Intelligent sensor cooperative supervision method and system based on Internet of Things

Abstract

The invention discloses an intelligent sensor cooperative supervision method and system based on the Internet of Things. The method comprises the following steps: collecting heterogeneous original monitoring data of various intelligent sensors and carrying out unified time sequence alignment; constructing a unified semantic mapping relationship to obtain standardized data; constructing a sensor state representation model based on a neural controlled differential equation, and obtaining a continuous hidden state vector; integrating newly obtained data in real time to update continuous hidden state vectors and calculating fusion feature representation; the current cooperation state is discriminated in real time through an anomaly detection discriminator; and when an exception is detected, starting a cooperative event disposal strategy, generating a supervision linkage disposal instruction, and executing closed-loop supervision. According to the invention, the real-time collaborative supervision of heterogeneous intelligent sensor data is realized, and the efficiency and accuracy of cross-platform data collaborative supervision are improved.

Description

Technical Field

[0001] This invention relates to the fields of Internet of Things (IoT) and sensor network technology, and in particular to a method and system for collaborative monitoring of smart sensors based on IoT. Background Technology

[0002] In recent years, with the rapid development of IoT technology, smart sensors, as key devices in the IoT sensing layer, have been widely used in numerous fields such as industrial production, environmental monitoring, smart cities, and intelligent transportation. Typically, smart sensors are diverse in type, widely distributed across different monitoring scenarios, and characterized by their wide sensing range, strong real-time performance, and large information capacity. To achieve data sharing and collaborative monitoring among various smart sensors, a unified data format conversion or standardized processing technology is usually employed to map heterogeneous data to a unified semantic representation space. Then, machine learning or artificial neural network models are used to achieve data fusion and regulatory judgment.

[0003] Commonly used intelligent sensor collaborative monitoring methods in existing technologies include steps such as data temporal alignment, data semantic mapping, data fusion, and collaborative judgment. Data fusion is performed through manually defined rules or traditional neural network models. However, with the expansion of intelligent sensor networks and the increasing diversification of data types, existing technical solutions are gradually showing shortcomings in terms of real-time performance, accuracy, and generalization ability.

[0004] Therefore, how to provide a collaborative monitoring method and system for smart sensors based on the Internet of Things is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a collaborative monitoring method and system for intelligent sensors based on the Internet of Things. Addressing the problem of insufficient real-time collaborative discrimination capability in cross-platform collaborative monitoring of heterogeneous intelligent sensors in existing technologies, this invention proposes a technical solution based on unified alignment of data time sequence using neural controlled differential equations, standardized processing of semantic mapping and nonlinear scaling transformation, continuous-time state representation modeling, real-time discrimination of abnormal states, and coordinated handling. This invention possesses the advantages of high real-time performance, high accuracy of collaborative monitoring, and strong generalization ability.

[0006] The IoT-based smart sensor collaborative monitoring method according to embodiments of the present invention includes: Acquire heterogeneous raw monitoring data collected by various types of smart sensors, and perform time-series alignment processing to obtain preliminary aligned data; For the initially aligned data, a unified semantic mapping relationship is established to obtain standardized data; Based on standardized data, a continuous-time sensor state representation model is constructed to obtain a continuous hidden state vector; By using standardized data from newly acquired various types of smart sensors, the continuous hidden state vector is dynamically updated, and the fused feature representation is calculated. Based on the fusion feature representation, the trained anomaly detection discriminator is used to make real-time judgments on the current collaborative state, and the collaborative supervision anomaly judgment results are obtained. Based on the results of collaborative supervision anomaly identification, when an abnormal collaborative supervision status is identified, a pre-set collaborative event handling strategy is activated, and a joint supervision handling instruction is generated; Regulatory coordination instructions are issued to relevant execution terminals to conduct closed-loop supervision and coordinated response of the intelligent sensor network.

[0007] Optionally, obtaining the preliminary alignment data specifically involves: The system receives raw monitoring data collected at different times from multiple heterogeneous smart sensors through a smart sensor network transmission channel. The timestamp information and unique identifier information of the received raw monitoring data are parsed to extract the data sequence of the collection time corresponding to each smart sensor; Select the data sequence with the smallest time granularity as the reference sequence, and use the timestamp of the reference sequence as the reference time node to generate a target alignment sequence with a unified time dimension. Interpolation processing is performed on other smart sensor data sequences whose timestamps are inconsistent with the reference time node to form an interpolated data sequence that is consistent with the target aligned sequence. Based on the unique identification information of the smart sensor, the interpolated data sequence and the data points of the reference sequence are matched one by one to obtain preliminary aligned data.

[0008] Optionally, the step of establishing a unified semantic mapping relationship for the initially aligned data to obtain standardized data specifically involves: Based on the unique identification information of the smart sensors and the type of sensor monitoring data, a mapping correspondence between the data collected by each smart sensor and standard numerical features is constructed. Probability density analysis is performed on the numerical distribution characteristics of the preliminary aligned data corresponding to each smart sensor to determine the distribution probability value of each sensor data in a preset information interval; Based on the probability distribution value, the standardized weight coefficients of each smart sensor data are obtained through dynamic weighting calculation. Adaptive nonlinear scaling transformation is performed on the preliminary aligned data corresponding to each smart sensor using standardized weighting coefficients. The data after nonlinear scaling is mapped to a standard numerical feature representation space to obtain standardized data under a unified semantics.

[0009] Optionally, based on standardized data, a continuous-time sensor state representation model is constructed to obtain a continuous hidden state vector, specifically as follows: Using standardized data as model input, the standardized data is divided into multiple adjacent data time windows according to a uniformly set continuous time series division method, and the start and end points of continuous time corresponding to each data time window are set. For the standardized data within each data time window, a set of ordinary differential equations describing the continuous change of the continuous hidden state vector over time is established using neural controlled differential equations. Based on the numerical characteristics and physical meaning of the standardized data, the initial state conditions corresponding to the neural controlled differential equations are determined, and the initial state conditions are input into the ordinary differential equation system. The initial solution of the continuous hidden state vector corresponding to each data time window is obtained through continuous integration. The initial solution of the continuous hidden state vector is substituted into the neural controlled differential equation of the corresponding data time window to calculate the predicted data sequence, and the difference between the predicted data sequence and the actual input standardized data is defined as the residual. The target error is determined based on the residual, and an acceptable threshold for the target error is set. The gradient descent method is used to iteratively optimize the neural controlled differential equation, continuously reducing the target error until the acceptable threshold is reached. Based on the optimized neural-controlled differential equation, a sensor state representation model is constructed; The system receives and inputs standardized data at the current moment in real time, performs continuous integration operations through the sensor state representation model, and obtains a continuous hidden state vector.

[0010] Optionally, the sensor state representation model includes a state space definition module, a state mapping construction module, a state evolution modeling module, a state feature fusion module, and a real-time model update module. The state space definition module determines the dimension of the continuous hidden state vector and the numerical representation range of the state representation space based on the optimized neural controlled differential equation. The state mapping construction module establishes a bidirectional nonlinear mapping relationship between the standardized data space and the continuous hidden state space by training a state encoding neural network and a state decoding neural network. The state encoding neural network maps the standardized data to the continuous hidden state space in real time, and the state decoding neural network maps the continuous hidden state vector back to the standardized data space in real time. The state evolution modeling module uses optimized neural controlled differential equations to dynamically model the evolution process of continuous hidden state vectors in continuous hidden state space; it achieves real-time updating of continuous hidden state vectors through continuous integration operations, thereby obtaining the hidden state evolution trajectory that characterizes the dynamic state of the sensor. The state feature fusion module, based on the bidirectional nonlinear mapping relationship obtained by the state mapping construction module, fuses the real-time input features of standardized data with the latent state evolution trajectory features to generate a fusion feature representation that characterizes the real-time state of the sensor. The real-time model update module updates the parameters of the sensor state representation model in real time based on the fused feature representation and outputs the continuously updated hidden state vector.

[0011] Optionally, the step of dynamically updating the continuous hidden state vector and calculating the fused feature representation using standardized data from newly acquired various types of smart sensors specifically involves: It receives standardized data newly collected by multiple smart sensors in real time and completes unified semantic mapping, and classifies the data according to the unique identification information of the sensors. The continuous hidden state vector obtained in the previous time step is used as the initial state condition for the continuous integration operation in the current time step and input into the optimized neural controlled differential equation. Using the standardized data received in real time at the current moment as the input driving signal, the continuous integration operation of the neural controlled differential equation is performed to calculate the updated continuous hidden state vector at the current moment. Based on the updated continuous hidden state vector, the state decoding neural network is used to map it back to the normalized data space to obtain the real-time prediction data sequence; The difference between the real-time predicted data sequence and the standardized data sequence of the actual input at the current time is calculated to obtain the real-time predicted residual that characterizes the current cooperative state. The fused feature representation is calculated based on the numerical characteristics of the real-time predicted residuals and the continuously hidden state vector updated at the current time.

[0012] Optionally, based on the fused feature representation, the trained anomaly detection discriminator is used to perform real-time discrimination of the current collaborative state to obtain the collaborative supervision anomaly discrimination result, specifically as follows: Using the obtained historical fusion feature representation, an anomaly detection discriminator is trained, and the loss function for anomaly discrimination and the training convergence condition are set during the training process; The system receives the fusion feature representation calculated at the current moment in real time and inputs it into the trained anomaly detection discriminator. The deep feature distribution of the fused feature representation data is extracted by an anomaly detection discriminator and compared with the standard distribution pattern of the historical fused feature representation in the normal collaborative state. Calculate the distribution difference between the fused feature representation data at the current moment and the standard distribution pattern of the normal cooperative state to obtain the abnormal deviation degree of the current cooperative state; The judgment threshold is determined based on the degree of abnormal deviation. When the degree of abnormal deviation exceeds the judgment threshold, the current collaborative status is judged as an abnormal collaborative supervision. If the abnormal deviation does not exceed the discrimination threshold, the current collaborative status is judged as normal collaborative supervision, and the corresponding collaborative supervision abnormal discrimination result is output.

[0013] Optionally, based on the collaborative supervision anomaly judgment result, when an abnormal collaborative supervision state is judged, a pre-set collaborative event handling strategy is activated to generate a regulatory linkage handling instruction, specifically as follows: Based on the results of collaborative supervision anomaly identification, confirm the anomaly type and anomaly level corresponding to the current collaborative supervision anomaly status; Based on the anomaly type and anomaly level, match the corresponding handling plan in the pre-stored collaborative event handling strategy library; Invoke the emergency response plan corresponding to the exception type and level, and obtain the specific response measures and parameters contained in the emergency response plan; Based on the unique identification information of each smart sensor involved when an abnormal state of collaborative supervision occurs, the relevant execution terminals participating in the joint response are identified. Based on the disposal measures, disposal parameters, and information on the execution terminals involved in the joint disposal, a regulatory joint disposal instruction is generated, which includes disposal measure instructions, execution parameters, and execution terminal address information. The generated regulatory linkage response instructions undergo security verification. Once confirmed to be correct, the final regulatory linkage response instructions are generated.

[0014] Optionally, the step of issuing regulatory linkage instructions to relevant execution terminals to conduct closed-loop supervision and linkage of the intelligent sensor network specifically includes: Based on the execution terminal address information contained in the regulatory joint response instructions, a network communication connection is established with each execution terminal through the intelligent sensor network transmission channel; The system sends coordinated response instructions to each execution terminal via network communication connection and monitors the sending status and receiving feedback information of the coordinated response instructions in real time. After each execution terminal receives the joint handling instruction, it receives the instruction confirmation information returned by each execution terminal in real time and confirms the delivery status of the joint handling instruction. It receives real-time feedback from each execution terminal regarding the instruction execution process, and monitors the progress of the response measures and the changes in sensor status after the implementation of the response measures. By receiving real-time execution process information and sensor status changes, it can be determined whether the execution terminal has correctly completed the handling measures specified in the linkage handling instruction; Once all execution terminals have completed their handling measures, a confirmation instruction for the completion of the joint handling is sent to each execution terminal, and information on the entire process of handling measures is recorded as closed-loop monitoring data.

[0015] The IoT-based intelligent sensor collaborative monitoring system includes: The data acquisition unit is used to receive heterogeneous raw monitoring data collected by various types of smart sensors in real time, and perform unified time-series alignment processing based on the unique identification information of the smart sensors to obtain preliminary aligned data under a unified time dimension. The semantic mapping unit is used to construct the mapping correspondence between the data collected by each smart sensor and the standard numerical features based on the unique identification information of the smart sensor and the type of sensor monitoring data. It also performs probability density analysis and adaptive nonlinear scaling transformation on the preliminary aligned data to generate standardized data under a unified semantic. The sensor state representation modeling unit is used to construct a set of ordinary differential equations describing the continuous evolution of the continuous hidden state vector over time based on the optimized neural controlled differential equations, using standardized data as input, and to update and output the continuous hidden state vector in real time. The fusion feature calculation unit is used to receive newly acquired standardized data in real time, perform continuous integration operations to dynamically update the continuous hidden state vector, and calculate the real-time prediction residual to obtain a fusion feature representation that reflects the overall collaborative state of the current smart sensor network. The anomaly detection and discrimination unit is used to receive the fused feature representation, extract the deep feature distribution of the fused features, compare it with the standard distribution pattern of the normal collaborative state in real time, obtain the anomaly deviation degree, determine the anomaly discrimination threshold based on the anomaly deviation degree, and output the collaborative supervision anomaly discrimination result in real time. The collaborative event handling unit is used to receive the collaborative supervision anomaly identification results, confirm the anomaly type and anomaly level corresponding to the collaborative supervision anomaly status, call the corresponding handling plan and generate supervision linkage handling instructions. The closed-loop linkage response unit is used to establish a network communication connection with the execution terminal according to the regulatory linkage response instructions, issue linkage response instructions and monitor the execution of the response measures of the execution terminal in real time, and record the information of the entire process of the response measures as closed-loop regulatory process data.

[0016] The beneficial effects of this invention are: (1) This invention realizes the real-time continuous state update of heterogeneous intelligent sensor data by constructing a sensor state representation model based on neural controlled differential equations and continuous hidden state vectors. This effectively improves the real-time performance and accuracy of sensor data fusion and collaborative supervision and judgment, and enhances the reliability and accuracy of cross-platform data collaborative supervision.

[0017] (2) This invention achieves unified semantic mapping of heterogeneous data through standardized data processing of unified data time alignment and nonlinear scale transformation, which significantly improves the generalization ability and stability of data fusion, and shows better adaptability and robustness in real-time collaborative supervision scenarios of intelligent sensor networks.

[0018] (3) In terms of real-time collaborative supervision of intelligent sensors, this invention effectively solves the problems of insufficient accuracy, poor real-time performance and lagging supervision of heterogeneous data fusion processing in the prior art by using dynamic fusion feature representation and real-time judgment and linkage handling mechanism for abnormal states. It breaks through the bottleneck of generalization ability of traditional data fusion methods, realizes efficient real-time fusion and accurate collaborative supervision of heterogeneous data, and effectively improves the application effect of intelligent sensor networks in industrial production and environmental monitoring. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of the IoT-based intelligent sensor collaborative monitoring method and system proposed in this invention. Detailed Implementation

[0020] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0021] refer to Figure 1 IoT-based intelligent sensor collaborative supervision methods include: Acquire heterogeneous raw monitoring data collected by various types of smart sensors, and perform unified time-series alignment processing on the heterogeneous raw monitoring data based on the unique identifier of each smart sensor to obtain preliminary aligned data under a unified time dimension. For the initial aligned data, a unified semantic mapping relationship is established to convert the heterogeneous raw monitoring data collected by various types of smart sensors into a unified standard numerical feature representation, thereby obtaining standardized data under a unified semantic framework. Based on standardized data under a unified semantic framework, a continuous-time sensor state representation model is constructed to obtain a continuous hidden state vector that can be updated in real time. By receiving and integrating standardized data from various types of smart sensors in real time, the continuous hidden state vector is dynamically updated, and a fusion feature representation reflecting the current cooperative state is calculated. Based on the fusion feature representation, the trained anomaly detection discriminator is used to make real-time judgments on the current collaborative state, and the collaborative supervision anomaly judgment results are obtained. Based on the results of collaborative supervision anomaly identification, when an abnormal collaborative supervision status is identified, a pre-set collaborative event handling strategy is activated, and a joint supervision handling instruction is generated; Regulatory coordination instructions are issued to relevant execution terminals to conduct closed-loop supervision and coordinated response of the intelligent sensor network.

[0022] In this embodiment, obtaining the preliminary alignment data specifically refers to: The system receives raw monitoring data collected at different times from multiple heterogeneous smart sensors through a smart sensor network transmission channel. The timestamp information and unique identifier information of the received raw monitoring data are parsed to extract the data sequence of the collection time corresponding to each smart sensor; Based on a unified clock reference, the data sequence with the smallest time granularity is selected as the reference sequence, and the timestamp of the reference sequence is used as the reference time node to generate a target alignment sequence with a unified time dimension. Interpolation processing is performed on other smart sensor data sequences whose timestamps are inconsistent with the reference time node to form an interpolated data sequence that is consistent with the target aligned sequence. Based on the unique identification information of the smart sensor, the data points of the interpolated data sequence and the reference sequence are matched one by one to obtain preliminary aligned data under a unified time dimension.

[0023] In this embodiment, the step of establishing a unified semantic mapping relationship for the initially aligned data to obtain standardized data specifically involves: Based on the unique identification information of the smart sensors and the type of sensor monitoring data, a mapping correspondence between the data collected by each smart sensor and standard numerical features is constructed. The model and category of each smart sensor are determined based on its unique identification information, as well as the physical meaning and numerical range of the monitoring data collected by each model of smart sensor. Based on the physical meaning of the monitoring data, the feature dimensions in the unified standard numerical feature representation space are determined, and the physical quantity type and standardized value range corresponding to each feature dimension are determined one by one. The physical meaning of the data collected by each smart sensor is associated with the feature dimensions in the standard numerical feature representation space one by one, forming a one-to-one mapping relationship between the physical meaning of each sensor model and the standardized feature dimensions. Probability density analysis is performed on the numerical distribution characteristics of the preliminary aligned data corresponding to each smart sensor to determine the distribution probability value of each sensor data in a preset information interval; For the initial alignment data of each smart sensor, the numerical range of the initial alignment data corresponding to each smart sensor is determined, and the numerical range is divided into several continuous statistical intervals with a fixed interval step size. The frequency of data occurrence in each statistical interval is counted one by one, and the frequency value of each statistical interval is calculated based on the data frequency in each interval and the total data volume. Kernel density estimation is performed using the frequency values ​​of each interval to obtain the probability density function curve of the preliminary aligned data of each sensor within the preset numerical range; The probability distribution value of each sensor data in a preset information interval is determined based on the probability density function curve; Based on the probability distribution value, the standardized weight coefficients of each smart sensor data are obtained through dynamic weighting calculation. The determination of weights in the dynamic weighted calculation is based on the distribution probability value of the preliminary alignment data corresponding to each smart sensor within a preset information interval. The distribution probability value is normalized and then used as the initial weight. The initial weights are dynamically adjusted in real time based on the magnitude of the fluctuation in the sensor monitoring data within the current acquisition period. The greater the fluctuation, the higher the weight value, and vice versa. The dynamic normalized weighting coefficients for adaptive nonlinear scaling are obtained; Adaptive nonlinear scaling transformation is performed on the preliminary aligned data corresponding to each smart sensor using standardized weighting coefficients. The adaptive nonlinear scaling transformation is a process of weighting the initial aligned data of each smart sensor according to the dynamic standardized weighting coefficients to obtain a weighted data sequence. Based on the skewness characteristics of the numerical distribution in the weighted data sequence, by setting a nonlinear transformation function, regions with large data skewness are mapped to relatively wide sub-intervals in the standard numerical feature interval, and regions with small data skewness are mapped to relatively narrow sub-intervals in the standard numerical feature interval. By detecting changes in the initial aligned data numerical distribution in real time, the parameters of the nonlinear transformation function are dynamically adjusted to continuously adapt to changes in data distribution characteristics. The data after nonlinear scaling is mapped to a standard numerical feature representation space to obtain standardized data under a unified semantics.

[0024] In this embodiment, the step of constructing a continuous-time sensor state representation model based on standardized data to obtain a continuous hidden state vector specifically involves: Using standardized data as model input, the standardized data is divided into multiple adjacent data time windows according to a uniformly set continuous time series division method, and the start and end points of continuous time corresponding to each data time window are set. The continuous time series partitioning method determines the initial length of the data time window based on the time dimension length of the standardized data and the sampling frequency of the smart sensor data. Using a fixed-duration sliding window approach, starting from the beginning of the time series, multiple consecutive and adjacent data time windows are divided sequentially by sliding forward with a set fixed step size. The starting point of each data time window coincides with the ending point of the previous data time window to ensure that there is no time interval or overlap between data time windows. During the partitioning process, data fluctuations are detected in real time based on the dynamic trend of data changes within the data time window. When the detected data fluctuation amplitude exceeds the preset threshold, the length of the data time window is dynamically adjusted so that the partitioned data time window can better adapt to the real-time fluctuation characteristics of the data, thereby achieving adaptive continuous time partitioning of the data time window. For the standardized data within each data time window, a set of ordinary differential equations describing the continuous change of the continuous hidden state vector over time is established using neural controlled differential equations. Standardized data is used as the input driving signal for neural controlled differential equations. A neural network model is used to fit and describe the dynamic evolution of the continuous hidden state vector over time. The time derivative of the continuous hidden state vector is expressed as a nonlinear function relationship jointly determined by the current continuous hidden state vector and the standardized data. Based on this function relationship, a set of ordinary differential equations is constructed with the continuous hidden state vector as the variable to be solved. The nonlinear function relationship is parameterized by a pre-constructed neural network model. The input of the neural network model is the current continuous hidden state vector and the corresponding standardized data, and the output is the time derivative of the continuous hidden state vector at the current time. Based on the numerical characteristics and physical meaning of the standardized data, the initial state conditions corresponding to the neural controlled differential equations are determined, and the initial state conditions are input into the ordinary differential equation system. The initial solution of the continuous hidden state vector corresponding to each data time window is obtained through continuous integration. The initial state conditions for determining the neural controlled differential equation are determined by using the data at the start time of the standardized data within each data time window as the initial input. A pre-trained initial state generation neural network is used to map the standardized data at the corresponding time to a continuous hidden state space, obtaining an initial state vector corresponding to the standardized data. The initial state generation neural network is trained through supervised learning. During training, the initial time data of the standardized data is used as the network input, and the true initial solution of the continuous hidden state vector within the corresponding data time window is used as the target output. The neural network parameters are optimized until the error between the network output and the true initial solution meets a preset threshold. After training, the standardized data at the time when the initial state conditions are to be determined is input into the initial state generation neural network, which can generate the initial state conditions that meet the requirements for solving the neural controlled differential equation in real time. The initial solution of the continuous hidden state vector is substituted into the neural controlled differential equation of the corresponding data time window to calculate the predicted data sequence, and the difference between the predicted data sequence and the actual input standardized data is defined as the residual. The calculation and prediction data sequence uses the initial solution of the continuous hidden state vector as the starting state condition of the neural controlled differential equation, and performs continuous integration operation one by one with each time node in the corresponding data time window as the integration interval to obtain the continuous hidden state vector corresponding to each time node. The obtained continuous hidden state vectors are input into the pre-trained state decoding neural network. After nonlinear mapping, the continuous hidden state vectors are mapped from the hidden state space back to the standardized data space, thereby obtaining the predicted data values ​​of each time node within the corresponding data time window. All the predicted data values ​​are arranged in order according to the time nodes to form a predicted data sequence corresponding to the actual standardized data sequence. The target error is determined based on the residual, and an acceptable threshold for the target error is set. The gradient descent method is used to iteratively optimize the neural controlled differential equation, continuously reducing the target error until the acceptable threshold is reached. The acceptable threshold is set as follows: Based on the numerical range and distribution characteristics of the actual input standardized data, the numerical variance of the standardized data sequence within the corresponding data time window is calculated. Based on the numerical variance, determine the statistical distribution characteristics of the residuals between the predicted data sequence and the actual input standardized data, calculate the historical statistical average of the prediction error within multiple continuous data time windows, and calculate the confidence interval of the historical statistical average. The upper bound of the confidence interval of the historical statistical average is used as the acceptable threshold. Based on the optimized neural-controlled differential equation, a sensor state representation model is constructed; The system receives and inputs standardized data at the current moment in real time, performs continuous integration operations through the sensor state representation model, and obtains a continuous hidden state vector.

[0025] In this embodiment, the sensor state representation model includes a state space definition module, a state mapping construction module, a state evolution modeling module, a state feature fusion module, and a real-time model update module. The state space definition module determines the dimension of the continuous hidden state vector and the numerical representation range of the state representation space based on the optimized neural controlled differential equation. The state mapping construction module establishes a bidirectional nonlinear mapping relationship between the standardized data space and the continuous hidden state space by training a state encoding neural network and a state decoding neural network. The state encoding neural network maps the standardized data to the continuous hidden state space in real time, and the state decoding neural network maps the continuous hidden state vector back to the standardized data space in real time. The state evolution modeling module uses optimized neural controlled differential equations to dynamically model the evolution process of continuous hidden state vectors in continuous hidden state space; it achieves real-time updating of continuous hidden state vectors through continuous integration operations, thereby obtaining the hidden state evolution trajectory that characterizes the dynamic state of the sensor. The state feature fusion module, based on the bidirectional nonlinear mapping relationship obtained by the state mapping construction module, fuses the real-time input features of standardized data with the latent state evolution trajectory features to generate a fusion feature representation that characterizes the real-time state of the sensor. The real-time model update module updates the parameters of the sensor state representation model in real time based on the fused feature representation, ensuring the model's real-time adaptive capability to changes in standardized data input, and outputs the continuously updated hidden state vector.

[0026] In this embodiment, the step of dynamically updating the continuous hidden state vector and calculating the fused feature representation using standardized data from newly acquired various types of intelligent sensors specifically involves: It receives standardized data newly collected by multiple smart sensors in real time and completes unified semantic mapping, and classifies the data according to the unique identification information of the sensors. The continuous hidden state vector obtained in the previous time step is used as the initial state condition for the continuous integration operation in the current time step and input into the optimized neural controlled differential equation. Using the standardized data received in real time at the current moment as the input driving signal, the continuous integration operation of the neural controlled differential equation is performed to calculate the updated continuous hidden state vector at the current moment. Based on the updated continuous hidden state vector, the state decoding neural network is used to map it back to the normalized data space to obtain the real-time prediction data sequence; The difference between the real-time predicted data sequence and the standardized data sequence of the actual input at the current time is calculated to obtain the real-time predicted residual that characterizes the current cooperative state. The fusion feature representation is calculated based on the numerical characteristics of the real-time predicted residuals and the continuously hidden state vector updated at the current time. The specific calculation process for the computational fusion feature representation is as follows: Based on the continuously hidden state vector updated at the current moment, extract the state feature values ​​corresponding to each dimension, and calculate the real-time prediction error of each smart sensor based on the real-time prediction residual. The dynamic contribution weight of each smart sensor in the overall collaborative state is determined based on the magnitude of the real-time prediction error. The dynamic contribution weights are used to perform weighted fusion calculations on the state feature values ​​of each dimension to obtain the fused feature values. The fused feature values ​​are mapped to a standardized collaborative state feature space to generate a fused feature representation that can characterize the overall collaborative state of the intelligent sensor network at the current moment.

[0027] In this embodiment, the step of using a trained anomaly detection discriminator to perform real-time discrimination of the current collaborative state based on the fused feature representation and obtaining the collaborative supervision anomaly discrimination result is specifically as follows: Using the obtained historical fusion feature representation, an anomaly detection discriminator is trained, and the loss function for anomaly discrimination and the training convergence condition are set during the training process; The system receives the fusion feature representation calculated at the current moment in real time and inputs it into the trained anomaly detection discriminator. The deep feature distribution of the fused feature representation data is extracted by an anomaly detection discriminator and compared with the standard distribution pattern of the historical fused feature representation in the normal collaborative state. Calculate the distribution difference between the fused feature representation data at the current moment and the standard distribution pattern of the normal cooperative state to obtain the abnormal deviation degree of the current cooperative state; The specific calculation method for the distribution difference is as follows: Extract the feature vector corresponding to the fused feature representation data at the current time, and calculate the Euclidean distance between the feature vector and the feature vector corresponding to the standard distribution pattern of the normal cooperative state. Calculate the likelihood probability of the fused feature representation data under the standard distribution pattern of normal collaborative state at the current moment, and then perform weighted fusion of the Euclidean distance and the likelihood probability to obtain a distribution difference value that comprehensively reflects the degree of difference between the two. The judgment threshold is determined based on the degree of abnormal deviation. When the degree of abnormal deviation exceeds the judgment threshold, the current collaborative status is judged as an abnormal collaborative supervision. The method for determining the discrimination threshold is as follows: Under normal collaborative conditions, the numerical distribution characteristics of the abnormal deviation of the historical fusion feature representation data are statistically analyzed, and the statistical mean and standard deviation of the historical abnormal deviation numerical distribution are calculated. A confidence level is set based on the statistical mean and standard deviation of historical anomalies, and an upper limit of the statistical deviation is determined based on the confidence level. This upper limit is then used as the anomaly detection threshold. If the abnormal deviation does not exceed the discrimination threshold, the current collaborative status is judged as normal collaborative supervision, and the corresponding collaborative supervision abnormal discrimination result is output.

[0028] In this embodiment, the step of activating a pre-set collaborative event handling strategy and generating a regulatory linkage handling instruction when an abnormal collaborative regulatory state is detected, based on the collaborative regulatory anomaly identification result, specifically involves: Based on the results of collaborative supervision anomaly identification, confirm the anomaly type and anomaly level corresponding to the current collaborative supervision anomaly status; Based on the abnormal deviation value in the collaborative supervision anomaly identification results, query the pre-set anomaly type classification rules and anomaly level classification thresholds; The specific anomaly type of the current collaborative supervision anomaly status is determined by comparing the anomaly deviation value with the corresponding value range of each anomaly type. Within the defined anomaly type, the anomaly level corresponding to the current collaborative monitoring anomaly status is determined based on the sub-interval into which the anomaly deviation value falls. Based on the anomaly type and anomaly level, match the corresponding handling plan in the pre-stored collaborative event handling strategy library; Invoke the emergency response plan corresponding to the exception type and level, and obtain the specific response measures and parameters contained in the emergency response plan; Based on the unique identification information of each smart sensor involved when an abnormal state of collaborative supervision occurs, the relevant execution terminals participating in the joint response are identified. Based on the disposal measures, disposal parameters, and information on the execution terminals involved in the joint disposal, a regulatory joint disposal instruction is generated, which includes disposal measure instructions, execution parameters, and execution terminal address information. The generated regulatory linkage response instructions undergo security verification. Once confirmed to be correct, the final regulatory linkage response instructions are generated.

[0029] In this embodiment, the step of issuing regulatory linkage instructions to relevant execution terminals to conduct closed-loop monitoring and linkage of the intelligent sensor network specifically includes: Based on the execution terminal address information contained in the regulatory joint response instructions, a network communication connection is established with each execution terminal through the intelligent sensor network transmission channel; The system sends coordinated response instructions to each execution terminal via network communication connection and monitors the sending status and receiving feedback information of the coordinated response instructions in real time. After each execution terminal receives the joint handling instruction, it receives the instruction confirmation information returned by each execution terminal in real time and confirms the delivery status of the joint handling instruction. It receives real-time feedback from each execution terminal regarding the instruction execution process, and monitors the progress of the response measures and the changes in sensor status after the implementation of the response measures. By receiving real-time execution process information and sensor status changes, it can be determined whether the execution terminal has correctly completed the handling measures specified in the linkage handling instruction; Once all execution terminals have completed their handling measures, a confirmation instruction for the completion of the joint handling is sent to each execution terminal, and information on the entire process of handling measures is recorded as closed-loop monitoring data.

[0030] The IoT-based intelligent sensor collaborative monitoring system includes: The data acquisition unit is used to receive heterogeneous raw monitoring data collected by various types of smart sensors in real time, and perform unified time-series alignment processing based on the unique identification information of the smart sensors to obtain preliminary aligned data under a unified time dimension. The semantic mapping unit is used to construct the mapping correspondence between the data collected by each smart sensor and the standard numerical features based on the unique identification information of the smart sensor and the type of sensor monitoring data. It also performs probability density analysis and adaptive nonlinear scaling transformation on the preliminary aligned data to generate standardized data under a unified semantic. The sensor state representation modeling unit is used to construct a set of ordinary differential equations describing the continuous evolution of the continuous hidden state vector over time based on the optimized neural controlled differential equations, using standardized data as input, and to update and output the continuous hidden state vector in real time. The fusion feature calculation unit is used to receive newly acquired standardized data in real time, perform continuous integration operations to dynamically update the continuous hidden state vector, and calculate the real-time prediction residual to obtain a fusion feature representation that reflects the overall collaborative state of the current smart sensor network. The anomaly detection and discrimination unit is used to receive the fused feature representation, extract the deep feature distribution of the fused features, compare it with the standard distribution pattern of the normal collaborative state in real time, obtain the anomaly deviation degree, determine the anomaly discrimination threshold based on the anomaly deviation degree, and output the collaborative supervision anomaly discrimination result in real time. The collaborative event handling unit is used to receive the collaborative supervision anomaly identification results, confirm the anomaly type and anomaly level corresponding to the collaborative supervision anomaly status, call the corresponding handling plan and generate supervision linkage handling instructions. The closed-loop linkage response unit is used to establish a network communication connection with the execution terminal according to the regulatory linkage response instructions, issue linkage response instructions and monitor the execution of the response measures of the execution terminal in real time, and record the information of the entire process of the response measures as closed-loop regulatory process data.

[0031] Example 1: To verify the feasibility of this invention in practice, it was applied to a remote collaborative monitoring task for intelligent equipment in an industrial park. For environmental monitoring, production equipment operation status monitoring, and safety protection monitoring scenarios within the park, various types of intelligent sensors, including temperature and humidity sensors, gas sensors, vibration sensors, current and voltage sensors, and photosensors, were deployed to acquire multi-source heterogeneous monitoring data in real time, enabling real-time monitoring and early warning of the park's equipment operation status and environmental safety status. In this practical scenario, traditional methods often employ simple time-series alignment and unified standard conversion rules, utilizing static rule-based discrimination methods or general machine learning algorithms for collaborative monitoring. However, facing complex scenarios such as a wide variety of park equipment, frequent dynamic changes in monitoring data, and diverse types of heterogeneous sensors, existing technologies have significant shortcomings in terms of the real-time performance, accuracy, and generalization of data standardization and fusion discrimination. Especially when equipment or the environment experiences abnormal states, monitoring delays or misjudgments often occur.

[0032] In practical operation, this embodiment first acquires heterogeneous raw monitoring data from different types of smart sensors within the park in real time through the smart sensor network transmission channel. Each sensor corresponds to unique identification information. The system performs unified time-series alignment processing on the collected heterogeneous raw data, using the sensor sequence with the highest sampling frequency as the benchmark, and interpolating the data from other sensors to obtain preliminary aligned data under a unified time dimension. Based on this, the system constructs a unified semantic mapping relationship according to the physical meaning and numerical range of the data collected by different sensor models. The system determines the distribution probability value of each sensor's data through probability density analysis and uses dynamic weighting coefficients for adaptive nonlinear scaling transformation, ultimately obtaining standardized data under a unified semantic dimension, ensuring that heterogeneous data is accurately mapped to a unified standard space.

[0033] The system constructs a continuous-time sensor state representation model based on neural controlled differential equations. It divides the continuous-time series into multiple adjacent data time windows using a uniformly defined method. Standardized data from these windows serves as model input. Initial solutions for the continuous hidden state vector are obtained through continuous integration, and model parameters are optimized using gradient descent. Finally, a state representation model capable of real-time updating the continuous hidden state vector is established. In actual operation, the system receives standardized data at the current moment, dynamically updates the continuous hidden state vector through integration, calculates the real-time predicted data sequence using a state decoding network, and obtains a real-time fused feature representation by combining the real-time prediction residuals. Furthermore, the system uses an anomaly detection discriminator trained based on historical fused feature representations to determine the current collaborative state in real time, extracts deep feature distributions, calculates the anomaly deviation, determines the anomaly detection threshold using historical data, and finally outputs the collaborative monitoring anomaly detection result in real time.

[0034] Once an abnormal state in collaborative monitoring is identified as abnormal in real time, the system automatically confirms the anomaly type and level, invokes the pre-set handling plan from the collaborative event handling strategy library, generates a regulatory linkage handling instruction containing specific handling measures and parameters, and issues it to each execution terminal after security verification. The execution terminals provide real-time feedback on the status of the instruction execution process and sensor status changes to ensure the closed-loop completion of the regulatory process.

[0035] After verification through actual field applications, the technical solution of this embodiment demonstrates significant advantages in terms of regulatory response speed, anomaly detection accuracy, and real-time monitoring. Specific performance characteristics are shown in Table 1. Table 1 Comparison of Abnormal State Monitoring Performance between the Invention and Traditional Methods

[0036] As can be seen from the data comparison results in Table 1 above, the method of the present invention significantly reduces the discrimination time to less than 3 seconds in various abnormal state monitoring tasks, which is significantly lower than the discrimination delay of traditional methods. At the same time, the accuracy of abnormal state monitoring is steadily improved to over 97%, while the false alarm rate and false negative rate are significantly reduced to below 2%, demonstrating the accurate and efficient identification capability of the method of the present invention in actual monitoring scenarios.

[0037] In summary, this embodiment demonstrates that the present invention, through continuous state modeling of neural controlled differential equations, unified semantic adaptive standardization, and real-time discrimination method of fusion features, achieves high-precision and high-real-time collaborative monitoring of heterogeneous data from various types of intelligent sensors in complex environments. It exhibits significant advantages and applicability in actual industrial park equipment and environmental monitoring application scenarios, effectively solving the problems of low accuracy, slow response, and frequent false alarms and missed alarms existing in traditional technologies, fully demonstrating the advanced nature and practical application value of the technical solution of the present invention.

[0038] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for collaborative supervision of intelligent sensors based on the Internet of Things, characterized by, The application relates to a method for realizing closed-loop supervision and linkage treatment of an intelligent sensor network. The method comprises the following steps: acquiring heterogeneous original monitoring data collected by multiple types of intelligent sensors, and performing time sequence alignment processing to obtain preliminary alignment data; establishing a unified semantic mapping relationship for the preliminary alignment data to obtain standardized data; constructing a continuous-time sensor state representation model according to the standardized data to obtain a continuous hidden state vector; dynamically updating the continuous hidden state vector and calculating a fusion feature representation through the newly acquired standardized data of the multiple types of intelligent sensors; performing real-time discrimination on a current cooperative state by using a trained abnormality detection discriminator according to the fusion feature representation to obtain a cooperative supervision abnormality discrimination result; starting a pre-set cooperative event treatment strategy to generate a supervision linkage treatment instruction when the cooperative supervision abnormality discrimination result is discriminated as an abnormal state; 2.The IoT-based intelligent sensor collaborative supervision method according to claim 1, characterized in that, issuing the supervision linkage treatment instruction to relevant execution terminals to perform closed-loop supervision and linkage treatment on the intelligent sensor network. The preliminary alignment data is obtained in the following manner: receiving original monitoring data collected by multiple heterogeneous intelligent sensors at different time points through an intelligent sensor network transmission channel; analyzing the time stamp information and the unique identification information of the received original monitoring data to extract a data sequence collected at each time point by each intelligent sensor; selecting a data sequence collected at the smallest time granularity as a reference sequence, taking the time stamp of the reference sequence as a reference time node, and generating a target alignment sequence with a unified time dimension; performing interpolation processing on other intelligent sensor data sequences with time stamps inconsistent with the reference time node to form an interpolation data sequence consistent with the target alignment sequence; 3.The IoT-based intelligent sensor collaborative supervision method according to claim 1, characterized in that, merging the interpolation data sequence and the data points of the reference sequence one by one according to the unique identification information of the intelligent sensors to obtain the preliminary alignment data. The unified semantic mapping relationship is established for the preliminary alignment data to obtain the standardized data in the following manner: constructing a mapping corresponding relationship between the data collected by each intelligent sensor and a standard numerical feature according to the unique identification information of the intelligent sensors and the type of the sensor monitoring data; performing probability density analysis on the numerical distribution characteristics of the preliminary alignment data corresponding to each intelligent sensor to determine the distribution probability value of the sensor data in a pre-set signal interval; obtaining the standardization weight coefficient of the data of each intelligent sensor through dynamic weighting calculation according to the distribution probability value; performing adaptive nonlinear scale transformation on the preliminary alignment data corresponding to each intelligent sensor by using the standardization weight coefficient; 4.The IoT-based intelligent sensor collaborative supervision method according to claim 1, characterized in that, mapping the data after the nonlinear scale transformation to a standard numerical feature representation space to obtain standardized data under a unified semantic. The continuous-time sensor state representation model is constructed according to the standardized data to obtain the continuous hidden state vector in the following manner: taking the standardized data as model input, dividing the standardized data into multiple adjacent data time windows according to a unified set continuous time sequence division method, and setting the continuous time starting point and the ending point corresponding to each data time window; for the standardized data in each data time window, a neural controlled differential equation is used to establish a system of ordinary differential equations describing the continuous change rule of the continuous hidden state vector with time. According to the numerical characteristics and physical meanings of the standardized data, initial state conditions corresponding to the neural controlled differential equation are determined, and the initial state conditions are input into the system of ordinary differential equations to obtain initial solutions of the continuous hidden state vector corresponding to each data time window through continuous integral operation; The initial solutions of the continuous hidden state vector are substituted into the neural controlled differential equation corresponding to the data time window to calculate a predicted data sequence, and a difference between the predicted data sequence and the actual input standardized data is defined as a residual error; A target error is determined according to the residual error, an acceptable threshold of the target error is set, and a gradient descent method is used to iteratively optimize the neural controlled differential equation to continuously reduce the target error until the acceptable threshold is reached; A sensor state representation model is constructed according to the optimized neural controlled differential equation; Real-time standardized data at the current time is received and input, and continuous integral operation is performed on the sensor state representation model to obtain a continuous hidden state vector. 5.The IoT-based intelligent sensor collaborative supervision method according to claim 4, characterized in that, The sensor state representation model includes a state space definition module, a state mapping construction module, a state evolution modeling module, a state feature fusion module, and a model real-time updating module: The state space definition module determines the dimension size of the continuous hidden state vector and the numerical representation range of the state representation space according to the optimized neural controlled differential equation; The state mapping construction module establishes a bidirectional nonlinear mapping relationship between the standardized data space and the continuous hidden state space by training a state encoding neural network and a state decoding neural network, the state encoding neural network maps the standardized data to the continuous hidden state space in real time, and the state decoding neural network maps the continuous hidden state vector back to the standardized data space in real time; The state evolution modeling module dynamically models the evolution process of the continuous hidden state vector in the continuous hidden state space by using the optimized neural controlled differential equation; Real-time updating of the continuous hidden state vector is realized through continuous integral operation to obtain a hidden state evolution trajectory representing the dynamic state of the sensor; The state feature fusion module fuses real-time input features of the standardized data and hidden state evolution trajectory features based on the bidirectional nonlinear mapping relationship obtained by the state mapping construction module to generate a fusion feature representation representing the real-time state of the sensor; The model real-time updating module updates the parameters of the sensor state representation model based on the fusion feature representation in real time and outputs the real-time updated continuous hidden state vector. 6.The IoT-based intelligent sensor collaborative supervision method according to claim 1, characterized in that, The continuous hidden state vector is dynamically updated and the fusion feature representation is calculated by using the newly obtained standardized data of the multiple types of intelligent sensors, specifically: Real-time receive the standardized data newly collected and completed uniform semantic mapping by the multiple intelligent sensors at the current time, and classify them according to the unique sensor identification information; The continuous hidden state vector obtained at the last time is used as the initial state condition for continuous integral operation at the current time, and is input into the optimized neural controlled differential equation; The continuous integral operation of the neural controlled differential equation is performed with the real-time received standardized data at the current time as the input driving signal to calculate the updated continuous hidden state vector at the current time. According to the updated continuous hidden state vector, a real-time prediction data sequence is obtained by mapping back to the standardized data space through a state decoding neural network; The difference between the real-time prediction data sequence and the actual input standardized data sequence at the current time is calculated to obtain a real-time prediction residual representing the current collaborative state; According to the numerical characteristics of the real-time prediction residual and the updated continuous hidden state vector at the current time, a fusion feature representation is calculated. 7.The IoT-based intelligent sensor collaborative supervision method according to claim 1, characterized in that, According to the fusion feature representation, a trained anomaly detection discriminator is used to perform real-time discrimination on the current collaborative state to obtain a collaborative supervision anomaly discrimination result, specifically: Using the obtained historical fusion feature representation, the anomaly detection discriminator is trained, and an anomaly discrimination loss function and a training convergence condition are set during the training process; The fusion feature representation calculated at the current time is received in real time and input into the trained anomaly detection discriminator; The anomaly detection discriminator extracts the deep feature distribution of the fusion feature representation data and compares it with the standard distribution pattern of the historical fusion feature representation of the normal collaborative state; The distribution difference between the fusion feature representation data at the current time and the standard distribution pattern of the normal collaborative state is calculated to obtain an anomaly deviation degree of the current collaborative state; According to the anomaly deviation degree, a discrimination threshold is determined, and when the anomaly deviation degree exceeds the discrimination threshold, the current collaborative state is determined to be a collaborative supervision anomaly; When the anomaly deviation degree does not exceed the discrimination threshold, the current collaborative state is determined to be a collaborative supervision normal, and the corresponding collaborative supervision anomaly discrimination result is output. 8.The IoT-based intelligent sensor collaborative supervision method according to claim 1, wherein, According to the collaborative supervision anomaly discrimination result, when a collaborative supervision anomaly state is discriminated, a pre-set collaborative event handling strategy is started to generate a supervision linkage handling instruction, specifically: According to the collaborative supervision anomaly discrimination result, the anomaly type and anomaly level corresponding to the current collaborative supervision anomaly state are determined; According to the anomaly type and anomaly level, the corresponding handling plan in the pre-stored collaborative event handling strategy library is matched; The handling plan corresponding to the anomaly type and anomaly level is called to obtain the specific handling measures and handling parameters contained in the handling plan; According to the unique identification information of each intelligent sensor involved when the collaborative supervision anomaly state occurs, the related execution terminals participating in the linkage handling are determined; Based on the handling measures, handling parameters, and execution terminal information participating in the linkage handling, a supervision linkage handling instruction including handling measure instructions, execution parameters, and execution terminal address information is generated; The generated supervision linkage handling instruction is subjected to security verification, and the final supervision linkage handling instruction is generated after verification. 9.The IoT-based intelligent sensor collaborative supervision method according to claim 1, wherein, The supervision linkage handling instruction is sent to the related execution terminals to perform closed-loop supervision and linkage handling on the intelligent sensor network, specifically: According to the execution terminal address information contained in the supervision linkage handling instruction, network communication connections between each execution terminal are established through the intelligent sensor network transmission channel; The linkage handling instruction is sent to each execution terminal through the network communication connection, and the sending state of the linkage handling instruction and the receiving feedback information are monitored in real time; After the linkage treatment instruction is received by each execution terminal, the instruction confirmation information returned by each execution terminal is received in real time, and the delivery state of the linkage treatment instruction is confirmed; The instruction execution process information fed back by each execution terminal is received in real time, and the progress of the treatment measure execution process and the sensor state change after the treatment measure execution are monitored; Through the real-time received execution process information and sensor state change, it is determined whether the execution terminal has correctly completed the treatment measure specified in the linkage treatment instruction; When all execution terminals complete the treatment measures, linkage treatment end confirmation instructions are sent to each execution terminal, and the information of the whole process of the treatment measure execution is recorded as closed-loop supervision process data.

10. The system for the collaborative supervision of smart sensors based on the Internet of Things, which performs the method for the collaborative supervision of smart sensors based on the Internet of Things according to any one of claims 1 to 9, characterized in that, Comprise: A data acquisition unit for receiving real-time heterogeneous raw monitoring data collected by multiple types of intelligent sensors, performing unified time alignment processing according to the unique identification information of the intelligent sensors, and obtaining preliminary aligned data under a unified time dimension; A semantic mapping unit for constructing a mapping correspondence between the data collected by each intelligent sensor and the standard numerical characteristic according to the unique identification information of the intelligent sensor and the type of sensor monitoring data, and performing probability density analysis and adaptive nonlinear scale transformation on the preliminary aligned data to generate standardized data under a unified semantic; A sensor state representation modeling unit for taking the standardized data as input, constructing a system of ordinary differential equations describing the continuous evolution law of the continuous hidden state vector with time based on the optimized neural controlled differential equation, and updating and outputting the continuous hidden state vector in real time; A fusion feature calculation unit for receiving newly obtained standardized data in real time, performing continuous integral operation to dynamically update the continuous hidden state vector, calculating real-time prediction residuals, and obtaining fusion feature representation reflecting the overall collaborative state of the intelligent sensor network; An abnormality detection and discrimination unit for receiving the fusion feature representation, extracting the deep feature distribution of the fusion feature, comparing it with the normal collaborative state standard distribution mode in real time, obtaining the abnormal deviation degree, determining the abnormal discrimination threshold according to the abnormal deviation degree, and outputting the collaborative supervision abnormality discrimination result in real time; A collaborative event treatment unit for receiving the collaborative supervision abnormality discrimination result, confirming the abnormal type and abnormal level corresponding to the collaborative supervision abnormal state, calling the corresponding treatment plan, and generating a supervision linkage treatment instruction; A closed-loop linkage treatment unit for establishing network communication connection with the execution terminal according to the supervision linkage treatment instruction, issuing the linkage treatment instruction and monitoring the treatment measure execution of the execution terminal in real time, and recording the whole process information of the treatment measure execution as closed-loop supervision process data.

Images