Device connection management method and system based on Internet of Things

Abstract

The present application provides a device connection management method and system based on the Internet of Things. After obtaining the device monitoring data stream, a load overload prediction neural network is obtained to perform load overload reasoning, and a first load overload estimated confidence and a second load overload estimated confidence of the IoT device to be analyzed are obtained. The first load overload estimated confidence and the second load overload estimated confidence are combined for collaborative reasoning to obtain the target load overload estimated confidence of the IoT device to be analyzed, so that when the IoT device to be analyzed performs load overload reasoning, it combines multi-dimensional data for collaborative learning, and simultaneously completes the overload reasoning of a single IoT device based on the federated model, and effectively protects the data privacy of IoT devices based on the federated model. In other words, the present application is not only conducive to achieving effective load balancing and effective device connection management, but also ensures data privacy.

Description

Technical Field

[0001] The present application relates to, but is not limited to, the field of data processing technology, and in particular to a device connection management method and system based on the Internet of Things. Background Art

[0002] With the rapid development of IoT technology, more and more devices are connected to the network, forming a huge IoT ecosystem. While providing convenience, these devices also face the risk of load overload. Load overload may not only lead to device performance degradation, but may even cause device failure, which will seriously affect the stability and reliability of the entire IoT system. Therefore, how to effectively manage and predict the load overload of devices has become an urgent problem to be solved in the field of IoT. Traditional device connection management methods usually rely on manual monitoring and manual adjustment, which is not only inefficient but also difficult to cope with the complex management needs of large-scale IoT devices. In addition, traditional methods often ignore the mutual influence between devices and changes in external environmental factors, resulting in inaccurate prediction results. In recent years, with the development of machine learning and deep learning technologies, using neural networks to predict device load overload has become a new solution. However, most of the existing prediction methods based on neural networks only rely on the data of the device itself, and the prediction results are not comprehensive enough. At the same time, these methods are also insufficient in protecting data privacy, and it is difficult to meet the high requirements of IoT devices for data security. Summary of the invention

[0003] In view of this, the embodiment of the present application at least provides a device connection management method and system based on the Internet of Things. The technical solution of the embodiment of the present application is implemented as follows:

[0004] On the one hand, an embodiment of the present application provides a device connection management method based on the Internet of Things, including: obtaining a device monitoring data stream, wherein the device monitoring data stream is composed of performance indicator data generated by different operating cycles of the Internet of Things device to be analyzed during operation; obtaining a load overload prediction neural network corresponding to the Internet of Things device to be analyzed in a distributed collaborative network framework, performing load overload inference on the device monitoring data stream, and obtaining a first load overload estimated confidence of the Internet of Things device to be analyzed; the distributed collaborative network framework includes a load overload prediction neural network corresponding to a linkage device of the Internet of Things device to be analyzed; obtaining a second load overload estimated confidence for the Internet of Things device to be analyzed in the linkage device, wherein the second load overload estimated confidence is the linkage monitoring data stream obtained through the linkage device, and is obtained by performing load overload inference on the linkage monitoring data stream based on the load overload prediction neural network corresponding to the linkage device; performing collaborative inference on the first load overload estimated confidence of the Internet of Things device to be analyzed and the second load overload estimated confidence, to obtain a target load overload estimated confidence of the Internet of Things device to be analyzed.

[0005] In a second aspect, the present application provides a computer system, comprising a memory and a processor, wherein the memory stores a computer program executable on the processor, and the processor implements the steps in the above-described method when executing the program.

[0006] The beneficial effects of the present application include at least: the device connection management method and system based on the Internet of Things provided by the present application, after obtaining the device monitoring data stream, obtains the load overload prediction neural network in the IoT device to be analyzed to perform load overload reasoning on the device monitoring data stream, obtains the first load overload estimated confidence of the IoT device to be analyzed, and obtains the second load overload estimated confidence for the IoT device to be analyzed in the linkage device of the IoT device to be analyzed, after the IoT device to be analyzed obtains the first load overload estimated confidence and the second load overload estimated confidence, it can be combined with the first load overload estimated confidence and the second load overload estimated confidence for collaborative reasoning to obtain the target load overload estimated confidence of the IoT device to be analyzed, so that the IoT device to be analyzed can be combined with multi-dimensional data for collaborative learning when performing load overload reasoning, and at the same time completes the overload reasoning of a single IoT device based on the federated model, and effectively protects the data privacy of the IoT device based on the federated model. In other words, the present application is not only conducive to achieving effective load balancing and effective device connection management, but also ensures data privacy. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] The drawings herein are incorporated into the specification and constitute a part of the specification. These drawings illustrate embodiments consistent with the present application and are used together with the specification to illustrate the technical solution of the present application.

[0008] Figure 1 A schematic diagram of the implementation flow of a device connection management method based on the Internet of Things provided in an embodiment of the present application.

[0009] Figure 2 A hardware entity schematic diagram of a computer system provided in an embodiment of the present application. DETAILED DESCRIPTION

[0010] In order to make the purpose, technical solutions and advantages of the present application clearer, the technical solutions of the present application are further elaborated in detail below in conjunction with the drawings and embodiments. The described embodiments should not be regarded as limiting the present application. All other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present application.

[0011] The embodiment of the present application provides a device connection management method based on the Internet of Things, which can be executed by a processor of a computer system. The computer system can refer to a device with data processing capabilities such as a server, a laptop, a tablet computer, and a desktop computer.

[0012] Figure 1 A schematic diagram of an implementation flow of a device connection management method based on the Internet of Things provided in an embodiment of the present application is shown in FIG. Figure 1 As shown, the method includes:

[0013] Step S10: Acquire a device monitoring data stream, where the device monitoring data stream is composed of performance indicator data generated by different operation cycles of the IoT device to be analyzed during operation.

[0014] In the embodiment of the present application, step S10 obtains the real-time performance index data of the IoT device to be analyzed during operation, and these data together constitute the device monitoring data stream. In the process of device connection management, the computer system will first connect to the IoT device to be analyzed, such as a production line server in a smart factory or a smart home gateway in a smart home. These devices will generate a large amount of performance index data during normal operation, which reflects the operating status and performance of the device. Specifically, the computer system continuously collects various performance index data from the IoT device to be analyzed. For example, for the production line server, it can collect data on the use of hardware resources such as CPU usage, memory occupancy, disk I / O read and write speed, as well as network performance-related data such as network bandwidth usage, network delay, packet loss rate, and connection establishment time. These data are generated in real time and are continuously updated with the operation cycle of the device. Taking the production line server as an example, the CPU usage reflects the server's ability to process tasks, the memory occupancy shows the current memory occupancy, and the disk I / O read and write speed determines the efficiency of data reading and writing. Indicators such as network bandwidth usage, network delay, packet loss rate, and connection establishment time are directly related to the communication efficiency and stability of the server and other devices or systems.

[0015] The way to collect this data is usually through the sensors and monitoring modules built into the IoT devices. These sensors and monitoring modules can monitor the operating status of the equipment in real time and transmit the collected data to the computer system through the IoT network. After receiving this data, the computer system will perform necessary preprocessing and sorting to form a complete equipment monitoring data stream. The equipment monitoring data stream is the basis for subsequent analysis and prediction, and it provides comprehensive information about the performance and status of the equipment. By analyzing these data streams, the computer system can understand the real-time operating status of the equipment, predict possible load overload conditions, and take corresponding management measures to optimize the operation and connection management of the equipment.

[0016] Step S20: Obtain a load overload prediction neural network corresponding to the IoT device to be analyzed in the distributed collaborative network framework, perform load overload inference on the device monitoring data stream, and obtain a first load overload estimation confidence of the IoT device to be analyzed; the distributed collaborative network framework includes a load overload prediction neural network corresponding to the linkage device of the IoT device to be analyzed.

[0017] In an embodiment of the present application, step S20 uses the load overload prediction neural network in the distributed collaborative network framework (i.e., the federated model architecture) to infer the device monitoring data stream of the IoT device to be analyzed, thereby obtaining the first load overload estimation confidence of the device. It should be clarified that the distributed collaborative network framework is a federated model architecture that allows multiple IoT devices to jointly train and optimize machine learning models while maintaining data localization. In this framework, each IoT device (including the IoT device to be analyzed and its linkage device) has its own corresponding load overload prediction neural network. These neural networks may be based on different machine learning algorithms, such as recurrent neural networks (RNN) or their variants, such as long short-term memory networks (LSTM) or gated recurrent units (GRU), which perform well in processing time series data and are suitable for analyzing performance indicator data of IoT devices.

[0018] In step S20, the computer system first obtains the device monitoring data stream from the IoT device to be analyzed. These data streams are composed of performance indicator data generated by different operation cycles of the device during operation, such as CPU usage, memory usage, disk I / O read and write speed, network bandwidth usage, etc. After preprocessing, these data are input into the load overload prediction neural network corresponding to the IoT device to be analyzed. The neural network has been trained with historical data and machine learning algorithms, and can identify patterns and features related to load overload in performance indicator data. During the reasoning process, the neural network will perform feature extraction and pattern recognition on the input device monitoring data stream, and then calculate the first load overload estimation confidence of the IoT device to be analyzed based on these features and patterns. This confidence represents the prediction probability of the neural network for the device to be overloaded. It is worth noting that the distributed collaborative network framework also includes a load overload prediction neural network corresponding to the linkage device of the IoT device to be analyzed. These linkage devices may be upstream and downstream IoT devices, or they may be devices for monitoring external events or user behaviors. Their role is to provide additional information and data related to the IoT device to be analyzed, helping to more accurately predict the load overload of the device. For example, if the IoT device to be analyzed is a server on a production line, its linkage device may include upstream sensor devices (such as temperature sensors, pressure sensors, etc.) and downstream data processing devices. The data generated by these devices can provide information about the operating status of the production line, the supply of raw materials, product quality, etc. By inputting this information into the load overload prediction neural network corresponding to the linkage device, a second load overload estimation confidence for the IoT device to be analyzed can be obtained. This confidence can be collaboratively reasoned with the first load overload estimation confidence to obtain a more accurate prediction result.

[0019] Step S20 uses the load overload prediction neural network in the distributed collaborative network framework to infer the device monitoring data stream of the IoT device to be analyzed, and obtains the first load overload estimation confidence of the device. This confidence provides an important reference for subsequent device performance analysis and load overload management. In practical applications, the computer system can take corresponding measures based on this confidence to optimize the operation and connection management of the device, thereby improving the stability and efficiency of the entire IoT system.

[0020] Step S30: Obtain a second load overload estimation confidence for the IoT device to be analyzed in the linkage device, where the second load overload estimation confidence is obtained by performing load overload inference on the linkage monitoring data stream obtained through the linkage device and by obtaining a load overload prediction neural network corresponding to the linkage device.

[0021] In the embodiment of the present application, step S30 obtains the second load overload estimation confidence for the IoT device to be analyzed from the linkage device. As mentioned above, the linkage device refers to other IoT devices or systems associated with the IoT device to be analyzed, which may include upstream and downstream IoT devices, devices for monitoring external events or user behaviors, etc. These linkage devices play an important role in the device connection management network. They can provide additional information and data related to the IoT device to be analyzed, thereby helping to more comprehensively evaluate the load status of the device. In step S30, the computer system first obtains the linkage monitoring data stream from the linkage device. These data streams are composed of performance indicator data generated by the linkage device during operation, and they may be directly or indirectly related to the IoT device to be analyzed. For example, if the IoT device to be analyzed is a production line server of a smart factory, then its upstream raw material supply equipment or downstream product testing equipment may be a linkage device. These linkage devices will generate data about the raw material supply status, product testing results, etc., which will be collected and transmitted to the computer system to form a linkage monitoring data stream.

[0022] Next, the computer system uses the load overload prediction neural network corresponding to the linkage device to infer the load overload on the linkage monitoring data stream. These neural networks are designed specifically for linkage devices and may be based on different machine learning algorithms, such as recurrent neural networks (RNN), long short-term memory networks (LSTM), or convolutional neural networks (CNN). These neural networks have been trained with historical data and machine learning algorithms to identify patterns and features related to load overload in the linkage monitoring data stream.

[0023] During the inference process, the neural network will extract features and recognize patterns from the input linkage monitoring data stream. They will analyze various indicators in the data stream, such as the stability of raw material supply, the qualified rate of product testing, etc., and calculate the linkage device's confidence in the estimated second load overload of the IoT device to be analyzed based on these indicators. This confidence represents the linkage device's predicted probability of an impending load overload of the IoT device to be analyzed based on its own data and machine learning model.

[0024] For example, if the linkage device is a sensor network that monitors external weather conditions, they will generate weather data such as temperature, humidity, wind speed, etc. These weather data may be related to the load condition of the IoT device to be analyzed. For example, high temperature weather may cause poor heat dissipation of the device and thus cause load overload. The computer system inputs these weather data into the load overload prediction neural network corresponding to the linkage device, and obtains the second load overload estimation confidence through reasoning. This confidence will serve as an additional source of information and be collaboratively reasoned with the load overload estimation confidence of the IoT device to be analyzed, so as to obtain a more accurate prediction result.

[0025] Step S30 obtains the linkage monitoring data stream from the linkage device and uses the load overload prediction neural network corresponding to the linkage device for reasoning to obtain the second load overload estimation confidence for the IoT device to be analyzed. This confidence provides an important reference for subsequent device performance analysis and load overload management, and helps to achieve more comprehensive device connection management.

[0026] Step S40: performing collaborative reasoning through the first load overload estimation confidence level and the second load overload estimation confidence level of the IoT device to be analyzed, to obtain the target load overload estimation confidence level of the IoT device to be analyzed.

[0027] In the embodiment of the present application, step S40 will perform collaborative reasoning through the first load overload estimation confidence of the IoT device to be analyzed and the second load overload estimation confidence of the linkage device, so as to calculate the target load overload estimation confidence of the IoT device to be analyzed. The first load overload estimation confidence is based on the monitoring data stream of the IoT device to be analyzed itself, and is inferred through its corresponding load overload prediction neural network (such as RNN). This confidence reflects the impact of the device's own operating status on the possibility of its load overload. The second load overload estimation confidence is obtained from the linkage device, which may be upstream and downstream IoT devices or devices for monitoring external events and user behavior. They perform reasoning through their respective monitoring data streams and corresponding load overload prediction neural networks, providing indirect load overload prediction information for the IoT device to be analyzed. In step S40, the computer system takes these two estimated confidences as input and integrates them through a collaborative reasoning algorithm. This collaborative reasoning algorithm may be based on a variety of strategies, such as weighted average, Bayesian fusion, etc. In a specific implementation, the computer system first assigns a weight to each estimated confidence level, which reflects the importance of the confidence level in the final prediction. The weight assignment may be based on factors such as the accuracy of historical data, the relevance of the linkage device to the device to be analyzed, etc.

[0028] Next, the computer system combines the two estimated confidences based on these weights. For example, if a weighted average strategy is used, the computer system multiplies each estimated confidence by its corresponding weight, and then adds the results to obtain the target load overload estimated confidence. This confidence combines the status of the IoT device to be analyzed and the information provided by the linkage device, so it is more accurate and reliable than the estimated confidence from a single source.

[0029] For example, in a smart factory, the IoT device to be analyzed is a server on a production line, and its first load overload estimation confidence is based on the server's own CPU usage, memory usage and other indicators. The linkage device may include upstream raw material supply equipment and downstream product inspection equipment. These devices provide the server with a second load overload estimation confidence through their own monitoring data streams and corresponding load overload prediction neural networks. For example, if there is a supply delay in the upstream equipment, it may cause the server to process more tasks to be processed, thereby increasing the risk of load overload; and abnormal detection results of downstream equipment may also affect the load condition of the server.

[0030] Through the collaborative reasoning of step S40, the computer system merges the first load overload estimation confidence of the server and the second load overload estimation confidence of the linkage device to obtain a more comprehensive target load overload estimation confidence. This confidence not only takes into account the operating status of the server itself, but also incorporates the relevant information provided by the linkage device, so it can more accurately predict the possibility of load overload on the server. Based on this prediction result, factory managers can take corresponding measures in a timely manner, such as adjusting production plans, optimizing resource allocation, etc., to ensure the stable operation of the production line.

[0031] As an implementation mode, step S10, obtaining a device monitoring data stream, includes:

[0032] Step S11: Obtain resource occupancy information data stream of the IoT device to be analyzed and device status information data stream for the IoT device to be analyzed, and obtain a load overload mark of the IoT device to be analyzed, wherein the load overload mark includes a mark generation time, and the mark generation time represents the time when the load overload occurs in the IoT device to be analyzed.

[0033] In the embodiment of the present application, step S11 mainly performs the following three subtasks: collecting resource occupancy information data streams of the IoT devices to be analyzed, collecting device status information data streams, and obtaining the load overload mark of the device. First, the computer system starts to collect resource occupancy information data streams of the IoT devices to be analyzed. This information directly reflects the consumption of various resources by the device during operation. For example, on the production line server of a smart factory, resource occupancy information may include CPU usage, memory usage, etc. CPU usage reflects the server's ability to process tasks, and memory usage shows the current memory usage. These data usually exist in the form of time series, and the computer system obtains these data streams from the device in real time through IoT protocols (such as MQTT, CoAP, etc.).

[0034] Next, the computer system also collects the device's status information data stream. This information describes the current operating status and health of the device. Taking the production line server as an example, the device status information may include the server's health status, temperature, fan speed, etc. The health status may be a comprehensive indicator, indicating whether the server is running normally; the temperature reflects the server's heat dissipation. Too high a temperature may cause performance degradation or even equipment damage; the fan speed is an important indicator of the cooling system, affecting the device's heat dissipation efficiency. These status information also exist in the form of time series, and together with the resource occupancy information, they constitute a complete operating portrait of the device. In addition to the real-time data stream, the computer system can also obtain the load overload tags of the IoT device to be analyzed from the historical records. These tags are generated when the device actually has a load overload event, and are used to identify the time and nature of the overload event. For example, if the production line server responds slowly at a certain moment due to excessive CPU occupancy, the computer system will generate a load overload tag. This tag usually contains two key information: the time when the tag is generated (such as a timestamp) and the tag value (such as Lt, Lt is 0 for no overload, and Lt is 1 for overload). The marker generation time represents the specific time when the overload event occurs, while the marker value clearly indicates whether the event occurs or not.

[0035] By executing step S11, the computer system can comprehensively collect the resource usage information, device status information and load overload mark of the IoT device to be analyzed. These data provide rich training samples and verification basis for the subsequent load overload prediction model, which helps the model to more accurately identify the overload risk of the device and take preventive measures in advance.

[0036] Step S12: According to the load overload mark and the mark generation time contained in the load overload mark, the load overload mark is matched with the resource occupation information data stream and the device status information data stream respectively.

[0037] In the embodiment of the present application, step S12 accurately matches the load overload mark with the resource occupancy information data stream and the device status information data stream. The purpose of this step is to associate the load overload state of the device at a specific time point with its operating data at that time, so as to provide a strong basis for subsequent analysis and prediction. First, the computer system obtains the load overload mark of the IoT device to be analyzed. These marks include the time when the mark is generated, that is, the specific time point when the device is overloaded. These time points are usually expressed in the form of timestamps, accurate to the second or millisecond level, to ensure that they can accurately correspond to the device operation data.

[0038] Next, the computer system matches these overload marks with the resource occupancy information data stream and the device status information data stream. The resource occupancy information data stream and the device status information data stream are collected in real time and contain the operating status and performance data of the device at different time points. These data streams usually exist in the form of time series, and each time point corresponds to a set of data, including key indicators such as CPU usage, memory occupancy, device temperature, and health status. During the matching process, the computer system first finds the data at the corresponding time point in the resource occupancy information data stream and the device status information data stream according to the time when the mark in the overload mark is generated. This process is usually achieved by comparing timestamps to ensure that the operating status data of the device when the overload occurs can be accurately located. Once the data at the corresponding time point is found, the computer system will associate this data with the overload mark. Specifically, the value of the overload mark (such as 0 for no overload and 1 for overload) is combined with the resource occupancy information and device status information at the corresponding time point to form a complete data record. This data record not only contains the operating status data of the device at a specific time point, but also clarifies whether a load overload event has occurred at that time point.

[0039] For example, suppose there is a production line server in a smart factory. At a certain time point (such as 2023-10-23 14:30:05), a load overload event occurs, and the automatic monitoring system generates a load overload mark (such as Lt=1, and the mark is generated at 2023-10-23 14:30:05). At the same time, the computer system continuously collects the resource occupancy information data stream and the device status information data stream of the server. In step S12, the computer system finds the data record of the corresponding time point (2023-10-23 14:30:05) in the data stream, including indicators such as CPU usage, memory usage, and device temperature at that time. Then, the computer system associates these indicator data with the load overload mark (Lt=1) to form a complete data record. This data record will be used for subsequent analysis and prediction to help managers understand the specific operating conditions of the equipment when it is overloaded, and take corresponding measures to optimize equipment performance or prevent similar events from happening again.

[0040] Step S13: Generate a device monitoring data stream based on the resource occupancy information data stream after matching the load overload mark and the device status information data stream after matching the load overload mark.

[0041] In the embodiment of the present application, step S13 integrates the resource occupancy information data stream and the device status information data stream after matching the load overload mark into a device monitoring data stream. The purpose of this step is to construct a data stream containing comprehensive operating status information of the device when a specific load overload event occurs, for subsequent analysis and prediction. In step S13, the computer system first reviews the result of step S12, that is, the resource occupancy information data stream and the device status information data stream that have been matched with the load overload mark. These data streams contain key operating parameters of the device when a load overload event occurs, such as CPU usage, memory usage, device temperature, health status, etc. Each data point corresponds to a specific timestamp to ensure the accuracy and timing of the data. Next, the computer system integrates these matched data streams. The integration process mainly aligns the resource occupancy information data stream and the device status information data stream according to the timestamp to ensure that they describe the device status at the same time. If there is a missing or inconsistent data at certain times, the computer system can use interpolation, smoothing or other data processing techniques to fill or correct these data. After the integration is completed, the computer system generates a complete device monitoring data stream. This data stream not only contains the resource usage and device status information of the device when the overload event occurs, but also retains the timestamp information, making the data stream time-series. This time-series is crucial for subsequent overload prediction and analysis, because it can reflect the changes in the operating status of the device at different time points and the potential correlation between these changes and overload events.

[0042] As an implementation mode, the resource occupancy information data stream is composed of the resource occupancy information data in the time interval [t0, tn], and the device status information data stream is composed of the device status information data in the time interval [t0, tn], tn is a positive integer greater than 0, and if the mark generation time of the load overload mark is x, x∈[t0, tn]. Based on this, step S12, according to the load overload mark and the mark generation time contained in the load overload mark, matches the load overload mark with the resource occupancy information data stream and the device status information data stream respectively, which may specifically include:

[0043] Step S121: Based on the mark generation time x of the load overload mark, the load overload mark is matched with the resource occupancy information data flow in the time interval [t0, tn] as the resource occupancy information data flow associated with the load overload mark.

[0044] In the embodiment of the present application, step S121 accurately matches the load overload mark with the resource occupancy information data stream. The purpose of this step is to find the device resource occupancy corresponding to the time when the load overload event occurs, so as to provide accurate data support for subsequent analysis and prediction. First of all, the resource occupancy information data stream refers to the data stream about resource occupancy generated by the Internet of Things device within a certain time interval [t0, tn]. These data include key indicators such as CPU usage and memory occupancy. The load overload mark is generated by the monitoring system when the device is overloaded. It contains the time x when the mark is generated, which indicates the specific time point when the device actually experiences a load overload.

[0045] In step S121, the computer system first obtains the mark generation time x of the load overload mark. Then, it traverses the resource occupancy information data stream and searches for the data point corresponding to the timestamp closest to or equal to the mark generation time x. Since the data stream is recorded continuously, it can be considered that the data point corresponding to this timestamp reflects the resource occupancy of the device when the load overload event occurs.

[0046] For example, in a smart factory, there is a production line server responsible for processing production data. This server will generate resource occupancy information data streams during operation, including data such as CPU usage and memory usage. At the same time, the monitoring system will monitor the operating status of the server in real time. Once it detects that the server is overloaded, it will generate a load overload mark and record the time x when the mark is generated. Suppose on a certain day, the monitoring system detects that the server is overloaded at 14:30:05 and generates a load overload mark. The time x when the mark is generated is 14:30:05. At this time, the computer system starts to execute step S121. It will traverse the resource occupancy information data stream of the server in the past period of time (such as the past hour) and find the data point closest to the timestamp of 14:30:05. Suppose a data point with a timestamp of 14:30:04 is found, which shows that the CPU usage is 95% and the memory usage is 80%. This data point is considered to be a resource occupancy information data stream associated with the load overload event.

[0047] Through the precise matching of step S121, the computer system can accurately find the resource occupancy corresponding to the time when the load overload event occurs. This provides strong data support for the subsequent analysis of the performance bottleneck of the device when the load is overloaded, the optimization of resource allocation, etc. At the same time, since the matching process is based on the timestamp, the accuracy and timing of the data can be ensured, providing a reliable basis for subsequent analysis and prediction.

[0048] Step S122: Based on the mark generation time x of the load overload mark, the load overload mark is matched with the device status information data stream in the time interval [t0, tn], and determined as the device status information data stream after matching the load overload mark.

[0049] In the embodiment of the present application, the goal of step S122 is to find the device status information corresponding to the time when the load overload event occurs, so as to provide detailed device context for subsequent analysis. As before, the device status information data stream is a data stream generated by the IoT device about its health status, temperature, fault and other key indicators within a certain time interval [t0, tn]. The load overload mark contains the mark generation time x, which is the specific time point when the load overload actually occurs in the device. In step S122, the computer system first obtains the mark generation time x of the load overload mark. Then, it traverses the device status information data stream and finds the data point corresponding to the timestamp closest to or equal to the mark generation time x. Since the data stream is recorded continuously, it can be considered that the data point corresponding to this timestamp reflects the status information of the device when the load overload event occurs. For example, in a smart factory, there is an important production equipment that continuously generates a device status information data stream, including key indicators such as temperature, vibration frequency, and health status. At the same time, the monitoring system monitors the operating status of the equipment in real time, and generates a load overload mark when the equipment is overloaded, and the mark generation time is x. Assume that the monitoring system detects that the device is overloaded at 14:30:05 and generates a corresponding overload mark. At this point, the computer system starts to execute step S122. It will search for the data point closest to the timestamp of 14:30:05 from the device status information data stream. Assume that a data point with a timestamp of 14:30:04 is found, which shows that the temperature of the device is 70 degrees Celsius (the normal range is 50-60 degrees Celsius), the vibration frequency is abnormally increased, and the health status indicator shows a warning. This data point is considered to be the device status information data stream associated with the load overload event. It provides detailed status information of the device when the load overload event occurs, including key indicators such as abnormal temperature increase and abnormal vibration frequency. This information is of great significance for subsequent analysis of the cause of the device overload, optimization of device configuration, and prevention of similar events from happening again.

[0050] As an implementation mode, step S13, based on the resource occupancy information data stream after matching the load overload mark and the device status information data stream after matching the load overload mark, generates the device monitoring data stream of the IoT device to be analyzed, which may specifically include:

[0051] Step S131: performing data stream splitting operations on the resource occupancy information data stream after matching the load overload mark, and performing data stream splitting operations on the device status information data stream after matching the load overload mark, respectively, to obtain at least one resource occupancy information data stream sample, and at least one device status information data stream sample.

[0052] In the embodiment of the present application, step S131 involves splitting the resource occupancy information data stream and the device status information data stream that have been matched with the load overload mark to generate data samples for subsequent analysis. In step S131, the computer system first locates the resource occupancy information data stream and the device status information data stream that match the load overload mark. These data streams contain real-time operating status data of the device when a load overload event occurs, which is an important basis for analyzing device performance and load overload causes.

[0053] Next, the computer system splits these matched data streams. The purpose of splitting is to cut the continuous data stream into several independent data samples for subsequent feature extraction, model training and predictive analysis. The splitting operation is usually based on a certain time window or data change rate to ensure that each data sample contains enough information to reflect the operating status of the device.

[0054] Taking the resource usage information data stream as an example, assume that the data stream contains key indicators such as CPU usage and memory usage. The system can take data within a certain time range forward and backward as a sample based on the time when the load overload mark is generated. For example, taking the time when the mark is generated as the center, take 1 minute of data forward and 1 minute of data backward to form a 2-minute data sample. This sample contains the resource usage of the device in a period of time before and after the load overload event occurs.

[0055] Similarly, the data stream of device status information can also be split similarly. Assuming that the data stream contains indicators such as the temperature and vibration frequency of the device, the system can also select data within a certain time range before and after the load overload mark generation time as a sample.

[0056] Through the data stream splitting operation in step S131, the computer system can generate multiple resource occupancy information data stream samples and equipment status information data stream samples. These samples not only contain the key operating status data of the equipment when the overload event occurs, but also have a certain time span, which can reflect the changes in the operating status of the equipment before and after the event. These samples will serve as basic data for subsequent analysis and prediction, helping managers to gain a deep understanding of the performance of the equipment and the causes of overload, so as to take corresponding optimization measures.

[0057] For example, suppose a production line server in a smart factory experiences a load overload event at 3 p.m. The computer system obtains the resource usage information data stream and device status information data stream of the server before and after the event through the monitoring system, and performs a split operation. The system generates multiple data samples, one of which contains the CPU usage rate, memory usage rate, temperature, and vibration frequency data of the server between 2:58 p.m. and 3:02 p.m. This sample reflects the key operating status information of the server before and after the load overload event, providing strong data support for subsequent analysis and prediction.

[0058] Step S132: Determine any one of the at least one resource occupancy information data stream sample and any one of the at least one device status information data stream sample as the device monitoring data stream of the IoT device to be analyzed.

[0059] In the embodiment of the present application, step S132 involves selecting specific data samples for subsequent analysis from multiple resource occupancy information data stream samples and device status information data stream samples to form a device monitoring data stream of the IoT device to be analyzed. In step S131, the computer system has split the resource occupancy information data stream and the device status information data stream after matching the load overload mark, and generated multiple data samples. These samples respectively contain the operating status information of the device in different time periods before and after the load overload event occurs.

[0060] In step S132, the computer system selects the most representative or relevant information from these samples as the equipment monitoring data stream. The basis for the selection may include factors such as the time span of the samples, the volatility of the data, and the correlation with the load overload event. Specifically, the computer system may tend to select samples that contain the time when the load overload event occurs and the data changes significantly, because these samples are more likely to contain the key information that causes the load overload.

[0061] For example, assume that after step S131, the system generates 5 resource occupancy information data stream samples and 5 device status information data stream samples. These samples cover a period of time before and after the load overload event occurs, and each sample contains data on key indicators such as CPU usage, memory usage, device temperature, and vibration frequency.

[0062] In step S132, the computer system analyzes the characteristics of these samples and selects them based on certain strategies. For example, the computer system may first filter out samples containing the time when the overload event occurs, and then further compare the volatility of the data in these samples. If the data of a sample changes significantly before and after the overload event occurs, especially if key indicators such as CPU usage or memory usage increase sharply, then this sample may be selected as the device monitoring data stream.

[0063] In addition, the computer system may also consider the time span of the sample. If the time span of a sample is long and can cover a long time range before and after the load overload event occurs, then this sample may also be preferred because it provides more comprehensive equipment operation status information.

[0064] Finally, the computer system selects the most suitable resource occupancy information data stream samples and equipment status information data stream samples from multiple samples according to the above strategy, and combines them into equipment monitoring data streams. This data stream will serve as the basic data for subsequent analysis and prediction, helping managers to gain a deeper understanding of equipment performance and load overload causes, so as to take corresponding optimization measures.

[0065] In this way, step S132 ensures the effectiveness and pertinence of the device monitoring data stream, providing strong data support for subsequent analysis and prediction.

[0066] As an implementation method, the load overload prediction neural network corresponding to the IoT device to be analyzed in the distributed collaborative network framework includes a resource occupancy inference network corresponding to the IoT device to be analyzed and a device state inference network corresponding to the IoT device to be analyzed, and the device monitoring data stream includes a resource occupancy information data stream sample and a device state information data stream sample. Based on this, step S20 obtains the load overload prediction neural network corresponding to the IoT device to be analyzed in the distributed collaborative network framework, performs load overload inference on the device monitoring data stream, and obtains the first load overload estimation confidence of the IoT device to be analyzed, which may specifically include:

[0067] Step S21: Obtain a resource occupancy inference network corresponding to the IoT device to be analyzed, perform load overload inference on the resource occupancy information data stream sample, and obtain an occupancy-type load overload estimation confidence of the IoT device to be analyzed.

[0068] In step S21, the computer system first calls the resource occupancy inference network corresponding to the IoT device to be analyzed. This inference network is a machine learning-based model, for example, constructed using a deep learning architecture such as a recurrent neural network (RNN), a long short-term memory network (LSTM), or a gated recurrent unit (GRU). The network has been trained with a large amount of historical data, and has learned to extract key features from resource occupancy information and identify potential associations between these features and load overload. In step S21, the computer system inputs a sample of the resource occupancy information data stream into the resource occupancy inference network. These samples typically contain a series of resource occupancy data corresponding to timestamps, such as CPU usage, memory usage, disk I / O read and write speeds, etc. Each sample reflects the resource occupancy of the device in a specific time period.

[0069] After receiving the resource occupancy information data stream sample, the resource occupancy inference network will perform a series of inference calculations. This process includes data preprocessing, feature extraction, pattern recognition and other steps. The network will encode the input data stream sample, extract the key feature vectors, and then use the learned model parameters to transform and calculate these feature vectors, and finally output an occupancy type load overload estimation confidence. This estimation confidence is a value between 0 and 1, which indicates the probability of the device being overloaded under the current resource occupancy state. The higher the value, the greater the risk of load overload.

[0070] For example, suppose there is a production line server in a smart factory, which generates a large amount of resource usage data during operation. During a certain period of time, the server's CPU usage rate continued to remain high, and the memory usage rate was close to saturation. The computer system obtains a sample of the resource usage information data stream during this period of time and inputs it into the resource usage inference network. After calculation, the inference network outputs an occupancy-type load overload estimation confidence, such as 0.85. This value indicates that based on the current resource usage, there is an 85% probability that the server will be overloaded. Based on this estimated confidence, factory managers can take timely measures, such as adjusting production plans, optimizing resource allocation, etc., to avoid performance degradation or failure caused by server overload.

[0071] Step S22: Obtain a device state inference network corresponding to the IoT device to be analyzed, perform load overload inference on the device state information data stream sample, and obtain the state-type load overload estimation confidence of the IoT device to be analyzed.

[0072] In the embodiment of the present application, step S22 focuses on predicting the possibility of load overload of the equipment through equipment status information, thereby providing decision support for management personnel.

[0073] In step S22, the computer system uses the device state inference network corresponding to the IoT device to be analyzed to infer the device state information data stream sample. This inference network is also a machine learning-based model, which may have been trained with a large amount of historical data, learned to extract key features from device state information, and identify the potential relationship between these features and load overload.

[0074] The device status information data stream samples contain the operating status data of the device in a specific time period, such as temperature, vibration frequency, health status, etc. These data reflect the health status and potential problems of the device. In step S22, the computer system inputs these samples into the device status inference network.

[0075] After receiving the device status information data stream sample, the device status inference network will perform a series of inference calculations. This process includes data preprocessing, feature extraction, pattern recognition and other steps. The network will encode the input data stream sample, extract the key feature vectors, and then use the learned model parameters to transform and calculate these feature vectors, and finally output a state-based load overload estimation confidence. This estimation confidence is a value between 0 and 1, which indicates the probability of the device being overloaded in the current state. The higher the value, the greater the risk of the device being overloaded.

[0076] For example, suppose that in a smart factory, there is an important production equipment that continuously generates a data stream of equipment status information, including key indicators such as temperature, vibration frequency, and health status. During a certain period of time, the temperature of the equipment rises abnormally, the vibration frequency also fluctuates abnormally, and the health status indicator shows a warning. The computer system obtains a sample of the equipment status information data stream during this period of time and inputs it into the equipment status inference network. After calculation, the inference network outputs a state-based load overload estimation confidence, such as 0.9. This value indicates that according to the current equipment status information, there is a 90% probability that the equipment will be overloaded. Based on this estimated confidence, factory managers can respond quickly and take necessary measures to prevent equipment from being overloaded. This may include adjusting the operating parameters of the equipment, increasing heat dissipation measures, and performing maintenance and repairs. Through timely intervention, managers can reduce the risk of equipment failure and ensure the stable operation of the production line.

[0077] By implementing step S22, the computer system can evaluate the load overload risk based on the device status information, providing strong support for device connection management. This prediction method based on device status can be combined with the prediction method based on resource occupancy to form a more comprehensive load overload risk assessment system.

[0078] Step S23: taking the occupancy-type load overload estimation confidence and the state-type load overload estimation confidence as the first load overload estimation confidence of the IoT device to be analyzed.

[0079] In the embodiment of the present application, step S23 is a key step of integrating information on resource occupancy and device status to form a first load overload estimation confidence level of the IoT device to be analyzed.

[0080] In step S23, the computer system integrates the occupancy-type load overload estimation confidence and the state-type load overload estimation confidence obtained in steps S21 and S22, respectively. These two estimation confidences are derived based on the resource occupancy information and the device state information of the device, respectively, and they reflect the possibility of load overload of the device from different perspectives. The integration method may vary depending on the specific application scenario, but usually involves weighted or average processing of the two estimation confidences. Weighted processing can assign different weights to the two estimation confidences according to actual conditions to reflect their relative importance in the prediction. Averaging is a simplified method that assumes that the two estimation confidences have the same value in the prediction.

[0081] For example, suppose that in a smart factory, a production line server has a potential risk of overload. Through the resource occupancy reasoning network, the computer system obtains an occupancy-type load overload estimation confidence, indicating that the server has an 80% probability of overload under the current resource occupancy state. At the same time, through the device state reasoning network, the computer system obtains a state-type load overload estimation confidence, indicating that the server has a 75% probability of overload under the current device state. In step S23, the computer system integrates these two estimated confidences. Assuming that a simple arithmetic average method is used for integration, the first load overload estimation confidence is (80% + 75%) / 2 = 77.5%. This value combines information on resource occupancy and device status, providing factory managers with a more comprehensive equipment load overload risk assessment result.

[0082] Based on the first load overload estimate confidence, factory managers can make more accurate decisions. If the estimate confidence is high (such as exceeding a certain set threshold), managers can immediately take measures to prevent equipment from being overloaded, such as adjusting production plans, optimizing resource allocation, increasing heat dissipation measures, etc. If the estimate confidence is low, managers can continue to observe the operating status of the equipment and further analyze or adjust the prediction model as needed. Through the implementation of step S23, the computer system can effectively integrate information on resource occupancy and equipment status to form a more comprehensive and accurate equipment load overload estimate confidence. This provides strong decision-making support for equipment connection management and helps ensure stable operation and efficient management of equipment.

[0083] As an implementation mode, the linkage device of the Internet of Things device to be analyzed includes at least one of upstream and downstream Internet of Things devices and external demand monitoring devices, and the load overload prediction neural network corresponding to the linkage device includes at least one of upstream and downstream reasoning networks and external demand reasoning networks; the linkage monitoring data stream includes at least one of upstream and downstream device information data streams obtained based on upstream and downstream Internet of Things devices, and at least one of external demand information data streams obtained based on external demand monitoring devices; the second load overload estimation confidence obtained in the linkage device includes: at least one of the upstream and downstream load overload estimation confidences obtained by load overload reasoning through upstream and downstream device information data streams based on the upstream and downstream reasoning network, and at least one of the external load overload estimation confidences obtained by load overload reasoning through external demand information data streams based on the external demand reasoning network.

[0084] In the embodiment of the present application, a more comprehensive load overload prediction mechanism is introduced in consideration of the mutual influence between devices and external environmental factors. This mechanism not only considers the status and resource occupancy of the IoT device to be analyzed, but also incorporates the information of the linkage device, including upstream and downstream IoT devices and external demand monitoring devices.

[0085] As mentioned above, linkage devices refer to other IoT devices or systems that are directly or indirectly related to the IoT device to be analyzed. They may include upstream and downstream IoT devices, such as upstream raw material supply equipment and downstream product testing equipment on a production line; they may also include external demand monitoring devices for monitoring external events (such as seasonal demand changes, holiday impacts, special events, etc.) or user behavior. These linkage devices help the computer system to more comprehensively evaluate the load overload risk of the IoT device to be analyzed by providing additional information and data. In order to use the information provided by the linkage devices for load overload prediction, the computer system is equipped with corresponding load overload prediction neural networks for these linkage devices. These neural networks may include upstream and downstream reasoning networks and external demand reasoning networks. The upstream and downstream reasoning networks focus on analyzing the operating status and data of upstream and downstream IoT devices to evaluate their impact on the load overload of the IoT device to be analyzed; while the external demand reasoning network focuses on changes in external events and user behavior to predict how these changes affect the load of the device. The linkage monitoring data stream refers to the real-time or historical data obtained from the linkage device, including the upstream and downstream device information data stream based on the upstream and downstream IoT devices and the external demand information data stream based on the external demand monitoring device. These data streams provide the necessary input for the load overload prediction neural network.

[0086] After obtaining the linkage monitoring data stream, the computer system inputs it into the corresponding load overload prediction neural network for reasoning. Based on the upstream and downstream reasoning network, the computer system obtains the upstream and downstream load overload estimation confidence through the upstream and downstream equipment information data stream; based on the external demand reasoning network, the computer system obtains the external load overload estimation confidence through the external demand information data stream. These estimation confidences respectively reflect the degree of influence of upstream and downstream equipment and external environmental factors on the load overload of the IoT device to be analyzed.

[0087] For example, in the production line of a smart factory, there is a core server responsible for processing production data. The server is connected to the upstream raw material supply equipment and the downstream product testing equipment, and is also affected by external factors such as seasonal demand changes and holidays. In this scenario, the computer system first obtains the occupancy-type load overload estimation confidence and state-type load overload estimation confidence of the server to be analyzed through the resource occupancy reasoning network and the equipment status reasoning network. Then, the computer system obtains the data stream of the raw material supply status from the upstream raw material supply equipment, and inputs it into the upstream and downstream reasoning network to obtain the upstream and downstream load overload estimation confidence. At the same time, the computer system also obtains the data stream of seasonal demand changes and holiday effects through the external demand monitoring device, and inputs it into the external demand reasoning network to obtain the external load overload estimation confidence. Finally, the computer system integrates these estimation confidences to form a comprehensive load overload estimation confidence. This confidence comprehensively considers the status of the server itself, the influence of upstream and downstream equipment, and the changes in external environmental factors, providing factory managers with a more comprehensive and accurate equipment load overload risk assessment result. Based on this result, managers can make more reasonable decisions, such as adjusting production plans, optimizing resource allocation, etc., to ensure stable operation and efficient management of the production line.

[0088] As an implementation, the process of performing load overload reasoning on the data stream sample through the corresponding reasoning network may include:

[0089] Step S201: pre-generate the corresponding inference network's characterization information influence coefficient, intermediate network influence coefficient, and intermediate network parameter variable.

[0090] In the embodiment of the present application, step S201 is the first step to start the load overload reasoning process, which involves pre-generating or initializing key parameters in the reasoning network. These parameters include the characterization information influence coefficient, the intermediate network influence coefficient, and the intermediate network parameter variable, which together determine the way the reasoning network processes the data stream sample and the accuracy of the prediction result.

[0091] The characterization information influence coefficient is also called the feature weight, which reflects the degree of influence of different features in the input data on the load overload prediction result. In the device connection management scenario, these features may include performance indicators such as CPU usage, memory occupancy, and disk I / O speed. In step S201, the computer system first assigns initial weights to these features. These initial weights are usually generated based on experience or randomly, and they are adjusted and optimized according to the performance of the model in the subsequent reasoning process.

[0092] For example, suppose a resource usage inference network needs to process data stream samples containing two features: CPU usage and memory usage. During the initialization phase, the computer system can assign a relatively high weight (such as 0.6) to CPU usage because it is usually a key factor affecting device performance. The weight of memory usage can be set to a slightly lower value (such as 0.4), but the specific value will be adjusted according to the actual scenario and needs.

[0093] The intermediate network influence coefficient, i.e., the weight of the intermediate layer, or the hidden layer weight, determines the connection strength between the layers within the inference network. In a multi-layer neural network, each layer receives input data from the previous layer and performs linear transformation and nonlinear activation function processing through a weight matrix. In step S201, the computer system assigns initial values ​​to these weight matrices. These initial values ​​are usually randomly generated to ensure that the network can learn the characteristics of the data from different angles during the training process.

[0094] Taking the resource occupancy inference network as an example, the network may contain multiple hidden layers, each with its own weight matrix. During the initialization phase, the computer system assigns a random value to each element in these weight matrices (such as sampling from a standard normal distribution). These random values ​​will serve as the starting point for network learning and will be adjusted and optimized through the back-propagation algorithm during the subsequent training process.

[0095] Intermediate network parameters, i.e., intermediate layer parameters, are adjustable parameters in the inference network, which are used to control the transformation and calculation process of data. These parameters may include bias terms, parameters in activation functions, etc. In step S201, the computer system will also assign initial values ​​to these parameters. These initial values ​​are usually randomly generated to ensure that the network has sufficient flexibility to adapt to different data distributions and task requirements.

[0096] Take the activation function in the resource occupancy inference network as an example. This function may be a Sigmoid function or a ReLU function with parameters. During the initialization phase, the computer system assigns initial values ​​to the parameters in these activation functions (such as the slope of the Sigmoid function or the negative slope of the ReLU function). These initial values ​​will affect the shape and output range of the activation function, and thus affect the prediction performance of the entire network.

[0097] By pre-generating or initializing the key parameters in the inference network in step S201, the computer system lays the foundation for the subsequent load overload inference process. These initial parameters will be continuously adjusted and optimized during the training process to improve the prediction accuracy and generalization ability of the inference network.

[0098] Step S202: The first data stream sample of the data stream sample and the pre-generated intermediate network parameter are fused by using the pre-generated characterization information influence coefficient and the pre-generated intermediate network influence coefficient to obtain a first load overload estimation confidence.

[0099] In an embodiment of the present application, step S202 involves using the pre-generated characterization information influence coefficient and the intermediate network influence coefficient, combining the first data flow sample of the data flow sample and the pre-generated intermediate network parameter variable, and obtaining the first load overload estimation confidence through fusion calculation.

[0100] In step S202, the computer system first obtains the first data stream sample of the data stream sample. This sample usually contains a set of eigenvalues, which reflect the operating status and resource occupancy of the device at a certain moment. For example, for a production line server, the first data stream sample may contain performance indicator data such as CPU usage, memory occupancy, and disk I / O speed. Next, the computer system uses the pre-generated characterization information influence coefficient to weight these eigenvalues. The feature weight reflects the importance of different features to the load overload prediction results. The system multiplies each eigenvalue by the corresponding weight to obtain a weighted eigenvalue. This process is equivalent to adjusting the importance of the original eigenvalue, so that the features that have a greater impact on the prediction results obtain greater weights in subsequent calculations.

[0101] At the same time, the computer system also uses the pre-generated intermediate network influence coefficient and the pre-generated intermediate network parameter variables for data processing. The intermediate layer weight determines the connection strength between the layers inside the inference network, while the intermediate layer parameters are used to control the transformation and calculation process of the data. The system will fuse the weighted eigenvalues ​​with the intermediate layer weights and intermediate layer parameters, which is usually achieved through matrix multiplication and nonlinear activation functions. The result of the fusion calculation will be used as the output of the inference network, that is, the first load overload estimation confidence. Taking the production line server as an example, assume that the CPU usage rate in the first data stream sample is 80%, the memory usage rate is 70%, and the disk I / O speed is 100MB per second. The system will first weight these data according to the feature weights, for example, setting the weight of the CPU usage rate to 0.5, the weight of the memory usage rate to 0.3, and the weight of the disk I / O speed to 0.2. The weighted eigenvalues ​​are 40, 21, and 20 respectively. Then, the computer system fuses these weighted eigenvalues ​​with the intermediate layer weights and intermediate layer parameters to obtain a specific value as the first load overload estimation confidence. This confidence represents the probability of server overload based on the current data stream sample and the parameter settings of the inference network.

[0102] Through the fusion calculation process of step S202, the computer system can generate a preliminary load overload estimation confidence based on the first data stream sample and the pre-generated reasoning network parameters. This confidence will serve as the basis for subsequent reasoning steps to help the system more accurately predict the load overload risk of the equipment.

[0103] Step S203: Determine the first load overload estimation confidence as a next-generation intermediate network parameter variable, and perform load overload reasoning on the second data stream sample of the joint data stream sample.

[0104] In the embodiment of the present application, step S203 uses the load overload estimation confidence obtained by reasoning the previous data stream sample as the intermediate network parameter of the next generation, and combines it with the next data stream sample to perform a new round of reasoning, thereby gradually optimizing and updating the estimation confidence.

[0105] In step S202, the computer system has obtained the first load overload estimation confidence through fusion calculation based on the first data stream sample and the pre-generated inference network parameters. This estimation confidence reflects the probability of the device being overloaded given the first data stream sample. However, since IoT devices generate continuous data streams during operation, reasoning based on only one data stream sample may not fully reflect the load overload risk of the device.

[0106] In order to more accurately predict the load overload of the equipment, the computer system executes step S203. In this step, the computer system uses the first load overload estimation confidence as the intermediate network parameter of the next generation. This means that the first estimation confidence will serve as the starting point or reference for a new round of reasoning, affecting the reasoning results of subsequent data stream samples. Next, the computer system obtains the second data stream sample of the data stream sample. This sample also contains data on the operating status and resource occupancy of the device at a certain moment. The system combines these data with the intermediate network parameter of the next generation (i.e., the first load overload estimation confidence) and the pre-generated reasoning network parameters to perform a new round of fusion calculations. This calculation process is similar to step S202, but at this time, since the reasoning result of the previous sample is introduced as a reference, the new round of reasoning will be more accurate and reliable.

[0107] By continuously iterating step S203, the computer system can gradually optimize and update the load overload estimation confidence. As the number of processed data stream samples increases, the accuracy of the estimation confidence will continue to improve. This iterative reasoning method fully considers the characteristics of continuous operation of IoT devices, can more comprehensively reflect the load overload risk of the device, and provide more accurate and reliable decision support for device connection management.

[0108] As an implementation manner, step S40, performing collaborative reasoning through the first load overload estimation confidence and the second load overload estimation confidence of the IoT device to be analyzed to obtain the target load overload estimation confidence of the IoT device to be analyzed, may specifically include:

[0109] Step S41: Obtain a first reasoning influence coefficient assigned to a first load overload estimation confidence level and a second reasoning influence coefficient assigned to a second load overload estimation confidence level in a distributed collaborative network framework.

[0110] In the embodiment of the present application, step S41 is an important step in the collaborative reasoning process, which involves obtaining the reasoning influence coefficient (i.e., a weight) assigned to the load overload estimation confidence from different sources from the distributed collaborative network framework. These reasoning influence coefficients determine the importance of different estimation confidences in the final decision, thereby ensuring that the final load overload estimation confidence can fully and accurately reflect the load status of the device.

[0111] In step S41, the computer system will first connect to a distributed collaborative network framework, which contains load overload prediction neural networks corresponding to the IoT devices to be analyzed and their linkage devices. These neural networks have been trained with historical data and can perform load overload prediction based on the device's own data and the data provided by the linkage devices, respectively. Next, the computer system obtains from the framework a first reasoning influence coefficient assigned to the first load overload estimation confidence. The first load overload estimation confidence is obtained through a resource occupancy reasoning network and a device status reasoning network based on the resource occupancy information and device status information of the IoT device to be analyzed. The first reasoning influence coefficient reflects the importance of these estimated confidences based on the device's own data in the final decision. The specific value of this coefficient may be determined based on factors such as the application scenario, the accuracy of historical data, and the correlation between different estimated confidences.

[0112] At the same time, the computer system also obtains the second reasoning influence coefficient assigned to the second load overload estimation confidence from the framework. The second load overload estimation confidence is obtained through the upstream and downstream reasoning network or the external demand reasoning network based on the data provided by the linkage device. These linkage devices may include upstream and downstream IoT devices and external demand monitoring devices, and the data they provide provide additional information and perspectives for the load overload estimation of the IoT device to be analyzed. The second reasoning influence coefficient reflects the importance of these estimation confidences based on the linkage device data in the final decision.

[0113] For example, suppose that in a smart factory, there is a production line server as an IoT device to be analyzed. The server is connected to the upstream raw material supply equipment and the downstream product inspection equipment, and is also affected by the external demand monitoring device. In the collaborative reasoning process, the computer system first obtains the first load overload estimation confidence based on the server's own data through the resource occupancy reasoning network and the device status reasoning network, which is assumed to be 0.8. Then, the computer system obtains the second load overload estimation confidence based on the linkage device data through the upstream and downstream reasoning network and the external demand reasoning network, which is assumed to be 0.7. In this example, the first reasoning influence coefficient obtained by the computer system from the distributed collaborative network framework may be 0.6, and the second reasoning influence coefficient may be 0.4. This means that in the final decision, the estimated confidence based on the server's own data will occupy a greater weight, but the estimated confidence based on the linkage device data will also be fully considered. This weight distribution method can be adjusted and optimized according to the specific application scenario to ensure that the final load overload estimation confidence can fully and accurately reflect the server's load status.

[0114] Step S42: The first load overload estimation confidence and the second load overload estimation confidence are respectively integrated by the first inference influence coefficient and the second inference influence coefficient to obtain the target load overload estimation confidence of the IoT device to be analyzed, and the target load overload estimation confidence represents the confidence that the IoT device to be analyzed will experience load overload.

[0115] In an embodiment of the present application, step S42 utilizes the first inference influence coefficient and the second inference influence coefficient to fuse the load overload estimation confidences from different sources, thereby obtaining the target load overload estimation confidence of the IoT device to be analyzed.

[0116] In step S42, the computer system first obtains the first inference influence coefficient and the second inference influence coefficient determined in step S41. These inference influence coefficients reflect the importance of the estimated confidences from different sources in the final decision. Generally, the determination of these coefficients is based on factors such as the accuracy of historical data, the correlation between different estimated confidences, and the specific needs of the current application scenario.

[0117] Next, the computer system uses these inference influence coefficients to fuse the first load overload estimation confidence and the second load overload estimation confidence. The first load overload estimation confidence is obtained through the corresponding inference network based on the resource occupancy information and device status information of the IoT device to be analyzed, which reflects the impact of the device's own status on the load overload. The second load overload estimation confidence is obtained through the corresponding inference network based on the data provided by the linkage device, which takes into account the impact of factors such as upstream and downstream device status and external demand changes on the device load. The fusion process is to multiply each estimated confidence by its corresponding inference influence coefficient, and then add the results. This process is equivalent to weighting different estimated confidences to ensure that their contribution to the final decision matches their importance.

[0118] For example, suppose that in a smart factory, there is a production line server as an IoT device to be analyzed. Through the resource occupancy reasoning network and the device status reasoning network, the computer system obtains a first load overload estimated confidence of 0.8 based on the server's own data. At the same time, through the upstream and downstream reasoning network and the external demand reasoning network, the computer system obtains a second load overload estimated confidence of 0.7 based on the linkage device data. In step S41, the computer system determines that the first reasoning influence coefficient is 0.6 and the second reasoning influence coefficient is 0.4. In step S42, the computer system performs a fusion calculation: target load overload estimated confidence = 0.8 (first estimated confidence) * 0.6 (first reasoning influence coefficient) + 0.7 (second estimated confidence) * 0.4 (second reasoning influence coefficient) = 0.48 + 0.28 = 0.76.

[0119] As an embodiment, the method further includes:

[0120] Step S110: Obtain a load overload mark that matches the IoT device to be analyzed.

[0121] In the embodiment of the present application, step S110 involves obtaining load overload tags that match the IoT device to be analyzed. These tags are essential supervisory information in the training process and are used to guide the network framework to learn how to accurately predict load overload conditions from the monitoring data stream of the device.

[0122] In step S110, the computer system first accesses the historical records or real-time monitoring system to retrieve the load overload event data related to the IoT device to be analyzed. These event data usually contain the timestamp, category, degree and other information of the load overload event that actually occurred during the operation of the device, and this information is recorded by the system as a load overload mark. Specifically, the load overload mark may be a binary label (such as 0 for no overload and 1 for overload), or a multi-classification label (indicating different levels of load overload), or a continuous value (indicating the probability or confidence of load overload). These marks reflect the actual load status of the device at different time points and are an important basis for evaluating the prediction performance of the network framework.

[0123] For example, suppose that in a smart factory, there is a production line server as an IoT device to be analyzed. The factory manager uses a monitoring system to monitor the operating status of the server in real time and record each load overload event. Whenever the server is overloaded due to excessive resource usage, slow response speed, etc., the monitoring system generates a load overload tag and stores it in the historical records. These tags may include the specific timestamp of the event (such as "2023-10-23 14:30:05"), the category of load overload (such as "CPU overload", "memory overload", etc.), and the possible degree of load overload (such as "mild overload", "heavy overload", etc.). When training the distributed collaborative network framework, the computer system accesses these historical records and extracts the load overload tags that match the production line server. These tags will be used as the label part of the training data and compared with the prediction results of the network framework to evaluate the prediction performance of the network framework and guide its training process. By continuously iterating and optimizing the parameters and structure of the network framework, the computer system can eventually obtain a distributed collaborative network framework that can accurately predict the load overload of the equipment.

[0124] Step S120: Compare the load overload mark that matches the IoT device to be analyzed with the target load overload estimation confidence of the IoT device to be analyzed.

[0125] In the embodiment of the present application, step S120 compares the actual load overload condition of the IoT device to be analyzed with the result predicted by the network framework.

[0126] In step S120, the computer system first obtains the load overload mark that matches the IoT device to be analyzed. These marks are usually generated by the monitoring system when a load overload event actually occurs on the device. They record the load status of the device at a specific time point, such as whether an overload occurs, the degree of overload, etc. These marks are real-world feedback and represent the actual operating status of the device.

[0127] Next, the computer system compares the acquired load overload mark with the target load overload estimated confidence of the IoT device to be analyzed. The target load overload estimated confidence is obtained by inferring the monitoring data stream of the device through the load overload prediction neural network in the distributed collaborative network framework, which represents the probability of the device being overloaded in the current state. This estimated confidence is the prediction of the future state of the device by the network framework based on the learned knowledge and rules. The comparison process is essentially to compare the prediction results of the network framework with the actual load state of the device. If the estimated confidence is consistent with the load overload mark, that is, the network framework accurately predicts the load state of the device, then it means that the performance of the network framework is good and can effectively identify the load overload risk of the device. On the contrary, if there is a large difference between the estimated confidence and the load overload mark, then it means that there is an error in the prediction of the network framework and further optimization is needed.

[0128] For example, suppose that in a smart factory, there is a production line server as an IoT device to be analyzed. The factory manager uses a monitoring system to monitor the operating status of the server in real time and record each load overload event. At a certain moment, the monitoring system detects that a load overload event has occurred on the server and generates a load overload mark, indicating that the server is indeed overloaded at that moment. At the same time, the computer system infers the monitoring data stream of the server through the load overload prediction neural network in the distributed collaborative network framework, and obtains a target load overload estimation confidence, assuming it is 0.85. This estimated confidence indicates that according to the prediction of the network framework, there is an 85% probability that the server will be overloaded.

[0129] In step S120, the computer system compares the load overload mark with the target load overload estimated confidence. Since the monitoring system has confirmed that the server is indeed overloaded at this moment, the load overload mark is true (such as marked as 1). At this time, if the estimated confidence is also close to or exceeds a preset threshold (such as 0.7), the prediction of the network framework is considered to be accurate. In this example, the estimated confidence is 0.85, which exceeds the preset threshold, so it can be considered that the network framework performs well in this prediction.

[0130] However, if there is a large difference between the estimated confidence and the load overload mark (e.g., the estimated confidence is far below the threshold or opposite to the actual load status), it means that there is an error in the prediction of the network framework, and its parameters and structure need to be further adjusted and optimized to improve the accuracy and reliability of the prediction.

[0131] Step S130: training the distributed collaborative network framework through comparison results until a trained distributed collaborative network framework is obtained.

[0132] In the embodiment of the present application, step S130 adjusts and optimizes the network framework according to the comparison result between the load overload mark and the target load overload estimated confidence until a trained model is obtained.

[0133] In step S130, the computer system compares the load overload mark obtained from step S120 with the estimated confidence of the target load overload of the IoT device to be analyzed. This comparison result reflects the current prediction performance of the network framework, that is, the consistency between its prediction results and the actual load overload situation. If the comparison result shows that the prediction performance of the network framework is good, that is, the estimated confidence is consistent or close to the load overload mark, then it means that the network framework already has a certain prediction ability and does not need large-scale adjustments. However, in order to further improve the prediction accuracy and generalization ability, the computer system can choose to continue training the network framework and use more training data to fine-tune its parameters and structure. If the comparison result shows that there are errors in the prediction of the network framework, that is, there is a large difference between the estimated confidence and the load overload mark, then the computer system will start a training process to optimize the performance of the network framework. This process, for example, includes the following steps:

[0134] Forward propagation: The computer system uses the network framework to perform inference calculations on the input data (i.e., the monitoring data stream of the equipment) and obtains the output result (i.e., the estimated confidence level of the target load overload).

[0135] Calculate the loss function: Calculate the loss function value based on the difference between the output result and the true label (i.e., the load overload mark). The loss function is used to quantify the size of the network framework prediction error and is the basis for optimizing the network framework.

[0136] Back propagation: Based on the value of the loss function, the back propagation algorithm is used to calculate the gradient of each parameter in the network framework, that is, the degree of influence of the parameter on the loss function.

[0137] Update parameters: Use an optimization algorithm (such as the gradient descent algorithm) to update the parameters in the network framework according to the calculated gradient to reduce the loss function value and optimize the prediction performance.

[0138] This process is repeated for multiple iterations, with each iteration using new training data to update the parameters of the network framework until the performance of the network framework reaches the expected level.

[0139] For example, suppose that in a smart factory, a production line server is often overloaded. The factory manager uses a monitoring system to monitor the server's operating status in real time and record each overload event. When training the distributed collaborative network framework, the computer system first obtains the overload mark in the historical record from the monitoring system as the true label. Then, the computer system uses the network framework to infer the input data (i.e., the server's monitoring data stream) to obtain the target load overload estimated confidence. Next, the computer system compares the estimated confidence with the true label and calculates the loss function value. If the loss function value is large, it means that there is an error in the prediction of the network framework. At this time, the computer system uses the back propagation algorithm to calculate the gradient and uses the optimization algorithm to update the parameters in the network framework. This process will be repeated for many iterations until the loss function value converges to a smaller value, indicating that the prediction performance of the network framework has been optimized. By continuously iterating and optimizing the parameters and structure of the network framework, the computer system can eventually obtain a trained distributed collaborative network framework. This network framework can accurately predict the load overload of the equipment, provide effective decision support for factory managers, and help them take timely measures to avoid performance degradation or failure caused by load overload of the equipment.

[0140] As an implementation manner, step S130, training the distributed collaborative network framework through comparison results, may specifically include:

[0141] Step S131: If the comparison result indicates that the load overload mark is different from the target load overload estimated confidence, the network error function of the IoT device to be analyzed is determined based on the device monitoring data stream, and the first inference influence coefficient for the first load overload estimated confidence allocated by the distributed collaborative network framework is optimized through the network error function.

[0142] In the embodiment of the present application, step S131 adjusts and optimizes the parameters of the network framework according to the comparison result between the prediction result and the actual load overload situation.

[0143] In step S131, the computer system compares the load overload mark obtained from step S120 with the target load overload estimated confidence of the IoT device to be analyzed. This comparison result reflects the current prediction performance of the network framework, that is, the consistency between its prediction results and the actual load overload situation. If the comparison result shows that there is a significant difference between the target load overload estimated confidence and the load overload mark, that is, the prediction result of the network framework does not match the actual situation, then the system needs to adjust and optimize the network framework. At this time, the computer system executes step S131, and the specific process is as follows:

[0144] First, the computer system determines the network error function (also called loss function) of the IoT device to be analyzed based on the device monitoring data stream. The network error function is used to quantify the size of the network framework prediction error and is the basis for optimizing the network framework. In the device connection management scenario, commonly used network error functions include mean square error (MSE), cross entropy loss, etc. These functions can calculate a specific value based on the difference between the output result of the network framework (i.e., the target load overload estimation confidence) and the true label (i.e., the load overload mark) as a measure of the network framework prediction error.

[0145] Taking the mean square error (MSE) as an example, assuming that the load overload mark is y (actual value, such as 1 for overload and 0 for no overload), and the target load overload estimate confidence is y' (predicted value, such as 0.85), the calculation formula of MSE is: MSE = (y-y') 2 If y is 1 and y' is 0.85, the MSE is a small positive value, indicating that there is a certain error in the prediction result.

[0146] Next, the computer system optimizes the network framework using back-propagation algorithms and optimization algorithms (such as gradient descent algorithms). Specifically, the computer system calculates the gradients of various parameters in the network framework (i.e., the degree of influence of the parameters on the network error) based on the value of the network error function, and then updates the parameters of the network framework based on the gradients. In this process, the computer system pays special attention to the first inference influence coefficient for the first load overload estimated confidence. This coefficient determines the importance of the estimated confidence based on the device's own data in the final decision. By adjusting the first inference influence coefficient, the computer system can optimize the network framework's ability to process the device's own data, thereby improving the accuracy of the prediction.

[0147] For example, suppose that in a smart factory, production line servers are often overloaded. Factory managers use a monitoring system to monitor the operating status of the server in real time and record each overload event. When training the distributed collaborative network framework, the computer system first obtains the overload mark in the historical record from the monitoring system as the true label. Then, the computer system uses the network framework to perform inference calculations on the input data (i.e., the monitoring data stream of the server) to obtain the target load overload estimated confidence. If the comparison result shows that there is a large difference between the estimated confidence and the true label (such as the estimated confidence is 0.7 and the true label is 1), the computer system calculates the gradient according to the network error function and adjusts the network framework parameters such as the first inference influence coefficient. After multiple iterations of training, the performance of the network framework is improved, and it can more accurately predict the server load overload.

[0148] Step S132: Transmit the comparison result to the linkage device, and let the linkage device determine the matching network error function based on the comparison result and the linkage monitoring data flow, and optimize the second inference influence coefficient of the second load overload estimation confidence allocated by the distributed collaborative network framework through the matching network error function, so as to realize the training of the distributed collaborative network framework.

[0149] In the embodiment of the present application, step S132 is an important part of the distributed collaborative network framework training process, which involves using the information provided by the linkage device to further optimize the parameters of the network framework. In step S132, if the comparison result (i.e., the comparison of the load overload mark and the target load overload estimated confidence) indicates that there is an inconsistency, that is, there is a difference between the prediction result of the network framework and the actual load overload situation, the computer system transmits the comparison result to the linkage device. The linkage device usually includes upstream and downstream IoT devices and external demand monitoring devices, etc., which can provide additional information and data related to the IoT device to be analyzed. Once the linkage device receives the comparison result, it will determine a matching network error function based on its own linkage monitoring data stream and the received comparison result. This matching network error function is used to quantify the error size of the data provided by the linkage device in the network framework prediction. Similar to the network error function of the IoT device to be analyzed, the matching network error function may also adopt common forms such as mean square error (MSE) and cross entropy loss, but the specific selection will be determined according to the data characteristics and application scenarios of the linkage device.

[0150] After determining the matching network error function, the linkage device uses back-propagation algorithms and optimization algorithms (such as gradient descent) to optimize the parameters in the distributed collaborative network framework. In particular, the linkage device pays attention to the second inference influence coefficient for the confidence of the second load overload estimate. This coefficient determines the importance of the estimated confidence based on the linkage device data in the final decision. By adjusting the second inference influence coefficient, the linkage device can optimize the network framework's ability to process the linkage device data, thereby improving the accuracy of the prediction.

[0151] For example, suppose that in a smart factory, there is a production line server as an IoT device to be analyzed, which is connected to the upstream raw material supply equipment and the downstream product inspection equipment. These linkage devices can provide information about the raw material supply status and product inspection results, which is of great significance for predicting the load overload of the server. In the process of training the distributed collaborative network framework, if the comparison result shows that there is a difference between the target load overload estimation confidence of the server and the actual load overload mark, the computer system transmits this comparison result to the linkage device. The linkage device determines a matching network error function based on its own monitoring data stream (such as raw material supply speed, product inspection pass rate, etc.) and the received comparison result. Then, the linkage device uses the back propagation algorithm and the optimization algorithm to adjust the parameters in the distributed collaborative network framework, especially the second inference influence coefficient for the second load overload estimation confidence. Through this process, the linkage device can optimize the network framework's processing ability of the linkage device data, so that the network framework can more comprehensively consider the mutual influence between the server and the linkage device, thereby improving the accuracy of the prediction.

[0152] By implementing step S132, the computer system can not only use the data of the IoT device to be analyzed to train the network framework, but also make full use of the information provided by the linkage device to further optimize the performance of the network framework. This enables the distributed collaborative network framework to more accurately predict the load overload of the device and provide strong support for device connection management.

[0153] As an implementation method, the network error function of the IoT device to be analyzed includes the error value of the resource occupancy reasoning network corresponding to the IoT device to be analyzed, the error value of the device state reasoning network corresponding to the IoT device to be analyzed, and the collaborative error value between the resource occupancy reasoning network and the device state reasoning network; the matching network error function includes the error value of the upstream and downstream reasoning networks, the error value of the external demand reasoning network, the collaborative error value between the upstream and downstream reasoning networks and the external demand reasoning network, the collaborative error value between the resource occupancy reasoning network and the upstream and downstream reasoning networks, the collaborative error value between the resource occupancy reasoning network and the external demand reasoning network, the collaborative error value between the device state reasoning network and the upstream and downstream reasoning networks, and the collaborative error value between the device state reasoning network and the external demand reasoning network.

[0154] In the embodiments of the present application, the network error function and its components play a vital role in training the distributed collaborative network framework. These error functions not only take into account the resource occupancy and device status information of the IoT device to be analyzed, but also incorporate the data provided by the linkage device and the synergy between them. The following is a detailed explanation of these network error functions and their components, and is illustrated in conjunction with specific application scenarios and exemplary calculation formulas.

[0155] The network error function of the IoT device to be analyzed consists of three parts: the error value of the resource occupancy reasoning network, the error value of the device state reasoning network, and the collaborative error value between the two reasoning networks.

[0156] Error value of resource occupancy inference network: This error value measures the difference between the prediction result of resource occupancy inference network and the actual load overload mark. It is usually calculated based on resource occupancy information data stream samples (such as CPU usage, memory usage, etc.) and the corresponding load overload mark. For example, the mean square error (MSE) can be used as the calculation formula:

[0157] ;

[0158] Where N is the number of samples, yi is the true load overload mark of the i-th sample, is the prediction result of the resource occupancy inference network for the i-th sample.

[0159] Error value of the device state inference network: This error value measures the difference between the prediction result of the device state inference network and the actual load overload mark. It is calculated based on the data stream samples of the device state information (such as temperature, health status, etc.), and can also be calculated using the mean square error as the calculation formula:

[0160] ;

[0161] in, is the prediction result of the device state inference network for the i-th sample.

[0162] Collaborative error value: This error value measures the collaborative effect of the resource occupancy reasoning network and the device state reasoning network in the prediction process. It reflects the error that may occur when the two networks fuse the prediction results. The calculation method of the collaborative error value can be set according to the specific application scenario and requirements, and can be the average of the absolute difference between the prediction results of the two networks:

[0163] ;

[0164] The matching network error function takes into account the data provided by the linkage devices (such as upstream and downstream IoT devices and external demand monitoring devices) and their synergy with the IoT devices to be analyzed. It includes the error values ​​of the upstream and downstream reasoning networks, the error values ​​of the external demand reasoning network, and the collaborative error values ​​between them.

[0165] Error values ​​of the upstream and downstream reasoning networks and the error value of the external demand reasoning network: The calculation method of these two error values ​​is similar to the error values ​​of the resource occupancy reasoning network and the device status reasoning network of the IoT device to be analyzed, and they are calculated based on the upstream and downstream device information data streams and the external demand information data streams respectively.

[0166] Collaborative error value: This part includes multiple collaborative error terms, which measure the mutual influence and interaction of different inference networks in the prediction process. The following are several exemplary collaborative error value calculation formulas:

[0167] Collaborative error value between upstream and downstream reasoning networks and external demand reasoning networks: This error value reflects the difference in prediction results between the upstream and downstream reasoning networks and the external demand reasoning network. It can be calculated using the average of the absolute differences similar to the collaborative error value:

[0168] ;

[0169] in, is the prediction result of the upstream and downstream reasoning networks for the i-th sample, It is the prediction result of the external demand reasoning network.

[0170] The collaborative error value between the resource occupancy inference network and the upstream and downstream inference networks: This error value measures the collaborative effect of the resource occupancy inference network and the upstream and downstream inference networks in the prediction process. It can be calculated using some combination or difference of the prediction results of the two networks, depending on the application scenario and requirements.

[0171] Other collaborative error values ​​(such as the collaborative error value between the device state inference network and the upstream and downstream inference networks) are calculated in a similar way, all of which are based on the differences or combinations between the prediction results of different inference networks to measure the synergy between them.

[0172] By comprehensively considering these network error functions and their components, the computer system can more comprehensively optimize network parameters when training the distributed collaborative network framework, improving the accuracy and reliability of predictions. This helps the device connection management system to more accurately predict the load overload of the device, so as to take corresponding management measures to optimize device performance and ensure the stable operation of the system.

[0173] As an implementation method, the first reasoning influence coefficient includes the reasoning influence coefficient of the resource occupancy reasoning network and the reasoning influence coefficient corresponding to the device status reasoning network; the second reasoning influence coefficient includes the reasoning influence coefficient of the upstream and downstream reasoning networks and the reasoning influence coefficient corresponding to the external demand reasoning network; wherein, any reasoning influence coefficient is used to fuse the load overload estimation confidence of the corresponding reasoning network.

[0174] In the embodiment of the present application, the inference influence coefficient determines the importance of the load overload estimation confidence from different sources in the final decision, thereby ensuring the accuracy and reliability of the prediction results.

[0175] The first inference influence coefficient includes the inference influence coefficient of the resource occupancy inference network and the inference influence coefficient corresponding to the device state inference network. These coefficients respectively reflect the importance of the estimated confidence based on the device's own resource occupancy information and device state information in the final decision.

[0176] Reasoning influence coefficient of resource occupancy inference network: This coefficient measures the contribution of resource occupancy information (such as CPU usage, memory usage, etc.) in predicting device overload. In the device connection management scenario, resource occupancy information is usually one of the key indicators to determine whether the device is overloaded. Therefore, the reasoning influence coefficient of the resource occupancy inference network is often assigned a higher value to ensure that the estimated confidence based on this information plays an important role in the final decision.

[0177] Reasoning influence coefficient corresponding to the device state reasoning network: Similar to resource occupancy information, device state information (such as temperature, health status, etc.) is also an important basis for judging the load overload of the device. The reasoning influence coefficient of the device state reasoning network reflects the importance of this information in the prediction process. Although its contribution may be slightly lower than that of resource occupancy information, device state information also has an important impact on the prediction results.

[0178] The second inference influence coefficient includes the inference influence coefficient of the upstream and downstream inference network and the inference influence coefficient corresponding to the external demand inference network. These coefficients measure the contribution of the data provided by the linkage device in the prediction process.

[0179] Inference influence coefficient of upstream and downstream inference network: In the IoT device connection management scenario, the mutual influence between upstream and downstream devices is inevitable. The inference influence coefficient of the upstream and downstream inference network reflects the importance of the estimated confidence based on the upstream and downstream device information (such as raw material supply speed, product inspection pass rate, etc.) in the final decision. This information helps to more fully understand the operating environment and load conditions of the device to be analyzed, thereby improving the accuracy of the prediction.

[0180] Reasoning influence coefficient corresponding to the external demand reasoning network: External demand factors (such as seasonal demand changes, holiday impacts, etc.) also have an important impact on the load of the equipment. The reasoning influence coefficient of the external demand reasoning network measures the contribution of the estimated confidence based on these external demand information in the prediction process. By adjusting this coefficient, the computer system can optimize the network framework's ability to respond to external demand changes, thereby more accurately predicting the load overload of the equipment.

[0181] When training the distributed collaborative network framework, the computer system first determines the initial values ​​of the first inference influence coefficient and the second inference influence coefficient. As the training progresses, the computer system continuously adjusts these coefficients based on the comparison results (i.e., the comparison of the load overload mark with the target load overload estimated confidence). For example, if the estimated confidence based on resource occupancy information shows high accuracy in multiple iterations, the system can increase the inference influence coefficient of the resource occupancy inference network; conversely, if the estimated confidence based on external demand information plays a key role in the prediction, the system can increase the corresponding inference influence coefficient of the external demand inference network.

[0182] In this way, the computer system can dynamically adjust the inference influence coefficient according to the contribution of information from different sources, thereby ensuring that the distributed collaborative network framework can accurately predict the load overload of the equipment and provide strong support for device connection management.

[0183] As an implementation method, the target load overload estimation confidence level represents the target confidence level of load overload of the IoT device to be analyzed. The method further includes:

[0184] Step S50: Compare the target load overload estimation confidence level representing the target confidence level of load overload of the IoT device to be analyzed with the confidence threshold.

[0185] In the embodiment of the present application, step S50 compares the target load overload estimation confidence with the preset confidence threshold to determine whether further preventive measures need to be taken. The target load overload estimation confidence is obtained through collaborative reasoning of multiple reasoning networks (such as resource occupancy reasoning network, device state reasoning network, upstream and downstream reasoning network, external demand reasoning network, etc.) in the distributed collaborative network framework. It integrates the device's own state information, the data provided by the linkage device, and the mutual influence between them, and comprehensively reflects the possibility of load overload of the IoT device to be analyzed in the current state.

[0186] The confidence threshold is a value set according to the actual application scenario and requirements, which represents the system's tolerance for the risk of equipment overload. For example, in a production environment with extremely high requirements for equipment performance, the confidence threshold can be set lower so that the system can promptly detect and respond to potential load overload risks; in an environment with relatively loose performance requirements, the confidence threshold can be set slightly higher. In step S50, the computer system compares the target load overload estimated confidence with the confidence threshold. If the estimated confidence is greater than or equal to the threshold, it means that the risk of equipment overload is high, and the computer system takes further measures to reduce this risk. On the contrary, if the estimated confidence is less than the threshold, it means that the current load condition of the equipment is relatively safe, and the computer system can continue to observe the operating status of the equipment.

[0187] For example, suppose that in a smart factory, the production line server is used as the IoT device to be analyzed. Through the collaborative reasoning of the distributed collaborative network framework, the computer system obtains the target load overload estimate confidence of the server as 0.85. According to the actual needs of the factory and the emphasis on equipment performance, the confidence threshold set by the computer system is 0.8. Since the estimated confidence is greater than the threshold, the computer system determines that the server has a high risk of load overload. Next, the computer system will execute corresponding preventive measures based on this judgment result, such as outputting load overload warning information, reducing the resource allocation of the server, etc., to ensure the stable operation of the server and the continuity of factory production.

[0188] Step S60: When the target confidence is not less than the confidence threshold, output load overload warning information, and reduce resource allocation of the IoT device to be analyzed based on the load overload warning information.

[0189] When the target confidence is not less than the confidence threshold, the computer system immediately outputs the load overload warning information. This warning information contains key information such as the identification of the IoT device to be analyzed, the current load overload estimated confidence, and the possible risk level. The output of the warning information can be achieved in a variety of ways, such as sending an email, pushing a mobile phone notification, displaying a warning sign on the monitoring interface, etc., to ensure that the equipment manager or related system can quickly receive this important information. The output of the warning information not only reminds the relevant personnel to pay attention to the load status of the equipment, but also provides them with a basis for taking further measures. The equipment manager can judge whether immediate intervention is needed or take other preventive measures to reduce the load pressure of the equipment based on the content of the warning information and the actual situation. In addition to outputting the warning information, the computer system will automatically reduce the resource allocation of the IoT device to be analyzed based on the load overload warning information in step S60. The reduction of resource allocation usually involves the adjustment of key resources such as the CPU usage rate, memory occupancy rate, and network bandwidth of the device.

[0190] The purpose of reducing resource allocation is to relieve the load pressure of the device and prevent its performance degradation or failure due to overload. Specifically, the computer system can reduce resource allocation by adjusting the operating parameters of the device, optimizing task scheduling, limiting the number of concurrent connections, etc. These measures can be flexibly adjusted according to the actual situation of the device and the confidence level of the load overload estimate to ensure that the device can maintain normal operation while reducing the load pressure.

[0191] For example, suppose that in a smart factory, the production line server is used as the IoT device to be analyzed. Through the collaborative reasoning of the distributed collaborative network framework, the computer system obtains the target load overload estimate confidence of the server as 0.9 (far exceeding the preset confidence threshold of 0.8). In this case, the computer system immediately outputs the load overload warning information to the equipment manager and automatically reduces the resource allocation of the server at the same time. Specifically, the computer system can limit the CPU usage of the server, reduce the number of high-load tasks running, or optimize the task scheduling strategy to reduce the memory occupancy of the server. Through the implementation of these measures, the computer system can effectively reduce the load pressure of the server and prevent its performance degradation or failure due to overload. At the same time, the equipment manager can also further evaluate the load status of the equipment based on the warning information and the measures taken by the system, and decide whether to take other necessary preventive measures.

[0192] Figure 2 A hardware entity diagram of a computer system provided in an embodiment of the present application is shown in FIG. Figure 2As shown, the hardware entity of the computer system 1000 includes: a processor 1001 and a memory 1002, wherein the memory 1002 stores a computer program that can be run on the processor 1001, and the processor 1001 implements the steps in the method of any of the above embodiments when executing the program.

[0193] The above are only implementation methods of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application.

Claims

1. A device connection management method based on the Internet of Things, characterized in that: include: Acquire a device monitoring data stream, wherein the device monitoring data stream is composed of performance indicator data generated by different operation cycles of the IoT device to be analyzed during operation; Obtaining a load overload prediction neural network corresponding to the IoT device to be analyzed in the distributed collaborative network framework, and performing load overload reasoning on the device monitoring data stream to obtain a first load overload estimation confidence of the IoT device to be analyzed; the distributed collaborative network framework includes a load overload prediction neural network corresponding to the linkage device of the IoT device to be analyzed; A second load overload estimation confidence for the IoT device to be analyzed is obtained in the linkage device, where the second load overload estimation confidence is obtained by performing load overload inference on the linkage monitoring data stream obtained by the linkage device and obtaining a load overload prediction neural network corresponding to the linkage device; Performing collaborative reasoning on the first load overload estimation confidence level and the second load overload estimation confidence level of the IoT device to be analyzed to obtain a target load overload estimation confidence level of the IoT device to be analyzed; Wherein, the acquiring device monitoring data stream includes: Obtaining a data stream of resource occupancy information of the IoT device to be analyzed and a data stream of device status information for the IoT device to be analyzed, and obtaining a load overload mark of the IoT device to be analyzed, wherein the load overload mark includes a mark generation time, and the mark generation time indicates the time when the load overload occurs in the IoT device to be analyzed; According to the load overload mark and the mark generation time contained in the load overload mark, the load overload mark is matched with the resource occupation information data stream and the device status information data stream respectively; Generate the device monitoring data stream based on the resource occupancy information data stream after matching the load overload mark and the device status information data stream after matching the load overload mark; The step of performing collaborative reasoning on the first load overload estimation confidence level and the second load overload estimation confidence level of the IoT device to be analyzed to obtain the target load overload estimation confidence level of the IoT device to be analyzed includes: Obtaining a first reasoning influence coefficient assigned to the first load overload estimation confidence level and a second reasoning influence coefficient assigned to the second load overload estimation confidence level in the distributed collaborative network framework; The first load overload estimation confidence and the second load overload estimation confidence are respectively integrated by the first inference influence coefficient and the second inference influence coefficient to obtain the target load overload estimation confidence of the IoT device to be analyzed, and the target load overload estimation confidence represents the confidence that the IoT device to be analyzed is load overloaded.

2. The method according to claim 1, characterized in that The resource occupancy information data stream is composed of resource occupancy information data in the time interval [t0, tn], and the device status information data stream is composed of device status information data in the time interval [t0, tn], where tn is a positive integer greater than 0. If the mark generation time of the load overload mark is x, x∈[t0, tn]; The matching of the load overload mark with the resource occupancy information data stream and the device status information data stream respectively according to the load overload mark and the mark generation time contained in the load overload mark includes: Based on the mark generation time x of the load overload mark, the load overload mark is matched with the resource occupancy information data flow in the time interval [t0, tn] as the resource occupancy information data flow associated with the load overload mark, and based on the mark generation time x of the load overload mark, the load overload mark is matched with the device status information data flow in the time interval [t0, tn], and is determined as the device status information data flow matching the load overload mark; The method of generating the device monitoring data stream of the IoT device to be analyzed based on the resource occupancy information data stream after matching the load overload mark and the device status information data stream after matching the load overload mark includes: Respectively performing a data stream splitting operation on the resource occupancy information data stream after matching the load overload mark, and performing a data stream splitting operation on the device status information data stream after matching the load overload mark, to obtain at least one resource occupancy information data stream sample, and at least one device status information data stream sample, respectively; Any one of the at least one resource occupancy information data stream sample and any one of the at least one device status information data stream sample are determined as the device monitoring data stream of the Internet of Things device to be analyzed.

3. The method according to claim 1, characterized in that The load overload prediction neural network corresponding to the IoT device to be analyzed in the distributed collaborative network framework includes a resource occupancy inference network corresponding to the IoT device to be analyzed and a device state inference network corresponding to the IoT device to be analyzed, and the device monitoring data stream includes a resource occupancy information data stream sample and a device state information data stream sample; the load overload prediction neural network corresponding to the IoT device to be analyzed in the distributed collaborative network framework is obtained by performing load overload inference on the device monitoring data stream to obtain a first load overload estimation confidence of the IoT device to be analyzed, including: Obtaining a resource occupancy inference network corresponding to the IoT device to be analyzed, performing load overload inference on the resource occupancy information data stream sample, and obtaining an occupancy type load overload estimation confidence of the IoT device to be analyzed; Obtaining a device state inference network corresponding to the IoT device to be analyzed, performing load overload inference on the device state information data stream sample, and obtaining a state-type load overload estimation confidence of the IoT device to be analyzed; The occupancy-type load overload estimation confidence and the state-type load overload estimation confidence are used as the first load overload estimation confidence of the Internet of Things device to be analyzed.

4. The method according to claim 1, characterized in that The linkage device of the IoT device to be analyzed includes at least one of upstream and downstream IoT devices and an external demand monitoring device, and the load overload prediction neural network corresponding to the linkage device includes at least one of an upstream and downstream reasoning network and an external demand reasoning network; the linkage monitoring data stream includes at least one of an upstream and downstream device information data stream obtained based on the upstream and downstream IoT devices, and an external demand information data stream obtained based on the external demand monitoring device; The second load overload estimation confidence obtained in the linkage device includes: obtaining the upstream and downstream load overload estimation confidence by inferring the upstream and downstream device information data flow load overload based on the upstream and downstream inference network, and obtaining the external load overload estimation confidence by inferring the external demand information data flow load overload based on the external demand inference network; The process of performing load overload reasoning on data stream samples through the corresponding reasoning network includes: Pre-generate the corresponding inference network's representation information influence coefficient, intermediate network influence coefficient, and intermediate network parameter variable; By using the pre-generated characterization information influence coefficient and the pre-generated intermediate network influence coefficient, the first data stream sample of the data stream sample and the pre-generated intermediate network parameter variable are fused to obtain the first load overload estimation confidence; The first load overload estimation confidence is determined as a next-generation intermediate network parameter variable, and a second data stream sample combined with the data stream sample is used to perform load overload reasoning.

5. The method according to claim 1, characterized in that The method further comprises: Obtaining a load overload mark matching the IoT device to be analyzed; Compare the load overload mark matched with the IoT device to be analyzed with the target load overload estimation confidence of the IoT device to be analyzed; The distributed collaborative network framework is trained by comparing the results until a trained distributed collaborative network framework is obtained.

6. The method according to claim 5, characterized in that The step of training the distributed collaborative network framework by comparing the results includes: If the comparison result indicates that the load overload mark is different from the target load overload estimated confidence, then determining a network error function of the IoT device to be analyzed based on the device monitoring data stream, and optimizing a first inference influence coefficient for the first load overload estimated confidence allocated by the distributed collaborative network framework through the network error function; and; The comparison result is transmitted to the linkage device, and the linkage device is instructed to determine a matching network error function based on the comparison result and the linkage monitoring data stream, and the second reasoning influence coefficient for the second load overload estimation confidence level assigned by the distributed collaborative network framework is optimized by the matching network error function, so as to realize training of the distributed collaborative network framework; Among them, the network error function of the IoT device to be analyzed includes the error value of the resource occupancy reasoning network corresponding to the IoT device to be analyzed, the error value of the device state reasoning network corresponding to the IoT device to be analyzed, and the collaborative error value between the resource occupancy reasoning network and the device state reasoning network; the matching network error function includes the error value of the upstream and downstream reasoning networks, the error value of the external demand reasoning network, the collaborative error value between the upstream and downstream reasoning network and the external demand reasoning network, the collaborative error value between the resource occupancy reasoning network and the upstream and downstream reasoning network, the collaborative error value between the resource occupancy reasoning network and the external demand reasoning network, the collaborative error value between the device state reasoning network and the upstream and downstream reasoning networks, and the collaborative error value between the device state reasoning network and the external demand reasoning network; The first reasoning influence coefficient includes the reasoning influence coefficient of the resource occupancy reasoning network and the reasoning influence coefficient corresponding to the device status reasoning network; the second reasoning influence coefficient includes the reasoning influence coefficient of the upstream and downstream reasoning networks and the reasoning influence coefficient corresponding to the external demand reasoning network; wherein, any reasoning influence coefficient is used to fuse the load overload estimation confidence of the corresponding reasoning network.

7. The method according to claim 1, characterized in that The target load overload estimation confidence level represents the target confidence level of load overload of the IoT device to be analyzed; the method further includes: Compare the target confidence level of the target load overload estimation confidence level representing the load overload of the IoT device to be analyzed with a confidence threshold; When the target confidence is not less than the confidence threshold, load overload warning information is output, and resource allocation of the to-be-analyzed Internet of Things device is reduced based on the load overload warning information.

8. A computer system comprising a memory and a processor, wherein the memory stores a computer program executable on the processor, wherein: When the processor executes the program, the steps in the method according to any one of claims 1 to 7 are implemented.

Images