Machine learning based array steam heat setting process optimization method

Abstract

The application discloses an array type steam heat setting process optimization method based on machine learning, which comprises the following steps: step one, constructing a disturbance voxel atlas; step two, constructing a regional disturbance response matrix; step three, generating a candidate micro-intervention control unit set; step four, generating a regional control intention vector based on the current cloth running state; step five, executing an auction scheduling, each region bidding according to the control intention vector, adopting a Vickrey auction mechanism to determine the winning region, and constructing a winning control atlas; step six, performing strategy re-parameter sampling according to the winning control atlas; step seven, when an abnormal state appears on the cloth surface, backtracking the winning control atlas and extracting an intervention control unit sequence; and step eight, adjusting the regional control intention vector and the bidding sequence based on the intervention control unit sequence, and forming a new winning control atlas. The application combines voxel modeling and the Vickrey auction mechanism, and realizes intelligent optimization of the array type steam heat setting process.

Description

Technical Field

[0001] This invention relates to the field of textile heat setting process optimization technology, and in particular to an array-based steam heat setting process optimization method based on machine learning. Background Technology

[0002] With the continued growth in market demand for high-end functional fabrics, differentiated fiber products, and precision heat-treated fabrics, array-type steam heat setting equipment is increasingly widely used in high-efficiency, wide-width, and highly uniform heat setting processes. However, existing steam heat setting control systems mainly rely on static empirical parameters, regional thresholding strategies, or traditional single optimization algorithms to regulate the steam behavior in the chamber, which generally suffers from the following problems under complex operating conditions:

[0003] Existing control methods lack systematic modeling and feedback adjustment mechanisms for the coupling effects of thermal and moisture disturbances between multi-chamber steam behaviors. This leads to problems such as uneven tension, abnormal heat spot diffusion, and pattern deformation in the fabric after multi-regional intervention. The thermal disturbance propagation paths between steam chambers are complex, and the disturbance response exhibits significant hysteresis and nonlinear characteristics, making it difficult to achieve global optimization using traditional independent regional control methods. The lack of dynamic competition and differentiated response capability measurement in the intervention resource allocation process results in insufficient steam response in high-demand areas and excessive intervention in low-load areas, leading to low resource utilization efficiency. Control strategy adjustments rely on fixed rules or optimization objectives constructed based on empirical weights, lacking data-driven feedback correction mechanisms, making it difficult to adapt to actual production scenarios such as fabric batch differences, fabric thickness variations, or fluctuations in operating status. Furthermore, current methods often rely on post-event machine shutdown adjustments or experience-based recovery strategies for handling abnormal operating conditions, failing to achieve rapid location of the source of the anomaly and accurate backtracking of responsible intervention units. This affects the overall machine operating efficiency and the stability of the setting quality, hindering the development of array-type steam heat setting processes towards higher precision, adaptability, and intelligence.

[0004] Therefore, how to provide a machine learning-based optimization method for array-type steam thermal setting process is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a machine learning-based optimization method for array-type steam heat setting processes. This invention integrates voxel modeling and the Vickery auction mechanism to achieve intelligent optimization of the entire array-type steam heat setting process. By constructing a perturbation voxel map, the dynamic coupling relationship of steam intervention between multiple chambers is accurately captured, generating a highly efficient and controllable micro-intervention control unit. Furthermore, the Vickery auction is used to achieve differentiated allocation of intervention resources and the optimal bidding strategy, significantly improving fabric setting quality, control accuracy, and system resource utilization efficiency. This method is suitable for intelligent steam control tasks involving complex fabrics and variable operating conditions.

[0006] An array-type steam thermal setting process optimization method based on machine learning according to an embodiment of the present invention includes the following steps:

[0007] Step 1: Divide each steam chamber along the fabric running path in the array-type steam heat setting equipment into three-dimensional voxel units according to the spatial and temporal dimensions, and construct a perturbation voxel map;

[0008] Step 2: Perform sequential perturbation excitation, activate the steam control unit of each steam chamber in turn, obtain the tension response and thermographic response of adjacent and non-adjacent steam chambers, and construct the regional perturbation response matrix;

[0009] Step 3: Discretize the control behavior of each steam chamber into several micro-intervention control units. Each micro-intervention control unit includes a control time segment, steam opening level and corresponding three-dimensional voxel unit index. Combine the perturbation voxel map and the regional perturbation response matrix to generate a set of candidate micro-intervention control units.

[0010] Step 4: Generate a region control intent vector based on the current fabric running status;

[0011] Step 5: Perform auction scheduling for each candidate micro-intervention control unit. Each region bids according to the control intent vector. The winning region is determined by the Vickrey auction mechanism, and the winning control map is constructed.

[0012] Step 6: Perform strategy reparameter sampling based on the winning control map to generate a control strategy for controlling the behavior of the steam chamber;

[0013] Step 7: When an abnormal state occurs on the fabric surface, based on the index mapping relationship of the abnormal area in the perturbation voxel map, backtrack the corresponding indexed control map and extract the sequence of intervention and control units that caused the abnormality;

[0014] Step 8: Based on the intervention control unit sequence, adjust the regional control intent vector and the corresponding bidding order, re-execute the resource auction scheduling, and form a new winning bid control map.

[0015] Optionally, the construction of the perturbation voxel map specifically involves:

[0016] The steam chambers along the fabric running path in the array-type steam heat setting equipment are divided into several regions. Each region is divided along the fabric running direction according to a set length, and the state of each region is recorded in the time dimension at a set sampling period to form a three-dimensional voxel unit with spatial position index and time index.

[0017] The data fields in the three-dimensional voxel unit include the steam valve opening and closing status, steam opening value, tension sensor reading value, infrared thermal imaging temperature value, humidity sensor data, and fabric image texture feature value.

[0018] All three-dimensional voxel units are organized according to spatial region numbering and chronological order to form a perturbation voxel map with a three-dimensional coordinate structure.

[0019] Optionally, step two specifically includes:

[0020] The sequential perturbation excitation activates the steam control unit of each steam chamber in the array-type steam heat setting equipment in turn. During each activation process, only the steam valve of the current target steam chamber is opened, while the steam valves of the other chambers remain closed. The steam excitation lasts for a preset time.

[0021] During each steam activation process, tension sensors placed at the inlet, center, and outlet of each steam chamber collect the fabric tension values ​​at multiple consecutive time points starting from the current activation moment, and calculate the tension changes in the corresponding tension channels of adjacent and non-adjacent steam chambers to form a multi-region tension response corresponding to the activation event.

[0022] By using infrared thermal imaging devices installed above and on the side walls of each steam chamber, thermal image sequences of the fabric in each region are collected. The temperature gradient changes, hot spot diffusion range and maximum temperature rise location of each region before and after steam excitation are extracted to form thermal response corresponding to adjacent and non-adjacent steam chambers.

[0023] Based on the tension response and thermographic response data of all regions obtained after each steam chamber excitation, a regional perturbation response matrix is ​​constructed by combining the indices of the excitation source steam chamber and the response target steam chamber. Each element of the regional perturbation response matrix corresponds to the perturbation effect of one steam chamber on another steam chamber region, which is used to quantify the thermal and moisture perturbation coupling relationship between each steam chamber in the array structure.

[0024] Optionally, step three specifically includes:

[0025] The complete steam control cycle of each steam chamber is divided into multiple non-overlapping time segments, the duration of which is a preset control time interval;

[0026] For each time segment, multiple steam opening levels are set. The steam opening level is a discrete set of values ​​divided in increments of 10%, including 10%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, 90%, and 100%, to form a set of steam action parameter combinations within each time segment.

[0027] For each steam chamber, under the combination of time segments and steam opening level, a corresponding micro-intervention control unit is formed by combining the spatial position index of the steam chamber. Each micro-intervention control unit is bound to a unique control time segment, steam opening level, and three-dimensional voxel unit index in the perturbation voxel map.

[0028] For all micro-intervention control units, based on the disturbance influence in the regional disturbance response matrix, the tension response and heatmap response generated by each micro-intervention control unit at the corresponding voxel are retrieved, and a disturbance propagation score is generated by weighted linear combination of values.

[0029] All micro-intervention control units are sorted according to the disturbance propagation score, and those micro-intervention control units below the preset disturbance impact threshold are eliminated. They are then clustered and integrated according to the three-dimensional voxel unit index to form a candidate micro-intervention control unit set, which is used as the resource bidding object in the auction scheduling process.

[0030] Optionally, step four: generating a region control intent vector based on the current fabric operating state, specifically:

[0031] For each steam chamber in the array-type steam heat setting equipment, a control intent vector for the corresponding area is generated based on the current operating status data of the fabric in that area.

[0032] The current operating status data includes the difference between the real-time tension value collected by the regional tension sensor and the historical tension baseline value, the fabric temperature change gradient value collected by the infrared thermal imaging device, the proportion of fabric defects detected by the fabric image recognition module, and the perturbation hysteresis weight of the region in the perturbation voxel map.

[0033] After standardizing the current operating status data, a regional control intent vector is constructed by combining the data. The control intent vector is a set of real-valued vectors, including tension deviation, temperature change gradient, fabric defect ratio, and disturbance hysteresis weight.

[0034] Optionally, the auction scheduling of each candidate micro-intervention control unit, with each region bidding according to the control intent vector, and the winning region determined by the Vickrey auction mechanism, specifically involves:

[0035] Each candidate micro-intervention control unit is set as an independent auction unit, and each auction unit contains a unique steam chamber number, control time period, steam opening level and three-dimensional voxel unit index.

[0036] A Vickery auction execution instance is initialized for each auction unit. The Vickery auction execution instance is configured with a single round, closed-ended, and second-highest-bid rule. After the auction starts, all participating regions submit the bid value of the auction unit in a non-interactive state, and modification or withdrawal is prohibited.

[0037] The bid value of the auction unit is generated by inputting the corresponding control intention vector into the regional bid calculation module. The regional bid calculation module combines the tension deviation, temperature change gradient, fabric defect ratio and disturbance hysteresis weight into a floating-point scalar value according to a preset weight, which serves as the region's bid input for the current candidate micro-intervention control unit.

[0038] Within each auction unit, the bid values ​​of all auction units are sorted, and the area with the highest bid value is determined as the winning bid area. The bid value of the area with the second highest bid value is used as the payment price for the winning bid area, and the unique assignment relationship between the winning bid area and the candidate micro-intervention control unit is recorded.

[0039] Optionally, the construction of the winning bid control map specifically includes:

[0040] All candidate micro-intervention control units in the winning areas are aggregated into the winning control unit set, sorted according to time index and voxel spatial index, and a structured winning control map is constructed. The winning control map is a two-dimensional mapping structure, with the time dimension corresponding to the control time period and the spatial dimension corresponding to the three-dimensional voxel position. Each cell of the winning control map includes the steam chamber number, steam opening value and winning area number.

[0041] Optionally, the strategy re-parameter sampling extracts all control time periods from the winning control map, constructs a dynamic control time series index, uses each control time period as an independent sampling window, and initializes the control parameter template in each sampling window according to the configuration of the winning control unit.

[0042] Within each sampling window, a strategy reparameter sampling process is performed, including: extracting the steam chamber number and corresponding spatial location of the winning control unit, and locking the current control area to be sampled; extracting the winning steam opening level as the initial value of the control intensity; based on the fabric disturbance voxel state within the current sampling window, inputting the disturbance voxel state features into the strategy adjustment model, adjusting the steam opening level and time offset to form the final action parameters, and generating a control strategy for controlling the behavior of the steam chamber. The strategy adjustment model adopts a parameter generator based on supervised training, with the disturbance voxel state features as input and the disturbance adjustment value of the control action parameters as output, which is used to disturb the initial value of the winning control intensity to achieve customized control.

[0043] The control strategy consists of a steam chamber number, a control time period, an actual steam opening level, and a corresponding three-dimensional voxel index quadruple. This set of control instructions is sent to the execution end of the array-type steam heat setting equipment to drive the corresponding steam chambers to perform steam intervention behavior according to the bidding order and the final action parameters.

[0044] Optionally, step seven specifically includes:

[0045] During the operation of the array-type steam heat setting equipment, abnormal conditions of the fabric surface are detected in real time. When an abnormality occurs in the fabric surface, the specific spatial location of the abnormality area is determined by infrared thermal imaging and fabric surface image recognition.

[0046] Based on the spatial location of the abnormal area on the fabric surface and the time index of the time when the abnormality occurred, find the three-dimensional voxel unit index in the perturbation voxel map that corresponds to the abnormal area in terms of spatial location and time.

[0047] Starting from the three-dimensional voxel unit index, the process traces back along the time axis to a preset backtracking period before the time of the anomaly occurrence, and searches step by step for candidate micro-intervention control units that have been executed on the abnormal region in the winning control map.

[0048] The retrieved candidate micro-intervention control units are sorted in order of execution time from most recent to oldest, forming a sequence of intervention control units that cause abnormal fabric conditions.

[0049] Optionally, step eight specifically includes:

[0050] Based on the intervention control unit sequence, determine the steam chamber number and spatial location corresponding to each candidate micro-intervention control unit in the intervention control unit sequence, and count the number of interventions and the degree of contribution to the anomaly in the area corresponding to each steam chamber before the anomaly occurred.

[0051] Based on the number of interventions and the degree of abnormal contribution, the tension deviation, temperature change gradient, fabric defect ratio and disturbance hysteresis weight of the corresponding area in the original regional control intention vector are corrected according to the preset weights.

[0052] The bid value for each region for each candidate micro-intervention control unit is recalculated based on the revised regional control intent vector, and the Vickrey auction mechanism is re-executed based on the updated regional bid value to determine the new winning region.

[0053] All candidate micro-intervention control units in the newly selected regions are aggregated into the set of selected control units, and sorted according to the time index and voxel spatial index to form a new selected control map.

[0054] The beneficial effects of this invention are:

[0055] This invention addresses the challenges of quantifying multi-chamber coupled disturbances, control strategy response lag, and uneven resource allocation in array-type steam heat setting equipment by constructing a dynamic modeling framework that combines perturbation voxel maps and regional perturbation response matrices. It employs three-dimensional space-time voxelization modeling technology to accurately capture the propagation path and time delay characteristics of steam excitation on tension and temperature changes. In the micro-intervention control unit generation stage, high-impact control candidates are screened through perturbation propagation scores, significantly compressing the control action space and improving control accuracy. Furthermore, a resource competition scheduling method based on the Vickrey auction mechanism is introduced to encode the fabric operating state into control parameters. The intent vector serves as the bidding basis for each region, enabling differentiated competition and adaptive control resource allocation among chambers, preventing high-intervention-demand regions from being disturbed by low-demand regions. A strategy reparameter sampling mechanism is introduced in the strategy generation stage, using a parameter generator driven by disturbance state characteristics to fine-tune the control strength and execution time, achieving dynamic and refined adaptive adjustment of the control strategy. During fabric anomaly handling, index mapping enables reverse tracing of the responsible control unit in the anomaly region, and retrospective updates of the region's control intent and bidding strategy are performed based on the anomaly contribution, forming a closed-loop self-correction mechanism driven by anomalies. Ultimately, this achieves dynamic optimization control across chambers, multiple time periods, and regions in the array-type steam heat setting process, improving steam resource utilization efficiency and overall system robustness while ensuring fabric heat setting quality. Attached Figure Description

[0056] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0057] Figure 1 This is an overall flowchart of an array-type steam thermal shaping process optimization method based on machine learning proposed in this invention.

[0058] Figure 2 This is a flowchart of the auction scheduling process based on the Vickrey auction mechanism for an array-type steam thermal shaping process optimization method proposed in this invention.

[0059] Figure 3 This is a flowchart illustrating the operation of generating a control strategy through strategy reparameter sampling in an array-based steam thermal shaping process optimization method based on machine learning proposed in this invention. Detailed Implementation

[0060] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0061] refer to Figures 1-3An optimization method for array-type steam thermal setting process based on machine learning includes the following steps:

[0062] Step 1: Divide each steam chamber along the fabric running path in the array-type steam heat setting equipment into three-dimensional voxel units according to the spatial and temporal dimensions, and construct a perturbation voxel map;

[0063] Step 2: Perform sequential perturbation excitation, activate the steam control unit of each steam chamber in turn, obtain the tension response and thermographic response of adjacent and non-adjacent steam chambers, and construct the regional perturbation response matrix;

[0064] Step 3: Discretize the control behavior of each steam chamber into several micro-intervention control units. Each micro-intervention control unit includes a control time segment, steam opening level and corresponding three-dimensional voxel unit index. Combine the perturbation voxel map and the regional perturbation response matrix to generate a set of candidate micro-intervention control units.

[0065] Step 4: Generate a region control intent vector based on the current fabric running status;

[0066] Step 5: Perform auction scheduling for each candidate micro-intervention control unit. Each region bids according to the control intent vector. The winning region is determined by the Vickrey auction mechanism, and the winning control map is constructed.

[0067] Step 6: Perform strategy reparameter sampling based on the winning control map to generate a control strategy for controlling the behavior of the steam chamber;

[0068] Step 7: When an abnormal state occurs on the fabric surface, based on the index mapping relationship of the abnormal area in the perturbation voxel map, backtrack the corresponding indexed control map and extract the sequence of intervention and control units that caused the abnormality;

[0069] Step 8: Based on the intervention control unit sequence, adjust the regional control intent vector and the corresponding bidding order, re-execute the resource auction scheduling, and form a new winning bid control map.

[0070] In this embodiment, the construction of the perturbation voxel map specifically involves:

[0071] The steam chambers along the fabric running path in the array-type steam heat setting equipment are divided into several regions. Each region is divided along the fabric running direction according to a set length, and the state of each region is recorded in the time dimension at a set sampling period to form a three-dimensional voxel unit with spatial position index and time index.

[0072] The data fields in the three-dimensional voxel unit include the steam valve opening and closing status, steam opening value, tension sensor reading value, infrared thermal imaging temperature value, humidity sensor data, and fabric image texture feature value.

[0073] All three-dimensional voxel units are organized according to spatial region numbering and chronological order to form a perturbation voxel map with a three-dimensional coordinate structure.

[0074] In this embodiment, step two specifically includes:

[0075] The sequential perturbation excitation activates the steam control unit of each steam chamber in the array-type steam heat setting equipment in turn. During each activation process, only the steam valve of the current target steam chamber is opened, while the steam valves of the other chambers remain closed. The steam excitation lasts for a preset time.

[0076] During each steam activation process, tension sensors placed at the inlet, center, and outlet of each steam chamber collect the fabric tension values ​​at multiple consecutive time points starting from the current activation moment, and calculate the tension changes in the corresponding tension channels of adjacent and non-adjacent steam chambers to form a multi-region tension response corresponding to the activation event.

[0077] By using infrared thermal imaging devices installed above and on the side walls of each steam chamber, thermal image sequences of the fabric in each region are collected. The temperature gradient changes, hot spot diffusion range and maximum temperature rise location of each region before and after steam excitation are extracted to form thermal response corresponding to adjacent and non-adjacent steam chambers.

[0078] In the specific implementation process, to obtain information on the physical response of the fabric to the steam excitation behavior, tension sensors are placed at the inlet, center, and outlet of each steam chamber to collect tension change data in real time during each steam excitation process. At least three tension sensors are placed in each steam chamber to record the fabric tension values ​​at multiple time points from the start of steam excitation to the end of excitation, with a time sampling frequency of no less than 10Hz to ensure the capture of short-term disturbance effects. The collected data, after filtering and normalization, is then compared with the tension baseline value before steam excitation to calculate the change in tension of the fabric segments in the current excitation chamber in response to the current excitation chamber and its adjacent and non-adjacent chambers, thus forming tension response data.

[0079] In addition, to obtain the changes in heat distribution during the steam excitation process, infrared thermal imaging devices were installed above and on the side walls of each steam chamber. These devices acquired real-time thermal image sequences of the fabric surface at a frame rate of no less than 30fps. The images before and after excitation were processed using inter-frame differencing and image enhancement techniques to extract the temperature change curves of each fabric region along the time axis. Furthermore, image analysis algorithms were used to calculate the temperature gradient change value, hot spot diffusion area, and coordinates of the maximum temperature rise point of the hot spot, thereby constructing thermal response features to reflect the thermal disturbance behavior induced by steam excitation on the fabric surface.

[0080] Tension response and thermal response data are synchronously bound to the steam excitation control signal via timestamps and mapped to voxel indices in the perturbation voxel map based on the correspondence between the chamber numbers of each sensor or imaging device and the fabric position. Ultimately, the steam excitation behavior and the tension and thermal response data induced in different chamber regions constitute a multi-region intervention response pair, used to construct a regional perturbation response matrix. This regional perturbation response matrix quantifies the degree of coupling influence of any steam chamber excitation behavior on the physical state of the fabric in other regions, reflecting the lateral diffusion and longitudinal conduction characteristics of steam perturbation in an array structure.

[0081] Based on the tension response and thermographic response data of all regions obtained after each steam chamber excitation, a regional perturbation response matrix is ​​constructed by combining the indices of the excitation source steam chamber and the response target steam chamber. Each element of the regional perturbation response matrix corresponds to the perturbation effect of one steam chamber on another steam chamber region, which is used to quantify the thermal and moisture perturbation coupling relationship between each steam chamber in the array structure.

[0082] In this embodiment, step three specifically includes:

[0083] The complete steam control cycle of each steam chamber is divided into multiple non-overlapping time segments, the duration of which is a preset control time interval;

[0084] For each time segment, multiple steam opening levels are set. The steam opening level is a discrete set of values ​​divided in increments of 10%, including 10%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, 90%, and 100%, to form a set of steam action parameter combinations within each time segment.

[0085] For each steam chamber, under the combination of time segments and steam opening level, a corresponding micro-intervention control unit is formed by combining the spatial position index of the steam chamber. Each micro-intervention control unit is bound to a unique control time segment, steam opening level, and three-dimensional voxel unit index in the perturbation voxel map.

[0086] For all micro-intervention control units, based on the disturbance influence in the regional disturbance response matrix, the tension response and heatmap response generated by each micro-intervention control unit at the corresponding voxel are retrieved, and a disturbance propagation score is generated by weighted linear combination of values.

[0087] ;

[0088] in, The score represents the perturbation propagation score, and α and β represent weighting coefficients used to adjust the contribution ratio of the tension response and the heatmap response to the total score. Indicates tension response, Indicates the heatmap response;

[0089] All micro-intervention control units are sorted according to the disturbance propagation score, and those micro-intervention control units below the preset disturbance impact threshold are eliminated. They are then clustered and integrated according to the three-dimensional voxel unit index to form a candidate micro-intervention control unit set, which is used as the resource bidding object in the auction scheduling process.

[0090] In this embodiment, step four: generating a region control intent vector based on the current fabric running state, specifically includes:

[0091] For each steam chamber in the array-type steam heat setting equipment, a control intent vector for the corresponding area is generated based on the current operating status data of the fabric in that area.

[0092] The current operating status data includes the difference between the real-time tension value collected by the regional tension sensor and the historical tension baseline value, the fabric temperature change gradient value collected by the infrared thermal imaging device, the proportion of fabric defects detected by the fabric image recognition module, and the perturbation hysteresis weight of the region in the perturbation voxel map.

[0093] The disturbance hysteresis weight measures the physical response (tension change or temperature change) of a region after receiving steam intervention relative to the time delay of the intervention action. It is used to characterize the region's response sensitivity and control difficulty. The acquisition process is as follows: In the historical operating data, for each execution record of the micro-intervention control unit, the time delay experienced by it inducing tension change or temperature change in the corresponding region is calculated. The historical intervention response time delays of all regions are accumulated to calculate the average hysteresis time. Then, it is normalized according to the preset maximum allowable hysteresis time to obtain the disturbance hysteresis weight.

[0094] After standardizing the current operating status data, a regional control intent vector is constructed. The control intent vector is a set of real-valued vectors, including tension deviation, temperature change gradient, fabric defect ratio, and disturbance hysteresis weight, which are used to characterize the region's intervention needs and urgency for the micro-intervention control unit.

[0095] In this embodiment, the auction scheduling of each candidate micro-intervention control unit, with each region bidding according to the control intent vector, and the winning region determined by the Vickrey auction mechanism, specifically involves:

[0096] Each candidate micro-intervention control unit is set as an independent auction unit, and each auction unit contains a unique steam chamber number, control time period, steam opening level and three-dimensional voxel unit index.

[0097] A Vickery auction execution instance is initialized for each auction unit. The Vickery auction execution instance is configured with a single round, closed-ended, and second-highest-bid rule. After the auction starts, all participating regions submit the bid value of the auction unit in a non-interactive state, and modification or withdrawal is prohibited.

[0098] The bid value of the auction unit is generated by inputting the corresponding control intention vector into the regional bid calculation module. The regional bid calculation module combines the tension deviation, temperature change gradient, fabric defect ratio and disturbance hysteresis weight into a floating-point scalar value according to a preset weight, which serves as the region's bid input for the current candidate micro-intervention control unit.

[0099] Within each auction unit, the bid values ​​of all auction units are sorted, and the area with the highest bid value is determined as the winning bid area. The bid value of the area with the second highest bid value is used as the payment price for the winning bid area, and the unique assignment relationship between the winning bid area and the candidate micro-intervention control unit is recorded.

[0100] In traditional array-type steam heat setting processes, the intervention behavior of each steam chamber is usually controlled by static rules or adjusted by a single optimization objective. This fails to reasonably reflect the coupling relationship and resource competition between complex multi-regions, resulting in low control accuracy, frequent disturbances, and difficulty in meeting the requirements of high-quality fabric setting processes.

[0101] This invention introduces the Vickrey auction mechanism into the array-type steam thermal shaping control process. By setting each candidate micro-intervention control unit as an independent auction unit, a dynamic competition mechanism for micro-scale control resources is formed, thereby breaking the static constraints of traditional control technology on the allocation of intervention resources.

[0102] Specifically, in this invention, the regional control intent vector is formed by quantifying and calculating the current fabric tension deviation, temperature change gradient, fabric defect ratio, and disturbance hysteresis weight, and integrating multiple physical state information to form a unified regional bidding basis. Then, resource competition is carried out in a single-round closed Vickrey auction, with each region bidding independently and the right to use resources determined by the second highest price. This effectively avoids the problem of strategic overestimation or underestimation of demand between regions, prompting each region to objectively and accurately express its own intervention needs, and significantly improving the process optimization efficiency and control stability under complex multi-region coupling conditions.

[0103] In this embodiment, the construction of the winning bid control map specifically involves:

[0104] All candidate micro-intervention control units in the winning areas are aggregated into the winning control unit set, sorted according to the time index and voxel spatial index, and a structured winning control map is constructed. The winning control map is a two-dimensional mapping structure, with the time dimension corresponding to the control time period and the spatial dimension corresponding to the three-dimensional voxel position. Each cell of the winning control map includes the steam chamber number, steam opening value and winning area number.

[0105] The winning control map serves as a template input source for subsequent control strategies. It is used to define the steam intervention execution behavior and regional correspondence of each chamber during the steam heat setting process, ensuring that the control behavior and priority order originate from the competitive allocation results under the Vickrey auction mechanism.

[0106] In this embodiment, the strategy re-parameter sampling extracts all control time periods from the winning bid control map, constructs a dynamic control time series index, uses each control time period as an independent sampling window, and initializes the control parameter template in each sampling window according to the configuration of the winning bid control unit.

[0107] Within each sampling window, a strategy reparameter sampling process is performed, including: extracting the steam chamber number and corresponding spatial location of the winning control unit, and locking the current control area to be sampled; extracting the winning steam opening level as the initial value of the control intensity; based on the fabric disturbance voxel state within the current sampling window, inputting the disturbance voxel state features into the strategy adjustment model, adjusting the steam opening level and time offset to form the final action parameters, and generating a control strategy for controlling the behavior of the steam chamber. The strategy adjustment model adopts a parameter generator based on supervised training, with the disturbance voxel state features as input and the disturbance adjustment value of the control action parameters as output, which is used to disturb the initial value of the winning control intensity to achieve customized control.

[0108] The strategy adjustment model is a set of lightweight neural network modules trained by supervised learning, which are set up according to the steam chamber numbering. The strategy adjustment model structure includes an input layer, at least one hidden layer and an output layer. The input layer receives the perturbation voxel state features, which include the tension gradient of the current voxel, the heat map temperature gradient, the historical response hysteresis value and the fabric surface anomaly prediction index. After standardization, they are used as the model input vector.

[0109] The output of the strategy adjustment model is two continuous real values: steam opening disturbance correction amount and time offset adjustment amount. The output value is combined with the steam opening level and execution time of the winning control unit to generate the final parameter configuration of the control action.

[0110] The control strategy consists of a steam chamber number, a control time period, an actual steam opening level, and a corresponding three-dimensional voxel index quadruple. This set of control instructions is sent to the execution end of the array-type steam heat setting equipment to drive the corresponding steam chambers to perform steam intervention behavior according to the bidding order and the final action parameters.

[0111] In this embodiment, step seven specifically includes:

[0112] During the operation of the array-type steam heat setting equipment, abnormal conditions of the fabric surface are detected in real time. When an abnormality occurs in the fabric surface, the specific spatial location of the abnormality area is determined by infrared thermal imaging and fabric surface image recognition.

[0113] Based on the spatial location of the abnormal area on the fabric surface and the time index of the time when the abnormality occurred, find the three-dimensional voxel unit index in the perturbation voxel map that corresponds to the abnormal area in terms of spatial location and time.

[0114] Starting from the three-dimensional voxel unit index, the process traces back along the time axis to a preset backtracking period before the time of the anomaly occurrence, and searches step by step for candidate micro-intervention control units that have been executed on the abnormal region in the winning control map.

[0115] The retrieved candidate micro-intervention control units are sorted in order of execution time from most recent to oldest, forming a sequence of intervention control units that cause abnormal fabric conditions.

[0116] In this embodiment, step eight specifically includes:

[0117] Based on the intervention control unit sequence, determine the steam chamber number and spatial location corresponding to each candidate micro-intervention control unit in the intervention control unit sequence, and count the number of interventions and the degree of contribution to the anomaly in the area corresponding to each steam chamber before the anomaly occurred.

[0118] Based on the number of interventions and the degree of abnormal contribution, the tension deviation, temperature change gradient, fabric defect ratio and disturbance hysteresis weight of the corresponding area in the original regional control intention vector are corrected according to the preset weights.

[0119] The bid value for each region for each candidate micro-intervention control unit is recalculated based on the revised regional control intent vector, and the Vickrey auction mechanism is re-executed based on the updated regional bid value to determine the new winning region.

[0120] All candidate micro-intervention control units in the newly selected regions are aggregated into the set of selected control units, and sorted according to the time index and voxel spatial index to form a new selected control map.

[0121] The degree of anomaly contribution is an indicator used to quantify the intensity of the influence of each corresponding area of ​​the steam chamber on fabric anomalies. The specific determination method is as follows:

[0122] Based on the sequence of intervention control units extracted in step seven, obtain the number of the steam chamber, steam opening level, execution time point and corresponding three-dimensional voxel unit index of each intervention control unit;

[0123] Using the perturbation voxel map constructed in step one, we trace back and calculate the changes in the tension deviation, temperature change gradient, and fabric defect ratio of the fabric in the corresponding spatial area after each intervention control unit is executed. The magnitude of the changes in the above three values ​​is used as the single contribution value of the intervention control unit to the fabric abnormality.

[0124] For each area corresponding to a steam chamber, the single contribution values ​​of all intervention control units are accumulated within a preset time range before the anomaly occurs to obtain the anomaly contribution value of that area.

[0125] Calculate the sum of the abnormal contribution values ​​of all regions, and then normalize the abnormal contribution value of each region by dividing it by the sum of the abnormal contribution values ​​of all regions to obtain the degree of abnormal contribution of that region. This value represents the proportion of the contribution of the intervention behavior in that region to the overall abnormality.

[0126] After calculating the degree of abnormal contribution, the system corrects the four indicators in the original regional control intent vector of each region based on the number of interventions executed and the degree of abnormal contribution: tension deviation, temperature change gradient value, fabric defect ratio, and disturbance hysteresis weight.

[0127] Specifically, a set of non-negative weighting coefficients is pre-defined. For each region, the original values ​​of the above four indicators are weighted linearly adjusted using the number of interventions and the degree of abnormal contribution, i.e.:

[0128] The revised indicator value = original indicator value × (1 + weight of intervention execution times × execution times ratio + weight of abnormal contribution × degree of abnormal contribution);

[0129] Among them, the execution frequency ratio represents the proportion of the number of interventions executed in this region to the total number of interventions in all regions. The weight of the number of interventions executed and the weight of abnormal contributions are non-negative real numbers preset based on historical data analysis experience, used to control the strength of the two influencing factors.

[0130] Finally, a new regional control intent vector is constructed using the corrected tension deviation, temperature change gradient, fabric defect ratio, and disturbance hysteresis weight. This vector is then used to recalculate the output value of candidate micro-intervention control units during the next resource auction scheduling process.

[0131] Example 1:

[0132] To verify the feasibility of this invention in practice, it was applied to a large-scale intelligent textile heat setting production line. This production line is equipped with 20 steam chambers, which are closely arranged and support independent steam control. The processed fabrics are four typical industrial fabrics: heavy knitted fabric, lightweight silk fabric, elastic blended fabric, and coarse-textured cotton-linen fabric. Each type of fabric is 120 meters long, and the initial tension fluctuates unstablely, with a fabric defect rate higher than the industry average.

[0133] Before the experiment, the steam heat-setting equipment was modified according to the requirements of this invention, including the installation of tension sensors and thermal imaging devices, the configuration of edge computing nodes for voxel modeling and disturbance response analysis, and the deployment of a server for control auction scheduling. During implementation, the coupled effects of thermal and moisture disturbance propagation between different steam chambers were first obtained by constructing a disturbance voxel map and a regional disturbance response matrix. Subsequently, a regional control intent vector was generated based on the real-time tension of the fabric and the thermal map status.

[0134] This invention introduces a Vickrey auction mechanism to schedule steam resources. Different regions bid for candidate micro-intervention control units based on their own conditions, and the system automatically determines the most suitable winning region, forming a dynamic bidding control map. The control strategy is generated using a reparameter sampling method, and the supervised training model dynamically adjusts the opening level and time offset according to the state of the perturbation voxels to achieve fine control.

[0135] Traditional uniform distribution strategies in steam heat setting evenly allocate steam resources (such as steam opening degree and action time) to each steam chamber or action area. Comparative experiments involving 50 fabrics in each of four batches revealed that under the traditional uniform distribution strategy, fabric tension fluctuations were high, heat setting temperature uniformity was low, and hot spot accumulation was prone to occur. However, under the strategy of this invention, fabric tension stability was significantly improved, infrared thermal imaging showed a more uniform temperature distribution, and the dimensional stability of the fabric after heat setting was significantly improved.

[0136] Table 1. Comparison and Analysis of Control Strategies between Traditional Methods and the Method of This Invention

[0137]

[0138] Analyzing the data in Table 1 above, the average tension deviation data shows that the tension deviation of all fabric types under the control strategy of this invention is significantly lower than that of the traditional strategy. For example, the average tension deviation of heavy knitted fabric decreased from the traditional 18.4N to 9.8N, a reduction of 46.7%; while that of elastic blended fabric decreased from 15.6N to 7.9N, a reduction of nearly 50%. This indicates that by constructing a perturbation voxel map and a regional control intention vector, this invention effectively alleviates the problem of excessive tension fluctuation in traditional heat setting and improves the stability of the fabric tension distribution.

[0139] Regarding defect rates, this invention also demonstrates superior control. Taking thin silk fabric as an example, the average defect rate under the traditional strategy was 7.2%, while after optimization it dropped to 4.1%, a reduction of nearly 43%; the average defect rate of coarse-textured cotton and linen fabric also decreased from 6.1% to 3.7%. This indicates that the refined intervention mechanism based on strategy reparameter sampling and Vickrey auction can improve heat setting uniformity and reduce surface defects such as wrinkles, shrinkage, and edge curling caused by uneven steam excitation.

[0140] Table 2 Comparison of Regional Temperature Control Uniformity

[0141]

[0142] Table 2 above, through a comparison of the temperature uniformity of the five steam-affected zones (A~E), comprehensively verifies the advantages of the Vickrey auction mechanism control strategy adopted in this invention in temperature distribution control. The data shows that under the traditional uniform scheduling control strategy, the regional temperature uniformity is generally maintained between 84.6% and 87.3%, while under the dynamic resource allocation method based on the Vickrey auction mechanism proposed in this invention, the regional temperature uniformity is significantly improved, with the average value stabilizing between 92.5% and 95.1%.

[0143] Specifically, region A achieved a temperature uniformity of 86.1% under the traditional strategy, while this increased to 94.6% using the strategy of this invention, representing an improvement of 8.5%. Regions B and D achieved improvements of 7.7% and 7.9%, respectively. Region C showed the most outstanding performance, achieving a temperature uniformity of 95.1% under the Vickrey strategy, a 7.8% improvement over the conventional strategy. Region E also saw a significant improvement of 8.4%. Overall, the temperature control uniformity improvement in all regions exceeded 7%, and the standard deviation was significantly reduced, demonstrating high consistency and stability.

[0144] This fully demonstrates that the mechanism introduced in this invention, which uses region-controlled intent vector bidding and a Vickrey auction mechanism to determine the intervention resource allocation strategy, can more intelligently coordinate the intensity and timing of steam chamber interventions. This enables more precise and adaptive dynamic adjustment of heat energy distribution during fabric heat setting, solving problems such as heat spots, heat accumulation, and insufficient edge heating that easily occur in traditional control methods. It is particularly suitable for setting heterogeneous fabrics and fabrics with complex textures. The implementation of this technology significantly improves the consistency and controllability of the final fabric heat treatment quality, possessing extremely high value for widespread application.

[0145] This embodiment achieves intelligent control of the array-type steam heat setting process by introducing perturbation voxel mapping modeling, regional control intent vector expression, and Vickrey auction mechanism, effectively improving tension control stability and temperature distribution uniformity. Compared with the traditional average allocation strategy, the method of this invention can accurately match the micro-intervention control unit according to the real-time fabric status, making steam resource allocation more targeted and flexible, avoiding the problems of heat energy waste and excessive intervention. At the same time, through the anomaly backtracking and dynamic correction mechanism of control intent vector, it has the ability to quickly respond to abnormal situations and adaptively optimize, significantly enhancing the robustness of the system and the stability of fabric heat setting. Overall, this method has the comprehensive advantages of high intelligence, excellent control accuracy, strong stability, and wide adaptability, and is suitable for complex fabrics and high-quality heat setting scenarios.

[0146] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A machine learning-based optimization method for array-type steam thermal setting process, characterized in that, Includes the following steps: Step 1: Divide each steam chamber along the fabric running path in the array-type steam heat setting equipment into three-dimensional voxel units according to the spatial and temporal dimensions, and construct a perturbation voxel map; Step 2: Perform sequential perturbation excitation, activate the steam control unit of each steam chamber in turn, obtain the tension response and thermographic response of adjacent and non-adjacent steam chambers, and construct the regional perturbation response matrix; Step 3: Discretize the control behavior of each steam chamber into several micro-intervention control units. Each micro-intervention control unit includes a control time segment, steam opening level and corresponding three-dimensional voxel unit index. Combine the perturbation voxel map and the regional perturbation response matrix to generate a set of candidate micro-intervention control units. Step 4: Generate a region control intent vector based on the current fabric running status; Step 5: Perform auction scheduling for each candidate micro-intervention control unit. Each region bids according to the control intent vector. The winning region is determined by the Vickrey auction mechanism, and the winning control map is constructed. Step 6: Perform strategy reparameter sampling based on the winning control map to generate a control strategy for controlling the behavior of the steam chamber; Step 7: When an abnormal state occurs on the fabric surface, based on the index mapping relationship of the abnormal area in the perturbation voxel map, backtrack the corresponding indexed control map and extract the sequence of intervention and control units that caused the abnormality; Step 8: Based on the intervention control unit sequence, adjust the regional control intent vector and the corresponding bidding order, re-execute the resource auction scheduling, and form a new winning bid control map; The auction scheduling of each candidate micro-intervention control unit, with each region bidding according to its control intent vector, and the winning region determined using the Vickrey auction mechanism, is specifically as follows: Each candidate micro-intervention control unit is set as an independent auction unit, and each auction unit contains a unique steam chamber number, control time period, steam opening level and three-dimensional voxel unit index. A Vickery auction execution instance is initialized for each auction unit. The Vickery auction execution instance is configured with a single round, closed-ended, and second-highest-bid rule. After the auction starts, all participating regions submit the bid value of the auction unit in a non-interactive state, and modification or withdrawal is prohibited. The bid value of the auction unit is generated by inputting the corresponding control intention vector into the regional bid calculation module. The regional bid calculation module combines the tension deviation, temperature change gradient, fabric defect ratio and disturbance hysteresis weight into a floating-point scalar value according to a preset weight, which serves as the region's bid input for the current candidate micro-intervention control unit. Within each auction unit, the bid values ​​of all auction units are sorted, the area with the highest bid value is determined as the winning bid area, the bid value of the area with the second highest bid value is used as the payment price of the winning bid area, and the unique assignment relationship between the winning bid area and the candidate micro-intervention control unit is recorded. The strategy re-parameter sampling extracts all control time periods from the winning bid control map, constructs a dynamic control time series index, uses each control time period as an independent sampling window, and initializes the control parameter template in each sampling window according to the configuration of the winning bid control unit. Within each sampling window, a strategy reparameter sampling process is performed, including: extracting the steam chamber number and corresponding spatial location of the winning control unit, and locking the current control area to be sampled; extracting the winning steam opening level as the initial value of the control intensity; based on the fabric disturbance voxel state within the current sampling window, inputting the disturbance voxel state features into the strategy adjustment model, adjusting the steam opening level and time offset to form the final action parameters, and generating a control strategy for controlling the behavior of the steam chamber. The strategy adjustment model adopts a parameter generator based on supervised training, with the disturbance voxel state features as input and the disturbance adjustment value of the control action parameters as output, which is used to disturb the initial value of the winning control intensity to achieve customized control. The control strategy consists of a steam chamber number, a control time period, an actual steam opening level, and a corresponding three-dimensional voxel index quadruple. This set of control instructions is sent to the execution end of the array-type steam heat setting equipment to drive the corresponding steam chambers to perform steam intervention behavior according to the bidding order and the final action parameters.

2. The method for optimizing an array-type steam thermal setting process based on machine learning according to claim 1, characterized in that, The construction of the perturbation voxel map specifically involves: The steam chambers along the fabric running path in the array-type steam heat setting equipment are divided into several regions. Each region is divided along the fabric running direction according to a set length, and the state of each region is recorded in the time dimension at a set sampling period to form a three-dimensional voxel unit with spatial position index and time index. The data fields in the three-dimensional voxel unit include the steam valve opening and closing status, steam opening value, tension sensor reading value, infrared thermal imaging temperature value, humidity sensor data, and fabric image texture feature value. All three-dimensional voxel units are organized according to spatial region numbering and chronological order to form a perturbation voxel map with a three-dimensional coordinate structure.

3. The method for optimizing an array-type steam thermal setting process based on machine learning according to claim 1, characterized in that, Step two specifically involves: The sequential perturbation excitation activates the steam control unit of each steam chamber in the array-type steam heat setting equipment in turn. During each activation process, only the steam valve of the current target steam chamber is opened, while the steam valves of the other chambers remain closed. The steam excitation lasts for a preset time. During each steam activation process, tension sensors placed at the inlet, center, and outlet of each steam chamber collect the fabric tension values ​​at multiple consecutive time points starting from the current activation moment, and calculate the tension changes in the corresponding tension channels of adjacent and non-adjacent steam chambers to form a multi-region tension response corresponding to the activation event. By using infrared thermal imaging devices installed above and on the side walls of each steam chamber, thermal image sequences of the fabric in each region are collected. The temperature gradient changes, hot spot diffusion range and maximum temperature rise location of each region before and after steam excitation are extracted to form thermal response corresponding to adjacent and non-adjacent steam chambers. Based on the tension response and thermographic response data of all regions obtained after each steam chamber excitation, a regional perturbation response matrix is ​​constructed by combining the indices of the excitation source steam chamber and the response target steam chamber. Each element of the regional perturbation response matrix corresponds to the perturbation effect of one steam chamber on another steam chamber region, which is used to quantify the thermal and moisture perturbation coupling relationship between each steam chamber in the array structure.

4. The method for optimizing an array-type steam thermal setting process based on machine learning according to claim 1, characterized in that, Step three specifically involves: The complete steam control cycle of each steam chamber is divided into multiple non-overlapping time segments, the duration of which is a preset control time interval; For each time segment, multiple steam opening levels are set, and the steam opening level is a discrete set of values ​​divided in increments of 10%. For each steam chamber, under the combination of time segments and steam opening level, a corresponding micro-intervention control unit is formed by combining the spatial position index of the steam chamber. Each micro-intervention control unit is bound to a unique control time segment, steam opening level, and three-dimensional voxel unit index in the perturbation voxel map. For all micro-intervention control units, based on the disturbance influence in the regional disturbance response matrix, the tension response and heatmap response generated by each micro-intervention control unit at the corresponding voxel are retrieved, and a disturbance propagation score is generated by weighted linear combination of values. All micro-intervention control units are sorted according to the disturbance propagation score, and those micro-intervention control units below the preset disturbance impact threshold are eliminated. They are then clustered and integrated according to the three-dimensional voxel unit index to form a candidate micro-intervention control unit set, which is used as the resource bidding object in the auction scheduling process.

5. The method for optimizing an array-type steam thermal setting process based on machine learning according to claim 1, characterized in that, Step four: Generating a region control intent vector based on the current fabric running state, specifically: For each steam chamber in the array-type steam heat setting equipment, a control intent vector for the corresponding area is generated based on the current operating status data of the fabric in that area. The current operating status data includes the difference between the real-time tension value collected by the regional tension sensor and the historical tension baseline value, the fabric temperature change gradient value collected by the infrared thermal imaging device, the proportion of fabric defects detected by the fabric image recognition module, and the perturbation hysteresis weight of the region in the perturbation voxel map. After standardizing the current operating status data, a regional control intent vector is constructed by combining the data. The control intent vector is a set of real-valued vectors, including tension deviation, temperature change gradient, fabric defect ratio, and disturbance hysteresis weight.

6. The method for optimizing an array-type steam thermal setting process based on machine learning according to claim 1, characterized in that, The construction of the winning bid control map specifically involves: All candidate micro-intervention control units in the winning areas are aggregated into the winning control unit set, sorted according to time index and voxel spatial index, and a structured winning control map is constructed. The winning control map is a two-dimensional mapping structure, with the time dimension corresponding to the control time period and the spatial dimension corresponding to the three-dimensional voxel position. Each cell of the winning control map includes the steam chamber number, steam opening value and winning area number.

7. The method for optimizing an array-type steam thermal setting process based on machine learning according to claim 1, characterized in that, Step seven specifically involves: During the operation of the array-type steam heat setting equipment, abnormal conditions of the fabric surface are detected in real time. When an abnormality occurs in the fabric surface, the specific spatial location of the abnormality area is determined by infrared thermal imaging and fabric surface image recognition. Based on the spatial location of the abnormal area on the fabric surface and the time index of the time when the abnormality occurred, find the three-dimensional voxel unit index in the perturbation voxel map that corresponds to the abnormal area in terms of spatial location and time. Starting from the three-dimensional voxel unit index, the process traces back along the time axis to a preset backtracking period before the time of the anomaly occurrence, and searches step by step for candidate micro-intervention control units that have been executed on the abnormal region in the winning control map. The retrieved candidate micro-intervention control units are sorted in order of execution time from most recent to oldest, forming a sequence of intervention control units that cause abnormal fabric conditions.

8. The method for optimizing an array-type steam thermal setting process based on machine learning according to claim 1, characterized in that, Step eight specifically involves: Based on the intervention control unit sequence, determine the steam chamber number and spatial location corresponding to each candidate micro-intervention control unit in the intervention control unit sequence, and count the number of interventions and the degree of contribution to the anomaly in the area corresponding to each steam chamber before the anomaly occurred. Based on the number of interventions and the degree of abnormal contribution, the tension deviation, temperature change gradient, fabric defect ratio and disturbance hysteresis weight of the corresponding area in the original regional control intention vector are corrected according to the preset weights. The bid value for each region for each candidate micro-intervention control unit is recalculated based on the revised regional control intent vector, and the Vickrey auction mechanism is re-executed based on the updated regional bid value to determine the new winning region. All candidate micro-intervention control units in the newly selected regions are aggregated into the set of selected control units, and sorted according to the time index and voxel spatial index to form a new selected control map.

Images