Multi-source big data analysis and ai-driven adaptive control system for industrial internet

Abstract

The present application relates to the field of industrial big data, and more particularly to a multi-source big data analysis and AI-driven adaptive control system for the industrial internet, comprising a multi-source data perception module, a data governance and feature engineering module, a mechanism-embedded AI modeling module, an adaptive decision control module, and a fluctuation-driven self-iteration module; the present application adopts a scheme of deeply embedding an artificial intelligence model from two dimensions of network structure and loss function by taking industrial mechanism equations and process constraints as prior knowledge, matches the production physical process on the network topology and embeds a dedicated mechanism calculation layer, adds a mechanism consistency regularization term and a process constraint penalty term in the training process, and has the advantages of naturally conforming the model output to the industrial physical law and process safety requirements, effectively inhibiting physical contradictions caused by pure data fitting, and significantly improving the model interpretability and small sample scene adaptation capability.

Description

Technical Field

[0001] This invention relates to the field of industrial big data, and in particular to a multi-source big data analysis and AI-driven adaptive control system for the industrial internet. Background Technology

[0002] With the deep advancement of intelligent manufacturing and industrial internet technologies, massive amounts of multi-source heterogeneous production data have been accumulated in industrial production sites. Adaptive control systems based on big data and artificial intelligence are gradually becoming core technical solutions for improving product quality stability and optimizing production efficiency.

[0003] Existing industrial AI control systems mostly adopt a purely data-driven modeling approach, relying on historical production data to train prediction and control models to achieve closed-loop regulation of the production process. However, such solutions have significant shortcomings in actual industrial scenarios: First, pure data-driven models lack the rigid constraints of industrial physical mechanisms, and their output results are prone to violating basic physical laws and process safety requirements. The models have weak interpretability, and the prediction accuracy under small sample conditions is difficult to guarantee, resulting in insufficient reliability for engineering implementation. Second, for product quality fluctuations during the production process, existing technologies mostly rely on manual investigation of root causes or only locate the causes through data correlation analysis. The traceability results lack physical logic support and have limited accuracy. At the same time, the cost of collecting and labeling abnormal fluctuation samples is high and the quantity is scarce, making it difficult for the model to quickly adapt to changes in operating conditions and limiting its adaptive capabilities. Third, model iteration and updates often adopt the full data retraining method, which consumes a lot of computing power and has a long iteration cycle. Full parameter updates are prone to destroying the original steady-state control performance and pose a risk of process parameters going out of bounds. It is difficult to simultaneously take into account steady-state control accuracy and dynamic operating condition adaptability. Fourth, industrial mechanism models and AI models often run independently and lack a two-way empowerment linkage mechanism. As equipment ages and operating conditions drift, the mechanism deviation and model accuracy decay problems cannot be continuously corrected, resulting in insufficient stability of the system in long-term operation. Summary of the Invention

[0004] To overcome the problems mentioned in the background art, this invention proposes a multi-source big data analysis and AI-driven adaptive control system for the industrial internet.

[0005] The technical solution of the present invention is: a multi-source big data analysis and AI-driven adaptive control system for the industrial Internet, including a multi-source data sensing module, a data governance and feature engineering module, a mechanism-embedded AI modeling module, an adaptive decision control module, and a fluctuation-driven self-iterative module; The multi-source data sensing module is used to connect with various industrial equipment and business systems in the industrial field to collect multi-source heterogeneous industrial data throughout the entire production chain. The data governance and feature engineering module is used to preprocess and extract features from the collected multi-source heterogeneous industrial data, and output a standardized process feature dataset. The mechanism-embedded AI modeling module is used to embed industrial mechanism equations and process constraints as prior knowledge into the AI ​​model from two dimensions: network structure and loss function, respectively, to construct a process prediction model with physical consistency. The adaptive decision control module is used to solve for the optimal control parameters within the process constraints based on the output of the process prediction model, generate control commands and send them to the field execution unit to realize closed-loop optimization control of the production process. The fluctuation-driven self-iterative module is used to detect product quality fluctuations in real time, combine the mechanism model to complete the root cause of fluctuations, and generate enhanced training samples that conform to physical laws based on the root cause parameters, thereby driving the incremental update and adaptive evolution of the process prediction model.

[0006] Preferably, the industrial equipment and business systems connected to the multi-source data sensing module include PLCs, DCSs, field sensors, quality inspection equipment, and MES systems. The multi-source heterogeneous industrial data collected includes equipment operating parameters, process data, product quality data, raw material attribute data, and environmental condition data.

[0007] Preferably, the preprocessing and feature extraction of the data governance and feature engineering module specifically includes: performing time alignment, outlier cleaning and normalization on multi-source heterogeneous industrial data, extracting time domain, frequency domain and statistical features, constructing a mapping feature set corresponding to process parameters, equipment status and quality results, and outputting a standardized process feature dataset.

[0008] Preferably, the mechanism embedding AI modeling module embeds mechanism priors from the network structure dimension, specifically including: adopting a graph neural network architecture that matches the production process flow, embedding the mechanism calculation logic of the corresponding process into each process node, and the connection between process nodes corresponding to the material and energy transfer relationship; setting a mechanism mapping layer in the network hidden layer to complete feature transformation with industrial mechanism equations; and setting a truncated activation function with process upper and lower limits in the output layer to limit the model output within the compliant parameter range.

[0009] Preferably, the mechanism embedding AI modeling module embeds mechanism priors from the perspective of loss function, specifically including: constructing a composite loss function composed of data fitting loss, mechanism consistency loss, and process constraint penalty loss; the data fitting loss is used to calculate the deviation between the model's predicted value and the actual detected value, the mechanism consistency loss is used to calculate the residual norm after substituting the model output into the mechanism equation, and the process constraint penalty loss is used to apply gradient penalty to the model output that exceeds the process constraints.

[0010] Preferably, the fluctuation-driven self-iterative module includes a fluctuation detection unit and a root cause tracing unit; The fluctuation detection unit is used to monitor key product quality indicators in real time based on statistical process control methods, determine abnormal fluctuation events, and collect full-link process data before and after the fluctuation. The root cause tracing unit is used to screen a set of candidate parameters associated with quality fluctuations through causal inference, then substitute the candidate parameters into the mechanism equation for reverse solution and simulation reproduction, quantitatively calculate the contribution of each parameter to quality fluctuations, and output the root cause conclusion and quantified influence weight.

[0011] Preferably, the fluctuation-driven self-iterative module further includes a sample enhancement generation unit, which uses the root cause parameters obtained from the location as the core variables, performs gradient adjustment within the feasible parameter domain of the mechanism model, calculates the corresponding quality output results through mechanism simulation, generates process input-quality output sample pairs under different fluctuation amplitudes and parameter combinations, and adds them to the training sample library after mechanism consistency verification.

[0012] Preferably, the fluctuation-driven self-iterative module further includes a model incremental update unit, which is used to freeze the parameters of the mechanism embedding layer in the process prediction model, fine-tune the training of only the high-level feature extraction layer and the output layer, and complete the incremental optimization of the model using the newly added fluctuation sample set; after training, it is automatically replaced by the online process prediction model after offline verification and low-volume online testing.

[0013] As a preferred option, it also includes a mechanism knowledge base calibration module, which is used to periodically correct the key physical properties and process parameters of the mechanism model based on full production data, correct mechanism deviations caused by equipment aging and operating condition drift, and the calibrated mechanism model is synchronously updated to the mechanism embedded AI modeling module and the fluctuation-driven self-iteration module.

[0014] As a preferred option, the system adopts a cloud-edge collaborative deployment architecture; a lightweight process prediction model, a multi-source data sensing module, and a fluctuation detection unit are deployed at the edge to achieve local real-time closed-loop control; a full-scale mechanism knowledge base, a root cause tracing unit, a sample enhancement generation unit, and a model incremental update unit are deployed in the cloud to complete the model iterative training using cloud computing power, and the lightweight model is periodically distributed to the edge.

[0015] The beneficial effects of this invention are: 1. Compared to existing technologies that mostly employ purely data-driven AI modeling schemes, relying solely on historical production data for parameter fitting without deeply integrating industrial physical laws and process rules into the model construction process, this invention addresses the drawbacks of model outputs easily violating basic physical laws, poor compliance with process constraints, insufficient prediction accuracy under small sample conditions, strong black-box attributes, and weak interpretability. Instead, it adopts a scheme that deeply embeds industrial mechanism equations and process constraints as prior knowledge into the AI ​​model from two dimensions: network structure and loss function. It matches the production physical process to the network topology and embeds a dedicated mechanism calculation layer. During training, it adds mechanism consistency regularization terms and process constraint penalty terms. This approach offers advantages such as model outputs naturally conforming to industrial physical laws and process safety requirements, effectively suppressing physical contradictions caused by pure data fitting, and significantly improving model interpretability and adaptability to small sample scenarios.

[0016] 2. Compared to existing technologies that mostly use pure data correlation analysis to trace the root causes of quality fluctuations, and whose abnormal fluctuation samples rely on on-site collection and manual annotation, this invention has the disadvantages of easy causal confusion in the traceability results, lack of physical evidence support, scarcity and high cost of obtaining abnormal working condition samples, and difficulty in quickly adapting the model to changes in working condition fluctuations. This invention adopts a two-step method for tracing the root causes of fluctuations that combines data screening and mechanism verification. Based on the root cause parameters obtained from the location, it combines the mechanism model to generate a batch of compliant enhanced training samples. First, the range of candidate parameters is narrowed down through causal analysis, and then the causal relationship between parameters and quality fluctuations is quantitatively verified through mechanism simulation. At the same time, a multi-gradient fluctuation sample set is constructed within the feasible domain of the mechanism. This method has the advantages of accurate location of the root cause of fluctuations with complete physical logic, effectively solving the industry pain point of insufficient industrial abnormal samples, and rapidly expanding the coverage dimension of model training samples.

[0017] 3. Compared with existing technologies that mostly use full data retraining or full parameter fine-tuning to complete model updates and iterations, which have disadvantages such as high computational consumption, long iteration cycles, easy destruction of the original steady-state control accuracy of the model, and risk of process parameters going out of bounds after the update, this invention adopts a model update scheme that freezes the embedding layer parameters of the mechanism and only performs incremental fine-tuning on the high-level feature extraction module and the output layer. Combined with a tiered deployment mechanism of offline verification and low-volume online testing, it has the advantages of significantly reducing the computational and time costs of model iteration, fully preserving the original steady-state control capability of the model, and ensuring the safety and stability of the control process while rapidly optimizing the performance in fluctuating scenarios.

[0018] 4. Compared with existing technologies that mostly use fixed mechanism models and independent artificial intelligence models to run in parallel, the two lack a two-way empowerment linkage mechanism. This has the disadvantages of the mechanism model deviation continuously expanding with equipment aging and operating condition drift, the accuracy of the artificial intelligence model gradually decreasing, and insufficient long-term system stability. This invention adopts a two-way calibration scheme that uses production data to periodically correct key mechanism parameters and the calibrated mechanism model to synchronously feed back into artificial intelligence modeling and sample generation. This forms a positive cycle system in which data optimizes the mechanism and the mechanism guides artificial intelligence. It has the advantages of effectively offsetting the accuracy decay caused by equipment aging and operating condition drift, ensuring the control accuracy and reliability of the system throughout its entire life cycle, and realizing the continuous autonomous evolution of the system's control capabilities. Attached Figure Description

[0019] Figure 1 The diagram shows the architecture of the multi-source big data analysis and AI-driven adaptive control system for the industrial internet of the present invention. Figure 2 The diagram shown is a schematic representation of the internal structure of the mechanism-embedded AI modeling module in the multi-source big data analysis and AI-driven adaptive control system for the industrial internet of the present invention. Figure 3 The diagram shown is a schematic of the system cloud-edge collaborative deployment architecture in the multi-source big data analysis and AI-driven adaptive control system for the industrial internet of the present invention. Figure 4 The diagram shown is a schematic of the triple-progressive closed-loop operation logic of the multi-source big data analysis and AI-driven adaptive control system for the industrial internet of the present invention. Detailed Implementation

[0020] The technical solution of the present invention will be clearly and completely described below with reference to specific embodiments. 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.

[0021] The multi-source big data analysis and AI-driven adaptive control system for the Industrial Internet described in this embodiment is deployed in a general industrial continuous production scenario. This scenario includes multiple sequentially connected production processes, each equipped with corresponding detection instruments, actuators, and control systems. The system as a whole achieves high-precision adaptive control of the production process through a closed-loop process of multi-source data acquisition, mechanism fusion modeling, real-time optimization control, and fluctuation self-iterative evolution. Its overall architecture and workflow are as follows.

[0022] like Figure 1 and Figure 4As shown, this system comprises five core modules: a multi-source data perception module, a data governance and feature engineering module, a mechanism-embedded AI modeling module, an adaptive decision control module, and a fluctuation-driven self-iterative module. It can also be expanded to include a mechanism knowledge base calibration module. The overall system adopts a cloud-edge collaborative deployment architecture. The system forms a complete closed loop from data access to closed-loop control and then to self-iterative evolution, wherein: The multi-source data sensing module provides the system with raw industrial data from all dimensions; The data governance and feature engineering module completes data standardization and feature construction; The mechanism is embedded in the AI ​​modeling module as the core computing unit, and outputs process prediction results with physical consistency. The adaptive decision control module generates optimal control commands based on the prediction results to achieve real-time closed-loop control; The fluctuation-driven self-iteration module uses quality fluctuations as the trigger point to complete root cause analysis, sample augmentation, and incremental model updates. The mechanism knowledge base calibration module periodically corrects the mechanism parameters to ensure the long-term accuracy of the system.

[0023] Specifically, the multi-source data sensing module is the system's data input unit, used to connect with various industrial equipment and business systems in the industrial field to collect multi-source heterogeneous industrial data throughout the entire production chain.

[0024] In practical implementation, this module uses multiple standard industrial communication protocols such as OPCUA, ModbusTCP, MQTT, and Industrial Ethernet to interface with field PLC programmable logic controllers, DCS distributed control systems, various analog / digital field sensors, online quality inspection equipment, and MES manufacturing execution systems, achieving unified data aggregation across devices and systems. All acquired data is timestamped using the NTP network time protocol, supporting sampling periods down to 1 second, forming raw data sequences with precise time stamps.

[0025] The collected multi-source heterogeneous industrial data are specifically divided into five categories: Equipment operating parameters: including motor speed, operating power, operating current, regulating valve opening, delivery pump frequency, and other operating status parameters of the equipment actuator; Process data includes process status parameters such as temperature, pressure, material flow rate, liquid level, pH value, and reaction residence time for each step. Product quality data includes quality indicators such as dimensional accuracy, component purity, mechanical strength, surface defect rate, and batch yield, covering both online real-time testing data and offline laboratory testing data. Raw material attribute data: including basic parameters of raw materials such as the proportion of raw material components, particle size distribution, moisture content, and purity of the raw material batch; Environmental operating condition data: including environmental impact parameters such as ambient temperature, relative humidity, and atmospheric pressure in the production workshop.

[0026] The collected raw data is transmitted to the data governance and feature engineering module via the industrial intranet.

[0027] Specifically, the data governance and feature engineering module is used to preprocess and extract features from the collected multi-source heterogeneous industrial data, outputting a standardized process feature dataset. The specific implementation process is as follows: Data preprocessing: First, an outlier identification algorithm combining the 3σ criterion and box plot method is used to remove outlier data points caused by sensor malfunctions and signal interference. Short-term missing data is filled using linear interpolation, and long-term missing data is marked and segmented. Then, based on a unified timestamp benchmark, multi-source data with different sampling frequencies are time-aligned to form a parameter matrix with the same time dimension. Finally, a min-max normalization method is used to map all parameters with different dimensions to the [0,1] numerical range, eliminating the impact of dimension differences on model training.

[0028] Feature extraction: Based on the preprocessed data, three types of features are extracted: time-domain features include statistical features such as the mean, peak, valley, volatility, and slope of each parameter; frequency-domain features are extracted through fast Fourier transform to extract periodic features such as the dominant frequency and harmonic components of the parameters; statistical correlation features include correlation features such as the Pearson correlation coefficient and covariance between core parameters.

[0029] Feature set construction: Following the causal logic of "process input - process status - quality output", the extracted features are correlated and mapped according to the sequence of processes to construct a standardized process feature dataset. Each sample contains all input features and corresponding quality output labels within the corresponding time window. The sample time window can be adjusted according to the lag characteristics of the production process, and is usually set to 5~30 minutes.

[0030] like Figure 2 As shown, specifically, the mechanism-embedded AI modeling module is used to embed industrial mechanism equations and process constraints as prior knowledge into the AI ​​model from two dimensions: network structure and loss function, respectively, to construct a process prediction model with physical consistency.

[0031] 4.1 Mechanism Embedding in Network Structure Dimension The model uses a graph neural network (GNN) as its basic architecture, and the network topology perfectly matches the actual production process. The specific design is as follows: Topology matching design: Each production process is abstracted as a graph node, with the number of nodes matching the actual number of processes. Directed edges between nodes correspond to the material and energy flow directions between processes, ensuring the network's information transmission paths are perfectly aligned with the actual material flow, avoiding meaningless cross-process parameter associations. Each graph node contains its own mechanism calculation logic specific to that process, prioritizing the internal physical laws of the process during feature aggregation.

[0032] Mechanism mapping layer embedding: An independent mechanism mapping calculation layer is set up in the hidden layer of each graph node, replacing part of the black-box calculations of the fully connected layers. The mechanism mapping layer uses fixed industrial mechanism equations as its calculation logic. For example, for reaction-type processes, the Arrhenius equation is embedded to calculate the reaction rate; for fluid transportation processes, the Bernoulli equation is embedded to calculate the relationship between flow rate and pressure; and for heat exchange processes, Fourier's heat transfer law is embedded to calculate the heat transfer. The basic parameters of the mechanism mapping layer are physical constants and process setpoints, which remain fixed during training. Only the weight parameters of the fully connected layers are updated to ensure that the core physical relationships are not distorted by pure data fitting.

[0033] Hard-coded constraint output: The model output layer uses a sigmoid activation function with upper and lower limits truncation. The upper and lower limits of the activation function's output correspond one-to-one with the upper and lower limits of the parameters of the process hard constraints. For example, if the allowable range of reaction temperature for a certain process is [80℃, 120℃], then the output of the activation function is directly mapped to this range, so that the predicted value of the control parameters output by the model is naturally limited to the rated range of the equipment and the safety compliance range, thus eliminating the output of infeasible solutions from the structural level.

[0034] 4.2 Mechanism Embedding of Loss Function Dimension The model training uses a multi-objective composite loss function, and the formula for calculating the total loss L is: ; in, For data fitting loss, For mechanism consistency loss, Penalty losses due to process constraints; , , The weighting coefficients for each loss term can be adjusted according to the needs of the scenario; the standard value is [value to be filled in]. , , .

[0035] Data fitting loss The deviation between the predicted quality indicators and process state quantities and the actual measured values ​​is calculated using the mean square error (MSE) model. The formula is as follows: ; in, This refers to the batch sample size. For the first The model prediction value for each sample. For the first The true detection value of each sample. This loss ensures the model's accuracy in fitting historical normal operating data.

[0036] Mechanism consistency loss The intermediate state quantities and final predicted values ​​output by the model are substituted into the corresponding process mechanism equations, and the L2 norm of the equation residuals is calculated as the physical regularization term. For example, for the material conservation equation, the predicted feed and discharge rates of each component are substituted into the conservation equation, and the residual of the quality deviation is calculated; the smaller the residual, the more the model output conforms to physical laws. The formula is: ; in, For the mechanism equation for the first Predicted values ​​for each sample The residual calculation function. The output of this loss constraint model always satisfies the basic physical laws, avoiding physical contradictions that may occur with pure data fitting.

[0037] Process constraint penalty loss A hinge loss term is used to construct the constraint penalty term, applying gradient penalties to results where the model output exceeds the upper and lower limits of the process constraints. For strong constraints such as safety red lines, an exponential penalty weight is set, significantly increasing the loss value for non-compliant outputs. The formula is: ; in, The upper limit of the parameter is allowed. This is the lower bound allowed for the parameters. This loss term guides the model to gradually converge to the constraint interval during training.

[0038] During model training, the Adam optimizer is used to minimize the composite loss function. The initial learning rate is set to 0.001. The model is considered to have converged when the loss value decreases by less than 1e-5 for 10 consecutive iterations, thus obtaining the final process prediction model.

[0039] Specifically, the adaptive decision control module is used to solve for the optimal control parameters within the process constraints based on the output of the process prediction model, generate control commands and issue them to the field execution unit to realize closed-loop optimization control of the production process.

[0040] The specific implementation process is as follows: The module receives real-time collected process parameters and raw material parameters, inputs them into the process prediction model, and obtains the product quality prediction results within the next 1 to 5 time windows; with "minimum product quality deviation and minimum production energy consumption" as the dual optimization objectives, and with process hard constraints and equipment rated parameters as boundary conditions, the optimal combination of control parameters, including adjustable parameters such as regulating valve opening, heating power, feed flow rate, and stirring speed, is solved in the feasible parameter space using particle swarm optimization algorithm or sequential quadratic programming algorithm.

[0041] The optimal control commands obtained from the solution are sent to the field PLC and actuators via the OPCUA protocol to adjust the operating parameters of the production process in real time, stabilizing product quality within the target range while also optimizing production energy efficiency. The frequency of control command issuance is consistent with the data acquisition frequency, achieving second-level closed-loop adjustment.

[0042] Specifically, the fluctuation-driven self-iterative module is used to detect product quality fluctuations in real time, combine the mechanism model to trace the root causes of fluctuations, and generate enhanced training samples that conform to physical laws based on the root cause parameters, driving the incremental update and adaptive evolution of the process prediction model. This module includes a fluctuation detection unit, a root cause tracing unit, a sample enhancement generation unit, and a model incremental update unit.

[0043] 6.1 Fluctuation Detection Unit and Root Cause Tracing Unit The fluctuation detection unit, based on the Statistical Process Control (SPC) method, generates control charts for key product quality indicators, calculates the historical mean and standard deviation of the indicators, and sets ±3σ as the upper and lower control limits. When a quality indicator exceeds the control limits or shows a unidirectional upward / downward trend for seven consecutive sampling points, it is determined to be an abnormal fluctuation event. After a fluctuation event is triggered, the unit automatically extracts the entire process parameters, equipment status, and raw material data for 30 minutes before and after the fluctuation occurs, forming a time-stamped fluctuation event dataset, which is then input into the root cause analysis unit.

[0044] Root cause tracing unit: Employs a two-step tracing method of "initial data screening + mechanism verification": The first step is data-driven initial screening: through Granger causality test and SHAP value sensitivity analysis, the top 10 candidate parameters with the highest correlation to quality fluctuations are screened from dozens to hundreds of process parameters, irrelevant parameters are eliminated, and the scope of tracing is narrowed. The second step is mechanism verification and quantitative positioning: Substitute the candidate parameter set into the mechanism equation of the corresponding process, and adjust the value of each candidate parameter one by one through reverse solving and digital simulation to reproduce the amplitude and trend of quality fluctuation. Quantitatively calculate the contribution ratio of each parameter fluctuation to the final quality fluctuation, distinguish between primary and secondary parameters, clarify the complete physical causal path, and finally output the root cause parameter list and the corresponding quantitative influence weight.

[0045] 6.2 Sample Augmentation Generation Unit The sample enhancement generation unit is used to perform gradient adjustment within the feasible parameter domain of the mechanism model, with the root cause parameters obtained from the location as the core variables. It calculates the corresponding quality output results through mechanism simulation, generates process input-quality output sample pairs under different fluctuation amplitudes and parameter combinations, and adds them to the training sample library after mechanism consistency verification.

[0046] The specific implementation process is as follows: Using the primary causal parameters obtained from root cause analysis as the core variables, and within the process-feasible range of these parameters, gradient values ​​are generated at equal step sizes. For example, for the temperature parameter, multiple sets of values ​​are generated within the allowable range with a step size of 2℃. Simultaneously, other secondary parameters are set to their actual values ​​during fluctuation periods. Each set of parameters is input into the mechanistic model for forward simulation calculation, yielding the corresponding quality output results, forming a sample pair of "process input parameters - quality output results".

[0047] All generated samples undergo mechanism consistency verification. The verification criteria are that the sample input and output satisfy basic mechanism equations such as material conservation and energy conservation, and the residual is lower than a set threshold to be considered a qualified sample. False samples that do not conform to the production logic are eliminated. Finally, the real fluctuation samples are merged with the qualified mechanism-generated samples to construct a special training subset for fluctuation conditions, which is then added to the global training sample library.

[0048] 6.3 Model Incremental Update Unit The incremental update unit is used to freeze the parameters of the mechanism embedding layer in the process prediction model, and only fine-tune the training of the high-level feature extraction layer and the output layer. The model incremental optimization is completed using the newly added fluctuation sample set. After training, the model is automatically replaced by the online process prediction model after offline verification and low-volume online testing.

[0049] The specific implementation process is as follows: Using a transfer learning and incremental training framework, firstly, all parameters of the mechanism mapping layer and output constraint layer in the process prediction model are frozen. These parameters correspond to fixed physical laws and process constraints and do not require adjustment. Only the weight parameters of the high-level graph neural network feature extraction layer and the fully connected output layer are fine-tuned. During training, a newly added training subset specifically for fluctuating operating conditions is used as input, and iterative training is performed with a small learning rate of 0.0001 to quickly optimize the model's prediction accuracy for fluctuating operating conditions.

[0050] After training, the model is first verified to meet the preset indicators in terms of prediction accuracy and constraint compliance under normal and fluctuating operating conditions through offline test sets. Then, a small-scale trial run is conducted online, with 10% of the production flow connected to the new model for comparison and verification. After confirming that the control effect meets the standards and there are no violations, the online process prediction model is automatically replaced with the full model, completing one adaptive iteration.

[0051] Specifically, the system also includes a mechanism knowledge base calibration module, which is used to periodically correct the key physical properties and process parameters of the mechanism model based on full production data, correct mechanism deviations caused by equipment aging and operating condition drift, and the calibrated mechanism model is synchronously updated to the mechanism embedded AI modeling module and the fluctuation-driven self-iteration module.

[0052] In practice, the system calls upon all historical production data at a fixed interval of one month and uses a nonlinear least squares parameter identification method to correct key parameters in the mechanism model. These parameters include physical and process properties that are prone to change due to equipment aging, catalyst deactivation, and operating condition drift, such as reaction activation energy, heat transfer coefficient, pipeline friction coefficient, and material diffusion coefficient. The corrected mechanism parameters are then synchronously updated to the mechanism mapping layer of the mechanism-embedded AI modeling module and the mechanism simulation unit of the fluctuation-driven self-iterative module, forming a two-way reinforcing cycle of "data feeding back to the mechanism and the mechanism guiding AI," ensuring the accuracy and reliability of the system's long-term operation.

[0053] like Figure 3 As shown, in this embodiment, the system adopts a cloud-edge collaborative deployment architecture, which takes into account both the real-time requirements of industrial control and the computing power requirements of model iteration: Edge deployment: Lightweight process prediction models, multi-source data sensing modules, data governance and feature engineering modules, adaptive decision control modules, and fluctuation detection units are deployed on edge gateways or local industrial control computers in industrial settings. The edge device possesses complete local closed-loop control capabilities; data processing and control command generation are both completed locally, with a response latency of less than 100ms, meeting the real-time requirements of industrial control. It can operate independently and stably even when offline, ensuring production continuity.

[0054] Cloud Deployment: On the industrial internet cloud platform, a full mechanism knowledge base, root cause analysis unit, sample enhancement generation unit, model incremental update unit, and mechanism knowledge base calibration module are deployed. The cloud utilizes ample computing resources to complete computationally intensive tasks such as complex mechanism simulation, large sample set generation, full model training, and parameter calibration. The lightweight model that has been trained and optimized is periodically delivered to the edge via OTA, enabling continuous iterative upgrades to the edge control capabilities.

[0055] Data exchange is conducted between the cloud and the edge via dedicated industrial internet lines. The edge devices regularly upload production data and fluctuation event data to the cloud, and the cloud distributes updated models and configuration parameters to the edge devices, forming a collaborative whole.

[0056] In summary, the complete operation process of this system encompasses three progressive closed loops: steady-state control, fluctuation self-iteration, and periodic calibration, as detailed below: Steady-state control closed loop: After the system starts, the multi-source data sensing module collects real-time industrial data from all dimensions on site. After preprocessing by the data governance and feature engineering module, the data is input into the process prediction model. The model predicts the product quality trend in real time, and the adaptive decision control module solves for the optimal control parameters and issues them for execution, so that the production process remains stable under optimal conditions and completes the closed-loop control of daily production.

[0057] Fluctuation Iteration Closed Loop: When the fluctuation detection unit identifies abnormal fluctuations in product quality, it automatically triggers the root cause tracing process to locate the cause of the fluctuation; then the sample enhancement generation unit generates a batch of enhanced samples based on the root cause to expand the training set; after the model incremental update unit completes model fine-tuning and verification, it replaces the online model to improve the system's ability to cope with similar fluctuations and completes an adaptive evolution.

[0058] Mechanism calibration closed loop: The system starts the mechanism knowledge base calibration module at fixed intervals, corrects the mechanism model parameters based on accumulated production data, and updates the mechanism calculation logic of the entire system in a synchronous manner to ensure the accuracy of the system's long-term operation. Specific Implementation This embodiment uses a general industrial continuous liquid-phase reaction production line as an application scenario. This production line sequentially comprises four processes: raw material mixing and pretreatment, catalytic main reaction, heat exchange and cooling, and online quality inspection. Each process is equipped with corresponding detection sensors and regulating actuators, and basic control is achieved through a DCS system, while production data management is achieved through a MES system. The specific implementation process of this system on this production line is as follows: Multi-source data perception stage The system interfaces with the on-site DCS system and online quality analyzer via the OPCUA protocol, and connects to distributed sensors on-site via the ModbusTCP protocol. End-to-end time synchronization is achieved using the NTP protocol, with a data sampling period of 1 second. Collected parameters include: feed flow rate, raw material component concentration, and stirring speed in the raw material pretreatment process; reactor temperature, internal pressure, catalyst feed rate, and reaction residence time in the main reaction process; cooling water flow rate and inlet / outlet temperature in the heat exchange and cooling process; purity of main components and impurity content in the finished product testing process; and ambient temperature and humidity data from the workshop. All collected data is timestamped and transmitted in real-time to the data governance and feature engineering module.

[0060] Data governance and feature engineering phase Firstly, use 3 The criteria identify and remove abnormal data points caused by sensor jumps, use linear interpolation to fill in short-term missing data points of less than 5 sampling points, and segment and mark long-term missing data of more than 30 seconds. Then, all parameters are time-aligned with a uniform time step of 1 second to form a two-dimensional data matrix with dimensions of [time step, number of parameters]. The min-max normalization method is used to map all parameters to the [0,1] interval.

[0061] Based on this, time-domain features are extracted, including the moving mean, moving standard deviation, and slope of change of each parameter, with the sliding window set to 10 minutes. Following the causal order of "raw material input - reaction process - cooling output - quality result", the features are associated with the corresponding product purity labels to construct a standardized feature dataset with a single sample length corresponding to a 10-minute process window.

[0062] Mechanism Embedded in AI Modeling Stage A graph neural network architecture is constructed to correspond one-to-one with the four processes, and four process nodes are set. The directed edges between the nodes correspond to the material flow direction and the energy transfer direction.

[0063] At the network structure level: the Arrhenius equation is embedded inside the main reaction process node as a mechanism mapping layer, which shows the quantitative relationship between the solidification reaction rate and temperature and activation energy. The mechanism layer parameters are kept fixed during training. The Fourier heat transfer law is embedded inside the heat exchange and cooling process node as a mechanism mapping layer, which shows the correspondence between the solidification heat exchange and temperature difference and heat transfer coefficient. The model output layer uses a truncated Sigmoid activation function to hard-constrain the output range of the main reaction temperature within the process allowable range of 80℃~120℃ and the reaction pressure within the range of 0.2MPa~0.5MPa.

[0064] At the loss function level: a composite loss function is used for training, with weight coefficients set to... The mechanism consistency loss is verified using the material conservation equation, calculating the residual between the predicted total feed mass and the total discharge mass. The process constraint penalty loss employs hinged loss for strongly constrained parameters such as temperature and pressure, imposing an exponential penalty when exceeding a safety threshold. The Adam optimizer is used for training, with an initial learning rate of 0.001. When the loss decreases by less than [a certain value] over 10 consecutive training rounds... Training is stopped at a certain point to obtain a process prediction model, which can predict the product purity 10 minutes later based on the current process parameters.

[0065] Adaptive decision control phase The system inputs the current process parameters into the process prediction model in real time to obtain the predicted product purity value for the next 10 minutes. With the dual optimization objectives of "minimizing product purity deviation and minimizing heating energy consumption," and using the upper and lower limits of each parameter as constraints, the system employs a particle swarm optimization algorithm to solve for the optimal combination of control parameters within the feasible region. These parameters include raw material feed flow rate, reactor heating power, and cooling water regulating valve opening. The optimal control command obtained is sent to the field DCS system via the OPCUA protocol, driving the actuators to complete parameter adjustments. The control cycle is consistent with the sampling cycle, which is 1 second, achieving real-time closed-loop optimization of the production process.

[0066] Fluctuation-driven self-iterative phase The fluctuation detection unit calculates the mean and standard deviation of product purity based on historical normal production data, and sets ±3. To control the upper and lower limits, when the product purity shows a unidirectional decreasing trend for 7 consecutive sampling points and deviates from the mean by more than 2... When an abnormal fluctuation event is detected, the entire process data within 30 minutes before and after the fluctuation occurs is automatically captured.

[0067] The root cause analysis unit first screened out the two candidate parameters with the highest correlation, reaction temperature and feed concentration, from all process parameters through SHAP value sensitivity analysis. Then, the two parameters were substituted into the master reaction mechanism equation for inverse simulation. The parameter values ​​were adjusted one by one to reproduce the purity fluctuation range. The calculation showed that the contribution of feed concentration fluctuation was 72%, which is the main cause of the fluctuation, while the contribution of reaction temperature fluctuation was 21%, which is the secondary factor.

[0068] The sample enhancement generation unit uses the raw material feed concentration as the core variable. Within the concentration range allowed by the process, it generates 20 sets of gradient values ​​with a step size of 0.5%, while the remaining parameters retain the true values ​​during the fluctuation period. Each set of parameters is input into the mechanism model for forward simulation to calculate the corresponding product purity output, resulting in 20 sets of "process input-quality output" sample pairs. All samples are verified by material conservation, and those with a residual of less than 0.1% are considered qualified. After removing unqualified samples, 19 sets of qualified enhanced samples are merged with the actual fluctuation samples, resulting in a total of 20 sets of samples added to the fluctuation condition subset of the training sample library.

[0069] The incremental update unit freezes the parameters of all mechanism mapping layers and output constraint layers in the model, and only fine-tunes the high-level feature aggregation layer and fully connected output layer of the graph neural network. It uses a small learning rate of 0.0001 for incremental training and stops after 50 iterations. After offline testing, the purity prediction error of the model for similar concentration fluctuation conditions decreased from 8.2% to 2.1%. Subsequently, a small online flow test with 10% flow was conducted. After confirming that the control effect met the standard and there were no parameter out-of-bounds errors, the online model was automatically fully replaced, completing one adaptive iteration.

[0070] Mechanism calibration and cloud-edge deployment The system performs a mechanism parameter calibration once a month, calling up all historical production data for the month and using a nonlinear least squares parameter identification method to correct easily drifting parameters such as the apparent activation energy of the main reaction and the total heat transfer coefficient of the heat exchange process. The calibrated parameters are then synchronously updated to the model mechanism layer and the sample generation module.

[0071] The system adopts a cloud-edge collaborative deployment: the on-site edge industrial control computer deploys a lightweight prediction model, data acquisition and real-time control module, with a local control response latency of less than 80ms, and can run independently when the network is disconnected; the cloud industrial internet platform deploys root cause tracing, sample generation, incremental training and mechanism calibration modules, uses cloud computing power to complete computationally intensive tasks, and distributes the optimized lightweight model to the edge every week to achieve continuous system evolution.

[0072] As verified by this embodiment, after adopting this system, the fluctuation range of key product quality indicators under normal production conditions is reduced by 23%, the recovery time after abnormal fluctuations is shortened by 75%, and the model adaptation cycle after changes in operating conditions is shortened by more than 70% compared with the pure data-driven solution, achieving the expected technical effect.

[0073] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A multi-source big data analysis and AI-driven adaptive control system for the industrial internet, characterized in that, It includes a multi-source data perception module, a data governance and feature engineering module, a mechanism-embedded AI modeling module, an adaptive decision control module, and a fluctuation-driven self-iterative module; The multi-source data sensing module is used to connect with various industrial equipment and business systems in the industrial field to collect multi-source heterogeneous industrial data throughout the entire production chain. The data governance and feature engineering module is used to preprocess and extract features from the collected multi-source heterogeneous industrial data, and output a standardized process feature dataset. The mechanism-embedded AI modeling module is used to embed industrial mechanism equations and process constraints as prior knowledge into the AI ​​model from two dimensions: network structure and loss function, respectively, to construct a process prediction model with physical consistency. The adaptive decision control module is used to solve for the optimal control parameters within the process constraints based on the output of the process prediction model, generate control commands and send them to the field execution unit to realize closed-loop optimization control of the production process. The fluctuation-driven self-iterative module is used to detect product quality fluctuations in real time, combine the mechanism model to complete the root cause of fluctuations, and generate enhanced training samples that conform to physical laws based on the root cause parameters, thereby driving the incremental update and adaptive evolution of the process prediction model.

2. The multi-source big data analysis and AI-driven adaptive control system for the industrial internet according to claim 1, characterized in that, The multi-source data sensing module interfaces with industrial equipment and business systems including PLCs, DCSs, field sensors, quality inspection equipment, and MES systems. The collected multi-source heterogeneous industrial data includes equipment operating parameters, process data, product quality data, raw material attribute data, and environmental condition data.

3. The multi-source big data analysis and AI-driven adaptive control system for the industrial internet according to claim 1, characterized in that, The preprocessing and feature extraction of the data governance and feature engineering module specifically includes: performing time alignment, outlier cleaning and normalization on multi-source heterogeneous industrial data, extracting time domain, frequency domain and statistical features, constructing a mapping feature set corresponding to process parameters, equipment status and quality results, and outputting a standardized process feature dataset.

4. The multi-source big data analysis and AI-driven adaptive control system for the industrial internet according to claim 1, characterized in that, The mechanism embedding AI modeling module embeds mechanism priors from the network structure dimension, specifically including: adopting a graph neural network architecture that matches the production process flow, embedding the mechanism calculation logic of the corresponding process into each process node, and the connection between process nodes corresponding to the material and energy transfer relationship; setting a mechanism mapping layer in the network hidden layer to complete feature transformation with industrial mechanism equations; and setting a truncated activation function with process upper and lower limits in the output layer to limit the model output within the compliant parameter range.

5. The multi-source big data analysis and AI-driven adaptive control system for the industrial internet according to claim 1, characterized in that, The mechanism-embedded AI modeling module embeds mechanism priors from the perspective of loss function, specifically including: constructing a composite loss function composed of data fitting loss, mechanism consistency loss, and process constraint penalty loss; the data fitting loss is used to calculate the deviation between the model's predicted value and the actual detected value, the mechanism consistency loss is used to calculate the residual norm after substituting the model output into the mechanism equation, and the process constraint penalty loss is used to apply gradient penalty to the model output that exceeds the process constraints.

6. The multi-source big data analysis and AI-driven adaptive control system for the industrial internet according to claim 1, characterized in that, The fluctuation-driven self-iterative module includes a fluctuation detection unit and a root cause tracing unit; The fluctuation detection unit is used to monitor key product quality indicators in real time based on statistical process control methods, determine abnormal fluctuation events, and collect full-link process data before and after the fluctuation. The root cause tracing unit is used to screen a set of candidate parameters associated with quality fluctuations through causal inference, then substitute the candidate parameters into the mechanism equation for reverse solution and simulation reproduction, quantitatively calculate the contribution of each parameter to quality fluctuations, and output the root cause conclusion and quantified influence weight.

7. The multi-source big data analysis and AI-driven adaptive control system for the industrial internet according to claim 6, characterized in that, The fluctuation-driven self-iterative module also includes a sample enhancement generation unit, which uses the root cause parameters obtained from the location as the core variables, performs gradient adjustment within the feasible parameter domain of the mechanism model, calculates the corresponding quality output results through mechanism simulation, generates process input-quality output sample pairs under different fluctuation amplitudes and parameter combinations, and adds them to the training sample library after mechanism consistency verification.

8. The multi-source big data analysis and AI-driven adaptive control system for the industrial internet according to claim 7, characterized in that, The fluctuation-driven self-iterative module also includes a model incremental update unit, which is used to freeze the parameters of the mechanism embedding layer in the process prediction model, fine-tune the training of only the high-level feature extraction layer and the output layer, and complete the incremental optimization of the model using the newly added fluctuation sample set. After training, the model is validated offline and tested online with a small volume of data. Once it meets the requirements, it automatically replaces the process prediction model running online.

9. The multi-source big data analysis and AI-driven adaptive control system for the industrial internet according to claim 1, characterized in that, It also includes a mechanism knowledge base calibration module, which is used to periodically correct the key physical properties and process parameters of the mechanism model based on full production data, correct mechanism deviations caused by equipment aging and operating condition drift, and the calibrated mechanism model is synchronously updated to the mechanism embedded AI modeling module and the fluctuation-driven self-iteration module.

10. The multi-source big data analysis and AI-driven adaptive control system for the industrial internet according to claim 1, characterized in that, The system adopts a cloud-edge collaborative deployment architecture; lightweight process prediction models, multi-source data sensing modules, and fluctuation detection units are deployed at the edge to achieve local real-time closed-loop control; a full-scale mechanism knowledge base, root cause tracing unit, sample enhancement generation unit, and model incremental update unit are deployed in the cloud to complete model iterative training using cloud computing power, and the lightweight model is periodically distributed to the edge.

Images