Method for online detection of measured object by using metering device of Internet of Things system

Abstract

The invention, which relates to the technical field of the metering instrument, discloses a method for on-line detection of a detected object by a metering instrument using an Internet of Things system, the Internet of Things system at least comprises a sensor sensing module, an edge computing module, a compatible communication module, a cloud platform and a user terminal. Physical parameters are collected in real time through a sensor deployed on a measured object, a sensor sensing module receives sensor data, the sensor data are preprocessed through an edge computing module, the preprocessed data are transmitted to a cloud platform through a compatible communication module, and the cloud platform is used for monitoring the measured object. The cloud platform receives and uses the distributed database to store data, constructs an anomaly detection model of a detected object through a self-supervised learning algorithm model based on DINO, identifies the anomaly condition of the detected object, integrates the anomaly detection model of the detected object into the metering instrument through the compatible communication interface, and stores the anomaly detection model of the detected object in the metering instrument. The detection capability of the metering instrument is improved.

Description

Technical Field

[0001] The present invention relates to metering instrument technology, and more specifically to a method for online detection of a measured object by a metering instrument of an Internet of Things system. Background Art

[0002] With the rapid development of Internet of Things technology, metering instruments have been widely used in many fields such as industry, agriculture, and energy. These instruments provide important support for the efficient operation and management of the system by real-time monitoring and control of various physical parameters such as temperature, pressure, and flow. However, there are still some problems that need to be solved in the practical application of traditional metering instrument systems: Traditional systems rely on fixed wired or wireless transmission protocols, which can easily lead to low data transmission efficiency or even data loss when the network is unstable or the transmission distance is long; traditional anomaly detection methods mostly use rule-based or simple statistical analysis methods, which are difficult to cope with complex and changeable actual working conditions, resulting in poor accuracy and reliability of anomaly detection, and prone to false alarms or missed alarms; traditional systems usually lack real-time alarm and remote control functions, and users need to check the status of equipment regularly. They cannot detect and handle abnormal situations in time, which may lead to equipment failure or production safety hazards; many existing systems lack intelligent analysis and decision support functions and cannot fully utilize the advantages of IoT technology. Summary of the invention

[0003] In view of the shortcomings of the above-mentioned technology, the purpose of the present invention is to provide a method for online detection of the measured object by a metering instrument of an Internet of Things system, utilizing the Internet of Things system and a cloud platform, and improving the measurement capability through a self-supervised learning algorithm based on DINO.

[0004] To achieve the above purpose, the technical solution adopted by the present invention is as follows: A method for online detection of a measured object by a metering instrument of an Internet of Things system, wherein the Internet of Things system comprises at least a sensor perception module, an edge computing module, a compatible communication module, a cloud platform and a user terminal, comprising the following steps: Step 1: collect physical parameters in real time through sensors deployed on the object to be measured, and transmit the data to the sensor sensing module through the wireless communication module, wherein the physical parameters include at least temperature, pressure, flow, frequency, power factor, active power, reactive power and harmonic content; Step 2: The sensor perception module receives sensor data, pre-processes the sensor data through the edge computing module, and transmits the pre-processed data to the cloud platform through the compatible communication module, wherein the pre-processing includes denoising, filtering, compression and data classification calculation; Step 3: The compatible communication module dynamically adjusts the transmission parameters according to the network conditions through the adaptive communication protocol; Step 4, the cloud platform receives and uses a distributed database to store data, and constructs an anomaly detection model of the object under test through a DINO-based self-supervised learning algorithm model to identify abnormal conditions of the object under test. The DINO-based self-supervised learning algorithm model includes a data receiving module, a feature extraction module, a DINO self-supervised learning module, an anomaly detection module, a dynamic target update module, a multi-task learning module, an anomaly output module and a compatible communication interface; Step 5: Integrate the abnormality detection model of the measured object into the meter through the compatible communication interface, and the meter immediately sends an alarm notification to the user when abnormal data is detected; Step 6: The user terminal receives the alarm notification through the supporting application, and uses the visual monitoring function of the application to view the status and historical data of the measured object in real time; Step seven: The user adjusts and operates the object to be measured through remote control according to the state of the object to be measured.

[0005] As a further embodiment of the present invention, in step three, the working method of the adaptive communication protocol is: Step (1), obtaining network bandwidth usage and network quality monitored by network sensors in real time; Step (2), marking the data packet by defining the transmission priority of different types of data, and adding a checksum at the end of the data packet. During the data transmission process, scheduling is performed according to the preset priority; Step (3), the receiver verifies the data packet according to the check code. When the receiver detects an error or missing data packet, it automatically sends a retransmission request through the automatic repeat request ARQ. Step (4), using multiple paths for data transmission, and selecting the best transmission path through a load balancer according to network congestion, path cost, and data transmission requirements; Step (5): adjusting the protocol parameters through the congestion control algorithm BBR according to the real-time network status, wherein the adjusted parameters at least include the congestion window, the sending rate, the minimum RTT and the gain coefficient.

[0006] As a further embodiment of the present invention, in the step 4, the data receiving module is used to read the physical parameters of the measured object received by the cloud platform, and temporarily store the data by using the cache space; The feature extraction module is used to extract data features through an autoencoder and reduce data dimensions through recursive feature elimination; the DINO self-supervised learning module is used to extract normal data feature representation by deploying a static teacher network and a dynamic student network through a dual network architecture; The anomaly detection module is used to set an anomaly detection threshold by calculating the mean and standard deviation of the difference measurement of the normal data feature representation, and identify the abnormal data and the degree of abnormality of the abnormal data according to the threshold; The dynamic target update module is used to adjust the parameters and weights of the static teacher network and the dynamic student network through the gradient descent method according to the user and system feedback data; The multi-task learning module is used to process multiple learning tasks in parallel and optimize the joint learning process by learning the dependency relationship between different tasks through association rules; The anomaly output module is used to perform multi-level scoring according to the degree of anomaly, and integrate the rule engine to interpret and output the anomaly detection results; The compatible communication interface is used to exchange data with an external system via a compatible RESTful API; The output end of the data receiving module is connected to the input end of the feature extraction module, the output end of the feature extraction module is connected to the input end of the DINO self-supervised learning module, the output end of the DINO self-supervised learning module is connected to the input end of the anomaly detection module, the output end of the anomaly detection module is connected to the input ends of the anomaly output module and the dynamic target update module, the output end of the dynamic target update module is connected to the input end of the DINO self-supervised learning module, the multi-task learning module is bidirectionally connected to the DINO self-supervised learning module, and the output end of the anomaly output module is connected to the input end of the compatible communication interface.

[0007] As a further embodiment of the present invention, The DINO-based self-supervised learning algorithm model includes a data receiving module, a feature extraction module, a DINO self-supervised learning module, an anomaly detection module, a dynamic target update module, a multi-task learning module, an anomaly output module and a compatible communication interface; wherein the output end of the data receiving module is connected to the input end of the feature extraction module, the output end of the feature extraction module is connected to the input end of the DINO self-supervised learning module, the output end of the DINO self-supervised learning module is connected to the input end of the anomaly detection module, the output end of the anomaly detection module is connected to the input end of the dynamic target update module, the output end of the dynamic target update module is connected to the input end of the multi-task learning module, the output end of the multi-task learning module is connected to the input end of the anomaly output module, and the output end of the anomaly output module is connected to the input end of the compatible communication interface; The working method of the DINO self-supervised learning module is: Step (1), obtain the feature data set D, and randomly select 80% of the feature data set as the training set , 20% as test set ; Step (2), constructing a static teacher network T and a dynamic student network S, wherein the parameters of the static teacher network T remain fixed, and the parameters of the dynamic student network S are updated through iterative training; Step (3): the training set The data x in is input into the static teacher network T and the dynamic student network S; the feature generated by the static teacher network T is represented as ; The feature representation generated by the dynamic student network S is ; Step (4), using the loss function L to measure the difference between the feature representations generated by the static teacher network T and the dynamic student network S, the formula of the loss function L is: In formula (1), is the feature representation of the static teacher network for the ith data point, is the feature representation of the dynamic student network of the ith data point, is the Euclidean distance, is the cross entropy loss, is the KL divergence, P is the probability distribution of feature representation, and are the weight parameters of cross entropy loss and KL divergence respectively; Step (5), using the momentum gradient descent algorithm to minimize the loss function L, update the parameters of the dynamic student network S, the update formula is: In formula (2), is the momentum term at the current moment, is the momentum coefficient, is the loss function L on the dynamic student network parameters The gradient of is the learning rate, are the parameters of the dynamic student network; Step (6), repeat steps (3), (4) and (5) to perform iterative training, so that the feature representation of the dynamic student network S Gradually approach the feature representation of the static teacher network T ; Step (7), calculate the feature representation of the dynamic student network S by Euclidean distance And the feature representation of the static teacher network T When the similarity reaches the set threshold, the feature representation of the dynamic student network S is output. It is the normal data characteristic representation of the object being tested.

[0008] As a further embodiment of the present invention, the working method of the anomaly detection module is as follows: step (1), using the loss function L of the DINO self-supervised learning module as the difference metric d of the normal data feature representation, and setting the threshold of anomaly detection by calculating the mean and standard deviation of the difference metric d, the threshold formula is: In formula (3), is the threshold for anomaly detection, N is the number of normal data, is the difference measure of the i-th normal data, is the mean of the difference measure for normal data, is the standard deviation of the difference measure of normal data, and k is the set threshold coefficient; Step (2), input the new data collected by the sensor in real time into the DINO self-supervised learning module to obtain input data The feature representation and , and obtain the current input data by calculating the loss function L The difference measure; Step (3), when the input data The difference measure Greater than threshold , determine the input data is the abnormal data, and calculates the input data The abnormality degree is: In formula (4), A is the input data The degree of abnormality, is the current input data The difference measure of is the mean of the variance measure for normal data, is the standard deviation of the variance measure for normal data.

[0009] As a further embodiment of the present invention, the multi-task learning module decomposes the training process of the DINO self-supervised learning module into multiple subtasks through a task scheduling algorithm. Each subtask is independently processed and executed in parallel, and the resource allocation of the task is dynamically adjusted according to the utilization of real-time computing resources and the urgency of the task. During the task learning process, the multi-task learning module creates a shared memory to share the intermediate feature representation and learning results of each subtask, thereby reducing redundant calculations.

[0010] As a further embodiment of the present invention, in step six, the visual monitoring function intuitively displays data through charts and dashboards, and supports playback and analysis of real-time data and historical data. The data analysis conducts a comprehensive evaluation of the data through statistical analysis and trend analysis, and generates data reports through the automated report generation tool PythonReportLab to provide decision support information.

[0011] As a further embodiment of the present invention, in step seven, the remote control is implemented jointly by a remote controlled module installed on the object under test and a remote control module installed on the user terminal, the remote control module includes automatic control and manual control, the automatic control mode automatically adjusts the parameters of the object under test according to the state of the object under test through a preset control strategy, the manual control mode manually sets the parameters of the object under test through the operation interface of the remote control module, and the remote controlled module receives instructions from the remote control module and performs corresponding operations.

[0012] As a further embodiment of the present invention, The dynamic target update module improves the data update capability through the target adaptation function; the working principle of the target adaptation function is: setting the target adaptation data vertex ,Adaptation Path Set to: In formula (5), is the data set before data adaptation, Map data features to data and calculate functions after adaptation for: In formula (6), represents the unit vector of the load characteristic quantity during adaptation, e represents the load characteristic quantity during adaptation, represents the angle between the adaptation path update path and the standard setting data path, The throughput that expresses the data update time, represents the external environment influencing conditions, v represents the data type, x represents the parameter adjustment data information, and P represents the data distribution auxiliary parameter. is the adaptation component feature.

[0013] As a further technical solution of the present invention, the working principle of the multi-task learning module is: Identify and define the attributes or labels of the temperature, pressure, flow, frequency, power factor, active power, reactive power and harmonic content data information in the instrument measurement object. These tasks can be of different types such as classification, regression, and detection. The shared part in data features or knowledge is defined as the current viewpoint. On the basis of shared parameters, each parameter learning task sets parameters with distinguishing technical characteristics. By jointly training all tasks, the gradients between different tasks will be shared, and the shared tasks will be recorded as expected viewpoints. Task consistency regularization ensures that useful features are extracted instantly by dynamically updated data tasks, and the correlation of different data information is learned through correlation functions.

[0014] The present invention collects physical parameters in real time through sensors deployed on the measured object, and transmits the data to the sensor perception module through the wireless communication module; the sensor perception module receives the sensor data, pre-processes the sensor data through the edge computing module, and transmits the pre-processed data to the cloud platform through the compatible communication module. Through the real-time data collection and edge computing processing of the Internet of Things system, the real-time and accuracy of data processing are improved, and abnormal situations can be responded to and handled quickly; the compatible communication module dynamically adjusts the transmission parameters according to the network conditions through the adaptive communication protocol, optimizes the data transmission process, reduces data loss and errors, and improves the overall communication efficiency; the cloud platform receives and uses a distributed database to store data, and uses a self-supervised learning algorithm model based on DINO The method constructs an anomaly detection model of the measured object, identifies the abnormal situation of the measured object, improves the accuracy of anomaly detection through self-supervised learning, and promptly warns the user, thereby reducing potential risks and losses; the anomaly detection model of the measured object is integrated into the metering instrument through the compatible communication interface, and the metering instrument immediately sends an alarm notification to the user when abnormal data is detected; the user terminal receives the alarm notification through the supporting application, and views the status and historical data of the measured object in real time through the visual monitoring function of the application; the user adjusts and operates the measured object through remote control according to the status of the measured object, so that the user can manage the measured object more intuitively and conveniently, optimize the operation process and improve the user experience, and realize efficient, accurate and intelligent online detection of the measured object by the metering instrument. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 A schematic diagram of the working process of a method for online detection of a measured object by a metering instrument of an Internet of Things system according to the present invention; Figure 2 A schematic diagram of an adaptive communication protocol workflow of a method for online detection of a measured object by a metering instrument of an Internet of Things system according to the present invention; Figure 3A structural schematic diagram of a self-supervised learning algorithm model based on DINO in a method for online detection of a measured object by a metering instrument of an Internet of Things system according to the present invention; Figure 4 The present invention is a schematic diagram of the workflow of the DINO self-supervised learning module of a method for online detection of a measured object using a metering instrument of an Internet of Things system.

[0016] Figure 5 The present invention is a schematic diagram of the working process of an abnormality detection module of a method for online detection of a measured object by a metering instrument of an Internet of Things system.

[0017] Figure 6 It is a schematic diagram of the principle of a dynamic target update module of a method for online detection of a measured object by a metering instrument of an Internet of Things system according to the present invention; Figure 7 A schematic diagram of the principle of a multi-task learning module in a method for online detection of a measured object by a metering instrument of an Internet of Things system of the present invention. DETAILED DESCRIPTION

[0018] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0019] like Figure 1 As shown, the present invention provides a method for online detection of a measured object by a metering instrument of an Internet of Things system, wherein the Internet of Things system at least includes a sensor perception module, an edge computing module, a compatible communication module, a cloud platform and a user terminal, and includes the following steps: Step 1: collect physical parameters in real time through sensors deployed on the object to be measured, and transmit the data to the sensor sensing module through the wireless communication module, wherein the physical parameters include at least temperature, pressure, flow, frequency, power factor, active power, reactive power and harmonic content; Step 2: The sensor perception module receives sensor data, pre-processes the sensor data through the edge computing module, and transmits the pre-processed data to the cloud platform through the compatible communication module, wherein the pre-processing includes denoising, filtering, compression and data classification calculation; Step 3: The compatible communication module dynamically adjusts the transmission parameters according to the network conditions through the adaptive communication protocol; Step 4, the cloud platform receives and uses a distributed database to store data, and constructs an anomaly detection model of the object under test through a DINO-based self-supervised learning algorithm model to identify abnormal conditions of the object under test. The DINO-based self-supervised learning algorithm model includes a data receiving module, a feature extraction module, a DINO self-supervised learning module, an anomaly detection module, a dynamic target update module, a multi-task learning module, an anomaly output module and a compatible communication interface; Step 5: Integrate the abnormality detection model of the measured object into the meter through the compatible communication interface, and the meter immediately sends an alarm notification to the user when abnormal data is detected; Step 6: The user terminal receives the alarm notification through the supporting application, and uses the visual monitoring function of the application to view the status and historical data of the measured object in real time; Step seven: The user adjusts and operates the object to be measured through remote control according to the state of the object to be measured.

[0020] In a specific embodiment, in the step one, the sensors deployed on the object to be measured include a temperature sensor, a pressure sensor and a flow sensor. The temperature sensor adopts DS18B20, and the measuring range is -55°C to +125°C, and the accuracy is ±0.5°C. The pressure sensor adopts BMP280, and the measuring range is 300 hPa to 1100 hPa, and the accuracy is ±1 hPa. The flow sensor is a YF-S201 liquid flow sensor, and the measuring range is 1-30 L / min, and the accuracy is ±1%. When the sensors are deployed, the temperature sensor is installed at the key temperature monitoring point of the object to be measured to ensure good contact with the surface or interior of the object to be measured. The pressure sensor is installed on the pressure pipe or container of the object to be measured to ensure sealing and stability. The flow sensor is installed in the fluid pipe of the object to be measured to ensure that the fluid can pass through the sensor stably.

[0021] In a specific embodiment, in step 2, the sensor sensing module supports multiple sensor interfaces, including I2C, SPI, UART and 1-Wire, etc., and can receive data from different types of sensors. Data denoising uses a low-pass filter to remove high-frequency noise in the sensor data, data filtering uses a sliding average filter to smooth the data and reduce data fluctuations, and data compression uses Huffman coding to compress the data, reduce the amount of data transmission, and improve transmission efficiency.

[0022] In the above embodiment, if Figure 2 As shown, in the step three, the working method of the adaptive communication protocol is: step (1), real-time acquisition of network bandwidth usage and network quality monitored by network sensors; Step (2), marking the data packet by defining the transmission priority of different types of data, and adding a checksum at the end of the data packet. During the data transmission process, scheduling is performed according to the preset priority; Step (3), the receiver verifies the data packet according to the check code. When the receiver detects an error or missing data packet, it automatically sends a retransmission request through the automatic repeat request ARQ. Step (4), using multiple paths for data transmission, and selecting the best transmission path through a load balancer according to network congestion, path cost, and data transmission requirements; Step (5): adjusting protocol parameters through the congestion control algorithm BBR according to the real-time network status, wherein the adjusted parameters at least include congestion window, sending rate, minimum RTT and gain coefficient.

[0023] In a specific embodiment, the network sensor uses a network interface controller that supports SNMP to obtain network bandwidth usage in real time, evaluates network quality by measuring parameters such as network delay, packet loss rate and jitter, and uses multi-WAN routers to achieve data transmission on multiple paths through MPTCP.

[0024] In the above embodiment, if Figure 3 As shown, in the step 4, the data receiving module is used to read the physical parameters of the object under test received by the cloud platform, and temporarily store the data by using the cache space; The feature extraction module is used to extract data features through an autoencoder and reduce data dimensions through recursive feature elimination; The DINO self-supervised learning module is used to extract normal data feature representation by deploying a static teacher network and a dynamic student network through a dual network architecture; The anomaly detection module is used to set an anomaly detection threshold by calculating the mean and standard deviation of the difference measurement of the normal data feature representation, and identify the abnormal data and the degree of abnormality of the abnormal data according to the threshold; The dynamic target update module is used to adjust the parameters and weights of the static teacher network and the dynamic student network through the gradient descent method according to the user and system feedback data; The multi-task learning module is used to process multiple learning tasks in parallel and optimize the joint learning process by learning the dependency relationship between different tasks through association rules; The anomaly output module is used to perform multi-level scoring according to the degree of anomaly, and integrate the rule engine to interpret and output the anomaly detection results; The compatible communication interface is used to exchange data with an external system via a compatible RESTful API; The output end of the data receiving module is connected to the input end of the feature extraction module, the output end of the feature extraction module is connected to the input end of the DINO self-supervised learning module, the output end of the DINO self-supervised learning module is connected to the input end of the anomaly detection module, the output end of the anomaly detection module is connected to the input ends of the anomaly output module and the dynamic target update module, the output end of the dynamic target update module is connected to the input end of the DINO self-supervised learning module, the multi-task learning module is bidirectionally connected to the DINO self-supervised learning module, and the output end of the anomaly output module is connected to the input end of the compatible communication interface.

[0025] In a specific embodiment, Figure 4 As shown, the working method of the DINO self-supervised learning module is: Step (1), obtain the feature data set D, and randomly select 80% of the feature data set as the training set , 20% as test set ; Step (2), constructing a static teacher network T and a dynamic student network S, wherein the parameters of the static teacher network T remain fixed, and the parameters of the dynamic student network S are updated through iterative training; Step (3): the training set The data x in is input into the static teacher network T and the dynamic student network S; the feature generated by the static teacher network T is represented as ; The feature representation generated by the dynamic student network S is ; Step (4), using the loss function L to measure the difference between the feature representations generated by the static teacher network T and the dynamic student network S, the formula of the loss function L is: In formula (1), is the feature representation of the static teacher network of the i-th data point, which is used to provide a stable target feature representation; is the feature representation of the dynamic student network of the i-th data point, which is used to gradually approach the target feature representation through training; is the Euclidean distance, which is used to measure the difference between the feature representations generated by the static teacher network and the dynamic student network; is the cross entropy loss, which is used to measure the difference between the feature representations generated by the static teacher network and the dynamic student network in the classification task; is the KL divergence, which is used to measure the difference between the probability distributions of feature representations generated by the static teacher network and the dynamic student network; P is the probability distribution of feature representations, which is used to convert feature representations into probability distributions for the use of cross entropy and KL divergence; and They are the weight parameters of cross entropy loss and KL divergence, which are used to control the relative importance of cross entropy loss and KL divergence in the total loss; Step (5), using the momentum gradient descent algorithm to minimize the loss function L, update the parameters of the dynamic student network S, the update formula is: In formula (2), is the momentum term at the current moment, which is used to represent the update amount of the dynamic student network parameters at the current moment t; is the momentum coefficient, which is used to control the influence of the momentum term at the previous moment on the current update. The momentum coefficient usually takes a value between 0 and 1. A larger momentum coefficient can accelerate convergence and reduce oscillation; is the loss function L on the dynamic student network parameters The gradient of the loss function L with respect to the dynamic student network parameters The gradient of indicates the direction of change of the loss function in the parameter space; The learning rate is used to control the step size of each parameter update and determines the speed of parameter update. A learning rate that is too large may lead to unstable training, while a learning rate that is too small may lead to slow convergence. are the parameters of the dynamic student network, which are used to gradually optimize the performance of the dynamic student network by updating these parameters; Step (6), repeat steps (3), (4) and (5) to perform iterative training, so that the feature representation of the dynamic student network S Gradually approach the feature representation of the static teacher network T ; Step (7), calculate the feature representation of the dynamic student network S by Euclidean distance And the feature representation of the static teacher network T When the similarity reaches the set threshold, the feature representation of the dynamic student network S is output. It is the normal data characteristic representation of the object being tested.

[0026] In a specific embodiment, Figure 5 As shown, the working method of the anomaly detection module is as follows: Step (1), using the loss function L of the DINO self-supervised learning module as the difference metric d of the normal data feature representation, and setting the threshold of anomaly detection by calculating the mean and standard deviation of the difference metric d. The threshold formula is: In formula (3), is the threshold of anomaly detection, which is used to determine whether a data point is abnormal. When the difference metric of a data point exceeds this threshold, it is considered abnormal. N is the number of normal data, which indicates the total number of normal data points used to calculate the mean and standard deviation. is the difference metric of the i-th normal data, which represents the difference metric of each data point calculated by the loss function of the DINO self-supervised learning module; It is the mean of the difference measure of normal data, which represents the average value of the difference measure of all normal data points and describes the central tendency of the difference measure of normal data points; It is the standard deviation of the difference measure of normal data, indicating the degree of dispersion of the difference measure of normal data points, and describing the distribution range of the difference measure of normal data points; k is the set threshold coefficient, which is used to control the sensitivity of the threshold. By adjusting the value of k, the sensitivity of anomaly detection can be changed. A larger k value will make the threshold higher, thereby reducing false positives, but may miss some anomalies; a smaller k value will make the threshold lower, thereby increasing the number of detected anomalies, but may increase false positives; Step (2), input the new data collected by the sensor in real time into the DINO self-supervised learning module to obtain input data The feature representation and , and obtain the current input data by calculating the loss function L The difference measure of Step (3), when input data The difference measure Greater than threshold , determine the input data is the abnormal data, and calculates the input data The abnormality degree is: In formula (4), A is the input data The degree of abnormality is used to indicate the degree of abnormality of the input data relative to the normal data. The greater the degree of abnormality, the more the input data deviates from the normal data distribution. is the current input data , which represents the input data point calculated by the loss function of the DINO self-supervised learning module The difference measure of It is the mean of the difference measure of normal data, which represents the average value of the difference measure of all normal data points and describes the central tendency of the difference measure of normal data points; It is the standard deviation of the difference measure of normal data, which indicates the degree of dispersion of the difference measure of normal data points and describes the distribution range of the difference measure of normal data points.

[0027] In a specific embodiment, the anomaly detection module performs anomaly detection on 5 detected objects, and the detection results are recorded as follows: In a specific embodiment, the abnormal level includes: Normal to abnormal process , no abnormality is detected in the measured object, and the physical parameters of the measured object are within the normal range; Mild abnormality is abnormal degree , a slight abnormality is detected in the object being measured, and the physical parameters of the object being measured slightly deviate from the normal range; Moderate abnormality is abnormal , the measured object detected moderate abnormality, the physical parameters of the measured object obviously deviated from the normal range and needed further monitoring; Severe abnormality is abnormal degree , serious abnormalities are detected in the object being measured, the physical parameters of the object being measured seriously deviate from the normal range, there is an obvious safety threat, and safety measures need to be taken immediately.

[0028] In the above embodiment, the multi-task learning module decomposes the training process of the DINO self-supervised learning module into multiple subtasks through a task scheduling algorithm. Each subtask is processed independently and executed in parallel, and the resource allocation of the task is dynamically adjusted according to the utilization of real-time computing resources and the urgency of the task. During the task learning process, the multi-task learning module creates a shared memory to share the intermediate feature representation and learning results of each subtask, thereby reducing redundant calculations.

[0029] In a specific embodiment, the multi-task learning module ensures efficient processing of large amounts of data with the support of high-performance server configurations, AMD EPYC series multi-core processors, NVIDIA Tesla V100 model GPUs, high-speed memory and storage devices, and high-speed network environments.

[0030] In the above embodiment, in step six, the visual monitoring function intuitively displays data through charts and dashboards, and supports playback and analysis of real-time data and historical data. The data analysis conducts a comprehensive evaluation of the data through statistical analysis and trend analysis, and generates data reports through the automated report generation tool PythonReportLab to provide decision support information.

[0031] In a specific embodiment, the visual monitoring function builds an interactive dashboard through the modern visualization tool Grafana to display key performance indicators, anomaly detection results and system status. Real-time data is transmitted and processed through Apache Flink to ensure the real-time and accuracy of the data. The data analysis part conducts a comprehensive evaluation of the data through statistical analysis and trend analysis. Statistical analysis is performed through descriptive statistics, correlation analysis and regression analysis, and trend analysis is performed through time series analysis, moving average and seasonal decomposition methods. The visual monitoring function generates a comprehensive data report through the automated report generation tool PythonReportLab. The report includes detailed statistical charts, trend analysis results and anomaly detection summaries, which can provide decision support information, and the report can be generated regularly and automatically sent to relevant personnel, or generated in real time according to demand, to ensure that decision makers can obtain the latest system status and analysis results in a timely manner.

[0032] In the above embodiment, in step seven, the remote control is implemented by a remote controlled module installed on the object under test and a remote control module installed on the user terminal. The remote control module includes automatic control and manual control. The automatic control mode automatically adjusts the parameters of the object under test according to the state of the object under test through a preset control strategy. The manual control mode manually sets the parameters of the object under test through the operation interface of the remote control module. The remote controlled module receives instructions from the remote control module and performs corresponding operations.

[0033] In a specific embodiment, the automatic control mode of the remote control monitors the status data of the object under test in real time through sensors and anomaly detection models of the object under test. The status data of the object under test is transmitted to the remote control module through a compatible communication module of the Internet of Things system. The remote control module applies a preset control strategy based on the status data of the object under test to automatically adjust the parameters of the object under test. The remote control module generates a control instruction and sends it to the remote controlled module through a secure encrypted communication channel TLS / SSL. The remote controlled module receives the instruction and performs corresponding operations.

[0034] In a specific embodiment, the manual control mode of the remote control manually sets the parameters of the object under test through the remote control module operation interface installed on the user terminal. The operation interface is in the form of a Web application or a mobile application APP to provide a friendly and intuitive user experience. After the user sets the parameters on the operation interface, the remote control module sends the manual control instructions to the remote controlled module through the compatible communication module of the Internet of Things system. The remote controlled module receives the instructions and performs corresponding operations to ensure that the object under test operates according to the parameters set by the user.

[0035] In a specific embodiment, the accuracy comparison results of the method for online detection of the measured object by a metering instrument using an Internet of Things system of the present invention and the method for online detection of the measured object by a traditional metering instrument are as follows: The above experiments show that the method of this study has positive and beneficial technical effects.

[0036] The working principle of the dynamic target update module is: The main function of the dynamic target update module is to update and calculate different types of detection data when measuring the instrument of the measured object, so as to dynamically sense the measured data object in real time and comprehensively evaluate the working status of the measured object.

[0037] like Figure 6 As shown, the data information of the measured object is first mapped in the three-dimensional XY plane to train different types of parameter data information, such as temperature, pressure, flow, frequency, power factor, active power, reactive power and harmonic content in the instrument measurement object, and these data information are processed and divided into A grid in three-dimensional space The grid is . It represents the location or point of different types of data information in coordinates. The data parameter grid range is It represents the position of different data information in coordinates or the distance between points, and in this way, dynamic update and calculation of different types of data can be achieved. The closer the area, the higher the accuracy requirement of the spatiotemporal data loaded by the IoT sensing data information. Choose the appropriate grid segmentation distance Direction splits the grid into a spatiotemporal grid of data attributes before classification and the spatiotemporal grid of classified data attributes Filter out redundant data information.

[0038] Dynamic target update is the dynamic target update of the detected parameter data information. During the data loading process, the accelerator, data mapping recognition protocol module, data resolution calculation module and data mapping conversion protocol module are used to dynamically obtain the updated data 24 hours a day, and feature learning is performed on the updated data.

[0039] In a further specific embodiment, the data loading process is that the system continuously loads new data information so as to update the target parameters in real time. The accelerator module is responsible for optimizing the data processing process and improving the efficiency of data loading and processing. This may include parallel processing, caching mechanism or algorithm optimization. The data mapping identification protocol module is responsible for parsing and identifying the target information in the data. It may use specific protocols or standards to ensure the consistency and accuracy of the data. The data resolution calculation module calculates and determines the data resolution according to the characteristics and requirements of the data. This may involve the evaluation of data quality, accuracy and real-time performance. The data mapping conversion protocol module is responsible for converting the raw data into a format that can be processed within the system. This may include steps such as data standardization, format conversion or data cleaning. 24-hour dynamic acquisition of updated data runs around the clock, monitors the data source in real time, and ensures that the latest data can be obtained in a timely manner. When updating data processing, the updated data obtained is processed, which may include data verification, error detection and exception handling. Machine learning or deep learning algorithms are used during feature learning to extract and identify key features from the updated data. These features are essential for target detection, classification or other subsequent analysis tasks.

[0040] Through this dynamic update mechanism, the system can maintain the real-time and accuracy of data, thereby improving decision-making efficiency and system performance. Dynamically reflect the above data attributes in three-dimensional space. And improve the learning ability of different types of data through the adaptation module. The dynamic target update module improves the data update ability through the target adaptation function; the working principle of the target adaptation function is: Set the target adapter data vertex ,Adaptation Path Set to: In formula (5), is the data set before data adaptation, Map data features to data and calculate functions after adaptation for: In formula (6), represents the unit vector of the load characteristic quantity during adaptation, e represents the load characteristic quantity during adaptation, represents the angle between the adaptation path update path and the standard setting data path, The throughput that expresses the data update time, represents the external environment influencing conditions, v represents the data type, x represents the parameter adjustment data information, and P represents the data distribution auxiliary parameter. is the adaptation component feature.

[0041] In a specific embodiment, feature adaptation and parameter adjustment are common steps to improve the performance of the model. In data mining tasks, feature selection and conversion are key steps to extract useful information. Feature adaptation may be used to improve the quality of the signal or extract important signal components. The data information such as temperature, pressure, flow, frequency, power factor, active power, reactive power and harmonic content in the instrument measurement object can reflect the operating status of different instruments from different angles. In a specific embodiment, the programmable controller is adapted to adjust the data type or data information in a timely manner.

[0042] like Figure 7 As shown, the working principle of the multi-task learning module is: Identify and define the attributes or labels of the temperature, pressure, flow, frequency, power factor, active power, reactive power and harmonic content data information in the instrument measurement object. These tasks can be of different types such as classification, regression, and detection. The shared part in data features or knowledge is defined as the current viewpoint. On the basis of shared parameters, each parameter learning task sets parameters with distinguishing technical characteristics. By jointly training all tasks, the gradients between different tasks will be shared, and the shared tasks will be recorded as expected viewpoints. Task consistency regularization ensures that useful features are extracted instantly by dynamically updated data tasks, and the correlation of different data information is learned through correlation functions.

[0043] In a further specific embodiment, the module first identifies and defines multiple data attributes or labels of the measured object in the metering, which may include temperature, pressure, flow, frequency, power factor, active power, reactive power and harmonic content. These tasks may involve different types, such as classification (identifying which category the data belongs to), regression (predicting numerical or continuous variables), detection (detecting whether a specific condition or anomaly exists), etc. In the shared feature or knowledge learning task, there is a shared part of the data features or knowledge. These shared features or knowledge are defined as "current viewpoints", which means that they are important for multiple tasks. On the basis of shared parameters, each parameter learning task sets parameters with distinguishing technical characteristics. This means that each task has its own specific parameters to handle the task-specific parts. These specific parameters allow the model to be fine-tuned for each task while still using shared parameters to capture common patterns in the data. All tasks are performed through joint training, which means that the model optimizes the parameters of all tasks simultaneously during the training process. In joint training, the gradients between different tasks are shared, which helps information transfer and learning between tasks. Shared tasks are recorded as "expected viewpoints", that is, the model expects to achieve the performance of all tasks under shared parameters. Task consistency regularization is used to ensure that the performance of all tasks is consistent under shared parameters, which helps prevent some tasks from overperforming on shared parameters and harming other tasks. In dynamic data extraction and feature learning, useful features are extracted instantly from dynamically updated data tasks, which means that the model can adapt to changes in data and learn from new data. Through the association function (which may be a series of feature associations or similarity measures), the model learns the associations between different data information.

[0044] This working principle enables the multi-task learning module to more effectively utilize the shared information in the data while processing the specific requirements of each task, thereby improving the performance of the overall model. In this way, the multi-task learning module can effectively utilize the correlation between tasks and improve the performance of the model on multiple tasks, thereby improving the learning ability of the instrument being measured.

[0045] Although the specific embodiments of the present invention are described above, it should be understood by those skilled in the art that these specific embodiments are only illustrative, and those skilled in the art may omit, replace, and change the details of the above methods and systems in various ways without departing from the principles and essence of the present invention. For example, merging the above method steps so as to perform substantially the same functions in substantially the same manner to achieve substantially the same results is within the scope of the present invention. Therefore, the scope of the present invention is limited only by the appended claims.

Claims

1. A method for online detection of a measured object using a metering instrument of an Internet of Things system, characterized in that: The Internet of Things system at least includes a sensor perception module, an edge computing module, a compatible communication module, a cloud platform and a user terminal. The method for online detection of the object under test includes the following steps: Step 1: collect physical parameters in real time through sensors deployed on the object to be measured, and transmit the data to the sensor sensing module through the wireless communication module, wherein the physical parameters include at least temperature, pressure, flow, frequency, power factor, active power, reactive power and harmonic content; Step 2: The sensor perception module receives sensor data, pre-processes the sensor data through the edge computing module, and transmits the pre-processed data to the cloud platform through the compatible communication module, wherein the pre-processing includes denoising, filtering, compression and data classification calculation; Step 3: The compatible communication module dynamically adjusts the transmission parameters according to the network conditions through the adaptive communication protocol; Step 4: The cloud platform receives and uses the distributed database to store data, and builds an anomaly detection model for the object under test through a self-supervised learning algorithm model based on DINO to identify abnormal conditions of the object under test. Step 5: Integrate the abnormality detection model of the measured object into the meter through the compatible communication interface, and the meter immediately sends an alarm notification to the user when abnormal data is detected; Step 6: The user terminal receives the alarm notification through the supporting application, and uses the visual monitoring function of the application to view the status and historical data of the measured object in real time; Step seven: The user adjusts and operates the object to be measured through remote control according to the state of the object to be measured.

2. According to claim 1, a method for online detection of a measured object using a metering instrument of an Internet of Things system, characterized in that: In step three, the working method of the adaptive communication protocol is: Step (1), obtaining network bandwidth usage and network quality monitored by network sensors in real time; Step (2), marking the data packet by defining the transmission priority of different types of data, and adding a checksum at the end of the data packet. During the data transmission process, scheduling is performed according to the preset priority; Step (3), the receiver verifies the data packet according to the check code. When the receiver detects an error or missing data packet, it automatically sends a retransmission request through the automatic repeat request ARQ. Step (4), using multiple paths for data transmission, and selecting the best transmission path through a load balancer according to network congestion, path cost, and data transmission requirements; Step (5): adjusting the protocol parameters through the congestion control algorithm BBR according to the real-time network status, wherein the adjusted parameters at least include the congestion window, the sending rate, the minimum RTT and the gain coefficient.

3. According to claim 1, a method for online detection of a measured object using a metering instrument of an Internet of Things system, characterized in that: In the step 4, the DINO-based self-supervised learning algorithm model includes a data receiving module, a feature extraction module, a DINO self-supervised learning module, an anomaly detection module, a dynamic target update module, a multi-task learning module, an anomaly output module and a compatible communication interface; wherein the output end of the data receiving module is connected to the input end of the feature extraction module, the output end of the feature extraction module is connected to the input end of the DINO self-supervised learning module, the output end of the DINO self-supervised learning module is connected to the input end of the anomaly detection module, the output end of the anomaly detection module is connected to the input end of the dynamic target update module, the output end of the dynamic target update module is connected to the input end of the multi-task learning module, the output end of the multi-task learning module is connected to the input end of the anomaly output module, and the output end of the anomaly output module is connected to the input end of the compatible communication interface; The data receiving module is used to read the physical parameters of the measured object received by the cloud platform and temporarily store the data by using the cache space; The feature extraction module is used to extract data features through an autoencoder and reduce data dimensions through recursive feature elimination; The DINO self-supervised learning module is used to extract normal data feature representation by deploying a static teacher network and a dynamic student network through a dual network architecture; The anomaly detection module is used to set an anomaly detection threshold by calculating the mean and standard deviation of the difference measurement of the normal data feature representation, and identify the abnormal data and the degree of abnormality of the abnormal data according to the threshold; The dynamic target update module is used to adjust the parameters and weights of the static teacher network and the dynamic student network through the gradient descent method according to the user and system feedback data; The multi-task learning module is used to process multiple learning tasks in parallel and optimize the joint learning process by learning the dependency relationship between different tasks through association rules; The anomaly output module is used to perform multi-level scoring according to the degree of anomaly, and integrate the rule engine to interpret and output the anomaly detection results; The compatible communication interface is used to exchange data with an external system via a compatible RESTful API.

4. According to claim 3, a method for online detection of a measured object using a metering instrument of an Internet of Things system, characterized in that: The working method of the DINO self-supervised learning module is: Step (1), obtain the feature data set D, and randomly select 80% of the feature data set as the training set , 20% as test set ; Step (2), constructing a static teacher network T and a dynamic student network S, wherein the parameters of the static teacher network T remain fixed, and the parameters of the dynamic student network S are updated through iterative training; Step (3): the training set The data x in is input into the static teacher network T and the dynamic student network S; the feature generated by the static teacher network T is represented as ; The feature representation generated by the dynamic student network S is ; Step (4), using the loss function L to measure the difference between the feature representations generated by the static teacher network T and the dynamic student network S, the formula of the loss function L is: In formula (1), is the feature representation of the static teacher network for the ith data point, is the feature representation of the dynamic student network of the ith data point, is the Euclidean distance, is the cross entropy loss, is the KL divergence, P is the probability distribution of feature representation, and are the weight parameters of cross entropy loss and KL divergence respectively; Step (5), using the momentum gradient descent algorithm to minimize the loss function L, update the parameters of the dynamic student network S, the update formula is: In formula (2), is the momentum term at the current moment, is the momentum coefficient, is the loss function L on the dynamic student network parameters The gradient of is the learning rate, are the parameters of the dynamic student network; Step (6), repeat steps (3), (4) and (5) to perform iterative training, so that the feature representation of the dynamic student network S Gradually approach the feature representation of the static teacher network T ; Step (7), calculate the feature representation of the dynamic student network S by Euclidean distance And the feature representation of the static teacher network T When the similarity reaches the set threshold, the feature representation of the dynamic student network S is output. It is the normal data characteristic representation of the object being tested.

5. According to claim 3, a method for online detection of a measured object using a metering instrument of an Internet of Things system, characterized in that: The working method of the anomaly detection module is as follows: step (1), using the loss function L of the DINO self-supervised learning module as the difference metric d of the normal data feature representation, and setting the threshold of anomaly detection by calculating the mean and standard deviation of the difference metric d. The threshold formula is: In formula (3), is the threshold for anomaly detection, N is the number of normal data, is the difference measure of the i-th normal data, is the mean of the difference measure for normal data, is the standard deviation of the difference measure of normal data, and k is the set threshold coefficient; Step (2), input the new data collected by the sensor in real time into the DINO self-supervised learning module to obtain input data The feature representation and , and obtain the current input data by calculating the loss function L The difference measure; Step (3), when the input data The difference measure Greater than threshold , determine the input data is the abnormal data, and calculates the input data The abnormality degree is: In formula (4), A is the input data The degree of abnormality, is the current input data The difference measure of is the mean of the difference measure for normal data, is the standard deviation of the variance measure for normal data.

6. According to claim 3, a method for online detection of a measured object using a metering instrument of an Internet of Things system, characterized in that: The multi-task learning module decomposes the training process of the DINO self-supervised learning module into multiple subtasks through a task scheduling algorithm. Each subtask is processed independently and executed in parallel, and the resource allocation of the task is dynamically adjusted according to the utilization of real-time computing resources and the urgency of the task. During the task learning process, the multi-task learning module creates a shared memory to share the intermediate feature representation and learning results of each subtask, thereby reducing redundant calculations.

7. According to claim 1, a method for online detection of a measured object by a metering instrument using an Internet of Things system, characterized in that: In step six, the visual monitoring function intuitively displays data through charts and dashboards, and supports playback and analysis of real-time data and historical data. The data analysis conducts a comprehensive evaluation of the data through statistical analysis and trend analysis, and generates data reports through the automated report generation tool PythonReportLab to provide decision support information.

8. According to claim 1, a method for online detection of a measured object by a metering instrument using an Internet of Things system, characterized in that: In step seven, the remote control is implemented by a remote controlled module installed on the object under test and a remote control module installed on the user terminal. The remote control module includes automatic control and manual control. The automatic control mode automatically adjusts the parameters of the object under test according to the state of the object under test through a preset control strategy. The manual control mode manually sets the parameters of the object under test through the operation interface of the remote control module. The remote controlled module receives instructions from the remote control module and performs corresponding operations.

9. According to claim 1, a method for online detection of a measured object by a metering instrument using an Internet of Things system, characterized in that: The dynamic target update module improves the data update capability through the target adaptation function; the working principle of the target adaptation function is: Set the target adapter data vertex ,Adaptation Path Set to: In formula (5), is the data set before data adaptation, Map data features to data and calculate functions after adaptation for: In formula (6), represents the unit vector of the load characteristic quantity during adaptation, e represents the load characteristic quantity during adaptation, represents the angle between the adaptation path update path and the standard setting data path, The throughput that expresses the data update time, represents the external environment influencing conditions, v represents the data type, x represents the parameter adjustment data information, and P represents the data distribution auxiliary parameter. is the adaptation component feature.

10. According to claim 1, a method for online detection of a measured object by a metering instrument using an Internet of Things system, characterized in that: The working principle of the multi-task learning module is: Identify and define the attributes or labels of the temperature, pressure, flow, frequency, power factor, active power, reactive power and harmonic content data information in the instrument measurement object. These tasks can be of different types such as classification, regression, and detection. The shared part in data features or knowledge is defined as the current viewpoint. On the basis of shared parameters, each parameter learning task sets parameters with distinguishing technical characteristics. By jointly training all tasks, the gradients between different tasks will be shared, and the shared tasks will be recorded as expected viewpoints. Task consistency regularization ensures that useful features are extracted instantly by dynamically updated data tasks, and the correlation of different data information is learned through correlation functions.

Images