Method and system for online regulation of quality of enzymatic processing of hides for non-leather applications

Abstract

The application discloses a method and system for online regulation and control of the quality of enzyme processing of animal hides for non-leather use, and the core lies in establishing a closed-loop control system of "perception-prediction-decision-execution-evolution". The method comprises the following steps: firstly, through image acquisition and processing in a processing device, pixel-level semantic segmentation and target detection are realized by using a multi-task convolutional neural network, and quality characteristics such as the residual rate of hair roots, the residual area percentage of epidermis and the density of grain damage are accurately quantified; secondly, a hybrid digital twin body is constructed by fusing a mechanism model and a data-driven model, based on real-time process parameters, quality characteristic data and initial parameters of raw materials, the future quality state change trend and process completion time are predicted through time series deep learning; then, process parameter adjustment strategies are generated according to the prediction results, and are converted into control instructions to drive the frequency conversion, temperature control and metering actuators of the processing device; finally, the regulation and control effect is fed back to the digital twin model in real time, and continuous learning evolution is realized.

Description

Technical Field

[0001] This invention belongs to the fields of food processing process control, automation of biomaterial preparation and intelligent manufacturing technology, specifically relating to an online method and system for quality control of enzymatic processing of animal hides for non-leather purposes. Background Technology

[0002] Current enzymatic processing (such as enzymatic hair removal) mainly relies on human experience for process control: operators periodically stop and turn on the processing equipment, and rely on visual observation of the hair root loss, the degree of epidermal removal and the state of grain damage on the animal hide, and then subjectively judge whether to adjust the temperature, speed or add enzyme preparation.

[0003] However, this kind of manual, experience-based control has significant drawbacks: First, the detection process is inherently delayed and subjective. Manual inspection requires shutting down and restarting the equipment, disrupting the continuity of the process, and the detection frequency is limited (usually once every 30-60 minutes), making it difficult to capture transient characteristics of quality changes. Differences in experience among different operators lead to inconsistent judgment standards, and descriptions such as "hair roots removed" and "exfoliation completely removed" lack quantitative indicators, resulting in large batch-to-batch quality fluctuations (CV is usually >15%).

[0004] Second, process control is passive and inefficient. Because future quality trends cannot be predicted, adjustments are only made when quality deviates significantly (e.g., grain surface damage has occurred). By this time, it is often too late, resulting in excessive grain removal damaging the grain surface or incomplete removal requiring rework. Adjustment decisions rely on simple rules (e.g., "increase the temperature by 2 degrees"), without considering the comprehensive balance between energy consumption, chemical consumption, and quality risks. Summary of the Invention

[0005] Therefore, in order to address the above shortcomings, this invention provides an online method and system for controlling the quality of enzymatic processing of animal hides for non-leather purposes. This invention establishes a closed-loop chain of "perception-prediction-decision-execution-evolution," achieves multi-dimensional and precise quantification of animal hide quality through multi-task deep learning, completes intelligent deduction from current state diagnosis to future trend prediction based on a hybrid digital twin, and implements dynamic optimization and adaptive control of process parameters based on the prediction results. Finally, the control effect is fed back to the digital twin model in real time to achieve continuous learning and evolution.

[0006] On the one hand, this invention provides an online method for quality control of enzymatic processing of animal hides for non-leather purposes, comprising: S1. Acquire an image of the animal hide located inside the processing equipment and process the image; S2. Construct a multi-task convolutional neural network model to perform pixel-level semantic segmentation and target detection on the preprocessed image, and extract and quantify the quality features of the animal hide, including the hair root residue rate, the percentage of the epidermal residue area, and the density and level of grain surface damage. S3. Establish a hybrid digital twin that integrates the fusion mechanism model and the data-driven model. Using real-time process parameters, the quantitative data of quality characteristics obtained in step S2, and the initial parameters of raw materials as inputs, predict the future trend of quality status changes and process completion time through a time-series deep learning model. S4. Based on the prediction results of the digital twin, construct a process parameter adjustment strategy; S5. Convert the optimal process parameter adjustment strategy into control commands and send them to the frequency converter, temperature control and precision metering actuators of the processing equipment. Feed back the adjusted process status to step S3 to update the digital twin model.

[0007] On the other hand, the present invention provides an online quality control system for enzymatic processing of animal hides for non-leather purposes, comprising: The feature acquisition module is used to acquire images of animal hides located within the processing device and to process the images. The feature extraction module is used to construct a multi-task convolutional neural network model, perform pixel-level semantic segmentation and target detection on the preprocessed image, and extract and quantify the quality features of the animal hide, including the hair root residue rate, the percentage of the epidermal residue area, and the density and grade of grain surface damage. The prediction module is used to establish a hybrid digital twin that integrates the mechanism model and the data-driven model. It takes real-time process parameters, the quantitative data of quality characteristics obtained in step S2, and the initial parameters of raw materials as inputs, and uses a time-series deep learning model to predict the future trend of quality status changes and process completion time. The strategy module is used to construct a process parameter adjustment strategy based on the prediction results of the digital twin; The execution module is used to convert the optimal process parameter adjustment strategy into control commands, send them to the frequency converter, temperature control and precision metering actuators of the processing equipment, and feed back the adjusted process status to step S3 to update the digital twin model.

[0008] The present invention has the following advantages: This invention establishes a closed-loop chain of "perception-prediction-decision-execution-evolution," achieving multi-dimensional and precise quantification of animal hide quality through multi-task deep learning. It relies on a hybrid digital twin to perform intelligent deduction from current state diagnosis to future trend prediction, and dynamically optimizes and adaptively controls process parameters based on the prediction results. Finally, the control effects are fed back to the digital twin model in real time for continuous learning and evolution. This transforms the traditional offline, lagging, and extensive processing mode, which relies on human experience, into a data-driven, online, real-time, accurate predictive, and proactively preventative intelligent production method, significantly improving the quality stability, resource utilization efficiency, and process adaptability of enzymatic processing.

[0009] This invention is applicable to the processing of animal hides such as cattle, pigs, and sheep using enzymatic methods for hair removal and cleaning, and can be used to produce high-value-added products such as food ingredients, medical adhesive raw materials, and bio-based packaging materials. This invention achieves precise control and endpoint determination of the processing through real-time monitoring and feedback control of key parameters such as hair root residue rate, percentage of residual epidermal area, and density and grade of grain damage. This ensures consistent product quality and safety, distinguishing it from process control methods in the leather industry that primarily rely on chemical treatment and focus on controlling the physical properties of the hides.

[0010] At the same time, this invention is fundamentally different from existing control methods in the leather industry: (1) Different control targets: This invention uses the quality indicators of food / pharmaceutical raw materials (such as residual hair root rate) as the control target; the leather industry uses the tanning adaptability of the hide, such as the degree of swelling, softness, and tanning agent penetration depth, as the control target.

[0011] (2) Different sensor configurations: This invention integrates an online visual inspection system (for hair root residue identification); leather tanning process control usually only monitors basic parameters such as temperature, drum speed, and bath pH. In recent years, although some studies have applied machine vision to the detection of surface defects in leather (such as scratches and color difference detection in finished leather), there are the following limitations: First, it only stays at the offline detection stage and does not form a closed loop with process control; second, it lacks the ability to accurately identify hair root residue, a microscopic feature unique to leather tanning.

[0012] (3) Different criteria for determining the endpoint: This invention uses visual recognition of hair removal completion to comprehensively determine the processing endpoint.

[0013] (4) Different data traceability requirements: The system of the present invention has complete data traceability function, which meets the traceability requirements of the food / pharmaceutical industry for the raw material processing process; the leather industry has no such mandatory traceability standard. Attached Figure Description

[0014] Figure 1 This is a flowchart illustrating an online quality control method for enzymatic processing of animal hides for non-leather purposes. Figure 2 This is a system block diagram of an online quality control system for enzymatic processing of animal hides for non-leather purposes.

[0015] In the diagram: 100, Feature Acquisition Module; 200, Feature Extraction Module; 300, Prediction Module; 400, Strategy Module; 500, Execution Module. Detailed Implementation

[0016] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0017] In this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, without necessarily requiring or implying any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0018] As described in the background section, manual, experience-based control has significant drawbacks: First, the detection process is inherently delayed and subjective. Manual inspection requires shutting down and restarting the equipment, disrupting the continuity of the process, and the detection frequency is limited (usually once every 30-60 minutes), making it difficult to capture transient characteristics of quality changes. Differences in experience among different operators lead to inconsistent judgment standards, and descriptions such as "hair roots removed" and "exfoliation completely removed" lack quantitative indicators, resulting in large batch-to-batch quality fluctuations (CV is usually >15%).

[0019] Second, process control is passive and inefficient. Because future quality trends cannot be predicted, adjustments are only made when quality deviates significantly (e.g., grain surface damage has occurred). By this time, it is often too late, resulting in excessive grain removal damaging the grain surface or incomplete removal requiring rework. Adjustment decisions rely on simple rules (e.g., "increase the temperature by 2 degrees"), without considering the comprehensive balance between energy consumption, chemical consumption, and quality risks.

[0020] In recent years, although some studies have applied machine vision to the detection of defects on the surface of leather (such as scratches and color differences in finished leather), there are the following limitations: First, it only stays at the offline detection stage and has not formed a closed loop with process control; second, it lacks the ability to accurately identify the microscopic feature of leather making, which is hair root residue.

[0021] For the reasons mentioned above, such as Figure 1 As shown, this embodiment provides an online method for quality control of enzymatic processing of animal hides for non-leather purposes, including: S1. Acquire an image of the animal hide located in the processing device and process the image; For example, the processing device is a rotary drum. Acquiring the hide image can be achieved by configuring an industrial camera. The method for configuring the industrial camera is as follows: an industrial camera equipped with a light system and an illumination system is configured with an observation window on the side of the rotary drum as the image acquisition point. The industrial camera can be a Basler acA2500-60gm industrial camera, equipped with a Sony IMX250 sensor (5 megapixels, global shutter), GigE interface, dynamic range 72dB, and frame rate 60fps. To protect the camera from the humid environment (relative humidity >95%) and liquid splashes inside the equipment, a custom IP67-rated stainless steel housing (316L material) is used. The housing is connected to the observation window flange of the processing equipment, and the connection uses a fluororubber sealing ring.

[0022] The optical system can be configured with a Computar M1614-MP2 telecentric lens (16 mm focal length, F1.4) to ensure constant magnification at different object distances. A polarizing filter (linearly polarized, extinction ratio >400:1) is installed at the front of the lens to eliminate specular reflections from wet animal hides.

[0023] The lighting system can consist of a ring of LED light sources (180 mm in diameter) arranged around the observation window, using a low-angle lighting design (30° incident angle of light) to create shadows by utilizing the height difference between the hair roots and the skin surface, thus enhancing the contrast of the hair roots. The LED light source is equipped with constant current drive, adjustable brightness (0-100%), color temperature of 6000K (close to natural light), and color rendering index Ra>95.

[0024] Meanwhile, a cleaning device can be installed in front of the stainless steel housing. For example, an air knife (AirKnife) can be installed at the front of the housing and connected to compressed air (0.4MPa) to automatically blow away moisture from the lens surface before each batch of processing. It is also equipped with an electric rotating wiper (rubber blade) that automatically cleans once every set time.

[0025] When acquiring images, the following two modes can be used: Mode 1, Time Trigger: For example, by default, data is automatically collected every X minutes. The value of X is determined according to the actual situation. The shutter is triggered when the hide is in the center of the observation window. Mode 2, Event Trigger: When an anomaly is detected (such as a sudden increase in torque of 20% or a temperature deviation from the set value of >1℃), the acquisition is immediately triggered to obtain the skin image under the abnormal condition.

[0026] A single acquisition captures a RAW format image (2448×2048 pixels, 12-bit depth) and transmits it to Jetson Orin via the GigE interface.

[0027] The method for processing the image in step S1 is as follows: Perform the following processing flow on Jetson Orin (accelerated using OpenCV 4.6 and CUDA): Illumination normalization: The multi-scale Retinex algorithm (MSR) is used to set the Gaussian blur scale. s 1 =15, s 2 =80, s 3 =250, enhancing the dynamic range of the image and eliminating shadows caused by non-uniform LED illumination inside the drum.

[0028] Geometric distortion correction: Pixel-physical coordinate mapping is pre-established using a calibration plate (checkerboard, 10×10 mm) while the processing equipment is stationary. Since the inner wall of the drum is cylindrical, and the skin becomes curved after attachment, cylindrical projection correction is used. ; ; in: r The radius of the drum. x It is the x-axis. This is the corrected x-axis. y It is the ordinate. It is the corrected ordinate, which converts the curved surface projection into a planar orthographic projection to ensure accurate area calculation.

[0029] Water vapor interference suppression: Combine hardware suppression of polarization filters with software algorithms (Dark Channel Prior algorithm) to further eliminate residual water vapor scattering.

[0030] Contrast Enhancement: Applying the CLAHE algorithm, setting clipLimit=4.0 and tileGridSize=(8,8), adaptive enhancement is performed on the hair root area (grayscale value usually <80), increasing its contrast with the exposed skin background (grayscale value 120-180) by more than 40%.

[0031] S2. Construct a multi-task convolutional neural network model to perform post-processing on the preprocessed image. The post-processing includes pixel-level semantic segmentation and target detection, and extracts and quantifies the quality features of the animal hide. The quality features include hair root residue rate, percentage of epidermal residue area, and density and grade of grain surface damage.

[0032] For example, a method for constructing a multi-task convolutional neural network model is: We constructed an MT-CNN model based on the improved Mask R-CNN (using the PyTorch 1.13 framework), with the following architecture: Backbone: EfficientNet-B4 is used as the feature extraction network. The top fully connected layer is removed, reducing the output feature map resolution to 1 / 32 of the input (76×64×1792). After pre-training on ImageNet, domain adaptation is performed using images of animal hides and hides from non-leather processing applications.

[0033] Microstructure Attention Module (MSA): An MSA module is inserted into the P3 layer (1 / 8 resolution, suitable for small object detection) of the FPN (Feature Pyramid Network). This module contains three parallel branches: Branch 1: 3×3 convolution (extracting local details) Branch 2: 5×5 convolution (extracting contextual information) Branch 3: 1×1 convolution (channel attention, SE-Net structure) The outputs of the three branches are fused by adding them pixel by pixel and then activated by ReLU to enhance the sensitivity to hair root features with a diameter <0.5 mm.

[0034] RPN and ROI Align: The Region Proposal Network (RPN) generates candidate boxes, and ROI Align extracts a fixed-size (14×14) feature map.

[0035] Multi-task Headers: Classification branch: The fully connected layer outputs class probabilities (hair root / skin / damage / background), and uses Softmax activation; Bounding box regression branch: Outputs target location coordinates ( x ,y, w , h ), using Smooth L1 Loss; Mask segmentation branch: The fully convolutional network (FCN) outputs a 28×28 pixel-level mask for accurate area calculation.

[0036] Loss function design: Total loss L = L cls + L box + L mask + L aux , in: Lcls : Classification cross-entropy loss; L box Bounding box regression loss; L mask The mask segmentation uses Dice Loss. This addresses the problem of the model biasing to predict the background due to the extreme imbalance between the foreground and background of the hair roots (foreground ratio <2%). L aux Auxiliary loss, using Focal Loss ( c =2.0) Further focus on fuzzy hair root samples that are difficult to classify.

[0037] For example, the inference and post-processing methods for the model are as follows: The input image is accelerated for inference using Jetson TensorRT, with a single image processing time of <80 ms. Morphological opening operations (kernel size 3×3) are performed on the output mask to remove isolated noise points, and connected component analysis is used to calculate: Hair root residue density: number of connected components / physical area of ​​the image (calibrated and converted to pixel resolution of 0.05 mm / pixel). Hair root residue rate: Number of hair root mask pixels / Total number of skin pixels × 100%; Skin residue rate: number of skin mask pixels / total pixels × 100%; Damage density and grade: The number of each grade is counted based on the area of ​​the damage mask (<1mm² is grade 1, 1-5mm² is grade 2, and so on).

[0038] The output results are uploaded to the time series database (InfluxDB) via the MQTT protocol and aligned with the process parameter timestamps.

[0039] S3. Establish a hybrid digital twin that integrates the fusion mechanism model and the data-driven model. Using real-time process parameters, the quantitative data of quality characteristics obtained in step S2, and the initial parameters of raw materials as inputs, predict the future trend of quality status changes and process completion time through a time-series deep learning model. For example, step S3 specifically includes: S31. Mechanism model establishment.

[0040] Based on the principles of mass transfer, a physical model of the enzymatic hair removal process is established: Root dissolution kinetics: The hydrolysis of connective proteins at the dermal papilla by the enzyme was described using a modified Michaelis-Menten kinetic equation. ; Where [H] is the root hair density (mg / cm²) and [E] is the effective enzyme concentration (U / mL). k cat The constant is the catalytic constant. K m It is the Michaelis constant. f ( T , pH ) is a temperature and pH correction function (the product of the Arrhenius equation and the pH activity curve).

[0041] Epidermal-dermal separation criterion: Establishing an interfacial shear stress model: ; in m For fluid viscosity, dv / day For the velocity gradient, t enzyme This is due to the decrease in interfacial adhesion strength caused by enzymatic hydrolysis. When t > t critical (Critical peel stress, determined experimentally, is typically 0.15 N / mm² for cowhide.) C E > C threshold At that time, it was determined that the epidermis could be removed.

[0042] Mass transfer-heat transfer model: Convection-diffusion equations for temperature field and enzyme concentration field within the processing equipment are established, taking into account forced convection caused by stirring rotation.

[0043] S32, Data-driven model building.

[0044] Employing a Temporal Fusion Transformer (TFT) architecture to handle long timing dependencies (4-8 hours per batch): Input variables: Static covariates: leather type (coded as 0=cowhide, 1=sheepskin, 2=pigskin), initial thickness (mm), initial fat content (%). Time-varying known inputs: processing time, target completion time; Time-varying observed inputs: temperature, pH value, rotation speed, torque, hair root residue rate, and epidermal residue rate.

[0045] Network structure: Variable Selection Networks: Automatically learn the importance weights of each input variable; LSTM encoder (Past LSTM): processes historical observation data (lookback window of 2 hours); LSTM decoder (Future LSTM): processes known future inputs (prediction window 1 hour); Multi-head self-attention: captures long-term dependencies in time-series data (such as the cumulative effect of temperature); Quantile Outputs: Predicts the 10th, 50th, and 90th quantiles of future hair root retention rates, providing an estimate of uncertainty.

[0046] S33, Hybrid Model Fusion and Training.

[0047] Embed the mechanistic model as a physical constraint into the loss function of the data-driven model: ; in L physics The physical residual loss includes: mass conservation residual. Monotonicity constraint: The hair root residual rate should decrease monotonically over time. If the predicted value violates this constraint, a penalty will be imposed. Boundary constraint: The predicted value must be within the range of [0%, 100%].

[0048] Supervised learning was performed using historical production data (≥1000 batches, covering different hide types, parts, seasons, and processing formulas). The Adam optimizer (learning rate 1e) was used during training. -4 (Cosine annealing), batch size 32, training epochs 200. After training on an NVIDIA A100 server, the model weights were converted to ONNX format and deployed to Jetson Orin for edge inference.

[0049] S34, Online calibration mechanism.

[0050] After each batch of processing is completed, the system will automatically execute: 1. Obtain the final actual quality test results for this batch (manual sampling or laboratory testing as the gold standard); 2. Calculate the deviation between the predicted and actual values. e = y pred - y true ; 3. If |e|> threshold (e.g., >5%), trigger incremental learning: use this batch of data to fine-tune the model for 5 epochs (learning rate reduced to 1e). -5 ), update model weights; 4. Use Bayesian neural networks (BNN) or ensemble learning (5 models voting) to estimate prediction uncertainty. When the variance is >0.1, mark it as "low confidence" and prompt manual review and increase the training data for this scenario.

[0051] S4. Based on the prediction results of the digital twin, construct a process parameter adjustment strategy.

[0052] For example, the specific method for step S4 is as follows: S41. Optimize problem construction; Deploy a multi-objective optimization decision engine on an industrial PC (or edge server). For the current moment... t Predicting the future based on digital twin models Δt Given a 30-minute timeframe, construct the following optimization problem: Decision variables: X =[ ΔT,ΔpH,Δω,ΔE This refers to the temperature adjustment, pH adjustment, rotation speed adjustment, and enzyme concentration.

[0053] Objective function: ; middle, w 1 - w 4 For the weighting coefficients, satisfying w 1 + w 2 + w 3 + w 4 =1, dynamically adjusted according to the production strategy: Quality-first mode: w 2 =0.6, w 1 =0.2, w 3 =0.1, w 4 =0.1 (to ensure that the hair root residue rate strictly meets the standard). Efficiency-first mode: w1=0.5, w2=0.3, w3=0.1, w4=0.1 (prioritize shortening processing time); Cost-first model: w3=0.5, w1=0.2, w2=0.2, w4=0.1 (prioritize reducing energy consumption and chemical consumption).

[0054] In the objective function, T pred ( t + Δt () is the predicted completion time. T target It is the target processing time. Δt It predicts the time domain. R hair ( t+Δt To predict the hair root residue rate, C energy It's the cost of energy consumption. C enzyme It's the cost of enzyme preparations. P damage ( t + D t ) represents the probability of damage risk.

[0055] The constraints of the objective function include: 1. Quality Constraints: R hair (t+Δt)≤2%; Physically, this means that the predicted hair root residue rate must meet the standard. 2. Temperature safety limit: 20℃≤ T ( t +ΔT≤35℃; the physical meaning is: the temperature is regulated within a safe range; 3. Speed ​​safety constraint: 5rpm≤ω(t)+ Give ≤15rpm; the physical meaning is: the speed is within the safe range.

[0056] 4. pH safety limit: 7.5≤ pH ( t )+ ΔpH ≤10.5; Physically, this means that the pH value is within a safe range.

[0057] 5. Resource constraints: Enzyme supplementation amount E add ≤0.5% (not exceeding the preset amount).

[0058] The single-objective optimization problem can be solved using either Sequential Quadratic Programming (SQP) or the Interior Point Method to obtain a unique optimal solution.

[0059] S42: Hierarchical decision-making strategy; L1 Level - Parameter Fine-Tuning: Model-based predictive control (MPC) or PID control is used to make small, continuous adjustments to the decision variables.

[0060] For example: if the current pH If the value deviates from the target value by more than the set value, start the peristaltic pump to add buffer solution for adjustment; If the current temperature deviates from the set value, the steam valve opening can be adjusted.

[0061] Level 2 - Mode Switching: When the digital twin's prediction of the future state deviates from the target, the control mode in the preset strategy library is invoked: Mode A (Acceleration Mode): If the predicted completion time delay is >30 minutes and the hair root residue rate is >3%, then the following instructions will be generated: such as increasing the temperature by 3°C (within the safety limit), increasing the rotation speed to the set value, and adding 0.5% enzyme preparation; Mode B (Protection Mode): If image recognition detects localized abnormal grain surface gloss (a precursor to over-damage) and torque... M >1.2 M rated If the command is changed to the intermittent mode of "rotate forward for 10 minutes - pause for 10 minutes - rotate in reverse for 10 minutes", the temperature will decrease by 3°C. Mode C (Economic Mode): If the forecast shows that the process can be completed ahead of schedule and the quality margin is sufficient, an instruction is generated: such as stopping heating (natural cooling), reducing the speed to the set value, and preparing to enter the next process.

[0062] Level 3 - Safety Intervention: The hard protection logic is independent of the system and is executed directly by the PLC. like M >1.2 M rated or T >35℃ or pH >11: Immediately disconnect the main motor power, turn on the emergency cooling water, and trigger the audible and visual alarm; If three consecutive image recognitions detect severe damage of grade >5, the enzymatic hydrolysis reaction will be terminated by automatically injecting a reaction terminator (diluted citric acid solution).

[0063] S43: Solving using an optimization algorithm; For example, the improved NSGA-III algorithm (i.e., non-dominated sorting genetic algorithm III) can be used to solve multi-objective problems: Population size: 100; Number of iterations: 200; Crossover operator: SBX (simulated binary crossover), distribution index or c =20; Mutation operators: polynomial mutation, distribution index or m =20; Reference points: Uniformly distributed reference points were generated using the Das and Dennis method (4 targets, each target divided into...). p =5 segments, total (Reference points).

[0064] To improve real-time performance, an offline Pareto front library can be established: for typical scenarios (such as 20 combinations like "thick cowhide + low temperature season" and "thin sheepskin + standard process"), the optimal Pareto front is pre-calculated offline and stored. During online runtime, the closest pre-stored front is quickly matched according to the current scenario, and local searches are performed only within a small range (reducing the number of iterations to 50), ensuring that the total latency from prediction to decision output is <100 ms.

[0065] S5. Convert the optimal process parameter adjustment strategy into control commands and send them to the frequency converter, temperature control and precision metering actuators of the processing equipment. Feed back the adjusted process status to step S3 to update the digital twin model.

[0066] For example, step S5 includes the following specific methods: S51, Control command issuance; The optimal parameter adjustment strategy generated by the decision engine is distributed to the PLC via the OPC UA protocol (based on publish / subscribe mode): Temperature setpoint: The steam flow rate entering the jacket of the processing equipment is controlled by adjusting the proportional control valve (positioner input 4-20 mA corresponding to 0-100% opening) through the analog output module (AO); Speed ​​setting value: The speed setting word is written to the inverter (ABB ACS880) via the PROFINET interface. The inverter adopts vector control mode, with a dynamic response time of <100 ms and a speed accuracy of ±5 rpm. Enzyme addition: A precision metering pump (flow range 0.1-50 L / h, accuracy ±0.5%) is driven by a servo motor, and the number of pulses is automatically calculated and executed according to the command volume.

[0067] S52, Feedback and Model Update; After the actuator moves, the system enters a new sampling cycle: 1. The sensor network acquires new temperature, pH, and torque data (10 Hz); 2. The vision system acquires a new image of the animal hide at the next sampling time (5 minutes later); 3. After preprocessing, the new data is used for real-time feedback of L1 level control (PID closed loop) and is also input into the digital twin model to update state variables. 4. The digital twin model re-predicts future trends based on the new state and initiates the next round of optimization decisions.

[0068] S53. Process completion judgment and knowledge accumulation; When the hair root residue rate is predicted twice consecutively (with a 10-minute interval), R hair <2% and epidermal residue rateR epi When the torque curve shows a stable plateau (indicating that the hair roots are fully loosened), the system determines that the process is complete, automatically generates a "process complete" signal, and triggers the drainage and washing procedures of the treatment equipment.

[0069] The complete data for this batch (process parameter trajectory, quality change curve, decision records) is automatically saved to the process database for: Incremental learning of the digital twin model (as described in step C4); Process knowledge graph construction: Discover the implicit association rules of "temperature-rotation speed-hair loosening rate" through association rule mining (Apriori algorithm); Production report generation: Automatically calculates the quality indicators, energy consumption, and chemical consumption per unit of this batch, and compares and analyzes them with the standard process.

[0070] On the other hand, the present invention provides an online quality control system for enzymatic processing of animal hides for non-leather purposes, such as... Figure 2 As shown, it includes: The feature acquisition module 100 is used to acquire images of animal hides located in the processing device and to process the images; The feature extraction module 200 is used to construct a multi-task convolutional neural network model, perform pixel-level semantic segmentation and target detection on the preprocessed image, and extract and quantify the quality features of the animal hide, including the hair root residue rate, the percentage of epidermal residue area, and the density and grade of grain surface damage. The prediction module 300 is used to establish a hybrid digital twin that integrates the mechanism model and the data-driven model. It takes real-time process parameters, the quantitative data of quality characteristics obtained in step S2, and the initial parameters of raw materials as inputs, and uses a time-series deep learning model to predict the future trend of quality status changes and process completion time. Strategy module 400 is used to construct a process parameter adjustment strategy based on the prediction results of the digital twin; The execution module 500 is used to convert the optimal process parameter adjustment strategy into control commands, send them to the frequency converter, temperature control and precision metering actuators of the processing equipment, and feed back the adjusted process status to step S3 to update the digital twin model.

[0071] This system is used to implement the above-mentioned online quality control method for enzymatic processing of animal hides for non-leather purposes, and has all the beneficial effects of the above-mentioned online quality control method for enzymatic processing of animal hides for non-leather purposes.

[0072] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for online quality control of enzymatic processing of animal hides for non-leather tanning purposes, characterized in that, include: S1. Acquire an image of the animal hide located in the processing device and process the image; S2. Construct a multi-task convolutional neural network model to perform pixel-level semantic segmentation and target detection on the preprocessed image, and extract and quantify the quality features of the animal hide, including the hair root residue rate, the percentage of the epidermal residue area, and the density and grade of grain surface damage. S3. Establish a hybrid digital twin that integrates the fusion mechanism model and the data-driven model. Using real-time process parameters, the quantitative data of quality characteristics obtained in step S2, and the initial parameters of raw materials as inputs, predict the future trend of quality status changes and process completion time through a time-series deep learning model. S4. Based on the prediction results of the digital twin, construct a process parameter adjustment strategy; S5. Convert the optimal process parameter adjustment strategy into control commands and send them to the frequency converter, temperature control and precision metering actuators of the processing equipment. Feed back the adjusted process status to step S3 to update the digital twin model.

2. The method for online quality control of animal skin enzymatic processing according to claim 1, characterized in that: The image processing includes: performing illumination normalization using the Retinex algorithm or homomorphic filtering, performing perspective transformation based on pre-calibrated drum surface parameters to correct geometric distortion, and applying the CLAHE algorithm to enhance the contrast between hair roots and dermal grain surfaces.

3. The method for online quality control of enzymatic processing of animal hides for non-leather purposes according to claim 1, characterized in that: The multi-task convolutional neural network model in step S2 adopts an improved Mask R-CNN or U-Net architecture; the loss function of the model adopts a combination of Dice Loss and Focal Loss.

4. The method for online quality control of enzymatic processing of animal hides for non-leather purposes according to claim 1, characterized in that: The quality feature quantification in step S2 includes: The number and morphology of residual hair roots per unit area are calculated by instance segmentation, and the hair root retention rate is output. The area of ​​the residual epidermis is calculated by pixel-level segmentation, and the percentage of the residual epidermis area is output. Based on the target detection algorithm, the types of surface damage are identified, including scratches, holes, and spots, and the damage density and damage level distribution are output according to area and depth.

5. The method for online quality control of enzymatic processing of animal hides for non-leather purposes according to claim 1, characterized in that: The real-time process parameters in step S3 include temperature, pH value, stirring speed, torque, and processing time. The initial parameters of the raw materials include the type of raw materials, thickness, and initial fat content; the digital twin model is calibrated online after each batch of processing is completed by using the deviation between the actual quality test results and the predicted values ​​through incremental learning or Bayesian update mechanism.

6. The method for online quality control of enzymatic processing of animal hides for non-leather purposes according to claim 1, characterized in that: The multi-objective optimization problem in step S4 is solved using NSGA-III or multi-objective particle swarm optimization algorithms, and the objective function is: ； in: T pred ( t + Δt () is the predicted completion time. T target It is the target processing time. Δt It predicts the time domain. R hair ( t +Δt To predict the hair root residue rate, C energy It's the cost of energy consumption. C enzyme It's the cost of enzyme preparations. P damage ( t + Δt ) represents the probability of damage risk. w 1 - w 4 The weights are dynamically adjustable.

7. The method for online quality control of enzymatic processing of animal hides for non-leather purposes according to claim 1, characterized in that: The process parameter adjustment strategy in step S4 adopts a hierarchical decision-making mechanism: L1 level parameter fine-tuning: Based on PID control or gradient descent, fine-tuning of temperature and speed; L2-level mode switching: When the prediction deviation exceeds the threshold, the preset control mode is invoked. The preset control mode includes intermittent rotation mode, accelerated enzymatic hydrolysis mode or mild treatment mode. Level L3 safety intervention: When the torque exceeds the set value or a risk of local over-damage is detected, an emergency shutdown is triggered or an automatic injection of a reaction terminator is used to terminate the enzymatic hydrolysis reaction.

8. An online quality control system for enzymatic processing of animal hides for non-leather purposes, including: The feature acquisition module is used to acquire images of animal hides located within the processing device and to process the images. The feature extraction module is used to construct a multi-task convolutional neural network model, perform pixel-level semantic segmentation and target detection on the preprocessed image, and extract and quantify the quality features of the animal hide, including the hair root residue rate, the percentage of the epidermal residue area, and the density and grade of grain surface damage. The prediction module is used to build a hybrid digital twin that integrates the mechanism model and the data-driven model. It takes real-time process parameters, acquired quality feature quantification data and raw material initial parameters as inputs, and uses a time-series deep learning model to predict the future trend of quality status changes and process completion time. The strategy module is used to construct a process parameter adjustment strategy based on the prediction results of the digital twin; The execution module is used to convert the optimal process parameter adjustment strategy into control commands, send them to the frequency converter, temperature control and precision metering actuators of the processing equipment, and feed back the adjusted process status to step S3 to update the digital twin model.

Images