A pet food production equipment adaptive control method

Abstract

The present application relates to the technical field of industrial control system, specifically to a pet food production equipment adaptive control method, including obtaining raw material type, melt temperature, bulk density and specific mechanical energy and other key parameters, when the change of raw material type is monitored, the three-dimensional control feasible region boundary between key quality attributes is calculated by using process mapping model; the feasible region boundary is used as control constraint and input into multi-objective optimization agent with real-time key quality attributes, target planning algorithm is executed, the optimal control point is calculated by using feasible region boundary combined with real-time key quality attributes, the optimal operation point agent corresponding to the optimal control point is determined, the optimal control point is input into three control loops, the process gain matrix is updated through local linearization, the output coordination signal is adjusted through feedforward decoupling matrix compensation, and the equipment is output. The present application improves product quality consistency through adaptive target determination and decoupling control.

Description

Technical Field

[0001] This invention relates to the field of industrial control system technology, specifically to an adaptive control method for pet food production equipment. Background Technology

[0002] In many complex manufacturing processes, such as the operation of thermomechanical reactors (e.g., extruders in pet food production), control systems face significant challenges.

[0003] First, the controlled object of this type is essentially a multi-input multi-output system, with strong nonlinear coupling between its multiple key output variables (such as melt temperature, bulk density, and specific mechanical energy). In existing technologies, even when standard multivariable control strategies (such as static decoupling networks combined with PID) are adopted, controlling one variable (e.g., adjusting the screw speed) will cause unexpected disturbances to other variables (e.g., causing melt temperature fluctuations), resulting in decreased stability and performance oscillations in the control loop, making it difficult to achieve coordinated stability of multiple objectives.

[0004] Secondly, and more fundamentally, the dynamic characteristics of these processes are not constant. In applications such as pet food production, random fluctuations in raw material properties (such as changes in composition and moisture content) are a common and unmeasured strong disturbance. This disturbance can alter the mathematical model of the controlled object in real time and nonlinearly, causing significant drifts in process gain, time constant, and coupling coefficients between multiple variables.

[0005] This directly leads to control model mismatch. Traditional control systems, whether based on fixed-parameter PID controllers or advanced controllers tuned by static models, design their control laws based on a nominal process model that is assumed to be invariant. When the actual process dynamics deviate from this nominal model due to fluctuations in raw materials, the performance of these fixed-parameter controllers deteriorates significantly, they lose robustness, and they cannot effectively suppress disturbances. Ultimately, this results in key quality attributes failing to stabilize at the setpoint, severely impacting product quality consistency.

[0006] Therefore, an adaptive control method for pet food production equipment is proposed. Summary of the Invention

[0007] The purpose of this invention is to provide an adaptive control method for pet food production equipment, which improves product quality consistency through adaptive target determination and decoupled control. This method includes acquiring key parameters such as raw material type, melt temperature, bulk density, and specific mechanical energy. When a change in raw material type is detected, a process mapping model is used to calculate the three-dimensional control feasible domain boundary between key quality attributes. This feasible domain boundary, along with real-time key quality attributes, is input into a multi-objective optimization agent. A target planning algorithm is executed, and the optimal control point is calculated using the feasible domain boundary combined with real-time key quality attributes. The agent determines the optimal operating point corresponding to the optimal control point, and this optimal control point is input into the Smith predictor, MPC, and PI three-loop system. The process gain matrix is ​​updated through local linearization, and after feedforward decoupling matrix compensation, a coordinated signal is output to regulate the equipment.

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] An adaptive control method for pet food production equipment includes:

[0010] Obtain the raw material type, key quality attributes, and control variables; key quality attributes include the melt temperature and bulk density of the finished product, and specific mechanical energy; control variables include die temperature, screw speed, and water injection rate.

[0011] When a change in raw material type is detected, a process mapping model is constructed using the raw material type and control variables to calculate the three-dimensional control feasible domain boundary between key quality attributes. The feasible domain boundary is used as a control constraint and input into the multi-objective optimization agent along with the real-time key quality attributes. The objective planning algorithm is executed, and the optimal control point is calculated using the feasible domain boundary combined with the real-time key quality attributes. The optimal operation point corresponding to the optimal control point is determined, and the process mapping model is used to linearize at the optimal operation point to generate an updated dynamic coupling model.

[0012] The optimal control point is input into three control loops; the three control loops calculate the first, second, and third intermediate control signals respectively; the three intermediate control signals are input into the feedforward control decoupling matrix; the control decoupling matrix receives and updates the dynamic coupling model, performs compensation calculations on the three intermediate control signals, outputs a coordinated control signal, and adjusts the equipment parameters.

[0013] Preferably, the process mapping model includes:

[0014] The raw material feature encoder receives raw material type data as input, processes it using an embedding layer and a fully connected network, and generates a static raw material feature vector.

[0015] The batch prediction submodule, which contains a virtual grid generator and a core prediction network, is used for:

[0016] The system receives the operating range of the screw speed, the water injection volume, and the die temperature, and creates a three-dimensional virtual process mesh using a virtual mesh generator. It then broadcasts and concatenates the static raw material feature vector with each data point in the three-dimensional virtual process mesh to form a batched combined input tensor. This batched combined input tensor is input into the core prediction network, outputting a three-dimensional prediction point cloud. Each point in the prediction point cloud contains a coordinate pair of the predicted melt temperature, bulk density, and specific mechanical energy.

[0017] The boundary extraction submodule receives the 3D predicted point cloud as input and contains an Alpha-Shape calculation unit. It performs boundary fitting on the predicted point cloud, identifies and connects the outermost non-convex boundary points of the point cloud, and finally outputs the 3D control feasible domain boundary.

[0018] Preferably, the process of calculating the optimal control point includes: defining a fixed ideal target point located outside the boundary of the three-dimensional control feasible domain within the agent; the ideal target point is set as the coordinates corresponding to the target values ​​of melt temperature, bulk density, and specific mechanical energy; the agent calculates the current three-dimensional deviation vector between the ideal target point and the current real-time melt temperature, bulk density, and specific mechanical energy data; the current three-dimensional deviation vector is input to a dynamic weight allocation module, which calculates and outputs a set of three-dimensional dynamic weights based on the magnitude of the deviation vector; the agent discretizes the calculated boundary of the three-dimensional control feasible domain to obtain a set of boundary candidate points; iteratively calculates the three-dimensional weighted Euclidean distance between the ideal target point and each point in the set of boundary candidate points, the Euclidean distance being calculated using the three-dimensional dynamic weights; the agent determines the boundary candidate point with the smallest weighted Euclidean distance and outputs the boundary candidate point as the optimal compromise point, the coordinates corresponding to the optimal compromise point being the optimal control point.

[0019] Preferably, the dynamic weight allocation module calculates and outputs the three-dimensional dynamic weights in real time based on the magnitude of the deviation vector, including: inputting the current three-dimensional deviation vector into a nonlinear amplification unit; the nonlinear amplification unit uses an exponential function to calculate the melt temperature deviation value, bulk density deviation value, and specific mechanical energy deviation value respectively to obtain an amplified intermediate vector; inputting the amplified intermediate vector into a normalization unit; the normalization unit performs a Softmax normalization operation on the amplified intermediate vector, calculates and outputs the three-dimensional dynamic weights.

[0020] Preferably, the first control loop of the three control loops is configured as a Smith predictor controller structure. The Smith predictor controller structure includes a main controller and an internal dynamic model. The parameters of the internal dynamic model are locally identified and updated by the process mapping model at the optimal operating point. The main controller is configured as a PI controller. The internal dynamic model is trained to characterize the high inertia and pure time delay response characteristics of the melt temperature to the first intermediate control signal. The main controller calculates the deviation between the target melt temperature value and the corrected feedback signal, and calculates the first intermediate control signal based on the deviation. The first intermediate control signal is sent to the feedforward decoupling matrix. The corrected feedback signal is calculated as follows: the first intermediate control signal is input into the internal dynamic model to generate a predicted melt temperature; the error correction term is added to the predicted melt temperature to obtain the corrected feedback signal. The error correction term is obtained by calculating the difference between the real-time melt temperature data and the simulated output of the internal dynamic model after a delay.

[0021] Preferably, the second control loop of the three control loops is configured as a model predictive controller; the controller calculates the current deviation between the target packing density value and the real-time packing density data; the controller retrieves a set of hard constraints, which are defined as the physical operating limits of the equipment; the controller uses a finite impulse response dynamic process model to perform optimization calculations in the optimization time domain, the dynamic process model being trained to characterize the response of the packing density to a second intermediate control signal; the optimization calculation uses an objective function to find the optimal internal control sequence while discarding all candidate control sequences that would trigger the hard constraints; the objective function is specifically calculated by: calculating the prediction deviation between the future packing density response and the target packing density value, and applying an asymmetric penalty term to the prediction deviation; the asymmetric penalty term is configured to: apply a penalty weight for packing density predictions higher than the target value. A penalty weight is applied to cases where the predicted packing density is lower than the target value. ;in, and It is a positive real number, and , The relative size can be configured according to the product's quality preference; the controller determines the optimal internal control sequence that minimizes the total cost calculated by the objective function within the optimization time domain, outputs the first control action of the optimal internal control sequence, and normalizes the action as the second intermediate control signal.

[0022] Preferably, the third control loop of the three control loops is configured as a PI controller; the PI controller is configured with a fast response parameter to match the fast time base and low inertia mechanical dynamic characteristics of the specific mechanical energy with high bandwidth; the PI controller calculates the instantaneous deviation between the target value of the specific mechanical energy and the real-time specific mechanical energy data; based on the instantaneous deviation, it calculates and outputs a third intermediate control signal at high speed, and the third intermediate control signal is sent to the feedforward decoupling matrix.

[0023] Preferably, the step of performing compensation calculation by the feedforward control decoupling matrix specifically includes: the dynamic coupling model built into the feedforward control decoupling matrix is ​​defined as a 3x3 process gain matrix. The parameters of the matrix are dynamically generated and updated online by linearizing the model near the optimal control point after the process mapping model calculates the optimal control point; the feedforward control decoupling matrix checks the condition number of the 3x3 process gain matrix. If the condition number is within a preset range, the inverse matrix of the 3x3 process gain matrix is ​​calculated; if the condition number exceeds the preset range, the pseudo-inverse matrix of the 3x3 process gain matrix is ​​calculated to obtain the 3x3 decoupling matrix; the feedforward control decoupling matrix performs matrix multiplication with the input vector containing the first, second, and third intermediate control signals; the feedforward control decoupling matrix outputs the result of the matrix multiplication as a cooperative control signal, simultaneously controlling the screw speed of the thermomechanical reactor, the water injection volume of the material pretreatment unit, and the die temperature of the thermomechanical reactor.

[0024] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0025] 1. By using a dynamic control model to calculate the boundary of the three-dimensional control feasible domain when a change in the type of raw materials is detected, and executing a target programming algorithm, an optimal compromise control point under the current physical constraints can be adaptively calculated. This solves the problem that the system attempts to track an unattainable static setpoint due to fluctuations in raw material characteristics, and avoids the control system from going out of control due to unreasonable targets.

[0026] 2. By using a feedforward control decoupling matrix to perform real-time compensation calculations on three heterogeneous parallel control loops, the physical coupling interference between three key quality attributes (melt temperature, bulk density, and specific mechanical energy) can be proactively offset, solving the problem of control loop conflicts caused by multivariable coupling. Thus, at the calculated optimal control point, the oscillations caused by fluctuations in raw material characteristics are significantly suppressed.

[0027] 3. By comprehensively and adaptively determining the optimal combination of objectives and the ability to stably decouple execution, when faced with fluctuations in raw material characteristics, it can enable the three key quality attributes (melt temperature, bulk density, and specific mechanical energy) to converge quickly and accurately to their dynamic optimal points, thus solving the problem of batch-to-batch inconsistencies in product quality caused by severe degradation of control system performance. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of an adaptive control method for pet food production equipment according to the present invention;

[0029] Figure 2 This is a schematic diagram of the process mapping model operation flow of the present invention;

[0030] Figure 3 This is a schematic diagram of the process for performing compensation calculations using the feedforward control decoupling matrix of the present invention. Detailed Implementation

[0031] 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.

[0032] Please see Figures 1 to 3 This invention provides an adaptive control method for pet food production equipment, the technical solution of which is as follows:

[0033] Example 1:

[0034] An adaptive control method for pet food production equipment, the specific process of which is as follows: Figure 1 As shown, it includes:

[0035] Obtain the raw material type, key quality attributes, and control variables; key quality attributes include the melt temperature and bulk density of the finished product, and specific mechanical energy; control variables include die temperature, screw speed, and water injection rate.

[0036] When a change in raw material type is detected, a process mapping model is constructed using the raw material type and control variables to calculate the three-dimensional control feasible domain boundary between key quality attributes. The feasible domain boundary is used as a control constraint and input into the multi-objective optimization agent along with the real-time key quality attributes. The objective planning algorithm is executed, and the optimal control point is calculated using the feasible domain boundary combined with the real-time key quality attributes. The optimal operation point corresponding to the optimal control point is determined, and the process mapping model is used to linearize at the optimal operation point to generate an updated dynamic coupling model.

[0037] The optimal control point is input into three control loops; the three control loops calculate the first, second, and third intermediate control signals respectively; the three intermediate control signals are input into the feedforward control decoupling matrix; the control decoupling matrix receives and updates the dynamic coupling model, performs compensation calculations on the three intermediate control signals, outputs a coordinated control signal, and adjusts the equipment parameters.

[0038] Furthermore, the process of acquiring data on raw materials, finished products, and the thermomechanical reactor includes: real-time scanning and interpretation of spectral data of the mixed raw materials using a near-infrared spectral sensor installed at the material inlet of the thermomechanical reactor to obtain the raw material type; real-time acquisition of the melt temperature using a melt temperature probe installed at the product outlet of the thermomechanical reactor; real-time measurement of the volume and dimensions of the finished product using a machine vision system installed at the product outlet, and calculation of the bulk density based on the volume and dimensions; real-time acquisition of the screw speed by reading the speed encoder of the main motor of the thermomechanical reactor; calculation of the specific mechanical energy by measuring the load of the main motor of the thermomechanical reactor and combining it with the real-time acquired screw speed; real-time acquisition of the die head temperature using a temperature sensor installed at the die head of the thermomechanical reactor; and real-time acquisition of the water injection volume using a mass flow meter installed on the water injection pipeline of the material pretreatment unit.

[0039] In one specific embodiment of this application, the data acquisition process begins with the monitoring of key interference variables. A near-infrared spectral sensor is installed at the material inlet of a thermomechanical reactor (e.g., a twin-screw extruder). This sensor continuously scans the mixed raw material about to enter the barrel and sends the acquired spectral data to an interpretation module. This module interprets the spectral data in real time using a pre-established chemometric model, thereby obtaining the digital characteristics of the current batch of raw material.

[0040] Simultaneously, at the product outlet of the thermomechanical reactor, the system acquires two key finished product attributes in parallel. First, it acquires the melt temperature in real time through a melt temperature probe (e.g., an armored thermocouple) installed inside the die before the material is extruded. The tip of the probe directly contacts the high-temperature and high-pressure molten material. Second, it acquires the melt temperature in real time through a machine vision system installed downstream of the cutter. This system uses a high-speed industrial camera to photograph the finished product particles, and the image processing software calculates the bulk density of the finished product in real time by analyzing the particle contour, volume, and size.

[0041] In addition, the system acquires three key data points from the thermomechanical reactor body. First, it obtains the screw speed in real time by reading the speed encoder built into the variable frequency drive that controls the main motor. Second, it measures the load of the main motor (e.g., output current or torque percentage) in real time through the communication interface of the same variable frequency drive. The control system then combines this real-time load value with the real-time screw speed and calculates the specific mechanical energy using a standard formula. Finally, it obtains the die head temperature in real time through a temperature sensor (e.g., a PT100 resistance temperature detector) installed inside the metal part of the thermomechanical reactor die head.

[0042] On auxiliary equipment, the water injection volume is obtained in real time and accurately through a mass flow meter installed on the water injection pipeline of the material pretreatment unit (e.g., a conditioner).

[0043] By comprehensively acquiring real-time data from raw materials, finished products, and thermomechanical reactors, the adaptive control system of this application can obtain accurate and complete process variables at the instant of disturbance occurrence and process fluctuation. This provides a reliable and timely data foundation for subsequent dynamic modeling and decoupled control, and is a necessary prerequisite for ensuring the accurate operation of the entire adaptive algorithm.

[0044] The process of detecting a change in raw material type includes: storing in the control system a reference spectral feature vector and its covariance matrix extracted and dimensionally reduced from the reference raw material using a chemometric model; the control system processes the raw material type, i.e., the current spectral data, acquired in real time by a near-infrared spectral sensor, using the chemometric model to calculate the current spectral feature vector; the control system uses the Mahalanobis distance algorithm, utilizing the reference spectral feature vector and covariance matrix, to calculate the statistical distance between the current spectral feature vector and the reference spectral feature vector, obtaining a change scalar value; comparing the change scalar value with a preset change threshold, and only triggering the recalculation of the three-dimensional control feasible domain boundary by the dynamic control model when the change scalar value exceeds the change threshold.

[0045] First, before system deployment, an offline calibration phase must be performed to generate a baseline. In this embodiment, the chemometric model is configured as a principal component analysis (PCA) model. A batch of samples defined as baseline raw materials are acquired and scanned extensively and repeatedly using a near-infrared spectral sensor to obtain a baseline spectral dataset. The PCA model then processes this dataset, extracting and calculating a low-dimensional baseline spectral feature vector (i.e., the mean vector of all baseline samples in the principal component space) and a covariance matrix (i.e., the statistical variance of the baseline samples in the principal component space) through mathematical dimensionality reduction. The baseline spectral feature vector and the covariance matrix are then fixed and stored in the control system.

[0046] Secondly, during the online operation phase of the production line, the system performs real-time monitoring. The near-infrared spectral sensor continuously (e.g., once per second) scans the currently flowing mixed raw material and acquires its current spectral data in real time. The calibrated chemometric model (i.e., PCA model) immediately projects this current spectral data onto the established principal component space to calculate a current spectral feature vector.

[0047] Next, the control system executes a statistical comparison algorithm, which calculates the statistical distance using a Mahalanobis distance algorithm. The inputs to this algorithm include: the current spectral feature vector; the stored reference spectral feature vector (i.e., the mean); and the stored covariance matrix (inverse matrix). The output of the algorithm is a variable scalar value that quantifies the statistical significance of the current raw material's deviation from the reference raw material.

[0048] Finally, the control system executes a trigger decision. A change threshold is preset in the control system (this threshold can be determined by technicians based on the statistical confidence level of the chi-square distribution). The control system continuously compares the calculated real-time change scalar value with this change threshold. If the change scalar value is lower than the threshold, the system determines that the raw material type has not changed significantly, and at this time, the subsequent recalculation of the feasible region boundary is not performed. However, once the change scalar value exceeds the change threshold, the system determines that the raw material type has changed significantly and immediately triggers the recalculation of the three-dimensional control feasible region boundary by the dynamic control model. This ensures that the computationally expensive (slow time-based) supervised optimization loop (i.e., the recalculation of the three-dimensional control feasible region boundary) is only triggered when absolutely necessary, thereby greatly saving computational resources and ensuring the real-time stability of the fast time-based control loop.

[0049] The process mapping model includes, for example, the specific operational flow as follows: Figure 2 As shown:

[0050] The raw material feature encoder receives raw material type data as input, processes it using an embedding layer and a fully connected network, and generates a static raw material feature vector.

[0051] The batch prediction submodule, which contains a virtual grid generator and a core prediction network, is used for:

[0052] The system receives the operating range of the screw speed, the water injection volume, and the die temperature, and creates a three-dimensional virtual process mesh using a virtual mesh generator. It then broadcasts and concatenates the static raw material feature vector with each data point in the three-dimensional virtual process mesh to form a batched combined input tensor. This batched combined input tensor is input into the core prediction network, outputting a three-dimensional prediction point cloud. Each point in the prediction point cloud contains a coordinate pair of the predicted melt temperature, bulk density, and specific mechanical energy.

[0053] The boundary extraction submodule receives the 3D predicted point cloud as input, performs boundary fitting on the predicted point cloud, identifies and connects the outermost non-convex boundary points of the point cloud, and finally outputs the 3D control feasible domain boundary.

[0054] The raw material feature encoder receives raw material type data (e.g., a 10-dimensional spectral feature vector reduced from the chemometric model) as input, wherein the output dimension of the embedding layer and the dimension of the static raw material feature vector are set to 128-dimensional floating-point vectors. The fully connected network ultimately generates a 128-dimensional static raw material feature vector.

[0055] The batch prediction submodule receives the operating ranges of the screw speed (operating range, for example, 200 rpm to 800 rpm), the water injection rate (operating range, for example, 1 L / min to 8 L / min), and the die temperature (operating range, for example, 100°C to 150°C). The virtual mesh generator uses a uniform, equally spaced sampling method to discretize the three operating variables: screw speed, water injection rate, and die temperature. For screw speed (operating range 200-800 rpm), 50 points are sampled at 12 rpm intervals; for water injection rate (operating range 1-8 L / min), 50 points are sampled at 0.14 L / min intervals; and for die temperature (operating range 100-150°C), 50 points are sampled at 1°C intervals. This sampling density was determined through prior sensitivity analysis, ensuring that the extraction accuracy of the three-dimensional control feasible domain boundary is not lower than the actual operating resolution. The final three-dimensional virtual process mesh contains 50×50×50=125,000 data points.

[0056] The core prediction network comprises one input layer and two fully connected hidden layers. The input layer receives a 131-dimensional combined tensor, formed by concatenating the static raw material feature vector (128-dimensional) with a data point (3-dimensional) from a 3D virtual process grid. The first hidden layer contains 256 neurons using the ReLU activation function. The second hidden layer contains 128 neurons using the ReLU activation function. The output layer contains three neurons, predicting melt temperature, bulk density, and specific mechanical energy, respectively. This output layer uses a linear activation function to support continuous numerical prediction. The network is trained offline using mean squared error (MSE) as the loss function and iteratively optimized using the Adam optimizer (with an initial learning rate of 0.001). The training and validation datasets are divided in an 8:2 ratio. Training stops and the best-performing model is saved when the MSE on the validation set no longer decreases for five consecutive epochs.

[0057] Before system deployment, a data acquisition and offline training phase must be conducted. First, a series of Design of Experiments (DOE) experiments are performed on the physical equipment (thermomechanical reactor), or simulations are performed using a high-fidelity CFD (Computational Fluid Dynamics) model. During this training data acquisition, the system is set to operate under multiple combinations of different controllable process variables (i.e., screw speed, water injection rate, and die temperature) and to produce using a variety of different, representative raw material types. In these experiments, the system fully records the causal relationship data pairs between the three controllable variables (inputs) and the three key quality properties (melt temperature, bulk density, and specific mechanical energy) (outputs).

[0058] Subsequently, these data pairs are used to train the process mapping model. The raw material type data (e.g., discrete type labels of raw materials or pre-extracted spectral feature vectors) are fed into the raw material feature encoder. The embedding layer and fully connected network convert this discrete raw material type label into a high-dimensional, dense static raw material feature vector. This static raw material feature vector is concatenated with the manipulable process variables (screw speed, water injection rate, die temperature) in the experimental data and input into the core prediction network (which consists of multiple fully connected hidden layers). This network is trained to learn how raw material features and operational inputs jointly predict melt temperature, bulk density, and specific mechanical energy. After training, the model is solidified and deployed into the online control system.

[0059] The online computation phase begins with the triggering event (i.e., detection of a new raw material type). When this event occurs, the current spectral feature vector calculated by the chemometric model is input to the raw material feature encoder. The encoder immediately processes it and outputs a static raw material feature vector. Simultaneously, the virtual mesh generator within the batch prediction submodule is activated. This generator acquires the full operational range (i.e., its physical minimum and maximum allowable values) of the three manipulated variables and creates a three-dimensional virtual process mesh containing tens of thousands of data points in the three-dimensional space. Subsequently, the system performs a broadcast stitching operation: the static raw material feature vector is copied and stitched onto each data point in the three-dimensional virtual process mesh, thereby forming a batched combined input tensor. The batched combined input tensor is input to the core prediction network at once. The core prediction network performs high-speed batch forward prediction on this batch of tensors and outputs a three-dimensional predicted point cloud of the predicted melt temperature, predicted bulk density, and predicted specific mechanical energy coordinate pairs.

[0060] Finally, the 3D predicted point cloud is passed to the boundary extraction submodule. The Alpha-Shape computing unit inside the submodule executes the Alpha-Shape computing algorithm on the 3D predicted point cloud. This algorithm can identify and connect the outermost non-convex boundary points of the point cloud, thereby generating an accurate and mathematically usable 3D control feasible region boundary. This calculated 3D control feasible region boundary is then passed to the multi-objective optimization agent as a dynamic constraint for its execution of the objective planning algorithm.

[0061] In a preferred embodiment, the process mapping model is further configured to: output the three-dimensional control feasible domain boundary and a prediction confidence interval calculated based on Monte Carlo Dropout, wherein the confidence interval characterizes the prediction confidence of the boundary; the multi-objective optimization agent is configured to execute a robust optimization algorithm, wherein when calculating the optimal control point, the agent searches within the entire prediction confidence interval of the boundary to determine a statistically robust control point that can still maintain acceptable performance in the worst case of model uncertainty, thereby obtaining the optimal control point.

[0062] After the process mapping model completes offline training, all hidden layers of its core prediction network (i.e., neural network) are configured to enable the Dropout mechanism. During the online operation phase, when the batch prediction submodule outputs a 3D prediction point cloud, the control system does not disable Dropout in a single prediction request, but instead randomly activates Dropout at a predetermined ratio (e.g., 20%). The control system performs N (e.g., N=100) independent forward prediction calculations on the same batch input tensor (i.e., the same virtual mesh and static raw material features). Due to the randomness of Dropout, each prediction outputs a slightly different 3D prediction point cloud. Subsequently, the control system calculates the standard deviation of the predicted values ​​of each 3D coordinate point (melt temperature, bulk density, specific mechanical energy) in these N prediction point clouds. This standard deviation quantifies the cognitive uncertainty of the model. The standard deviation is used to define a prediction confidence interval (e.g., defined as the prediction mean plus or minus two standard deviations). This confidence interval encompasses the original 3D control feasible domain boundary, thereby characterizing the confidence of the boundary in the prediction. The specific process of the multi-objective optimization agent executing the robust optimization algorithm includes: the objective of the robust optimization algorithm is to minimize the weighted Euclidean distance in the worst case. After the agent discretizes the boundary of the three-dimensional control feasible region, for each boundary candidate point P, the robust optimization algorithm no longer simply calculates the distance between its predicted mean and the ideal target point, but instead performs the following search: the robust optimization algorithm defines a worst-case point Pw within the prediction confidence interval corresponding to the boundary candidate point P. This worst-case point Pw refers to the point within the confidence interval that maximizes the weighted Euclidean distance between Pw and the fixed ideal target point. The agent determines this worst-case distance as the penalty value for the candidate point P. The agent iteratively calculates the worst-case distance of all boundary candidate points P and selects the boundary candidate point with the smallest worst-case distance as the final optimal control point. Through this robust optimization, the control system sacrifices some average performance, but in exchange, it can reliably maintain within an acceptable range of process targets even when there is uncertainty in the model prediction (e.g., when the raw material characteristics fluctuate drastically), greatly improving the safety and robustness of the adaptive control system.

[0063] Through this dynamic control model, this application can perform forward-looking batch simulations of a specific raw material type (represented by a static raw material feature vector) with all controllable operations (such as screw speed and water injection volume), thereby calculating all physically possible outcomes (predicted point cloud) for the raw material. Subsequently, the boundary extraction submodule uses the Alpha-Shape algorithm to reduce the dimensionality of this complex multidimensional point cloud and simplify it into a precise, mathematically usable three-dimensional controllable feasible domain boundary. This effect provides a clear and reliable computational foundation for subsequent target programming algorithms, ensuring that the optimization process is based on the real physical constraints of the current raw material.

[0064] Further, the process of calculating the optimal control point includes: defining a fixed ideal target point located outside the boundary of the three-dimensional control feasible domain within the agent; the ideal target point is set as the coordinates corresponding to the target values ​​of melt temperature, bulk density, and specific mechanical energy; the agent calculates the current three-dimensional deviation vector between the ideal target point and the current real-time melt temperature, bulk density, and specific mechanical energy data; the current three-dimensional deviation vector is input to a dynamic weight allocation module, which calculates and outputs a set of three-dimensional dynamic weights based on the magnitude of the deviation vector; the agent discretizes and samples the calculated boundary of the three-dimensional control feasible domain to obtain a set of boundary candidate points; iteratively calculates the three-dimensional weighted Euclidean distance between the ideal target point and each point in the set of boundary candidate points, the Euclidean distance being calculated using the three-dimensional dynamic weights; the agent determines the boundary candidate point with the smallest weighted Euclidean distance and outputs the boundary candidate point as the optimal compromise point, the coordinates corresponding to the optimal compromise point being the optimal control point.

[0065] A fixed ideal target point is defined within the intelligent agent. This ideal target point is a three-dimensional coordinate representing the desired perfect process result. Mathematically, this point is set to be located outside the boundary of the three-dimensional control feasible domain. The principle for setting the ideal target point is to set a fixed, unchanging, and most desirable value in each dimension that exceeds the current physical limits of the process or is technically impossible to achieve. This solves the problem of the system attempting to track an unattainable static setpoint due to raw material fluctuations and avoids the control system from going out of control due to unreasonable targets. For example, 155°C is a constant value higher than the upper limit of the process operation of all known raw material melt temperatures (e.g., 148°C), and 250 g / L is a constant value lower than the lower limit of the process operation of all known raw material bulk density (e.g., 350 g / L). In one specific embodiment, the ideal target point is set as (melt temperature = 155°C, bulk density = 250 g / L, specific mechanical energy = 0 kJ / kg), where 155°C is an unattainable value slightly above the upper limit of the process operation (e.g., 150°C), corresponding to maximizing the melt temperature (to ensure gelatinization); 250 g / L is an unattainable value significantly below the lower limit of the conventional process (e.g., below 380 g / L), corresponding to minimizing the bulk density (to ensure a chewable texture); and 0 kJ / kg corresponds to the desired value of minimizing specific mechanical energy (i.e., minimum energy consumption). The agent acquires the current real-time melt temperature data, bulk density data, and specific mechanical energy data from data acquisition and calculates the current three-dimensional deviation vector between the ideal target point and the current real-time data.

[0066] Then, the current three-dimensional deviation vector is input to the dynamic weight allocation module to perform the real-time calculation and output the three-dimensional dynamic weights: First, the current three-dimensional deviation vector is input to the nonlinear amplification unit; the nonlinear amplification unit first performs linear scaling with coefficients on each deviation value, that is, multiplies the original deviation value by a fixed scaling factor (e.g., 0.5) to control the speed of exponential function amplification and prevent numerical overflow. Then, the nonlinear amplification unit uses the natural exponential function to calculate the three scaled deviation values ​​respectively to obtain an amplified intermediate vector; second, the amplified intermediate vector is input to a normalization unit; the normalization unit performs Softmax normalization on the amplified intermediate vector, that is, divides each element in the intermediate vector by the sum of all elements, calculates and outputs the three-dimensional dynamic weights.

[0067] Simultaneously, the agent receives the three-dimensional controllable feasible domain boundary calculated by the dynamic control model. The agent discretizes this three-dimensional boundary surface (generated by the Alpha-Shape algorithm) by sampling (e.g., extracting its mesh vertices or generating thousands of sampling points on its surface) to obtain a set of boundary candidate points. Then, the agent performs iterative calculations. It traverses each point in the set of boundary candidate points. The weighted Euclidean distance is calculated as follows: First, the squared deviations between the candidate point and the ideal target point in the three dimensions of melt temperature, bulk density, and specific mechanical energy are calculated respectively. Then, the three-dimensional dynamic weights (w_temp, w_density, w_sme) output by the dynamic weight allocation module are multiplied by the squared deviations in their corresponding dimensions. Finally, the three weighted squared deviation values ​​are summed and their square root is taken to obtain the weighted Euclidean distance of the candidate point. The calculation uses three-dimensional dynamic weights for weighting.

[0068] The method for discretizing the boundary of the 3D controllable feasible domain by the intelligent agent is as follows: First, all triangular units and vertices of the 3D boundary surface generated by the Alpha-Shape algorithm are extracted. This boundary surface typically contains 5000-10000 triangles and 3000-7000 vertices. The intelligent agent uses these vertices as an initial candidate point set. To increase the coverage density of candidate points, the intelligent agent adds an additional sampling point at the center point of each triangle and the midpoint of each of the three sides, ultimately forming a sampling set containing 10000-20000 candidate points. The intelligent agent then uses an exhaustive search algorithm to traverse all candidate points and calculates... The weighted Euclidean distance from each candidate point to the ideal target point is calculated. During the search process, the agent saves the currently discovered minimum distance value and the corresponding candidate point. After all candidate point distances have been calculated, the agent determines the boundary candidate point corresponding to the minimum distance and outputs it as the optimal compromise point. In the discretization calculation, the agent records the distribution of all distance values. When the difference between the minimum and the second minimum value is less than 1%, it is considered that there are multiple approximate optimal points. At this time, the following strategy is adopted: all candidate points within 150% of the second minimum value form a candidate pool, and the one closest to the current actual control point is selected as the final optimal control point to achieve a smooth transition.

[0069] Finally, after the agent completes the distance calculation for all boundary candidate points, it performs a minimization search to determine the boundary candidate point with the smallest weighted Euclidean distance, and outputs this point as the optimal compromise point, i.e., the optimal control point. The coordinates corresponding to the optimal compromise point are the target combination, which is then sent to the three parallel control loops.

[0070] By employing this target planning algorithm based on ideal target points and dynamic weights, the agent in this application no longer blindly optimizes using fixed, preset weights. Instead, by calculating the current deviation vector in real time and inputting it into the dynamic weight allocation module, the system can adaptively adjust its control priority for three key quality attributes (melt temperature, bulk density, and specific mechanical energy). This effect allows the algorithm to dynamically and centrally address the control objective with the largest current deviation, thereby achieving a more intelligent and efficient real-time compromise in multi-objective conflicts, ensuring that the control system always corrects the most pressing problem first.

[0071] Furthermore, the dynamic weight allocation module calculates and outputs the three-dimensional dynamic weights in real time based on the magnitude of the deviation vector, including: inputting the current three-dimensional deviation vector into a nonlinear amplification unit; the nonlinear amplification unit uses an exponential function to calculate the melt temperature deviation value, bulk density deviation value, and specific mechanical energy deviation value respectively to obtain an amplified intermediate vector; inputting the amplified intermediate vector into a normalization unit; the normalization unit performs a Softmax normalization operation on the amplified intermediate vector, calculates and outputs the three-dimensional dynamic weights.

[0072] The calculation process of the dynamic weight allocation module has been optimized. First, unit normalization is performed. Before calculating the exponential function, each component of the current three-dimensional deviation vector is relativized. This means dividing the melt temperature deviation value by the allowable temperature fluctuation range (e.g., 20°C), the bulk density deviation value by the allowable bulk density fluctuation range (e.g., 100 g / L), and the specific mechanical energy deviation value by the allowable specific mechanical energy fluctuation range (e.g., 50 kJ / kg). This normalization process ensures that the deviations of three different dimensions are mapped to the same relative scale, so that the subsequent exponential amplification can truly reflect the relative deviation of each target, rather than being distorted by the dimensions. Subsequently, the nonlinear amplification unit performs exponential function processing on the normalized deviation value (with a base of e and a power of the normalized deviation value multiplied by a scaling factor of 2.0) to obtain the amplified intermediate vector. The scaling factor of 2.0 is chosen based on the following principle: it ensures that at a moderate deviation (approximately 0.5 after normalization), the exponential output is approximately e^1.0 ≈ 2.72, which, after Softmax, yields a weight of approximately 33%. This highlights the importance of the target without completely ignoring other targets. Finally, the normalization unit performs a Softmax operation on the amplified intermediate vector to obtain the three-dimensional dynamic weights. This improved process ensures that the weight allocation considers both the physical characteristics and acceptable error range of different quality attributes, and can be dynamically adjusted according to the current actual deviation magnitude.

[0073] The calculation process of the Softmax normalization operation is as follows: The normalization unit first calculates the sum of all elements in the amplified intermediate vector (i.e., the sum of the three amplified deviation values); then, the normalization unit divides each individual element in the intermediate vector (i.e., the amplified melt temperature deviation, bulk density deviation, and specific mechanical energy deviation) by the calculated sum; finally, the normalization unit outputs these three calculation results as the three-dimensional dynamic weights (i.e., melt temperature weight, bulk density weight, and specific mechanical energy weight), which are then used to calculate the weighted Euclidean distance.

[0074] By employing an exponential function and Softmax normalization, the dynamic weight allocation module achieves a non-linear amplification, ensuring that when a deviation is slightly larger than others, its corresponding weight increases disproportionately and exponentially, dominating the final weight allocation. This allows the entire goal programming algorithm to focus its computational resources more quickly and decisively on the most pressing control problem, rather than distributing its attention evenly among multiple deviations, thereby improving the convergence speed and decision-making efficiency of the adaptive response.

[0075] Furthermore, the first control loop of the three control loops is configured as a Smith predictor controller structure. The Smith predictor controller structure includes a main controller and an internal dynamic model. The parameters of the internal dynamic model are locally identified and updated by the process mapping model at the optimal operating point. The main controller is configured as a PI controller. The internal dynamic model is trained to characterize the high inertia and pure time delay response characteristics of the melt temperature to the first intermediate control signal. The main controller calculates the deviation between the target melt temperature value and the corrected feedback signal, and calculates the first intermediate control signal based on the deviation. The first intermediate control signal is sent to the feedforward decoupling matrix. The corrected feedback signal is calculated as follows: the first intermediate control signal is input into the internal dynamic model to generate a predicted melt temperature; the error correction term is added to the predicted melt temperature to obtain the corrected feedback signal. The error correction term is obtained by calculating the difference between the real-time melt temperature data and the simulated output of the internal dynamic model after a delay.

[0076] Specifically, in one embodiment of this application, the first control loop is configured as a Smith predictor controller structure, the purpose of which is to overcome the inherent high inertia (slow response) and pure hysteresis (response delay) characteristics of melt temperature. The implementation process of the controller includes two stages: offline model calibration and online real-time control.

[0077] First, an internal dynamic model is acquired offline. This model is calibrated through an open-loop step test. In a preferred embodiment, the system operates under stable conditions. At this point, the output of the first control loop (i.e., the first intermediate control signal) is manually subjected to a step change (e.g., from 20% to 30%). The control system then records the complete response curve of the melt temperature caused by this step input. Due to the high inertia and pure time delay characteristics of thermodynamics, this response curve will exhibit an S-shaped curve (i.e., it begins to rise slowly after a period of pure time delay and eventually reaches a new steady-state value). Subsequently, this S-shaped response curve is fitted to a standard first-order plus pure time delay process model. This fitting process calculates three key parameters: process gain (Kp), time constant (Kp), and time delay time constant (Kp). The calibrated first-order plus pure time delay process model, defined by τ (representing high inertia) and θ (representing response delay), is solidified as an internal dynamic model. To ensure the accuracy of the internal dynamic model, the adaptive update mechanism for the model parameters is as follows: When the multi-objective optimization agent calculates a new optimal operating point (which corresponds to the optimal control point and is linearized on the process mapping model), the control system retrieves the local linearization result of the process mapping model near that optimal operating point and uses this local model parameter (e.g., new process gain and time constant) to update the parameters of the Smith predictor controller's internal dynamic model online. This ensures that the internal model always matches the actual process dynamics at the optimal operating point under the current raw material characteristics.

[0078] During the online real-time control phase, the Smith predictor controller executes its calculation process in each control cycle (i.e., in the fast time-base tracking control cycle). The main controller (a PI controller) receives the melt temperature target value output by the multi-objective optimization agent. The main controller calculates the deviation (i.e., error) between the melt temperature target value and a corrected feedback signal. Based on this deviation, the main controller calculates and outputs the first intermediate control signal, which is immediately sent to the feedforward decoupling matrix.

[0079] Meanwhile, the Smith predictor controller performs parallel feedback correction calculations to generate a corrected feedback signal for use in the next cycle. The calculation process is as follows: the first intermediate control signal (i.e., the output of the main controller) is simultaneously input into the internal dynamic model, and the pure lag part in the model is removed, thereby generating a delay-free predicted melt temperature.

[0080] In a parallel computation, the simulation output of the internal dynamic model (including a pure time-lag component) is subtracted from the real-time melt temperature data (from a melt temperature probe) to calculate an error correction term that represents the difference between the model prediction and physical reality.

[0081] Finally, the error correction term is added to the delayed predicted melt temperature, and the result of this addition is the corrected feedback signal, which is sent back to the main controller to calculate the deviation for the next control cycle.

[0082] By configuring the first control loop as a Smith predictor controller architecture, and utilizing its internal dynamic model to actively compensate for the inherent high inertia and pure time delay of the melt temperature data, the problem of control oscillation and large overshoot caused by integral saturation in traditional PI controllers when facing long time delays is solved. This allows the main PI controller to operate safely with higher gain, thereby significantly improving the response speed and tracking accuracy of the melt temperature control loop without sacrificing system stability.

[0083] Furthermore, the second control loop of the three control loops is configured as a model predictive controller; the controller calculates the current deviation between the target packing density value and the real-time packing density data; the controller retrieves a set of hard constraints, which are defined as the physical operating limits of the equipment; the controller uses a finite impulse response dynamic process model to perform optimization calculations in the optimization time domain, and the dynamic process model is trained to characterize the response of the packing density to the second intermediate control signal; the optimization calculation uses an objective function to find the optimal internal control sequence while discarding all candidate control sequences that would trigger the hard constraints; the objective function is specifically calculated by: calculating the prediction deviation between the future packing density response and the target packing density value, and applying an asymmetric penalty term to the prediction deviation; the asymmetric penalty term is configured to: apply a penalty weight for packing density predictions that are higher than the target value. A penalty weight is applied to cases where the predicted packing density is lower than the target value. ;in, and It is a positive real number, and , The relative size can be configured according to the product's quality preference; the controller determines the optimal internal control sequence that minimizes the total cost calculated by the objective function within the optimization time domain, outputs the first control action of the optimal internal control sequence, and normalizes the action as the second intermediate control signal.

[0084] In one specific embodiment of this application, the second control loop is configured as a model predictive controller (MPC) for the purpose of proactively controlling the packing density and avoiding quality defects (i.e., overly hard particles). The implementation process of the controller includes two stages: offline model calibration and online real-time control.

[0085] First, those skilled in the art must obtain the finite impulse response (FIR) dynamic process model in an offline phase. This model is trained to characterize the dynamic response of the packing density to the second intermediate control signal (i.e., the output of the MPC). The model is calibrated by an open-loop impulse test: under a stable operating condition, the second intermediate control signal (i.e., the signal sent to the feedforward decoupling matrix) is subjected to a brief impulse change, and the control system then records the complete response curve of the packing density (from the machine vision system) caused by the impulse input. The coefficients of the finite impulse response (FIR) dynamic process model are determined by discretely sampling the response curve.

[0086] During the online real-time control phase, the model predictive controller executes its optimization calculation process in each control cycle (i.e., in the fast time-base tracking control cycle). The controller first receives the target value of the packing density output by the multi-objective optimization agent and obtains real-time packing density data from the data acquisition to calculate the current deviation. Subsequently, the controller retrieves a set of hard constraints. In this embodiment, to ensure the smoothness of control, the hard constraints are defined as the limits on the maximum rate of change of the second intermediate control signal itself (e.g., a change of no more than 5% per control cycle) and the output amplitude (e.g., 0% to 100%). Then, the controller uses the calibrated finite impulse response dynamic process model to perform forward optimization calculations within an optimization time domain (the next 60 control cycles). This calculation iteratively searches for an optimal internal control sequence through an objective function. During the search process, the controller discards all candidate control sequences that would trigger the hard constraints (i.e., the rate of change or amplitude limits).

[0087] The objective function is calculated as follows: The controller calculates the prediction deviation between the future bulk density response (predicted by the FIR model) and the target bulk density value. Subsequently, the controller applies a penalty to the prediction deviation using the asymmetric penalty term, which is configured to apply a high penalty weight for all time points where the prediction deviation is positive (i.e., the predicted bulk density is higher than the target value, resulting in overly hard particles). Furthermore, a low penalty weight is applied to all time points where the prediction deviation is negative (i.e., the predicted packing density is lower than the target value); .in and It is a positive real number. and The relative size can be configured according to the product's quality preferences, and (For example, to prioritize punishing defects with excessively hard particles, it can be configured) > ,like =0.8, =0.2).

[0088] The specific configuration of the asymmetric penalty term is as follows: for all future time steps k (k=1, 2, ..., 60), if the predicted packing density is greater than the target value, a penalty weight is applied to the prediction deviation. =5.0; If the predicted packing density is less than the target value, a penalty weight is applied to the deviation. =1.0. Weighting ratio : The 5:1 ratio is based on the following process principle: excessively high bulk density (i.e., overly hard pellets) will severely affect a pet's chewing experience and digestion, and must be strictly avoided; while slightly lower bulk density (i.e., slightly softer pellets), although not ideal, is within an acceptable range. This weighting ratio remains constant across all ingredient types and does not change dynamically over time.

[0089] Finally, the controller determines the optimal internal control sequence that minimizes the total cost calculated by the objective function (i.e., the weighted total deviation) within the optimization time domain. The controller outputs only the first control action of the optimal internal control sequence and normalizes this action (a value between 0% and 100%) (e.g., converts it to a floating-point number between 0.0 and 1.0) as the second intermediate control signal to the feedforward decoupling matrix.

[0090] The weighting ratio of the asymmetric penalty item is set based on the process principle: (1) In the pet food industry, according to consumer feedback and product test data, overly hard granules (usually corresponding to a bulk density > 420 g / L) will cause the following problems: difficult to chew, easy to cause wear and tear on pet teeth, and reduced digestibility and absorption rate. This is a serious quality defect and the market return rate is high; (2) Overly soft granules (usually corresponding to a bulk density < 350 g / L) are not ideal, but will only cause slight deviations in quality indicators and will not cause market complaints. The product can still be sold; (3) Based on the above quality risk assessment, the weighting ratio is set as follows: : The standard weighting ratio is set to 5:1. However, to improve the versatility and robustness of the solution, the system allows technicians to adjust the weighting ratio according to specific pet food formulations and processing requirements. Specifically, the system reserves a configuration parameter during deployment: a weighting ratio configuration factor α, where... =5×α, =1×α. Technicians can adjust α between 0.5 and 2.0 based on actual quality data and market feedback to adapt to the requirements of different product lines. In addition, the system records the bulk density distribution of each production batch and the corresponding product feedback, forming a database that can be used to periodically review and optimize the rationality of the weighting ratio.

[0091] By configuring the second control loop as a model predictive controller, control actions can be planned proactively within an optimization time domain, and the physical operating limits (i.e., hard constraints) of the equipment can be explicitly addressed, ensuring operational safety. More importantly, by utilizing the bouncing loss function to configure an asymmetric cost function, the controller can prioritize the most critical process deviations, causing the system to disproportionately penalize control paths that lead to a packing density higher than the target value during optimization, thereby ensuring proactive avoidance of critical quality defects.

[0092] Furthermore, the third control loop of the three control loops is configured as a PI controller; the PI controller is configured with a fast response parameter to match the fast time base and low inertia mechanical dynamic characteristics of the specific mechanical energy with high bandwidth; the PI controller calculates the instantaneous deviation between the target value of the specific mechanical energy and the real-time specific mechanical energy data; based on the instantaneous deviation, it calculates and outputs a third intermediate control signal at high speed, and the third intermediate control signal is sent to the feedforward decoupling matrix.

[0093] In one specific embodiment of this application, the third control loop is configured as a PI controller. The purpose of this configuration is to control the specific mechanical energy, which is a fast time-base, low-inertia mechanical dynamic characteristic. The parameter tuning of the PI controller must match the fast time-base characteristic of the specific mechanical energy. During the system commissioning phase, the dynamic characteristics of the loop are obtained through open-loop or closed-loop testing (e.g., the Ziegler-Nichols method), and a high proportional gain and a low integral time are configured for the PI controller accordingly. The high proportional gain ensures that the controller has a high-speed, aggressive response to the instantaneous deviation of the SME (e.g., changes in motor load caused by instantaneous fluctuations in material viscosity), while the low integral time ensures that the integral term can quickly eliminate steady-state errors, so that the SME is accurately maintained at its target value.

[0094] During the online real-time control phase, the PI controller executes its calculation process in each control cycle (i.e., the fast time-base tracking control loop). The PI controller first receives the specific mechanical energy target value (from the target combination) output by the multi-objective optimization agent and obtains real-time specific mechanical energy data (from the calculation of motor load) from the data acquisition. Then, the PI controller calculates the instantaneous deviation between the specific mechanical energy target value and the real-time specific mechanical energy data. Finally, based on this instantaneous deviation and applying its configured high proportional gain and low integral time parameters, the PI controller calculates and outputs the third intermediate control signal at high speed. The third intermediate control signal (e.g., a normalized value between 0% and 100%) is immediately sent to the feedforward decoupling matrix as one of its three input vectors.

[0095] By configuring the third control loop as a high-gain, fast-time-base PI controller, this application matches the dynamic response of the loop with the inherent physical characteristics of "fast time-base and low inertia" of specific mechanical energy. This enables the control loop to quickly and accurately suppress high-frequency instantaneous deviations in specific mechanical energy caused by physical disturbances, ensuring the stability of this key mass attribute, specific mechanical energy, on a fast time-base. This provides crucial support for the rapid and coordinated response of the entire decoupled system.

[0096] Furthermore, the steps for performing compensation calculations using the feedforward control decoupling matrix specifically include the following operational flow: Figure 3 As shown: The dynamic coupling model built into the feedforward control decoupling matrix is ​​defined as a 3x3 process gain matrix. The parameters of the matrix are dynamically generated and updated online by linearizing the model near the optimal control point after the process mapping model calculates the optimal control point. The feedforward control decoupling matrix checks the condition number of the 3x3 process gain matrix. If the condition number is within a preset range, the inverse matrix of the 3x3 process gain matrix is ​​calculated. If the condition number exceeds the preset range, the pseudo-inverse matrix of the 3x3 process gain matrix is ​​calculated to obtain the 3x3 decoupling matrix. The feedforward control decoupling matrix performs matrix multiplication with the input vector containing the first, second, and third intermediate control signals. The feedforward control decoupling matrix outputs the result of the matrix multiplication as a cooperative control signal, which simultaneously controls the screw speed of the thermomechanical reactor, the water injection volume of the material pretreatment unit, and the die temperature of the thermomechanical reactor.

[0097] In one specific embodiment of this application, the feedforward control decoupling matrix is ​​the core execution module of the fast time-base tracking control loop. The dynamic coupling model built into the feedforward control decoupling matrix is ​​defined as a 3x3 process gain matrix. The parameters of this matrix are not fixed values, but are dynamically generated and updated online by the process mapping model after the multi-objective optimization agent calculates the optimal control point by linearizing the model near the corresponding optimal operation point.

[0098] The model linearization operation employs a numerical method, by applying small virtual perturbations (e.g., at the optimal operating point) A 1% change is applied to each manipulable variable, and then the steady-state change of the key quality properties of the output is calculated using a process mapping model. Each element of the process gain matrix is ​​obtained by calculating the steady-state ratio of the output change to the input disturbance. This mechanism ensures that the matrix parameters are always matched with the current raw material characteristics and the actual coupling dynamics at the operating point.

[0099] Upon receiving a new 3x3 process gain matrix, the feedforward control decoupling matrix immediately checks the condition number of the matrix to assess the degree of system coupling and the numerical stability of decoupling. If the condition number is within a preset safety range, the decoupling matrix is ​​deemed well-conditioned, and the inverse of the 3x3 process gain matrix is ​​calculated as the decoupling matrix. If the condition number exceeds a preset range (indicating that the matrix is ​​close to singular, the system is highly coupled, or there is an uncontrollable risk), the decoupling matrix is ​​deemed ill-conditioned. In this case, algorithms such as singular value decomposition are used to calculate the pseudo-inverse of the 3x3 process gain matrix, which is then used as the decoupling matrix. The use of the pseudo-inverse matrix ensures that even when the system faces numerical instability, it can still output a robust control compensation that is optimal in the least squares sense.

[0100] The feedforward control decoupling matrix performs the following condition number check process: First, the condition number cond(G) of the 3x3 process gain matrix G is calculated, which is defined as the ratio of the maximum singular value to the minimum singular value of the matrix. The preset safe range of the condition number is defined as [1, 1000]. When 1≤cond(G)≤1000, the matrix is ​​determined to be a well-state matrix, and the system calculates its inverse matrix as the decoupling matrix. A condition number of 1 indicates that the matrix is ​​an orthogonal matrix (optimal state). A condition number of 1000 indicates that the matrix is ​​slightly ill-conditioned, but the direct inverse matrix method can still be used. When cond(G)>1000, the matrix is ​​determined to be an ill-conditioned matrix, indicating that the coupling relationship between a certain manipulated variable and a certain output variable of the system is extremely complex or there is a risk of near uncontrollability. At this time, the system adopts the pseudo-inverse matrix method, which calculates the pseudo-inverse of the matrix through singular value decomposition. To prevent numerical instability, the pseudo-inverse calculation adopts Tikhonov regularization, that is, all values ​​less than the maximum singular value are multiplied by 10. -6 The singular values ​​are truncated and set to zero. This method ensures that even when the system is highly coupled or close to singular, it can still output a robust decoupling compensation signal that is optimal in the least squares sense.

[0101] During the online real-time control phase, the feedforward control decoupling matrix performs compensation calculations. This matrix receives first, second, and third intermediate control signals from three parallel control loops and combines them into an input vector. The calculated 3x3 decoupling matrix is ​​then multiplied by this input vector. The result of this multiplication is the coordinated control signal, which is immediately sent to the physical equipment to simultaneously control the screw speed of the thermomechanical reactor, the water injection rate of the material pretreatment unit, and the die temperature of the thermomechanical reactor, thereby proactively eliminating cross-interference between the three control loops.

[0102] By utilizing the inverse of the process gain matrix to perform real-time compensation calculations, the decoupling matrix of this application can proactively counteract the cross-interference between manipulated variables (such as screw speed) and three key quality attributes (such as bulk density), ensuring that the three parallel control loops do not cause significant disturbances to other loops when controlling their respective targets, thereby ensuring the synergy and ultimate control stability of the entire multivariable control system.

[0103] By adaptively calculating an optimal target combination under the current physical constraints when a change in raw material type is detected, and utilizing a multivariable decoupled control architecture consisting of three heterogeneous control loops and a feedforward control decoupling matrix, this application can proactively counteract the physical coupling interference between three key quality attributes and enable the system to quickly and stably track the optimal target combination. This solves the problems of control mismatch, system oscillation, and batch-to-batch inconsistency in product quality caused by raw material fluctuations, ensuring the robustness and consistency of production.

[0104] Example 2:

[0105] In one specific embodiment of this application, the control system of the pet food production equipment is operating stably, and the system is producing senior cat food using standard formula A, which has been calibrated as the baseline raw material. The fast time-base tracking control loop, i.e., the three parallel control loops and the feedforward control decoupling matrix, is actively running in each control cycle (e.g., every 200 milliseconds). The slow time-base supervised optimization loop, i.e., the calculation of the three-dimensional control feasible domain boundary, is in standby mode because it has not detected any change in the raw material type. At this time, the three control loops are precisely tracking an optimal control point for standard formula A, i.e., the optimal target combination A, which was previously calculated by the slow time-base loop. The target is a melt temperature of 140°C, a bulk density of 380 g / L, and a specific mechanical energy of 110 kJ / kg.

[0106] At T=10 minutes, the operator switches the raw material silo from standard formulation A to high-fiber formulation B. At T=10 minutes and 30 seconds, the high-fiber B raw material flows into the material inlet of the thermomechanical reactor. The near-infrared spectral sensor immediately scans the current spectral data of the new material, and the control system immediately executes the raw material type change monitoring process: the chemometric model (a PCA model) calculates the current spectral data into a current spectral feature vector. The Mahalanobis distance algorithm calculates the statistical distance between the current spectral feature vector and the stored reference spectral feature vector (A), obtaining a change scalar value, such as 5.8, which exceeds the preset change threshold of 3.0.

[0107] The aforementioned event immediately triggered the supervised optimization calculation. The system now knows the raw material type is B (high fiber). The system immediately executes the calculation process of the dynamic control model: the raw material feature encoder receives B high fiber data and outputs its static raw material feature vector. The virtual mesh generator of the batch prediction submodule creates a three-dimensional virtual process mesh containing combinations of screw speed, water injection volume, and die temperature. The system broadcasts and stitches the static raw material feature vector of B to every point in this mesh. The core prediction network performs high-speed prediction on this batch tensor, outputting a completely new three-dimensional prediction point cloud. Due to the higher melt viscosity of the high fiber raw material, the physical space represented by this new point cloud is completely different from that of A. The boundary extraction submodule (an Alpha-Shape computation unit) processes this new point cloud and outputs a completely new three-dimensional controllable domain boundary for B. The boundary extraction submodule uses a combination of an improved Alpha-Shape algorithm and physical constraint filtering. First, the Alpha-Shape computation unit performs standard non-convex hull extraction on the prediction point cloud to obtain a preliminary boundary surface. Subsequently, the submodule performs the following physical constraint filtering steps: (1) Temperature rationality check: filter out any boundary points where the predicted melt temperature exceeds 180°C, because this temperature range exceeds the thermal decomposition temperature of pet food raw materials and is physically infeasible; (2) Bulk density rationality check: filter out boundary points where the predicted bulk density is below 200 g / L or above 600 g / L, because these values ​​exceed the loose and compaction limits that the process equipment can achieve, respectively; (3) Specific mechanical energy rationality check: filter out boundary points where the predicted specific mechanical energy is negative or exceeds 500 kJ / kg, because negative energy consumption is physically impossible and excessive energy consumption will lead to equipment overload; (4) Multivariate consistency check: for each boundary candidate point, verify that its distance to the nearest internal (non-boundary) sampling point is not less than 1% of the range of manipulated variables to filter out false boundary points that may be caused by model prediction noise. After the above physical constraint filtering, the submodule obtains a more reliable and physically effective three-dimensional control feasible domain boundary, ensuring that the subsequent target planning algorithm will not calculate control points that are physically impossible to achieve.

[0108] The new three-dimensional control feasible domain boundary is immediately transmitted to the multi-objective optimization agent, and the system then executes the calculation process of the objective planning algorithm: the agent obtains the fixed ideal target point, which is set as the coordinates corresponding to maximizing melt temperature, minimizing bulk density, and minimizing specific mechanical energy. The agent calculates the current three-dimensional deviation vector between the ideal target point and the current real-time data (still 140°C, 380 g / L, 110kJ / kg). The dynamic weight allocation module (a Softmax normalization unit) calculates the three-dimensional dynamic weights based on the deviation vector. The agent performs discretization sampling on the new B-dimensional control feasible domain boundary. The agent uses the dynamic weights to calculate the weighted Euclidean distance and finds the optimal compromise point on the new boundary.

[0109] The slow time-base calculation is completed, and a new optimal target combination, namely optimal target combination B, is output. Its target values ​​are melt temperature 142°C, bulk density 395 g / L, and specific mechanical energy 125 kJ / kg. Optimal target combination B is immediately sent to the three parallel control loops, i.e., the fast time-base tracking control loop. Simultaneously, the system uses a process mapping model to perform local linearization at the optimal operating point and dynamically updates the 3x3 process gain matrix to the decoupling matrix. The first control loop (Smith predictor controller) receives the new melt temperature target value of 142°C. Its internal dynamic model has been updated online according to the latest 3x3 process gain matrix parameters. It calculates the deviation between the new target and the corrected feedback signal and outputs a new first intermediate control signal. The second control loop (model predictor controller) receives the new bulk density target value of 395 g / L. It utilizes its finite impulse response model and configures asymmetric penalty weights (…). and The objective function is used to calculate a new second intermediate control signal, which preferentially avoids the defect of excessively hard particles; the third control loop (PI controller) receives the new specific mechanical energy target value of 125kJ / kg, and uses its high-gain PI to calculate a new third intermediate control signal.

[0110] The method for local linearization of the process mapping model near the optimal operating point is as follows: After the multi-objective optimization agent determines the optimal control point, the system applies a small virtual perturbation around the optimal operating point. Specifically, a ±5% perturbation is applied to the first intermediate control signal, while keeping other intermediate control signals unchanged. The resulting change in melt temperature response is recorded. For example, if the baseline first intermediate control signal is 30%, it is set to 28.5% and 31.5% respectively. After stabilization (usually requiring 30 seconds), the corresponding melt temperature output is recorded. Based on the input and output data of these two virtual experiments, a new process gain is calculated. Similarly, the dynamic response of the melt temperature is observed through a brief pulse input, and the time constant and pure lag time are recalculated using a curve fitting method. The parameter update adopts a first-order low-pass filtering mechanism. This filtering coefficient ensures a smooth transition of parameters and avoids controller instability caused by drastic parameter fluctuations. The parameter update interval is triggered once each time the agent outputs a new optimal control point, and in principle, it should not exceed once every 2 minutes to balance adaptability and computational efficiency.

[0111] The feedforward control decoupling matrix receives three new intermediate control signals in each control cycle. The decoupling matrix first checks the condition number of the updated 3x3 process gain matrix and calculates the corresponding inverse or pseudo-inverse matrix based on the result. Then, it performs matrix multiplication with the three intermediate control signals. The decoupling matrix immediately outputs a coordinated control signal to simultaneously control the three actuators. To handle high-fiber raw materials and achieve the optimal target combination B, it may increase the screw speed (to achieve a new specific mechanical energy target), slightly increase the water injection rate (to help control a new bulk density target), and slightly decrease the die temperature (to precisely achieve a new melt temperature target) (due to the increased shear heat of specific mechanical energy).

[0112] Because the feedforward control decoupling matrix proactively compensates for the physical interference between the three actions, the system smoothly (without oscillation or overshoot) transitions from the optimal target combination A to the optimal target combination B. As a result, the adaptive control method of this application automatically calculates and executes a completely new optimal process window for the new raw material within seconds of detecting a change in the raw material, thereby solving the problem of batch-to-batch quality inconsistency caused by raw material variability.

[0113] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An adaptive control method for pet food production equipment, characterized in that, include: Obtain the raw material type, key quality attributes, and control variables; Key quality attributes include the melt temperature and bulk density of the finished product, and specific mechanical energy; control variables include die temperature, screw speed, and water injection volume. When a change in raw material type is detected, a process mapping model is constructed using the raw material type and control variables. The three-dimensional control feasible domain boundary between key quality attributes is calculated. The process of obtaining the three-dimensional control feasible domain boundary involves performing boundary fitting on the three-dimensional predicted point cloud generated by the process mapping model, identifying and connecting the outermost non-convex boundary points of the point cloud, and finally outputting the three-dimensional control feasible domain boundary. The feasible domain boundary is used as a control constraint and input into the multi-objective optimization agent along with the real-time key quality attributes. The objective planning algorithm is executed, and the optimal control point is calculated using the feasible domain boundary combined with the real-time key quality attributes. The optimal operation point corresponding to the optimal control point is determined, and the process mapping model is used to linearize at the optimal operation point to generate an updated dynamic coupling model. The optimal control point is input into three control loops; the three control loops calculate the first, second, and third intermediate control signals, respectively; the first control loop of the three control loops is configured as a Smith predictor controller structure; the Smith predictor controller structure includes a main controller and an internal dynamic model, the parameters of which are locally identified and updated by the process mapping model at the optimal operating point; the main controller is configured as a PI controller, and the internal dynamic model is trained to characterize the high inertia and pure time delay response characteristics of the melt temperature to the first intermediate control signal; the main controller calculates the deviation between the target melt temperature value and the corrected feedback signal, and calculates the first intermediate control signal based on the deviation, the second, and the third intermediate control signals, and so on. An intermediate control signal is sent to the feedforward decoupling matrix; the corrected feedback signal is calculated as follows: the first intermediate control signal is input into the internal dynamic model to generate a predicted melt temperature; the error correction term is added to the predicted melt temperature to obtain the corrected feedback signal; the error correction term is obtained by calculating the difference between the real-time melt temperature data and the simulated output of the internal dynamic model after a delay; the second control loop of the three control loops is configured as a model prediction controller; the controller calculates the current deviation between the target packing density value and the real-time packing density data; the controller retrieves a set of hard constraints, which are defined as the physical operating limits of the equipment; The controller utilizes a finite impulse response dynamic process model to perform optimization calculations in the optimization time domain. This dynamic process model is trained to characterize the response of the packing density to a second intermediate control signal. The optimization calculation uses an objective function to find the optimal internal control sequence while discarding all candidate control sequences that would trigger hard constraints. Specifically, the objective function is calculated by: calculating the prediction deviation between the future packing density response and the target packing density value, and applying an asymmetric penalty term to the prediction deviation. The asymmetric penalty term is configured to apply a penalty weight to packing density predictions that are higher than the target value. A penalty weight is applied to cases where the predicted packing density is lower than the target value. ;in, and It is a positive real number, and , The relative size can be configured according to the product's quality preference; the controller determines the optimal internal control sequence that minimizes the total cost calculated by the objective function within the optimization time domain, outputs the first control action of the optimal internal control sequence, and normalizes the action as the second intermediate control signal; the third control loop of the three control loops is configured as a PI controller; the PI controller is configured with fast response parameters to match the fast time base and low inertia mechanical dynamic characteristics of the specific mechanical energy with high bandwidth; the PI controller calculates the instantaneous deviation between the specific mechanical energy target value and the real-time specific mechanical energy data; based on the instantaneous deviation, it calculates and outputs the third intermediate control signal at high speed, and the third intermediate control signal is sent to the feedforward decoupling matrix; Three intermediate control signals are input into the feedforward control decoupling matrix; the control decoupling matrix receives and updates the dynamic coupling model, performs compensation calculations on the three intermediate control signals, outputs a coordinated control signal, and adjusts the equipment parameters.

2. The adaptive control method for pet food production equipment according to claim 1, characterized in that, The process mapping model include: The raw material feature encoder receives raw material type data as input, processes it using an embedding layer and a fully connected network, and generates a static raw material feature vector. The batch prediction submodule, which contains a virtual grid generator and a core prediction network, is used for: The system receives the operating range of the screw speed, the water injection volume, and the die temperature, and creates a three-dimensional virtual process mesh using a virtual mesh generator. It then broadcasts and splices the static raw material feature vector with each data point in the three-dimensional virtual process mesh to form a batched combined input tensor. This batched combined input tensor is then input into the core prediction network, outputting a three-dimensional prediction point cloud. Each point in the prediction point cloud contains a predicted melt temperature, bulk density, and specific mechanical energy coordinate pair. The boundary extraction submodule receives the 3D predicted point cloud as input and contains an Alpha-Shape calculation unit. It performs boundary fitting on the predicted point cloud, identifies and connects the outermost non-convex boundary points of the point cloud, and finally outputs the 3D control feasible domain boundary.

3. The adaptive control method for pet food production equipment according to claim 1, characterized in that, The process of calculating the optimal control point includes: defining a fixed ideal target point located outside the boundary of the three-dimensional control feasible domain within the agent; the ideal target point is set as the coordinates corresponding to the target values ​​of melt temperature, bulk density, and specific mechanical energy; the agent calculates the current three-dimensional deviation vector between the ideal target point and the current real-time melt temperature, bulk density, and specific mechanical energy data; the current three-dimensional deviation vector is input to a dynamic weight allocation module, which calculates and outputs a set of three-dimensional dynamic weights based on the magnitude of the deviation vector; the agent discretizes and samples the calculated boundary of the three-dimensional control feasible domain to obtain a set of boundary candidate points; iteratively calculates the three-dimensional weighted Euclidean distance between the ideal target point and each point in the set of boundary candidate points, the Euclidean distance being calculated using the three-dimensional dynamic weights; the agent determines the boundary candidate point with the smallest weighted Euclidean distance and outputs the boundary candidate point as the optimal compromise point, the coordinates corresponding to the optimal compromise point being the optimal control point.

4. The adaptive control method for pet food production equipment according to claim 3, characterized in that, The dynamic weight allocation module calculates and outputs the three-dimensional dynamic weights in real time based on the magnitude of the deviation vector, including: relativizing each component of the current three-dimensional deviation vector; inputting the processed current three-dimensional deviation vector into a nonlinear amplification unit; the nonlinear amplification unit uses an exponential function to calculate the melt temperature deviation, bulk density deviation, and specific mechanical energy deviation in the three-dimensional deviation vector to obtain an amplified intermediate vector; inputting the amplified intermediate vector into a normalization unit; and the normalization unit performs a Softmax normalization operation on the amplified intermediate vector, calculates and outputs the three-dimensional dynamic weights.

5. The adaptive control method for pet food production equipment according to claim 1, characterized in that, The specific steps of the feedforward control decoupling matrix performing compensation calculations include: the dynamic coupling model built into the feedforward control decoupling matrix is ​​defined as a 3x3 process gain matrix. The parameters of the matrix are dynamically generated and updated online by linearizing the model near the optimal control point after the process mapping model calculates the optimal control point; the feedforward control decoupling matrix checks the condition number of the 3x3 process gain matrix. If the condition number is within a preset range, the inverse matrix of the 3x3 process gain matrix is ​​calculated; if the condition number exceeds the preset range, the pseudo-inverse matrix of the 3x3 process gain matrix is ​​calculated to obtain the 3x3 decoupling matrix; the feedforward control decoupling matrix performs matrix multiplication with the input vector containing the first, second, and third intermediate control signals; the feedforward control decoupling matrix outputs the result of the matrix multiplication as a cooperative control signal, simultaneously controlling the screw speed of the thermomechanical reactor, the water injection volume of the material pretreatment unit, and the die temperature of the thermomechanical reactor.

Images