Multivariable intelligent predictive control system for drying process of dried konjac rice

Abstract

The application relates to the technical field of industrial automation control, and discloses a multivariable intelligent prediction control system for a konjac rice drying process, which comprises the following modules: a multimodal perception and soft measurement field construction module, which is used for collecting and fusing data to construct a dynamic soft measurement field; a twin dynamics characterization module, which is used for establishing a model and predicting a future material state based on the soft measurement field; a physical field-virtual field mapping and abnormality tracing module, which is used for comparing a real state with a predicted state to diagnose an abnormality root cause; an adaptive reinforcement learning prediction control module, which is used for integrating all information to generate an optimal control action; and a dynamic edge computing and model collaborative deployment module, which provides calculation and resource scheduling for the aforementioned modules. The application adopts a technical scheme in which multimodal perception data and deep learning are fused, a dynamic soft measurement field is constructed, and therefore key states such as a material internal moisture gradient and a surface crust degree can be quantified in real time and nondestructively.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial automation control, in particular to a multivariable intelligent predictive control system for the drying process of dried konjac rice. BACKGROUND

[0002] Dried konjac rice, as a popular healthy food, has an increasing market demand. In industrial production, drying is the core process that determines the final rehydration, taste, shelf life and production cost of the product. An ideal drying process needs to maintain the product quality (such as color, texture and nutritional ingredients) at the optimal level in the shortest possible time with the lowest energy consumption.

[0003] In current industrial practice, the drying process of dried konjac rice has introduced automation control technology. These systems are usually equipped with conventional process sensors such as temperature and humidity, and cooperate with PID (Proportional-Integral-Derivative) controllers or PLCs (Programmable Logic Controllers) to drive heaters, fans and other execution units. Compared with intermittent operation relying entirely on human experience, this automation solution realizes the stable maintenance and programmed control of the drying equipment environment parameters, ensuring that the production process under specific process settings can be accurately reproduced, thereby laying the foundation for large-scale and standardized production.

[0004] However, the inherent limitations of existing technology gradually emerge when pursuing higher levels of refinement and intelligent control. First, the perception dimension of such control systems is indirect, as they monitor the macro environment of the drying chamber rather than the internal state of the material itself. The control system cannot directly know the difference between the internal and surface moisture content of the material, the rate of water migration, or the risk of surface crust that is a precursor to quality deterioration, which leads to lag and blindness in control decisions. Second, the adaptability of existing models is very limited. Once the control parameters are set, they tend to be fixed. When faced with batches of raw materials with natural fluctuations in physical and chemical properties due to different origins and seasons, this fixed control logic cannot maintain optimal performance, leading to unstable product quality between batches. Finally, the system's ability to handle abnormal conditions is weak. Its diagnostic logic usually stays at the level of simple threshold alarms and cannot independently investigate the root cause of abnormalities. Whether it is sensor drift, actuator aging or sudden changes in raw material properties, human intervention is ultimately required, which limits the robustness and autonomy of the production process. SUMMARY

[0005] In view of the deficiencies of the prior art, the present application provides a multivariable intelligent predictive control system for the drying process of dry konjak rice, which solves the technical problem that the drying efficiency, product quality and energy consumption cannot be optimized simultaneously in the prior drying control method for konjak rice because the internal key state of the material cannot be obtained in real time and non-destructively, and the model cannot adapt to batch differences and process abnormalities.

[0006] To achieve the above object, the present application is implemented by the following technical solutions: a multivariable intelligent predictive control system for the drying process of dry konjak rice, comprising:

[0007] A multi-modal perception and soft measurement field construction module is used to collect multi-source heterogeneous data in the drying process of dry konjak rice in real time, and fuse and infer the collected data to construct a dynamic soft measurement field representing the internal state of the material and external environmental parameters;

[0008] A twin dynamics representation module is used to establish a dynamics model based on the real-time state provided by the dynamic soft measurement field and the input control action, and predict the material state change trajectory in the future period of time;

[0009] A physical field-virtual field mapping and abnormality tracing module is used to compare the real material state provided by the dynamic soft measurement field with the material state predicted by the twin dynamics representation module in real time, and identify and output the root cause of the deviation when the deviation exceeds the preset threshold;

[0010] An adaptive reinforcement learning predictive control module is used to integrate the real-time state provided by the dynamic soft measurement field, the material state change trajectory predicted by the twin dynamics representation module, and the root cause output by the physical field-virtual field mapping and abnormality tracing module, make a multi-objective optimization decision through a reinforcement learning algorithm, and generate a set of optimal multivariable control actions;

[0011] A dynamic edge computing and model collaborative deployment module is used to provide a computing framework and resource scheduling for the operation of the multi-modal perception and soft measurement field construction module, the twin dynamics representation module, the physical field-virtual field mapping and abnormality tracing module, and the adaptive reinforcement learning predictive control module.

[0012] As a preferred technical solution, the multi-modal perception and soft measurement field construction module constructs the dynamic soft measurement field by the following steps:

[0013] The multi-source heterogeneous data is collected by non-contact spectral sensors and conventional process sensors, and the multi-source heterogeneous data includes spectral data and conventional sensor data;

[0014] Baseline correction, smoothing filtering and scattering correction are performed on the collected spectral data, and normalization or standardization is performed on the conventional sensor data;

[0015] The spectral data and the conventional sensor data after baseline correction, smoothing filtering, scattering correction and normalization or standardization are input into a deep learning fusion model to infer an internal state vector containing material particle average internal moisture content, surface moisture content, internal and external moisture gradient, surface crust hardening degree and quality related index;

[0016] The internal state vector is integrated with external environmental parameters extracted from the conventional sensor data to form the dynamic soft measurement field.

[0017] As a preferred technical solution, the twin kinetic characterization module establishes the kinetic model and performs prediction, including the following steps:

[0018] A hybrid neural network model is constructed by integrating heat and mass transfer physical equations as inductive bias; in the offline training stage, based on the drying data of historical batches, a group of meta-parameters with rapid adaptation ability is trained by a meta-learning algorithm, the meta-learning algorithm can be a model-independent meta-learning algorithm, and the meta-parameters are a group of model initialization parameters;

[0019] In the online running stage, a small amount of initial data of a new batch is used to perform gradient descent based on the meta-parameters to quickly calculate task-specific parameters adapted to the current new batch, and the calculation method is to start from the meta-parameters, use a small amount of initial data of the new batch, calculate the gradient according to the prediction loss function, and update once or more times along the gradient direction to obtain;

[0020] Taking the real state at the current time as the initial condition, using the adapted model and the input control action, the state prediction sequence of multiple steps in the future is iteratively calculated by the autoregressive method.

[0021] As a preferred technical solution, the physical field-virtual field mapping and abnormality tracing module identifies and outputs the root cause, including the following steps:

[0022] The real-time deviation vector between the real state provided by the dynamic soft measurement field and the predicted state predicted by the twin kinetic characterization module is calculated;

[0023] The absolute value of the deviation of the preset key state variable in the real-time deviation vector is compared with the corresponding dynamic threshold value, and when any deviation exceeds the threshold value, an abnormal event is triggered;

[0024] When an abnormal event is triggered, a causal reasoning is performed on a pre-constructed knowledge graph describing the drying process to identify the most likely root cause leading to the current abnormality, the knowledge graph comprising a set of entities including drying equipment component entities, material batch entities and environmental factor entities, and a set of relationships including causal relationships and association relationships between the entities, the causal reasoning being a reasoning path search on the knowledge graph to identify the root cause that best matches the deviation vector pattern triggering the abnormality;

[0025] The identified root cause and corresponding confidence are encoded into a structured diagnostic information vector and output.

[0026] As a preferred technical solution, the adaptive reinforcement learning predictive control module generating the set of optimal multivariate control actions comprises the following steps:

[0027] The current state provided by the dynamic soft measurement field, the future state prediction sequence output by the twin dynamic characterization module, and the diagnostic information vector output by the physical field-virtual field mapping and abnormality tracing module together constitute a state space;

[0028] According to the real-time process state and diagnostic information in the state space, the dynamic weight coefficients of each reward or penalty in a multi-objective reward function are adjusted online, the multi-objective reward function being a weighted sum of a drying efficiency reward term, a product quality reward term, an energy consumption reward term, and an operation constraint penalty term, the adjustment logic of the dynamic weight coefficients being to increase the weight of the product quality reward term when a state variable representing quality risk in the dynamic soft measurement field exceeds a threshold, and to decrease the weight of a reward term in the multi-objective reward function that depends on data from an indicator sensor when the diagnostic information vector indicates a failure of the indicator sensor;

[0029] The state space is input to a trained policy network, and the policy network outputs the current optimal control action vector under the guidance of the adjusted multi-objective reward function.

[0030] As a preferred technical solution, the dynamic edge computing and model collaborative deployment module provides the computing framework and resource scheduling comprising the following steps:

[0031] A distributed heterogeneous computing architecture comprising edge computing units deployed on site and remote cloud computing clusters is adopted;

[0032] A task dynamic offloading scheduling algorithm is executed to dynamically determine whether a computing task is executed on the edge computing unit or the cloud computing cluster according to the real-time requirement and computing complexity of the task, the determination logic being:

[0033] If an expected execution delay of a task is less than a preset real-time delay threshold, and a current available computing resource of the edge computing unit is greater than a required computing resource of the task, the task is allocated to the edge computing unit for execution, otherwise, the task is unloaded to the cloud or the central computing cluster for execution.

[0034] A model library is established in the cloud computing cluster, and pruning and quantization optimization are performed on the models deployed to the edge computing unit, and online hot updating is supported.

[0035] The application provides a multivariable intelligent prediction control system for a dry konjak rice drying process.

[0036] 1、The application adopts a technical scheme of fusing multi-modal sensing data and deep learning, and constructs a dynamic soft measurement field, so that key states such as internal moisture gradient and surface crust degree of materials can be quantified in real time and losslessly, compared with a control mode of the prior art that depends on external environmental parameters or offline sampling inspection, the scheme solves the problems of blind control strategy and uneven product quality caused by the inability to understand the internal real state of materials.

[0037] 2、The application applies a dynamics modeling method based on meta-learning, and gives the twin dynamics characterization module the ability to quickly adapt, which can complete the adaptation to the current material characteristics according to a small amount of real-time data in a short time after the start of a new batch production, changes the condition that model parameters are fixed and difficult to adapt to differences between batches of raw materials in the prior art, and solves the problems of inaccurate prediction and unstable control effect caused by model mismatch.

[0038] 3、The application constructs a closed-loop control loop of abnormal diagnosis and adaptive decision, the system compares the deviation between the real physical field and the twin virtual field in real time, and performs causal reasoning combined with a knowledge graph, so that the root cause of process abnormalities can be traced, compared with a scheme of the prior art that depends on fixed threshold alarm and manual post-diagnosis, the application can autonomously identify the source of abnormalities, and dynamically adjust the reward function of reinforcement learning to change the control priority, and solves the problems of rigid decision-making and poor robustness in complex working conditions. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 It is a system function module structure diagram of the embodiment of the application.

[0040] Figure 2 It is a multi-modal sensing and soft measurement field construction module function and data flow diagram of the embodiment of the application.

[0041] Figure 3 It is a twin dynamics characterization module function and data flow diagram of the embodiment of the application.

[0042] Figure 4 The physical field-virtual field mapping and abnormality tracing module function and data flow diagram of the embodiment of the present application.

[0043] Figure 5 The adaptive reinforcement learning predictive control module function and data flow diagram of the embodiment of the present application.

[0044] Figure 6 The dynamic edge computing and model collaborative deployment module function and data flow diagram of the embodiment of the present application.

[0045] Among them, 10, multi-modal perception and soft measurement field construction module; 20, twin dynamic characterization module; 30, physical field-virtual field mapping and abnormality tracing module; 40, adaptive reinforcement learning predictive control module; 50, dynamic edge computing and model collaborative deployment module. DETAILED DESCRIPTION

[0046] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the specification of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0047] Referring to the drawings Figure 1 , Figure 1 It is a function module structure schematic diagram of a multivariable intelligent predictive control system for a dry konjak rice drying process according to an embodiment of the present application. The present application provides a multivariable intelligent predictive control system for a dry konjak rice drying process, which can include: a multi-modal perception and soft measurement field construction module 10, a twin dynamic characterization module 20, a physical field-virtual field mapping and abnormality tracing module 30, an adaptive reinforcement learning predictive control module 40, and a dynamic edge computing and model collaborative deployment module 50.

[0048] The multi-modal perception and soft measurement field construction module 10 is used to collect multi-source heterogeneous data in the drying process in real time, and fuse and state infer the collected data to construct a dynamic soft measurement field representing the internal state of the material and the external environmental parameters.

[0049] The twin dynamic characterization module 20 has a data input end connected with an output end of the multi-modal perception and soft measurement field construction module 10, and receives a control action sequence from the adaptive reinforcement learning predictive control module 40. The function of the twin dynamic characterization module 20 is to establish a dynamic model that can quickly adapt to the drying characteristics of different batches of materials, and based on the real-time state provided by the soft measurement field and the input control action, to predict the material state change trajectory in the future period of time.

[0050] The physical field-virtual field mapping and anomaly tracing module 30 is connected to the output of the multi-modal perception and soft measurement field construction module 10 and the output of the twin dynamics characterization module 20. The function of the physical field-virtual field mapping and anomaly tracing module 30 is to compare the real material state provided by the multi-modal perception and soft measurement field construction module 10 with the predicted material state provided by the twin dynamics characterization module 20 in real time, and when the deviation between the two exceeds the preset threshold, identify and output the root cause of the deviation through an anomaly tracing algorithm.

[0051] The adaptive reinforcement learning predictive control module 40 is connected to the outputs of the multi-modal perception and soft measurement field construction module 10, the twin dynamics characterization module 20, and the physical field-virtual field mapping and anomaly tracing module 30. The function of the adaptive reinforcement learning predictive control module 40 is to integrate real-time state, predicted trajectory, and anomaly diagnosis information, make multi-objective optimization decisions through a reinforcement learning algorithm with a dynamic reward mechanism, and generate a set of optimal multi-variable control actions.

[0052] The dynamic edge computing and model collaborative deployment module 50 provides computing power support and model management for the multi-modal perception and soft measurement field construction module 10, the twin dynamics characterization module 20, the physical field-virtual field mapping and anomaly tracing module 30, and the adaptive reinforcement learning predictive control module 40 as the computing framework for the operation of the aforementioned modules. The function of the dynamic edge computing and model collaborative deployment module 50 is to perform dynamic offloading scheduling of tasks, lightweight optimization of deployed models, and online hot updating of model versions.

[0053] In system operation, the aforementioned modules work together to form a complete closed-loop control link from multi-modal perception, adaptive modeling, anomaly diagnosis, intelligent decision-making to model deployment.

[0054] The specific implementation of each module in the multi-variable intelligent predictive control system for the drying process of konjac rice will be described in detail below with reference to the accompanying drawings.

[0055] Referring to the accompanying drawings Figure 2 , Figure 2 is a function and data flow diagram of the multi-modal perception and soft measurement field construction module 10 according to an embodiment of the present application. The core function of the multi-modal perception and soft measurement field construction module 10 is to obtain and quantify the internal physical and chemical state of the konjac rice material in the drying process in real time and losslessly, and integrate it with external environmental parameters to form a high-dimensional, dynamic state representation, i.e., a soft measurement field, to provide accurate and comprehensive data input for subsequent prediction and control modules.

[0056] The specific implementation of the multi-modal perception and soft-sensing field construction module 10 includes the following steps:

[0057] 1. Data acquisition and pre-processing: The multi-modal perception and soft-sensing field construction module 10 integrates and processes data from multiple types of sensors.

[0058] One type is a non-contact spectral sensor, which is selected based on the molecular structure characteristics of the core ingredient of konjac flour, glucomannan. This sensor is used to obtain information about the internal chemical composition and microstructure of the material, especially the binding state of glucomannan macromolecules and water molecules (free water, bound water ratio) and their changes under the action of hot air, which is core information that conventional sensors cannot capture. Specifically, it can include a near-infrared (NIR) spectrometer and a hyperspectral imaging (HSI) device deployed inside the drying equipment or observation port. The data collected by the NIR spectrometer is a vector representing the absorbance information of the material at multiple near-infrared wavelength points , which is directly related to the water content inside the material and the conformational changes of the glucomannan chain. The data collected by the HSI device is a data cube containing spectral information (a total of bands) of each pixel point in the spatial dimension . By capturing the optical property changes caused by the dehydration and solidification of glucomannan at specific bands, early and pixel-level warning of the key quality deterioration phenomenon of surface crust is achieved.

[0059] Another type is a conventional process sensor, which is used to obtain macroscopic physical parameters of the drying environment. Specifically, it includes temperature sensors, humidity sensors, anemometers, and online weighing devices, which respectively collect the air temperature , air relative humidity , hot air flow rate , and total mass of the material in the drying cabinet. The collected conventional sensor data is combined into a vector .

[0060] The collected raw data is input to the pre-processing unit. For spectral data and , pre-processing includes but is not limited to baseline correction, Savitzky-Golay smoothing filter to reduce noise, and multivariate scatter correction (MSC) or standard normal variate transformation (SNV) to eliminate physical interference caused by particle size and surface scattering. For conventional sensor data , normalization or standardization processing is performed to eliminate the influence of different physical dimensions.

[0061] 2. Multimodal data fusion and state inference: The pre-processed data is fed into a deep learning fusion model for extracting deep features from heterogeneous data and inferring the internal state of the material. The input of the fusion model is a concatenated feature vector :

[0062] ;

[0063] where and are feature extractors for NIR and HSI data respectively. In one embodiment, the feature extractors can be one-dimensional or three-dimensional convolutional neural networks (CNNs) for automatically learning effective feature representations in spectral or spectral-spatial data.

[0064] The specific implementation of the fusion model can be a Transformer network architecture. This architecture leverages its self-attention mechanism to effectively capture long-range dependencies and complex nonlinear couplings between different sensor data (e.g., spectral absorption peaks at specific wavelengths and ambient temperature). In another embodiment, the fusion model can also be a graph neural network (GNN), where different types of sensor data are treated as nodes in a graph, and their mutual influence relationships are explicitly modeled by learning edge weights between nodes.

[0065] The output of the fusion model is an internal state vector that contains multiple key internal state variables that cannot be directly measured by hardware:

[0066] ;

[0067] where represents the average internal moisture content of the material particles; represents the surface moisture content of the material particles; represents the moisture gradient from the material interior to the surface, which is a key indicator for judging the internal moisture migration rate; represents the degree of surface crust hardening of the material, which is inferred by fusing spectral and image data and directly quantifies the degree of dense inactivated layer formed by glucomannan on the surface; represents a key indicator related to the quality of the final product, such as a prediction indicator directly related to the elastic modulus or water absorption rate of the finished product after rehydration, which is the core commercial value of evaluating the quality of dried konjac rice; represents a transpose operation.

[0068] 3. Soft measurement field construction: Finally, the multi-modal perception and soft measurement field construction module 10 integrates the inferred internal state vector with the external environment parameters extracted from the conventional sensor data (e.g. , , ) to form a comprehensive, dynamic real state soft measurement field :

[0069] ;

[0070] This soft measurement field is output as a complete state vector to the twin dynamics characterization module 20, the physical field-virtual field mapping and anomaly tracing module 30, and the adaptive reinforcement learning predictive control module 40, providing real-time, accurate, and information-rich basis for subsequent analysis, prediction, and decision-making of the entire system. In this way, the multi-modal perception and soft measurement field construction module 10 solves the core technical problem of real-time, non-destructive acquisition of material internal state in the prior art.

[0071] Referring to the accompanying Figure 3 , Figure 3 is a functional and data flow diagram of the twin dynamics characterization module 20 according to an embodiment of the present application. The data input end of the twin dynamics characterization module 20 is connected to the output end of the multi-modal perception and soft measurement field construction module 10, and receives the control action sequence from the adaptive reinforcement learning predictive control module 40. The function of the twin dynamics characterization module 20 is to establish a dynamic model that can quickly adapt to the drying characteristics of different batches of materials according to real-time data, and based on this model, accurately predict the material state change trajectory in the future period of time, providing a model basis for subsequent anomaly tracing and control decision-making.

[0072] The specific implementation of the twin dynamics characterization module 20 includes the following steps:

[0073] 1. Hybrid model architecture with physical information enhancement: The core of the twin dynamics characterization module 20 is a hybrid neural network model , where are the learnable parameters of the model. The architecture design of the model incorporates the heat and mass transfer physical equations for this high-molecular gel material, dry konjac rice, as the inductive bias of the model, to ensure that the model's prediction conforms to the basic physical laws. In an embodiment, the physical constraints incorporated include a material internal moisture diffusion model based on Fick's second law and a surface heat and mass exchange model based on Newton's cooling law.

[0074] This hybrid neural network model a neural network part for learning the nonlinear parameters in the physical equation that are difficult to model accurately, such as the effective moisture diffusion coefficient, which varies with the batch characteristics of the material and the drying stage , and is a nonlinear parameter that varies dramatically with the internal moisture content and temperature and the contraction of the konjac starch gel network, which is the key point that the traditional model cannot accurately describe, and the surface heat and mass transfer coefficient and . The hybrid neural network model takes as input the soft measurement field state at the current time and a control action , and outputs the state prediction value at the next time , and the mathematical relationship is:

[0075] ;

[0076] In this way of combining physical information and data-driven, the hybrid neural network model can have the representation ability of complex nonlinear dynamics while ensuring the interpretability of the material.

[0077] 2. Meta-learning fast self-adaptation for batch differences: To solve the industry-wide problem that the fixed parameter model completely fails due to the significant differences in initial moisture content, particle size, and konjac starch content of different batches of dried konjac rice raw materials due to different origins, seasons, and processing techniques, the twin dynamic characterization module 20 is trained and adapted online using a meta-learning framework. This framework treats each drying batch as an independent learning task .

[0078] In the offline training phase, the module is trained based on a large amount of historical drying data of batches using a meta-learning algorithm (in an embodiment, the model-agnostic meta-learning MAML algorithm). The goal is to learn a set of optimal meta-parameters . The meta-parameters are not specific to any particular batch, but serve as an initialization model with fast adaptation capability.

[0079] In the online running phase, when a new batch of drying process starts, the twin dynamic characterization module 20 starts from the meta-parameters , uses a small amount of initial data of the new batch provided by the multi-modal perception and soft measurement field construction module 10 , and quickly calculates the task-specific parameters suitable for the current new batch by performing a small number of gradient descent steps inside the model . The mathematical expression of the adaptation process is:

[0080] ;

[0081] wherein, is a prediction loss function, in one embodiment a mean squared error loss; is an inner loop learning rate; is a loss function with respect to meta-parameters . Through this mechanism, the twin dynamic characterization module 20 is able to quickly generate a high-precision dynamic model dedicated to the batch in a short time at the beginning of each new batch .

[0082] 3. Multi-step future state prediction: After obtaining the model parameters adapted to the current batch, the twin dynamic characterization module 20 uses the model to perform multi-step future state prediction. The module receives the future H-step control action sequence generated by the adaptive reinforcement learning predictive control module 40.

[0083] With the real state soft measurement field at the current time as the initial condition, the twin dynamic characterization module 20 iteratively calls the model in a self-recurrent manner to sequentially calculate the future H-step state prediction sequence . The prediction calculation formula of the k+1 step is:

[0084] ;

[0085] wherein, the initial value of . The predicted future state sequence will be output to the physical-virtual field mapping and abnormality tracing module 30 and the adaptive reinforcement learning predictive control module 40, providing a basis for the system's abnormal diagnosis and optimization decision.

[0086] Referring to the accompanying Figure 4 , Figure 4 is a functional and data flow diagram of the physical-virtual field mapping and abnormality tracing module 30 according to one embodiment of the present application. The data input end of the physical-virtual field mapping and abnormality tracing module 30 is connected with the output end of the multi-modal perception and soft measurement field construction module 10 and the output end of the twin dynamic characterization module 20, respectively. The function of the physical-virtual field mapping and abnormality tracing module 30 is to monitor and diagnose the running state of the system in real time, identify process abnormalities by comparing the real physical state with the model-predicted virtual state, and trace the root cause of the abnormality, providing a decision basis for the adaptive reinforcement learning predictive control module 40.

[0087] The specific implementation of the physical-virtual field mapping and anomaly root cause tracing module 30 includes the following steps:

[0088] 1. Real-time deviation mapping between real and predicted states: The physical-virtual field mapping and anomaly root cause tracing module 30 continuously and synchronously receives the real state soft measurement field output by the multi-modal perception and soft measurement field construction module 10 , and the predicted state sequence under the same control action output by the twin dynamics characterization module 20. The physical-virtual field mapping and anomaly root cause tracing module 30 extracts the current time prediction value in the predicted state sequence , and calculates the real-time deviation vector between the real state and the predicted state :

[0089] ;

[0090] Where each component of the real-time deviation vector corresponds to the prediction error of each state variable in the soft measurement field .

[0091] The physical-virtual field mapping and anomaly root cause tracing module 30 performs anomaly detection on the preset key state variables (such as internal moisture gradient , surface crust degree ) in the real-time deviation vector . This detection is achieved by comparing the absolute value of the deviation of each key variable with a corresponding dynamic threshold . When the deviation of any key variable satisfies the following condition, an anomaly event is triggered:

[0092] s.t. ;

[0093] The dynamic threshold can be adaptively adjusted according to the statistical distribution of historical operation data (such as the three-sigma criterion) or the process stage.

[0094] 2. Abnormal root cause tracing based on causal reasoning and knowledge graph: When an anomaly event is triggered, when the actual value of the internal moisture gradient is much larger than the model prediction value, it directly indicates that the internal moisture migration rate is much lower than the surface evaporation rate, which is a strong precursor signal of the occurrence of surface crust. The physical-virtual field mapping and anomaly root cause tracing module 30 activates its internal anomaly root cause tracing algorithm. This algorithm performs causal reasoning based on a pre-constructed knowledge graph specifically for the drying process of dry amorphophellus konjac rice. The knowledge graph , where For entities, including dry equipment component entities (e.g. heater, fan, various sensors), material batch entities, environmental factor entities, etc. For relations, including causal relations (e.g. "influence", "cause") and association relations (e.g. "measure", "locate") between entities.

[0095] Anomaly root cause tracing algorithm receives the deviation vector triggering the anomaly and combines the context information in the system operation log, conducts reasoning path search on the knowledge graph to identify the most possible root cause leading to the current anomaly . The reasoning process can be represented as:

[0096] ;

[0097] wherein, is a causal reasoning function, which in one embodiment can be based on Bayesian network for probabilistic reasoning; is the context information.

[0098] Tracing result may point to various root causes, such as: data drift of non-contact spectral sensor, significant deviation of physical and chemical properties of current material batch from historical training data distribution, performance degradation of heater element, or unmodeled step change in external environment temperature and humidity.

[0099] 3. Diagnosis information generation and output: after completing the anomaly root cause tracing, the physical field-virtual field mapping and anomaly tracing module 30 encodes the tracing result and the corresponding confidence into a structured diagnosis information vector . The diagnosis information vector is output to the adaptive reinforcement learning predictive control module 40. The diagnosis information vector provides an additional dimension for the decision-making of the adaptive reinforcement learning predictive control module 40, enabling it to make more robust and safe decisions when formulating control strategies, not only considering the current state and future prediction, but also considering the diagnosed abnormal state currently existing in the system.

[0100] Referring to the accompanying Figure 5 , Figure 5is a functional and data flow diagram of the adaptive reinforcement learning predictive control module 40 according to an embodiment of the present application. The data input of the adaptive reinforcement learning predictive control module 40 is connected to the output of the multi-modal perception and soft-sensing field construction module 10, the twin dynamic representation module 20 and the physical-virtual field mapping and anomaly tracing module 30, respectively. The function of the adaptive reinforcement learning predictive control module 40 is to integrate the system real-time state, future predicted trajectory and process anomaly diagnosis information, make multi-objective optimization decision through a reinforcement learning algorithm with dynamic reward mechanism, and generate a set of optimal multivariable control action sequence.

[0101] The specific implementation of the adaptive reinforcement learning predictive control module 40 includes the following steps:

[0102] 1. Multi-objective optimization reinforcement learning control framework construction: The adaptive reinforcement learning predictive control module 40 models the dry konjac rice drying process as a Markov decision process. The state space , action space and reward function of the process are defined as follows.

[0103] The state space at time is composed of three parts: the current real state soft-sensing field output by the multi-modal perception and soft-sensing field construction module 10; the future H-step state prediction sequence output by the twin dynamic representation module 20; and the current diagnosis information vector output by the physical-virtual field mapping and anomaly tracing module 30. The mathematical representation is:

[0104] ;

[0105] wherein is the length of the prediction time domain.

[0106] The action space is the set of continuous control variables that can be adjusted within each control period, including heating power , fan speed and exhaust valve opening in an embodiment. The mathematical representation is:

[0107] ;

[0108] The reward function is designed as a multi-objective weighted sum to solve the internal contradictions among fast drying, good quality and low energy consumption in the dry konjac rice drying process, and its form is:

[0109] ;

[0110] wherein, is a drying efficiency reward term, is a product quality reward term, which is designed as a negative function of the surface crust hardening degree and the internal moisture gradient , i.e. the system will get a high reward for successfully suppressing crust and maintaining the balance of moisture gradient, is an energy consumption reward term, is an operation constraint penalty term. , , , are dynamic weight coefficients of each reward or penalty term.

[0111] 2. Dynamic reward mechanism fusing real-time state and abnormal diagnosis: To achieve adaptive adjustment of control objectives, the adaptive reinforcement learning predictive control module 40 contains a dynamic reward mechanism adaptive unit. The function of this unit is to adjust the weight coefficients in the reward function online according to real-time process state and abnormal diagnosis information.

[0112] In an embodiment, each dynamic weight coefficient is a function of the current state , and its update logic is:

[0113] ;

[0114] wherein, is a preset mapping function for the th weight.

[0115] Specifically, when the surface crust degree obtained from the soft measurement field exceeds a preset quality risk threshold, the mapping function will increase the value of the product quality reward weight , meaning that the control objective of the system will be dynamically and smoothly switched from efficiency priority to quality priority, sacrificing certain drying rate to avoid irreversible quality damage by reducing hot air temperature or adjusting air speed and other operations, and the mapping function will increase the penalty strength of the penalty term weight related to crust. When the diagnosis information vector received from the physical field-virtual field mapping and abnormal tracing module 30 indicates that a sensor is faulty, the mapping function will reduce the weight of the reward term strongly related to the data of the sensor to enhance the robustness of the control strategy.

[0116] 3. Multivariable control strategy generation and output: the decision-making agent inside the adaptive reinforcement learning predictive control module 40 learns a policy network to approximate the optimal control strategy, where are the learnable parameters of the policy network. In one embodiment, the decision-making agent is trained using the Proximal Policy Optimization (PPO) algorithm. The goal of this algorithm is to maximize the expected cumulative discounted reward :

[0117] ;

[0118] where, is the state-action trajectory generated by the policy interacting with the drying environment, is the total length of the drying process, is the discount factor.

[0119] At each control decision-making time , the adaptive reinforcement learning predictive control module 40 inputs the current state to the trained policy network , and the network outputs the current optimal control action vector :

[0120] ;

[0121] This control action vector is then sent to the underlying actuators of the drying equipment to complete the simultaneous control of multiple variables such as heating power, fan speed, and exhaust valve opening, forming a closed loop.

[0122] Referring to the accompanying Figure 6 , Figure 6 is the function and data flow diagram of the dynamic edge computing and model collaborative deployment module 50 according to one embodiment of the present application. The dynamic edge computing and model collaborative deployment module 50 provides a computing framework and resource scheduling for the operation of the multi-modal perception and soft measurement field construction module 10, the twin dynamics representation module 20, the physical field-virtual field mapping and anomaly tracing module 30, and the adaptive reinforcement learning predictive control module 40.

[0123] The specific implementation of the dynamic edge computing and model collaborative deployment module 50 includes the following steps:

[0124] 1. Heterogeneous computing architecture of edge-cloud collaboration: The dynamic edge computing and model collaborative deployment module 50 adopts a distributed heterogeneous computing architecture, which includes an edge computing unit deployed on the site of the drying equipment and a remote cloud or central computing cluster. The edge computing unit is an embedded computing device integrated with a central processing unit (CPU), a graphics processing unit (GPU), or a field programmable gate array (FPGA), and is configured with data interfaces for connecting various sensors and actuators. The cloud or central computing cluster is composed of high-performance servers with large-scale data storage and complex model training computing capabilities.

[0125] 2. Task dynamic offloading and model optimization for real-time control: The dynamic edge computing and model collaborative deployment module 50 internally executes a task dynamic offloading scheduling algorithm, which dynamically determines the execution location of the computing task based on the real-time requirements, data throughput, and computational complexity of the task. The decision logic is as follows:

[0126] ;

[0127] wherein, is the allocation location of the task; is the expected delay of executing the task; is a preset real-time delay threshold; is the current available computing resource of the edge computing unit; is the required computing resource of the task; is the edge computing unit; is the cloud or central computing cluster.

[0128] In one embodiment, tasks with high real-time requirements, such as data preprocessing and real-time inference of the multi-modal perception and soft-sensing field construction module 10, fast state prediction of the twin dynamic characterization module 20, and control action generation of the adaptive reinforcement learning predictive control module 40, are allocated to the edge computing unit for execution. Tasks that are computationally intensive and have high non-real-time requirements, such as meta-learning model updating of the twin dynamic characterization module 20 and policy network retraining of the adaptive reinforcement learning predictive control module 40, are offloaded to the cloud or central computing cluster for execution.

[0129] To adapt to the limited computing resources of the edge computing unit, models deployed to the edge are processed by a model pruning and quantization optimizer before deployment. Model pruning reduces the number of model parameters by removing low-contribution weight connections in the neural network; model quantization converts model weights and activation values from 32-bit floating-point numbers to 16-bit or 8-bit integer representations to reduce memory usage and accelerate computation.

[0130] 3. Model version management and online hot update: The dynamic edge computing and model collaborative deployment module 50 establishes a model library in the cloud or central computing cluster for storing and managing multiple training versions of the twin dynamic characterization module 20 and the adaptive reinforcement learning predictive control module 40. This module supports online hot update function. When a new version model with better performance is trained in the cloud or central computing cluster, the dynamic edge computing and model collaborative deployment module 50 can safely push and deploy the new version model file to the edge computing unit without interrupting the drying process control, seamlessly replacing the running old version model, and realizing the continuous iteration of system control capability.

[0131] To further illustrate the cooperative working process of the technical solutions of the present application, a specific working scenario will be described below.

[0132] In one embodiment, the multivariate intelligent predictive control system of the drying process of the present application is deployed on a line of dried konjac rice. When a new batch of dried konjac rice raw material whose physical and chemical properties are unknown is loaded into the drying equipment and the drying process is started, the system cooperates according to the following steps:

[0133] 1. System initialization and model adaptation: After the drying process is started, the multi-modal perception and soft-sensing field construction module 10 immediately starts collecting data through its integrated non-contact spectral sensors and conventional process sensors, and constructs an initial real-time soft-sensing field . At the same time, the twin dynamic characterization module 20 deployed on the edge computing unit of the dynamic edge computing and model collaborative deployment module 50 loads its meta-parameters obtained through offline meta-learning training . The twin dynamic characterization module 20 uses the small amount of initial data of the new batch provided by the multi-modal perception and soft-sensing field construction module 10 to execute its internal fast adaptation algorithm and generate a set of task-specific model parameters with high precision for the current batch in a short time .

[0134] 2. Steady-state optimization control phase: After the model completes adaptation, the system enters the steady-state optimization control phase. The twin dynamic characterization module 20 uses its adapted model to continuously predict the material state change trajectory in the future. The adaptive reinforcement learning predictive control module 40 integrates the real-time state provided by the multi-modal perception and soft-sensing field construction module 10 and the future prediction trajectory provided by the twin dynamic characterization module 20 to form the state space required for decision-making. Based on this state, the decision-making agent of the adaptive reinforcement learning predictive control module 40 outputs a set of multivariate control actions that optimize drying efficiency, product quality, and energy consumption as the comprehensive goal And the drying equipment is controlled through the underlying actuators.

[0135] 3. Process Anomaly Diagnosis and Control Strategy Adjustment: At a certain point during the drying process, the physical field-virtual field mapping and anomaly tracing module 30 detects the actual internal moisture gradient of the material provided by the multimodal sensing and soft measurement field construction module 10. , compared with the predictions made by the twin dynamics characterization module 20 The deviation between the two exceeded the preset dynamic threshold. The physical field-virtual field mapping and anomaly tracing module 30 then activated its anomaly root cause tracing algorithm. Through causal reasoning on the knowledge graph, it identified that the root cause of the deviation was the high initial moisture content of the current batch of materials, which caused the surface moisture evaporation rate to far exceed the internal moisture migration rate, posing a risk of surface crusting. A corresponding diagnostic information vector was then generated. .

[0136] The adaptive reinforcement learning predictive control module 40 receives a vector containing the diagnostic information in its next decision cycle. The updated state space. Its internal dynamic reward mechanism, an adaptive unit, is triggered, automatically adjusting the reward function based on the diagnostic information. The weighting. Specifically, increase the weighting of rewards related to product quality. In particular, the sub-items related to suppressing moisture gradients and avoiding crusting are weighted more leniently, while the weights of reward items related to drying efficiency are correspondingly reduced. .

[0137] 4. System Closed-Loop Feedback and Recovery: Guided by the adjusted reward function, the decision-making agent of the adaptive reinforcement learning predictive control module 40 generates a new set of control actions prioritizing mitigation of crust formation risk. For example, these new actions may include reducing heating power to slow surface evaporation and adjusting fan speed and dehumidification valve opening to optimize the temperature and humidity of the drying environment. Under the new control strategy, the internal and external moisture gradients of the material tend to balance, and the deviation monitored by the physical field-virtual field mapping and anomaly tracing module 30 gradually decreases and returns to the threshold range, allowing the system to return to steady-state optimization control.

[0138] All data throughout the process, including sensor data, model predictions, anomaly diagnosis, and control actions, is recorded and uploaded to the cloud or central computing cluster by the dynamic edge computing and model collaborative deployment module 50. This data can be used for offline retraining and continuous optimization of the models in the twin dynamics representation module 20 and the adaptive reinforcement learning prediction and control module 40 to enhance the system's ability to respond to similar future events.

[0139] While embodiments of the application have been shown and described, it is to be understood that the embodiments described are merely exemplary of the principles and application of the present application. Numerous modifications and adaptions can be effected without departing from the spirit and scope of the present application, which is not limited to the exact construction and arrangement described. It is intended, therefore, to cover all modifications and adaptions that fall within the scope of the claims and their equivalents.

Claims

1. A multivariable intelligent predictive control system for a process of drying konjac rice, characterized by, The method comprises the following steps: a multi-modal perception and soft measurement field construction module is used to collect multi-source heterogeneous data in real time during the drying process of dried konjac rice, and the collected data is fused and state inferred to construct a dynamic soft measurement field representing the internal state of the material and the external environmental parameters; a twin dynamic characterization module is used to establish a dynamic model based on the real-time state provided by the dynamic soft measurement field and the input control action, and to predict the material state change trajectory in the future period of time; a physical field-virtual field mapping and anomaly tracing module is used to compare the real material state provided by the dynamic soft measurement field with the material state predicted by the twin dynamic characterization module in real time, and when the deviation between the two exceeds a preset threshold, the root cause of the deviation is identified and output; an adaptive reinforcement learning predictive control module is used to integrate the real-time state provided by the dynamic soft measurement field, the material state change trajectory predicted by the twin dynamic characterization module, and the root cause output by the physical field-virtual field mapping and anomaly tracing module, to make multi-objective optimization decisions through a reinforcement learning algorithm, and to generate a set of optimal multivariate control actions; the adaptive reinforcement learning predictive control module generating a set of optimal multivariate control actions comprises: the current state provided by the dynamic soft measurement field, the future state prediction sequence output by the twin dynamic characterization module, and the diagnostic information vector output by the physical field-virtual field mapping and anomaly tracing module are collectively used to form a state space; the dynamic weight coefficients of each reward or penalty in the multi-objective reward function are adjusted online according to the real-time process state and diagnostic information in the state space; the multi-objective reward function is a weighted sum of a drying efficiency reward term, a product quality reward term, an energy consumption reward term, and an operation constraint penalty term; the state space is input into a trained policy network, and the policy network outputs the current optimal control action vector under the guidance of the adjusted multi-objective reward function; a dynamic edge computing and model collaborative deployment module is used to provide a computing framework and resource scheduling for the operation of the multi-modal perception and soft measurement field construction module, the twin dynamic characterization module, the physical field-virtual field mapping and anomaly tracing module, and the adaptive reinforcement learning predictive control module.

2. The multivariable intelligent predictive control system for the process of drying konjac rice according to claim 1, wherein, The multi-modal perception and soft measurement field construction module constructs the dynamic soft measurement field, which comprises: the multi-source heterogeneous data is collected by non-contact spectral sensors and conventional process sensors, and the multi-source heterogeneous data includes spectral data and conventional sensor data; the collected spectral data is subjected to baseline correction, smoothing filtering, and scattering correction, and the conventional sensor data is subjected to normalization or standardization processing; the spectral data and the conventional sensor data subjected to baseline correction, smoothing filtering, scattering correction, and normalization or standardization processing are input into a deep learning fusion model to infer an internal state vector containing the average internal moisture content, the surface moisture content, the internal and external moisture gradient, the surface crust hardening degree, and the quality-related indicators of the material particles; The internal state vector is integrated with external environment parameters extracted from conventional sensor data to form the dynamic soft measurement field.

3. The multivariable intelligent predictive control system for the process of drying konjac rice according to claim 1, wherein, The twin dynamics characterization module establishes the dynamic model and performs prediction, which includes: A hybrid neural network model is constructed by incorporating heat and mass transfer physical equations as inductive bias. In the offline training phase, based on historical batch drying data, a set of meta-parameters with rapid adaptation capability is trained by a meta-learning algorithm. In the online running phase, using a small amount of initial data of a new batch, gradient descent is performed based on the meta-parameters to quickly calculate task-specific parameters adapted to the current new batch. Using the real state at the current time as the initial condition, the adapted model and the input control action are used to iteratively calculate the state prediction sequence for multiple future steps in a self-recurrent manner.

4. The multivariable intelligent predictive control system for the process of drying konjac rice according to claim 1, wherein, The physical field-virtual field mapping and anomaly tracing module identifies and outputs the root cause, which includes: Calculating the real-time deviation vector between the real state provided by the dynamic soft measurement field and the predicted state predicted by the twin dynamics characterization module; Comparing the absolute value of the deviation of the pre-set key state variable in the real-time deviation vector with the corresponding dynamic threshold value, and triggering an abnormal event when any deviation exceeds the threshold value; When an abnormal event is triggered, causal reasoning is performed on a pre-constructed knowledge graph that describes the drying process to identify the most likely root cause of the current anomaly; The identified root cause and corresponding confidence are encoded into a structured diagnostic information vector and output.

5. The multivariable intelligent predictive control system for the process of drying konjac rice according to claim 1, wherein, The dynamic edge computing and model co-deployment module provides the computing framework and resource scheduling, which includes: Adopting a distributed heterogeneous computing architecture including edge computing units deployed on site and remote cloud computing clusters; Executing a task dynamic offloading scheduling algorithm to dynamically determine the execution of computing tasks on the edge computing unit or the cloud computing cluster according to the real-time requirements and computing complexity of the tasks; Establishing a model library in the cloud computing cluster, and performing pruning and quantization optimization on the models deployed to the edge computing unit, as well as supporting online hot updates.

6. The multivariable intelligent predictive control system for the process of drying konjac rice according to claim 3, wherein, The meta-learning algorithm is a model-independent meta-learning algorithm; The meta-parameters are a set of model initialization parameters; The calculation method of the task-specific parameters is: starting from the meta-parameters, using a small amount of initial data of a new batch, calculating the gradient according to the prediction loss function, and updating once or more times along the gradient direction.

7. The multivariable intelligent predictive control system for the process of drying konjac rice according to claim 1, wherein, The adjustment logic of the dynamic weight coefficient is: when the state variable representing quality risk in the dynamic soft measurement field exceeds the threshold value, the weight of the product quality reward term is increased; when the diagnostic information vector indicates a sensor failure, the weight of the reward term in the multi-objective reward function that depends on the sensor data for calculation is reduced.

8. The multivariable intelligent predictive control system for the process of drying konjac rice according to claim 4, wherein, The knowledge graph includes an entity set and a relationship set; The entity set includes drying equipment component entities, material batch entities, and environmental factor entities; The relationship set includes causal relationships and association relationships between entities; The causal reasoning is a reasoning path search on the knowledge graph to identify the root cause that best matches the deviation vector pattern that triggered the anomaly.

9. The multivariable intelligent predictive control system for the process of drying konjac rice according to claim 5, wherein, If an expected execution delay of a task is less than a preset real-time delay threshold, and a current available computing resource of the edge computing unit is greater than a required computing resource of the task, the task is assigned to the edge computing unit for execution; otherwise, the task is unloaded to the cloud or the central computing cluster for execution.

Images