Intelligent edge computing method, system and terminal based on AI artificial intelligence

Abstract

The present invention relates to the field of edge computing technology, and specifically, to an intelligent edge computing method, system and terminal based on AI artificial intelligence. It includes the following steps: collecting operation data through edge devices; using long short-term memory network models based on operation data to predict equipment failures, and training the long short-term memory network models on the cloud; deploying the trained long short-term memory network models to edge devices; the long short-term memory network models in the edge devices continuously analyze the newly collected data in real time and respond quickly. The design of the present invention can realize real-time analysis and rapid response to newly collected data by deploying the trained long short-term memory network model to the edge device. This greatly reduces the time delay for data to be transmitted to the cloud and then returned to the processing results, which is particularly important for application scenarios that require real-time decision-making.

Description

Technical Field

[0001] The present invention relates to the field of edge computing technology, and specifically, to an intelligent edge computing method, system and terminal based on AI artificial intelligence. Background Art

[0002] In the wave of digital transformation, data has become a key resource to drive innovation and growth in all walks of life. With the widespread deployment of Internet of Things (IoT) devices, we are witnessing an explosive growth in the amount of data generated, which comes from many fields such as smart cities, industrial automation, health monitoring, and personal devices. Although the traditional cloud computing model has achieved great success in centrally processing and analyzing large-scale data sets, it faces challenges such as increased latency, bandwidth limitations, and privacy protection. Therefore, an intelligent edge computing method, system, and terminal based on AI artificial intelligence are provided. Summary of the invention

[0003] The purpose of the present invention is to provide an intelligent edge computing method, system and terminal based on AI artificial intelligence to solve the problems of increased latency, bandwidth limitation and privacy protection in the traditional cloud computing model proposed in the above background technology.

[0004] To achieve the above objectives, the present invention provides an intelligent edge computing method based on AI artificial intelligence, comprising the following steps:

[0005] S1, collect operation data through edge devices;

[0006] S2. Use the LSTM network model to predict equipment failures based on operating data, and train the LSTM network model on the cloud. By integrating multi-source information, the training process of the LSTM network model is optimized to make the model better adapt to data drift.

[0007] S3, deploy the trained LSTM network model to the edge device;

[0008] S4. The long short-term memory network model in the edge device continuously analyzes the newly collected data in real time and responds quickly.

[0009] As a further improvement of the technical solution, in S2, the long short-term memory network model is used to predict equipment failure based on the operation data, and the long short-term memory network model is trained on the cloud, including the following steps:

[0010] S2.1, preprocessing the operation data;

[0011] S2.2, extracting features from the operating data that can reflect the status of the equipment;

[0012] S2.3. Determine the specific structure of the LSTM network, clarify the length of the time series input to the LSTM network and the output format, and set the hyperparameters;

[0013] S2.4, the preprocessed and feature-extracted running data set is divided into a training set, a validation set, and a test set;

[0014] S2.5. Use computing resources on the cloud platform to train the long short-term memory network model, define the loss function, optimize the loss function to address the data drift problem, and optimize the loss function by fusing multi-source information to make the model better adapt to data drift. During the training process, optimize the model algorithm, improve the efficiency of model training through distributed training technology, and protect the security of data transmission through TLS / SSL encryption protocol;

[0015] S2.6. Use the training set data to optimize the model parameters and use the validation set to adjust the hyperparameters.

[0016] As a further improvement of the technical solution, in S2.5, the loss function is:

[0017] ;

[0018] in, Indicates the degree of difference between the model prediction results and the actual results; Indicates the number of data samples contained in the data set involved in training; Indicates the number of types of faults; Indicates The sample in The true label value of each category; The model predicts The samples belong to The probability value of each category; An index indicating the number of categories; Indicates the sample index;

[0019] To address the data drift problem, the loss function is optimized:

[0020] ;

[0021] in, Indicates the degree of difference between the model prediction results and the actual results after addressing the data drift problem; Indicates the parameters that adjust the impact of data drift;

[0022] By fusing multi-source information to optimize the loss function, the model can better adapt to data drift:

[0023] ;

[0024] in, Indicates the degree of difference between the model prediction results and the actual results after integrating multi-source information; Indicates the number of additional information sources; Indicates the measurement The weight parameter of the importance of the additional information source in the entire loss function; Indicates of the additional information sources The data value of each sample; Represents the model for of the additional information sources The predicted value of a sample.

[0025] As a further improvement of the technical solution, in S2.5, the efficiency of model training is improved by using distributed training technology, including the following steps:

[0026] S2.51. According to computing resources, the training data set is evenly divided into multiple subsets, and each computing node processes one subset;

[0027] S2.52. Deploy a complete copy of the LSTM network model on each computing node;

[0028] S2.53, each node independently performs forward propagation on its assigned data subset and calculates the corresponding loss value;

[0029] S2.54, based on its own loss value, each node independently calculates the gradient;

[0030] S2.55. Use the hierarchical All-Reduce algorithm to aggregate the gradients calculated by all nodes. Since the more nodes there are, the greater the communication overhead and the longer the gradient aggregation time is, the communication complexity increases exponentially. The hierarchical All-Reduce algorithm is optimized.

[0031] S2.56. Based on the aggregated gradient information, the synchronous update algorithm is used to update the model parameters. The synchronous update algorithm is optimized to address the problem that the parameter update oscillates back and forth around the minimum value, which slows down the convergence speed.

[0032] As a further improvement of the technical solution, in S2.55, the hierarchical All-Reduce algorithm is optimized, including the following steps:

[0033] S2.551, divide all nodes into several groups, and the number of nodes in each group is close;

[0034] S2.552, perform All-Reduce operation within each group to obtain the local aggregation result of each group;

[0035] S2.553, perform All-Reduce operation on the local aggregation results of all groups again to obtain the global aggregation result;

[0036] S2.554, broadcast the global aggregation results back to all nodes.

[0037] As a further improvement of the technical solution, in S2.56, the synchronization update algorithm is:

[0038] ;

[0039] in, represents the updated model parameters; Represents the learning rate, which controls the step size of parameter update at each iteration; Represents the model parameters before updating; Indicates the batch size; Indicates The loss function on samples Parameters The gradient of

[0040] To solve the problem that parameter updates oscillate back and forth near the minimum value, which slows down the convergence speed, the synchronous update algorithm is optimized:

[0041] ;

[0042] in, Represents the model parameters after optimization and update; represents the momentum term coefficient; Represents the cumulative gradient at the last iteration.

[0043] As a further improvement of the technical solution, in S3, deploying the trained long short-term memory network model to the edge device in the production line includes the following steps:

[0044] S3.1, convert the model into a format suitable for edge deployment;

[0045] S3.2, deploy the model manager together with the trained LSTM network model to the edge device;

[0046] S3.3, when the cloud completes the training of the new version of the model and confirms that the performance of the new version of the model is better than the existing model, the new model is sent to the model manager on the edge device through a secure channel;

[0047] S3.4, the model manager receives the new model file and verifies its integrity and correctness;

[0048] S3.5, the model manager allows to unload old models and load new models;

[0049] S3.6, after the new model is loaded, it immediately starts to analyze the newly collected data, and the monitoring system continuously collects performance data and resource usage during the model execution;

[0050] S3.7. All communication interfaces adopt security measures.

[0051] As a further improvement of the technical solution, in S4, the long short-term memory network model in the edge device continuously analyzes the newly collected data in real time and responds quickly, including the following steps:

[0052] S4.1, the long short-term memory network model in the edge device continuously analyzes the newly collected data in real time;

[0053] S4.2. Make a quick response based on the analysis results;

[0054] S4.3. Define common interfaces, protocols and data formats;

[0055] S4.4. Asynchronous message processing is achieved through message queues and event-driven architecture.

[0056] On the other hand, the present invention provides an intelligent edge computing system based on AI artificial intelligence, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-mentioned intelligent edge computing method based on AI artificial intelligence;

[0057] Furthermore, the present invention also provides a terminal, including a processor, a memory, and a computer control unit stored in the memory and executable on the processor, for executing the AI ​​artificial intelligence-based intelligent edge computing method as described above.

[0058] Compared with the prior art, the present invention has the following beneficial effects:

[0059] 1. In this AI-based intelligent edge computing method, system and terminal, by deploying the trained long short-term memory network model to the edge device, it is possible to achieve instant analysis and rapid response to newly collected data. This greatly reduces the time delay for data transmission to the cloud and then returning the processing results, which is particularly important for application scenarios that require real-time decision-making (such as industrial automation, smart cities, and health monitoring). Edge computing can provide faster response speeds, ensure the efficient operation of the system, and can take quick action in emergency situations (such as equipment failure warnings).

[0060] 2. In this AI-based intelligent edge computing method, system and terminal, since most of the data processing occurs on the edge device, only a small amount of key information or abnormal situations will be sent to the cloud for further analysis, thus greatly reducing the risk of sensitive data leakage. At the same time, the security of the entire system is guaranteed by using TLS / SSL encryption protocol to protect the security of data transmission and digital signature technology to verify the integrity of the model file. This approach not only meets the current strict privacy regulations, but also increases users' trust in the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 The figure is a flow chart of the overall method of the present invention. DETAILED DESCRIPTION

[0062] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0063] Example 1: Please refer to Figure 1 As shown, this embodiment provides an intelligent edge computing method based on AI artificial intelligence, including the following steps:

[0064] S1. Collect operation data through edge devices (such as sensors, cameras, etc.)

[0065] In this embodiment, the operating data may come from various devices, including temperature, pressure, vibration, sound, visual information (such as through a camera), etc.;

[0066] S2. Use the long short-term memory network model to predict equipment failure based on the operating data, and train the long short-term memory network model on the cloud;

[0067] In this embodiment, a long short-term memory network model is used to predict equipment failure based on operation data, and the long short-term memory network model is trained on the cloud, including the following steps:

[0068] S2.1. Preprocess the operation data, including removing noise, outliers and incomplete records to ensure data quality;

[0069] S2.2. Extract the features that can reflect the equipment status from the operation data, including mean, variance, peak value, etc.;

[0070] S2.3. Determine the specific structure of the LSTM network, such as the number of layers, the number of units in each layer, etc., clarify the length of the time series input to the LSTM network and the output format (use statistical methods (such as autocorrelation function) to determine the significant time dependency in the data, and select the time series length accordingly; if the task is to predict a single future value (such as the temperature at the next time point), the output format can be a scalar value; if the task is to predict multiple future values ​​(such as temperature changes in the next 10 minutes), the output format can be a vector; if the task is to perform classification (such as fault detection), the output format can be the probability distribution of the category; if the task is to generate a new time series (such as predicting all temperature values ​​in the future), the output format can be a sequence), set hyperparameters such as learning rate, batch size, and number of iterations;

[0071] S2.4, the preprocessed and feature-extracted running data set is divided into a training set, a validation set, and a test set;

[0072] S2.5. Use computing resources on the cloud platform to train the long short-term memory network model, define the loss function, optimize the loss function to address the data drift problem (the distribution of input data changes due to changes in the working state of the long-term running equipment), and optimize the loss function by fusing multi-source information (data from different data sources) to make the model better adapt to data drift. Perform algorithm optimization on the model during training (algorithm optimization includes model pruning (removing unimportant weights), quantization (reducing the precision of weights and activations), and knowledge distillation (transferring knowledge from large models to small models)). During training, use various regularization techniques, such as dropout and weight decay, to prevent model overfitting, and use distributed training technology to improve the efficiency of model training. Use TLS / SSL encryption protocols to protect the security of data transmission.

[0073] Furthermore, the loss function is:

[0074] ;

[0075] in, It indicates the difference between the model prediction result and the actual result, and is the objective function that needs to be minimized during the model training process. Indicates the number of data samples contained in the data set involved in training; Indicates the number of types of faults; Indicates The sample in The true label value of each category; The model predicts The samples belong to The probability value of each category ranges from 0 to 1; An index indicating the number of categories; Indicates the sample index;

[0076] Over time, the working environment and status of the equipment may change, such as temperature, humidity, degree of wear, etc. These changes will cause the distribution of input data to change. If the model cannot adapt well to such changes, its prediction performance may decline. Taking data drift into account by optimizing the loss function can help the model maintain a high prediction accuracy; the optimized loss function allows the model to learn more generally applicable features, rather than relying solely on data characteristics of a specific time period, which helps to improve the model's generalization ability under different conditions, allowing the model to make accurate predictions even when faced with new or unseen data; for continuously running systems, it is very important to be able to automatically adapt to data drift. By continuously updating model parameters and adjusting the loss function according to the latest data distribution, the model can have stronger dynamic adaptability to ensure its long-term effectiveness;

[0077] To address the data drift problem (the distribution of input data changes due to changes in the working state of the device running for a long time), the loss function is optimized:

[0078] ;

[0079] in, Indicates the degree of difference between the model prediction results and the actual results after addressing the data drift problem; Indicates the parameters that adjust the impact of data drift. The larger the value of, the greater the adjustment of data drift;

[0080] In practical applications, the operating status of equipment may be affected by a variety of factors, such as environmental conditions, operating modes, maintenance history, etc. The data of these factors may come from different data sources (such as sensor readings, maintenance records, external weather data, etc.). Integrating these multi-source information into the loss function can help the model understand the state changes of the equipment more comprehensively, thereby improving its adaptability to complex environments; the information provided by different data sources can supplement the deficiencies of a single data source and provide richer and more comprehensive data features, which helps the model capture more potential correlations, thereby improving the accuracy of fault prediction. For example, by combining equipment maintenance records with real-time operation data, the model may be able to identify signs of impending failures earlier; data drift refers to the phenomenon that data distribution changes over time. By fusing information from different data sources, especially those that can reflect long-term trends or periodic changes (such as seasonal maintenance activities), the model can better adapt to such changes. For example, if the maintenance frequency of equipment changes with the seasons, then multi-source information including maintenance records can help the model adjust to such seasonal changes;

[0081] By fusing multi-source information (data from different data sources) to optimize the loss function, the model can better adapt to data drift:

[0082] ;

[0083] in, Indicates the degree of difference between the model prediction results and the actual results after integrating multi-source information; represents the number of additional information sources, that is, Different multi-source information is integrated into the loss function; Indicates the measurement The weight parameter of the importance of the additional information source in the entire loss function; Indicates of the additional information sources The data value of the sample, for example, if An additional source of information is equipment maintenance records. Indicates A relevant value in the maintenance record (such as maintenance time, maintenance cost, etc.); Represents the model for of the additional information sources The predicted value of the sample, for example, if An additional source of information is equipment maintenance records. It is a value such as maintenance time or maintenance cost predicted by the model.

[0084] Furthermore, fusing multi-source information includes the following steps: for each sample In addition to collecting the data required for the main task (such as equipment operation data), it is also necessary to collect data from all relevant additional information sources, organize these data into an appropriate form, and ensure that they can be compared and used at the same time point; extract useful features from various data sources, which may include the original data itself, or may include derived features calculated based on these data; design a model architecture that can handle multiple inputs, and ensure that the model can output two parts: one is the prediction for the main task (such as failure probability), and the other is the prediction for additional information sources (such as maintenance costs or weather conditions); define the above The loss function, where the first term focuses on the accuracy of the main task and the second term focuses on the accuracy of the additional information source; by adjusting to balance the importance of different information sources; during training, the model tries to minimize loss function, which means that the model should not only accurately predict the main task, but also predict the data from the additional information source as accurately as possible.

[0085] Furthermore, distributed training technology is to divide the data set into multiple parts, each of which is processed independently by different computing nodes, and then merge the results of these nodes in some way to update the global model parameters; distributed training technology uses multiple machines or multiple computing nodes to process data in parallel, which can significantly reduce the time required for model training; for very large data sets, a single machine may not be able to accommodate all the data, and distributed training allows the data to be divided into small pieces and processed by different computing nodes, so that larger data sets can be processed. Since the data is divided into multiple subsets and each computing node can process a subset independently, the total training time is greatly shortened; if a computing node fails, other nodes can continue to work, and even restart the failed task on a new node, thereby maintaining the stability of the entire system;

[0086] Improving the efficiency of model training through distributed training technology includes the following steps:

[0087] S2.51. Based on the available computing resources, the training data set is evenly divided into multiple subsets, and each computing node (such as a virtual machine or container on a cloud platform) processes one subset;

[0088] S2.52. Deploy a complete copy of the LSTM network model on each computing node to ensure that the initial model parameters of all nodes are consistent;

[0089] S2.53, each node independently performs forward propagation on its assigned data subset and calculates the corresponding loss value;

[0090] S2.54, based on its own loss value, each node independently calculates the gradient;

[0091] S2.55. Use the hierarchical All-Reduce algorithm to aggregate the gradients calculated by all nodes (the hierarchical All-Reduce algorithm is a common communication protocol used to efficiently exchange and merge gradient information between all nodes. It ensures that all nodes eventually have the same gradient value). This step ensures that all nodes can obtain global gradient information, thereby maintaining the consistency of the model. Since the more nodes there are, the greater the communication overhead, the longer the gradient aggregation time, and the exponential growth of communication complexity, the hierarchical All-Reduce algorithm is optimized;

[0092] Furthermore, as the number of nodes increases, each node needs to exchange gradient information with other nodes, which will cause significant communication delays. By using a hierarchical All-Reduce strategy, the nodes can be divided into small groups. Local aggregation is first performed within each group, and then global aggregation is performed between groups. This can reduce the number of nodes directly involved in global communication, thereby reducing communication delays. When the number of nodes is large, the communication complexity of the traditional All-Reduce algorithm will increase exponentially, resulting in a significant increase in the overall training time. Such optimization allows the communication cost to be kept low even in large-scale distributed systems. When the number of nodes increases, any problem with a single node or network connection may affect the entire training process. Hierarchical All-Reduce reduces the impact of single point failures by grouping, and can alleviate the impact of local failures by rescheduling tasks, thereby enhancing the stability and reliability of the system.

[0093] Optimizing the hierarchical All-Reduce algorithm includes the following steps:

[0094] S2.551. Divide all nodes into several groups, with the number of nodes in each group being equal or close;

[0095] S2.552, perform All-Reduce operation within each group to obtain the local aggregation result of each group;

[0096] S2.553, perform All-Reduce operation on the local aggregation results of all groups again to obtain the global aggregation result;

[0097] S2.554, broadcast the global aggregation results back to all nodes;

[0098] The specific mathematical expression is:

[0099] If there are a total of nodes, divided into groups, each with nodes;

[0100] The nodes in each group perform All-Reduce operations, and the communication complexity is ;

[0101] The local aggregation results of all groups are subjected to All-Reduce operation, and the communication complexity is ;

[0102] The total communication complexity is: ;

[0103] S2.56. Update the model parameters using the synchronous update algorithm based on the aggregated gradient information (the synchronous update algorithm is a parameter update method commonly used in distributed training, which ensures that all computing nodes use the latest global gradient information to update the model parameters at each iteration). Optimize the synchronous update algorithm to solve the problem that the parameter update oscillates back and forth around the minimum value, resulting in a slow convergence speed.

[0104] Furthermore, the synchronization update algorithm is:

[0105] ;

[0106] in, represents the updated model parameters; Represents the learning rate, which controls the step size of parameter update at each iteration; Represents the model parameters before updating; represents the batch size, that is, the number of samples used in one iteration; Indicates The loss function on samples Parameters The gradient of the loss function is the optimized loss function in step 4. ;

[0107] When the parameters are close to the minimum value of the loss function, the gradient usually becomes very small, which may result in a small parameter update amplitude, so that the model requires more iterations to reach the optimal solution; in some cases, especially in high curvature areas, gradient descent may cause the parameters to oscillate back and forth near the local minimum. This phenomenon is called "oscillation". By introducing the momentum term, the parameter update path can be smoothed, reducing this unnecessary oscillation, and making the model move more stably toward the optimal solution; the traditional gradient descent method is very sensitive to the choice of learning rate. Inappropriate learning rate setting will lead to unstable training and may even fail to converge. By using an optimization algorithm that includes momentum, this problem can be alleviated to a certain extent because it allows the use of a larger learning rate without causing training divergence;

[0108] To solve the problem that parameter updates oscillate back and forth near the minimum value, which slows down the convergence speed, the synchronous update algorithm is optimized:

[0109] ;

[0110] in, Represents the model parameters after optimization and update; Represents the momentum term coefficient, which determines the influence of historical gradient accumulation; Represents the cumulative gradient at the last iteration, usually initialized to 0 and updated to after each iteration ,here This will be used as the next iteration ;

[0111] S2.6. Use the training set data to optimize the model parameters, and use the validation set to adjust the hyperparameters to avoid overfitting.

[0112] S3, deploy the trained LSTM network model to edge devices, such as edge gateways or smart controllers;

[0113] In this embodiment, the trained long short-term memory network model is deployed to the edge device in the production line, including the following steps:

[0114] S3.1. Convert the model to a format suitable for edge deployment, such as ONNX or TensorFlow Lite, to ensure compatibility with the inference engine.

[0115] S3.2, deploy the model manager together with the trained LSTM network model to the edge device;

[0116] S3.3, when the cloud completes the training of the new version of the model and confirms that the performance of the new version of the model is better than the existing model, the new model is sent to the model manager on the edge device through a secure channel;

[0117] S3.4. The model manager receives the new model file and verifies its integrity and correctness (before the model file is sent, the file content is digitally signed using the private key. On the edge device, the corresponding public key is used to verify the validity of the signature. If the signature is valid, it means that the file has not been tampered with during transmission). The model manager is a key component that manages and maintains the machine learning models deployed on the edge devices.

[0118] S3.5. The model manager allows the old model to be unloaded and the new model to be loaded without restarting the entire system.

[0119] S3.6. After a new model is loaded, it immediately starts to analyze the newly collected data to ensure continuous service provision, and the monitoring system continuously collects performance data and resource usage during model execution to ensure that the system is in optimal working condition;

[0120] S3.7. A modular and abstract communication module design is adopted to support efficient and reliable communication. All communication interfaces adopt security measures, such as data encryption and authentication mechanisms, to ensure the security of data transmission.

[0121] S4, the long short-term memory network model in the edge device continuously analyzes the newly collected data in real time and responds quickly;

[0122] In this embodiment, the long short-term memory network model in the edge device continuously analyzes the newly collected data in real time and responds quickly, including the following steps:

[0123] S4.1, the long short-term memory network model in the edge device continuously analyzes the newly collected data in real time;

[0124] S4.2. Respond quickly based on the analysis results, such as issuing an alarm, adjusting equipment parameters, or triggering a maintenance process;

[0125] S4.3. Define common interfaces (RESTful API), protocols (MQTT), and data formats (JSON) so that different systems and services can interact with smart terminals in a consistent manner and ensure effective integration with Web-based services and applications;

[0126] S4.4. Asynchronous message processing is implemented through message queues (message queues are an asynchronous communication method in which producers send messages to queues and consumers get messages from queues and process them, such as RabbitMQ) and event-driven architectures (event-driven architecture is a software architecture pattern in which communication between components is triggered by events, which can be generated by external systems or internal components, such as using Node.js and EventEmitter to implement a simple event-driven architecture) to ensure the response speed and stability of the system.

[0127] Embodiment 2: This embodiment provides an intelligent edge computing system based on AI artificial intelligence, including a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor executes the computer program to implement the above-mentioned intelligent edge computing method based on AI artificial intelligence;

[0128] This embodiment also provides a terminal, including a processor, a memory, and a computer control unit stored in the memory and executable on the processor, for executing the above-mentioned AI-based intelligent edge computing method.

[0129] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. The above embodiments and descriptions are only preferred examples of the present invention and are not intended to limit the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. The intelligent edge computing method based on AI artificial intelligence is characterized by: The following steps are involved: S1, collect operation data through edge devices; S2. Use the LSTM network model to predict equipment failures based on operating data, and train the LSTM network model on the cloud. By integrating multi-source information, the training process of the LSTM network model is optimized to make the model better adapt to data drift. S3, deploy the trained LSTM network model to the edge device; S4, the long short-term memory network model in the edge device continuously analyzes the newly collected data in real time and responds quickly; In S2, the long short-term memory network model is used to predict equipment failure based on the operation data, and the long short-term memory network model is trained on the cloud, including the following steps: S2.1, preprocessing the operation data; S2.2, extracting features from the operating data that can reflect the status of the equipment; S2.

3. Determine the specific structure of the LSTM network, clarify the length of the time series input to the LSTM network and the output format, and set the hyperparameters; S2.4, the preprocessed and feature-extracted running data set is divided into a training set, a validation set, and a test set; S2.

5. Use computing resources on the cloud platform to train the long short-term memory network model, define the loss function, optimize the loss function to address the data drift problem, and optimize the loss function by fusing multi-source information to make the model better adapt to data drift. During the training process, optimize the model algorithm, improve the efficiency of model training through distributed training technology, and protect the security of data transmission through TLS / SSL encryption protocol; S2.

6. Use the training set data to optimize model parameters and use the validation set to adjust hyperparameters. In S2.5, the loss function is: ； in, Indicates the degree of difference between the model prediction results and the actual results; Indicates the number of data samples contained in the data set involved in training; Indicates the number of types of faults; Indicates The sample in The true label value of each category; The model predicts The samples belong to The probability value of each category; An index indicating the number of categories; Indicates the sample index; To address the data drift problem, the loss function is optimized: ； in, Indicates the degree of difference between the model prediction results and the actual results after addressing the data drift problem; Indicates the parameters that adjust the impact of data drift; By fusing multi-source information to optimize the loss function, the model can better adapt to data drift: ； in, Indicates the degree of difference between the model prediction results and the actual results after integrating multi-source information; Indicates the number of additional information sources; Indicates the measurement The weight parameter of the importance of the additional information source in the entire loss function; Indicates of the additional information sources The data value of each sample; Represents the model for of the additional information sources The predicted value of samples; In S2.5, the efficiency of model training is improved by using distributed training technology, including the following steps: S2.

51. According to computing resources, the training data set is evenly divided into multiple subsets, and each computing node processes one subset; S2.

52. Deploy a complete copy of the LSTM network model on each computing node; S2.53, each node independently performs forward propagation on its assigned data subset and calculates the corresponding loss value; S2.54, based on its own loss value, each node independently calculates the gradient; S2.

55. Use the hierarchical All-Reduce algorithm to aggregate the gradients calculated by all nodes. Since the more nodes there are, the greater the communication overhead and the longer the gradient aggregation time is, the communication complexity increases exponentially. The hierarchical All-Reduce algorithm is optimized. S2.

56. Based on the aggregated gradient information, the synchronous update algorithm is used to update the model parameters. The synchronous update algorithm is optimized to solve the problem that the parameter update oscillates back and forth around the minimum value, which slows down the convergence speed. In S2.55, the hierarchical All-Reduce algorithm is optimized, including the following steps: S2.551, divide all nodes into several groups, and the number of nodes in each group is close; S2.552, perform All-Reduce operation within each group to obtain the local aggregation result of each group; S2.553, perform All-Reduce operation on the local aggregation results of all groups again to obtain the global aggregation result; S2.554, broadcast the global aggregation results back to all nodes; In S2.56, the synchronization update algorithm is: ； in, represents the updated model parameters; Represents the learning rate, which controls the step size of parameter update at each iteration; Represents the model parameters before updating; Indicates the batch size; Indicates The loss function on samples Parameters The gradient of To solve the problem that parameter updates oscillate back and forth near the minimum value, which slows down the convergence speed, the synchronous update algorithm is optimized: ； in, Represents the model parameters after optimization and update; represents the momentum term coefficient; Represents the cumulative gradient at the last iteration; In S3, the trained long short-term memory network model is deployed to the edge device in the production line, including the following steps: S3.1, convert the model into a format suitable for edge deployment; S3.2, deploy the model manager together with the trained LSTM network model to the edge device; S3.3, when the cloud completes the training of the new version of the model and confirms that the performance of the new version of the model is better than the existing model, the new model is sent to the model manager on the edge device through a secure channel; S3.4, the model manager receives the new model file and verifies its integrity and correctness; S3.5, the model manager allows to unload old models and load new models; S3.6, after the new model is loaded, it immediately starts to analyze the newly collected data, and the monitoring system continuously collects performance data and resource usage during the model execution; S3.

7. All communication interfaces adopt security measures.

2. The AI-based intelligent edge computing method according to claim 1 is characterized in that: In S4, the long short-term memory network model in the edge device continuously analyzes the newly collected data in real time and responds quickly, including the following steps: S4.1, the long short-term memory network model in the edge device continuously analyzes the newly collected data in real time; S4.

2. Make a quick response based on the analysis results; S4.

3. Define common interfaces, protocols and data formats; S4.

4. Implement asynchronous message processing through message queues and event-driven architecture.

3. An intelligent edge computing system based on AI artificial intelligence, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: The processor executes a computer program to implement the AI ​​artificial intelligence-based intelligent edge computing method as described in any one of claims 1-2.

4. A terminal, characterized in that: It includes a processor, a memory, and a computer control unit stored in the memory and executable on the processor, and is used to execute the AI ​​artificial intelligence-based intelligent edge computing method as described in any one of claims 1-2.

Images