanical properties and process parameters of hot-rolled strip steel is needed to address these issues. Summary of the Invention

[0004] In view of the shortcomings of the existing technology, the purpose of this invention is to provide a method for reverse optimization of process parameters of mechanical properties of hot-rolled strip steel based on deep learning, so as to solve the problems existing in the background technology.

[0005] This invention is implemented as follows: a deep learning-based method for reverse optimization of process parameters for the mechanical properties of hot-rolled strip steel, comprising the following steps: The process parameters and corresponding mechanical property data of the hot strip rolling process are collected. After time alignment, steady-state screening and anomaly removal, a sample set consisting of process parameter vectors and mechanical property vectors is constructed. Using process parameter vectors as input and mechanical property vectors as output, a deep performance prediction field is trained to learn the mapping relationship between process parameters and mechanical properties. Using the sample set, a deep confidence field is trained, with process parameter vectors as input, and scalar confidence is output. Using the feasible process parameters in the sample set, a deep feasibility potential field is trained, which approaches zero in the feasible region. For a given target mechanical performance vector and current benchmark process parameter vector, a cooperative energy function is constructed. The cooperative energy function is composed of three parts: a confidence-weighted performance prediction deviation term, a feasibility potential term, and an anchoring smoothing term. Starting from the current benchmark process parameters and several cluster centers of process parameters corresponding to historically similar target performance, a multi-starting point constraint descent search is performed along the gradient direction of the cooperative energy function, and the results of each iteration are projected back into the preset process hard constraint boundary to obtain a set of candidate process parameters. Solutions with feasibility potential higher than the first threshold or confidence lower than the second threshold are removed from the candidate process parameter set. The process parameter vector with the smallest value of the anchoring smoothing term is selected from the remaining solutions as the final output of the inverse optimization.

[0006] As a further aspect of the present invention, the step of constructing a sample set specifically includes: Using the synchronous timestamps of the thickness gauge and pyrometer at the finish mill exit as a reference, the rolling parameters, temperature parameters, cooling parameters of each stand are aligned with the tensile test results; Records containing sensor malfunctions, rolling stoppages, and unsteady sections at the beginning and end of the strip are excluded; A local weighted regression smoothing method was used to detect and remove outlier sample points to obtain the process-performance sample set.

[0007] As a further aspect of the present invention, the step of training the deep performance prediction field specifically includes: Inject random disturbances that conform to the statistical characteristics of field instrument noise into the input process parameter vector to generate small disturbance samples; This ensures that the depth performance prediction field outputs consistent mechanical performance prediction values ​​for multiple perturbation samples at the same process parameter point; By minimizing the joint objective of prediction error and consistency constraint, the performance prediction field remains smooth and accurate at the sample point and its neighborhood.

[0008] As a further aspect of the present invention, the step of training the deep confidence field specifically includes: The process parameter points that are densely clustered in the sample set are marked as high-confidence samples; Randomly sample extrapolation points far from the sample clustering area in the process parameter space and mark them as low-confidence samples; Through comparative training, the confidence field is made to output a value close to 1 for high confidence samples and a value close to 0 for low confidence samples, with a smooth transition region.

[0009] As a further aspect of the present invention, the step of training a deep feasibility potential field specifically includes: The feasible process parameters in the sample set are used as positive samples; Three types of negative samples are generated: process parameter combinations that exceed equipment limits, associated combinations that violate process logic, and out-of-bounds points generated by perturbations along the edge of the feasible process manifold. Based on positive and negative samples, discriminative training with transition band smoothing constraints is introduced to make the feasible potential field output a low potential value near zero on positive samples and a high potential value on negative samples, and the potential field gradient is continuous at the boundary of the feasible region.

[0010] As a further aspect of the present invention, the cooperative energy function is: E(x) = C(x)·‖F(x)-y tgt || 2 +α·Φ(x)+β·‖xx cur || 2 S , x is the vector of process parameters to be optimized; F(x) is the vector of mechanical properties output by the depth performance prediction field; y tgt Let C(x) be the target mechanical performance vector; C(x) be the scalar confidence level output by the depth confidence field at x, serving as the pointwise weight of the performance deviation term; Φ(x) be the value of the depth feasibility potential field at x; α be the feasibility penalty coefficient; ||xx cur || 2 S The current baseline process parameter x cur The weighted squared distance to x is S, which is a diagonal weight matrix composed of the adjustment costs of each process parameter; β is the smoothing anchoring coefficient.

[0011] As a further aspect of the present invention, the step of performing a multi-starting-point constrained descent search specifically includes: Determine a set of starting points, which includes at least the current baseline process parameter vector and cluster centers of process parameter vectors corresponding to orders with similar target mechanical properties identified from historical data; For each starting point, the gradient of the cooperative energy function with respect to the process parameter vector is calculated, the process parameter vector is updated along the negative gradient direction with a fixed step size, and the components of the updated vector are projected into the hard boundary determined by the equipment capability and process specifications. The process is iterated until the convergence condition is met, resulting in multiple local energy minima, which form a set of candidate process parameters.

[0012] As a further aspect of the present invention, the method further includes: After production according to the final output process parameters, the measured mechanical properties of the strip steel are obtained, and the deviation between the measured mechanical properties and the target mechanical property vector is calculated. With a fixed depth performance prediction field and depth confidence field, starting from the currently used process parameters, local gradient descent with a minimum step size is performed on the cooperative energy function to generate correction increments and superimpose them to achieve fast offset compensation without repetition optimization. New production data is periodically incorporated into the sample set, and the deep performance prediction field, deep credibility field, and deep feasibility potential field are incrementally retrained to ensure that each deep field evolves synchronously with the production process.

[0013] Compared with the prior art, the beneficial effects of the present invention are: By introducing a deep confidence field, the optimization process is empowered to perceive and mitigate prediction uncertainties. The optimization automatically avoids unfamiliar process regions, searching only within regions where predictions are reliable, significantly improving the reliability of the back-optimization results. Constructing a deep feasibility potential field ensures that the output process parameters fall within equipment capabilities and conform to process logic, eliminating the need for post-implementation verification. Adding an anchoring smoothing term to the optimization objective prioritizes recommending the process solution closest to the current production status with the lowest adjustment cost, maximizing production line stability while meeting new order requirements. Attached Figure Description

[0014] Figure 1 This is a flowchart of a deep learning-based method for reverse optimization of process parameters for the mechanical properties of hot-rolled strip steel.

[0015] Figure 2 This is a flowchart illustrating the construction of a sample set in a deep learning-based method for reverse optimization of mechanical property process parameters of hot-rolled strip steel.

[0016] Figure 3 This is a flowchart of the training deep performance prediction field in the deep learning-based method for reverse optimization of mechanical property process parameters of hot-rolled strip steel.

[0017] Figure 4 This is a flowchart of the training depth confidence field in the deep learning-based method for reverse optimization of mechanical property process parameters of hot-rolled strip steel.

[0018] Figure 5 This is a flowchart of the training depth feasibility potential field in the deep learning-based method for reverse optimization of mechanical properties and process parameters of hot-rolled strip steel.

[0019] Figure 6 This is a flowchart of the multi-starting-point constraint descent search in the deep learning-based method for reverse optimization of mechanical properties and process parameters of hot-rolled strip steel.

[0020] Figure 7 This is a flowchart of the fast offset compensation method in the reverse optimization method of mechanical property process parameters of hot-rolled strip steel based on deep learning. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0022] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

[0023] like Figure 1 As shown in the figure, this invention provides a method for reverse optimization of process parameters for the mechanical properties of hot-rolled strip steel based on deep learning. The method includes the following steps: S100 collects the process parameters and corresponding strip mechanical property data of the hot continuous rolling process. After time alignment, steady-state screening and anomaly removal, a sample set consisting of process parameter vectors and mechanical property vectors is constructed.

[0024] In this embodiment of the invention, raw process parameters (such as temperature, pressure, and speed) and corresponding mechanical property test data (such as yield strength and tensile strength) are first collected from the hot continuous rolling production line. These data need to undergo time alignment (ensuring that the process parameters and performance data of each coil of steel correspond), steady-state screening (removing unstable data segments such as the start and end of rolling), and anomaly removal (removing erroneous data caused by instrument failure, etc.) to finally form a reliable process-performance sample library.

[0025] S200 uses process parameter vectors as input and mechanical property vectors as output to train a deep performance prediction field, enabling it to learn the mapping relationship between process parameters and mechanical properties.

[0026] In this embodiment of the invention, a deep performance prediction field needs to be trained. Its purpose is to establish an accurate forward mapping from process parameters to mechanical properties. Given a process formulation, it can predict the corresponding mechanical properties. This forms the basis for subsequent backpropagation.

[0027] S300: Using the sample set, train a deep confidence field, taking the process parameter vector as input, and output scalar confidence.

[0028] In this embodiment of the invention, a deep confidence field also needs to be trained to evaluate how reliable the prediction results are. It also takes process parameters as input, but outputs a scalar between 0 and 1. In historically data-rich and experienced process regions, it outputs high confidence (close to 1); in data-sparse or unexplored regions, it outputs low confidence (close to 0).

[0029] S400, using the feasible process parameters in the sample set, train a deep feasibility potential field that approaches zero in the feasible region.

[0030] In this embodiment of the invention, a deep feasibility potential field also needs to be trained to determine whether a process parameter is physically feasible and whether it is a normal production practice. Within the "feasible region" where production has been successfully completed historically, its value approaches zero; within the "infeasible region" where equipment capabilities are limited or where process logic conflicts, its value will be very high.

[0031] S500, for a given target mechanical performance vector and current benchmark process parameter vector, construct a cooperative energy function, which is composed of three parts: a confidence-weighted performance prediction deviation term, a feasibility potential energy term, and an anchoring smoothing term.

[0032] In this embodiment of the invention, the three independently trained fields are merged into a unified optimization objective. Given a process parameter x to be solved, the function calculates a total energy E. This energy consists of the sum of three parts, each corresponding to solving a real-world problem. The aim is to ensure high reliability and feasibility of the results, making the optimization results as close as possible to the current situation, avoiding drastic process changes, and ensuring production stability.

[0033] S600: Starting from the current benchmark process parameters and several cluster centers of process parameters corresponding to historically similar target performance, perform a multi-starting-point constraint descent search along the gradient direction of the cooperative energy function, and project the results of each iteration back into the preset process hard constraint boundary to obtain a set of candidate process parameters.

[0034] In this embodiment of the invention, given the objective function E, the solution becomes finding the x that minimizes E. To avoid getting trapped in local optima, instead of starting from a single point, multiple starting points are used simultaneously. These points include the current process and processes that historically produced similar performance. From each starting point, x is moved step by step along the gradient direction of the energy function (i.e., the direction of the fastest energy decrease), ensuring at each step that x does not exceed the physical hard boundaries of the equipment (such as the maximum rolling force). Ultimately, each starting point may converge to a local energy minimum point, and these points together constitute a set of candidate process parameters.

[0035] S700, remove solutions from the candidate process parameter set whose feasibility potential is higher than the first threshold or whose confidence is lower than the second threshold, and select the process parameter vector with the smallest value of the anchoring smoothing term from the remaining solutions as the final output of the reverse optimization.

[0036] In this embodiment of the invention, the candidate set obtained in the previous step is further optimized. First, a qualification screening is performed to eliminate solutions with excessively high feasibility potential (meaning the process is infeasible) or excessively low confidence (meaning the prediction is unreliable). Among the remaining qualified solutions, the one with the smallest anchoring smoothing term, i.e., the one that causes the least change to the current production state, is selected as the final unique solution output. This ensures that the final recommended process scheme not only meets performance standards and is reliable and feasible, but also minimizes disturbance to the production line.

[0037] like Figure 2 As shown, in a preferred embodiment of the present invention, the step of constructing a sample set specifically includes: S101, using the synchronous timestamps of the finishing mill exit thickness gauge and pyrometer as a reference, aligns the rolling parameters, temperature parameters, cooling parameters of each stand with the tensile test results; S102, records containing sensor failures, rolling stoppages, and unsteady sections at the beginning and end of the strip are excluded; S103 uses a local weighted regression smoothing method to detect and remove outlier sample points, resulting in a process-performance sample set.

[0038] In this embodiment of the invention, the temperature and pressure data during the rolling process on the hot strip mill production line are separated from the performance data obtained from the final sampling and testing. It is essential to use the time points recorded by the pyrometer and thickness gauge at the finishing mill exit as a benchmark. The process parameters of each stand recorded at the same time must be correlated with the tensile test results of the strip after cooling, using the strip length and speed to backtrack to that time point, achieving a precise one-to-one correspondence. Furthermore, it is necessary to clearly define which data is invalid, including erroneous data recorded during sensor malfunctions, meaningless data during mill downtime, and non-representative data generated due to tension and temperature instability during strip threading and throwing at the head and tail of the strip. Removing this data ensures that the sample set reflects a controllable and stable rolling state. Next, a local weighted regression smoothing method is used to identify isolated points that, although not caused by obvious faults, deviate from the overall trend of most normal data. These isolated points are then removed, resulting in a clean, high-confidence sample set.

[0039] like Figure 3 As shown, in a preferred embodiment of the present invention, the step of training a deep performance prediction field specifically includes: S201, inject random disturbances that conform to the statistical characteristics of field instrument noise into the input process parameter vector to generate a small disturbance sample; S202, which enables the deep performance prediction field to output consistent mechanical performance prediction values ​​for multiple perturbation samples at the same process parameter point; S203 minimizes the joint objective of prediction error and consistency constraint, ensuring that the performance prediction field remains smooth and accurate at the sample point and its neighborhood.

[0040] In this embodiment of the invention, the original data is not used directly during training. Instead, the small measurement errors inherent in the field sensors, conforming to certain patterns (such as a Gaussian distribution), are simulated. This noise is repeatedly added to the input process parameters, forming a set of perturbation samples around the original point. The training goal is not only to make the model accurately predict the original samples, but more importantly, to require it to provide highly consistent prediction results for all perturbation samples generated at the same process point. This consistency requirement is combined with the traditional prediction error minimization objective as an additional optimization objective. This makes the final performance prediction field not only accurate at the data points, but also very smooth and stable in its neighborhood, avoiding drastic jumps in performance prediction due to small input fluctuations, which is crucial for subsequent gradient-based optimization.

[0041] like Figure 4 As shown, in a preferred embodiment of the present invention, the step of training the deep confidence field specifically includes: S301, mark the process parameter points that are densely clustered in the sample set as high-confidence samples; S302, randomly sample extrapolation points far from the sample clustering area in the process parameter space and mark them as low confidence samples; S303, through comparative training, makes the confidence field output a value close to 1 for high confidence samples and a value close to 0 for low confidence samples, with a smooth transition region.

[0042] In this embodiment of the invention, regions with densely packed sample points in historical data are identified as high-confidence areas. These regions signify rich production experience, and the model should have high confidence in them. Therefore, density clustering is used to mark the densely clustered points in the sample set as positive samples for training the confidence field. Additionally, untrustworthy regions need to be identified. Random sampling is performed across the entire process parameter space; these regions, far from any historical production points, are considered low-confidence. Simultaneously, a perturbation can be performed outwards from the edge of the feasible region to include points whose slight deviation from the known feasible region could lead to drastic changes in performance prediction as negative samples. Using the constructed positive and negative sample pairs, the network is trained to distinguish between trustworthy and untrustworthy regions. The training objective forces the network to output 1 for positive samples and 0 for negative samples. Particular emphasis is placed on ensuring that the output value smoothly decreases from 1 to 0 in the transition region between positive and negative samples, making the confidence field differentiable everywhere and providing a smooth weight signal for subsequent gradient optimization.

[0043] like Figure 5 As shown, in a preferred embodiment of the present invention, the step of training a deep feasibility potential field specifically includes: S401, use the feasible process parameters in the sample set as positive samples; S402 generates three types of negative samples: process parameter combinations that exceed equipment limits, associated combinations that violate process logic, and out-of-bounds points generated by perturbations along the edge of the feasible process manifold. S403, based on positive and negative samples, introduces discriminative training with transition band smoothing constraints, so that the feasible potential field outputs a low potential value near the zero value on positive samples and a high potential value on negative samples, and the potential field gradient is continuous at the boundary of the feasible region.

[0044] In this embodiment of the invention, all historically feasible process parameters from successful production in the sample set are directly used as positive samples. Then, infeasible samples are identified, categorized into three types: The first type consists of points outside the physical hard limit, such as a reduction rate exceeding equipment capacity. The second type consists of points with conflicting process logic, such as extremely high final rolling temperatures and extremely low coiling temperatures, which are difficult to achieve thermodynamically and in terms of phase transformation kinetics. The third type consists of boundary violation points, generated by making small perturbations outward from the edge of the feasible region, used to finely characterize the boundary shape of the feasible region. The training objective is to make the potential field Φ a valid feasibility criterion. It is forced to output a low potential value close to zero on all positive samples and a high potential value on all negative samples. Furthermore, it achieves a smooth transition constraint, requiring that Φ's value cannot abruptly increase at the boundary of the feasible region where positive and negative samples intersect; it must grow smoothly. This makes the potential field differentiable everywhere, providing accurate navigation information for subsequent gradient descent searches.

[0045] In a preferred embodiment of the present invention, the cooperative energy function is: E(x) = C(x)·‖F(x)-y tgt || 2 +α·Φ(x)+β·‖xx cur || 2 S , x is the vector of process parameters to be optimized; F(x) is the vector of mechanical properties output by the depth performance prediction field; y tgt Let x be the target mechanical performance vector; C(x) be the scalar confidence level output by the depth confidence field at x, serving as the pointwise weight of the performance deviation term; Φ(x) be the value of the depth feasibility potential field at x; α be the feasibility penalty coefficient, the larger α is, the more conservative the optimization result, and the less likely it is to approach the infeasibility boundary; ||xx cur || 2 S For a complete smooth anchoring term, the current baseline process parameter x curThe weighted squared distance with respect to x is not a simple Euclidean distance, but rather the deviation from x in each dimension multiplied by an independent weight (stored on the diagonal of matrix S), and then the sum of squares is calculated. This means that the cost of changing certain key parameters (such as the final rolling temperature) can be set higher. S is a diagonal weight matrix composed of the adjustment costs of each process parameter; β is the smoothing anchoring coefficient. The larger β is, the more the optimization result is anchored to the current operating conditions, and the smaller the process changes.

[0046] like Figure 6 As shown, in a preferred embodiment of the present invention, the step of performing a multi-starting-point constraint descent search specifically includes: S601, determine the set of starting points, the set of starting points shall at least include the current benchmark process parameter vector and the cluster center of the process parameter vector corresponding to the order with similar mechanical properties to the target from the historical data; S602, For each starting point, calculate the gradient of the cooperative energy function with respect to the process parameter vector, update the process parameter vector along the negative gradient direction with a fixed step size, and project each component of the updated vector into the hard boundary determined by the equipment capability and process specifications. S603, iterative execution until the convergence condition is met, obtains multiple local energy minima, and forms a set of candidate process parameters.

[0047] In this embodiment of the invention, a starting point for the search is defined. Besides the current baseline process parameter x... cur Furthermore, it leverages historical databases to identify process parameters that have historically met similar performance targets, and uses clustering to find representative centers. Starting from these points with successful precedents increases the chance of finding the global optimum. Then, projective gradient descent is performed, calculating the gradient to find the direction that causes the energy E to decrease most rapidly at the current x point. Each update moves x a small step along that direction. The projection operation is a crucial safeguard, ensuring that no dimension of the updated x ever crosses a pre-defined physical hard boundary. Iteration stops when the energy no longer decreases significantly or the maximum number of steps is reached. Each starting point eventually converges to a local optimum. The convergence results from all starting points are combined to form a candidate set containing multiple competitive solutions.

[0048] like Figure 7 As shown, in a preferred embodiment of the present invention, the method further includes: S801: After production according to the final output process parameters, obtain the measured mechanical properties of the strip steel and calculate the deviation between it and the target mechanical property vector. S802, with a fixed depth performance prediction field and depth confidence field, starting from the currently used process parameters, performs local gradient descent with a very small step size on the cooperative energy function, generates correction increments and superimposes them, to achieve fast offset compensation without repetition optimization; S803 periodically incorporates new production data into the sample set and incrementally retrains the deep performance prediction field, deep credibility field, and deep feasibility potential field, so that each depth field evolves synchronously with the production process.

[0049] In this embodiment of the invention, if the measured performance still deviates slightly from the target after production according to the optimized process, there is no need to restart the entire global optimization. Instead, a lightweight compensation strategy is adopted: the trained prediction field and confidence field are fixed, and a local gradient descent with a very small step size is performed on the co-energy function, starting only from the currently used process x, to quickly calculate a correction amount and add it back. This allows the system to quickly compensate for subtle changes such as raw material batch fluctuations. In addition, newly accumulated production data is periodically added to the initial sample set, and these new data are used to incrementally retrain the deep performance prediction field, deep confidence field, and deep feasibility potential field. In this way, the model can continuously adapt to the data distribution drift caused by equipment aging, raw material changes, and the development of new steel grades, ensuring that the accuracy of back-optimization does not decay during long-term operation.

[0050] The above description only details the preferred embodiments of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0051] It should be understood that although the steps in the flowcharts of the various embodiments of the present invention are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the various embodiments may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least a portion of the sub-steps or stages of other steps.

[0052] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the disclosure in the specification and embodiments. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the claims.

Claims

1. A method for reverse optimization of process parameters for the mechanical properties of hot-rolled strip steel based on deep learning, characterized in that, The method includes the following steps: The process parameters and corresponding mechanical property data of the hot strip rolling process are collected. After time alignment, steady-state screening and anomaly removal, a sample set consisting of process parameter vectors and mechanical property vectors is constructed. Using process parameter vectors as input and mechanical property vectors as output, a deep performance prediction field is trained to learn the mapping relationship between process parameters and mechanical properties. Using the sample set, a deep confidence field is trained, with process parameter vectors as input, and scalar confidence is output. Using the feasible process parameters in the sample set, a deep feasibility potential field is trained, which approaches zero in the feasible region. For a given target mechanical performance vector and current benchmark process parameter vector, a cooperative energy function is constructed. The cooperative energy function is composed of three parts: a confidence-weighted performance prediction deviation term, a feasibility potential term, and an anchoring smoothing term. Starting from the current benchmark process parameters and several cluster centers of process parameters corresponding to historically similar target performance, a multi-starting point constraint descent search is performed along the gradient direction of the cooperative energy function, and the results of each iteration are projected back into the preset process hard constraint boundary to obtain a set of candidate process parameters. Solutions with feasibility potential higher than the first threshold or confidence lower than the second threshold are removed from the candidate process parameter set. The process parameter vector with the smallest value of the anchoring smoothing term is selected from the remaining solutions as the final output of the inverse optimization.

2. The method for reverse optimization of process parameters of hot-rolled strip steel based on deep learning according to claim 1, characterized in that, The steps for constructing a sample set include: Using the synchronous timestamps of the thickness gauge and pyrometer at the finish mill exit as a reference, the rolling parameters, temperature parameters, cooling parameters of each stand are aligned with the tensile test results; Records containing sensor malfunctions, rolling stoppages, and unsteady sections at the beginning and end of the strip are excluded; A local weighted regression smoothing method was used to detect and remove outlier sample points to obtain the process-performance sample set.

3. The method for reverse optimization of process parameters of hot-rolled strip steel based on deep learning according to claim 1, characterized in that, The steps for training a deep performance prediction field include: Inject random disturbances that conform to the statistical characteristics of field instrument noise into the input process parameter vector to generate small disturbance samples; This ensures that the depth performance prediction field outputs consistent mechanical performance prediction values ​​for multiple perturbation samples at the same process parameter point; By minimizing the joint objective of prediction error and consistency constraint, the performance prediction field remains smooth and accurate at the sample point and its neighborhood.

4. The method for reverse optimization of process parameters of hot-rolled strip steel based on deep learning according to claim 1, characterized in that, The steps for training a deep confidence field include: The process parameter points that are densely clustered in the sample set are marked as high-confidence samples; Randomly sample extrapolation points far from the sample clustering area in the process parameter space and mark them as low-confidence samples; Through comparative training, the confidence field is made to output a value close to 1 for high confidence samples and a value close to 0 for low confidence samples, with a smooth transition region.

5. The method for reverse optimization of process parameters of hot-rolled strip steel based on deep learning according to claim 1, characterized in that, The steps for training a deep feasible potential field include: The feasible process parameters in the sample set are used as positive samples; Three types of negative samples are generated: process parameter combinations that exceed equipment limits, associated combinations that violate process logic, and out-of-bounds points generated by perturbations along the edge of the feasible process manifold. Based on positive and negative samples, discriminative training with transition band smoothing constraints is introduced to make the feasible potential field output a low potential value near zero on positive samples and a high potential value on negative samples, and the potential field gradient is continuous at the boundary of the feasible region.

6. The method for reverse optimization of process parameters of hot-rolled strip mechanical properties based on deep learning according to claim 1, characterized in that, The cooperative energy function is: E(x) = C(x)·‖F(x) - y tgt || 2 +α·Φ(x)+β·‖xx cur || 2 S , x is the vector of process parameters to be optimized; F(x) is the vector of mechanical properties output by the depth performance prediction field; y tgt Let C(x) be the target mechanical performance vector; C(x) be the scalar confidence level output by the depth confidence field at x, serving as the pointwise weight of the performance deviation term; Φ(x) be the value of the depth feasibility potential field at x; α be the feasibility penalty coefficient; ||xx cur || 2 S For the current baseline process parameter x cur The weighted squared distance to x is S, which is a diagonal weight matrix composed of the adjustment costs of each process parameter; β is the smoothing anchoring coefficient.

7. The method for reverse optimization of process parameters of hot-rolled strip steel based on deep learning according to claim 1, characterized in that, The steps for performing a multi-starting-point constrained descent search specifically include: Determine a set of starting points, which includes at least the current baseline process parameter vector and cluster centers of process parameter vectors corresponding to orders with similar target mechanical properties identified from historical data; For each starting point, the gradient of the cooperative energy function with respect to the process parameter vector is calculated, the process parameter vector is updated along the negative gradient direction with a fixed step size, and the components of the updated vector are projected into the hard boundary determined by the equipment capability and process specifications. The process is iterated until the convergence condition is met, resulting in multiple local energy minima, which form a set of candidate process parameters.

8. The method for reverse optimization of process parameters of hot-rolled strip steel based on deep learning according to claim 1, characterized in that, The method further includes: After production according to the final output process parameters, the measured mechanical properties of the strip steel are obtained, and the deviation between the measured mechanical properties and the target mechanical property vector is calculated. With a fixed depth performance prediction field and depth confidence field, starting from the currently used process parameters, local gradient descent with a minimum step size is performed on the cooperative energy function to generate correction increments and superimpose them to achieve fast offset compensation without repetition optimization. New production data is periodically incorporated into the sample set, and the deep performance prediction field, deep credibility field, and deep feasibility potential field are incrementally retrained to ensure that each deep field evolves synchronously with the production process.

Images