Cloud edge cooperation system for industrial control and implementation method thereof

Abstract

The invention provides an industrial control-oriented cloud-side collaboration system and an implementation method thereof, and relates to the technical field of cloud-side collaboration, and the system comprises an equipment layer which comprises an industrial equipment access gateway and is responsible for collecting equipment operation data and executing a control instruction; the edge layer comprises a data preprocessing unit used for processing the real-time data uploaded by the equipment layer; the lightweight inference engine is used for loading a lightweight model issued by the cloud layer and generating a control decision based on the real-time data; the deterministic scheduling engine is used for hierarchically controlling tasks according to task time sensitivity and distributing execution positions; the cloud layer comprises an industrial knowledge model center used for constructing a scenarized knowledge model based on the equipment operation data; the model lightweight unit is used for converting the scenarized knowledge model into a lightweight model; and the decision distribution unit is used for issuing the lightweight model to the edge layer. The technical problems of knowledge expression, model adaptation and task allocation in an industrial control scene can be solved.

Description

Technical Field

[0001] The present invention relates to the technical field of cloud-edge collaboration, and particularly to a cloud-edge collaboration system for industrial control and its implementation method. Background Art

[0002] As an important development direction of the distributed computing architecture, the cloud-edge collaboration system realizes the optimal allocation of computing resources by integrating the powerful computing power of cloud computing and the low-latency characteristics of edge computing. This architecture has significant advantages in aspects such as real-time data processing, load balancing, and network bandwidth optimization. Currently, mainstream cloud-edge collaboration systems usually adopt static configuration or task allocation mechanisms based on simple rules, and determine the execution location of tasks according to factors such as computational complexity and data transmission volume, and have been widely applied in multiple fields including traffic monitoring and smart cities.

[0003] The industrial control field poses special requirements for computing systems, including millisecond-level response time, high reliability, and deterministic execution, etc. With the advancement of Industry 4.0, more and more industrial control systems have begun to adopt the cloud-edge collaboration architecture to meet the requirements of intelligent manufacturing for data processing and decision-making control. Existing industrial control cloud-edge systems usually deploy complex models in the cloud for training and optimization, and deploy simplified models at the edge for real-time control. Such systems have shown certain advantages in data center-type industrial applications, but still face challenges in industrial field environments with limited resources and complex network conditions.

[0004] However, there are obvious technical bottlenecks in the knowledge expression and model adaptation of the cloud-edge collaboration system in the industrial control scenario. The professional knowledge and data characteristics in the industrial field are complex and diverse, and it is difficult for the system to effectively integrate domain knowledge and data-driven models; at the same time, the heterogeneity and resource limitations of edge devices make it difficult to efficiently deploy complex models developed in the cloud to the edge. In addition, industrial control tasks have different requirements for time sensitivity, ranging from millisecond level to minute level, and existing systems lack a flexible task scheduling mechanism to balance real-time performance and computing resources. Summary of the Invention

[0005] In view of this, the present invention proposes a cloud-edge collaboration system for industrial control and its implementation method. By constructing a scenario-based knowledge model to integrate industrial domain knowledge and data characteristics, adopting a dedicated lightweight technology to achieve efficient migration of models between the cloud and the edge, and combining a task scheduling mechanism based on time sensitivity, the technical problems of knowledge expression, model adaptation, and task allocation in the industrial control scenario are solved, thereby improving the control accuracy, response speed, and adaptability of industrial control systems in complex and changeable working conditions.

[0006] The technical solution of the present invention is realized as follows:

[0007] On the one hand, the present invention provides a cloud-edge collaborative system for industrial control, including:

[0008] The device layer includes an industrial device access gateway, which is responsible for collecting device operation data and executing control instructions;

[0009] The edge layer includes: a data preprocessing unit for processing real-time data uploaded by the device layer; a lightweight inference engine for loading the lightweight model sent from the cloud layer and generating control decisions based on real-time data; a deterministic scheduling engine for classifying control tasks according to task time sensitivity and allocating execution locations;

[0010] The cloud layer includes: an industrial knowledge model center for constructing a scenario-based knowledge model based on device operation data; a model lightweighting unit for converting the scenario-based knowledge model into a lightweight model; a decision distribution unit for sending the lightweight model to the edge layer.

[0011] Preferably, the construction process of the scenario-based knowledge model includes:

[0012] Step 1: Obtain expert experience, device process specifications, and threshold parameters through the industrial device access gateway, establish an industrial entity and relationship classification system, and obtain an industrial domain concept framework;

[0013] Step 2: Summarize and clean the historical data uploaded by the device layer, perform standardization processing, extract time series statistical features and frequency domain features, and obtain a device state feature index set;

[0014] Step 3: Based on the industrial domain concept framework, use the ontology description language to construct an industrial domain ontology, extract an industrial mechanism model from expert experience and device process specifications, convert it into a parametric expression, construct a knowledge graph with the industrial domain ontology as the skeleton, and embed the parametric expression into the knowledge graph to form a static knowledge layer;

[0015] Step 4: Use the device state feature index set to calibrate the initial parameters of the parametric model in the static knowledge layer, determine the optimal parameters, and form a dynamic knowledge layer;

[0016] Step 5: Perform multi-scale time window analysis on the device state feature index set, combine it with the dynamic knowledge layer, construct a directed graph of causal relationships between parameters, and integrate the causal reasoning structure into the dynamic knowledge layer;

[0017] Step 6: Use the density-based clustering method to identify typical working scenarios, and train a prediction model that integrates the static knowledge layer and the dynamic knowledge layer for each working scenario using historical operation data to obtain a complete set of scenario-based knowledge models.

[0018] Preferably, the industrial mechanism model in the static knowledge layer adopts the following parametric expression:

[0019]

[0020] In the formula, P(x,θ) represents the output process parameter; x is the state variable vector; θ is the set of parameters to be calibrated; f k (x) is the k-th physical constraint function; K is the total number of physical constraint functions; γ k is the weight of the k-th physical constraint; C is the total number of domain knowledge functions; ψ c (x,θ c ) is the c-th domain knowledge function; λ c is the weight of the c-th domain knowledge; θ c is the adjustable parameter subset of the c-th domain knowledge function.

[0021] Preferably, in step 4, the initial parameter calibration of the parameterized model is optimized using the following objective function to determine the optimal parameters:

[0022]

[0023] In the formula, Ω is the objective function; θ is the set of parameters to be calibrated; x i and y i respectively represent the input features and the corresponding target observed values in the historical data; N is the number of features; represents the predicted output of the parameterized model in the static knowledge layer for the input x i ; ω i is the data quality weight; is the residual function of the r-th industrial mechanism constraint, and the number thereof is R; δ r is the weight of the mechanism constraint in the objective function; κ is the adjustment coefficient of the prior deviation; D(θ,θ ref ) is the prior deviation; θ ref is the parameter reference value.

[0024] Preferably, the prediction model that fuses the static knowledge layer and the dynamic knowledge layer adopts the following hybrid reasoning mechanism:

[0025] y = σ[β(s)·Φ static (x,θ static )+(1-β(s))·Φ dynamic (x,θ dynamic ,H(t))]

[0026] In the formula, y is the model prediction output; σ is the activation function; β(s) is the knowledge fusion coefficient in scenario s, and the value range is [0,1]; s is the current scenario identifier; Φ static is the reasoning function based on the static knowledge layer; θ static is the static knowledge layer parameter set; Φ dynamicis the inference function based on the dynamic knowledge layer; θ dynamic is the set of dynamic knowledge layer parameters; H(t) is the temporal context information, including the time series of historical data.

[0027] Preferably, the process of converting the scenario knowledge model into a lightweight model includes:

[0028] Step A: Perform mixed-precision quantization on the static knowledge layer parameters and dynamic knowledge layer parameters in the scenario knowledge model. The static knowledge layer retains the high-precision representation of the parameters related to key physical constraints, and the dynamic knowledge layer parameters adopt dynamic range quantization based on sliding window statistics to obtain a preliminary lightweight model;

[0029] Step B: Based on the directed graph of causal relationships between parameters, calculate the influence factors of each parameter node in the preliminary lightweight model, dynamically adjust the pruning threshold according to the clustering results of typical working conditions, and remove the redundant parameters with influence lower than the threshold to obtain a structure-simplified model;

[0030] Step C: Use the original scenario knowledge model as the teacher model and the structure-simplified model as the student model, and design a distillation loss function to guide the student model to learn the knowledge of the teacher model to obtain a distillation-optimized model;

[0031] Step D: For the distillation-optimized model, optimize the network structure according to the hardware characteristics of different edge devices to generate the final lightweight model.

[0032] Preferably, control tasks according to the task time sensitivity level and allocate execution locations, specifically including:

[0033] Classify industrial control tasks into three levels: high, medium, and low according to time sensitivity; high-time-sensitivity tasks are preferentially executed at the edge layer; medium-time-sensitivity tasks are dynamically determined to be executed at the edge layer or the cloud layer according to the current system load; low-time-sensitivity tasks are preferentially executed at the cloud layer.

[0034] Preferably, the lightweight model at the edge layer performs adaptive parameter optimization according to real-time data and adopts the following update mechanism:

[0035]

[0036] where θ j (t) is the estimated value of the jth parameter at time t, j ∈ {1, 2,..., N}, N is the total number of parameters; η(t) is the adaptive learning rate; L s (·) is the loss function calculated on a sliding window of length s; is the gradient of the loss function with respect to the parameter θ j ; q(t) is the data quality scoring factor, with a value range of [0, 1]; α is the regularization coefficient; R e(·) is an elastic constraint function; is the parameter θ j calibration value.

[0037] Preferably, the industrial device access gateway supports OPC UA, Modbus / TCP and Profinet protocols, and generates a standardized data format through a protocol conversion module.

[0038] On the other hand, the present invention also provides a method for implementing a cloud-edge collaborative system for industrial control, the method being used to implement the system described in any one of the above, and the method comprising:

[0039] S1. The device layer collects device operation data through the industrial device access gateway and uploads it to the edge layer;

[0040] S2. Clean, standardize and extract features from the collected real-time data;

[0041] S3. Based on the device operation data, expert experience and process specifications, construct a scenario-based knowledge model including a static knowledge layer and a dynamic knowledge layer;

[0042] S4. Convert the scenario-based knowledge model into a lightweight model adapted to edge devices;

[0043] S5. According to the hardware configuration and working scenario of the edge device, select and distribute the adapted lightweight model to the edge layer;

[0044] S6. Control each task according to the task time sensitivity level and allocate the execution location;

[0045] S7. The edge layer loads the lightweight model, generates a control decision based on the real-time data, and issues a control instruction to the device layer;

[0046] S8. The device layer receives and executes the control instruction, and at the same time continuously collects new operation data to form a closed-loop control.

[0047] The present invention has the following beneficial effects compared with the prior art:

[0048] (1) By constructing a cloud-edge collaborative system for industrial control, the present invention fully combines the computing power of the cloud and the real-time response advantage of the edge, solves the deficiencies of traditional centralized control systems in terms of real-time performance, reliability and flexibility, effectively realizes the efficient collaboration of data processing, model inference and task execution in complex industrial control scenarios, and improves the comprehensive performance of the system under variable working conditions;

[0049] (2) The scenario-based knowledge model proposed by the present invention can dynamically perform knowledge expression and update based on the professional knowledge in the industrial field and real-time working condition data through the structural design of the static knowledge layer and the dynamic knowledge layer, improving the generalization ability of the model and the adaptability to complex working conditions, and making up for the deficiency that the model in the traditional industrial control system is difficult to fully reflect the field characteristics.

[0050] (3) The present invention adopts a dedicated lightweight technology for industrial control characteristics. By means of causal perception pruning and hierarchical knowledge distillation, while maintaining the integrity of the key parameters of the model, it significantly reduces the computational complexity and resource occupancy of the model, enabling the lightweight model to efficiently adapt to resource-constrained edge devices in the industrial field, thereby improving the edge inference performance.

[0051] (4) The task scheduling mechanism of the present invention based on task time sensitivity effectively reduces the response delay of highly time-sensitive tasks and improves the deterministic scheduling performance of the overall system by distinguishing the priorities of different tasks and reasonably allocating the execution locations of tasks, meeting the real-time and reliability requirements in the industrial control scenario. Brief Description of the Drawings

[0052] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0053] Figure 1 It is the system framework diagram of the present invention;

[0054] Figure 2 It is the method flow chart of the present invention. Detailed Embodiments

[0055] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention.

[0056] As Figure 1 shown, the present invention provides a cloud-edge collaborative system for industrial control, including:

[0057] The device layer includes an industrial device access gateway, which is responsible for collecting device operation data and executing control instructions.

[0058] Edge layer, including: a data preprocessing unit for processing real-time data uploaded by the device layer; a lightweight inference engine for loading the lightweight model sent from the cloud layer and generating control decisions based on the real-time data; a deterministic scheduling engine for classifying and controlling tasks according to the task time sensitivity and allocating execution locations.

[0059] Cloud layer, including: an industrial knowledge model center for constructing a scenario-based knowledge model based on device operation data; a model lightweighting unit for converting the scenario-based knowledge model into a lightweight model; a decision distribution unit for sending the lightweight model to the edge layer.

[0060] The technical implementation process of the present invention is as follows: construct a cloud-edge collaborative system for industrial control, adopt a three-layer architecture design of device layer-edge layer-cloud layer, and effectively integrate industrial expertise and data-driven models by constructing a scenario-based knowledge model including a static knowledge layer and a dynamic knowledge layer in the cloud layer; then use dedicated lightweight technologies such as elastic quantization compression, causal perception pruning, and hierarchical knowledge distillation to convert the complex knowledge model into a lightweight model suitable for the edge computing environment; combine a task scheduling mechanism based on time sensitivity to reasonably allocate execution locations according to task characteristics; and optimize the model deployment and iteration process between the cloud and the edge through intelligent selection and incremental update strategies, and finally form a technical solution that can efficiently integrate domain knowledge, adapt to the edge computing environment, and meet the real-time requirements of industrial control.

[0061] Specifically, in an embodiment of the present invention, the industrial device access gateway supports OPC UA, Modbus / TCP, and Profinet protocols, and generates a standardized data format through a protocol conversion module.

[0062] In this embodiment, the device layer data acquisition process is mainly realized through the industrial device access gateway, which serves as the core bridge connecting industrial devices and the edge computing layer. The gateway system supports mainstream industrial communication protocols such as OPC UA, Modbus / TCP, and Profinet, and can adapt to various industrial control devices, including PLCs, DCSs, sensors, and actuators. In the data acquisition stage, the gateway first identifies the device type and communication protocol according to the preset configuration, and then establishes a communication connection with various industrial devices through the corresponding protocol interface. For devices using the OPC UA protocol, the gateway obtains device data point information through the client subscription mode; for Modbus / TCP devices, the register or coil status value is read through the polling mechanism; for Profinet devices, the IO data is obtained through real-time Ethernet frame exchange.

[0063] After the raw device data is collected, the protocol conversion module built into the industrial device access gateway immediately performs unified processing on the data format. This module first parses the data structures of different protocols, extracts the real data values and their attribute information; then performs data type conversion, mapping the data types in different protocols to a unified internal data type; then adds timestamps to the data and assigns quality labels to mark the data collection time and validity; finally, reorganizes and encapsulates the processed data according to the predefined standardized data format to generate a standardized data packet with a unified structure.

[0064] After completing the data standardization process, the industrial device access gateway transmits the data packet to the data preprocessing unit of the edge layer through a secure channel, while continuously monitoring the device status and timely collecting the change information of the operating parameters. The gateway also has a local caching function. When the network connection is unstable, it can temporarily store the collected data and automatically upload it after the connection is restored to ensure data integrity.

[0065] Specifically, in an embodiment of the present invention, the construction process of the scenario-based knowledge model in the industrial knowledge model center includes:

[0066] Step 1: Obtain expert experience, device process specifications, and threshold parameters through the industrial device access gateway, establish an industrial entity and relationship classification system, and obtain an industrial domain concept framework.

[0067] Specifically, the expert experience includes device maintenance procedures and fault diagnosis processes. The device process specifications include operation manuals and process flowcharts. The threshold parameters include temperature upper limits and pressure ranges. After obtaining these data, perform semantic analysis on them to extract core concepts and relationships. Subsequently, construct an industrial entity classification system, including device type hierarchies (such as pump → centrifugal pump → variable frequency centrifugal pump), component composition relationships, process flow stages, etc. Then establish a relationship classification system, including physical connection relationships (such as pipeline connections), functional dependency relationships (such as control loops), causal influence relationships (such as the impact of temperature on product quality), etc. Finally, form an industrial domain concept framework.

[0068] Step 2: Summarize and clean the historical data uploaded from the device layer, perform standardization processing, extract time series statistical features and frequency domain features, and obtain a set of device status characteristic indicators.

[0069] Specifically, data cleaning includes removing outliers, filling in missing values, removing redundant records, etc. Standardization processes unify data with different dimensions into a comparable scale, specifically using the min-max normalization method (for continuous variables) and one-hot encoding (for discrete variables). Subsequently, multi-dimensional time-series statistical features are extracted, including statistics such as mean, standard deviation, kurtosis, skewness, quartiles, etc., as well as dynamic features such as rate of change and trend slope. At the same time, frequency-domain feature extraction is carried out, and the fast Fourier transform (FFT) is applied to analyze the spectral features of signals such as vibration and current, and indicators such as the main frequency component, frequency energy distribution, and harmonic ratio are extracted. Finally, a complete set of equipment status feature indicators is obtained.

[0070] Step 3: Based on the industrial domain concept framework, use ontology description language to construct an industrial domain ontology, extract industrial mechanism models from expert experience and equipment process specifications, transform them into parameterized expressions, use the industrial domain ontology as the framework to construct a knowledge graph, and embed the parameterized expressions into the knowledge graph to form a static knowledge layer.

[0071] Specifically, by analyzing industrial literature, standards and specifications, and expert knowledge, a top-down and bottom-up combined method is used to construct a concept hierarchy tree, define top-level concept categories such as equipment classes, parameter classes, process classes, and fault classes, and perform multi-level refinement. At the same time, design the relationship types between concepts, including basic semantic relationships such as "is a kind of", "part-whole", etc., and domain-specific relationships such as "monitoring parameter", "influence", and "precondition", etc. Define semantic constraint conditions for each relationship type to ensure the rationality and consistency of its use. On this basis, define data attributes and object attributes for each concept, clarify the value range, unit, and physical meaning of the attributes, and establish the dependency relationships between the attributes. Finally, use standard languages such as OWL / RDF for ontology formal description, establish unique identifiers and naming spaces for concepts, construct a domain vocabulary and a synonym set, and process the equivalent relationships between terms.

[0072] Extract the equipment working principle and performance parameters from the equipment manufacturer's manual, obtain process parameter constraints and operating procedures from industry standards and specifications, capture implicit experience knowledge and heuristic rules through expert interviews, and extract diagnosis and treatment knowledge from historical fault cases. According to the theoretical basis, the mechanism models are divided into three categories: physical models, empirical models, and hybrid models. Then, the mechanism models are parameterized and expressed in the following forms:

[0073]

[0074] where P(x,θ) represents the output process parameters (such as pressure, temperature, flow rate, etc.); x is the state variable vector, including various observable quantities of the equipment operation; θ is the set of parameters to be calibrated, which is used to adjust the adaptability of the model to specific equipment; f k(x) is the k-th physical constraint function, which describes the basic physical laws, for example: Describing the square root relationship between flow rate and pressure difference, Describing the exponential relationship between temperature and reaction rate; K is the total number of physical constraint functions; γ k is the weight of the k-th physical constraint, reflecting the importance of this constraint in a specific scenario; C is the total number of domain knowledge functions; ψ c (x,θ c ) is the c-th domain knowledge function, for example: ψ1(x,θ1) = θ 11 ·max(0, x5 - θ 12 ), θ 11 represents the wear impact factor, θ 12 represents the wear critical threshold, and this function describes the impact of the equipment wear threshold. ψ2(x,θ2) = θ 21 ·sin(θ 22 ·x6), θ 21 represents the amplitude coefficient of the periodic influence, θ 22 represents the frequency coefficient of the periodic influence, and this function describes the periodic process influence; λ c is the weight of the c-th domain knowledge, reflecting the confidence of expert experience in the model; θ c is the subset of adjustable parameters of the c-th domain knowledge function. These parametric expressions are embedded into the ontology structure through the attribute relationships of the knowledge graph, and finally form the static knowledge layer.

[0075] Step 4: Use the equipment status characteristic index set to calibrate the initial parameters of the parametric model in the static knowledge layer, determine the optimal parameters, and form the dynamic knowledge layer.

[0076] Specifically, the initial parameter calibration of the parametric model is optimized using the following objective function to determine the optimal parameters:

[0077]

[0078] In the formula, Ω is the objective function; θ is the set of parameters to be calibrated; x i and y i respectively represent the input features and the corresponding target observed values in the historical data; N is the number of features; represents the predicted output of the parametric model in the static knowledge layer for the input x i ; ω i is the data quality weight; is the residual function of the r-th industrial mechanism constraint, and its number is R; δ r is the weight of the mechanism constraint in the objective function; κ is the adjustment coefficient of the prior deviation; D(θ,θ ref ) is the prior deviation; θref is the parameter reference value.

[0079] Based on the objective function Ω(θ), a phased iterative algorithm is used for parameter calibration; let the initial parameter be θ (0) , and each iteration is divided into the following steps:

[0080] A1. Calculate the gradient vector under the current parameter:

[0081]

[0082] A2. Update the parameter:

[0083]

[0084] where μ is the learning rate, which can be adaptively adjusted according to the scale of historical data and the complexity of the mechanism model.

[0085] A3. Judge the convergence condition. If ‖θ (t+1) -θ (t) ‖ or the decrease amplitude of Ω(θ (t+1) ) is lower than the threshold, stop the iteration; otherwise, continue the next round.

[0086] After the iteration converges, the obtained θ * is the optimal parameter set, which is applied to the parameterized model of the static knowledge layer to generate the dynamic knowledge layer.

[0087] Step 5. Conduct multi-scale time window analysis on the device state characteristic index set, and combine it with the dynamic knowledge layer to construct a directed graph of causal relationships between parameters, and integrate the causal inference structure into the dynamic knowledge layer.

[0088] Specifically, during the operation of the device, a large amount of historical state characteristic data collected and transmitted from the device layer and the edge layer to the cloud layer is used for multi-scale analysis. These data include process parameters (such as temperature, pressure, flow rate, etc.) and device performance indicators (such as vibration, energy consumption, fault frequency), and also cover the time series information of device operation. According to the change rules of different device state characteristics, time windows of different scales are set, including short-term windows (seconds to minutes), medium-term windows (hours to shift levels), and long-term windows (days to months). Within each time window, features are calculated through a sliding window to extract statistical and time-domain features such as mean, standard deviation, change rate, peak value, etc., and frequency-domain features are extracted through wavelet transform or FFT (Fast Fourier Transform). To avoid the computational complexity caused by the high-dimensional feature space, dimensionality reduction algorithms such as principal component analysis (PCA) or autoencoders are combined to reduce the dimensionality of the features extracted under the multi-scale time window, while retaining the key features that can best characterize the device state. The result forms a data set that can be used for subsequent causal construction.

[0089] Based on the modeling output of the dynamic knowledge layer, the present invention constructs a directed causal graph model based on the data analysis ability and knowledge support of cloud layers. Its specific implementation includes the following processes:

[0090] (1) Determine the prior structure of causal relationships

[0091] The dynamic knowledge layer contains a causal relationship framework constructed based on domain expert experience, historical law summary, or physical parameterization models. By extracting the physical coupling relationship or industrial mechanism relationship between parameters in the static knowledge layer, a preliminary causal candidate network is generated.

[0092] (2) Data-driven causal relationship correction and supplementation

[0093] Based on the prior structure, by analyzing the state feature data in multi-scale time windows, use time-series-based causal discovery algorithms to supplement and correct the causal framework. It mainly includes: Granger causality analysis: testing whether the past information of one variable can predict another variable to identify lagged causal effects. Information theory methods: quantifying the causal strength between parameters through mutual information or conditional mutual information to avoid the limitations of linear methods. Structure learning algorithms: applying algorithms such as PC algorithm and FCI algorithm to recursively identify the backbone structure of the causal model. Dynamic constraint detection: combining the causal priors in the dynamic knowledge layer to ensure that the discovered causal relationships conform to physical laws or domain constraints and prevent pseudo-association problems in the causal chain.

[0094] (3) Construct a directed causal graph

[0095] Through the above data-driven methods and dynamic knowledge constraints, the causal relationships between the selected parameters are represented in the form of a directed graph. The nodes of the graph are device state parameters (such as temperature, pressure) or performance indicators (such as failure rate, efficiency), and the edges represent the causal relationships and directions. The weights of the edges can represent the causal strength. The result forms a high-confidence parameter causal relationship network.

[0096] After that, integrate the causal reasoning structure into the dynamic knowledge layer to expand its function. The integration process first establishes a mapping relationship between the nodes of the causal network and the model parameters of the dynamic knowledge layer to ensure semantic consistency between the two. Then construct a causal reasoning engine to support forward causal reasoning (predicting the influence range of parameter changes) and backward causal reasoning (tracing the root cause of abnormal parameters).

[0097] Step 6: Use density-based clustering methods to identify typical working scenarios, and for each working scenario, train a prediction model that integrates the static knowledge layer and the dynamic knowledge layer using historical operation data to obtain a complete set of scenario-based knowledge models.

[0098] Specifically, the clustering algorithm uses the DBSCAN algorithm, regarding historical operation data as data points in a high-dimensional feature space. Each data point contains multi-dimensional features such as device status parameters, environmental condition parameters, and production indicators. The DBSCAN algorithm is used for clustering analysis. This algorithm can identify density regions of arbitrary shapes. In the clustering process, an adaptive density threshold determination method is adopted to dynamically adjust the density parameter according to the data distribution characteristics of different industrial scenarios.

[0099] After clustering is completed, the identified density regions need to be further characterized and semantically processed. Specifically, feature analysis is performed on each cluster center and its coverage area, and scene feature descriptors are extracted, including the mean vector of process parameters, the parameter fluctuation range, the intensity of correlation between parameters, and causal relationship features, etc. For each identified scene, a multi-dimensional feature portrait is generated, and through matching with process specifications and operation manuals in the static knowledge layer, semantic annotation of the scene is achieved, such as "normal production state", "start-up process", "load adjustment", or "equipment degraded operation", etc. At the same time, a scene conversion relationship graph is established to record the conversion conditions and probabilities between different working scenes, forming a working scene state transition model.

[0100] For each identified typical working scene, a three-level fusion architecture is designed: parameter-level fusion, model-level fusion, and inference-level fusion. Parameter-level fusion adjusts the weight coefficients of the parameterized expressions in the static knowledge layer to make it more conform to the actual data distribution in a specific scene; model-level fusion adopts an ensemble learning framework, taking the parameterized models in the static knowledge layer as one of the base learners, and jointly constructing a hybrid model ensemble with the machine learning models in the dynamic knowledge layer; inference-level fusion combines a causal inference structure to dynamically select the most suitable knowledge source and inference path during the inference process.

[0101] Based on the above fusion mechanism, a prediction model dedicated to the scene is constructed. The training process adopts a two-stage learning strategy: in the first stage, the parameters of the static knowledge layer are fixed, and only the parameters of the dynamic knowledge layer are trained to ensure that the model behavior conforms to physical laws; in the second stage, under the condition of meeting physical constraints, the parameter weights of the static knowledge layer are finely tuned to achieve a fine match between the model and the actual scene.

[0102] Specifically, the prediction model that fuses the static knowledge layer and the dynamic knowledge layer adopts the following hybrid inference mechanism:

[0103] y = σ[β(s)·Φ static (x,θ static )+(1 - β(s))·Φ dynamic (x,θ dynamic ,H(t))]

[0104] Where y is the predicted output of the model; σ is the activation function; β(s) is the knowledge fusion coefficient in scenario s, with a value range of [0,1]; s is the current scenario identifier; Φ static is the inference function based on the static knowledge layer; θ static is the set of static knowledge layer parameters; Φ dynamic is the inference function based on the dynamic knowledge layer; θ dynamic is the set of dynamic knowledge layer parameters; H(t) is the temporal context information, including the time series of historical data.

[0105] Specifically, in an embodiment of the present invention, the process of converting the scenario-based knowledge model into a lightweight model includes:

[0106] Step A: Perform mixed-precision quantization on the static knowledge layer parameters and dynamic knowledge layer parameters in the scenario-based knowledge model. The static knowledge layer retains the high-precision representation of the parameters related to key physical constraints, and the dynamic knowledge layer parameters adopt dynamic range quantization based on sliding window statistics to obtain a preliminary lightweight model.

[0107] Specifically, perform precision quantization on the static knowledge layer and dynamic knowledge layer parameters of the scenario-based knowledge model: The static knowledge layer retains the key parameters related to physical constraints as high precision (FP32 or FP16) to ensure the accurate expression of physical rules. The dynamic knowledge layer parameters adopt dynamic range quantization, and are quantized to low precision (such as INT8) by analyzing their distribution range to reduce storage and calculation overhead. At the same time, compensation training is performed for the errors introduced by quantization to ensure that the model performance is not significantly affected.

[0108] Step B: Based on the directed graph of causal relationships between parameters, calculate the influence factors of each parameter node in the preliminary lightweight model, dynamically adjust the pruning threshold according to the clustering results of typical working conditions, and remove the redundant parameters with influence lower than the threshold to obtain a structure-simplified model.

[0109] Specifically, combine the directed graph of causal relationships and the clustering results of typical working conditions to prune the parameters: Calculate the causal influence of each parameter node, and determine the pruning priority according to the influence factor. Dynamically adjust the pruning threshold for complex working conditions, remove redundant parameters, and retain the paths related to key parameters at the same time. Fine-tune the model after pruning to restore performance to obtain a structure-simplified model.

[0110] Step C: Use the original scenario-based knowledge model as the teacher model and the structure-simplified model as the student model, and design a distillation loss function to guide the student model to learn the knowledge of the teacher model to obtain a distillation-optimized model.

[0111] Specifically, knowledge distillation transfers the knowledge of the original model (teacher) to the lightweight model (student), and performs the transfer by designing a multi-objective distillation loss function:

[0112] L total = λ1·L soft + λ2·L feat + λ3·L phys + λ4·L hard

[0113] L soft = D KL (σ(z T / T), σ(z S / T))

[0114]

[0115] where L soft is the soft target distillation loss, which measures the difference in the output distributions of the teacher model and the student model using KL divergence, and introduces a temperature parameter T to adjust the softening degree of the output. z T and z S are the outputs of the teacher and student models respectively; L feat is the feature distillation loss, which selects key intermediate layer features for matching to ensure that the internal representation of the student model is consistent with that of the teacher model. FSP is the feature space transformation function, and A i , B i are the intermediate layer features of the teacher model, and A' i , B' i are the corresponding features of the student model; L phys is the physical consistency loss, which ensures that the student model satisfies the physical constraint conditions and maintains the physical interpretability of the industrial model, is the value of the industrial mechanism function on the parameters of the student model; L hard is the hard target loss, which is the traditional supervised learning loss. The weight coefficients λ1 to λ4 are dynamically adjusted according to the performance on the validation set to control the relative importance of each loss term. The distillation training adopts a three-stage strategy: first, perform basic knowledge transfer and freeze the backbone network of the student model; then, perform feature matching and gradually unfreeze some layers; finally, perform full model fine-tuning to optimize the comprehensive loss.

[0116] Step D: For the distilled and optimized model, optimize the network structure according to the hardware characteristics of different edge devices to generate the final lightweight model.

[0117] Specifically, optimize the lightweight model according to the hardware characteristics of different edge devices: optimize and fuse the model operators at the operator level for the computing power, memory limit, and dedicated accelerator characteristics of the device. Adjust the data layout and memory allocation strategy to minimize resource occupancy to the greatest extent. Design a heterogeneous device scheduling strategy to achieve load distribution among multiple computing units (such as CPU, GPU, NPU) and generate a dedicated model version adapted to the device.

[0118] In this embodiment, after generating the lightweight model, the decision-making distribution unit distributes the model according to the operating status of industrial equipment, the hardware capabilities of edge devices, and the actual application scenarios. The strategies are as follows:

[0119] According to the industrial process and actual operating conditions where the edge device is located, select the lightweight model that matches the current scenario from the set of scenario-based knowledge models in the cloud. By analyzing the scenario transition state diagram, predict the possible working conditions that may occur next, and distribute the matching standby scenario-based models in advance to ensure the continuity of processing.

[0120] Analyze the hardware capabilities of edge devices, including computing capabilities (CPU, GPU, NPU, etc.), memory size, power consumption limits, etc. Automatically select the lightweight model version that has been optimized for hardware (such as INT8 quantization version, FP16 acceleration version) to ensure that the model adapts to hardware resources and achieves the best efficiency.

[0121] In an embodiment of the present invention, the edge layer receives the lightweight model sent from the cloud and the real-time data uploaded by the device layer. First, preprocess the real-time data, and the preprocessing process includes data cleaning, normalization processing, and feature extraction. Then, the lightweight inference engine performs online learning on the lightweight model according to the real-time data. The specific process is as follows:

[0122] First, use the sliding window method to process the real-time data, and the window size is adaptively adjusted according to the dynamic characteristics of the device. Calculate the quality score q(t) for the data in each window, and its calculation formula is:

[0123] q(t) = α c ·C(t) + α v ·V(t) + α s ·S(t)

[0124] In the formula, C(t) is the data integrity score, V(t) is the data validity score, S(t) is the data stability score, and α c 、ɑ v 、α s are the corresponding weight coefficients. Among them, the example calculation of C(t) is: C(t) = 1 - (the number of missing values in the window / the total number of data points in the window); the example calculation of V(t) is: V(t) = the number of valid data points in the window / the total number of data points in the window); the example calculation of S(t) is: S(t) = 1 - min(1, σ(t) / σ threshold ), σ(t) is the standard deviation of the data in the window, and σ threshold is the predefined standard deviation threshold.

[0125] Based on the data quality score, the following adaptive parameter optimization update mechanism is adopted in the online learning process of the model:

[0126]

[0127] In the formula, θ j (t) is the j-th parameter estimation value at time t, where j ∈ {1, 2, ..., N} and N is the total number of parameters; η(t) is the adaptive learning rate, where η0 is the initial learning rate and δ is the decay coefficient; L s (·) is the loss function calculated on a sliding window of length s, which is defined as: where W is the window size, y i is the actual observed value, referring to the physical quantity measurement value actually collected from industrial equipment, such as temperature, pressure, flow rate, etc., is the model prediction value, referring to the prediction value of the same physical quantity using the model with the current parameter θ j (t);; is the gradient of the loss function with respect to the parameter θ j ; q(t) is the data quality scoring factor, with a value range of [0, 1]; α is the regularization coefficient; R e (·) is the elastic constraint function to prevent the parameters from deviating too much from the initial calibration values, where ρ is the regression intensity coefficient and v is the regression form index; is the calibration value of the parameter θ j , that is, the optimal calibrated parameter value in the cloud.

[0128] According to the above formula, the lightweight model parameters are continuously adjusted through an online learning mechanism to enable the model to reflect the latest working conditions. Its fine-tuning process can be summarized as follows: within a sliding window of fixed length (such as data in the recent few seconds or minutes), calculate the difference between the predicted output of the model and the actual measurement data to generate a loss metric. Use this loss to calculate the gradient information, and then fine-tune the parameters of the lightweight model.

[0129] After the model parameters are fine-tuned in real time to the current optimal state, use the model to generate control decisions, and the process is as follows:

[0130] The updated lightweight model receives the latest input data and predicts the key operating parameters of the equipment. For example, predict the pressure, temperature, flow rate, etc. of the equipment in the future for a period of time.

[0131] According to the production requirements and process specifications in the industrial field, a series of decision rules and objective functions are preset. Including: if it is predicted that the equipment temperature will exceed the safety threshold, then trigger the cooling mechanism; if it is predicted that the energy consumption will rise abnormally, then adjust the control strategy to reduce the load or optimize the operating curve; if the predicted working condition is stable, then maintain the current state or conduct energy efficiency optimization.

[0132] The predefined rules will perform comparison and trade - off based on the real - time prediction results, and combine with the objective function (such as minimum energy consumption, optimal response time or process stability) to calculate the optimal control action.

[0133] Based on the output of the current model, the prediction results are fused with historical decisions and actual control effects, and a continuous and stable control curve is formed using a smoothing method. Combining the physical and process constraints of the equipment, upper and lower limits and conversion rates are set for the control decision to ensure that the output control instructions are within the actual executable range.

[0134] Specifically, in an embodiment of the present invention, the deterministic scheduling engine controls tasks according to the task time sensitivity level and allocates execution locations, specifically set as:

[0135] Industrial control tasks are divided into three levels according to time sensitivity: high, medium, and low. High - time - sensitivity tasks are preferentially executed at the edge layer; medium - time - sensitivity tasks are dynamically determined to be executed at the edge layer or the cloud layer according to the current system load; low - time - sensitivity tasks are preferentially executed at the cloud layer.

[0136] The deterministic scheduling engine receives task requests from the device layer or the edge data pre - processing unit in real time, and collects basic information about the tasks. A set of static decision rules are preset inside the system, such as high - priority tasks passing directly through to the edge. At the same time, combined with dynamic indicators such as the current edge layer load, communication network status, task queue length and other information, a multi - dimensional scheduling decision model is constructed. The scheduling engine uses a priority - based decision model and a prediction algorithm to finely divide tasks: for high - sensitivity tasks, the task requests will immediately suspend the dedicated resources of the edge layer, and no matter how high the edge load is, it is ensured to be executed at the fastest speed. Medium - sensitivity tasks are dynamically scheduled by the engine according to the current load status of the edge and the cloud layer, and a load - balancing algorithm is used, such as considering task waiting time, system response speed and data transmission delay, to determine the specific execution location. Low - sensitivity tasks usually enter the cloud task queue and use the powerful computing power of the cloud for batch processing or scheduled execution to ensure the optimal utilization of the overall system resources. During the scheduling process, the engine monitors all the allocated tasks throughout the process, and collects the task execution status and feedback data in real time. Through the monitoring feedback, if it is found that there is a delay or the task is not executed as expected, the scheduling unit can dynamically adjust the resource allocation strategy or perform task migration when necessary.

[0137] In summary, the cloud-edge collaboration system of the present invention first constructs a scenario-based knowledge model in the industrial knowledge model center based on device operation data, and optimizes and generates a lightweight model suitable for edge devices through a model lightweighting unit; subsequently, the decision distribution unit intelligently distributes the optimized model to the edge layer, and the edge lightweight inference engine fine-tunes the model online based on real-time data and generates control decisions. Finally, under the coordination of the deterministic scheduling engine, hierarchical scheduling is performed according to the time sensitivity of tasks, realizing efficient collaboration and dynamic adjustment between the cloud and edge devices, and ensuring the safety, real-time, and stability of the industrial production process.

[0138] In addition, as Figure 2 shown, the present invention also provides a method for implementing a cloud-edge collaboration system for industrial control. The method is used to implement the system described in any one of the above, and the method includes:

[0139] S1. The device layer collects device operation data through an industrial device access gateway and uploads it to the edge layer;

[0140] S2. Clean, standardize, and extract features from the collected real-time data;

[0141] S3. Based on device operation data, expert experience, and process specifications, construct a scenario-based knowledge model including a static knowledge layer and a dynamic knowledge layer;

[0142] S4. Convert the scenario-based knowledge model into a lightweight model adapted to edge devices;

[0143] S5. According to the hardware configuration and working scenario of the edge device, select and distribute the adapted lightweight model to the edge layer;

[0144] S6. Hierarchically control each task according to the task time sensitivity and allocate the execution location;

[0145] S7. The edge layer loads the lightweight model, generates control decisions based on real-time data, and sends control instructions to the device layer;

[0146] S8. The device layer receives and executes the control instructions, and at the same time continuously collects new operation data to form a closed-loop control.

[0147] The implementation process of this method can be summarized as follows: First, the device layer collects and uploads the operation data of industrial devices through the gateway, and then the data is cleaned, standardized, and feature-extracted at the edge layer; Next, a scenario-based knowledge model is constructed based on device data, expert experience, and process specifications, and it is transformed into a lightweight model suitable for edge hardware; According to the hardware configuration and actual working scenario of the edge device, the appropriate lightweight model is selected and dispatched; After that, through the deterministic scheduling engine, tasks are hierarchically scheduled according to the time sensitivity of the tasks, and the execution locations are reasonably allocated; The edge layer loads this lightweight model, generates control decisions using real-time data, and feeds back the control instructions to the device layer; Finally, the device layer executes the instructions and continuously collects new data, thus forming a closed-loop control system of data collection, decision generation, and execution feedback.

[0148] The above is only the preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A cloud-edge collaborative system for industrial control, characterized in that It includes: The device layer includes an industrial device access gateway, which is responsible for collecting device operation data and executing control instructions. The edge layer includes: a data preprocessing unit for processing real-time data uploaded by the device layer; a lightweight inference engine for loading the lightweight model sent from the cloud layer and generating control decisions based on real-time data; a deterministic scheduling engine for classifying control tasks according to the task time sensitivity and allocating execution locations. The cloud layer includes: an industrial knowledge model center for constructing a scenario-based knowledge model based on device operation data; a model lightweighting unit for converting the scenario-based knowledge model into a lightweight model; a decision distribution unit for sending the lightweight model to the edge layer.

2. The cloud-edge collaborative system for industrial control according to claim 1, wherein The construction process of the scenario-based knowledge model includes: Step 1: Obtain expert experience, device process specifications, and threshold parameters through the industrial device access gateway, establish an industrial entity and relationship classification system, and obtain an industrial domain concept framework. Step 2: Summarize and clean the historical data uploaded by the device layer, perform standardization processing, extract time series statistical features and frequency domain features, and obtain a device status feature index set. Step 3: Based on the industrial domain concept framework, use ontology description language to construct an industrial domain ontology, extract an industrial mechanism model from expert experience and device process specifications, convert it into a parameterized expression, construct a knowledge graph with the industrial domain ontology as the skeleton, and embed the parameterized expression into the knowledge graph to form a static knowledge layer. Step 4: Use the device status feature index set to calibrate the initial parameters of the parameterized model in the static knowledge layer, determine the optimal parameters, and form a dynamic knowledge layer. Step 5: Perform multi-scale time window analysis on the device status feature index set, combine it with the dynamic knowledge layer, construct a directed graph of causal relationships between parameters, and integrate the causal inference structure into the dynamic knowledge layer. Step 6: Adopt a density-based clustering method to identify typical working scenarios, and use historical operation data to train a prediction model that fuses the static knowledge layer and the dynamic knowledge layer for each working scenario to obtain a complete set of scenario-based knowledge models.

3. The cloud-edge collaborative system for industrial control according to claim 2, characterized in that, The industrial mechanism model in the static knowledge layer adopts the following parameterized expression: Wherein, P(x,θ) represents the output process parameter; x is the state variable vector; θ is the set of parameters to be calibrated; f k (x) is the k-th physical constraint function; K is the total number of physical constraint functions; γ k is the weight of the k-th physical constraint; C is the total number of domain knowledge functions; ψ c (x,θ c ) is the c-th domain knowledge function; λ c is the weight of the c-th domain knowledge; θ c is a subset of adjustable parameters of the c-th domain knowledge function.

4. A cloud-edge collaboration system for industrial control according to claim 3, characterized in that, In Step 4, the following objective function is used to optimize the initial parameter calibration of the parameterized model to determine the optimal parameters: where, Ω is the objective function; θ is the set of parameters to be calibrated; x i and y i represent the input features and the corresponding target observations in the historical data respectively; N is the number of features; represents the predicted output of the parametric model in the static knowledge layer for the input x i ; ω i is the data quality weight; is the residual function of the r-th industrial mechanism constraint, and the number thereof is R; δ r is the weight of the mechanism constraint in the objective function; κ is the adjustment coefficient of the prior deviation; D(θ, θ ref ) is the prior deviation; θ ref is the parameter reference value.

5. A cloud-edge collaborative system for industrial control according to claim 4, characterized in that The prediction model that fuses the static knowledge layer and the dynamic knowledge layer adopts the following hybrid inference mechanism: y = σ[β(s)·φ static (x,θ static )+(1-β(s))·Φ dynamic (x,θ dynamic ,H(t))] In the formula, y is the model prediction output; σ is the activation function; β(s) is the knowledge fusion coefficient under scenario s, and its value range is [0,1]. s is the current scene identifier; Φ static is the inference function based on the static knowledge layer; θ static is the set of static knowledge layer parameters; Φ dynamic is an inference function based on the dynamic knowledge layer; θ dynamic is a set of dynamic knowledge layer parameters; H(t) is the temporal context information, including the time series of historical data.

6. The cloud-edge collaboration system for industrial control according to claim 2, wherein, The process of converting the scenario-based knowledge model into a lightweight model includes: Step A: Perform mixed-precision quantization on the parameters of the static knowledge layer and the dynamic knowledge layer in the scenario-based knowledge model. The static knowledge layer retains the high-precision representation of the parameters related to key physical constraints, and the parameters of the dynamic knowledge layer adopt dynamic range quantization based on sliding window statistics to obtain a preliminary lightweight model. Step B: Based on the directed graph of causal relationships between parameters, calculate the influence factors of each parameter node in the preliminary lightweight model, dynamically adjust the pruning threshold according to the clustering results of typical working conditions, and remove redundant parameters with influence lower than the threshold to obtain a structure-simplified model. Step C: Use the original scenario-based knowledge model as the teacher model and the structure-simplified model as the student model, and design a distillation loss function to guide the student model to learn the knowledge of the teacher model, obtaining a distillation-optimized model. Step D: For the distillation-optimized model, optimize the network structure according to the hardware characteristics of different edge devices to generate the final lightweight model.

7. The cloud-edge collaborative system for industrial control according to claim 1, characterized in that Control tasks according to the task time sensitivity level and allocate execution locations, specifically including: Classify industrial control tasks into three levels of high, medium, and low according to time sensitivity; high-time-sensitivity tasks are preferentially executed at the edge layer; medium-time-sensitivity tasks are dynamically determined to be executed at the edge layer or the cloud layer according to the current system load; low-time-sensitivity tasks are preferentially executed at the cloud layer.

8. The cloud-edge collaborative system for industrial control according to claim 2, wherein The lightweight model at the edge layer adaptively optimizes parameters according to real-time data, and adopts the following update mechanism: where, θ j (t) is the estimated value of the j-th parameter at time t, j ∈ {1, 2, ..., N}, and N is the total number of parameters; η(t) is the adaptive learning rate; L s (·) is the loss function calculated on a sliding window of length s; is the gradient of the loss function with respect to the parameter θ j ; q(t) is the data quality scoring factor, with a value range of [0, 1]; α is the regularization coefficient; R e (·) is the elastic constraint function; is the calibrated value of the parameter θ j .

9. A cloud-edge collaborative system for industrial control according to claim 1, characterized in that, The industrial device access gateway supports OPC UA, Modbus / TCP, and Profinet protocols, and generates a standardized data format through a protocol conversion module.

10. A method for implementing a cloud-edge collaboration system for industrial control, characterized in that, The method is used to implement the system according to any one of claims 1-9, and the method includes: S1: The device layer collects device operation data through the industrial device access gateway and uploads it to the edge layer. S2: Clean, standardize, and extract features from the collected real-time data. S3: Based on device operation data, expert experience, and process specifications, construct a scenario-based knowledge model including a static knowledge layer and a dynamic knowledge layer. S4: Convert the scenario-based knowledge model into a lightweight model adapted to edge devices. S5: According to the hardware configuration and working scenario of the edge device, select and send the adapted lightweight model to the edge layer. S6: Classify and control each task according to the task time sensitivity level and allocate the execution location. S7: The edge layer loads the lightweight model, generates a control decision based on real-time data, and sends the control instruction to the device layer. S8: The device layer receives and executes the control instruction, and at the same time continuously collects new operation data to form a closed-loop control.

Images