A dynamic optimization and adjustment system for process parameters in titanium alloy rolling.

Abstract

This invention relates to the field of titanium alloy rolling process parameter optimization, specifically disclosing a dynamic optimization and adjustment system for titanium alloy rolling process parameters. This invention significantly improves the detection accuracy and real-time performance of titanium alloy microstructure and defects through an innovative multi-feature fusion image recognition algorithm. Furthermore, it constructs an adaptive intelligent correction system based on reinforcement learning, achieving dynamic and precise control of rolling process parameters. This effectively reduces defect rate and energy consumption while improving product qualification rate and microstructure performance compliance rate. Through automated collaborative optimization throughout the entire process, it significantly reduces manual intervention, thereby improving overall production efficiency and system stability.

Description

Technical Field

[0001] This invention relates to the field of parameter optimization in titanium alloy rolling process, and specifically to a dynamic optimization and adjustment system for process parameters in titanium alloy rolling process. Background Technology

[0002] In titanium alloy manufacturing processes, especially in the rolling process, traditional production methods rely on manual sampling, empirical adjustments, and human intervention. These methods face the following technical challenges:

[0003] 1) Unevenness in manual sampling and unstable quality control:

[0004] Traditional manual sampling is susceptible to human error, resulting in uneven sampling, significant data errors, and difficulty in achieving high-precision quality control. Due to the unavoidable nature of manual operation, sampling may not fully represent all stages of the production process, leading to inconsistencies in product quality.

[0005] 2) Low timeliness of visual inspection methods:

[0006] Currently, the detection of surface defects in titanium alloys largely relies on manual or traditional visual inspection methods. These methods are not only inefficient but also lack high precision and real-time performance, and are prone to missed or false detections. In particular, minute defects such as surface cracks, bubbles, and pores are often difficult to detect using traditional methods, affecting product quality.

[0007] 3) Delayed or inaccurate adjustment of process parameters during production:

[0008] Traditional methods of adjusting process parameters typically rely on experience and manual judgment, and are inherently lagging, failing to provide real-time adjustments based on defects that arise during production. This not only leads to inconsistent quality in titanium alloy products but can also result in unnecessary waste and increased production costs.

[0009] 4) The problem of scarce training data and imbalanced data:

[0010] In the task of detecting and classifying defects in titanium alloys, the scarcity of training data and class imbalance are common challenges. A lack of sufficient defect images, especially for rarer defects, leads to poor model training performance and weak generalization ability.

[0011] 5) Optimization of inter-task collaboration and multi-task learning:

[0012] In traditional deep learning models, image recognition and process parameter adjustment are usually independent tasks, but these two are closely related in actual production. How to effectively combine image recognition and process parameter optimization tasks to improve the overall system performance is a technical challenge that urgently needs to be solved. Summary of the Invention

[0013] To address the technical problems mentioned above, this invention combines deep learning, automatic sampling, image recognition technology, and an intelligent feedback control system. Through real-time defect detection and dynamic process adjustment, it significantly improves the accuracy, production efficiency, and product quality of the titanium alloy production process. The following is a detailed description of the technical solution of this invention.

[0014] To achieve the above objectives, the present invention provides a dynamic optimization and adjustment system for process parameters in titanium alloy rolling, comprising:

[0015] The dynamic adaptive automatic sampling module is deployed at multiple stations on the rolling production line. It is used to dynamically adjust the sampling position and sampling amount according to the shape of the rolled piece and the rolling pass, automatically collect the rolled piece sample, and simultaneously record the real-time process parameters at the sampling time.

[0016] The multimodal image recognition and analysis module is connected to the automatic sampling module and is used to acquire images of the collected samples and process and analyze the acquired image data to obtain structural feature data and defect detection data that characterize the internal structure of the rolled piece.

[0017] The intelligent correction decision module is connected to the multimodal image recognition and analysis module. It is used to receive the tissue feature data, defect detection data and real-time process parameters, dynamically calculate the process parameters to be adjusted and their adjustment amount based on the preset optimization model, and generate control commands.

[0018] The execution control module connects the intelligent correction decision module to the rolling production line and is used to adjust the rolling process parameters in real time according to the control instructions, forming a closed-loop control system of "sampling-identification-decision-execution".

[0019] Preferably, the dynamically adaptable automatic sampling module includes:

[0020] The station adaptation unit is used to select a sampling device that matches the current rolling condition based on the roughing, intermediate, or finishing rolling station where the workpiece is located.

[0021] The sampling control unit, connected to the workstation adaptation unit, is used to determine the specific sampling location and sampling quantity based on the shape of the rolled piece and the current rolling pass, and to drive the sampling device to perform the sampling action.

[0022] The data association and storage unit is used to generate a unique code for each sample and bind the process parameters such as rolling temperature, rolling force, reduction rate, and rolling speed at the sampling time, as well as equipment status data, and store them in the associated database.

[0023] Preferably, the multimodal image recognition and analysis module includes:

[0024] Image acquisition unit, used to acquire the original image of the sample;

[0025] An image preprocessing unit, connected to the image acquisition unit, is used to perform layered adaptive noise reduction processing on the original image to suppress interference from oxide layers, oil stains, and edge noise.

[0026] The tissue segmentation and feature extraction unit, connected to the image preprocessing unit, is used to perform multi-feature fusion segmentation on the denoised image, identify and segment the α-phase region and β-phase region in the image, and calculate the tissue quantification features of α-phase volume fraction, α-phase average grain size and β-phase distribution uniformity.

[0027] The defect identification unit, connected to the tissue segmentation and feature extraction unit, is used to identify cracks, inclusions, and oxide layer defects based on the segmented image, and output the type, size, and distribution data of the defects.

[0028] Preferably, the tissue segmentation and feature extraction unit includes:

[0029] Connected component extraction subunit is used to extract multiple connected components from the denoised image as candidate grain regions;

[0030] The multi-feature calculation subunit is used to calculate the pixel features, shape features, and spatial features for each connected component.

[0031] The feature fusion and classification subunit, connected to the multi-feature calculation subunit, is used to perform weighted fusion of pixel features, shape features and spatial features of each connected region to obtain fused feature values, and to determine whether the connected region belongs to the α phase or the β phase according to a preset threshold or neighborhood voting rule.

[0032] Preferably, the intelligent correction decision module includes:

[0033] The state construction unit is used to receive tissue feature data and defect detection data output by the multimodal image recognition and analysis module, and combine them with real-time process parameters to construct a state vector representing the current production state.

[0034] The reinforcement learning strategy unit has a pre-trained reinforcement learning model built in. The reinforcement learning strategy unit takes the state vector as input and outputs an adjustment action amount for at least one of the process parameters, namely rolling temperature, rolling force, reduction rate and rolling speed.

[0035] The objective function calculation unit is used to calculate the comprehensive objective function value based on the state vector, and the reinforcement learning model aims to minimize the objective function value for training and optimization.

[0036] Preferably, the system further includes:

[0037] An emergency correction unit connects the defect identification unit and the intelligent correction decision module. When the severity of the identified defect exceeds a preset emergency threshold, an emergency response is triggered.

[0038] The emergency correction unit is configured to: invoke a preset emergency parameter adjustment command to immediately reduce the rolling speed, and access the historical defect-parameter association database to match and apply the corresponding emergency process parameter combination; if the defect is not eliminated after adjustment, a shutdown alarm signal is generated.

[0039] Preferably, the dynamic adaptive automatic sampling module, the multimodal image recognition and analysis module, the intelligent correction decision module, and the execution control module achieve real-time data interaction and collaboration through industrial Ethernet or industrial IoT to form an integrated closed-loop control architecture.

[0040] Preferably, the dynamic adaptive automatic sampling module is equipped with sampling devices with different structures at the roughing, intermediate, and finishing rolling stations, including a laser positioning hydraulic clamping sampling mechanism for the roughing station, a vacuum adsorption micro-cutting sampling mechanism for the intermediate rolling station, and a flexible clamping ultrasonic cutting sampling mechanism for the finishing rolling station.

[0041] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0042] (1) Image recognition algorithm innovation significantly improves accuracy and real-time performance: Through the hierarchical adaptive noise reduction algorithm, the image signal-to-noise ratio is improved by more than 75%; the multi-feature fusion segmentation algorithm improves the α / β phase recognition accuracy to more than 96%, the grain size measurement error is reduced to ±0.5μm, and the processing time of a single image is shortened to 0.3s, meeting the requirements of hot rolling online inspection.

[0043] (2) The intelligent correction system achieves adaptive control: Based on the dynamic control strategy of reinforcement learning, the pass rate of rolled parts is increased from 85.2% to 96.7%, and the α phase volume fraction compliance rate is increased by 18.5%; the emergency correction mechanism has a response time of <25s, the defect rate is reduced by 80.9%, and batches of non-conforming products are avoided.

[0044] (3) Full-process collaborative optimization improves production efficiency: The collaboration of automatic sampling, image recognition, and intelligent correction reduces manual intervention time by more than 30%, and the annual capacity of the TC4 hot rolling line increases by 12% (from 10,000 tons to 11,200 tons), while reducing energy consumption by 8% (due to parameter optimization to avoid overheating / rolling).

[0045] This invention solves the technical pain points of delayed microstructure identification and blind adjustment of process parameters in the hot rolling process of TC4 titanium alloy by innovating image recognition algorithms and precisely controlling the intelligent correction system, providing key technical support for the stable production of high-end titanium alloy rolled products. Attached Figure Description

[0046] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0047] Figure 1 This is a flowchart of the system of the present invention. Detailed Implementation

[0048] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0049] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0050] Example 1:

[0051] This embodiment provides a dynamic optimization and adjustment system for process parameters in titanium alloy rolling, including: a dynamic adaptive automatic sampling module, a multimodal image recognition and analysis module, an intelligent correction decision module, and an execution control module.

[0052] The parameters and workflow of each module of the present invention will be described in detail below with reference to this embodiment:

[0053] (a) Design of dynamic adaptive automatic sampling module.

[0054] 1. Sampling device deployment and operating condition adaptation:

[0055] Differential sampling devices were deployed at key workstations on the TC4 hot rolling line using workstation adaptation units.

[0056] Roughing station (rolling temperature 950-1000℃, reduction rate 30%-40%): For thick plate rolls (thickness 20-50mm), a "laser contour positioning + hydraulic clamping sampling mechanism" is adopted. The laser sensor (resolution 0.01mm) measures the cross-sectional dimensions of the roll in real time. According to the preset sampling interval of the rolling pass (once every 5 passes), a block sample of 15mm×15mm×roll thickness is cut out. After sampling, the cross-sectional image of the sample is captured by an industrial camera (12 million pixels, frame rate 30fps). Gray difference analysis (refer to the noise analysis logic of CN119399567A) is used to verify whether there is an oxide layer in the sample (gray value < 50 is judged as oxidation). If the oxidation area ratio is > 5%, the sample is resampled.

[0057] Intermediate rolling station (rolling temperature 900-950℃, reduction rate 20%-30%): For medium-thickness rolled parts (thickness 10-20mm), a "vacuum adsorption + micro-cutting sampling mechanism" is adopted, and the sampling amount is controlled at 1-2g to avoid affecting the subsequent forming of the rolled parts; the cutting depth is controlled by a displacement sensor (accuracy 0.001mm) to ensure that the sampling section is flat (roughness Ra≤0.8μm), providing clear samples for subsequent image recognition.

[0058] Finishing mill station (rolling temperature 850-900℃, reduction rate 10%-20%): For thin-walled / irregularly shaped rolled parts (thickness 3-10mm), a "flexible clamping + ultrasonic cutting" sampling method is adopted. The ultrasonic frequency is set to 20-30kHz, and the cutting force is <50N to avoid deformation of the rolled parts. After sampling, eddy current testing (refer to the defect detection logic of CN120655593A) is used to quickly check for micro-cracks on the sample surface (sensitivity ≥0.1mm depth). Unqualified samples trigger the resampling process.

[0059] 2. Correlation between sampling data and process parameters:

[0060] The sampling control unit determines the specific sampling location and sampling quantity based on the shape of the rolled piece and the current rolling pass, and drives the sampling device to perform the sampling action.

[0061] The data association and storage unit assigns a unique code to each sample and binds the process parameters (rolling temperature T, rolling force F, reduction rate ε, rolling speed v) and equipment status (motor current, roll gap value) at the time of sampling through the Industrial Internet of Things (IIoT). The data is then stored in a distributed database (MySQL + Redis) to form a "process parameter-sample image-organic feature" associated dataset, providing data support for image recognition algorithm training and intelligent correction.

[0062] (II) Multimodal Image Recognition and Analysis Module.

[0063] First, the original image of the sample is acquired using the image acquisition unit. The process includes:

[0064] 1. First-level processing: Layered adaptive noise reduction algorithm (for hot rolling image interference problems). The image preprocessing unit performs layered adaptive noise reduction on the original image to suppress interference from oxide layer, oil stains and edge noise.

[0065] The TC4 hot-rolled sample images exhibit three typical types of interference: "local dark spots" caused by the oxide layer, "uneven grayscale" caused by rolling oil contamination, and "edge noise" caused by sampling and cutting. The process for designing a layered noise reduction strategy is as follows:

[0066] Step 1: Calculate the global noise index.

[0067] The sample image is converted to a grayscale image (grayscale range 0-255). A 3×3 neighborhood grayscale co-occurrence matrix is ​​constructed for each pixel. The homogeneity and contrast of the matrix are calculated. Smooth regions (homogeneity > 0.8, contrast < 20) are selected through fuzzy comprehensive evaluation. The grayscale change coefficient of the smooth regions is calculated to obtain the global noise index.

[0068] ,

[0069] in, S is the set of pixels in the smooth region, N is the number of pixels, and ds is the global noise index, which reflects the overall noise intensity of the image.

[0070] Step 2: Calculate the pixel-level noise difference coefficient.

[0071] For each pixel, construct a 5×5 neighborhood centered on it, and calculate the difference between the grayscale value within the neighborhood and the global noise index:

[0072] ,

[0073] in, The grayscale mean of the 5×5 neighborhood. For smoothing terms, This is the noise difference coefficient (value range 0-1). A value greater than 0.6 is considered a high-noise pixel (such as oxide spots or edge burrs).

[0074] Step 3: Adaptive filtering.

[0075] For high-noise pixels, a "Gaussian-median hybrid filter" is used: first, a 3×3 Gaussian filter (σ=1.2) is used to smooth the gradual noise caused by oil stains, and then a 5×5 median filter is used to remove the pulse noise caused by cutting burrs; for low-noise pixels ( For samples with a resolution ≤0.6, a 2×2 mean filter is used to preserve grain boundary details. After layered noise reduction, the image signal-to-noise ratio (SNR) is improved from 20dB to over 35dB, laying the foundation for subsequent segmentation.

[0076] (2) Secondary processing: The multi-feature fusion α / β phase segmentation algorithm uses the tissue segmentation and feature extraction unit to perform multi-feature fusion segmentation on the denoised image, identify and segment the α phase region and β phase region in the image, and calculate the tissue quantification features of α phase volume fraction, α phase average grain size and β phase distribution uniformity.

[0077] Traditional segmentation algorithms rely solely on grayscale values ​​(e.g., α-phase grayscale value 180-220, β-phase grayscale value 80-120), failing to distinguish between "low grayscale α-phase (e.g., α-phase grayscale reduction caused by low-temperature rolling)" and "high grayscale β-phase (e.g., β-phase grayscale increase caused by localized oxidation)." This embodiment integrates "pixel features, shape features, and spatial features." This embodiment designs an improved multi-feature fusion segmentation algorithm, the process of which is as follows:

[0078] Step 1: Connected component extraction.

[0079] Initial connected regions were extracted using the region growing method (seed points were selected from pixels with a grayscale value of 150-200, and the growth threshold was ±15). Each connected region corresponded to a potential α-phase or β-phase grain, labeled C. k (k=1,2,...,n, where n is the total number of connected components).

[0080] Step 2: Multi-feature calculation.

[0081] Pixel features (distinguishing α / β phase grayscale characteristics): pixel numerical distribution g k Connected component C k The grayscale mean of all pixels within the range, α phase g k Typically >150, β phase g k Usually <130;

[0082] Pixel disorder distribution degree S k Connected component C k The variance of the internal gray value, the α phase is stable due to the crystal structure, S k <30, β phase due to uneven composition, S k >50;

[0083] Pixel feature G k =g k ×exp(-S k / 100), G k >120 is determined to be a candidate for α phase, G k <80 is considered a candidate for phase β, 80≤G k If the value is ≤120, proceed to subsequent feature verification.

[0084] ① Shape characteristics (distinguishing between α / β phase morphological differences):

[0085] During the hot rolling process of TC4, the α phase is mostly equiaxed or short strips (high roundness), while the β phase is mostly long strips or irregular shapes (high perimeter-to-area ratio).

[0086] Circularity R k =L k 2 / (4πA k ), A k Let L be the area of ​​the connected region. k Let R be the perimeter, and α phase be R. k >0.6, β phase R k <0.4;

[0087] Perimeter to Area Ratio P k =A k / L k α phase P k <0.1, β phase P k >0.2;

[0088] Shape feature S shape,k =R k ×exp(-P k ), S shape,k >0.3 indicates a candidate for α phase, S shape,k <0.1 indicates a candidate for phase β.

[0089] ② Spatial characteristics (distinguishing α / β phase distribution patterns):

[0090] The α phase tends to aggregate during hot rolling (especially when rolling at 900-950℃), while the β phase is mostly dispersed at the α phase grain boundaries: local distribution quantity N k : Using connected component C k Centered on the center point, the number of other connected components within a 10×10 range, α phase N k >5, β phase N k <3; Local location distance D space,k Connected component C k The sum of the distances between the center point and the center points of the surrounding connected domains, α phase D space,k <20 pixels, β phase D space,k >30 pixels;

[0091] Spatial features S space,k =N k ×exp(-D space,k / 50), S space,k >1.5 is judged as an α phase candidate, S space,k <0.8 indicates a candidate for β phase.

[0092] Step 3: Feature fusion and classification.

[0093] Constructing fusion features F k =w1G k +w2S shape,k +w3S space,k The weights are w1=0.5 (pixel feature dominant), w2=0.3 (shape feature auxiliary), and w3=0.2 (spatial feature verification). The threshold F is determined using the training set (1000 sets of labeled TC4 hot-rolled images). k >80 is determined to be the α phase, F k <50 is determined to be phase β, 50≤F k ≤80% is determined by neighborhood voting (if more than 80% of the surrounding connected components are α phase, it is determined to be α phase; otherwise, it is β phase). Testing showed that the algorithm achieves an α / β phase segmentation accuracy of 96.3%, an improvement of 11.2 percentage points compared to the traditional grayscale thresholding method.

[0094] (3) Three-level processing: including tissue quantification and defect detection. The defect identification unit identifies cracks, inclusions and oxide layer defects based on the segmented image and outputs the type, size and distribution data of the defects.

[0095] Among them, tissue quantification analysis involves quantifying tissue features in the segmented image, specifically including:

[0096] ①A α phase volume fraction V α The calculation formula is:

[0097] ,

[0098] For the hot rolling process of TC4 titanium alloy, V α The target range is 30%–40%.

[0099] ②Average grain size of α phase d α The calculation formula is:

[0100] ,

[0101] For the hot rolling process of TC4 titanium alloy, d α The target range is 5–10 μm.

[0102] ③ Calculate the uniformity of β phase distribution. The evaluation is carried out by statistically analyzing the ratio of the area of ​​the largest β phase aggregation zone to the total area of ​​β phase. The target is that the ratio is less than 20%, that is, to avoid excessive local aggregation of β phase.

[0103] Defect detection involves identifying defects in the segmented image, specifically including:

[0104] ① Crack detection: Linear edges are extracted using Canny edge detection (threshold 50-150). If the edge length is greater than 0.5mm and the width is less than 0.1mm, it is judged as a crack.

[0105] ② Impurity detection: Identify grayscale values ​​< 50 and areas > 0.05 mm² 2 The connected domains are identified as inclusions (such as Fe, Si, and other impurities).

[0106] ③Oxide layer detection: Identify areas with gray values ​​< 80 and continuous distribution at the edge of the image, determine them as oxide layers, and calculate the oxide layer thickness tox.

[0107] Image recognition algorithm performance verification:

[0108] The algorithm performance was verified using 100 sets of TC4 hot-rolled sample images (covering the roughing, intermediate, and finishing stages), as shown in Table 1:

[0109] Table 1:

[0110] .

[0111] (III) Intelligent Correction Decision Module.

[0112] a. Correction Target and State Space Definition. The state construction unit receives the tissue feature data and defect detection data output by the multimodal image recognition and analysis module, and combines them with real-time process parameters to construct a state vector representing the current production state.

[0113] Correction Objectives: With "optimal organizational quality + highest process stability" as the dual objectives, the objective function is constructed as follows:

[0114] ,

[0115] Among them, V α,target =35% (optimal α-phase volume fraction for TC4 hot rolling), d α,target =7μm (optimal grain size); F is the actual rolling force; F max To avoid overloading the equipment, use the maximum rolling force (e.g., 3000kN); D defect For defect density (number of cracks / inclusions / image area), there are <5 targets; weights ω1=0.4, ω2=0.3, ω3=0.2, ω4=0.1 (adjusted according to the requirements of TC4 for aviation).

[0116] State space S: S=[T,F,ε,v,V] α ,d α D defect ],

[0117] Among them, the rolling temperature T (850-1000℃), rolling force F (500-3000kN), reduction rate ε (10%-40%), and rolling speed v (0.5-2m / s); V α d α D defect The image is output in real time by the multimodal image recognition and analysis module.

[0118] Action space A: A=[ΔT,ΔF,Δε,Δv], adjustment range: ΔT: ±5-15℃ (roughing ±15℃, finishing ±5℃); ΔF: ±50-200kN (avoid excessive fluctuations in rolling force that could cause equipment impact); Δε: ±0.5-2% (±0.5% in the finishing stage to ensure dimensional accuracy); Δv: ±0.1-0.5m / s (speed adjustment and reduction rate work together to avoid slippage of the rolled piece).

[0119] b. Training the reinforcement learning model.

[0120] Reward function design: If J < 0.1 (excellent microstructure, stable process), a reward of R = +10 is given; if 0.1 ≤ J < 0.3 (acceptable microstructure, stable process), a reward of R = +5 is given; if J > 0.3 (unacceptable microstructure or overloaded process), a penalty of R = -10 is given. Here, J is the objective function value, and the reward R guides the model towards a "low J" direction. If J decreases by more than 10% after adjustment, an additional reward of +5 is given.

[0121] Training process:

[0122] 1) Offline training: The proximal policy optimization (PPO) algorithm is trained using 100,000 sets of historical data ("process parameters-organization characteristics" related data) to construct a policy network (input state S, output action probability distribution) and a value network (evaluating the long-term reward of state S); the training rounds are 5000 rounds, the learning rate is initially 0.001, and it decays to 80% of the previous round every 1000 rounds to avoid overfitting.

[0123] 2) Online fine-tuning: Deploy the trained model to the hot rolling line, and update the network parameters (learning rate 0.0001) every 10 rolling passes based on the actual image recognition results and reward feedback, so that the model can adapt to real-time changes in working conditions (such as fluctuations in raw material composition and rolling force deviations caused by equipment aging).

[0124] c. Dynamic control strategy for intelligent deviation correction. This embodiment uses a reinforcement learning strategy unit, which has a pre-trained reinforcement learning model built in. The reinforcement learning strategy unit takes the state vector as input and outputs an adjustment action for at least one process parameter among rolling temperature, rolling force, reduction rate, and rolling speed. Specifically:

[0125] (1) Layered control logic: Rough rolling stage (rapid microstructure forming): If image recognition shows V α <30% (excessive β phase), the model prioritizes adjusting the rolling temperature (increasing by 10-15℃) and simultaneously increasing the reduction rate (Δε=+1%) to promote the transformation of the β phase to the α phase; if the number of cracks is >10 / mm 2 If obvious cracks are present, reduce the rolling speed (Δv=-0.3m / s) to reduce stress concentration.

[0126] (2) Intermediate rolling stage (microstructure homogenization): If the grain size is greater than 100μm (coarse grains), the rolling force (ΔF=+100kN) and reduction rate (Δε=+1.5%) are adjusted in the model to refine the grains through dynamic recrystallization; if the β phase distribution uniformity index is <0.8 (β phase aggregation), the rolling temperature is finely adjusted (reduced by 3-5℃) to improve the β phase distribution.

[0127] (3) Finishing stage (dual control of dimensions and structure): If the thickness deviation of the rolled piece is >0.1mm, the reduction rate (Δε=±0.5%) should be adjusted first; if Vα deviates from the target (35%±2%), the rolling temperature should be finely adjusted (±2-4℃) to avoid dimensional deviation due to excessive parameter adjustment.

[0128] Finally, the objective function calculation unit calculates the comprehensive objective function value based on the state vector. The reinforcement learning model aims to minimize this objective function value for training and optimization.

[0129] This embodiment also includes an emergency correction unit, which connects the defect identification unit and the intelligent correction decision module. When the severity of the identified defect exceeds a preset emergency threshold, an emergency response is triggered.

[0130] Emergency correction mechanism: When the image identifies a "fatal defect" (such as crack length > 2mm), an emergency strategy is triggered: immediately reduce the rolling speed to below 0.5m / s to reduce defect propagation; call the "defect-parameter" association database (built offline) to match the historical best repair parameters (such as "when the crack length is 2-3mm, increase the rolling temperature by 15℃ and reduce the reduction rate by 2%)"; if the defect is not eliminated after 3 consecutive adjustments, the machine will automatically stop and push alarm information to the operator, generating a "defect cause-adjustment suggestion" report (such as "the crack is caused by the rolling temperature fluctuation, it is recommended to check the heating furnace temperature control system").

[0131] d. Performance verification of the intelligent correction system:

[0132] A three-month verification was conducted on a TC4 titanium alloy hot rolling production line (annual capacity of 10,000 tons). Key indicators before and after optimization were compared, as shown in Table 2.

[0133] Table 2:

[0134] .

[0135] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

[0136] Example 2:

[0137] The following will describe in detail, with reference to this embodiment, how the present invention solves the technical problems in practical work, and the process is as follows: Figure 1 As shown. This embodiment uses TC4 titanium alloy thick plate hot-rolled (20mm thick, for aerospace use).

[0138] First, automatic sampling is performed using a dynamically adaptable autosampling module, as detailed below:

[0139] (1) Rough rolling station (rolling temperature 980℃, reduction rate 35%): Sampling is done once every 5 passes, and a 20mm×15mm×20mm sample is cut. The cross-sectional image is taken by an industrial camera to verify that the oxide layer area is <3%, which is qualified.

[0140] (2) Intermediate rolling station (rolling temperature 930℃, reduction rate 25%): vacuum adsorption sampling is adopted, sampling is 1.2g, the cutting depth is controlled by displacement sensor to be 0.5mm, and the cross-sectional roughness Ra=0.6μm;

[0141] (3) Finishing rolling station (rolling temperature 880℃, reduction rate 15%): ultrasonic cutting sampling, sample amount 0.8g, eddy current testing did not detect cracks, qualified.

[0142] Next, image recognition is performed using the multimodal image recognition and analysis module. The specific process is as follows:

[0143] Layered noise reduction: After processing, the SNR of the rough rolling sample image increased from 22dB to 37dB, and the oxide spot removal rate was 98%; Multi-feature fusion segmentation: α-connected connected domain G k =135, S shape,k =0.42, S space,k =1.8, F k =0.5×135+0.3×0.42+0.2×1.8=68.1, determined to be phase α; β-connected domain F k =42, determined to be β phase; tissue quantification: V α =32% (close to the target of 35%), d α =8.2μm (close to the target 7μm), 0 (no cracks / inclusions).

[0144] The intelligent correction decision module performs intelligent correction, and the process is as follows:

[0145] state space ;

[0146] Objective function:

[0147] .

[0148] Finally, the execution control module outputs the action: The adjusted rolling temperature is 885℃, and the reduction rate is 15.5%.

[0149] Resampling and identification:

[0150] , This achieves the optimal state. The overall process is as follows: Figure 1 As shown.

[0151] Example 3:

[0152] The following will describe in detail, with reference to this embodiment, how the present invention solves the technical problems in practical work. This embodiment uses hot-rolled TC4 titanium alloy thin-walled tubes (outer diameter 10mm, wall thickness 1.5mm, for medical devices). The working modules are the same as in Embodiment 2, and will not be repeated here.

[0153] 1. Automatic sampling:

[0154] The finishing mill station uses "flexible clamping + ultrasonic cutting" with a sampling amount of 0.5g to avoid deformation of thin-walled tubes; an industrial camera is used to photograph the sample cross-section to verify that there is no oxide layer and that it is qualified.

[0155] 2. Image recognition:

[0156] Layered noise reduction: Due to strong edge effect noise in the thin-walled tube sample image, the SNR was improved from 18dB to 33dB after processing; Multi-feature fusion segmentation: Local regions F were identified. k =65(50≤F k ≤80), determined as α phase by neighborhood voting (90% of the surrounding connected domains are α phase); Defect detection: 6 microcracks with a length of 0.8 mm were found (slightly exceeding the target).

[0157] 3. Intelligent correction:

[0158] state space .

[0159] Objective function:

[0160] (Excellent organization, but with slightly more flaws); Action output: After adjusting the rolling speed to 0.8 m / s, stress concentration was reduced; a second sampling and identification showed that the crack length had been shortened to 0.3 mm, achieving the target.

Claims

1. A system for dynamic optimization adjustment of process parameters of a titanium alloy rolling process, characterized in that, The application relates to a dynamic adaptive automatic sampling module, a multi-modal image recognition and analysis module, an intelligent correction decision module and an execution control module. The dynamic adaptive automatic sampling module is arranged at multiple stations of a rolling production line and is used for dynamically adjusting a sampling position and a sampling amount according to a rolling piece shape and a rolling pass, automatically collecting a rolling piece sample and synchronously recording real-time process parameters at a sampling moment. The multi-modal image recognition and analysis module is connected with the automatic sampling module and is used for collecting images of the collected sample and processing and analyzing collected image data to obtain organization characteristic data and defect detection data representing an internal organization state of the rolling piece. The intelligent correction decision module is connected with the multi-modal image recognition and analysis module and is used for receiving the organization characteristic data, the defect detection data and the real-time process parameters, dynamically calculating process parameters to be adjusted and adjustment amounts of the process parameters based on a preset optimization model and generating a control instruction. The intelligent correction decision module comprises a state construction unit, a reinforcement learning strategy unit and a target function calculation unit. The state construction unit is used for receiving the organization characteristic data and the defect detection data output by the multi-modal image recognition and analysis module and combining the real-time process parameters to construct a state vector representing a current production state. The reinforcement learning strategy unit is internally provided with a pre-trained reinforcement learning model, takes the state vector as input and outputs an adjustment action amount of at least one process parameter of rolling temperature, rolling force, rolling reduction and rolling speed. The target function calculation unit is used for calculating a comprehensive target function value according to the state vector, and the reinforcement learning model takes minimizing the target function value as a training and optimization target. The execution control module is connected with the intelligent correction decision module and the rolling production line and is used for adjusting the rolling process parameters in real time according to the control instruction to form a closed-loop control system of ''sampling-recognition-decision-execution''.

2. The system of claim 1, wherein, The dynamic adaptive automatic sampling module comprises a station adaptation unit, a sampling control unit and a data association and storage unit. The station adaptation unit is used for selecting a sampling device matched with a current rolling working condition according to a rough rolling station, a medium rolling station or a finish rolling station where a rolling piece is located. The sampling control unit is connected with the station adaptation unit, is used for determining a specific sampling position and a sampling amount according to a rolling piece shape and a current rolling pass, and drives the sampling device to perform a sampling action. The data association and storage unit is used for generating a unique code for each sample, binding process parameters of rolling temperature, rolling force, rolling reduction and rolling speed and equipment state data at a sampling moment and storing the process parameters and the equipment state data into an association database.

3. The system of claim 1, wherein, The multi-modal image recognition and analysis module comprises an image acquisition unit, an image preprocessing unit, an organization segmentation and feature extraction unit and a defect recognition unit. The image acquisition unit is used for acquiring original images of the sample. The image preprocessing unit is connected with the image acquisition unit, is used for performing layered self-adaptive noise reduction processing on the original images to suppress interference of an oxidation layer, oil stains and edge noise. The organization segmentation and feature extraction unit is connected with the image preprocessing unit, is used for performing multi-feature fusion segmentation on the noise-reduced images, identifying and segmenting alpha phase regions and beta phase regions in the images and calculating organization quantitative features of an alpha phase volume fraction, an alpha phase average grain size and beta phase distribution uniformity. The defect recognition unit is connected with the organization segmentation and feature extraction unit, is used for identifying cracks, inclusions and oxidation layer defects based on the segmented images and outputting type, size and distribution data of the defects.

4. The system of claim 3, wherein, The tissue segmentation and feature extraction unit comprises: A connected domain extraction subunit configured to extract a plurality of connected domains from the denoised image as grain candidate regions; A multi-feature calculation subunit configured to calculate pixel features, shape features and spatial features of each connected domain respectively; A feature fusion and classification subunit connected to the multi-feature calculation subunit, configured to fuse the pixel features, shape features and spatial features of each connected domain by weighting to obtain a fusion feature value, and determine whether the connected domain belongs to alpha phase or beta phase according to a preset threshold or a neighborhood voting rule.

5. The system for dynamic optimization adjustment of process parameters of titanium alloy rolling process according to claim 3, characterized in that, The system further comprises: An emergency correction unit connected to the defect identification unit and the intelligent correction decision module, configured to trigger an emergency response when the identified defect severity exceeds a preset emergency threshold; The emergency correction unit is configured to call a preset emergency parameter adjustment instruction to immediately reduce the rolling speed, and access a historical defect-parameter association database to match and apply a corresponding emergency process parameter combination; if the defect is not eliminated after adjustment, a shutdown alarm signal is generated.

6. The system of any one of claims 1 to 5, wherein, The dynamic adaptive automatic sampling module, the multi-modal image recognition and analysis module, the intelligent correction decision module and the execution control module realize real-time data interaction and cooperation through industrial Ethernet or industrial Internet of Things, forming an integrated closed-loop control architecture.

7. The system for dynamic optimization and adjustment of process parameters for titanium alloy rolling process according to claim 1, wherein, The dynamic adaptive automatic sampling module is respectively provided with sampling devices with different structures at rough rolling, intermediate rolling and finishing rolling stations, including a laser positioning hydraulic clamping sampling mechanism for the rough rolling station, a vacuum adsorption micro-cutting sampling mechanism for the intermediate rolling station, and a flexible clamping ultrasonic cutting sampling mechanism for the finishing rolling station.

Images