Practical training platform, evaluation method and equipment for aluminum electrolysis process based on multi-modal data

Abstract

The invention provides an aluminum electrolysis process practical training platform based on multi-modal data, an evaluation method and equipment. The evaluation method comprises the following steps: acquiring log record data and video data in an aluminum electrolysis process practical training process, and acquiring sensor data of industrial equipment in an aluminum electrolysis workshop; an LSTM network is used for processing sequence data, the long-term dependency relationship in log record data can be captured, and the sequence and logic of aluminum electrolysis process operation can be understood; processing the video data by using a CNN network, and extracting key features in the video; processing sensor data using a full connection layer; and a full connection layer is used as a fusion layer, and outputs of the networks are fused. According to the method, different deep learning networks are adopted to process different data, the extracted features are more representative, and the reliability of a result obtained through fusion is higher; the actual operation effect of the aluminum electrolysis process can be reflected and evaluated more accurately, more valuable feedback is provided, and the aluminum electrolysis process is controlled and optimized.

Description

Technical Field

[0001] The present invention belongs to the technical field of aluminum electrolysis, and in particular relates to an aluminum electrolysis process training platform, an evaluation method and a device based on multimodal data. Background Art

[0002] As the core link of aluminum industry production, the efficiency and safety of aluminum electrolysis process directly affect the quality and production cost of aluminum products. The traditional aluminum electrolysis process is highly dependent on manual operation and experience judgment, which not only limits production efficiency but also increases safety risks. Although automation technology has developed in recent years, most existing systems still face challenges such as strong operation dependence, insufficient data utilization, many safety hazards, obvious efficiency bottlenecks and high maintenance costs.

[0003] Although attempts have been made to introduce intelligent technologies, such as using sensors to monitor the status of electrolytic cells and adopting simple automated control systems, most of these technologies are still in their infancy. Promoting the development of intelligent production decision-making in aluminum electrolysis is inseparable from a wide range of reliable data sources. Although the existing parameter detection technology in the aluminum electrolysis industry has made great progress and a large amount of data can be collected during the production process, data utilization is insufficient and the ability to integrate multi-source data is insufficient.

[0004] Patent CN111651890B discloses a data-driven aluminum electrolysis digital twin factory control method, which uses CNN neural network, t-SNE fusion tools, etc. to build a fusion decision model, obtains industrial operation decisions based on the fusion decision model and multi-source data, and controls the aluminum electrolysis workshop according to the industrial operation decisions. This method uses CNN neural network to process sequence data and video data, and the data processing is less targeted, the multi-source data integration capability is insufficient, and the reliability is low. Summary of the invention

[0005] The purpose of the present invention is to address the deficiencies in the prior art and provide an aluminum electrolysis process training platform, evaluation method and equipment based on multimodal data to improve the reliability and accuracy of aluminum electrolysis process evaluation.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is:

[0007] A method for evaluating an aluminum electrolysis process based on multimodal data comprises the following steps:

[0008] S1: Obtain log data and video data during the aluminum electrolysis process training, and obtain sensor data of industrial equipment in the aluminum electrolysis workshop;

[0009] S2: Preprocessing the log data, video data, and sensor data;

[0010] S3: using the preprocessed log data, video data, and sensor data as inputs of a deep learning network model, and training the deep learning network model, wherein the deep learning network model includes a CNN network, an LSTM network, a first fully connected layer, and a second fully connected layer, and the CNN network, the LSTM network, and the first fully connected layer are respectively connected to the second fully connected layer;

[0011] The training process of the deep learning network model includes:

[0012] The LSTM network is used to process the preprocessed log record data to obtain a first feature;

[0013] Using the CNN network to process the preprocessed video data to obtain a second feature;

[0014] Processing the preprocessed sensor data using the first fully connected layer to obtain a third feature;

[0015] fusing the first feature, the second feature, and the third feature through the second fully connected layer to obtain an evaluation result;

[0016] The hyperparameter combination is optimized through automatic parameter adjustment and the deep learning network model is iteratively optimized and trained to obtain an aluminum electrolysis process evaluation model.

[0017] The present invention uses a long short-term memory network (LSTM) to process sequence data, which can capture long-term dependencies in log data and understand the order and logic of aluminum electrolysis process operations; uses a convolutional neural network (CNN) to process video data, which can extract key features in the video; uses a fully connected layer (ANN) to process sensor data; and uses a fully connected layer as a fusion layer to fuse the outputs of the above networks. The present invention uses different deep learning networks to process different data according to the characteristics of different data, and can process complex multimodal data, achieve stronger feature learning capabilities, and the extracted features are more representative, and the reliability and accuracy of aluminum electrolysis process evaluation are higher. The combination of LSTM, CNN, and ANN enables the model to cope with a variety of different types of data (such as video, time series data, classification data, etc.), and improves the generalization ability of the model in different tasks.

[0018] Furthermore, the log record data includes operation logs and decision records of operators, technicians and administrators during the aluminum electrolysis process training; the video data includes operation videos of operators, technicians and administrators during the aluminum electrolysis process training; the sensor data includes the current, voltage, temperature, number of material unloading, material unloading amount, aluminum tapping time, and pole changing time of the aluminum electrolysis cell.

[0019] Through the operation data of personnel at different levels, the actual operation effect of the aluminum electrolysis process can be more accurately reflected and evaluated, more valuable feedback can be provided, and the accuracy and precision of the model can be improved.

[0020] Furthermore, the specific process of S2 includes:

[0021] Cleaning and formatting the log record data;

[0022] Performing size standardization and normalization processing on the video data;

[0023] The sensor data is smoothed and normalized.

[0024] By preprocessing the log data, invalid and redundant information is removed and then formatted into a sequence format suitable for input into the LSTM network; by preprocessing the video data, the efficiency and accuracy of subsequent feature extraction are improved; by preprocessing the sensor data, noise is reduced and important features are highlighted to facilitate subsequent feature extraction.

[0025] Furthermore, the LSTM network includes an input gate, a forget gate, an output gate, and a candidate cell state, and the expressions of the input gate, the forget gate, the output gate, and the candidate cell state are as follows:

[0026] F t =σ(W f ·[H t-1 ,X t ]+b f );

[0027] I t =σ(W i ·[H t-1 ,X t ]+b i );

[0028]

[0029] O t =σ(W o ·[H t-1 ,X t ]+b o );

[0030] H t =O t *tanh(C t );

[0031] Among them, σ is the sigmoid activation function, tanh is the hyperbolic tangent function, and F t is the forget gate at time t, I tis the input gate at time t, C t is the cell state at time t, is the candidate cell state at time t, O t is the output gate at time t, H t is the hidden state at time t, X t is the log record data at time t, H t-1 is the hidden state at time t-1, C t-1 is the cell state at time t-1, W f , W i , W o , W c are the weights of the forget gate, input gate, output gate and candidate cell state, respectively, and b f , b i , b o , b c They are the forget gate, input gate, output gate and bias of the candidate cell state respectively, and * indicates element-by-element multiplication.

[0032] Furthermore, the CNN network includes multiple cascaded convolutional networks, and the convolutional network includes a convolutional layer, a pooling layer, and a fully connected layer. The expressions of the convolutional layer, the pooling layer, and the fully connected layer are as follows:

[0033]

[0034] x l =σ(W l *Y2+b l ),l=1,2,…,L-1,

[0035] Among them, x0 is the video data, x is the input of the lth convolutional layer in the CNN network, is the output of the lth convolutional layer in the CNN network, K is the convolution kernel, k h is the height of the convolution kernel, k w is the width of the convolution kernel, C is the number of channels of the convolution kernel, i and j are the positions of the output feature maps, is the output of the lth pooling layer in the CNN network, x l is the output of the lth fully connected layer in the CNN network, σ is the sigmoid activation function, and b l is the bias of the lth fully connected layer in the CNN network, W l is the weight of the lth fully connected layer in the CNN network, is the size of the kth dimension of the video data, k = 1, 2, ..., d; d is the dimension of the video data, and L is the number of layers of the CNN network.

[0036] Furthermore, the expression of the first fully connected layer or the second fully connected layer is as follows:

[0037] y (g) =f(W (g) x (g-1) +b (g) )

[0038] Among them, y (g) is the output of the g-th layer, W (g) is the weight of the g-th layer, b (g) is the weight bias of the g-th layer, x (g-1) is the input of the g-th layer, and f is the activation function.

[0039] Furthermore, the log record data, video data, and sensor data are input into an aluminum electrolysis process evaluation model to obtain evaluation results, and the aluminum electrolysis process is controlled and optimized based on the evaluation results.

[0040] Based on the same inventive concept, the present invention also provides an aluminum electrolysis process training platform based on the above method, including a real-time monitoring unit, an aluminum electrolysis process training unit, a data storage unit, a model training optimization unit, and an evaluation feedback unit;

[0041] Real-time monitoring unit, used to monitor and collect sensor data of industrial equipment in aluminum electrolysis workshop in real time;

[0042] The aluminum electrolysis process training unit builds a one-to-one digital twin model based on the aluminum electrolysis cell workshop and constructs an aluminum electrolysis process training scenario for operators, technicians, and administrators to conduct aluminum electrolysis process training;

[0043] A data storage unit is used to integrate and store sensor data collected by the real-time monitoring unit, operation logs and decision records during aluminum electrolysis process training, and video data according to time;

[0044] A model training optimization unit is used to train an aluminum electrolysis process evaluation model using sensor data, operation logs and decision records, and video data as input data;

[0045] The evaluation feedback unit is used to control and optimize the aluminum electrolysis process according to the evaluation results output by the aluminum electrolysis process evaluation model.

[0046] Based on the same inventive concept, the present invention also provides an electronic device, including:

[0047] one or more processors;

[0048] A memory having one or more programs stored thereon, which, when executed by the one or more processors, enables the one or more processors to implement the steps of an aluminum electrolysis process evaluation method based on multimodal data.

[0049] Based on the same inventive concept, the present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of an aluminum electrolysis process evaluation method based on multimodal data.

[0050] Compared with the prior art, the present invention has the following beneficial effects:

[0051] The present invention uses a long short-term memory network (LSTM) to process sequence data, which can capture long-term dependencies in log data and understand the order and logic of aluminum electrolysis process operations; uses a convolutional neural network (CNN) to process video data, which can extract key features in the video; uses a fully connected layer (ANN) to process sensor data; and uses a fully connected layer as a fusion layer to fuse the outputs of the above networks. The present invention uses different deep learning networks to process different data according to the characteristics of different data, and can process complex multimodal data, achieve stronger feature learning capabilities, and extract more representative features. The reliability and accuracy of aluminum electrolysis process evaluation are higher, and the optimization control of aluminum electrolysis production processes is achieved. The combination of LSTM, CNN, and ANN enables the model to cope with a variety of different types of data (such as video, time series data, classification data, etc.), and improves the generalization ability of the model in different tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 It is a schematic diagram of the platform function planning of the aluminum electrolysis process training platform of the present invention;

[0053] Figure 2 A schematic diagram of the training scenario planning of the aluminum electrolysis process training platform of the present invention;

[0054] Figure 3 It is a framework diagram of the aluminum electrolysis process evaluation model of the present invention;

[0055] Figure 4 It is a schematic diagram of the effect of the aluminum electrolysis process training scenario of the present invention. DETAILED DESCRIPTION

[0056] The present invention will be described in detail below in conjunction with the embodiments. It should be noted that the embodiments and features of the embodiments of the present invention can be combined with each other without conflict. For the convenience of description, if the words "upper", "lower", "left" and "right" appear in the following, they only indicate that the upper, lower, left and right directions are consistent with the drawings themselves, and do not limit the structure.

[0057] Example 1

[0058] This embodiment proposes an aluminum electrolysis process training platform, which aims to achieve intelligent management and optimization of the aluminum electrolysis process by integrating advanced information technology and industrial automation technology.

[0059] The aluminum electrolysis process training platform includes real-time monitoring unit, aluminum electrolysis process training unit, data storage unit, model training optimization unit, evaluation feedback unit, three-dimensional visualization unit, BI customization unit, etc.

[0060] Real-time monitoring unit, used to monitor and collect sensor data of industrial equipment in aluminum electrolysis workshop in real time;

[0061] The aluminum electrolysis process training unit builds a one-to-one digital twin model based on the aluminum electrolysis cell workshop and constructs an aluminum electrolysis process training scenario for operators, technicians, and administrators to conduct aluminum electrolysis process training;

[0062] A data storage unit is used to integrate and store sensor data collected by the real-time monitoring unit, operation logs and decision records during aluminum electrolysis process training, and video data according to time;

[0063] A model training optimization unit is used to train an aluminum electrolysis process evaluation model using sensor data, operation logs and decision records, and video data as input data;

[0064] An evaluation feedback unit, used to control and optimize the aluminum electrolysis process according to the evaluation results output by the aluminum electrolysis process evaluation model;

[0065] The 3D visualization unit, based on the digital twin model, uses 3D visualization and other technical means to synchronously reflect the real-time status of the aluminum electrolysis workshop, providing a visualization interface for practical training for operators, technicians and administrators;

[0066] BI custom unit provides report design tools and rich visualization components, supporting custom report layout, chart types, data dimensions and key indicators.

[0067] The platform supports multiple communication protocols, such as Modbus, OPC UA and MQTT. These protocols enable seamless connection between industrial equipment in aluminum electrolysis workshops and various units of the platform, realizing real-time data collection and transmission.

[0068] like Figure 1 , Figure 2 , Figure 4 The core functional planning of the platform includes:

[0069] 1. Distributed architecture capabilities:

[0070] The aluminum electrolysis process training platform adopts an advanced distributed architecture, and realizes modularization and automated resource management of applications through microservice architecture and elastic scaling technology, thereby improving the flexibility and stability of the system. It supports public cloud and private cloud deployment, and combines native, Docker containers and Kubernetes (K8S) technology to provide diverse deployment options and efficient container management. In addition, the platform's SaaS service model further lowers the threshold for use, simplifies the maintenance process, and ensures the instant update and scalability of the software, thereby providing efficient, flexible and scalable services to meet production and management needs in different scenarios.

[0071] 2. Dynamic definition capability of digital twins:

[0072] The dynamic definition capability of the digital twin of the aluminum electrolysis process training platform provides a highly flexible and customizable virtual simulation environment. The core of this capability lies in the breadth and depth of its attribute definition, which covers multiple types such as strings, numbers, dates, dictionaries, Boolean values, relations, dynamic data, arrays, and time series data, ensuring a comprehensive description of the aluminum electrolysis process. Through multi-domain definitions, the platform can cover the entire product life cycle from design to maintenance, so that key attributes at each stage can be accurately managed and monitored.

[0073] The platform's attribute rendering function can automatically generate attribute forms based on attribute and domain definitions, and dynamically refresh time series data to ensure the real-time and accuracy of data. In addition, the platform also provides a wealth of industry templates, precipitating standards from different industries, and can quickly select or customize templates suitable for specific application scenarios. The integration of encoders ensures data compatibility and interoperability, thereby achieving unified management of data from different sources and formats.

[0074] The platform complies with ETSI (Telecommunications Standards Institute) standards and further enhances the platform's standardized description and integrated management capabilities through the Asset Management Shell (AAS) function, so that the attributes, functions and behaviors of industrial assets can be described and integrated in a standardized manner. These comprehensive features make the aluminum electrolysis process training platform a powerful tool that can meet the diverse needs of aluminum electrolysis production and management and improve the intelligence and automation level of industrial production.

[0075] 3. Multiple data source integration capabilities:

[0076] The ability to integrate multiple data sources of the aluminum electrolysis process training platform is a key component of its technical architecture, ensuring that the platform can be flexibly integrated with various devices and systems to achieve efficient data management and analysis. The platform supports multiple communication protocols, such as Modbus, OPC UA, and MQTT, which enable different types of devices and systems to seamlessly connect to the platform to achieve real-time data collection and transmission. In addition, the platform has the ability to convert data formats, which can convert the data collected by sensors into a format that matches the production control system, thereby facilitating data integration and analysis.

[0077] To ensure the accuracy and consistency of data, the platform provides a unified interface management module. This module manages the interfaces of all systems and devices connected to the platform, simplifying the management and maintenance of data. At the same time, the platform adopts security measures such as encrypted communication, identity authentication and permission management to ensure the security and privacy of data.

[0078] In terms of openness and flexibility, the platform design focuses on adapting to the ever-increasing number of systems and devices, and can quickly adapt to different industrial scenarios and needs. This design enables the platform to not only support existing industrial software, such as production control systems, process simulation software, data acquisition software, and data analysis software (such as Matlab / Simulink, Ansys, etc.), but also easily integrate new systems and equipment.

[0079] In particular, the platform has the ability to access various cell control systems, such as the Alwhit cell control system. This capability enables the platform to achieve precise control and monitoring of aluminum electrolytic cells, further improving the automation and intelligence level of the aluminum electrolysis process. Through these comprehensive features, the aluminum electrolysis process training platform provides a powerful, flexible and secure data integration and management solution.

[0080] 4. Dynamic expansion capability of model algorithm:

[0081] The dynamic expansion capability of the model algorithm of the aluminum electrolysis process training platform is another highlight of its technical architecture, providing full-process support from algorithm development to deployment. The platform's algorithm library contains a variety of commonly used algorithm modules, including both general libraries and libraries specifically for the aluminum electrolysis industry. It supports the integration of self-developed and third-party algorithm libraries, and covers multiple programming languages ​​such as Python, C / C++, .NET, and Go, ensuring the flexibility and diversity of algorithm development.

[0082] The platform provides a visual algorithm arrangement interface, which allows different algorithm modules to be combined by dragging and connecting to form a complete data processing flow. The input and output parameters, execution order and logic of the algorithm module can be customized to achieve a variety of data processing tasks. The algorithm execution engine is responsible for executing the entire process according to the algorithm combination, using lightweight container-based virtualization technology, such as Docker containerization, and supporting K8S deployment to improve the efficiency and scalability of algorithm execution.

[0083] Algorithm performance monitoring and optimization is another key feature of the platform. Through performance analysis tools, the platform can monitor key performance indicators during algorithm execution, such as running time and memory usage, so as to evaluate and optimize the algorithm and improve execution efficiency.

[0084] In addition, the platform's algorithm model management function supports version management, updating and release of algorithm models. It usually takes the form of model warehouse and model services, organically combining model management with algorithm orchestration functions to provide one-stop algorithm modeling and management services.

[0085] In terms of dynamic deployment of model algorithms, the platform adopts a hierarchical deployment strategy, deploying model algorithms at four levels: group, factory, workshop, and slot, with each level equipped with corresponding computing resources and storage space. The group level deploys big data platforms and machine learning algorithms to conduct summary analysis of factory data across the country and establish prediction models; the factory and workshop levels deploy data acquisition and production control systems, as well as real-time or historical machine learning algorithms to monitor and optimize the production process; the slot level deploys intelligent sensors and control equipment, as well as real-time data processing algorithms to achieve precise adjustment and optimization of the production process.

[0086] Finally, the platform's model algorithm virtual operation capability allows the evaluation of the effect and performance of the algorithm model without relying on real data. Through virtual data sets generated by technologies such as generative adversarial networks, the platform supports virtual operation control, analysis and evaluation, simulates the operation process of real production scenarios, and feeds back virtual production status and effects through digital twin models, testing and optimizing algorithms in a safe environment, thereby continuously improving the production process.

[0087] In summary, the dynamic expansion capability of the model algorithm of the aluminum electrolysis process training platform provides a comprehensive solution from algorithm development, orchestration, execution, monitoring, optimization to deployment, supports the rapid iteration and precise control of the algorithm, and lays a solid technical foundation for the intelligent and automated production of aluminum electrolysis.

[0088] 5. Online model training and optimization capabilities:

[0089] The online model training optimization capability of the aluminum electrolysis process training platform provides users with a comprehensive set of model development tools, aiming to achieve efficient and accurate model training through refined data processing and algorithm optimization. The platform's data management capabilities are extremely critical. It supports the import, cleaning, processing, storage and management of data, allowing users to quickly access and prepare the required data sets. On this basis, data preprocessing and feature engineering can be performed, which are critical to improving model performance.

[0090] Furthermore, the algorithm selection capability provides a wide range of algorithm libraries, including machine learning, deep learning, statistical learning, etc., to adapt to different scenarios and needs. You can choose the most appropriate algorithm framework according to specific problems, providing a solid algorithm foundation for model training.

[0091] Hyperparameter tuning is another major feature of the platform. It uses advanced technologies such as automatic parameter tuning and grid search to find the optimal hyperparameter combination, thereby significantly improving the performance and prediction accuracy of the model. This process automates the traditionally time-consuming manual parameter tuning work, making model training more efficient.

[0092] Model training capabilities further improve the efficiency and speed of model training by adopting modern technologies such as batch processing and GPU parallel computing. The application of these technologies makes training even for large-scale data sets feasible while ensuring the high efficiency of the model training process.

[0093] Finally, the model saving and calling capabilities ensure that the trained models can be easily saved, exported, and used on other platforms. In addition, the platform also supports online calling and deployment of models, which can be directly used in actual application scenarios, speeding up the transformation process from model development to actual application.

[0094] In general, the online model training and optimization capabilities of the aluminum electrolysis process training platform provide a one-stop service from data preparation to model deployment. Each link is aimed at improving the prediction accuracy and application effect of the model, fully demonstrating the professionalism and efficiency of the platform.

[0095] 6. 3D visualization capability:

[0096] The 3D visualization capability of the aluminum electrolysis process training platform provides a highly interactive and immersive virtual environment for simulating and monitoring the production process. The platform's 3D visualization technology is based on the Web, ensuring cross-platform accessibility and convenient experience, while achieving a good balance between performance and presentation effects.

[0097] The platform uses advanced 3D visualization technology to present the real-time status of the production site, including but not limited to the operating status of the equipment and the operating results of the model algorithm. This visualization not only enhances the understanding of the production process, but also makes the monitoring process more intuitive and easier to manage by displaying complex data in a graphical way.

[0098] Furthermore, the 3D virtual-real interaction function allows direct participation in the production process through the operation interface. It can monitor process parameters, adjust equipment status in real time, and perform remote control tasks, thereby achieving precise control of the production process. This interactive method not only improves the flexibility of operation, but also enhances the perception and response capabilities of the production environment.

[0099] Overall, the 3D visualization capability of the aluminum electrolysis process training platform provides a powerful, easy-to-operate, and responsive virtual training environment by combining high-performance Web technology, rich 3D displays, and intuitive virtual-real interaction. This not only improves the understanding of the aluminum electrolysis production process, but also provides strong technical support for the monitoring, optimization, and management of the production process.

[0100] 7. BI customization capabilities:

[0101] The aluminum electrolysis process training platform provides a comprehensive and flexible intelligent solution through its BI customization capabilities. The advantage lies in its report design tool, which supports customizing the layout, chart type, data dimensions and key indicators of reports to ensure the personalization and pertinence of reports.

[0102] At the same time, the platform provides rich visualization components that can further enhance the interactive experience, such as tables, charts, maps, and dashboards. These components not only support custom styles, but also customize interactive behaviors, making the presentation of data more intuitive and easy to understand.

[0103] To improve the accessibility and sharing of reports, the platform supports publishing reports to multiple platforms and devices, including the Web and mobile devices. Reports can also be easily shared to promote information exchange and collaboration. The platform has built-in powerful analysis tools that support multi-dimensional analysis, data mining, and predictive analysis to deeply explore the value of data and discover potential business insights. In addition, the platform also supports automated report generation and scheduled tasks, and can set specific conditions to automatically trigger report generation, improve work efficiency, and ensure the accuracy and timeliness of reports.

[0104] The BI customization capability of the aluminum electrolysis process training platform greatly enhances the capabilities in data management, analysis and decision support by providing flexible design tools, rich visualization components, convenient sharing mechanisms, in-depth analysis functions and automated report generation.

[0105] 8. Continuous data service capabilities:

[0106] The continuous data service capability of the aluminum electrolysis process training platform ensures stable and convenient data acquisition and use through a series of comprehensive functions. The data interfaces provided by the platform include API and other types of interfaces, which can be directly called programmatically to realize automatic data acquisition and processing, thereby improving the flexibility and efficiency of data processing.

[0107] In order to meet different data receiving preferences, the platform supports data distribution functions, allowing data to be transmitted through FTP, message queue (MQ), email, etc. This diverse distribution mechanism not only covers various usage scenarios, but also ensures that the most appropriate data receiving method can be selected according to needs.

[0108] The data subscription service further enhances the convenience of obtaining data. You can select the data you are interested in and subscribe to it in the system. The system will regularly push updated data to you, allowing you to focus on the information you care about and reduce the workload of information screening.

[0109] In addition, service monitoring is the key to ensuring the stability and reliability of data services. The platform monitors the performance and operating status of the system to promptly detect and resolve potential problems and ensure the continuity and accuracy of data services. Monitoring of data interfaces and distribution channels is also included to ensure smooth data transmission.

[0110] In summary, the continuous data service capability of the aluminum electrolysis process training platform provides a comprehensive, efficient and reliable data support system by providing direct data interface, flexible data distribution methods, convenient data subscription services, and solid service monitoring, which provides a strong guarantee for the data-driven decision-making process.

[0111] 9. Data security management capabilities:

[0112] The data security management capability of the aluminum electrolysis process training platform is a key component to ensure data security and integrity. It protects data and systems through multi-level security measures. First, the platform uses encryption technology for sensitive data, especially key model data, and uses advanced encryption algorithms to ensure the confidentiality and integrity of data during storage and transmission.

[0113] Secondly, the platform uses a permission management mechanism to set access rights according to roles, ensuring that only authorized persons can access corresponding data resources. This hierarchical permission control effectively reduces the risk of data leakage and abuse and enhances the security of the system.

[0114] In addition, the platform implements a security log recording and monitoring strategy. The system will record all operational behaviors and security events in detail, and through real-time monitoring and auditing, timely discover and respond to security threats. This mechanism provides continuous security protection for the platform, so that any abnormal behavior can be quickly identified and handled.

[0115] In general, the data security management capability of the aluminum electrolysis process training platform has built a comprehensive data protection system through data encryption, permission management, and security log recording and monitoring. These measures work together to ensure the security and integrity of the platform data and protect the interests from infringement.

[0116] On the other hand, the platform provides an aluminum electrolysis process training scenario, and through multi-level training scenario design, it provides comprehensive training and skills improvement channels for personnel at different levels.

[0117] The training at the operator level focuses on the cultivation of basic operating skills and safety awareness. Static training covers the introduction of key operating knowledge, such as pole changing, aluminum tapping, busbar lifting, etc., and provides intuitive learning materials in the form of pictures and on-site videos. At the same time, static training of operating procedures emphasizes the importance of operating procedures and specifications. The equipment knowledge module introduces the basic knowledge and maintenance points of various electrolytic cell systems through a 3D digital twin model. Basic safety and fault training improves operators' awareness of safety risks and fault handling through checklists, texts, videos and safety signs. Dynamic training focuses more on the evaluation and feedback of actual operations, such as key operation evaluation, overall tank condition evaluation and safety disposal evaluation. By connecting with the information corresponding to the on-site electrolytic cell, the operation process is rated and scored in real time to promote the improvement of operator skills.

[0118] The training content for technicians is more in-depth, covering all-round skills improvement from process basics to equipment maintenance. In static training, technicians master the ability to combine theory with practice by learning key process parameters, control basics, report analysis and decision-making. Dynamic training focuses on actual operations. Technicians can adjust process parameters, monitor control curves, make production decisions, and conduct simulation training on safety and troubleshooting in real time. Through 3D digital twin models and real-time data feedback, technicians can deeply understand the equipment status and production process, and comprehensively improve their professional skills and decision-making ability in aluminum electrolysis production.

[0119] The training at the administrator level focuses on the macro level, involving production cost composition, factors affecting production efficiency, report data training and series and enterprise-level decision-making foundations. Administrators master the calculation and control of production costs by learning cost components such as electricity prices, alumina costs, and anode costs. The training on factors affecting production efficiency introduces the impact of process parameter matching, raw material quality, and operation control on production efficiency through pictures and on-site videos. Report data training improves administrators' ability to analyze and make decisions on data by analyzing record reports and series production reports. Decision-making foundation training intuitively displays the effects of various decisions by building decision models. Dynamic training is based on Alwhit's control system and database support to record and evaluate the dynamic effects of production decisions and safety fault handling, and improve administrators' understanding of production management and decision-making.

[0120] For the above-mentioned personnel at different levels, in the aluminum electrolysis training scenario, an aluminum electrolysis process evaluation model based on multimodal data fusion is further established, such as Figure 3 As shown. Through this model, the aluminum electrolysis process training platform will be able to more accurately reflect the effects of actual operations and provide more valuable feedback to operators, technicians and administrators, thereby improving the quality and efficiency of the entire training process. The model establishment process is as follows:

[0121] (1) Collect the operation logs and decision records of operators, technicians and administrators during the training process as log record data; in the platform virtual training operation, record or intercept the video of the training process to capture the operation behavior and operation process, and the captured video is used as video data; the current, voltage, temperature, etc. collected in real time by sensors from the electrolytic cell and related equipment are used as sensor data;

[0122] The video data includes the operation videos of operators, technicians and administrators during the aluminum electrolysis process training; the sensor data includes the current, voltage, temperature, number of material unloading, material unloading amount, aluminum tapping time, and pole changing time of the aluminum electrolysis cell.

[0123] (2) Log data, video data, and sensor data are combined as a dataset and divided into training set, validation set, and test set;

[0124] (3) Model structure design: Sub-network A uses a long short-term memory network (LSTM) to process sequence data. This sub-network can capture long-term dependencies in log data and understand the order and logic of operations. Sub-network B uses a convolutional neural network (CNN) to extract key features in the video. Sub-network C uses a fully connected layer (ANN) to process sensor data. The fully connected layer is used as a fusion layer to fuse the outputs of the above sub-networks.

[0125] (4) Use the training set to train the model, use the validation set to evaluate the performance of the model, adjust the hyperparameters to optimize the accuracy and generalization ability of the model, and save the trained network model parameters;

[0126] (5) Using the log record data in the test set as the input of the trained sub-network A, the video data as the input of the trained sub-network B, and the sensor data as the input of the trained sub-network C, and then using the output of each sub-network as the input of the trained fusion layer to obtain dynamic training evaluation results;

[0127] The above model will score the training performance at each level based on the standardization of operations, the rationality of decisions, and the degree of understanding of equipment and processes. For operators, the focus will be on the accuracy and safety of operations; for technicians, the model will assess their proficiency in adjusting process parameters and their ability to optimize production processes; for managers, the assessment will focus on the comprehensiveness and strategic nature of decisions, as well as a deep understanding of the impact on production costs and efficiency.

[0128] In some embodiments, the LSTM network of subnetwork A is defined as follows:

[0129] F t =σ(W f ·[H t-1 ,X t ]+b f );

[0130] I t =σ(W i ·[H t-1 ,X t ]+b i );

[0131]

[0132] O t =σ(W o ·[H t-1 ,X t ]+b o );

[0133] H t =O t *tanh(C t );

[0134] Where σ is the sigmoid activation function, tanh is the hyperbolic tangent function, and F t is the forget gate at time t, I t is the input gate at time t, C t is the cell state at time t, is the candidate cell state at time t, O t is the output gate at time t, H t is the hidden state at time t, X t is the log record data at time t, H t-1 is the hidden state at time t-1, C t-1 is the cell state at time t-1, W f , W i , W o , W c are the weights of the forget gate, input gate, output gate and candidate cell state, respectively, and b f , b i , b o , b c They are the forget gate, input gate, output gate and bias of the candidate cell state respectively, and * indicates element-by-element multiplication.

[0135] In some embodiments, the definition of the CNN network of subnetwork B is as follows:

[0136] The CNN network consists of multiple cascaded convolutional networks. The convolutional network includes convolutional layers, pooling layers, and fully connected layers. The expressions of convolutional layers, pooling layers, and fully connected layers are as follows:

[0137]

[0138] Among them, x0 is the video data, x is the input of the lth convolutional layer in the CNN network, is the output of the lth convolutional layer in the CNN network, K is the convolution kernel, k h is the height of the convolution kernel, k w is the width of the convolution kernel, C is the number of channels of the convolution kernel, i and j are the positions of the output feature maps, is the output of the lth pooling layer in the CNN network, x l is the output of the lth fully connected layer in the CNN network, σ is the sigmoid activation function, and b l is the bias of the lth fully connected layer in the CNN network, W l is the weight of the lth fully connected layer in the CNN network, is the size of the kth dimension of the video data, k = 1, 2, ..., d; d is the dimension of the video data, and L is the number of layers of the CNN network.

[0139] In some embodiments, the sub-network C and the ANN network of the fusion layer are defined as follows:

[0140] y (g) =f(W (g) x (g-1) +b (g) )

[0141] Among them, y (g) is the output of the g-th layer, W (g) is the weight of the g-th layer, b (g) is the weight bias of the g-th layer, x (g-1) is the input of the gth layer, for the output layer, this will be the output data, and f is the activation function.

[0142] In some embodiments, the following steps are included: the log data needs to be cleaned to remove invalid and redundant information, and then formatted into a sequence format suitable for input into sub-network A; the video data needs to be preprocessed, such as size standardization and normalization, to improve the efficiency and accuracy of subsequent feature extraction; the sensor data is usually time series data, which needs to be smoothed and normalized to reduce noise and highlight important patterns and trends.

[0143] Overall, the training scenario design of the aluminum electrolysis process training platform not only covers all levels from operators to administrators, but also comprehensively improves the professional skills and emergency response capabilities of practitioners through a combination of static and dynamic training, ensuring the efficient, safe and sustainable development of aluminum electrolysis production.

[0144] Example 2

[0145] This embodiment provides the specific deployment, operation process and expected results of the aluminum electrolysis intelligent process training platform to reflect its efficiency and practicality in actual applications.

[0146] (1) Platform deployment

[0147] 1) System installation and configuration: The first step of the embodiment involves installing the basic software environment required for the training platform on a dedicated server. This includes the installation of the operating system, the configuration of the database system, and the deployment of middleware to ensure that the software foundation for the platform operation is solid and reliable.

[0148] 2) Implementation of microservice architecture: According to the principles of microservice architecture, the platform functions are split into independent microservice units and deployed on the server. Each microservice is responsible for a single function and communicates through well-defined APIs to achieve interoperability and independent scalability between services.

[0149] 3) Data source integration: The platform integrates multiple data sources, such as sensor data, production control systems, etc., to achieve real-time data collection and unified management. The integration of data sources adopts a design that adapts to multiple communication protocols to ensure seamless access to data from different sources.

[0150] 4) Algorithm deployment and expansion: Deploy the predefined algorithm model to the corresponding microservice, and provide an algorithm orchestration tool that can customize and expand the algorithm according to specific needs, and realize dynamic management and optimization of the algorithm.

[0151] (2) Operation process

[0152] 1) User login and authority verification: Different roles (operators, technicians, administrators) log in to the platform through a security authentication mechanism and access corresponding functions and data based on their authority.

[0153] 2) Training scenario selection: Select appropriate training scenarios based on their roles and training objectives. For example, operators’ pole-changing operation training, technicians’ process parameter setting training, or managers’ production decision-making dynamics training.

[0154] 3) Practical training and feedback: simulate actual operations in a 3D visualization interface to conduct skill training. The platform provides real-time operational feedback and guidance, simulates the real production environment, understands the impact of operations, and improves operational skills.

[0155] 4) Data analysis and decision support: Use built-in BI tools to analyze production data, generate reports, and obtain decision support, thereby improving the scientificity and accuracy of decision-making.

[0156] 5) Performance monitoring and optimization: The platform monitors algorithm and system performance in real time, analyzes performance bottlenecks, and provides optimization suggestions to ensure efficient and stable operation of the system and algorithm.

[0157] (3) Expected results

[0158] Through this training platform, it is expected that many positive effects will be achieved. Operators can significantly improve their operating skills and standardization by simulating the real operating environment. Technicians and managers can use data analysis tools to make more scientific process adjustments and production decisions. Automated control and optimized decision-making can help reduce waiting and delays in the production process, thereby improving production efficiency. Intelligent monitoring and early warning systems can help to timely discover and deal with potential safety issues and reduce safety risks. Predictive maintenance and automated troubleshooting can reduce equipment downtime and maintenance costs.

[0159] (4) Implementation effect evaluation

[0160] To ensure the effectiveness of the implementation, the platform records all operations and system feedback during the training process, providing detailed data support for effect evaluation. By comparing the production data and operations before and after the training, the training effect and skill improvement can be evaluated. In addition, feedback on the training platform, including operating experience, interface friendliness, and training effect, is collected to provide valuable information for the continuous optimization of the platform.

[0161] In summary, the intelligent process training platform for aluminum electrolysis can not only quickly improve professional skills through a highly simulated production environment, but also optimize production decisions, improve production efficiency and safety, and reduce maintenance costs, providing strong support for the intelligent upgrade of aluminum electrolysis technology.

[0162] Example 3

[0163] This embodiment provides an electronic device, including:

[0164] one or more processors;

[0165] A memory having one or more programs stored thereon, which, when executed by one or more processors, enables the one or more processors to implement the steps of an aluminum electrolysis process evaluation method based on multimodal data.

[0166] In some implementations, the memory may be a high-speed random access memory (RAM), and may also include a non-volatile memory, such as at least one disk memory.

[0167] In some other implementations, the processor may be a central processing unit (CPU), a digital signal processor (DSP), or other general-purpose processors of various types, which are not limited herein.

[0168] This embodiment provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of an aluminum electrolysis process evaluation method based on multimodal data.

[0169] The contents explained in the above embodiments should be understood as these embodiments are only used to more clearly illustrate the present invention, and are not used to limit the scope of the present invention. After reading the present invention, various equivalent forms of modifications to the present invention by those skilled in the art all fall within the scope defined by the claims attached to this application.

Claims

1. A method for evaluating aluminum electrolysis process based on multimodal data, characterized in that: The following steps are involved: S1: Obtain log data and video data during the aluminum electrolysis process training, and obtain sensor data of industrial equipment in the aluminum electrolysis workshop; S2: Preprocessing the log data, video data, and sensor data; S3: using the preprocessed log data, video data, and sensor data as inputs of a deep learning network model, and training the deep learning network model, wherein the deep learning network model includes a CNN network, an LSTM network, a first fully connected layer, and a second fully connected layer, and the CNN network, the LSTM network, and the first fully connected layer are respectively connected to the second fully connected layer; The training process of the deep learning network model includes: The LSTM network is used to process the preprocessed log record data to obtain a first feature; Using the CNN network to process the preprocessed video data to obtain a second feature; Processing the preprocessed sensor data using the first fully connected layer to obtain a third feature; fusing the first feature, the second feature, and the third feature through the second fully connected layer to obtain an evaluation result; The hyperparameter combination is optimized through automatic parameter adjustment and the deep learning network model is iteratively optimized and trained to obtain an aluminum electrolysis process evaluation model.

2. The aluminum electrolysis process evaluation method based on multimodal data according to claim 1 is characterized in that: The log record data includes the operation logs and decision records of operators, technicians and administrators during the aluminum electrolysis process training; the video data includes the operation videos of operators, technicians and administrators during the aluminum electrolysis process training; the sensor data includes the current, voltage, temperature, number of material unloading, material unloading amount, aluminum tapping time, and pole changing time of the aluminum electrolysis cell.

3. The aluminum electrolysis process evaluation method based on multimodal data according to claim 1 is characterized in that: The specific process of S2 includes: Cleaning and formatting the log record data; Performing size standardization and normalization processing on the video data; The sensor data is smoothed and normalized.

4. The aluminum electrolysis process evaluation method based on multimodal data according to claim 1 is characterized in that: The LSTM network includes an input gate, a forget gate, an output gate, and a candidate cell state. The expressions of the input gate, the forget gate, the output gate, and the candidate cell state are as follows: F t =σ(W f ·[H t-1 ,X t ]+b f ); I t =σ(W i ·[H t-1 ,X t ]+b i ); The t =σ(W o ·[H t-1 ,X t ]+b o ); H t =O t *tanh(C t ); Among them, σ is the sigmoid activation function, tanh is the hyperbolic tangent function, and F t is the forget gate at time t, I t is the input gate at time t, C t is the cell state at time t, is the candidate cell state at time t, O t is the output gate at time t, H t is the hidden state at time t, X t is the log record data at time t, H t-1 is the hidden state at time t-1, C t-1 is the cell state at time t-1, W f , W i , W o , W c are the weights of the forget gate, input gate, output gate and candidate cell state, respectively, and b f 、b i 、b o 、b c They are the forget gate, input gate, output gate and bias of the candidate cell state respectively, and * indicates element-by-element multiplication.

5. The aluminum electrolysis process evaluation method based on multimodal data according to claim 1 is characterized in that: The CNN network includes multiple cascaded convolutional networks, and the convolutional network includes a convolutional layer, a pooling layer, and a fully connected layer. The expressions of the convolutional layer, the pooling layer, and the fully connected layer are as follows: x l =σ(W l *Y2+b l ),l=1,2,…,L-1, Among them, x0 is the video data, x is the input of the lth convolutional layer in the CNN network, and Y1 (l) is the output of the lth convolutional layer in the CNN network, K is the convolution kernel, k h is the height of the convolution kernel, k w is the width of the convolution kernel, C is the number of channels of the convolution kernel, i and j are the positions of the output feature map, Y2 (l) is the output of the lth pooling layer in the CNN network, x l is the output of the lth fully connected layer in the CNN network, σ is the sigmoid activation function, and b l is the bias of the lth fully connected layer in the CNN network, W l is the weight of the lth fully connected layer in the CNN network, is the size of the kth dimension of the video data, k = 1, 2, ..., d; d is the dimension of the video data, and L is the number of layers of the CNN network.

6. The aluminum electrolysis process evaluation method based on multimodal data according to claim 1 is characterized in that: The expression of the first fully connected layer or the second fully connected layer is as follows: y (g) =f(W (g) x (g-1) +b (g) ) Among them, y (g) is the output of the g-th layer, W (g) is the weight of the g-th layer, b (g) is the weight bias of the g-th layer, x (g-1) is the input of the g-th layer, and f is the activation function.

7. The aluminum electrolysis process evaluation method based on multimodal data according to claim 1 is characterized in that: The log data, video data and sensor data are input into the aluminum electrolysis process evaluation model to obtain the evaluation results, and the aluminum electrolysis process is controlled and optimized according to the evaluation results.

8. An aluminum electrolysis process training platform based on the method according to any one of claims 1 to 7, characterized in that: It includes real-time monitoring unit, aluminum electrolysis process training unit, data storage unit, model training optimization unit, and evaluation feedback unit; Real-time monitoring unit, used to monitor and collect sensor data of industrial equipment in aluminum electrolysis workshop in real time; The aluminum electrolysis process training unit builds a one-to-one digital twin model based on the aluminum electrolysis workshop and constructs an aluminum electrolysis process training scenario for operators, technicians, and administrators to conduct aluminum electrolysis process training; A data storage unit is used to integrate and store sensor data collected by the real-time monitoring unit, operation logs and decision records during aluminum electrolysis process training, and video data according to time; A model training optimization unit is used to train an aluminum electrolysis process evaluation model using sensor data, operation logs and decision records, and video data as input data; The evaluation feedback unit is used to control and optimize the aluminum electrolysis process according to the evaluation results output by the aluminum electrolysis process evaluation model.

9. An electronic device, characterized in that: include: one or more processors; A memory having one or more programs stored thereon, which, when the one or more programs are executed by the one or more processors, enables the one or more processors to implement the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The computer program is stored therein, and when the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images