Edge router fast acquisition decision method and system
By employing meta-learning algorithms to construct pre-trained models and quantum filtering rules in edge routers, the redundancy and adaptability issues of edge router data acquisition are resolved, achieving efficient and intelligent data processing and resource optimization.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- GUANGZHOU ROBUSTEL CO LTD
- Filing Date
- 2025-12-25
- Publication Date
- 2026-05-12
AI Technical Summary
Existing edge routers suffer from problems such as excessive redundant data, network congestion, resource waste, and low accuracy of prediction models during data collection, and they are difficult to adapt quickly to new monitoring scenarios and devices.
A pre-trained model is built in the cloud using a meta-learning algorithm. A personalized data acquisition strategy model is generated using a small number of data samples. Combined with quantum filtering rules, data acquisition and filtering are optimized to achieve intelligent decision-making for edge routers.
This improves the adaptability and resource utilization of edge routers, reduces the amount of data reported, and enhances data quality and the accuracy of cloud-based predictive models.
Smart Images
Figure CN121397083B_ABST
Abstract
Description
Technical Field
[0001] This application belongs to the field of Internet of Things (IoT) technology, and more specifically, relates to an edge router fast acquisition decision method and system. Background Technology
[0002] With the deepening development of Industry 4.0 and the Internet of Things, edge routers, as key nodes connecting the physical and digital worlds, need to transmit massive amounts of data from various sensors to cloud servers via networks for analysis and prediction by cloud-based predictive models. For existing edge routers, to avoid missing crucial information, high-frequency, comprehensive data collection from sensors is necessary. This generates a large amount of redundant data, easily causing network congestion and wasting cloud storage and computing resources. Moreover, truly valuable information is often buried in the massive amounts of data, affecting the accuracy of cloud-based predictive models. Although some edge routers have filtering rules, relying solely on static rules or preset scripts for collection and filtering passively executes instructions, failing to proactively decide on collection and filtering based on the data's inherent value and real-time changing operating conditions. Furthermore, when edge routers are applied to new monitoring scenarios, extensive manual configuration and data sample training for model training and strategy adjustments are required, resulting in long adaptation cycles and failing to meet the rapid deployment and iteration needs of flexible manufacturing.
[0003] Therefore, the technical problem addressed by this solution is: how to improve the intelligence, adaptability, and resource utilization of edge routers. Summary of the Invention
[0004] The main objective of this application is to provide a rapid data collection decision-making method for edge routers. By employing a meta-learning algorithm, a pre-trained model is trained in the cloud. A small number of data samples from the target device are then input into the pre-trained model to obtain a personalized data collection strategy model for the target device. The data collection strategy model then outputs a data collection strategy for the target device, enabling the edge router to perform adaptive data collection for the target device, reducing the amount of data reported to the cloud, thereby improving the intelligence, adaptability, and resource utilization of the edge router.
[0005] In addition, an edge router rapid acquisition decision system is also provided.
[0006] To achieve the above objectives, the technical solution adopted in this application is as follows:
[0007] An edge router fast sampling decision method includes the following steps:
[0008] Step 1: Load the pre-trained model. The pre-trained model is obtained by the cloud server constructing a meta-model based on the meta-learning algorithm and training the meta-model using historical task data from multiple devices.
[0009] Step 2: Collect data samples from the target device, and use the data samples to iteratively adjust the pre-trained model to obtain the acquisition strategy model. The acquisition strategy model outputs an acquisition strategy mapping table, which contains the acquisition objects, acquisition frequencies, and thresholds corresponding to various operating states.
[0010] Step 3: Obtain the current operating status of the target device, map the current operating status information to the acquisition strategy mapping table, filter out the best acquisition strategy, and then acquire the sensor data of the target device according to the best acquisition strategy to obtain the target data;
[0011] Step 4: Construct quantum filtering rules based on the acquisition strategy mapping table, and filter the target data using the quantum filtering rules to obtain the reported data.
[0012] Preferably, the historical task data includes historical operating status and corresponding collected data; in step 1, the training process of the pre-trained model is as follows: the cloud server constructs a meta-model based on the meta-learning algorithm, uses the historical operating status of various devices and the corresponding collected data as the training set to train the meta-model, and then optimizes the parameters of the meta-model by minimizing the policy utility loss function to obtain the pre-trained model.
[0013] Preferably, the formula for minimizing the policy utility loss function is:
[0014] ;
[0015] in, for The weight, The contribution of the collected data to the cloud-based fault prediction model. for The weight, The weighted sum of bandwidth, CPU, and storage space consumed in collecting this data.
[0016] Preferably, step 2 includes the following sub-steps:
[0017] Step A1: Collect data samples from the target device. The data samples are the collection objects and collection frequencies corresponding to various operating states obtained by running the target device N times, where N is a positive integer.
[0018] Step A2: Input the data samples into the pre-trained model, and iteratively adjust the pre-trained model using the gradient descent method to obtain the data acquisition strategy model;
[0019] Step A3: The acquisition strategy model generates an acquisition strategy mapping table based on the various operating states of the target device. The acquisition strategy mapping table contains the acquisition objects, acquisition frequencies, and thresholds corresponding to various operating states.
[0020] Step A4: Store the acquisition strategy mapping table in the acquisition strategy library.
[0021] Preferably, in step 3, the current operating status of the target device is obtained in the following way:
[0022] Method 1: Obtain the control signals from the controller of the target device, and analyze the control signals to obtain the current operating status of the target device;
[0023] Method 2: When the controller's control signal cannot be obtained, obtain the sensor data of the target device and deduce the current operating status of the target device from the sensor data.
[0024] Preferably, step 4 specifically involves: obtaining the acquisition objects and corresponding thresholds in the acquisition strategy mapping table; the acquisition objects are various sensors; constructing quantum filtering rules based on the acquisition objects and corresponding thresholds; using the quantum filtering rules to filter the target data to obtain reported data; the reported data is used to upload to the prediction model of the peripheral device to predict the state of the target device.
[0025] Preferably, the quantum filtering rules include:
[0026] Rule 1: When the data collection object is a sensor that monitors a specific action, a baseline for the target data is set. The corresponding threshold is 105% to 115% of the baseline. When the target data exceeds the corresponding threshold, the target data is reported as data; otherwise, the target data is discarded.
[0027] Rule 2: When the data to be collected is from a sensor with small fluctuations in monitoring data, if the absolute value of the difference between the current target data and the previous target data is greater than the corresponding threshold, then the target data is reported; otherwise, the target data is discarded.
[0028] Rule 3: When the data collection object is a sensor that monitors changes in its state, if the target data is greater than the corresponding threshold, the target data will be reported as data; otherwise, the target data will be discarded.
[0029] Preferably, the method further includes step 5: uploading the reported data to the prediction model of the peripheral device to predict the status of the target device, calculating the resource cost consumed by the reported data, and the prediction model of the peripheral device outputs feedback information. Based on the resource cost consumed and the feedback information, the gradient descent method is used to adaptively update the acquisition strategy model.
[0030] In addition, an edge router rapid acquisition decision system is provided to implement the above-mentioned edge router rapid acquisition decision method, including the following modules:
[0031] Model loading module: used to load pre-trained models, which are obtained by cloud servers constructing meta-models based on meta-learning algorithms and training the meta-models using historical task data from various devices;
[0032] Model adaptation module: used to collect data samples from the target device, iteratively adjust the pre-trained model using the data samples to obtain the acquisition strategy model, and output the acquisition strategy mapping table, which contains the acquisition objects, acquisition frequencies and thresholds corresponding to various operating states;
[0033] Data acquisition module: used to obtain the current operating status of the target device, map the current operating status information to the acquisition strategy mapping table, filter out the best acquisition strategy, and then acquire the sensor data of the target device according to the best acquisition strategy to obtain the target data;
[0034] Data filtering module: Used to construct quantum filtering rules based on the acquisition strategy mapping table, and filter the target data through the quantum filtering rules to obtain the reported data.
[0035] One of the above-mentioned technical solutions in this application has at least one of the following advantages or beneficial effects:
[0036] The rapid acquisition decision method of this application constructs a pre-trained model in a cloud server using a meta-learning algorithm. Through the meta-learning algorithm, the pre-trained model can learn to learn. After receiving a small number of samples from the target device, it can obtain a personalized acquisition strategy model for the target device. In this way, when the edge router faces a new device or a new task, it can quickly generate an acquisition strategy with only a small number of samples, achieving "small sample, fast adaptation" and improving the adaptability of the edge router.
[0037] Furthermore, by constructing a data acquisition strategy model through pre-trained models, edge routers can intelligently determine the optimal data acquisition strategy based on data value and task objectives, thereby improving the intelligence of edge routers.
[0038] Secondly, by performing intelligent data collection and filtering at the edge router level, the edge router is transformed from a "data relay station" into a "data gatekeeper," ensuring that all data uploaded to the remote end is high-quality and high-value, reducing the amount of data processing required by the cloud and improving resource utilization. Furthermore, experiments have shown that processing massive amounts of data at the edge router before inputting it into the cloud-based predictive model not only reduces the network resources required for data upload but also improves the accuracy of the predictive model. Attached Figure Description
[0039] The present application will be further described below with reference to the accompanying drawings and embodiments;
[0040] Figure 1 This is a flowchart of the edge router fast sampling decision method in Example 1;
[0041] Figure 2 The ROC curves are used to compare the performance of the elevator cable fatigue prediction models in Example 1.
[0042] Figure 3 This is a block diagram of the edge router rapid acquisition decision system in Example 2. Detailed Implementation
[0043] The embodiments of this application are described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain this application, and should not be construed as limiting this application.
[0044] The following disclosure provides many different implementation methods or examples for different schemes of implementing this application.
[0045] refer to Figure 1-2 An edge router fast sampling decision method includes the following steps:
[0046] Step 1: Load the pre-trained model. The pre-trained model is obtained by the cloud server constructing a meta-model based on the meta-learning algorithm and training the meta-model using historical task data from multiple devices.
[0047] In this embodiment, the method of this application is applied in an edge router. Taking a newly installed elevator for intelligent data collection as an example, the purpose is to realize fault warning and predictive maintenance of the elevator.
[0048] The edge router loads the pre-trained model from the cloud server.
[0049] It should be noted that the pre-trained model is pre-built and trained on a cloud server, enabling the pre-trained model to be trained a second time based on a small number of data samples from the target device, thereby transforming the pre-trained model into a data acquisition strategy model for that target device.
[0050] In this embodiment, since an elevator is used as an example, various devices include low-speed elevators, high-speed elevators, and sightseeing elevators. Historical task data includes historical operating states and corresponding collected data. Historical operating states include stationary, accelerating upward, uniform downward, leveling, door opening / closing, and peak / off-peak periods. The corresponding collected data are the data collected in different historical operating states.
[0051] The training process of the pre-trained model is as follows: First, the cloud server constructs a meta-model based on the meta-learning algorithm. The historical operating states of various devices and the corresponding collected data are used as the training set and input into the meta-model. The meta-model learns and trains on the historical operating states and the corresponding collected data. Then, the parameters of the meta-model are tuned by minimizing the policy utility loss function to obtain the pre-trained model. At this point, the pre-trained model can learn how to output the general rules of different collection strategies according to the operating states of different devices, which is to say, "learning to learn".
[0052] The formula for minimizing the policy utility loss function is as follows:
[0053] ;
[0054] in, for The weight, (Data value) refers to the contribution of the collected data to the cloud-based fault prediction model. for The weight, (Resource cost) is the weighted sum of bandwidth, CPU, and storage space consumed in collecting this data.
[0055] In this embodiment, =0.01, =0.001, The AUC value is the performance metric of the prediction model for peripheral devices. The calculation formula is:
[0056] ;
[0057] Where A represents bandwidth usage, B represents CPU usage, and C represents storage usage. , and As the weight value, in this embodiment, =0.5, =0.3, =0.2.
[0058] This loss function allows us to find the optimal balance between accuracy and resource consumption, thereby achieving the highest accuracy of the peripheral prediction model with minimal resource consumption.
[0059] Step 2: Collect data samples from the target device, and use the data samples to iteratively adjust the pre-trained model to obtain the acquisition strategy model. The acquisition strategy model outputs an acquisition strategy mapping table, which contains the acquisition objects, acquisition frequencies, and thresholds corresponding to various operating states.
[0060] In this embodiment, after the edge router loads the pre-trained model, it performs personalized training on the pre-trained model using data samples from the target device, thereby obtaining a collection strategy model for the target device.
[0061] Preferably, step 2 includes the following sub-steps:
[0062] Step A1: Collect data samples from the target device. The data samples are the collection objects and collection frequencies corresponding to various operating states obtained by running the target device N times, where N is a positive integer.
[0063] In this embodiment, N=2, the target device is the newly installed elevator, the controller and sensors of the newly installed elevator (including motor vibration sensor, car accelerometer, door operator current sensor, temperature sensor and steel cable tension sensor) are connected to the edge router, and then the newly installed elevator runs up and down completely twice under no load, and the edge router collects data samples of the elevator running completely twice.
[0064] Step A2: Input the data samples into the pre-trained model, and iteratively adjust the pre-trained model using the gradient descent method to obtain the data acquisition strategy model;
[0065] Then, the data samples from two complete elevator runs are input into the pre-trained model. The pre-trained model is iteratively adjusted using the gradient descent method to obtain the acquisition strategy model. At this point, the acquisition strategy model is unique to the elevator and can generate personalized acquisition strategies based on the elevator. Moreover, the acquisition strategy model only needs a few minutes to complete training, which greatly improves adaptability and efficiency.
[0066] Specifically, the gradient descent method involves calculating the minimum value along the direction of gradient descent, and its iterative formula is as follows: ,in, This represents the current parameter value at the k-th iteration. The updated parameter values at the (k+1)th iteration. The direction of the gradient is negative. The step size is the search step size in the gradient direction. The gradient direction can be obtained by differentiating the loss function. The step size is determined by a linear search algorithm, that is, treating the coordinates of the next point as... The function is then found to satisfy f( minimum value That's all.
[0067] Step A3: The acquisition strategy model generates an acquisition strategy mapping table based on the various operating states of the target device. The acquisition strategy mapping table contains the acquisition objects, acquisition frequencies, and thresholds corresponding to various operating states.
[0068] In this embodiment, the acquisition strategy model generates an acquisition strategy mapping table based on the various operating states of the target device (in this embodiment, the various operating states of the target device include stationary, accelerating upward, uniform descent, and door opening / closing).
[0069] In this embodiment, the specific acquisition strategies in the acquisition strategy mapping table are as follows:
[0070] When the elevator is stationary, data from the motor vibration sensor, temperature sensor, and steel cable tension sensor are collected at a frequency of 1Hz.
[0071] During the elevator's accelerated ascent or uniform descent, data from the motor vibration sensor, temperature sensor, steel cable tension sensor, and car accelerometer are collected at a frequency of 500Hz.
[0072] The elevator door opening and closing status is monitored by collecting data from the door operator current sensor, temperature sensor, and cable tension sensor at a frequency of 100Hz.
[0073] Secondly, based on the type of the collected object and the sample data, the corresponding threshold is output.
[0074] Step A4: Store the acquisition strategy mapping table in the acquisition strategy library.
[0075] The generated acquisition policy mapping table is input into the pre-built acquisition policy library. When the corresponding acquisition policy is needed in the future, there is no need to call the acquisition policy model frequently. You only need to call the corresponding acquisition policy in the acquisition policy library to ensure the low-latency execution capability of the edge router.
[0076] Step 3: Obtain the current operating status of the target device, map the current operating status information to the acquisition strategy mapping table, filter out the best acquisition strategy, and then acquire the sensor data of the target device according to the best acquisition strategy to obtain the target data;
[0077] Preferably, in step 3, the current operating status of the target device is obtained in the following way:
[0078] Method 1: Obtain the control signals from the controller of the target device, and analyze the control signals to obtain the current operating status of the target device;
[0079] Method 2: When the controller's control signal cannot be obtained, obtain the sensor data of the target device and deduce the current operating status of the target device from the sensor data.
[0080] In this embodiment, the current operating state of the target device is obtained by receiving the control signal from the controller of the target device. Since the elevator's states such as stationary, accelerating upward, or descending at a constant speed are all controlled by the controller, it is only necessary to obtain the controller's control signal to know the current operating state of the elevator.
[0081] However, there are also situations where the controller's control signal cannot be directly obtained or the control signal is delayed. In such cases, it is necessary to infer the current state based on the physical characteristics (data changes) of the sensor readings. For example, if the car accelerometer detects an acceleration > 0, it can be determined that the elevator is in an accelerating upward state.
[0082] The system obtains the current operating status of the target device, maps the current operating status information to the acquisition strategy mapping table in the acquisition strategy library, pairs the current operating status with the acquisition strategy mapping table, and selects the optimal acquisition strategy. For example, if the current operating status is accelerating upward, the optimal acquisition strategy is to acquire data from the motor vibration sensor, the steel cable tension sensor, and the car accelerometer at a frequency of 500Hz. Then, based on the optimal acquisition strategy, the corresponding sensor (acquisition object) is called to monitor the elevator and acquire the data from that sensor, which is the target data.
[0083] Step 4: Construct quantum filtering rules based on the acquisition strategy mapping table, and filter the target data using the quantum filtering rules to obtain the reported data.
[0084] Preferably, step 4 specifically involves: obtaining the acquisition objects and corresponding thresholds in the acquisition strategy mapping table; the acquisition objects are various sensors; constructing quantum filtering rules based on the acquisition objects and corresponding thresholds; using the quantum filtering rules to filter the target data to obtain reported data; the reported data is used to upload to the prediction model of the peripheral device to predict the state of the target device.
[0085] Preferably, the quantum filtering rules include:
[0086] Rule 1: When the data collection object is a sensor that monitors a specific action, a baseline for the target data is set. The corresponding threshold is 105% to 115% of the baseline. When the target data exceeds the corresponding threshold, the target data is reported as data; otherwise, the target data is discarded.
[0087] For example, consider a door operator current sensor. This sensor monitors the current during door opening and closing. The current for opening and closing has a corresponding design value, or baseline. Therefore, the corresponding threshold is 110% of the baseline. This means that only when the door operator current sensor data exceeds the baseline by 10% is the target data reported. Thus, rule one applies to data with a design value (meaning that the value has a clearly defined numerical value during normal operation).
[0088] Rule 2: If the absolute value of the difference between the current target data and the previous reported data is greater than the corresponding threshold, then the target data will be reported; otherwise, the target data will be discarded.
[0089] For example, temperature sensors monitor the temperature of core elevator components. As the elevator's core components are used, their temperature will gradually rise, but will eventually tend to a maximum value. Therefore, based on this situation, the current data can be compared with the previous data. As long as the fluctuation is not large, it proves that the temperature is normal.
[0090] Taking specific data as an example, the corresponding threshold is 0.5℃. If the previously reported data was 25℃, then subsequent readings of 25.1℃, 25.3℃, 24.8℃, etc., will be filtered out until the data reaches or exceeds 25.5℃, or reaches or falls below 24.5℃, at which point the new data will be reported. Rule two is for data with gradual changes.
[0091] Rule 3: When the data collection object is a sensor that monitors changes in its state, if the target data is greater than the corresponding threshold, the target data will be reported as data; otherwise, the target data will be discarded.
[0092] For example, the car accelerometer primarily acquires data on the elevator's acceleration, which reflects changes in its state. During periods of rest or constant speed, the car accelerometer data is zero, resulting in a large amount of redundant data. Therefore, regarding car accelerometer data, when the absolute value of the acceleration data exceeds 0.1 m / s²,... 2 Only then will the target data be used as the reported data, filtering out acceleration data with an absolute value less than 0.1 m / s². 2 The data is insufficient because the acceleration data from the car accelerometer is less than 0.1 m / s². 2 The data provided is not helpful for early warning and predictive maintenance of target equipment, thus ensuring the quality of the reported data. Rule 3 applies only to data exceeding a certain absolute value.
[0093] This approach, combining proactive data collection and intelligent filtering, not only reduces the amount of data to be reported but also improves the quality of the reported data.
[0094] Step 5: Upload the reported data to the prediction model of the peripheral device to predict the status of the target device, calculate the resource cost consumed by this data report, and the prediction model of the peripheral device outputs feedback information. Based on the resource cost consumed and the feedback information, the gradient descent method is used to adaptively update the acquisition strategy model.
[0095] After the data is uploaded, the resource cost consumed in this data upload can be calculated (i.e., After the peripheral device's prediction model makes a prediction, it outputs feedback information. Then, the consumed resource cost and feedback information are input back into the acquisition strategy model, which is then adaptively updated using the gradient descent method. For example, if the feedback information indicates "missed critical vibration," the acquisition strategy model will automatically lower the corresponding threshold or increase the sampling frequency. Alternatively, if the resource cost consumed is too high, a cost warning signal will be generated, and the acquisition strategy model will automatically adjust the corresponding parameters. The acquisition strategy model then updates the acquisition strategy mapping table, and the updated acquisition strategy takes effect immediately, entering the next cycle.
[0096] In this embodiment, the prediction model of the peripheral device is an elevator cable fatigue prediction model. The method of this application reduces the data transmission volume from the elevator machine room to the cloud by approximately 90% compared to the traditional method (direct transmission of raw data). More importantly, due to the extremely high value density of the collected data, the performance of the elevator cable fatigue prediction model in the cloud is significantly improved, with an accuracy increase of approximately 15%. Figure 2 As shown.
[0097] Example 2
[0098] refer to Figure 3 An edge router rapid acquisition decision system, used to implement the above-mentioned edge router rapid acquisition decision method, includes the following modules:
[0099] Model loading module: used to load pre-trained models, which are obtained by cloud servers constructing meta-models based on meta-learning algorithms and training the meta-models using historical task data from various devices;
[0100] Model adaptation module: used to collect data samples from the target device, iteratively adjust the pre-trained model using the data samples to obtain the acquisition strategy model, and output the acquisition strategy mapping table, which contains the acquisition objects, acquisition frequencies and thresholds corresponding to various operating states;
[0101] Data acquisition module: used to obtain the current operating status of the target device, map the current operating status information to the acquisition strategy mapping table, filter out the best acquisition strategy, and then acquire the sensor data of the target device according to the best acquisition strategy to obtain the target data;
[0102] Data filtering module: Used to construct quantum filtering rules based on the acquisition strategy mapping table, and filter the target data through the quantum filtering rules to obtain the reported data.
[0103] In this embodiment, the specific workflow of the decision-making system is as follows: the model loading module loads the pre-trained model from the cloud server and then sends the pre-trained model to the model adaptation module;
[0104] The model adaptation module collects data samples obtained from two complete runs of the target device, inputs the data samples as the training set into the pre-trained model, trains the pre-trained model, and fine-tunes the pre-trained model using the gradient descent method to obtain the acquisition strategy model. The acquisition strategy model outputs an acquisition strategy mapping table, which is then sent to the data acquisition module and the data filtering module respectively.
[0105] The data acquisition module obtains the current operating status of the target device, finds the acquisition strategy corresponding to the current operating status in the acquisition strategy mapping table, and obtains the optimal acquisition strategy. Based on the optimal acquisition strategy, it acquires the sensor data of the target device to obtain the target data, and sends the target data to the data filtering module.
[0106] The data filtering module constructs quantum filtering rules based on the acquisition strategy mapping table, determines which quantum filtering rule the target data belongs to, filters it according to the rule, obtains the reported data, and sends the reported data to the cloud.
[0107] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.
Claims
1. A method for edge router fast sampling decision-making, characterized in that, Includes the following steps: Step 1: Load the pre-trained model. The pre-trained model is obtained by the cloud server constructing a meta-model based on the meta-learning algorithm and training the meta-model using historical task data from multiple devices. Step 2: Collect data samples from the target device, and use the data samples to iteratively adjust the pre-trained model to obtain the acquisition strategy model. The acquisition strategy model outputs an acquisition strategy mapping table, which contains the acquisition objects, acquisition frequencies, and thresholds corresponding to various operating states. Step 3: Obtain the current operating status of the target device, map the current operating status information to the acquisition strategy mapping table, filter out the best acquisition strategy, and then acquire the sensor data of the target device according to the best acquisition strategy to obtain the target data; Step 4: Construct quantum filtering rules based on the acquisition strategy mapping table, and filter the target data using the quantum filtering rules to obtain the reported data.
2. The edge router fast sampling decision method according to claim 1, characterized in that, Historical task data includes historical operating status and corresponding collected data; in step 1, the training process of the pre-trained model is as follows: the cloud server constructs a meta-model based on the meta-learning algorithm, uses the historical operating status and corresponding collected data of various devices as the training set to train the meta-model, and then optimizes the parameters of the meta-model by minimizing the policy utility loss function to obtain the pre-trained model.
3. The edge router fast sampling decision method according to claim 2, characterized in that, The formula for the minimized policy utility loss function is: ； in, for The weight, The contribution of the collected data to the cloud-based fault prediction model. for The weight, The weighted sum of bandwidth, CPU, and storage space consumed in collecting this data.
4. The edge router fast sampling decision method according to claim 1, characterized in that, Step 2 includes the following sub-steps: Step A1: Collect data samples from the target device. The data samples are the collection objects and collection frequencies corresponding to various operating states obtained by running the target device N times, where N is a positive integer. Step A2: Input the data samples into the pre-trained model, and iteratively adjust the pre-trained model using the gradient descent method to obtain the data acquisition strategy model; Step A3: The acquisition strategy model generates an acquisition strategy mapping table based on the various operating states of the target device. The acquisition strategy mapping table contains the acquisition objects, acquisition frequencies, and thresholds corresponding to various operating states. Step A4: Store the acquisition strategy mapping table in the acquisition strategy library.
5. The edge router fast sampling decision method according to claim 1, characterized in that, In step 3, the current operating status of the target device is obtained in the following way: Method 1: Obtain the control signals from the controller of the target device, and analyze the control signals to obtain the current operating status of the target device; Method 2: When the controller's control signal cannot be obtained, obtain the sensor data of the target device and deduce the current operating status of the target device from the sensor data.
6. The edge router fast sampling decision method according to claim 1, characterized in that, Step 4 specifically involves: obtaining the acquisition objects and corresponding thresholds from the acquisition strategy mapping table; the acquisition objects are various sensors; constructing quantum filtering rules based on the acquisition objects and corresponding thresholds; using the quantum filtering rules to filter the target data to obtain reported data; the reported data is used to upload to the prediction model of the peripheral device to predict the state of the target device.
7. The edge router fast sampling decision method according to claim 6, characterized in that, The quantum filtering rules include: Rule 1: When the data collection object is a sensor that monitors a specific action, a baseline for the target data is set. The corresponding threshold is 105% to 115% of the baseline. When the target data exceeds the corresponding threshold, the target data is reported as data; otherwise, the target data is discarded. Rule 2: When the data to be collected is from a sensor with small fluctuations in monitoring data, if the absolute value of the difference between the current target data and the previous reported data is greater than the corresponding threshold, then the target data will be reported; otherwise, the target data will be discarded. Rule 3: When the data collection object is a sensor that monitors changes in its state, if the target data is greater than the corresponding threshold, the target data will be reported as data; otherwise, the target data will be discarded.
8. The edge router fast sampling decision method according to claim 1, characterized in that, It also includes step 5: uploading the reported data to the prediction model of the peripheral device to predict the status of the target device, calculating the resource cost consumed by this reported data, and the prediction model of the peripheral device outputs feedback information. Based on the resource cost consumed and the feedback information, the gradient descent method is used to adaptively update the acquisition strategy model.
9. An edge router rapid acquisition decision system, characterized in that, The edge router fast sampling decision method for implementing any one of claims 1-8 includes the following modules: Model loading module: used to load pre-trained models, which are obtained by cloud servers constructing meta-models based on meta-learning algorithms and training the meta-models using historical task data from various devices; Model adaptation module: used to collect data samples from the target device, iteratively adjust the pre-trained model using the data samples to obtain the acquisition strategy model, and output the acquisition strategy mapping table, which contains the acquisition objects, acquisition frequencies and thresholds corresponding to various operating states; Data acquisition module: used to obtain the current operating status of the target device, map the current operating status information to the acquisition strategy mapping table, filter out the best acquisition strategy, and then acquire the sensor data of the target device according to the best acquisition strategy to obtain the target data; Data filtering module: Used to construct quantum filtering rules based on the acquisition strategy mapping table, and filter the target data through the quantum filtering rules to obtain the reported data.