Low-code configuration capability driven fault diagnosis algorithm customization system and method

Abstract

The application relates to a low-code configuration capability driven fault diagnosis algorithm customization system and method, which comprises a fault algorithm knowledge base, an algorithm customization front end, a fault algorithm kernel, an algorithm analysis micro-service platform, a visual service platform and an interactive parameter optimization system; the fault algorithm knowledge base communicates with the visual service platform, the algorithm customization front end and the algorithm analysis micro-service platform; the visual service platform, the interactive parameter optimization system and the algorithm analysis micro-service platform communicate with the algorithm customization front end; and the fault algorithm kernel communicates with the algorithm analysis micro-service platform.The application reduces the requirements for the knowledge reserves of maintenance personnel in artificial intelligence and computer programming and reduces the code repetition rate; realizes the whole life cycle management, realizes the high-performance iterative updating of the algorithm, realizes the requirements for the more flexible and reliable intelligent fault diagnosis algorithm customization, and guarantees the high precision, high reliability and fast landing of the intelligent fault diagnosis algorithm.

Description

Technical Field

[0001] This invention belongs to the field of mechanical equipment fault diagnosis technology, and in particular relates to a fault diagnosis algorithm customization system and method driven by low-code configuration capability. Background Technology

[0002] Bearings, gears, and motors are core components of modern industrial manufacturing equipment, responsible for energy conversion and transmission. However, these rotating mechanical components operate in harsh environments for extended periods, leading to frequent component failures. Equipment malfunctions caused by rotating component failures can range from impacting production to causing injuries or fatalities, while also incurring significant maintenance and repair costs. Therefore, fault diagnosis technology has become crucial for ensuring the reliable operation of mechanical equipment.

[0003] With the development of artificial intelligence, fault diagnosis technology has advanced from the traditional "manual feature extraction + expert knowledge diagnosis" stage to the "end-to-end feature extraction and fault diagnosis" stage of intelligent fault diagnosis. Intelligent fault diagnosis greatly reduces the workload of maintenance and repair and the demand for professional personnel, enabling data-driven intelligent operation and maintenance of mechanical equipment.

[0004] Low-code development platforms offer a more sophisticated and user-friendly alternative to traditional IDEs like IDEA and VS. Developers are no longer required to write code manually; instead, they can complete development tasks more efficiently through graphical drag-and-drop interfaces and parameter configuration.

[0005] Currently, traditional fault diagnosis development algorithms face the following technical challenges:

[0006] Poor adaptability to dynamic scenarios: Equipment in industrial scenarios varies greatly and has a complex structure. Traditional fault diagnosis algorithms need to be designed and developed specifically according to the characteristics of the equipment or components, resulting in poor adaptability to dynamic scenarios.

[0007] Long development cycle and poor reusability: Stable and accurate fault diagnosis algorithms require steps such as data collection, model training, algorithm optimization, and deployment. Traditional text-based coding methods greatly reduce the readability and reusability of the algorithm, thus extending the development cycle.

[0008] Poor responsiveness and slow deployment: Traditional algorithm development is a one-time process, requiring the collection of large amounts of data to ensure the algorithm's accuracy and stability before deployment. However, data collection often takes a significant amount of time, resulting in traditional algorithms being characterized by poor development responsiveness and slow deployment.

[0009] Lack of full lifecycle management: In real-world scenarios, faults are uncertain and gradual. If diagnostic algorithms are not iteratively optimized in a timely manner, the accuracy of the original diagnostic algorithms may decrease or even become invalid. Currently, the development of diagnostic algorithms is often based on the deployment and release of the algorithm as the end point, lacking full lifecycle management.

[0010] Insufficient visualization: Traditional algorithm development requires developers to visualize and encode a large number of training logs or metrics. The visualization interfaces vary greatly and are inefficient, making it difficult to objectively reflect the actual changes in the algorithm and uncover the root causes of failures.

[0011] High labor costs: Developing a diagnostic algorithm not only requires maintenance personnel to have strong engineering practice skills, but also requires strong knowledge of computer science, mathematical statistics and artificial intelligence, which increases the company's labor costs. Summary of the Invention

[0012] The purpose of this invention is to provide a fault diagnosis algorithm customization system and method driven by low-code configuration capabilities. Utilizing the concept of full-scenario, full-lifecycle management of diagnostic algorithms, it addresses the development problem of complex fault diagnosis algorithms in varying time and space (diagnostic requirements change with the actual working scenarios and service life of equipment). Through low-code development and a model-based development approach, it combines data flow, model flow, and equipment flow to achieve higher code reusability and faster model deployment (due to the complex structure of components within equipment and mutual constraints and interference between equipment, there are problems with multi-point fault diagnosis. Furthermore, the poor correlation and mutual interference between diagnostic algorithms at different points result in significant code redundancy in traditional imperative text coding development schemes. The separation of data acquisition and model design strategies leads to difficulties in model deployment). By introducing a low-code development scheme with graphical programming and parameter configuration, the algorithm design process is made visible. Visualization of training and testing logs and indicators ensures real-time model visualization, helping maintenance personnel analyze and handle faults in real time.

[0013] This invention provides a fault diagnosis algorithm customization system driven by low-code configuration capabilities, including a fault algorithm knowledge base, an algorithm customization front-end, a fault algorithm kernel, an algorithm parsing microservice platform, a visualization service platform, and an interactive parameter optimization system; the fault algorithm knowledge base communicates with the visualization service platform, the algorithm customization front-end, and the algorithm parsing microservice platform; the visualization service platform, the interactive parameter optimization system, and the algorithm parsing microservice platform communicate with the algorithm customization front-end; and the fault algorithm kernel communicates with the algorithm parsing microservice platform.

[0014] The fault algorithm knowledge base includes an expert knowledge base and an algorithm database. The expert knowledge base stores expert knowledge graph data containing spatiotemporal modeling information of faulty components. The algorithm database stores algorithm development data, including fault data, model parameters, training configurations, test results, and deployment information.

[0015] The algorithm customization front end includes a web client and a desktop client, both of which share a set of algorithm visualization programming system, which is used to encapsulate each algorithm in the fault diagnosis algorithm kernel into a task node and represent it with a visual graphic, including task node unit, process design unit and result visualization unit;

[0016] The fault diagnosis algorithm kernel includes six sub-modules: data processing, model design, training configuration, testing and verification, deployment optimization, and extension components.

[0017] The algorithm parsing microservice platform includes an algorithm process parsing system and an algorithm microservice platform. The algorithm process parsing system is used to parse the diagnostic algorithm development configuration with a workflow engine as the main body, and parse it into a process sequence composed of organizational structure, process, node, and redirection rules. The fault algorithm microservice platform is used to register the fault diagnosis algorithm development kernel to obtain microservice instances, provide algorithm microservices to the algorithm process parsing system, and retrieve corresponding fault data, algorithm models, and parameter configuration data from the fault algorithm database to complete the diagnostic algorithm development work.

[0018] The visualization service platform includes a visualization module, a data retrieval module, and a custom module. The visualization module consists of an out-of-the-box component library, used to render various commonly used charts and models. The data retrieval module is used to retrieve relevant data from the fault algorithm support library according to the needs of the visualization module. The custom module provides users with extended APIs for custom rendering components, allowing them to complete custom rendering by importing a small amount of code in a specified format.

[0019] The interactive parameter optimization system includes a search configuration module, a learning optimization module, a performance analysis module, and a proxy interaction module. The search configuration module configures the search space and enumeration type for hyperparameters involved in algorithm development. The learning optimization module performs algorithm self-optimization through embedded optimizers, including grid search, random search, Bayesian optimization, and reinforcement learning optimizers. The performance analysis module performs prior analysis on the performance of a single parameter configuration model. The proxy interaction module includes a proxy module and an interaction module. The proxy module connects to the algorithm customization frontend and the algorithm parsing microservice platform, proxying the algorithm training and presenting the results. The interaction module includes an algorithm access module and a parameter configuration module. The algorithm access module provides real-time visualization of the algorithm training process. The parameter configuration module provides professionals with parameter couplers and adjusters. The parameter couplers implement custom topological constraints for hyperparameters, and the adjusters dynamically adjust the hyperparameter search space based on the prior knowledge of professionals.

[0020] Furthermore, the expert knowledge graph data includes three-dimensional modeling information and mathematical modeling information of the components; the expert knowledge graph, as prior data for the fault diagnosis algorithm, is combined with sensor data to form hybrid data, which is used as input to the diagnostic model to transmit the spatiotemporal information of the faulty components to the model.

[0021] Furthermore, the algorithm's customized front end uses a client / server architecture, with a web client and a desktop client as the client, and a user's small server or cloud server as the server, communicating with a server cluster via WebSocket.

[0022] Furthermore, the task node unit includes a data processing node, a model design node, a training configuration node, a testing and verification node, and a deployment and launch node. Each node is composed of its own task sub-nodes, and each task node and its sub-nodes form a deep tree data structure.

[0023] The process design unit consists of a data flow module and a model flow module. The data flow module is used to control the flow of fault data and feature maps, and the model flow module is used to control the flow of model deployment.

[0024] The result visualization unit is used to call the visualization service platform to visualize the corresponding results. The visualization results include the digital twin of the equipment, the dynamic signal diagram of the data, the network structure diagram of the model, and the change diagram of the test indicators.

[0025] Furthermore, the fault diagnosis algorithm kernel is obtained by reorganizing and secondary development of deep learning and data processing framework. The data processing, model design, training configuration, testing and verification, and deployment optimization sub-modules expose the internally integrated APIs. The extended components are used to provide user-defined options.

[0026] The present invention also provides a method for customizing fault diagnosis algorithms using the above-described system, comprising the following steps:

[0027] Step 1: Backend Configuration and Interface Rendering

[0028] The backend configuration steps include:

[0029] Register the APIs for data processing, model design, training configuration, testing and verification, and deployment optimization in the fault algorithm kernel to the algorithm microservice platform to complete the registration of the algorithm kernel microservice;

[0030] The interactive parameter optimization system integrates the search configuration module, learning optimization module, performance analysis module, and proxy interaction module with the algorithm customization front-end and algorithm parsing microservice platform.

[0031] The interface rendering steps include:

[0032] Retrieve algorithm design data from the algorithm database, including fault data, model initialization parameters, and configuration parameter data, perform model design initialization work, and render it to the algorithm customization front-end page;

[0033] The data retrieval module of the visualization service platform retrieves the model expert knowledge graph from the algorithm database, parses the 3D modeling information and data modeling information, and passes it to the visualization module for equipment digital twin pre-rendering to the front-end result visualization module.

[0034] Step Two: Front-end Algorithm Customization

[0035] Launch the algorithm customization front-end, log in to the developer account, and after logging in, use the task node units and process node units of the algorithm customization front-end to allow developers to perform visual programming and attribute configuration of fault diagnosis algorithms on the front-end page;

[0036] Based on a client / server architecture and WebSocket cluster communication, a collaborative design of multi-concurrency fault diagnosis algorithms was implemented; the task node unit adopts a deep tree structure with task nodes and sub-task nodes superimposed.

[0037] Based on the process design unit, different data processing flows are filtered, merged, and output; on the same node, different developers set different development priorities, and conflict alarm confirmation is set for the same priority, which is voted on by developers of the same priority or selected by developers of higher priority.

[0038] Developers use task node units and process design units to complete the design of data processing, model design, training configuration, test optimization and deployment, and generate process configuration form documents, which are sent to the algorithm parsing microservice platform and saved to the algorithm database in the fault algorithm knowledge base.

[0039] Step 3: Algorithm parsing and invocation

[0040] The algorithm process parsing system of the algorithm parsing microservice platform receives and parses the process configuration form document of the algorithm customization front end, parses it into individual sub-task nodes, and calls the data processing, model design, training configuration, testing and verification, and deployment optimization microservices registered in the algorithm microservice.

[0041] Step 4: Full Lifecycle Management

[0042] After the algorithm is deployed and launched, the algorithm parsing microservice platform provides monitoring microservices to the equipment to monitor the equipment's working status and algorithm performance in real time, which is then displayed by the visualization service platform. The algorithm development process documents are dynamically saved in the algorithm support library. Developers monitor the algorithm based on the data dashboards provided by the visualization service platform. If the algorithm's performance degrades or fails, they retrieve the process configuration forms and trained model parameters from the algorithm database, optimize and iterate the algorithm, and then deploy and launch it again.

[0043] Furthermore, the specific process of step three is as follows:

[0044] 1) Parse the data processing task nodes from the process configuration form document, call the data processing microservice series, retrieve fault data from the algorithm database of the fault algorithm knowledge base, complete the data processing work of data segmentation, data labeling, data augmentation and data fusion, and save the processing results;

[0045] 2) Parse the model architecture from the model design nodes, and complete the initialization of the model structure and parameters through the microservices of the model design, forming and saving the static computation graph of the model;

[0046] 3) Based on the training configuration node, the training configuration parameters are parsed out, the data processing results are input into the initialized model static graph, and the model is iteratively trained. Through the interactive parameter optimization system, the model is self-optimized and real-time performance is analyzed. After the training process is completed, the trained model parameters are saved to the algorithm database.

[0047] 4) Verify the microservice through testing. Retrieve test data from the algorithm database to test and verify the trained model using multiple performance metrics, ensuring the model's generalization and robustness, and save the test results to the algorithm database.

[0048] 5) Perform deployment optimization on the trained model according to the configuration of the deployment node; perform lossless compression operations on the model, including model pruning, quantization and distillation, convert the model into an ONNX standard description to calculate a static graph, perform operations such as environment deployment and inference acceleration through ONNX runtime, and then deploy the model online;

[0049] The algorithm parsing microservice platform timely feeds back the call results of each process to the customized front end, and the front end calls the visualization service platform for real-time visualization to display the data processing, model training, testing and verification processes.

[0050] With the above solution, the fault diagnosis algorithm customization system and method driven by low-code configuration capability reduce the knowledge reserve requirements of maintenance personnel for artificial intelligence, computer programming and the like, and reduce the code repetition rate; realize the full life cycle management of the design, development and deployment stages of the intelligent fault diagnosis algorithm, and realize high-performance iterative update of the algorithm; meet more flexible and reliable customization requirements of the intelligent fault diagnosis algorithm; ensure the characteristics of the intelligent fault diagnosis algorithm such as high precision, high reliability and fast implementation.

[0051] The above description is only an overview of the technical solutions of the present invention. In order to understand the technical means of the present invention more clearly and implement them according to the content of the specification, the preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings. Brief Description of the Drawings

[0052] Figure 1 is a schematic diagram of the interaction of each module of the fault diagnosis algorithm customization system driven by low-code configuration capability according to the present invention;

[0053] Figure 2 shows the algorithm customization front-end web page end in an embodiment of the present invention, including diagrams a, b and c. Detailed Description of the Embodiments

[0054] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. The following embodiments are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0055] Reference Figure 1 As shown, different from the traditional imperative text development of fault diagnosis algorithms, this embodiment provides a fault diagnosis algorithm customization platform system driven by low-code configuration capability. The platform includes a fault algorithm knowledge base, an algorithm customization front end, a fault algorithm kernel, an algorithm parsing microservice platform, a visualization service platform, and an interactive parameter optimization system. The interaction between each module is as shown in Figure 1 as shown in the figure.

[0056] (1) Fault algorithm knowledge base:

[0057] The fault algorithm knowledge base consists of an expert knowledge base and an algorithm database. The expert knowledge base uses a graph database as its basic database and stores expert knowledge graph data containing spatiotemporal modeling information of faulty components (or equipment). The expert knowledge graph data includes the component's three-dimensional modeling information (geometric information, constraint information, and environmental information) and mathematical modeling information (fitting data, mechanical analysis, and evolution mechanisms).

[0058] Expert knowledge graphs, serving as prior data for fault diagnosis algorithms, are combined with sensor data to form hybrid data that acts as input to the diagnostic model, thereby transmitting spatiotemporal information about the faulty component to the model. Furthermore, the visualization service platform promptly renders the expert knowledge graph of the selected faulty component, facilitating iterative operations such as data feedback, model reshaping, and deployment optimization for developers in the faulty component diagnosis algorithm.

[0059] The algorithm database consists of a relational database and a basic database, storing algorithm development data such as fault data, model parameters, training configurations, test results, and deployment information. Fault data includes historical and real-time data; model parameters include pre-trained and fine-tuned parameters; training configurations include information on data processing, loss functions, and optimization functions; test results include metrics such as accuracy, precision, recall, and F1 score; and deployment information includes model compression, model packaging, and device information.

[0060] (2) Algorithm Customization Front-end:

[0061] The algorithm customization front-end consists of a web-based and a desktop-based interface, both sharing a common algorithm visualization programming system. This system encapsulates each algorithm within the fault diagnosis algorithm kernel into a task node and represents it graphically. It includes: task node units, process design units, and result visualization units. A diagram of the web-based algorithm customization front-end is shown in Figure 2.

[0062] The task node unit consists of data processing nodes, model design nodes, training and configuration nodes, testing and verification nodes, and deployment and launch nodes. Each node contains its own task sub-nodes. The task nodes and their sub-nodes form a deep tree data structure.

[0063] The process design unit consists of a data flow module and a model flow module. The data flow module controls the flow of fault data and feature maps, while the model flow module controls the model deployment flow. Through the process design unit, higher code reuse can be achieved, diagnostic model parameters can be reduced, and data utilization can be improved.

[0064] The algorithm customization frontend uses a client / server architecture, with the customization frontend (web and desktop versions) as the client and a user's small server or cloud server as the server, communicating with a server cluster via WebSocket. Currently, the server starts a process for each collaboratively edited document, and each user in the collaboration is associated with this document. Whenever a user opens the algorithm design document, the client starts and downloads a copy of the file. From this point onward, any bidirectional document modifications are synchronized via the WebSocket connection. This allows developers to go offline (exit the collaboration state) at any time and continue editing. When a developer comes back online, the client downloads the latest copy of the file, and all changes made while the developer was offline are applied to this latest copy. Then, a WebSocket connection is established and synchronized.

[0065] The results visualization unit visualizes the results by calling a visualization service platform. The visualization results include: a digital twin of the equipment, dynamic signal graphs of the data, network structure diagrams of the model, and graphs showing changes in test indicators, etc.

[0066] (3) Fault algorithm kernel:

[0067] The fault diagnosis algorithm kernel consists of six sub-modules: data processing, model design, training configuration, testing and verification, deployment optimization, and extension components. The algorithm kernel is derived from the reorganization and secondary development of current deep learning and data processing frameworks. The first five sub-modules, including data processing, expose their internally integrated APIs, while the extension components provide user-defined options.

[0068] (4) Algorithm parsing microservice platform:

[0069] The algorithm parsing microservice platform consists of two parts: an algorithm process parsing system and an algorithm microservice platform. The algorithm process parsing system, based on a workflow engine, parses the diagnostic algorithm development configuration into a process sequence composed of organizational structure, workflow, nodes, and redirection rules. The fault diagnosis algorithm microservice platform registers the fault diagnosis algorithm development kernel to obtain a microservice instance, provides algorithm microservices to the algorithm parsing system, and retrieves relevant fault data, algorithm models, and parameter configurations from the fault algorithm database to complete the diagnostic algorithm development work.

[0070] (5) Visualization service platform:

[0071] The visualization service platform consists of a visualization module, a data retrieval module, and a custom module. The visualization module comprises a ready-to-use component library, capable of rendering various commonly used charts and models. The data retrieval module retrieves relevant data from the fault algorithm support library based on the needs of the visualization module. The custom module provides users with extended APIs for custom rendering components, allowing for customized rendering tasks by importing a small amount of code in a specified format.

[0072] (6) Interactive parameter optimization system:

[0073] The interactive parameter optimization system consists of a search configuration module, a learning optimization module, a performance analysis module, and a proxy interaction module. The search configuration module configures the search space and enumeration types for hyperparameters involved in algorithm development. The learning optimization module is a self-optimizing algorithm module that embeds optimizers such as grid search, random search, Bayesian optimization, and reinforcement learning. The performance analysis module performs prior analysis on the performance of the model after a single parameter configuration, saving parameter optimization time. The proxy interaction module is divided into a proxy module and an interaction module. The proxy module connects to the algorithm customization frontend and the algorithm parsing microservice platform, acting as a proxy to complete algorithm training and present the results. The interaction module is divided into an algorithm access module and a parameter configuration module. The algorithm access module allows real-time visualization of the algorithm training process, such as gradient optimization switches, early-stop switches, and training start / stop switches; the parameter configuration module provides professionals with parameter couplers and adjusters. The couplers can implement custom topological constraints for hyperparameters, and the adjusters can dynamically adjust the hyperparameter search space based on the professional's prior knowledge.

[0074] The process for customizing fault diagnosis algorithms using this system is as follows:

[0075] Step 1: Backend configuration and interface rendering.

[0076] The backend configuration first involves registering the APIs for data processing, model design, training configuration, testing and verification, and deployment optimization within the fault algorithm kernel to the algorithm microservice platform, thus completing the registration of the algorithm kernel microservice. Then, it integrates the search configuration module, learning optimization module, performance analysis module, and proxy interaction module within the interactive parameter optimization system with the algorithm customization frontend and algorithm parsing microservice platform. Interface rendering includes retrieving fault data, model initialization parameters, configuration parameters, and other algorithm design data from the algorithm database, performing model design initialization, and rendering it to the algorithm customization frontend page. Secondly, through the data retrieval module of the visualization service platform, it retrieves the model expert knowledge graph from the algorithm database, parses the 3D modeling information and data modeling information, and passes it to the visualization module for equipment digital twin pre-rendering to the frontend result visualization module.

[0077] Step 2: Customize the front-end algorithm.

[0078] First, launch the algorithm customization frontend (web or desktop) and log in to your developer account. After logging in, developers can perform visual programming and attribute configuration of fault diagnosis algorithms on the frontend page through task node units and process node units. Based on a client / server architecture and WebSocket cluster communication, collaborative design of multi-concurrency fault diagnosis algorithms is realized. Furthermore, the task node unit adopts a deep tree structure with superimposed task nodes and sub-task nodes. Therefore, multiple developers can distribute different sub-tasks on a single task node. For example, the data processing part can be divided into sub-tasks such as data cleaning, data integration, data reduction, and data transformation, which can be developed by several people. Finally, based on the process design unit, different data processing flows are filtered, merged, and output. On the same node, different developers set different development priorities. Conflict alarm confirmation is set for developers with the same priority, and the decision is made by voting among developers with the same priority or by developers with higher priority. Finally, developers use the task node unit and process design unit to complete the design of data processing, model design, training configuration, test optimization, and deployment, and generate a process configuration form document, which is sent to the algorithm parsing microservice platform and saved to the algorithm database in the fault algorithm knowledge base.

[0079] Step 3: Algorithm parsing and invocation.

[0080] The algorithm parsing microservice platform's algorithm process parsing system receives and parses the process configuration form document from the algorithm customization front-end, parses it into individual sub-task nodes, and calls the microservices registered in the algorithm microservice, such as data processing, model design, training configuration, testing and verification, and deployment optimization. The specific process is as follows: First, parse the data processing task nodes from the process configuration form document, call the data processing microservices, retrieve fault data from the algorithm database of the fault algorithm knowledge base (hereinafter referred to as the algorithm database), complete data processing tasks such as data segmentation, data labeling, data augmentation, and data fusion, and save the processing results; Second, parse the model architecture from the model design node, complete the initialization of the model structure and parameters through the model design microservice, form the model static computation graph and save it; Third, based on the training configuration node, parse the training configuration parameters, input the data processing results into the initialized model static graph, perform iterative training of the model, perform parameter self-optimization and real-time performance analysis of the model through the interactive parameter optimization system, and save the trained model parameters to the algorithm database after the training process is completed; Fourth, through the test and verification microservice, retrieve test data from the algorithm database, test and verify the trained model with multiple performance indicators to ensure the generalization and robustness of the model, and save the test results to the algorithm database; Fifth, deploy and optimize the trained model according to the configuration of the deployment node. The model undergoes lossless compression operations such as model pruning, quantization, and distillation, and is then converted into an ONNX standard descriptive computation static graph. After environment deployment and inference acceleration via ONNX runtime, the model is deployed online. Furthermore, developers can customize various parts of the above process based on extended component microservices, such as configuring separate data processing modules. The results of each process call are promptly fed back to the customized frontend by the algorithm parsing microservice platform. The frontend then calls the visualization service platform for real-time visualization, displaying the data processing, model training, and testing verification processes.

[0081] Step 4: Full lifecycle management.

[0082] After the algorithm is deployed and launched, the algorithm parsing microservice platform provides monitoring microservices to the equipment, monitoring the equipment's operational status and algorithm performance in real time, which is then displayed by the visualization service platform. Algorithm development process documents are dynamically stored in the algorithm support library. Developers can monitor the algorithm using the data dashboards provided by the visualization service platform. If a performance degradation or failure is detected, they can retrieve the process configuration forms and trained model parameters from the algorithm database to optimize and iterate the algorithm before re-deploying it.

[0083] This invention reduces the knowledge requirements for maintenance personnel in areas such as artificial intelligence and computer programming, and decreases code duplication; it enables full lifecycle management of the design, development, and deployment stages of intelligent fault diagnosis algorithms, achieving high-performance iterative updates; it fulfills the requirements for more flexible and reliable customized intelligent fault diagnosis algorithms; it ensures the high precision, high reliability, and rapid deployment characteristics of intelligent fault diagnosis algorithms, and has the following features:

[0084] 1. This invention provides a method for customizing algorithms across all scenarios and throughout their entire lifecycle.

[0085] Because faults are often sporadic and gradual, and their occurrence is closely related to the equipment's operating environment and timeframe, fault diagnosis requirements must change over time and space. This necessitates customizing fault diagnosis algorithms from a full-scenario, full-lifecycle perspective. This invention, based on existing deep learning frameworks and incorporating expert knowledge of scenario modeling and fault evolution, customizes fault diagnosis algorithms across all scenarios and lifecycles, including data feedback, model reshaping, and deployment optimization.

[0086] 2. This invention is a low-code driven, high-efficiency algorithm development platform.

[0087] In real-world industrial scenarios, multiple pieces of equipment often work together, cooperating to complete product production. This means that equipment failures are often accompanied by a complex failure of multiple pieces of equipment. Local failures are influenced and constrained by other equipment or components, and the complex failure is formed by the superposition and coupling of local failures. This invention utilizes a low-code development platform, based on the aforementioned algorithm customization method covering the entire scenario and lifecycle, and leverages algorithm model-driven, graphical programming and parameter configuration for efficient development of diagnostic algorithms. It possesses both fine-grained local fault diagnosis capabilities and global comprehensive fault diagnosis capabilities.

[0088] 3. This invention is a full-process visualization service system for diagnostic algorithms.

[0089] The diagnostic algorithm visualization system proposed in this invention includes modules such as dynamic display of real-time data, rendering of low-code development process, and self-drawing of training and testing indicators. It provides real-time visualization of each stage of the model development process, further improving the user-friendliness of diagnostic algorithm development for maintenance personnel.

[0090] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A fault diagnosis algorithm customization system driven by low-code configuration capabilities, characterized in that, It includes a fault algorithm knowledge base, an algorithm customization front-end, a fault algorithm kernel, an algorithm parsing microservice platform, a visualization service platform, and an interactive parameter optimization system; The fault algorithm knowledge base communicates with the visualization service platform, the algorithm customization front-end, and the algorithm parsing microservice platform. The visualization service platform, the interactive parameter optimization system, and the algorithm parsing microservice platform communicate with the algorithm customization front-end. The fault algorithm kernel communicates with the algorithm parsing microservice platform. The fault algorithm knowledge base includes an expert knowledge base and an algorithm database. The expert knowledge base stores expert knowledge graph data containing spatiotemporal modeling information of faulty components. The algorithm database stores algorithm development data, including fault data, model parameters, training configurations, test results, and deployment information. The algorithm customization front end includes a web client and a desktop client, both of which share a set of algorithm visualization programming system, which is used to encapsulate each algorithm in the fault diagnosis algorithm kernel into a task node and represent it with a visual graphic, including task node unit, process design unit and result visualization unit; The fault diagnosis algorithm kernel includes six sub-modules: data processing, model design, training configuration, testing and verification, deployment optimization, and extension components. The algorithm parsing microservice platform includes an algorithm process parsing system and an algorithm microservice platform. The algorithm process parsing system is used to parse the diagnostic algorithm development configuration with a workflow engine as the main body, and parse it into a process sequence composed of organizational structure, process, node, and redirection rules. The fault algorithm microservice platform is used to register the fault diagnosis algorithm development kernel to obtain microservice instances, provide algorithm microservices to the algorithm process parsing system, and retrieve corresponding fault data, algorithm models and parameter configuration data from the fault algorithm database to complete the diagnostic algorithm development work. The visualization service platform includes a visualization module, a data retrieval module, and a custom module. The visualization module consists of an out-of-the-box component library, used to render various commonly used charts and models. The data retrieval module is used to retrieve relevant data from the fault algorithm support library according to the needs of the visualization module. The custom module provides users with extended APIs for custom rendering components, allowing them to complete custom rendering by importing a small amount of code in a specified format. The interactive parameter optimization system includes a search configuration module, a learning optimization module, a performance analysis module, and a proxy interaction module. The search configuration module configures the search space and enumeration type for hyperparameters involved in algorithm development. The learning optimization module performs algorithm self-optimization through embedded optimizers, including grid search, random search, Bayesian optimization, and reinforcement learning optimizers. The performance analysis module performs prior analysis on the performance of a single parameter configuration model. The proxy interaction module includes a proxy module and an interaction module. The proxy module connects to the algorithm customization frontend and the algorithm parsing microservice platform, proxying the algorithm training and presenting the results. The interaction module includes an algorithm access module and a parameter configuration module. The algorithm access module provides real-time visualization of the algorithm training process. The parameter configuration module provides professionals with parameter couplers and adjusters. The parameter couplers implement custom topological constraints for hyperparameters, and the adjusters dynamically adjust the hyperparameter search space based on the prior knowledge of professionals.

2. The fault diagnosis algorithm customization system based on low-code configuration capability as described in claim 1, characterized in that, The expert knowledge graph data includes 3D modeling information and mathematical modeling information of the components; the expert knowledge graph, as prior data for the fault diagnosis algorithm, is combined with sensor data to form hybrid data, which is used as input to the diagnostic model to transmit the spatiotemporal information of the faulty components to the model.

3. The fault diagnosis algorithm customization system based on low-code configuration capability as described in claim 2, characterized in that, The algorithm's customized front end uses a client / server architecture, with a web client and desktop client as the client, and a user's small server or cloud server as the server, communicating with a server cluster via WebSocket.

4. The fault diagnosis algorithm customization system based on low-code configuration capability as described in claim 3, characterized in that, The task node unit includes data processing nodes, model design nodes, training configuration nodes, testing and verification nodes, and deployment and online nodes. Each node is composed of its own task sub-nodes, and each task node and its sub-nodes form a deep tree data structure. The process design unit consists of a data flow module and a model flow module. The data flow module is used to control the flow of fault data and feature maps, and the model flow module is used to control the flow of model deployment. The result visualization unit is used to call the visualization service platform to visualize the corresponding results. The visualization results include the digital twin of the equipment, the dynamic signal diagram of the data, the network structure diagram of the model, and the change diagram of the test indicators.

5. The fault diagnosis algorithm customization system based on low-code configuration capability as described in claim 4, characterized in that, The fault diagnosis algorithm kernel is obtained by reorganizing and secondary development of deep learning and data processing framework. The data processing, model design, training configuration, testing and verification, and deployment optimization sub-modules expose the internally integrated APIs. The extension components are used to provide user-defined options.

6. A method for customizing a fault diagnosis algorithm using the system described in any one of claims 1 to 5, characterized in that, Includes the following steps: Step 1: Backend Configuration and Interface Rendering The backend configuration steps include: Register the APIs for data processing, model design, training configuration, testing and verification, and deployment optimization in the fault algorithm kernel to the algorithm microservice platform to complete the registration of the algorithm kernel microservice; The interactive parameter optimization system integrates the search configuration module, learning optimization module, performance analysis module, and proxy interaction module with the algorithm customization front-end and algorithm parsing microservice platform. The interface rendering steps include: Retrieve algorithm design data from the algorithm database, including fault data, model initialization parameters, and configuration parameter data, perform model design initialization work, and render it to the algorithm customization front-end page; The data retrieval module of the visualization service platform retrieves the model expert knowledge graph from the algorithm database, parses the 3D modeling information and data modeling information, and passes it to the visualization module for equipment digital twin pre-rendering to the front-end result visualization module. Step Two: Front-end Algorithm Customization Launch the algorithm customization front-end, log in to the developer account, and after logging in, use the task node units and process node units of the algorithm customization front-end to allow developers to perform visual programming and attribute configuration of fault diagnosis algorithms on the front-end page; Based on a client / server architecture and WebSocket cluster communication, a collaborative design of multi-concurrency fault diagnosis algorithms was implemented; the task node unit adopts a deep tree structure with task nodes and sub-task nodes superimposed. Based on the process design unit, different data processing flows are filtered, merged, and output; on the same node, different developers set different development priorities, and conflict alarm confirmation is set for the same priority, which is voted on by developers of the same priority or selected by developers of higher priority. Developers use task node units and process design units to complete the design of data processing, model design, training configuration, test optimization and deployment, and generate process configuration form documents, which are sent to the algorithm parsing microservice platform and saved to the algorithm database in the fault algorithm knowledge base. Step 3: Algorithm parsing and invocation The algorithm process parsing system of the algorithm parsing microservice platform receives and parses the process configuration form document of the algorithm customization front end, parses it into individual sub-task nodes, and calls the data processing, model design, training configuration, testing and verification, and deployment optimization microservices registered in the algorithm microservice. Step 4: Full Lifecycle Management After the algorithm is deployed and launched, the algorithm parsing microservice platform provides monitoring microservices to the equipment to monitor the equipment's working status and algorithm performance in real time, which is then displayed by the visualization service platform. The algorithm development process documents are dynamically saved in the algorithm support library. Developers monitor the algorithm based on the data dashboards provided by the visualization service platform. If the algorithm's performance degrades or fails, they retrieve the process configuration forms and trained model parameters from the algorithm database, optimize and iterate the algorithm, and then deploy and launch it again.

7. The fault diagnosis algorithm customization method according to claim 6, characterized in that, The specific process for step three is as follows: 1) Parse the data processing task nodes from the process configuration form document, call the data processing microservice series, retrieve fault data from the algorithm database of the fault algorithm knowledge base, complete the data processing work of data segmentation, data labeling, data augmentation and data fusion, and save the processing results; 2) Parse the model architecture from the model design nodes, and complete the initialization of the model structure and parameters through the microservices of the model design, forming and saving the static computation graph of the model; 3) Based on the training configuration node, the training configuration parameters are parsed out, the data processing results are input into the initialized model static graph, and the model is iteratively trained. Through the interactive parameter optimization system, the model is self-optimized and real-time performance is analyzed. After the training process is completed, the trained model parameters are saved to the algorithm database. 4) Verify the microservice through testing. Retrieve test data from the algorithm database to test and verify the trained model using multiple performance metrics, ensuring the model's generalization and robustness, and save the test results to the algorithm database. 5) Optimize the deployment of the trained model according to the configuration of the deployment nodes; perform lossless compression operations on the model, including model pruning, quantization and distillation, and convert the model into a static graph describing the ONNX standard. After environment deployment and inference acceleration operations through the ONNX runtime, the model is deployed online. The algorithm analysis microservice platform promptly feeds back the call results of each process to the customized front end. The front end then calls the visualization service platform for real-time visualization, showcasing the data processing, model training, and testing verification processes.

Images