Method and system for adjusting carrier frequency point, carrier bandwidth and like

Abstract

The invention provides a method and a system for adjusting carrier frequency points, carrier bandwidth and the like. Performing multi-stage compression optimization on the original model through deep learning pruning or knowledge distillation to obtain a model after multi-stage compression optimization; and the like. The sensor calibration method and system have the capability of being deployed in the sensor terminal device, the base station and the cloud server respectively, application in multiple types of sensor devices and multiple scenes of the sensor devices is achieved, and simple and efficient inspection of calibration results is achieved through multi-stage cooperative calibration.

Description

CROSS-REFERENCE TO RELATED APPLICATIONS

[0001] This application claims priority to U.S. patent applications with application number US 63 / 239420, filed on September 1, 2021, and invention name “Awireless system”, application number US 63 / 240965, filed on September 5, 2021, and invention name “Awireless system”, application number US 63 / 325613, filed on March 31, 2022, and invention name “IOTNetworks”, and application number US 63 / 353816, filed on June 20, 2022, and invention name “An IOT System”; at the same time, this application claims priority to Chinese patent application with application number CN202210571576.8, filed on May 24, 2022, and invention name “Internet of Things Data Application and Deep Learning Method”, all of which are incorporated herein by reference. Technical Field

[0002] The present disclosure involves the Internet of Things system and the multiple layers it covers, including the terminal layer, transmission layer, support layer, artificial intelligence business platform layer and urban operation comprehensive IOC layer; it also includes: security management platform, unified operation and maintenance management platform and IT resource services. Specifically, it involves industry terminals, edge computing, data intelligent fusion, artificial intelligence, streaming media, blockchain security management, digital twins, converged communications, intelligent inspection, unified operation and maintenance, cloud management and other technologies. Background Art

[0003] At present, the Internet of Things technology has not yet achieved the interconnection of all things, there is no effective integration between data, and no data warehouse that can be used for the entire industry has been formed. In addition, the Internet of Things systems in related technologies still have specific problems such as high latency, high power consumption, incomplete network coverage, low data carrying capacity, different communication protocols, and insecure data. There is no effective allocation of communication resources in the application terminals. These are obstacles to the development of the Internet of Things technology field.

[0004] The integration of the Internet of Things and vertical industries will be a comprehensive network and scenario with multiple devices, multiple networks, multiple applications, interconnection and mutual integration. The standardization of device interface standards, communication protocols, and management protocols is a systematic technological innovation work. Only by solving the above problems can the Internet of Things technology be promoted and applied.

[0005] The key to the aforementioned integration is to apply IoT technology to vertical industries and effectively solve problems to achieve the "Internet of Everything", that is, to rely on the scenarios applicable to the IoT and effectively solve practical problems. At present, the "data islands" and "industry chimneys" that have emerged in this field, from a technical point of view, mainly have the following problems to be solved: 1. In industrial applications, there are problems such as unreliable network connections (especially in remote areas), high operating and maintenance costs, inconsistent access technologies, and insecure industry data. There is a lack of a ubiquitous, dynamic, and real-time network, and the problem of "Internet of Everything" needs to be solved; 2. Terminals of different manufacturers and models with different communication protocols, access authentication methods, network bandwidth requirements, and application protocols lack a safe and unified access method to access the network, and the problem of "ubiquitous access and unified management" needs to be solved; 3. There is a lack of a unified security protection system, including the access security issues of various types of IoT edge devices, and multi-mode The security protection issues of diversified transmission channels, the potential safety hazards brought about by multi-scenario business coupling, data sharing and data interaction require solutions to the "Internet of Things security" issue; 4. The lack of a data fusion platform requires the unified collection, aggregation, data standardization, storage management, analysis and mining, fusion algorithms, on-demand services and other core functions of multi-source data, which require solutions to the "information islands and application chimneys" problems between smart applications; 5. The "end-cloud" technical solution based on cloud computing can effectively utilize the powerful computing and storage resources of the cloud, but it is difficult to meet the low latency requirements of many real-time applications, and new technical solutions are urgently needed; 6. Intelligent AI technology needs to be able to quickly meet the comprehensive management of artificial intelligence models for classification, clustering, prediction, and correlation analysis of smart applications. Summary of the invention

[0006] The Internet of Things system or industrial Internet system provided by the present disclosure is built for the smart twin / smart empowerment of various industries, covering multiple levels. The overall system can be divided into five horizontal and three vertical layers. The five horizontal layers are terminal layer, transmission layer, support layer, artificial intelligence business platform layer, and comprehensive IOC layer for urban operations from bottom to top. The three vertical layers are security, operation and maintenance, and IT resource services. Among them, security and operation and maintenance run vertically through all horizontal layers, providing full-chain, end-to-end services; IT resource services provide services for the support layer, artificial intelligence business platform layer, and comprehensive IOC layer for urban operations.

[0007] The terminal layer includes thousands of different types of sensing, linkage, mobile, video and other terminals from different industries.

[0008] Perception terminals can ubiquitously, in real time and dynamically perceive the multi-dimensional status of the city, such as water, gas, electricity, soil, sound, fire, etc. The perception data is uploaded to the central platform through a communication network that dynamically adjusts any communication parameters according to industry requirements and / or physical location.

[0009] The linkage terminal can realize the linkage from edge perception to execution based on the communication network that dynamically adjusts any communication parameters according to industry requirements and / or physical location, such as linkage alarm, linkage shouting, linkage control valve / door, linkage SMS / email notification, etc.

[0010] Fusion communication terminals provide transmission and interconnection that is dynamically adjusted according to industry requirements and / or physical location for sensors that do not have communication transmission capabilities. It supports composite sensing technology, multi-sensor data fusion, and unified access to sensing devices from different manufacturers. Sensing technology combined with edge computing technology realizes edge correction and self-correction of sensing data, and derives optimized sampling strategies, such as dynamically changing the sampling interval, sampling accuracy, and transmission frequency according to the rate of change of sensing data, preset thresholds, network conditions, etc., so that response time, overall machine power consumption, and network bandwidth occupancy can be taken into account at the same time.

[0011] The following describes a terminal device sensor in the general inventive concept, and the inventive concept of the Internet of Things or Industrial Internet system, wherein a sensor calibration method and system are provided.

[0012] In order to extend the service life of sensors, reduce maintenance costs, and improve sensor data accuracy and sensitivity analysis, the present invention uses a deep learning calibration algorithm to train the accurate values ​​of historical data reported by sensors and at least part of the corresponding historical data collected by standard sensors to obtain an original model, and combines the computing power characteristics of sensor terminal devices, base stations, and cloud servers to perform deep learning pruning or knowledge distillation on the trained original model, so as to achieve a balance between the ratio of accuracy and response speed.

[0013] The calibration algorithm based on deep learning in the present disclosure adopts a transformer model (Transformer model) based on a multi-head attention mechanism (Multi-HeadAttention). The Transformer model is an Encoder-Decoder model that is completely based on the attention mechanism. The original model obtained is further subjected to deep learning pruning or knowledge distillation, so that the model can be compressed and optimized without significantly reducing the accuracy, so that it has the ability to be deployed in sensor terminal devices, base stations and cloud servers respectively.

[0014] The technical problem solved by the present disclosure is to provide a sensor calibration method and system thereof, which realizes hierarchical, efficient, intelligent calibration and multi-level collaborative calibration of the sensor.

[0015] In this context, embodiments of the present disclosure are intended to provide a method and system for calibrating a sensor based on deep learning.

[0016] The present disclosure provides a sensor calibration method based on deep learning, the calibration method comprising the following steps:

[0017] Sensors collect historical data in chronological order;

[0018] Acquiring accurate values ​​of at least part of the corresponding historical data by means of standard sensors;

[0019] providing the historical data and the accurate value to a deformer model;

[0020] The deformer model trains the historical data and the accurate value to obtain an original model;

[0021] Performing multi-level compression optimization on the original model through deep learning pruning or knowledge distillation to obtain a multi-level compressed optimized model;

[0022] The original data subsequently collected by the sensor is calibrated according to the original model or the model after multi-stage compression optimization.

[0023] The present disclosure also provides a sensor calibration system based on deep learning, the calibration system comprising: a sensor, which is used to collect historical data in chronological order; a standard sensor, which is used to collect accurate values ​​of at least part of the corresponding historical data; a training device, which is used to receive the historical data and the accurate values, and train the historical data and the accurate values ​​to obtain an original model; a compression optimization device, which is used to perform multi-level compression optimization on the original model through deep learning pruning or knowledge distillation to obtain a multi-level compression optimized model; a calibration device, which calibrates the original data collected by the sensor afterwards according to the original model or the multi-level compression optimized model.

[0024] The present disclosure also provides a deep learning processing method, the processing method comprising the following steps: collecting historical data in chronological order; collecting accurate values ​​of at least part of the corresponding historical data; providing the historical data and the accurate values ​​to a deformer model; the deformer model trains the historical data and the accurate values ​​to obtain an original model; performing multi-level compression optimization on the original model through deep learning pruning or knowledge distillation to obtain a multi-level compressed optimized model, deploying the first-level compressed optimized model obtained through knowledge distillation on a terminal device, deploying the second-level compressed optimized model obtained through deep learning pruning on a base station, and deploying the original model on The cloud server, the processing accuracy of the model after the secondary compression optimization is higher than the processing accuracy of the model after the primary compression optimization, and lower than the processing accuracy of the original model, the response speed of the model after the secondary compression optimization is lower than the response speed of the model after the primary compression optimization, and higher than the response speed of the original model, the data calculation amount of the model after the secondary compression optimization is higher than the data calculation amount of the model after the primary compression optimization, and lower than the data calculation amount of the original model; according to the processing accuracy requirements, response speed and / or data calculation amount, determine to use the deployed model in the terminal device, base station or cloud server to process the raw data collected later.

[0025] The present disclosure also provides an application of a sensor calibration method based on deep learning, the application comprising the following steps: the sensor collects raw data in real time; the standard sensor collects accurate values ​​corresponding to the raw data in real time;

[0026] Extracting a certain amount of the raw data and the accurate value corresponding thereto according to the sampling rate, and uploading the extracted raw data and the accurate value to the base station; the base station compares the certain amount of the raw data and the accurate value corresponding thereto;

[0027] If the difference between the two is greater than a certain accuracy threshold and the proportion is less than the ratio threshold, the sensor is marked as a sensor in a normal state, all the raw data is accepted and uploaded to the cloud server.

[0028] According to the sensor calibration method based on deep learning and the system thereof provided by the present disclosure, the following are achieved: first, a Transformer model based on a multi-head attention mechanism is adopted. This deep learning model can not only effectively learn and imitate the characteristics of time series data, but also use the multi-head attention mechanism to help the Transformer model capture richer sensor characteristics and information, and further perform comprehensive processing on the captured sensor characteristics and information. It can not only learn the data characteristics of sensors of the same type, but also learn to discover the data correlation between sensors of different types, alarm for abnormal values, and has extremely strong filtering capabilities, thus realizing the application in multiple types of sensor devices; second, hierarchical calibration, deep learning pruning or knowledge distillation is performed on the trained original model, so that the model can be compressed and optimized without significantly reducing the accuracy, so that it has the ability to be deployed in sensor terminal devices, base stations and cloud servers respectively, combined with the computing power characteristics of sensor terminal devices, base stations and cloud servers, Intelligently match the calibration position to achieve a balance between the ratio of accuracy and response speed, perform low-precision fast calibration on sensor terminal devices with weak computing power, and perform high-precision calibration on cloud servers with strong computing power, realizing application in multiple scenarios; Third, multi-level collaborative calibration, upload low-level calibration results and raw data to high-level equipment or environment, and perform high-level calibration on at least part of the raw data in the high-level equipment or environment, such as performing a first-level calibration on the raw data in the sensor terminal device, and uploading the obtained first-level calibrated data and raw data to the base station, and performing a second-level calibration on at least part of the raw data at the base station to obtain at least part of the second-level calibrated data, and compare the at least part of the second-level calibrated data with the corresponding first-level calibrated data. If the difference between the two is less than a certain error threshold, all the first-level calibrated data is accepted, otherwise, the received raw data is calibrated by the model optimized by the second-level compression to obtain all the second-level calibrated data. Multi-level collaborative calibration can use multi-level calibration models with different precisions to calibrate part of the raw data, spot check whether the calibration results reported by the low-level calibration model are qualified, and realize simple and efficient inspection of the received calibration results. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 It is the Internet of Things composition diagram;

[0030] Figure 2 It is the global relationship diagram of the Internet of Things;

[0031] Figure 3 This is an example diagram of IoT business flow;

[0032] Figure 4 is a flowchart of an implementation of a method for calibrating a sensor based on deep learning according to the present application;

[0033] Figure 5 It is a schematic diagram of hierarchical deployment after different processing of the original model;

[0034] Figure 6 is the sensor network topology diagram;

[0035] Figure 7 This is the flowchart of the Transformer model training the original model;

[0036] Figure 8 is a flow chart of a multi-stage collaborative calibration method for a sensor based on deep learning according to the present application;

[0037] Fig. 9 This is the flowchart of retraining the updated original model;

[0038] Fig.10 This is a diagram of the relationship between temperature and humidity learned by the Transformer model;

[0039] Fig.11 is a block diagram of a sensor calibration system based on deep learning according to the present application;

[0040] Fig.12 is a structural block diagram of the compression optimization device 940;

[0041] Fig.13 is a block diagram of a calibration device 950;

[0042] Fig.14 is a flow chart of an exemplary application of the deep learning-based calibration method to the calibration of sensors of the same type;

[0043] Fig.15 It is a flow chart for the use of adaptive network topology numerical calibration of LSTM neural network;

[0044] Fig.16 It is a composition diagram of an adaptive multi-point collaboration system with multiple gateways and multiple terminals coexisting;

[0045] Fig.17 This is a flow chart of terminal data calibration technology based on edge computing mode;

[0046] Fig.18 This is a technical flow chart of dynamically adjusting sensor coefficients based on edge computing mode;

[0047] Fig.19 This is a technical flow chart of the interval control of terminal device data reporting based on the cloud;

[0048] Fig. 20 It is the terminal data reporting flow chart; DETAILED DESCRIPTION

[0049] The present invention will be further described below with reference to the accompanying drawings and embodiments:

[0050] Please also read Figure 1 and Figure 2 The terminal layer contains thousands of different types of sensing, linkage, multi-mode heterogeneous communication, mobile and / or video terminals in different industries. The sensing terminals can ubiquitously, in real time and dynamically sense the multi-dimensional state of the city, such as water, gas, electricity, soil, sound, fire, etc. The sensing data is uploaded to the central platform through a multi-mode heterogeneous network that dynamically adjusts any communication parameters according to industry requirements and / or physical locations. Furthermore, the sensing terminal can be a variety of sensors with data acquisition functions or electronic devices with sensors, for example, temperature sensors, smoke sensors, atmospheric pressure sensors, sound wave sensors, image sensors, cameras, etc.

[0051] In this embodiment, the linkage terminal can realize the linkage from edge perception to execution based on a multi-mode heterogeneous network that dynamically adjusts any communication parameters according to industry requirements or / and physical location, such as linkage alarm, linkage shouting, linkage control valve / door, linkage SMS / email notification, etc. Multi-mode heterogeneous communication terminals provide transmission and interconnection that are dynamically adjusted according to industry requirements and / or physical location for sensors that do not have communication transmission capabilities. It supports composite perception technology, multi-sensor data fusion, and supports unified access to perception devices from different manufacturers. Perception technology combined with edge computing technology realizes edge correction and self-correction of perception data, and derives optimized sampling strategies, such as dynamically changing the sampling interval, sampling accuracy, and sending frequency according to the rate of change of perception data, preset thresholds, network conditions, etc., so that response time, overall machine power consumption, and network bandwidth occupancy can be taken into account at the same time. Further, the edge computing method includes: collecting data by several sensing terminals; judging whether the data collected by the sensing terminal is abnormal; when abnormal, the first device connected to the sensing terminal generates a first alarm message, and sends the first alarm message to all second devices connected to the first device; the second device sends a second alarm message to all alarm devices connected to the second device. Among them, the first device and the second device can be edge devices or intermediate devices. In this embodiment, the edge device can be used for data packet transmission between access devices and core / backbone network devices, and can be switches, routers, routing switches, gateways, IADs, and various MAN / WAN devices installed on the edge network. With the addition of edge computing, the data collected by the sensing terminal does not need to shuttle between the local and central servers, so that the local device can know which function to execute. This can save operating costs and storage device investment. In addition, the information that needs to be uploaded by the communication terminal (such as a sensing terminal or a gateway or a base station) and its matching communication and network are judged by edge computing or fog computing or an algorithm platform. The information may include but is not limited to data difference, feature value and / or feature value of the image and video collected by the sensing terminal. It should be understood that the matching communications and networks are dynamically allocated and output by the multi-mode heterogeneous network. For example, when the data difference of the data collected by the perception terminal exceeds the threshold range, or meets the specific image, video feature value, or meets the specific sound wave feature, it can be determined that the collected data is abnormal, and then an alarm message is generated and other further operations are performed (at the same time, the multi-mode heterogeneous network dynamically allocates network and communication resources). When the collected data is not abnormal, the perception terminal can reduce the frequency of collecting data and transmit it through the matching communications and networks, which reduces the waste of communication resources and saves the energy of the communication terminal.

[0052] refer to Figure 2 , the present invention provides an example of global relationships for the next generation of the Internet of Things.

[0053] The present invention provides an Internet of Things platform, which comprises: a terminal layer, a communication layer, a support layer and an application layer which are arranged in sequence from bottom to top.

[0054] In the terminal layer, the terminal participates in forming the network as a node of the communication network, obtains network status information, determines the perception and execution strategy according to the network status information, and adjusts the sampling mode according to the business needs and the communication environment to obtain the perception data. The sampling mode includes: sampling frequency, sampling accuracy, source coding and / or compression mode. As an embodiment, the air quality monitoring station for dust monitoring at the construction site uses a wireless gateway to transmit data back, in which a particle sensor is used to evaluate the dust situation. High-frequency sampling data can obtain a more detailed dust change curve, but if the system deploys multiple different types of terminals and the gateway capacity is limited, the sampling and return frequency can be appropriately reduced to reduce the communication resource occupation. When there are construction site operations during the day, a slightly higher sampling frequency should be used. If the wind speed is large and the dust diffusion changes quickly, a higher sampling frequency should be used to adapt to accurate change tracking. At night, when there is no construction at the construction site, the sampling frequency, sampling accuracy and return frequency can be further reduced. If the dust change is small, the source coding can be changed to a difference mode, that is, using fewer data bits to represent the difference from the previous value, or multiple sampling data can be accumulated and compressed using a compression algorithm and sent uniformly. Fig.19 This is the technical flow chart of the terminal equipment data reporting interval control. Fig. 20 This is the flow chart of terminal data reporting in compression mode. Fig.18 It is a technical flow chart of the terminal dynamically adjusting the sensor coefficient based on the data reported by the surrounding terminals.

[0055] In some embodiments, in the terminal layer, the terminal can participate in the establishment of the network as a node of the communication network. The terminal can obtain network status information and then change the perception or execution strategy. The terminal performs edge computing according to business needs and the communication environment to adjust parameters such as sampling frequency, sampling accuracy, source coding, and compression method. Fig.19 The composition principle of the adaptive multi-point cooperation to select the best gateway communication by the terminal to obtain network status information is shown. Fig. 20 The terminal calibration flow chart based on the surrounding terminal data is shown. Fig.18The flowchart of the terminal dynamically adjusting the data reporting frequency is shown. The terminal layer adds edge computing function, and the perception data can make local decisions on the edge side, thereby directly linking the execution terminal; due to the existence of edge computing, it is possible to integrate multiple perception device data for more complex and high-level decisions; the results of edge computing can enhance the importance and urgency of the corresponding perception data, and through the dynamic adjustment of communication parameters and network transmission rules, the terminal can obtain faster, more reliable and more suitable communication and network resources. When the perception data changes significantly and needs continuous attention, more network resources can be applied to the communication layer and the support layer through data transmission as needed, while increasing the sampling frequency and accuracy. The multi-mode heterogeneous network realizes resource allocation through dynamic adjustment of communication. The dynamically adjustable factors include carrier frequency, carrier bandwidth, modulation mode, channel coding, transmission power, receiving sensitivity, etc. As an embodiment, in the urban fire protection application, multiple smoke sensor terminals are installed on each floor of a building. In addition, the terminal has temperature recognition capability and uses one or more wireless gateways for communication. During normal operation, all terminals have been detected at a certain frequency, but they only send data to the gateway once every few hours to report the device status (such as battery power, temperature, etc.). In order to reduce the power consumption of the terminal, the communication is sent using spread spectrum modulation, low transmission power and medium receiving sensitivity. When a fire occurs on a certain floor, all terminals on that floor will send smoke concentration and temperature data at a faster frequency. These data are added by the algorithm to calculate the spread of the fire and the best escape route. In order to cope with the temporary increase in communication resources, the communication between the gateway and the device opens multiple carrier frequencies and uses the GFSK debugging method that supports higher rates to increase the transmission power and receiving sensitivity to ensure the communication distance.

[0056] As another embodiment, the air quality monitoring station for monitoring dust on construction sites uses a wireless gateway to transmit data back, in which a particulate matter sensor is used to evaluate the dust situation. High-frequency sampling data can obtain a more detailed dust change curve, but if the system deploys multiple different types of terminals and the gateway capacity is limited, then the sampling and return frequency can be appropriately reduced to reduce the use of communication resources. When there are construction site operations during the day, a slightly higher sampling frequency should be used. If the wind speed is high and the dust diffusion changes quickly, a higher sampling frequency should be used to accommodate accurate change tracking. At night, when there is no construction on the construction site, the sampling frequency, sampling accuracy and return frequency can be further reduced. If the dust changes slightly, the source encoding can be changed to a difference method, that is, fewer data bits are used to represent the difference from the previous value. It is also possible to accumulate multiple sampling data and use a compression algorithm to compress and send them uniformly.

[0057] refer to Figure 3 , the present invention provides an example of next generation Internet of Things business flow.

[0058] like Figure 3As shown, the sensing device 3 (represented by sensing 3 in the figure, similarly, the sensing device 2 is represented by sensing 2, and the sensing device 1 is represented by sensing 1) has the edge computing function, which can directly generate edge decisions (represented by edge computing 3 in the figure) through the sensing data to directly link the control 3, and the sensing data is sent to the gateway / base station 1 through the self-organizing network communication mode. In this embodiment, the sensing data collected by the sensing device can be water, gas, electricity, soil, sound, fire, etc. The sensing device can be a mobile terminal such as a handheld device, a walkie-talkie, a vehicle-mounted device, a positioning terminal, a wearable terminal, etc.; the sensing data can also be voice / video / text, etc. For example, the sensing device can be a diversified video sensing terminal such as a camera, thermal imaging, and a hyperspectral device.

[0059] Furthermore, the sensing device 1 obtains the sensing data from the sensing device 2 through point-to-point communication, and fuses the sensing data of the two devices. With the help of edge computing 1, if the specified conditions are met, the sensing strategy is automatically adjusted (such as increasing the sampling frequency, increasing the sampling progress, etc.) to control the sampling behavior of the sensing device 1, and request to adjust the communication parameters and strategies to obtain a higher rate and highly reliable transmission through multi-mode communication, and at the same time, the linkage control 1 performs the corresponding action. Among them, the sensing data collected by the sensing device is uploaded to the edge computing module. The edge computing module can be a part of the terminal, or a different device connected to the terminal in close range communication, such as a device connected to the terminal through ZigBee, Wifi, Bluetooth, etc. The edge computing module forms an edge decision based on the processing of the sensing data. Among them, the formation of the edge decision is based on the data processing of one or more sensing devices. For example, the edge computing module fuses the sensing data of more than one terminal to form an edge decision, Fig. 20 The terminal data calibration technical process based on the edge computing mode is displayed. In some embodiments, the edge computing module obtains processed data based on the processing of the sensed data, such as realizing edge correction and self-correction of the sensed data, and the processed data is further used for uploading or forming edge decisions. The edge decision of the edge computing module includes adjusting the terminal's perception strategy based on the processing of the sensed data, such as dynamically changing the sampling interval, sampling accuracy, and transmission frequency according to the rate of change of the sensed data, preset thresholds, network conditions, etc., so that the response time, overall power consumption, and network bandwidth occupancy can be taken into account at the same time.

[0060] Multi-mode heterogeneous communication terminals provide transmission and interconnection that is dynamically adjusted according to industry requirements and / or physical location for sensors that do not have communication transmission capabilities. It supports composite sensing technology, multi-sensor data fusion, and unified access to sensing devices from different manufacturers. Sensing technology combined with edge computing technology realizes edge correction and self-correction of sensing data, and derives optimized sampling strategies from this. Typical examples include dynamically changing the sampling interval, sampling accuracy, and transmission frequency according to the rate of change of sensing data, preset thresholds, network conditions, etc., so that response time, overall machine power consumption, and network bandwidth occupancy can be taken into account at the same time.

[0061] The existing sensor implementation methods based on the Internet of Things have problems such as single function, inability to learn big data, inability to make decisions and respond quickly, inability to process large amounts of data concurrently, etc. In order to solve the problems existing in the above-mentioned existing sensor implementation methods based on the Internet of Things, the present disclosure provides a variety of sensor implementation methods based on the Internet of Things and devices thereof.

[0062] With the widespread application of sensor technology, the sensor maintenance problem caused by sensor drift during its use also needs to be solved urgently. At present, there are four main maintenance solutions for sensors: one is to conduct aging experiments before leaving the factory to simulate the exposure process of gas sensors in the atmosphere, and then generate compensation algorithms to pre-correct the sensor response to achieve a certain degree of anti-aging and self-calibration after installation; the second is to regularly maintain the micro air station, replace new sensors, and return the original gas sensors to the factory for secondary calibration in the laboratory; the third is the traditional on-site manual calibration method, that is, technicians go to the site to use tools for calibration; the fourth is to use a deep learning model to calibrate the sensor. In the related technology, the deep learning model is a residual neural network based on self-attention. This method uses data augmentation methods to expand data samples. The proposed sensor drift calibration method includes two parts: drift feature extraction and drift calibration, which correspond to the drift feature extraction module and the calibration module respectively. The drift feature extraction module extracts the key drift features of time and frequency hidden in the data at different scales through a multi-scale convolution layer, laying the foundation for the calibration module; the calibration module uses a one-dimensional residual convolutional neural network based on self-attention to effectively use the data correlation between adjacent sensors to compensate for the drift of drift data, and can calibrate the drift of multiple sensors in the sensor group at the same time. The relevant technology only supports the deployment of the calibration model on the cloud server. The system at least includes: an acquisition module, a standard sensor and a cloud server. The acquisition module at least includes a concentration sensor, a humidity sensor, a temperature sensor and an electrical parameter acquisition module for collecting electrical parameters of the concentration sensor. The cloud server at least includes a modeling module and a calibration module. The modeling module establishes a calibration model according to a deep learning algorithm based on the original concentration parameters, humidity parameters, temperature parameters, electrical parameters sent by the acquisition module and the standard data sent by the standard sensor. Exemplarily, a high-dimensional nonlinear model based on the least squares fitting algorithm, or a BP neural network algorithm is used for deep learning to obtain a calibration model. The calibration module calibrates the original concentration parameters sent by the acquisition module based on the calibration model. By incorporating the electrical data of the gas sensor acquisition state into the calibration model as a calibration influencing factor, the calibration rate of the calibration model is improved and more accurate gas concentration parameters are obtained.

[0063] The compensation algorithm generated by the aging simulation before leaving the factory in Solution 1 can extend the service life of the sensor to a certain extent, but due to the unpredictability of changes in atmospheric temperature, humidity, and gas concentration, it is difficult for the compensation algorithm to maintain accurate compensation after the sensor has been used for a long time. The method of returning the sensor to the factory for recalibration or directly replacing it with a new sensor in Solution 2 is costly, inefficient, and very time-consuming. The on-site manual calibration method in Solution 3 is time-consuming and labor-intensive, and various errors will be introduced into the manual sensor calibration due to factors such as measurement methods and personnel operations. The residual neural network of the self-attention mechanism in Solution 4, the high-dimensional nonlinear model of the least squares fitting algorithm, or the BP neural network algorithm need to be improved in terms of the prediction effect of time series; and deep learning, due to its computational complexity or parameter redundancy, limits the deployment of corresponding models on sensor terminal devices and base stations. Solution 4 cannot match the calibration position for the calibration model according to the computational amount, has no hierarchical deployment capability, and only supports the deployment of the calibration model on the cloud server; and it is impossible to calibrate and check outliers for different types of sensors, and is only effective for a single sensor of the same type.

[0064] In response to the above problems with sensors, the present disclosure provides a sensor calibration method and system based on deep learning.

[0065] In the present disclosure, the Transformer model based on the multi-head attention mechanism can not only consider the factors of time series, but also capture richer sensor features and information, learn the data features of sensors of the same type, and learn to discover the data correlation between sensors of different types, and has extremely strong filtering capabilities, which enables the application in multiple types of sensor devices; hierarchical calibration realizes the intelligent matching of calibration positions by the calibration model, achieves a balance between the ratio of accuracy and response speed, and realizes the application in multiple scenarios. Multi-level collaborative calibration uses multi-level calibration models of different accuracy to calibrate part of the original data, spot-checks whether the calibration results reported by the low-level calibration model are qualified, and realizes a simple and efficient inspection of the received calibration results.

[0066] like Figure 4The flowchart shown is an implementation method of the sensor calibration method based on deep learning according to the present application, including the following steps: S101: the sensor collects historical data in chronological order; S102: the value of at least part of the corresponding historical data is collected through a standard sensor; S103: the historical data and the value are provided to the Transformer model; S104: the Transformer model trains the historical data and the value to obtain the original model; S105: the original model is subjected to multi-level compression optimization through deep learning pruning or knowledge distillation to obtain a multi-level compressed optimized model; S106: the original data collected by the sensor thereafter is calibrated according to the original model or the multi-level compressed optimized model.

[0067] Among them, the sensor is included in the sensor terminal device, and the type of sensor may include but is not limited to temperature sensors, humidity sensors, gas sensors, pressure sensors, vibration sensors, distance sensors, infrared sensors, optical sensors and displacement sensors. There is no limitation here. As long as it can sense the measured information and can convert the sensed information into electrical signals or other required forms of information output according to a certain rule, it can meet the requirements of the sensor terminal device for information transmission, processing, storage, display, recording and control.

[0068] The sensor may include a combination of the following different sensors: at least one target sensor; at least one target sensor in the same environment, and at least one sensor of the same type but with different accuracy or different usage duration; at least one target sensor in the same environment, and at least one sensor of a different type; or at least one target sensor in the same environment, at least one sensor of the same type but with different accuracy or different usage duration, and at least one sensor of a different type.

[0069] The target sensor refers to the sensor that currently needs to be deeply learned and a calibration model is established for subsequent calibration. During the deep learning process, the value of the historical data corresponding to the target sensor is collected through the standard sensor.

[0070] The same environment refers to two or more points whose distance does not exceed a distance threshold or whose difference in environmental attribute values ​​such as temperature and humidity does not exceed a corresponding environmental difference threshold. For example, multiple sensors included in a detection box can be considered to be in the same environment.

[0071] For sensors of the same type but with different precision or different usage time in the same environment, for example, if the target sensor is a temperature sensor with a precision of Class C, the data of Class B temperature sensors and Class A temperature sensors in the same environment can be added to the system for training and calibrating the measurement values ​​of the Class C temperature sensor. The precision of Class A temperature sensors is higher than that of Class B, and that of Class B is higher than that of Class C.

[0072] The historical data refers to time series data composed of sensor data collected and recorded at certain time intervals within a certain period of time. For example, the temperature sensor measures and records the temperature data of the room every 10 minutes from 9:00 to 14:00 Beijing time on April 11, 2022. These temperature data are composed of time series data in the order of collection and recording, which is the historical data of this temperature sensor.

[0073] The sensor data includes but is not limited to: the measurement value of the sensor, the sensor device ID, the measurement value collection time of the sensor, the geographical location information of the sensor, the current weather information of the sensor at the collection time point, and other environmental information of the sensor at the collection time point.

[0074] The standard sensor is a high-precision sensor in the same environment as the sensor, which is responsible for providing the value of the sensor during the deep learning process, and the value is sensor data with an accuracy not lower than the standard threshold.

[0075] The Transformer model is a Transformer model based on a multi-head attention mechanism, which can be used for machine translation. Based on the overall architecture of the Transformer model, which includes an encoder Encoder and a decoder Decoder, the Transformer model is an Encoder-Decoder structure, which is formed by stacking several encoders and decoders. The encoder consists of multi-head attention and a feedforward neural network, which is used to convert input data into feature vectors. The decoder, whose input is the output of the encoder and the predicted results, consists of masked multi-head attention, multi-head attention and a feedforward neural network, which is used to output the conditional probability of the final result. Since there is no recursion or convolution in the Transformer model, the information of the absolute (or relative) position of each token in the sequence is represented by position encoding. The linear layer in the Transformer model is a simple fully connected neural network, and the vector generated by the Decoder is projected into a larger vector to become a logarithmic vector. After the linear layer is a Softmax layer, the Softmax layer can convert the score into a probability by conversion, select the highest probability as the index, and then find the data as the output through the index.

[0076] In the existing Encoder-Decoder framework, all are implemented based on CNN or RNN. However, the Transformer model abandons CNN and RNN and only uses attention to implement it. It can be compared to the role of using multiple filters at the same time in CNN. Therefore, Transformer is an Encoder-Decoder model that is completely based on the attention mechanism.

[0077] The multiple heads are multiple inputs, and the multiple inputs include the input of the target sensor in the same environment, the value of the target sensor, and the input of other sensors of the same or different types in the same environment. The input of the target sensor in the same environment is the historical data of the target sensor in the same environment. The input of other sensors of the same or different types in the same environment is the historical data of other sensors of the same or different types in the same environment. The sensors of the same type are sensors of different precisions or different usage time. For example, the target sensor is a temperature sensor with a precision of Class B, and the sensors of the same type in the same environment are two temperature sensors with precisions of Class A and Class AA, respectively. Intuitively speaking, multi-head attention helps the network capture richer features and information. The so-called multi-head attention mechanism is to send the input obtained by each head to the attention mechanism and comprehensively utilize the features and information of various aspects. The multi-head attention mechanism can extract and learn the intrinsic relationship between the data. When there is sufficient data support, the Transformer model can not only learn the data features of sensors of the same type and realize the establishment of calibration models of sensors of the same type by using the multi-head attention mechanism, but also learn to discover the data correlation between sensors of different types and realize the establishment of calibration models of sensors of different types, that is, to obtain the original model, and alarm for abnormal values.

[0078] The entire training process is carried out in the cloud server. The original model is a model obtained through training that meets the accuracy requirements. Therefore, the Transformer model can not only effectively learn and imitate the characteristics of time series data, but also use the multi-head attention mechanism to help the Transformer model capture richer sensor features and information, and further comprehensively process the captured sensor features and information. It can not only learn the data features of sensors of the same type, but also learn to discover the data correlation between sensors of different types, and has extremely strong filtering capabilities. The Transformer model has been applied to multiple types of sensor devices. The Transformer model supports the establishment of multiple types of calibration models, including but not limited to: the establishment of a calibration model for a single sensor; the establishment of a calibration model for sensors of the same type; the establishment of a calibration model for sensors of different types.

[0079] Furthermore, the original model after deep learning has high computational complexity and high requirements for hardware storage space and computing power. It can only be deployed on cloud servers with high computing power, which limits the deployment of corresponding models in some scenarios and devices. For example, the original model cannot be deployed on base stations and sensor terminal devices, which have lower computing power than cloud servers. In order to achieve hierarchical calibration and enable the model after deep learning to be deployed on sensor terminal devices, base stations and cloud servers respectively, it is necessary to break through the bottleneck with the help of compression optimization, namely model compression, optimization acceleration and other methods. Compression optimization can effectively reduce the redundancy of model parameters, thereby reducing storage occupancy, communication bandwidth and computational complexity, which is conducive to the application deployment of deep learning.

[0080] The sensor terminal device refers to a device that collects data and sends data to the network layer. The base station is an information bridge between the sensor terminal and the cloud server and is a multi-channel transceiver.

[0081] The multi-level compression optimization method provided by the present disclosure includes but is not limited to: knowledge distillation and deep learning pruning. Knowledge distillation refers to the method of using the knowledge learned from the complex model to guide the training of the small model, so that the small model has performance comparable to the complex model, but the number of parameters is greatly reduced, thereby achieving model compression and acceleration. A complex model is a single complex network or a collection of several networks with good performance and generalization capabilities. The core idea of ​​knowledge distillation is to first train a complex network model, and then use the output of this complex network and the real label of the data to train a smaller network. Therefore, the knowledge distillation framework usually includes a complex model (Teacher model) and a small model (Student model). The first-level compressed and optimized model obtained by knowledge distillation in this application is deployed on the sensor terminal device.

[0082] Deep learning pruning refers to a method of pruning a deep learning model into a streamlined network model by converting dense connections in a large network into sparse connections due to the sparsity or overfitting tendency of the deep learning model. During the training process, the parameters with smaller weights are gradually set to 0, and then those with weight values ​​0 are removed, that is, some computing costs with too low benefits are deleted. The model after secondary compression optimization obtained by deep learning pruning in this application is deployed in the base station. The computing power of the base station is higher than that of the sensor terminal device and lower than that of the cloud server.

[0083] Through deep learning pruning and knowledge distillation of the original model, the model can be compressed and optimized at multiple levels without significantly reducing the accuracy, so that it can be deployed in sensor terminal devices, base stations and cloud servers respectively. Combined with the computing power characteristics of sensor terminal devices, base stations and cloud servers, the present invention can intelligently match the calibration position to achieve a balance between the ratio of accuracy and response speed, and realize application in multiple scenarios. Figure 5 As shown, the original model after knowledge distillation, the lightweight model, that is, the model after the first-level compression optimization, is deployed on the sensor terminal device with weak computing power; the original model without compression optimization, the large model, is deployed on the cloud server with strong computing power; the original model after deep learning pruning, the medium model, that is, the model after the second-level compression optimization, is deployed on the base station with medium computing power.

[0084] Data calculation amount: the original model is larger than the model after the second-level compression optimization, which is larger than the model after the first-level compression optimization; calibration accuracy: the original model is larger than the model after the second-level compression optimization, which is larger than the model after the first-level compression optimization; response speed: the model after the first-level compression optimization is larger than the model after the second-level compression optimization, which is larger than the original model.

[0085] Further, according to the calibration accuracy requirements, response speed and / or data calculation amount, it is determined that the sensor terminal device, base station or cloud server uses the deployed model to perform hierarchical calibration. For example, a fast first-level calibration with a small amount of data calculation and / or low accuracy is performed on the sensor terminal device, and a third-level calibration with a large amount of data calculation and / or high accuracy is performed on the cloud server.

[0086] If the base station's secondary compression optimized model is required for secondary calibration, the sensor terminal device will select a nearby base station for calibration. If the nearby base station is unavailable, the sensor terminal device will enable a backup line and submit sensor data to other base stations. If all base stations are unavailable, the sensor terminal device will use the primary compression optimized model for calibration. Figure 6 The solid lines shown are lines between the sensor and the nearest base station, and the dotted lines shown are lines between the sensor and other base stations, ie, backup lines.

[0087] In summary, the complete process of steps S101-S106 is as follows: Step S101 is specifically the sensor data collected and recorded at regular intervals in a certain time period in a time sequence, forming time series data, i.e., the historical data, wherein the sensor data includes but is not limited to: the measurement value of the sensor, the sensor device ID, the measurement value collection time of the sensor, the geographical location information of the sensor, the environmental meteorological information of the sensor, and other environmental information of the sensor. For example, it can also include (1) air quality parameters in the environmental protection industry: NO 2 、SO2 ,CO,O 3 , PM2.5, PM10, PM1.0, TVOC and other parameters; (2) Hydrological parameters in the environmental protection industry: flow rate, flow, water level, water volume; (3) Water quality parameters in the environmental protection industry: water temperature, dissolved oxygen, pH value, conductivity, turbidity, total phosphorus, total nitrogen, ammonia nitrogen, permanganate index, chemical oxygen demand CODcr; (4) Soil parameters in the forest fire protection industry: conductivity, humidity, salinity, temperature; (5) Various equipment parameters in the urban management or security industry: hydrogen sulfide, NH 3 , smoke sensor, gas sensor, pressure, geomagnetic parameters, smart trash can parameters, smart manhole cover parameters, smart sound and light alarm parameters. By maintaining the matrix information of the time series of the reported data, inferring whether the data drifts, and combining the calculation amount, intelligently selecting the calibration location (sensor end, gateway end and cloud end) for inference calibration.

[0088] Among them, the types of sensors may include but are not limited to temperature sensors, humidity sensors, gas sensors, pressure sensors, vibration sensors, distance sensors, infrared sensors, optical sensors and displacement sensors, which are not limited here. The sensors include: target sensors; target sensors in the same environment and sensors of the same type but with different precisions or different lengths of use; target sensors in the same environment and sensors of different types; or target sensors in the same environment, sensors of the same type but with different precisions or different lengths of use, and sensors of different types. The target sensor refers to the sensor that currently needs to be deeply learned and a calibration model is established for subsequent calibration. In the process of deep learning, the value of the historical data corresponding to the target sensor is collected by the standard sensor. The same environment refers to two points or more whose distance does not exceed the distance threshold or the difference of the environmental attribute values ​​such as temperature and humidity does not exceed the corresponding environmental difference threshold.

[0089] Figure 7The specific process of steps S102-S104 is further shown, and the value of the historical data corresponding to the target sensor is collected through the standard sensor, and the historical data and the value are provided to the Transformer model. The standard sensor is a high-precision sensor in the same environment corresponding to the sensor, which is responsible for providing the value of the sensor in the deep learning process, and the value is the sensor data with an accuracy not lower than the standard threshold. The Transformer model trains the historical data and the value, wherein the multi-head attention mechanism of the Transformer model sends multiple heads, i.e., multiple inputs, including the historical data of the target sensor in the same environment, the value of the target sensor, and the historical data of other sensors of the same or different types in the same environment, to the attention mechanism, and comprehensively utilizes the features and information of various aspects of the multiple inputs to extract and learn the intrinsic relationship between various data. Among them, sensors of the same type are sensors with different precision or different usage time. By learning the historical data and values ​​of the same sensor, the calibration model of a single sensor is established. The data characteristics of sensors of the same type but with different accuracy or different usage time in the same environment are used to establish the calibration model of sensors of the same type. By learning the data correlation between different types of sensors in the same environment, the calibration models of different types of sensors are established. It is further judged whether the accuracy of the trained model meets the requirements. If so, it is output as the original model. If not, training continues until its accuracy meets the requirements.

[0090] Step S105 specifically comprises performing first-level compression optimization on the original model through knowledge distillation to obtain a first-level compression optimized model; performing second-level compression optimization on the original model through deep learning pruning to obtain a second-level compression optimized model. The first-level compression optimized model has low computing power requirements and is deployed in the sensor terminal device. The second-level compression optimized model has medium computing power requirements and is deployed in the base station.

[0091] The original model that has not been compressed and optimized requires high computing power and is deployed on a cloud server. Therefore, the multi-level compression optimization of the original model implements hierarchical calibration, so that the model after deep learning has the ability to be deployed in sensor terminal devices, base stations, and cloud servers respectively. According to the calibration accuracy requirements, response speed, and / or data calculation amount, the model after deep learning, that is, the model after the first-level compression optimization, the model after the second-level compression optimization, and the original model can intelligently match the calibration position. A balance is achieved between the ratio of accuracy and response speed, and application in multiple scenarios is realized.

[0092] Further, step S106 calibrates the raw data subsequently collected by the target sensor according to the original model deployed on the cloud server, the secondary compression optimized model deployed on the base station, or the primary compression optimized model deployed on the sensor terminal device.

[0093] In addition, the obtained calibration models of different precision types, namely the original model, the model after the first-level compression optimization and the model after the second-level compression optimization, can not only be used for hierarchical calibration, and different calibration models can be selected according to different needs; they can also be used for multi-level collaborative calibration, where the advanced model calibrates at least part of the original data, and the obtained at least part of the advanced calibrated data is compared with the corresponding low-level calibrated data. If the difference between the two is less than a certain error threshold, the sampling inspection is qualified, and all the low-level calibrated data are accepted. Otherwise, the advanced calibration is performed on all the original data to obtain all the advanced calibrated data. By utilizing the different precisions of the calibration models, the low-level calibrated data is spot-checked through the advanced model, which realizes a simpler and more efficient inspection of the received calibration results.

[0094] like Figure 8 The present invention is a flowchart of a multi-level collaborative calibration method for a sensor calibration method based on deep learning according to the present application, comprising the following steps: S601: the sensor terminal device performs a first-level calibration on the raw data collected by the sensor using a model optimized by a first-level compression to obtain the first-level calibrated data; S602: the sensor terminal device uploads the raw data and the first-level calibrated data to the base station; S603: the base station performs a second-level calibration on at least part of the received raw data using a model optimized by a second-level compression to obtain at least part of the second-level calibrated data; S604: compare the at least part of the second-level calibrated data with the corresponding first-level calibrated data. If the difference between the two is less than a certain error threshold, accept all the first-level calibrated data; otherwise, use the second-level compression to optimize the data. The calibrated model performs a second-level calibration on the received original data to obtain all the second-level calibrated data; S605: the base station uploads the original data and all the received first-level calibrated data, or the original data and all the second-level calibrated data to the cloud server; S606: the cloud server performs a third-level calibration on at least part of the received original data using the original model to obtain at least part of the third-level calibrated data; S607: compare the third-level calibrated data with the corresponding first-level calibrated data or the second-level calibrated data, if the difference between the two is less than a certain error threshold, accept all the first-level calibrated data or all the second-level calibrated data, otherwise, perform a third-level calibration on the received original data using the original model to obtain all the third-level calibrated data.

[0095] The calibration accuracy of the original model obtained through deep learning will decrease to varying degrees after a certain period of time has passed since the last model was deployed. Therefore, when the calibration accuracy is lower than a certain accuracy threshold or a certain period of time has passed since the last model was deployed, the updated original model needs to be retrained to replace the original model. For example, the target sensor is a temperature sensor. On the third day of each month, another high-precision temperature sensor is used to accurately measure the temperature of the environment where the target sensor is located. When it is found that the difference between the model calibration result and the accurate measurement result is higher than a certain accuracy threshold, it means that the error of the calibration result of the original model is large, and the original model cannot continue to work. It needs to be retrained, and the original model that has been retrained and updated is used to replace the previous original model; or after the original model has been used for 3 months, the original model that has been retrained and updated is used to replace the previous original model.

[0096] like Fig. 9 It is a flowchart of retraining the updated original model, including the following steps: S701: the sensor collects historical data in chronological order. The historical data required for retraining is different from the historical data used for the last training model. The historical data within a period of time immediately before the current training should be used for this training, so that a more accurate model can be obtained; S702: the standard sensor collects at least part of the corresponding historical data; S703: the historical data and the values ​​are provided to the Transformer model; S704: the Transformer model trains the historical data and the values ​​to obtain the updated original model; S705: the updated original model is subjected to multi-level compression optimization through deep learning pruning or knowledge distillation to obtain an updated multi-level compression optimized model; S706: the original data collected by the sensor later is calibrated according to the updated original model or the updated multi-level compression optimized model.

[0097] As before, the updated first-level compressed and optimized model obtained through knowledge distillation is deployed on the sensor terminal device; the updated second-level compressed and optimized model obtained through deep learning pruning is deployed on the base station; and the updated original model is deployed on the cloud server.

[0098] As an optional implementation, 4 temperature sensors with different accuracies within a radius of 10 meters (the accuracies are Class C, Class B, Class A, and Class AA, and the accuracies increase from left to right, with Class C having the lowest accuracies and Class AA having the highest accuracies), 1 humidity sensor, and 1 pressure sensor collect and record sensor data every 10 minutes from 9:00 to 14:00 Beijing time on April 11, 2022 in chronological order, forming time series data, i.e., the historical data of each sensor. The sensor data includes, but is not limited to: the measurement value of the sensor, the sensor device ID, the measurement value collection time of the sensor, the geographic location information of the sensor, the environmental meteorological information of the sensor, and other environmental information of the sensor. The measurement value of the temperature sensor is the temperature value, the measurement value of the humidity sensor is the humidity value, and the measurement value of the pressure sensor is the pressure of the measuring medium, i.e., the liquid or gas.

[0099] The C-level temperature sensor with a precision is the target temperature sensor, and the value corresponding to the historical data of the C-level temperature sensor is collected through a standard sensor, and the historical data of the four temperature sensors, the humidity sensor and the pressure sensor and the value of the C-level temperature sensor are provided to the Transformer model. The Transformer model is trained on the above data, wherein the multi-head attention mechanism of the Transformer model sends the historical data of the C-level temperature sensor, the value of the C-level temperature sensor, the historical data of the B-level temperature sensor, the historical data of the A-level temperature sensor, the historical data of the AA-level temperature sensor, the historical data of the humidity sensor and the historical data of the pressure sensor to the attention mechanism, and comprehensively utilizes the combination of at least part of the data of the temperature, humidity, pressure value and the temperature value in multiple inputs to learn and obtain different calibration models.

[0100] The first calibration model is established by learning the historical data and values ​​of the target sensor, such as the historical data and values ​​of the C-class temperature sensor; the second calibration model is established by learning the data characteristics of the target sensor and sensors of the same type but with different precision or different usage time within a radius of 10 meters, such as the historical data of the four temperature sensors and the values ​​of the C-class temperature sensor; the third calibration model is established by learning the data correlation between the target sensor and different types of sensors within a radius of 10 meters, for example, by learning the historical data and values ​​of the C-class temperature sensor, the historical data of the humidity sensor and the historical data of the pressure sensor, and discovering the data correlation between these three types of sensors. Fig.10The figure shows the correlation between temperature and humidity learned by the Transformer model. It can be seen from the figure that humidity decreases as temperature increases. When the temperature rises to the highest point, the humidity also drops to the lowest point, and then the humidity increases as the temperature decreases. Further outlier detection for different types of sensors includes: if there is only one humidity sensor and several temperature sensors in the same environment, and the temperature continues to rise over a period of time, but the reported humidity data suddenly increases after a continuous decline, that is, Fig.10 If the trend does not match, an alarm will be triggered.

[0101] refer to Figures 11 to 15 By learning the target sensor, sensors of the same type but with different precision or different usage time within a radius of 10 meters, and the data association between sensors of different types, for example, by learning the historical data of the four temperature sensors and the value of the C-level temperature sensor, the historical data of the humidity sensor and the historical data of the pressure sensor, the fourth calibration model is established. Further continue to determine whether the accuracy of the trained model meets the requirements, if so, output it as the corresponding original model; if not, continue training until its accuracy meets the requirements.

[0102] Since different calibration models can be trained based on different historical data inputs, each model has different complexity and requires different input data. The first to fourth original calibration models can be deployed to sensor terminal devices, base stations or cloud servers according to specific circumstances.

[0103] When calibrating, if the input data only comes from the C-level temperature sensor, the first model mentioned above needs to be selected for calibration; if the input data comes from the C-level temperature sensor, humidity sensor and pressure sensor, the third model mentioned above needs to be selected for calibration.

[0104] Since the original model has a high computational complexity, it can only be deployed on a cloud server with high computing power. However, the computing power of the cloud server is large and the responsiveness is low. In order to achieve fast calibration, the original model is further compressed and optimized by knowledge distillation to obtain a model after first-level compression optimization; the original model is compressed and optimized by deep learning pruning to obtain a model after second-level compression optimization. The model after first-level compression optimization has low computing power requirements and is deployed in sensor terminal devices. The model after second-level compression optimization has medium computing power requirements and is deployed in base stations.

[0105] The calibration accuracy of the model after the secondary compression optimization is higher than the calibration accuracy of the model after the primary compression optimization, and lower than the calibration accuracy of the original model; the response speed of the model after the secondary compression optimization is lower than the response speed of the model after the primary compression optimization, and higher than the response speed of the original model; the data calculation amount of the model after the secondary compression optimization is higher than the data calculation amount of the model after the primary compression optimization, and lower than the data calculation amount of the original model.

[0106] This implementation requires that the calibration response speed should be as fast as possible, so it is determined that the model after the first-level compression optimization deployed in the sensor terminal device or the first calibration model in the above-mentioned original model performs first-level calibration on the 100 original data collected after the C-level temperature sensor, and obtains 100 first-level calibrated temperature sensor data; the sensor terminal device uploads the 100 original data and the 100 first-level calibrated temperature sensor data to the base station; in order to spot-check whether the 100 data calibrated by the sensor terminal device are accurate, the base station uses the model after the second-level compression optimization to perform second-level calibration on any 10 of the original data, and obtains 10 second-level calibrated temperature sensor data; compares these 10 second-level calibrated temperature sensor data with their corresponding 10 first-level calibrated temperature sensor data, and obtains 10 differences. The average value of these 10 differences is greater than a certain error threshold that has been set, that is, the spot check fails, and continues to use the model after the second-level compression optimization to perform second-level calibration on the remaining 90 original data to obtain all second-level calibrated temperature sensor data.

[0107] The base station uploads the 100 original data and the 100 secondary calibrated temperature sensor data to the cloud server; in order to spot check whether the 100 data calibrated by the base station are accurate, the cloud server uses the original model to perform three-level calibration on any 10 original data to obtain 10 three-level calibrated temperature sensor data; compares these 10 three-level calibrated temperature sensor data with their corresponding 10 secondary calibrated temperature sensor data to obtain 10 differences, and the average value of these 10 differences is less than a set error threshold, that is, the spot check passes, and all the secondary calibrated temperature sensor data are accepted.

[0108] The calibration accuracy of the original model, the model after the first level compression optimization, and the model after the second level compression optimization has been greatly reduced after 3 months since the last model deployment. It is necessary to retrain the updated original model to replace the original model whose calibration accuracy cannot meet the requirements. The specific training method is the same as the method for training the original model for the first time. Similarly, the updated original model is subjected to multi-level compression optimization to obtain an updated multi-level compression optimized model, and the updated first-level compression optimized model obtained by knowledge distillation is deployed on the sensor terminal device; the updated second-level compression optimized model obtained by deep learning pruning is deployed on the base station; and the updated original model is deployed on the cloud server.

[0109] The Transformer model based on the multi-head attention mechanism in this embodiment can not only capture the historical data and values ​​of the target sensor to establish the first calibration model; it can also capture the data characteristics of the target sensor and multiple temperature sensors with different accuracies in the same environment to establish the second calibration model; it can also establish the third calibration model by learning the data correlation between the target sensor and the humidity sensor and pressure sensor in the same environment; it can also establish the fourth calibration model by learning the target sensor, sensors of the same type but with different accuracies or different usage times in the same environment, and the data correlation between the humidity sensor and the pressure sensor in the same environment.

[0110] And through knowledge distillation, the original model is compressed and optimized at the first level to obtain a first-level compressed and optimized model that can be deployed on the sensor terminal device, realizing the rapid calibration of the original data collected by the target temperature sensor. The multi-level compression optimization of the original model achieves a balance between the ratio of calibration accuracy and response speed. It breaks through the singleness of the deep learning model that can only be deployed on the cloud server due to the high computational complexity.

[0111] Multi-level collaborative calibration can use multi-level calibration models with different precision to calibrate part of the original data, and spot-check whether the calibration results reported by the low-level calibration model are qualified, thus realizing a simple and efficient verification of the received calibration results.

[0112] The present disclosure also provides a more general deep learning processing method, which generates a multi-level simplified model using the original model, and deploys it in a multi-level distributed network device according to the processing capacity of the network device, so as to realize the corresponding recognition, calibration and other big data processing applications using the deployed models at each level in the multi-level network device. The processing method comprises the following steps:

[0113] Collect historical data in chronological order;

[0114] collecting values ​​of at least some of the corresponding historical data;

[0115] providing the historical data and the numerical value to a deformer model;

[0116] The deformer model trains the historical data and the numerical value to obtain an original model;

[0117] The original model is subjected to multi-level compression optimization through deep learning pruning or knowledge distillation to obtain a multi-level compression optimized model, the first-level compression optimized model obtained through knowledge distillation is deployed on the terminal device, the second-level compression optimized model obtained through deep learning pruning is deployed on the base station, and the original model is deployed on the cloud server, the processing accuracy of the second-level compression optimized model is higher than the processing accuracy of the first-level compression optimized model, and lower than the processing accuracy of the original model, the response speed of the second-level compression optimized model is lower than the response speed of the first-level compression optimized model, and higher than the response speed of the original model, the data calculation amount of the second-level compression optimized model is higher than the data calculation amount of the first-level compression optimized model, and lower than the data calculation amount of the original model; according to the processing accuracy requirements, response speed and / or data calculation amount, determine to use the deployed model in the terminal device, base station or cloud server to process the original data collected later.

[0118] The terminal device performs primary processing on the raw data collected later using the model after the primary compression optimization to obtain primary processed data;

[0119] The terminal device uploads the original data and the primary processed data to the base station;

[0120] The base station performs secondary processing on at least part of the received original data using the model optimized by the secondary compression to obtain at least part of the secondary processed data;

[0121] Compare the at least part of the secondary processed data with the corresponding primary processed data, and if the difference between the two is less than a certain error threshold, accept all the primary processed data; otherwise, perform secondary processing on the received original data using the model optimized by secondary compression to obtain all the secondary processed data;

[0122] The base station uploads the original data and all the received primary processed data, or the original data and all the secondary processed data to the cloud server;

[0123] The cloud server performs three-level processing on at least part of the received original data using the original model to obtain at least part of the three-level processed data;

[0124] The tertiary processed data is compared with the corresponding primary processed data or the secondary processed data. If the difference between the two is less than a certain error threshold, all the primary processed data or all the secondary processed data are accepted. Otherwise, the received original data is subjected to tertiary processing using the original model to obtain all the tertiary processed data.

[0125] Retrain the updated original model based on the processing accuracy falling below a certain accuracy threshold or a certain amount of time having passed since the last model deployment, including the following steps:

[0126] Collect historical data in chronological order;

[0127] Collecting values ​​of at least part of the corresponding historical data;

[0128] providing the historical data and the numerical value to the deformer model;

[0129] The deformer model trains the historical data and the numerical value to obtain the updated original model;

[0130] Perform multi-level compression optimization on the updated original model through deep learning pruning or knowledge distillation to obtain an updated multi-level compression optimized model, deploy the updated first-level compression optimized model obtained through knowledge distillation on the terminal device, deploy the updated second-level compression optimized model obtained through deep learning pruning on the base station, and deploy the updated original model on the cloud server;

[0131] The raw data collected subsequently is processed according to the updated original model or the updated multi-stage compression optimized model.

[0132] The present disclosure is not limited to the above-mentioned distributed network structure of three-level network devices, but can also be applied to two-level, four-level or even more-level network structures, so that each level of network equipment can be applied to the trained versions of the models at each level to improve the efficiency and accuracy of data processing.

[0133] Exemplary Systems

[0134] The embodiments of the present disclosure are described in detail below with reference to the accompanying drawings.

[0135] Fig.11 This is a structural block diagram of a sensor calibration system based on deep learning according to the present application.

[0136] The calibration system 900 includes: a sensor 910 for collecting historical data in chronological order;

[0137] A standard sensor 920 for collecting values ​​of at least part of the corresponding historical data;

[0138] A training device 930, which is used to receive the historical data and the numerical value, and train the historical data and the numerical value to obtain an original model;

[0139] A compression optimization device 940, which is used to perform multi-level compression optimization on the original model through deep learning pruning or knowledge distillation to obtain a multi-level compressed optimized model;

[0140] The calibration device 950 calibrates the raw data subsequently collected by the sensor according to the original model or the model after multi-stage compression optimization.

[0141] Sensor 910 includes: a target sensor;

[0142] Target sensors in the same environment, and sensors of the same type but with different accuracy or different usage time;

[0143] Target sensors in the same environment, and sensors of different types; or

[0144] Target sensors in the same environment, sensors of the same type but with different accuracy or different usage duration, and sensors of different types.

[0145] The training device 930 is a deformer model, which uses a multi-head attention mechanism to learn the data features of sensors of the same type and learn to discover the data associations between sensors of different types to obtain the original model; wherein

[0146] The multiple heads are multiple inputs, and the multiple inputs include inputs of target sensors in the same environment, values ​​of the target sensors, and inputs of other sensors of the same or different types in the same environment.

[0147] like Fig.12 As shown, the compression optimization device 940 also includes a deployment module 1041, which is used to deploy the first-level compression optimized model obtained through knowledge distillation on the sensor terminal device; deploy the second-level compression optimized model obtained through deep learning pruning on the base station; and deploy the original model on the cloud server.

[0148] The calibration accuracy of the model after the secondary compression optimization is higher than the calibration accuracy of the model after the primary compression optimization, and lower than the calibration accuracy of the original model; the response speed of the model after the secondary compression optimization is lower than the response speed of the model after the primary compression optimization, and higher than the response speed of the original model; the data calculation amount of the model after the secondary compression optimization is higher than the data calculation amount of the model after the primary compression optimization, and lower than the data calculation amount of the original model.

[0149] like Fig.13 As shown, the calibration device 950 also includes a model determination module 1151 , a multi-level calibration scheduling module 1152 and an update command module 1153 .

[0150] The model determination module 1151 determines to perform calibration using the deployed model in the sensor terminal device, base station or cloud server according to the calibration accuracy requirement, response speed and / or data calculation amount.

[0151] The multi-level calibration scheduling module 1152 performs the following control operations:

[0152] The sensor terminal device uses the model after the first-level compression optimization to perform a first-level calibration on the raw data collected by the sensor to obtain the first-level calibrated data, and the sensor terminal device uploads the raw data and the first-level calibrated data to the base station;

[0153] The base station uses the model optimized by the secondary compression to perform secondary calibration on at least part of the received original data to obtain at least part of the secondary calibrated data, and compares the at least part of the secondary calibrated data with the corresponding primary calibrated data. If the difference between the two is less than a certain error threshold, all the primary calibrated data are accepted. Otherwise, the model optimized by the secondary compression is used to perform secondary calibration on the received original data to obtain all the secondary calibrated data, and the original data and all the accepted primary calibrated data, or the original data and all the secondary calibrated data are uploaded to the cloud server.

[0154] The cloud server uses the original model to perform three-level calibration on at least part of the received original data to obtain at least part of the three-level calibrated data, and compares the three-level calibrated data with the corresponding first-level calibrated data or the second-level calibrated data. If the difference between the two is less than a certain error threshold, all the first-level calibrated data or all the second-level calibrated data are accepted. Otherwise, the original model is used to perform three-level calibration on the received original data to obtain all the three-level calibrated data.

[0155] The update command module 1153 determines to retrain the updated original model according to the calibration accuracy being lower than a certain accuracy threshold or a certain time after the last model deployment, and performs the following control operations:

[0156] The sensor 910 collects historical data in chronological order;

[0157] Allowing the standard sensor 920 to collect at least part of the corresponding values ​​of the historical data;

[0158] The training device 930 receives the historical data and the numerical value, and performs training on the historical data and the numerical value to obtain an original model;

[0159] The compression optimization device 940 performs multi-level compression optimization on the updated original model through deep learning pruning or knowledge distillation to obtain an updated multi-level compression optimized model;

[0160] The calibration device 950 calibrates the raw data subsequently collected by the sensor according to the updated original model or the updated multi-stage compression optimized model.

[0161] The deployment module 1041 deploys the updated first-level compressed and optimized model obtained through knowledge distillation on the sensor terminal device; deploys the updated second-level compressed and optimized model obtained through deep learning pruning on the base station; and deploys the updated original model on the cloud server.

[0162] Example Applications

[0163] An application of a sensor calibration method based on deep learning, the application comprising the following steps:

[0164] Sensors collect raw data in real time;

[0165] Collecting values ​​corresponding to the raw data in real time through standard sensors;

[0166] Extract a certain amount of the raw data and the corresponding values ​​according to the sampling rate, and upload the extracted raw data and the values ​​to the base station;

[0167] The base station compares a certain amount of the original data with a numerical value corresponding thereto;

[0168] If the difference between the two is greater than a certain accuracy threshold and the proportion is less than the ratio threshold, the sensor is marked as a sensor in a normal state, all the raw data is accepted and uploaded to the cloud server.

[0169] If the proportion of the difference between the two is greater than a certain accuracy threshold is greater than or equal to a ratio threshold, the base station sends a first-level calibration instruction to the sensor terminal device so that the sensor terminal device uses the first-level compression optimized model to perform a first-level calibration on the original data to obtain the first-level calibrated data. The sensor terminal device uploads the first-level calibrated data to the base station. The base station takes out a certain amount of the first-level calibrated data according to the sampling rate, and compares it with the corresponding numerical value. If the difference between the two is greater than a certain accuracy threshold and the proportion is less than the ratio threshold, the sensor is marked as a sensor that requires local calibration, accepts all the first-level calibrated data, and uploads all the first-level calibrated data to the cloud server.

[0170] The base station takes out a certain amount of the first-level calibrated data according to the sampling rate, and compares it with the corresponding value. If the difference between the two is greater than a certain accuracy threshold and the proportion is greater than or equal to the proportion threshold, the sensor is marked as an abnormal sensor;

[0171] The base station performs secondary calibration on the raw data of the abnormal sensor using the secondary compression optimized model, obtains the secondary calibrated data and uploads all the secondary calibrated data to the cloud server;

[0172] The abnormal sensor is retrained to obtain its updated original model. Before obtaining the updated original model, the base station is required to perform secondary calibration on the original data of the abnormal sensor using the model after secondary compression optimization.

[0173] As an optional implementation, Fig.14 Shown is a flowchart of an exemplary application of a deep learning-based calibration method to calibration of sensors of the same type.

[0174] Three temperature sensors with different accuracies in the same region (the accuracies are Class C, Class B, and Class A, and the accuracies increase from left to right, with Class C having the lowest accuracies and Class A having the highest accuracies) collect 100 raw data in real time, and deploy three standard sensors to collect 300 values ​​corresponding to the raw data in real time. The sampling rate is set to 10% according to changes in the external environment, such as seasonal changes, the arrival of the rainy season, and continuous fog. Ten raw data are randomly selected from the 100 raw data of the Class A temperature sensor, the 100 raw data of the Class B temperature sensor, and the 100 raw data of the Class C temperature sensor. The 30 raw data taken out above and their corresponding values ​​are uploaded to the base station. The base station further calculates the difference between these 30 original data and their corresponding values ​​to obtain 30 difference values, among which only 1 difference value in the A-level temperature sensor is greater than the accuracy threshold (that is, the proportion of data that does not meet the accuracy threshold is 10%), 4 difference values ​​in the B-level temperature sensor are greater than the accuracy threshold (that is, the proportion of data that does not meet the accuracy threshold is 40%), and 8 difference values ​​in the C-level temperature sensor are greater than the accuracy threshold (that is, the proportion of data that does not meet the accuracy threshold is 80%).

[0175] In this embodiment, the ratio threshold is set to 20% according to the change of the external environment. If the ratio of Class A temperature sensors that do not meet the accuracy threshold is 10% and is less than the ratio threshold of 20%, the base station marks the Class A temperature sensor as a sensor in normal state. The Class A sensor uploads 100 raw data to the base station, and the base station receives all the raw data of the Class A sensor and uploads it to the cloud server.

[0176] If 40% of Class B temperature sensors do not meet the accuracy threshold, which is greater than the ratio threshold of 20%, the base station marks the Class B temperature sensors as sensors that need calibration. If 80% of Class C temperature sensors do not meet the accuracy threshold, which is greater than the ratio threshold of 20%, the base station marks the Class C temperature sensors as sensors that need calibration.

[0177] The base station further sends a first-level calibration instruction to the Class B temperature sensor and the Class C temperature sensor, so that the Class B temperature sensor terminal device uses the model after the first-level compression optimization to perform a first-level calibration on 100 original data of the Class B temperature sensor, and obtains 100 first-level calibrated data of the Class B temperature sensor. The Class C temperature sensor terminal device uses the model after the first-level compression optimization to perform a first-level calibration on 100 original data of the Class C temperature sensor, and obtains 100 first-level calibrated data of the Class C temperature sensor.

[0178] According to a sampling rate of 10%, 10 data are randomly selected from 100 first-level calibrated data of Class B temperature sensors and 100 first-level calibrated data of Class C temperature sensors, respectively. The 20 first-level calibrated data and their corresponding values ​​are uploaded to the base station.

[0179] The base station further calculates the difference between the 20 first-level calibrated data and their corresponding values ​​to obtain 20 difference values, among which only one difference value in the B-level temperature sensor is greater than the accuracy threshold (that is, the proportion of data that does not meet the accuracy threshold is 10%), and 5 difference values ​​in the C-level temperature sensor are greater than the accuracy threshold (that is, the proportion of data that does not meet the accuracy threshold is 50%).

[0180] If the proportion of Class B temperature sensors that do not meet the accuracy threshold is 10% and is less than the proportion threshold of 20%, the base station will mark the Class B temperature sensors as sensors that need local calibration. Class B sensors upload 100 first-level calibrated data to the base station, and the base station receives all the first-level calibrated data of Class B sensors and uploads them to the cloud server.

[0181] The proportion of Class C temperature sensors that do not meet the accuracy threshold is 50%, which is greater than the proportion threshold of 20%. The base station marks the Class C temperature sensors as abnormal sensors.

[0182] The base station performs secondary calibration on the original data of the C-class temperature sensor using the model optimized by secondary compression, obtains the secondary calibrated data and uploads all the secondary calibrated data to the cloud server.

[0183] The C-level temperature sensor is retrained to obtain its updated original model. Before obtaining the updated original model, the base station needs to perform secondary calibration on the original data collected in real time by the C-level temperature sensor using the model optimized by secondary compression.

[0184] As another optional implementation, different from the above implementation which selects 100 raw data from the same sensor (such as 100 raw data from 1 Class A temperature sensor), this implementation selects 1 raw data from each of 100 sensors of the same accuracy and the same type (such as 100 Class A temperature sensors each collect 1 raw data).

[0185] 100 Class A temperature sensors, 100 Class B temperature sensors, and 100 Class C temperature sensors in the same region collect 1 raw data each, and 300 standard sensors are deployed to collect 300 values ​​corresponding to the raw data in real time. 10 raw data of 100 Class A temperature sensors, 100 Class B temperature sensors, and 100 Class C temperature sensors are randomly selected. The 30 raw data and their corresponding values ​​are uploaded to the base station. The base station further calculates the difference between these 30 raw data and their corresponding values ​​to obtain 30 difference values, among which only 1 difference value of Class A temperature sensors is greater than the accuracy threshold (that is, the proportion of sensors that do not meet the accuracy threshold is 10%), 4 difference values ​​of Class B temperature sensors are greater than the accuracy threshold (that is, the proportion of sensors that do not meet the accuracy threshold is 40%), and 8 difference values ​​of Class C temperature sensors are greater than the accuracy threshold (that is, the proportion of sensors that do not meet the accuracy threshold is 80%).

[0186] In this embodiment, the proportion that does not meet the accuracy threshold is the proportion of the number of such sensors, for example, 1 out of 10 Class A temperature sensors does not meet the accuracy threshold, and the proportion of sensors that do not meet the accuracy threshold is 10%; this is different from the proportion that does not meet the accuracy threshold in the previous embodiment is the proportion of sensor data of this sensor, for example, 1 out of 10 sensor data of 1 Class A temperature sensor does not meet the accuracy threshold, and the proportion of data that does not meet the accuracy threshold is 10%.

[0187] The above two embodiments respectively determine whether it is necessary to calibrate the sensor according to the proportion of data with deviations in the data of the same sensor, and determine whether to calibrate a batch of sensors according to the proportion of sensors with deviations in the same batch of sensors. The different judgment criteria both fall within the scope of the present disclosure. Ordinary technicians in this field can flexibly determine their judgment criteria, accuracy thresholds, ratio thresholds, etc. according to specific needs.

[0188] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit it. Although the present application has been described in detail with reference to the aforementioned embodiments, a person of ordinary skill in the art should understand that the technical solutions described in the aforementioned embodiments can still be modified, or some of the technical features can be replaced by equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for collecting data by a sensing terminal, comprising: Collect data through multiple sensing terminals; Adjusting a sampling mode, wherein the sampling mode includes one or more of the following: sampling frequency, sampling accuracy, source coding and / or compression method; as well as Determine whether the data collected by the perception terminal is abnormal; when abnormal, the first device connected to the perception terminal generates a first alarm message, and sends the first alarm message to all second devices connected to the first device; the second device sends a second alarm message to all alarm devices connected to the second device; wherein the first device and the second device are edge devices or intermediate devices of the IOT network.

2. The method according to claim 1, wherein the method further comprises: The sensing strategy is determined and / or the sampling mode is adjusted according to business needs and / or the network status.

3. The method according to claim 1, wherein: When the data difference of the data collected by the sensing terminal exceeds a threshold range, and / or meets specific image or video feature values, or specific sound wave features, an alarm message is generated.

4. The method according to claim 1, wherein the method further comprises: When the perceived data changes and / or requires attention, communication resources are allocated by adjusting communication parameters, and the adjustment parameters include one or more of the following: carrier frequency, carrier bandwidth, modulation mode, channel coding, transmit power, and receive sensitivity.

5. A method of edge computing, comprising: Receive data from the terminal; analyzing a rate of change of the data; Compared with adjusting the terminal data reporting frequency threshold; Determine whether the reporting frequency needs to be adjusted based on the comparison results; as well as If the terminal reported data is in a fast-changing interval, a sensor parameter adjustment instruction is sent to the terminal to reduce the terminal data reporting time interval; If the terminal reported data changes slowly, a sensor parameter adjustment instruction is sent to the terminal to lengthen the terminal data reporting time interval.

6. The method according to claim 5, wherein the method further comprises: When the data difference of the data collected by the perception terminal exceeds a specific value, and / or meets specific image, video feature values, and / or sound wave features, an alarm message is generated.

7. The method according to claim 6, wherein the method further comprises: When the data collected by the perception terminal is normal, reduce the frequency of data collection by the perception terminal.

8. The method according to claim 6, wherein the method further comprises: Allocate network communication resources and / or acquisition modes; The sampling mode includes one or more of the following: sampling frequency, sampling accuracy, source coding and / or compression method.

9. The method according to claim 5, wherein the method further comprises: Data is collected by several sensing terminals; it is determined whether the data collected by the sensing terminals is abnormal; when an abnormality occurs, a first device connected to the sensing terminal generates a first alarm message, and sends the first alarm message to all second devices connected to the first device; the second device sends a second alarm message to all alarm devices connected to the second device.

10. A method for multi-stage compression optimization, the method comprising: Sensors collect historical data in chronological order; Acquiring accurate values ​​of at least part of the corresponding historical data by means of standard sensors; providing the historical data and the accurate value to a deformer model; The deformer model trains the historical data and the accurate value to obtain an original model; Performing multi-level compression optimization on the original model through deep learning pruning or knowledge distillation to obtain a multi-level compressed optimized model; The original data subsequently collected by the sensor is calibrated according to the original model or the model after multi-stage compression optimization.

Images