Visual construction method for industrial interconnection application model

Abstract

The invention discloses a visual construction method for an industrial interconnection application model, which comprises the following steps of: redefining an interface and a structure of each model by utilizing an existing atomic model, and establishing a graphical development environment to package each model and a corresponding algorithm into a model component; a user can select a required model component through the dragging module to quickly construct an industrial interconnection application model adapted to a new scene; in combination with an input data automatic acquisition algorithm, an interconnection application model iterative output algorithm, a model component topological sorting algorithm and a system-level overall collaborative operation algorithm, collaborative operation of a model component dynamic system is realized, and a parallel training method is utilized to train and optimize model parameters to adapt to a new scene. Therefore, the development cycle of the industrial model in a new scene is remarkably shortened, the development cost is reduced, the model reuse efficiency is improved, and the application deployment speed is increased.

Description

Technical Field

[0001] The present invention belongs to the field of industrial intelligent technology. Specifically, it relates to a method for visually constructing industrial Internet application models related to the combination processing and operation optimization of intelligent algorithm component models of cloud-edge collaborative platforms. Background Art

[0002] With the rapid development of the industrial sector, atomic models for various production processes, equipment operation, quality control, energy management, fault diagnosis, and production scheduling have become relatively mature. These atomic models cover a wide range of aspects, including basic control (such as PID control and fuzzy control), data analysis (such as regression analysis and cluster analysis), optimization and decision-making (such as linear programming and dynamic programming), prediction and early warning (such as time series prediction and anomaly detection), and diagnostic reasoning (such as expert systems and fault trees).

[0003] However, under the new landscape of the Industrial Internet and intelligent manufacturing, production methods are increasingly moving towards flexibility and intelligence, demand for customized production is increasing, new processes and technologies are constantly emerging, environmental protection and energy conservation requirements are becoming increasingly stringent, and product quality standards are constantly improving. These changes have made it difficult for traditional model development methods to quickly respond to the needs of new application scenarios.

[0004] How to fully utilize existing atomic models and quickly build application models that adapt to new scenarios through standardized interface specifications, efficient interconnection mechanisms, and intelligent combination methods has become a key issue that needs to be addressed in the industrial field. By establishing an efficient model interconnection and combination mechanism. Summary of the Invention

[0005] The present invention proposes a method for visually constructing industrial interconnected application models, which fully utilizes existing atomic models and quickly constructs application models that adapt to new scenarios through standardized interface specifications, efficient interconnection mechanisms and intelligent combination methods. This can significantly shorten the model development cycle, reduce development costs, improve model reuse efficiency, accelerate application deployment, and better support digital transformation and intelligent upgrading in the industrial field.

[0006] The present invention is achieved by adopting the following technical solutions:

[0007] A visualization construction method for industrial Internet application models is proposed, including:

[0008] Build a visual industrial Internet application model;

[0009] Automatically obtain input data and iteratively input the interconnected application model

[0010] The topological sorting of model components and the overall coordinated operation of the system level are realized to realize the coordinated operation of the dynamic system of model components;

[0011] Update and optimize model parameters through parallel training based on scenario requirements.

[0012] In some embodiments of the present invention, building a visual industrial internet application model specifically includes:

[0013] Define the data path and logical relationship between each model component and its upstream and downstream model components, which are divided into predecessor interface and successor interface. The predecessor interface is used to receive information from the upstream model component, and the successor interface is used to send information to the downstream model component.

[0014] Define the structural relationship of model components and implement the execution mode of model components by establishing control relationship nodes between model components;

[0015] Add a field selector component to get streaming data for different fields required for the model component to run.

[0016] In some embodiments of the present invention, building a visual industrial internet application model specifically includes:

[0017] Establish a web-based graphical development environment, encapsulate each model and algorithm into model components through object diagrams and activity diagrams, so that the required model components can be selected by dragging the module, and the input and output relationships are connected according to the logical relationship in the application; and

[0018] Visually assemble model components according to industrial system and production process templates.

[0019] In some embodiments of the present invention, in the collaborative operation design, the dependency relationships and input and output interface connections between the components of the industrial interconnection application model are described based on the system diagram network, including:

[0020] Each node in the graph represents a model component, and the edges represent the input and output connections between model components;

[0021] Use the connection graph matrix connect_map[i, j] to indicate whether there is an input-output relationship between edge i and edge j;

[0022] The input graph matrix input_map represents the vector obtained by input_map×u, which can map the external input to the internal input edge of the network graph; where u is the external input data vector at the current moment;

[0023] The output map matrix output_map is used to represent the overall external output of the industrial Internet application model obtained by output_map×ylist; among them, ylist is a vector composed of the output values of each model component calculated at the current moment.

[0024] In some embodiments of the present invention, in the collaborative operation design, the input data required for the operation of the precursor model component is recursively extracted by inputting an automatic extraction function, including:

[0025] When the current node is a field selector, it returns the time series data of the specified field and range of the dataset;

[0026] Find all input nodes adjacent to the current node in the system graph network, and recursively call the input extraction function on each input node to obtain output sequence data;

[0027] When the current node is the target component node, the dynamic output sequence data of all input nodes is returned; otherwise, the dynamic output sequence data of all input nodes are concatenated to calculate the forward dynamic output of the current node.

[0028] In some embodiments of the present invention, in the collaborative operation design, the system dynamic operation of the industrial Internet application model during operation is obtained based on the automatic extraction of functions and status updates of model components based on input, including:

[0029] Collect all external input data of the industrial Internet application model at the current time t;

[0030] Calculate the overall output of the current state industrial interconnection application model using the external input data at the current time t;

[0031] Use the external input data at the current time t to update the status of all components in the industrial interconnection application model at the current time;

[0032] If a component with a prediction function is included, the following steps are performed to make a forward prediction for a period of time (t, t+h) before the current moment:

[0033] Collect external inputs of the interconnected application model at the current prediction time (t+i), where i <h;

[0034] Calculate the overall output of the industrial interconnection application model at the current forecast state using external input at the current forecast time (t+i);

[0035] Use the external input data at the current prediction time to update the status of all components within the industrial interconnection application model at the current prediction time (t+i);

[0036] Restore the internal state of the industrial Internet application model to the current time t, return to step 1, and continue to calculate the output stage state at the next time.

[0037] In some embodiments of the present invention, in the collaborative operation design, for the model components forming a closed loop, the transient outputs of other internal model components are used as the inputs of the current model component through the interconnected application model iterative output method, and then the components in the loop repeatedly perform iterative calculations until all internal inputs and outputs converge and remain unchanged. Specifically, the method includes:

[0038] Set the maximum number of iterations to the number of components in the interconnected application model;

[0039] When the number of iterations does not exceed the maximum, the iteration loop begins and the following steps are performed to traverse all model components in the industrial interconnection application model in the set order:

[0040] For each model component, collect all the input data of the component at the current time t and calculate the output vector ylist in the current state;

[0041] Send the output of each model component along the component connection line to the subsequent model component as input;

[0042] Monitor all input edge vectors within the interconnected application model: connect_map × ylist + input_map × u, where u is an external input data vector. End when it no longer changes, and use output_map × ylist as the output of the industrial interconnected application model at this time.

[0043] When the number of iterations exceeds the maximum value, the output detects an algebraic loop and exits abnormally.

[0044] In some embodiments of the present invention, in collaborative design, reducing the number of iterations by topologically sorting model components includes:

[0045] Calculate the out-degree and in-degree information of each node based on the system graph network;

[0046] Use the Kahn algorithm to perform topological sorting of model components in the following steps:

[0047] Select a point with an in-degree of 0 and whose successor node's input does not contain an inner edge of the ring as the starting point, and add the starting point to the queue;

[0048] When the queue is not empty, enter the loop: take out the first node of the queue, add it to the topology sequence, mark the node as visited, and add all the output edges of the node that have not been visited to the queue, and return to the loop execution;

[0049] After exiting the loop, for nodes with in-degree 0 and that have not been visited, insert them into the position before their successor node in the topological sequence.

[0050] In some embodiments of the present invention, in the parallel training design, a directed acyclic graph is used to describe the dependency relationship of the industrial Internet application model, where each node in the graph represents a model component, and the edges between the nodes represent the operation sequence or control connection between the model components; wherein the node consists of a name, a description of the operation performed, and a reference to the implemented object; the operations performed include Transform and Estimate, Transform indicates that the node is only used for data transformation, and Estimate indicates that the model of the node contains model parameters to be trained; = { Indicates that from the root node One to leaf node The assembly line path ends with Train the model parameters of the Estimate node of this path and minimize the total training loss after the data is transformed by the Transform node.

[0051] In some embodiments of the present invention, in parallel training design, the K-Fold cross-validation strategy is used to calculate the path Training performance metrics include:

[0052] Divide the input dataset into K folds of equal size;

[0053] K-1 folds of data are used to train the assembly path ,The data of the remaining folds are used for validation;

[0054] Compare the prediction results of each fold with the true value to get the total training loss;

[0055] The above process is repeated for all folds one by one, and the average loss is calculated as the final training performance metric.

[0056] Compared with the existing technology, the advantages and positive effects of the present invention are: in the visualization construction method of the industrial Internet application model proposed by the present invention, the existing atomic model is used to redefine the interface and structure of each model, and a graphical development environment is established to encapsulate each model and its corresponding algorithm as a model component, so that users can quickly build an industrial Internet application model that adapts to new scenarios by dragging the module to select the required model component; combined with the input data automatic acquisition algorithm, the interconnected application model iterative output algorithm, the model component topological sorting algorithm and the system-level overall collaborative operation algorithm, the collaborative operation of the model component dynamic system is realized, and the model parameters are trained and optimized using a parallel training method to adapt to the new scenario, thereby significantly shortening the development cycle of the industrial model under the new scenario, reducing development costs, improving model reuse efficiency, and accelerating application deployment.

[0057] Other features and advantages of the present invention will become more apparent after reading the detailed description of the embodiments of the present invention in conjunction with the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0059] Figure 1 This is a schematic diagram of the steps of the method for visually constructing an industrial interconnection application model proposed in the present invention;

[0060] Figure 2 This is a schematic diagram of the combined structure of the industrial joint application model in the present invention;

[0061] Figure 3 Schematic diagram of the training and cross-validation process of the model components in the present invention;

[0062] Figure 4 Schematic diagram of the parallel training process of the model components in the present invention. DETAILED DESCRIPTION

[0063] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0064] The method for visually constructing industrial Internet application models proposed in the present invention changes the existing industrial model of industrial model development. It can use the model assembly line to quickly build an interconnected application model in a "building blocks" manner, so that developers do not have to focus on the programming and implementation of the basic model, but instead focus on the model combination design for specific application scenarios. It involves three aspects: constructing a visual industrial Internet application model, collaborative operation, and optimized training, which are explained in detail below.

[0065] 1. Build a visual industrial Internet application model.

[0066] 1. Industrial model component definition and interface design.

[0067] Industrial model components are independent intelligent units with complex functions, including but not limited to transfer functions, dynamic models represented by state-space equations, and time series deep learning models. They characterize how industrial models change over time. In this invention, model components are constructed as standardized, universal, reusable components that can be interconnected and exchange data. The interconnection and interoperability of multiple model components form an industrial interconnected application model.

[0068] The industrial model component interface is used to define the data path and logical relationship between the component and its upstream and downstream components. It is divided into a predecessor interface and a successor interface. The predecessor interface is used to receive information from the upstream component, and the successor interface is used to send information to the downstream component.

[0069] By defining the structural relationship of model components, the combination rules between industrial model components are described, such as Figure 2 The illustrated implementation of establishing execution relationships implements the execution of model components by establishing control relationship nodes between them. Control nodes include judgment, branching, looping, and parallelism. Low-code scripts can also be embedded for more complex operational logic control, enabling model combinations such as parallel integration, weighted integration, series integration, nested integration, structural network integration, and partial method substitution integration. Therefore, the topological structure formed by interconnection usually contains loops.

[0070] Furthermore, in order to unify the data processing method, a special field selector component is added to obtain streaming data of different fields required for component operation.

[0071] 2. And the visualization construction of industrial Internet application models.

[0072] To lower the barrier to entry and improve model development efficiency, this paper establishes a web-based graphical agile development environment. Models and algorithms are encapsulated as model components through object diagrams, activity diagrams, and other forms. Users select the desired model components by dragging modules and then connect input / output relationships based on the application's logical relationships. Furthermore, the system features a model assembly line, allowing for visual assembly of model components based on industrial systems and production process templates, rapidly completing the development of industrial interconnected application models.

[0073] 2. Based on the automatic acquisition of input data, iterative output of interconnected application models, topological sorting of model components and overall collaborative operation method at the system level, the collaborative operation of the dynamic system of model components is realized.

[0074] The industrial interconnected application model demonstrates its overall functionality through the collaborative operation of model components. Since most industrial model components are dynamic models that include time factors, the combined industrial model is essentially a complex dynamic system. For this system, the present invention provides a component collaborative operation technology for a system graph network, which includes the following key algorithms: automatic acquisition of input data, iterative output of the interconnected application model, topological sorting of model components, and overall collaborative operation at the system level.

[0075] Visual assembly is a convenient operation method provided to users, while the system diagram network is a directed graph network model that describes the interdependencies between components of the industrial Internet application model and the connections between input and output interfaces. It is the internal data structure for the construction and operation of the industrial Internet application model and includes the following contents:

[0076] (1) Each node in the graph represents a model component, and the edges represent the input and output connections between components.

[0077] (2) Connection map matrix: connect_map, shape (number of input edges, number of output edges), indicating that for each input edge i, if it is the output of the jth edge, then connect_map[i, j]=1.

[0078] (3) Input graph matrix: input_map, shape (number of input edges, number of external input fields). Let u be the external input data vector at the current moment, then the vector obtained by input_map×u can map the external input to the internal input edge of the network graph.

[0079] (4) Output map matrix, output_map, shape (number of output edges, number of external output points), let ylist be the vector composed of the output values of each model component calculated at the current moment, obtained by the output edge, then output_map×ylist is the external output of the entire industrial Internet application model.

[0080] In the industrial interconnected application model, the input of a target model component is the output of one or more predecessor model components. Therefore, in the system diagram network of the interconnected application model, the present invention recursively extracts the input data required for the operation of the predecessor component by inputting an automatic extraction function. The main algorithm steps of the automatic extraction function include:

[0081] 1. If the current node is a field selector, it returns the time series data of the specified field and range of the dataset;

[0082] 2. Find all input nodes adjacent to the current node in the system graph network, and recursively call the input extraction function on each input node to obtain the output sequence data;

[0083] 3. If the current node is the target component node, the dynamic output sequence data of all input nodes is returned; otherwise, the dynamic output sequence data of all input nodes are concatenated to calculate the forward dynamic output of the current node.

[0084] Since industrial model components can be dynamic models represented by transfer functions, state-space equations, or time series deep learning models, all of which have states, the present invention utilizes a dynamic system model conversion method to convert each heterogeneous model into a unified state-space equation form, including state output and updates, and merges the operating state vectors of each component to form the state vector of the industrial interconnection application model. In order to be compatible with model components with prediction functions, it is necessary to be able to predict a period of time in advance during operation. In summary, using the input automatic extraction function and the state update of the model component, the overall steps of the system-level dynamic operation (i.e., the overall system-level coordinated operation) of the industrial interconnection application model during operation t=0,1,2…T are obtained as follows:

[0085] 1. Collect all external input data of the industrial Internet application model at the current time t;

[0086] 2. Calculate the overall output of the current state industrial interconnection application model using the external input data at the current time t;

[0087] 3. Use the external input data at the current time t to update the status of all components in the interconnected application model at the current time;

[0088] 4. If a component with prediction function is included, then continue to perform 5-7 steps for a period of time before the current moment (t, t+h) to make a forward prediction;

[0089] 5. Collect external inputs of the interconnected application model at the current prediction time (t+i), where i <h;

[0090] 6. Calculate the overall output of the interconnected application model for the current forecast state using the external input at the current forecast time (t+i);

[0091] 7. Use the external input data at the current prediction time to update the status of all components within the interconnected application model at the current prediction time (t+i);

[0092] 8. Restore the internal state of the interconnected application model to the current time t, return to step 1, and continue to calculate the output stage state at the next time.

[0093] As described above, for step 2, the overall output of the interconnected application model at the current state at time t is calculated. The output of the interconnected application model is the collection of outputs of each model component.

[0094] In most cases, model components are interconnected in a closed loop. The inputs to a model component can be either external data or the outputs of other internal model components. The latter is unknown before the initial run of the industrial interconnection application model. Therefore, multiple rounds of iteration are required, using the transient outputs of other internal model components as inputs for the current component. Iterative calculations are then repeated within each component within the loop until all internal inputs and outputs converge and remain unchanged.

[0095] The main steps of the interconnected application model iterative output algorithm designed by the present invention include:

[0096] 1. Set the maximum number of iterations to the number of components in the interconnected application model;

[0097] 2. When the number of iterations does not exceed the maximum value, the iteration loop is entered and steps 3-5 are executed to traverse all model components in the industrial Internet application model in the set order;

[0098] 3. For each model component in the industrial Internet application model, collect all the input data of the component at the current time t and calculate the output vector ylist in the current state;

[0099] 4. Send the component output along the component connection line to the subsequent model component as input;

[0100] 5. Monitor all input edge vectors within the interconnected application model: connect_map × ylist + input_map × u, where u is an external input data vector. End when it no longer changes (i.e., converges), and use output_map × ylist as the output of the interconnected application model at this time.

[0101] 6. When the number of iterations exceeds the maximum value, the output detects an algebraic loop and exits abnormally.

[0102] The above method iterates the model components within the loop over multiple rounds until convergence, resulting in a stable interconnected application model output. However, this method also imposes an additional computational burden, as each iteration requires traversing all components. In fact, the order of model components in the iterations determines the number of iterations required for model convergence. Therefore, topological sorting of the model components can be used to reduce the number of iterations and shorten the run time. Specific methods include:

[0103] 1. Calculate the out-degree and in-degree information of each node based on the system graph network;

[0104] 2. Use the Kahn algorithm (a topological sorting algorithm for directed acyclic graphs) to topologically sort the model components according to steps 3-6;

[0105] 3. Select a point with an in-degree of 0 and whose successor node's input does not contain an inner edge of the ring as the starting point, that is, a purely external input point as the starting point, and add the starting point to the queue;

[0106] 4. Enter the loop when the queue is not empty;

[0107] 5. Take out the first node of the queue, add it to the topology sequence, and mark the node as visited;

[0108] 6. If all outgoing edges of the node have not been visited, add them to the queue and return to step 5;

[0109] 7. After exiting the loop, for nodes with an in-degree of 0 and that have not been visited, insert them into the position before their successor node in the topological sequence.

[0110] The resulting topological sorting achieves the optimal structural ordering based on the directed graph of the interconnected application model. If the directed graph is acyclic, the resulting node sorting starts with only external input data and runs sequentially according to node output connections. This way, the output of each node is exactly the input of the next node, reducing the total number of iterative convergence cycles to the theoretical lower bound (1). If the system contains cycles, the nodes are accessed in the order of the loop, significantly reducing the total number of iterative convergence cycles.

[0111] 3. Update and optimize model parameters through parallel training based on scenario requirements.

[0112] In the present invention, for the industrial Internet application model constructed by assembly modeling, the model parameters are updated and optimized through parallel training to adapt to the needs of new scenarios.

[0113] The dependencies of the model assembly line can be described by a directed acyclic root graph (DAG), denoted as G(V,E). Each vertex vi∈V in the assembly graph G is called a model assembly node, representing a model component, and the edge ei∈E represents the operation sequence or control connection between vertices. The assembly node consists of a name and a description of the operation it performs and a reference to the object it implements. It consists of a tuple Nodes are labeled using a node name ID. Each node in an assembly diagram should have a unique name. The node name is a tag that allows users to provide additional information to control node behavior. Pipeline nodes (nodes that represent a specific step or operation in data processing or task execution) perform two types of operations: Transform or Estimate. The former indicates that the node is only used for data transformation and does not participate in training; the latter indicates that the node's model contains model parameters to be trained and must participate in training.

[0114] An assembly line path of graph G(V, E) is denoted as = { , is from the root node Start, go to leaf node A directed path that ends at the end of a component, that is, the path links component-related operations together.

[0115] The goal of assembly line training is to train the model parameters of the Estimate node of a user-specified path P for a given dataset D, and to minimize the total training loss after the data is transformed by the Transform node. This can be expressed as follows:

[0116] ;

[0117] Figure 3 The figure illustrates the process of calculating the training performance metric of the assembly path Pi using the K-Fold cross-validation strategy. The input dataset D is divided into K folds of equal size (see the upper left corner). Next, data from the K−1 folds are used to train the assembly path , the data from the remaining (single) fold is used for validation, and the prediction results of each fold are compared with the true value to obtain the total training loss. This process is repeated for all other folds one by one, and the average loss is calculated as the final training performance measure.

[0118] To implement parallel distributed training of the assembly line, it is necessary to design an evaluator, an optimizer module, and a task scheduling module to manage the common tasks required for distributed sampling and gradient optimization training.

[0119] Model training involves a variety of data types, including files, pictures, images, mirrors, address tags, relational data and other forms, which need to be stored in different storage nodes to ensure performance security. The present invention continues to use the distributed architecture model for management and control. Its technical characteristics are to solve complex problems by decomposing tasks step by step, designing a scheduler, and allowing multiple processors or multiple computer nodes to participate in the calculation to solve the problem. The distributed architecture system can support multiple hosts to access shared files and storage directories through the network at the same time, allowing multiple users on multiple computers to share files and storage resources. Its architecture can better support the integration, storage and processing of platform model self-built pools and other industrial model pools, and support various business applications.

[0120] The "data parallel" approach to distributed training involves splitting the data and training in parallel across multiple nodes. In the synchronous case, the gradients for different batches of data are computed separately on each node, but are averaged across each node so that consistent updates are applied to the model replica in each node, e.g. Figure 4 shown.

[0121] Data parallel distributed training includes the following steps:

[0122] Run multiple copies of the training script, each of which reads a mini-batch of data, runs it through the model components, and computes model updates (gradients).

[0123] 2. Calculate the average gradient across multiple copies;

[0124] 3. Update the model;

[0125] 4. Repeat the above steps.

[0126] The trainer takes the form of a number of built-in SGD (Stochastic Gradient Descent) strategies, including a choice between synchronous and asynchronous execution, and the user can specify an SGD strategy or let the optimizer automatically select one.

[0127] In order to make reasonable use of these gradient optimization algorithms, it is necessary to implement the Evaluator abstract class. The evaluator is a class that defines the following methods, where X is the experience vector, θ is the model parameter, and ev is the evaluator:

[0128] ;

[0129] The Evaluator abstraction has the following advantages: (1) The optimizer can make decisions based on positional information from the Evaluator. (2) The Evaluator interface fully encapsulates the model graph and loss function, allowing SGD implementations to be improved and reused in different deep learning frameworks. All gradient-based algorithms use this flexible SGD interface. (3) In task-based mode, it is easy to extract the distributed SGD abstraction because the parallelism of the SGD component can be fully encapsulated. It is also encapsulated by providing a gradient descent SGD module, which allows all deep learning libraries with composable components to use high-performance components to achieve rapid development and prototyping of new algorithms, facilitating the rapid development of new deep learning model components.

[0130] It should be noted that, in the specific implementation process, part of the above-mentioned method can be implemented by a hardware processor executing computer execution instructions in software form stored in the memory, which will not be elaborated here, and the programs corresponding to the executed actions can be stored in the system's computer-readable storage medium in software form, so that the processor can call and execute the operations corresponding to the above modules.

[0131] The computer-readable storage medium mentioned above may include volatile memory, such as random access memory; may also include non-volatile memory, such as read-only memory, flash memory, hard disk or solid-state drive; may also include a combination of the above types of memory.

[0132] The processor mentioned above can also be a collective term for multiple processing elements. For example, the processor can be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs) or other programmable logic devices (PLDs), discrete gate or transistor logic devices (LDDs), discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor, etc., and can also be a special-purpose processor.

[0133] It should be pointed out that the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by ordinary technicians in this technical field within the essential scope of the present invention should also fall within the scope of protection of the present invention.

Claims

1. A method for visually constructing an industrial Internet application model, characterized in that: include: Build a visual industrial Internet application model; Based on automatic acquisition of input data, iterative output of interconnected application models, topological sorting of model components, and overall system-level collaborative operation methods, the collaborative operation of dynamic systems of model components is achieved; Update and optimize model parameters through parallel training based on scenario requirements.

2. The method for visually constructing an industrial interconnection application model according to claim 1, Its characteristics are The construction of a visual industrial Internet application model specifically includes: Define the data path and logical relationship between each model component and its upstream and downstream model components, which are divided into predecessor interface and successor interface. The predecessor interface is used to receive information from the upstream model component, and the successor interface is used to send information to the downstream model component. Define the structural relationship of model components and implement the execution mode of model components by establishing control relationship nodes between model components; Add a field selector component to get streaming data for different fields required for the model component to run.

3. The method for visually constructing an industrial interconnection application model according to claim 1, wherein: Building a visual industrial Internet application model specifically includes: Establish a web-based graphical development environment, encapsulate each model and algorithm into model components through object diagrams and activity diagrams, so that the required model components can be selected by dragging the module, and the input and output relationships are connected according to the logical relationship in the application; and Visually assemble model components according to industrial system and production process templates.

4. The method for visually constructing an industrial Internet application model according to claim 1, wherein: In collaborative operation design, the dependencies between industrial Internet application model components and the input and output interface connections are described based on the system diagram network, including: Each node in the graph represents a model component, and the edges represent the input and output connections between model components; Use the connection graph matrix connect_map[i, j] to indicate whether there is an input-output relationship between edge i and edge j; The input graph matrix input_map represents the vector obtained by input_map×u, which can map the external input to the internal input edge of the network graph; where u is the external input data vector at the current moment; The output map matrix output_map is used to represent the overall external output of the industrial Internet application model obtained by output_map×ylist; among them, ylist is a vector composed of the output values of each model component calculated at the current moment.

5. The method for visually constructing an industrial Internet application model according to claim 4, characterized in that: In the collaborative operation design, the input data required for the operation of the precursor model component is recursively extracted by inputting the automatic extraction function, including: When the current node is a field selector, it returns the time series data of the specified field and range of the dataset; Find all input nodes adjacent to the current node in the system graph network, and recursively call the input extraction function on each input node to obtain output sequence data; When the current node is the target component node, the dynamic output sequence data of all input nodes is returned; otherwise, the dynamic output sequence data of all input nodes are concatenated to calculate the forward dynamic output of the current node.

6. The method for visually constructing an industrial Internet application model according to claim 5, characterized in that: In collaborative operation design, the system dynamic operation of the industrial Internet application model during operation is obtained by automatically extracting functions and updating the status of model components based on input, including: Collect all external input data of the industrial Internet application model at the current time t; Calculate the overall output of the current state industrial interconnection application model using the external input data at the current time t; Use the external input data at the current time t to update the status of all components in the industrial interconnection application model at the current time; If a component with a prediction function is included, the following steps are performed to make a forward prediction for a period of time (t, t+h) before the current moment: Collect external inputs of the interconnected application model at the current prediction time (t+i), where i <h; Calculate the overall output of the industrial interconnection application model at the current forecast state using external input at the current forecast time (t+i); Use the external input data at the current prediction time to update the status of all components within the industrial interconnection application model at the current prediction time (t+i); Restore the internal state of the industrial Internet application model to the current time t, return to step 1, and continue to calculate the output stage state at the next time.

7. The method for visually constructing an industrial Internet application model according to claim 4, characterized in that: In collaborative operation design, for model components that form a closed loop, the transient outputs of other internal model components are used as the inputs of the current model component through the interconnected application model iterative output method. Then, iterative calculations are repeatedly performed on each component in the loop until all internal inputs and outputs converge and remain unchanged. Specifically, Set the maximum number of iterations to the number of components in the interconnected application model; When the number of iterations does not exceed the maximum, the iteration loop begins and the following steps are performed to traverse all model components in the industrial interconnection application model in the set order: For each model component, collect all the input data of the component at the current time t and calculate the output vector ylist in the current state; Send the output of each model component along the component connection line to the subsequent model component as input; Monitor all input edge vectors within the interconnected application model: connect_map × ylist + input_map × u, where u is an external input data vector. End when it no longer changes, and use output_map × ylist as the output of the industrial interconnected application model at this time. When the number of iterations exceeds the maximum value, the output detects an algebraic loop and exits abnormally.

8. The method for visually constructing an industrial Internet application model according to claim 7, characterized in that: In collaborative design, the number of iterations can be reduced by topologically sorting model components, including: Calculate the out-degree and in-degree information of each node based on the system graph network; Use the Kahn algorithm to perform topological sorting of model components in the following steps: Select a point with an in-degree of 0 and whose successor node's input does not contain an inner edge of the ring as the starting point, and add the starting point to the queue; When the queue is not empty, enter the loop: take out the first node of the queue, add it to the topology sequence, mark the node as visited, and add all the output edges of the node that have not been visited to the queue, and return to the loop execution; After exiting the loop, for nodes with in-degree 0 and that have not been visited, insert them into the position before their successor node in the topological sequence.

9. The method for visually constructing an industrial Internet application model according to claim 1, wherein: In the parallel training design, a directed acyclic graph is used to describe the dependency relationship of the industrial Internet application model. Each node in the graph represents a model component, and the edges between nodes represent the operation sequence or control connection between model components. The node consists of a name, a description of the operation performed, and a reference to the implemented object. The operations performed include Transform and Estimate. Transform means that the node is only used for data transformation, and Estimate means that the node model contains model parameters to be trained. = { Indicates that from the root node One to leaf node The assembly line path ends with Train the model parameters of the Estimate node of this path and minimize the total training loss after the data is transformed by the Transform node.

10. The method for visually constructing an industrial Internet application model according to claim 9, characterized in that: In parallel training design, the K-Fold cross-validation strategy is used to calculate the path Training performance metrics include: Divide the input dataset into K folds of equal size; K-1 folds of data are used to train the assembly path ,The data of the remaining folds are used for validation; Compare the prediction results of each fold with the true value to get the total training loss; The above process is repeated for all folds one by one, and the average loss is calculated as the final training performance metric.

Images