Power distribution cabinet for realizing environment monitoring based on sensor network

Abstract

The invention relates to the technical field of power distribution cabinet monitoring, in particular to a power distribution cabinet for realizing environment monitoring based on a sensor network, which realizes omnibearing real-time monitoring of the surrounding environment of the power distribution cabinet by arranging multi-point temperature and humidity sensors on the outer side of a power distribution cabinet body and utilizing a cloud monitoring module to uniformly manage data acquisition and storage. A dynamic weight distribution algorithm is adopted, a comprehensive weight is generated according to the historical abnormal frequency, the position and the environment change rate of the sensor, and a temperature and humidity characteristic matrix is given. The data relevance and weight change are learned through the graph attention network, the global feature vector of the environment state is generated, and the analysis capability of the complex environment data is greatly improved. Meanwhile, in combination with a Gaussian process regression prediction model, the future environment trend is accurately predicted. The alarm score and the alarm level are generated by comparing the deviation interval of the predicted value and the real-time value and combining the parameter change rate, and the real-time alarm information is provided, so that the manual inspection cost is effectively reduced, and the monitoring accuracy is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power distribution cabinet monitoring, and in particular to a power distribution cabinet for realizing environmental monitoring based on a sensor network. Background Art

[0002] As the scale of modern power distribution systems continues to expand and the complexity increases, the stability and reliability of power distribution equipment has become one of the key factors in ensuring power supply security. As an important device for power distribution, the distribution cabinet is exposed to a changing environment for a long time, such as high temperature, humidity, and sudden natural disasters (such as typhoons and floods). It is very susceptible to environmental factors, which can lead to equipment performance degradation or even failure. This will not only affect the stability of power supply, but also may cause serious safety hazards. The environmental monitoring of traditional distribution cabinets mainly relies on manual inspections and single-point sensor measurements. This method has the following shortcomings: Limited monitoring range: Traditional methods usually only focus on a single monitoring point of the distribution cabinet, and cannot fully understand the environmental status around the distribution cabinet; Lack of real-time performance: The frequency of manual inspections is limited, and it cannot respond quickly to sudden environmental changes; Traditional systems lack analysis and judgment of monitoring data, and the predictive ability error is high. Summary of the invention

[0003] To solve the above problems, the present invention provides a distribution cabinet for realizing environmental monitoring based on a sensor network. By setting up a sensor network and conducting deep learning by considering environmental parameters, environmental parameter orientations, historical abnormal frequencies and current environmental change rates, automatic monitoring of the distribution cabinet is realized and the accuracy of predictive analysis is effectively improved.

[0004] To achieve the above object, the technical solution adopted by the present invention is: A power distribution cabinet for realizing environmental monitoring based on a sensor network, comprising: a power distribution cabinet body, a sensor network, and a cloud monitoring module, wherein the cloud monitoring module comprises a plurality of temperature sensors and humidity sensors arranged at different points outside the power distribution cabinet body, and the sensor network is communicatively connected with the cloud monitoring module; The cloud monitoring module includes a sensor management unit, a parameter selection unit, an environmental vector generation unit and a predictive maintenance unit; The sensor management unit is used to store the type of each sensor in the sensor network, the position relative to the power distribution cabinet body, and the temperature parameter or humidity parameter obtained by the corresponding sensor; The parameter selection unit is used to perform weighted fusion of the relative position of the power distribution cabinet body, the historical abnormal frequency and the current environmental change rate based on a dynamic weight allocation algorithm to obtain a comprehensive weight, and weight the temperature parameter and the humidity parameter based on the comprehensive weight to generate a temperature feature matrix and a humidity feature matrix; The environmental vector generation unit is used to use the temperature feature matrix and the humidity feature matrix as inputs of the graph attention network at the same time, learn the correlation and weight changes between temperature and humidity parameters at specific positions through the node attention mechanism, model the global characteristics of the environmental state around the distribution cabinet, and generate a multi-dimensional feature vector for describing the environmental state; The predictive maintenance unit is used to construct a Gaussian process regression prediction model based on a multidimensional feature vector, calculate the abnormal probability distribution by comparing the deviation interval between the predicted value generated by the Gaussian process regression prediction model and the real-time value, and generate an alarm level in combination with the parameter change rate, and send maintenance information to the maintenance personnel according to the alarm level.

[0005] Furthermore, the weighted fusion of the relative position of the power distribution cabinet body, the historical abnormal frequency and the current environmental change rate based on the dynamic weight allocation algorithm includes the following steps: Based on the orientation of the sensor relative to the power distribution cabinet body stored in the sensor management unit, combined with the key area of ​​the power distribution cabinet and the preset weight area influence factor, the initial position weight of each sensor is calculated and stored in the position weight set; Count the abnormal frequency and type of sensors in historical data, calculate the abnormal impact factor of each sensor based on the weighted time decay function, and generate a historical abnormal frequency weight set; Call the historical temperature and humidity parameters, use the sliding window difference algorithm to calculate the change rate of continuous data, and map the change rate to a rate weight to generate a current environment change rate weight set; Based on the location weight set, historical abnormal frequency weight set and rate weight set, the weights are fused through a dynamic weight allocation algorithm to generate a comprehensive weight for each sensor.

[0006] Furthermore, the step of weighting the temperature parameters and the humidity parameters based on the comprehensive weights to generate the temperature characteristic matrix and the humidity characteristic matrix includes: Extract the orientation information of each sensor relative to the power distribution cabinet body from the sensor management unit, map each sensor position into a coordinate point using a preset two-dimensional plane rectangular coordinate system, and generate an initial coordinate matrix containing the sensor position; Match the real-time collected temperature parameters and humidity parameters with the initial coordinate matrix respectively, and generate an initial temperature matrix and an initial humidity matrix containing humidity data; The comprehensive weights are called and multiplied point by point with the corresponding positions of the initial temperature matrix and the initial humidity matrix according to the position index to generate the temperature feature matrix and the humidity feature matrix.

[0007] Furthermore, the graph attention network is trained by the following steps: Extract the historical temperature feature matrix and humidity feature matrix as node feature input data, combine the relative position and environmental correlation between sensors, build the graph structure of the sensor network, and initialize the node set and edge set; Initialize the node features based on the parameter values ​​of each node in the temperature feature matrix and the humidity feature matrix, and initialize the edge weight matrix according to the preset edge weight calculation function; Through the node attention mechanism of the graph attention network, the attention score is calculated based on the node features and the edge weight matrix, and the weight values ​​of the adjacent nodes in the graph structure are updated; Use attention scores to perform weighted calculations on adjacent node features to generate a multi-dimensional feature vector; Based on the mean square error loss function, the actual environmental parameter change trend in the training data is taken as the target to optimize the model output prediction value, and the model parameters of the graph attention network are updated through the back propagation algorithm.

[0008] Furthermore, the historical temperature feature matrix and humidity feature matrix are obtained by the following steps: The temperature parameters and humidity parameters under different seasons, weather, natural events and man-made events are collected and a temperature characteristic matrix and a humidity characteristic matrix are generated through a parameter selection unit. The natural events include typhoons, floods and return of the south wind, and the man-made events include pouring water on the distribution cabinet body from different directions.

[0009] Furthermore, the construction of a Gaussian process regression prediction model based on a multidimensional feature vector comprises the following steps: extracting a multi-dimensional feature vector from an environment vector generating unit, and performing normalization processing on the multi-dimensional feature vector; Based on Gaussian process regression, the covariance function is constructed. The radial basis kernel function is selected in combination with the bias term. The preprocessed multidimensional feature vector is used as the training data of the Gaussian process regression prediction model. The covariance matrix between any two points in the training data is calculated using the covariance function, and the Gaussian process distribution is constructed in combination with the mean of the sampled historical data. Optimize hyperparameters in Gaussian process regression prediction model using gradient descent method; The optimized Gaussian process regression prediction model is used to predict the input real-time multidimensional feature vector to generate the predicted values ​​of temperature parameters and humidity parameters at future time points and their corresponding abnormal probability distributions.

[0010] Furthermore, the formula of the Gaussian process regression prediction model is as follows: ; in, Forecast point The predicted value of and are the mean functions of the prediction points and training points respectively; is the covariance matrix between the prediction point and the training data; is the covariance matrix between training data; is the prediction distribution variance; is the identity matrix; is the observed value of the training data.

[0011] Furthermore, the generation of the alarm level comprises the following steps: Obtain the predicted value from the Gaussian process regression prediction model, compare it with the real-time monitoring value, and calculate the deviation between the two; According to the calculated deviation value and the preset deviation range, the deviation value is divided into normal, slight deviation, moderate deviation and severe deviation; Using the variance of the prediction distribution provided by the Gaussian process regression model, the cumulative probability within the deviation interval is calculated; The cumulative probability and parameter change rate within the deviation interval are weighted to obtain the alarm score; Match the alarm level according to the alarm score.

[0012] Furthermore, sending maintenance information to the maintenance personnel according to the alarm level includes the following steps: Acquire the generated alarm level from the predictive maintenance unit, where the alarm level includes normal, slightly abnormal, moderately abnormal, and severely abnormal; If the alarm level is normal, no maintenance information will be sent to the maintenance personnel; If the alarm level is a minor abnormality, a moderate abnormality, or a serious abnormality, a maintenance message is sent to the maintenance personnel.

[0013] The beneficial effect of the present invention is that the present invention arranges temperature sensors and humidity sensors at different points outside the distribution cabinet body, and uniformly manages the data collection and storage of these sensors through the cloud monitoring module, thereby overcoming the limitation of the traditional method that only focuses on a single monitoring point, and realizing all-round real-time monitoring of the environmental state around the distribution cabinet. By adopting a dynamic weight allocation algorithm, the historical abnormal frequency, position and environmental change rate of the sensor are used as key parameters for generating comprehensive weights. The weighted temperature feature matrix and humidity feature matrix are input through the graph attention network to fully learn the correlation and weight changes of sensor data, and generate a global feature vector of the environmental state. This method significantly improves the ability to analyze complex environmental data, and combines the Gaussian process regression prediction model to accurately predict future trends. By comparing the deviation interval between the Gaussian process regression prediction value and the real-time value, and combining the parameter change rate to generate an alarm score, the alarm level is further generated and an alarm is issued according to the alarm score, which significantly reduces the cost of manual inspection and the false alarm rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 It is a structural schematic diagram of a power distribution cabinet for realizing environmental monitoring based on a sensor network in the present invention.

[0015] Figure 2 It is a flow chart of the training steps of the graph attention network in the present invention. DETAILED DESCRIPTION

[0016] See also Figure 1-2 As shown, the present invention relates to a power distribution cabinet for realizing environmental monitoring based on a sensor network, comprising: a power distribution cabinet body, a sensor network, and a cloud monitoring module, wherein the cloud monitoring module comprises a plurality of temperature sensors and humidity sensors arranged at different points outside the power distribution cabinet body, and the sensor network is communicatively connected with the cloud monitoring module; The cloud monitoring module includes a sensor management unit, a parameter selection unit, an environmental vector generation unit and a predictive maintenance unit; The sensor management unit is used to store the type of each sensor in the sensor network, the position relative to the power distribution cabinet body, and the temperature parameter or humidity parameter obtained by the corresponding sensor; The parameter selection unit is used to perform weighted fusion of the relative position of the power distribution cabinet body, the historical abnormal frequency and the current environmental change rate based on a dynamic weight allocation algorithm to obtain a comprehensive weight, and weight the temperature parameter and the humidity parameter based on the comprehensive weight to generate a temperature feature matrix and a humidity feature matrix; The environmental vector generation unit is used to use the temperature feature matrix and the humidity feature matrix as inputs of the graph attention network at the same time, learn the correlation and weight changes between temperature and humidity parameters at specific positions through the node attention mechanism, model the global characteristics of the environmental state around the distribution cabinet, and generate a multi-dimensional feature vector for describing the environmental state; The predictive maintenance unit is used to construct a Gaussian process regression prediction model based on a multidimensional feature vector, calculate the abnormal probability distribution by comparing the deviation interval between the predicted value generated by the Gaussian process regression prediction model and the real-time value, and generate an alarm level in combination with the parameter change rate, and send maintenance information to the maintenance personnel according to the alarm level.

[0017] Specifically, the power distribution cabinet system consists of a power distribution cabinet body, a sensor network and a cloud monitoring module. The power distribution cabinet body is made of high-strength galvanized steel plate and coated with an anti-corrosion coating on the surface to meet the long-term use requirements in harsh environments. The cabinet meets the IP54 protection level, has dustproof and waterproof properties, and can protect the internal equipment from external interference under changing environmental conditions. Multiple temperature sensors and humidity sensors are arranged outside the cabinet. These sensors are distributed according to accurately calculated monitoring points to ensure coverage of key areas, such as high-heat areas (top of the power distribution cabinet) and areas susceptible to moisture (bottom and sides of the power distribution cabinet). The sensor network adopts a hybrid communication architecture, mainly consisting of a wired RS485 bus and a wireless LoRa module. Among them, the wired network is responsible for real-time data transmission of sensors with high data throughput, which is suitable for areas with dense sensors, while the wireless module is used in areas far away from the cabinet or with limited wiring to achieve low-power long-distance data transmission. Each sensor node is equipped with a high-performance microcontroller (MCU), such as the STM32 series chip, which is used to collect environmental data in real time and perform preliminary filtering to remove high-frequency noise in the collected signal. As the core hardware unit of the whole system, the cloud monitoring module includes a sensor management unit, a parameter selection unit, an environmental vector generation unit, and a predictive maintenance unit. The sensor management unit is implemented by an embedded server, which uses a high-performance ARM architecture processor to run the Linux operating system and is equipped with a large-capacity solid-state storage device to store sensor types, installation locations, and real-time collected temperature and humidity parameter data. The unit communicates with the sensor network through the Modbus TCP protocol to achieve low-latency data transmission and storage. The parameter selection unit is implemented by FPGA hardware, which uses its parallel computing power to perform weighted calculations on a large amount of sensor data. Based on the dynamic weight allocation algorithm, the unit comprehensively considers the sensor location, historical abnormal frequency, and environmental change rate, generates a comprehensive weight, and transmits the result to the subsequent module at high speed. The hierarchical pipeline structure design is adopted inside the FPGA, which significantly improves the real-time performance of the weight calculation. The environmental vector generation unit is composed of high-performance GPU (such as NVIDIA Jetson series) hardware, which is mainly used for deep learning modeling of the weighted temperature and humidity feature matrix. Through the parallel computing power optimized by the CUDA architecture, the unit runs the graph attention network model, learns the correlation and weight distribution between sensor data, and generates a multi-dimensional feature vector for describing the global state of the environment. The predictive maintenance unit is based on industrial-grade server hardware and deploys a Gaussian process regression model. The server uses a multi-core x86 architecture processor. By calling the feature vector output by the environment vector generation unit, the predictive maintenance unit makes a high-precision prediction of the future environmental state. Its main tasks include real-time calculation of deviation values, evaluation of abnormal probability, and generation of alarm levels in combination with parameter change rates. Alarm information is sent to the maintenance terminal via the Ethernet module, including the alarm level, possible abnormal cause, and recommended maintenance measures.

[0018] Furthermore, the weighted fusion of the relative position of the power distribution cabinet body, the historical abnormal frequency and the current environmental change rate based on the dynamic weight allocation algorithm includes the following steps: Based on the orientation of the sensor relative to the power distribution cabinet body stored in the sensor management unit, combined with the key area of ​​the power distribution cabinet and the preset weight area influence factor, the initial position weight of each sensor is calculated and stored in the position weight set; Count the abnormal frequency and type of sensors in historical data, calculate the abnormal impact factor of each sensor based on the weighted time decay function, and generate a historical abnormal frequency weight set; Call the historical temperature and humidity parameters, use the sliding window difference algorithm to calculate the change rate of continuous data, and map the change rate to a rate weight to generate a current environment change rate weight set; Based on the location weight set, historical abnormal frequency weight set and rate weight set, the weights are fused through a dynamic weight allocation algorithm to generate a comprehensive weight for each sensor.

[0019] In some embodiments, taking the temperature sensor arranged on the top of the power distribution cabinet as an example, the temperature parameters collected by this position are of great reference value for environmental monitoring because it is close to the heat source area (such as switches and cables). First, according to the physical position of the sensor, its initial position weight is calculated. Assuming that the importance of the top area is higher than that of other areas, the physical position of the sensor is weighted by the regional influence factor so that its weight accounts for a larger proportion in the position set. This process is realized by a preset regional weight matrix, and the matrix value is determined in combination with historical experience data, and can be further adjusted according to the actual use environment. Subsequently, the historical data of the temperature sensor is analyzed to count the frequency and type of its abnormal events. For example, if the sensor detects abnormal values ​​that exceed the normal temperature range many times, based on the weighted time decay model, a higher impact weight is given to recent abnormalities to generate historical abnormal frequency weights. In this way, the system can dynamically identify the severity of abnormal events and their importance to overall environmental monitoring. In addition, the system also calls the temperature change data of the sensor in real time and calculates the rate of change of the current environment using a sliding window difference algorithm. Within a set time interval, the difference of temperature change is calculated point by point and the average rate of change and the maximum rate of change are counted. Taking a typhoon weather as an example, the system detected that the temperature parameters fluctuated greatly in a short period of time. By mapping the relationship between rate and weight, the change rate weight was generated. This weight can reflect the dynamic characteristics of the current environment and provide early warning support for sudden environmental changes. After completing the calculation of the initial position weight, historical abnormal frequency weight and change rate weight, the system merges the three through a dynamic weight allocation algorithm to generate a comprehensive weight. The fusion algorithm adjusts the weight allocation ratio according to the specific needs of the sensor environment. For example, in a stable environment, the position weight and historical abnormality weight account for the main proportion, while in an emergency environment, the weight of the change rate weight will be dynamically increased to ensure that the comprehensive weight can reflect the current monitoring needs in a timely manner. The generated comprehensive weight is used to weight the temperature and humidity parameters to form a feature matrix. With the feature matrix as input, the system further uses a deep learning model (such as the attention network in the figure) to learn and model the correlation and weight distribution between data. Combined with the Gaussian process regression prediction model, the system can achieve high-precision prediction of future environmental parameters based on these features, thereby providing strong support for anomaly detection and maintenance planning.

[0020] Furthermore, the step of weighting the temperature parameters and the humidity parameters based on the comprehensive weights to generate the temperature characteristic matrix and the humidity characteristic matrix includes: Extract the orientation information of each sensor relative to the power distribution cabinet body from the sensor management unit, map each sensor position into a coordinate point using a preset two-dimensional plane rectangular coordinate system, and generate an initial coordinate matrix containing the sensor position; Match the real-time collected temperature parameters and humidity parameters with the initial coordinate matrix respectively, and generate an initial temperature matrix and an initial humidity matrix containing humidity data; The comprehensive weights are called and multiplied point by point with the corresponding positions of the initial temperature matrix and the initial humidity matrix according to the position index to generate the temperature feature matrix and the humidity feature matrix.

[0021] In some embodiments, first, the stored sensor position data is extracted from the sensor management unit. The relative position of each sensor is based on the spatial layout of the distribution cabinet body and is mapped through a preset two-dimensional rectangular coordinate system. Assuming that the distribution cabinet body takes the lower left corner of the bottom as the origin and the upper right corner of the top as the diagonal point of the coordinate system, the relative position of each sensor (such as the top, bottom, and side) is converted into a specific coordinate point. For example, the temperature sensor position in the center of the top is mapped to the coordinate point (0.5, 1.0), and the humidity sensor position on the right side of the bottom is mapped to the coordinate point (1.0, 0.0). The position data of all sensors are stored in a matrix to form an initial coordinate matrix. Secondly, the system collects temperature parameters and humidity parameters from the sensor network in real time to generate two independent data sets. These parameters are matched point by point with the initial coordinate matrix according to the unique identifier of the sensor, and the environmental data is associated with the physical space position. The matched temperature data and humidity data are stored as the initial temperature matrix and the initial humidity matrix, respectively, where the rows correspond to the spatial positions of the sensors and the columns record the corresponding environmental parameter values. For example, the data in the i-th row and j-th column of the temperature matrix collected at a certain moment represents the temperature reading of the sensor located at the coordinate point (0.5, 1.0). Next, the system calls the comprehensive weight to assign the calculated weight to each sensor to characterize the importance of the sensor. The comprehensive weight combines the regional influence factor of the sensor location, the historical abnormal frequency, and the environmental change rate. These weights have been dynamically calculated in the previous step. The comprehensive weight is stored in the form of a one-dimensional vector, where each element corresponds to a sensor. The system performs a weighting operation on each element of the initial temperature matrix and the initial humidity matrix in turn. The weighting operation is completed in the form of matrix point-by-point multiplication: through the position index, the comprehensive weight is multiplied with the data point at the corresponding position of the initial matrix to generate the weighted temperature feature matrix and humidity feature matrix. For example, if the comprehensive weight of the coordinate point (0.5, 1.0) is 1.5, the corresponding initial temperature value is 30°C, and the final generated feature value is 30×1.5=45. The calculation process of the humidity feature matrix is ​​similar.

[0022] Furthermore, the graph attention network is trained by the following steps: Extract the historical temperature feature matrix and humidity feature matrix as node feature input data, combine the relative position and environmental correlation between sensors, build the graph structure of the sensor network, and initialize the node set and edge set; Initialize the node features based on the parameter values ​​of each node in the temperature feature matrix and the humidity feature matrix, and initialize the edge weight matrix according to the preset edge weight calculation function; Through the node attention mechanism of the graph attention network, the attention score is calculated based on the node features and the edge weight matrix, and the weight values ​​of the adjacent nodes in the graph structure are updated; Use attention scores to perform weighted calculations on adjacent node features to generate a multi-dimensional feature vector; Based on the mean square error loss function, the actual environmental parameter change trend in the training data is taken as the target to optimize the model output prediction value, and the model parameters of the graph attention network are updated through the back propagation algorithm.

[0023] In some embodiments, first, the system extracts the historical temperature feature matrix and the humidity feature matrix as input data. These feature matrices are generated by the historical temperature and humidity parameters collected by the sensor, and have been weighted by the previous step, integrating the information of the sensor's spatial position, historical abnormal frequency and environmental change rate. These feature matrices are mapped to the node feature space, where each node represents a sensor, and the feature vector of each node contains the temperature and humidity parameters of the sensor and its comprehensive weight. Based on the relative position between sensors and their environmental relevance, the system constructs a graph structure of the sensor network, including a node set and an edge set, where the node represents the sensor and the edge represents the association between the sensors. In the initialization stage, the system assigns a feature vector to each node according to the parameter value in the temperature and humidity feature matrix, and generates an edge weight matrix according to the relative position of the sensor and the environmental relevance. The edge weight calculation uses a preset function, such as a Gaussian kernel function, to calculate the weight value based on the Euclidean distance between sensors. The higher the weight value, the stronger the correlation between the two sensors, for example, the weight between sensors located in the same area is higher, while the weight between long-distance sensors is lower. The initialized graph structure can intuitively reflect the relationship between the spatial position and environmental parameters in the sensor network. During the training process, the system calculates the attention score between each node and its neighboring nodes through the node attention mechanism of the graph attention network. The node attention mechanism dynamically adjusts the weight values ​​of neighboring nodes by jointly calculating node features and edge weight matrices. For example, for areas with large environmental fluctuations, the attention scores of related nodes will be increased to enhance sensitivity to environmental changes in key areas. Through the calculation of attention scores, the system updates the weight values ​​of neighboring nodes in the graph structure to ensure that the representation ability of the graph is closer to actual environmental changes. After the weight is updated, the system uses the attention score to perform weighted calculations on the features of neighboring nodes to generate multi-dimensional feature vectors. These feature vectors integrate the environmental information of the node itself and its neighboring nodes, and are a high-level representation of the overall environmental state of the sensor network. For example, for areas where the temperature and humidity parameters between adjacent nodes are strongly correlated, the feature vector will more prominently reflect the environmental state of the area. During the training process, the output of the model is compared with the actual trend of environmental parameter changes, and the mean square error (MSE) is used as the loss function to optimize the prediction results. Through the back propagation algorithm, the system dynamically adjusts the model parameters of the graph attention network to gradually reduce the prediction error. For example, for consecutive model training iterations, the system identifies the difference between the current prediction results and the target value and further optimizes by adjusting the weights of nodes and edges.

[0024] Furthermore, the historical temperature feature matrix and humidity feature matrix are obtained by the following steps: The temperature parameters and humidity parameters under different seasons, weather, natural events and man-made events are collected and a temperature characteristic matrix and a humidity characteristic matrix are generated through a parameter selection unit. The natural events include typhoons, floods and return of the south wind, and the man-made events include pouring water on the distribution cabinet body from different directions.

[0025] In some embodiments, first, in order to build complete historical data, the system deploys multiple temperature sensors and humidity sensors, which are arranged at key points outside the body of the distribution cabinet, including the top, side and bottom, which are easily affected by the environment. The sensors collect data in real time, and their sampling frequency is dynamically adjusted according to the event type. For example, the sampling frequency can be set to once an hour in stable seasons, while the sampling frequency is increased to once a minute under natural events or human intervention conditions to capture rapidly changing environmental conditions. Data collection covers a variety of weather conditions under the change of seasons, including sunny days, rainy days, high humidity environments, etc. At the same time, the impact of special natural events is recorded, such as the impact of strong winds and heavy rains brought by typhoons on the external environment of the distribution cabinet; the long-term high humidity state that may occur under flood conditions; and the condensation water problem caused by the return of the south (high humidity weather in spring). In these cases, the system can capture the sharp increase in external humidity in real time through high-precision humidity sensors, and synchronously record the temperature fluctuations caused by the evaporation of condensed water through temperature sensors. In addition, in order to simulate the impact of emergencies on the distribution cabinet environment, a variety of human intervention experiments were carried out. For example, water is poured on the distribution cabinet body from different directions to simulate possible external liquid leakage or rainwater intrusion; at the same time, the amount and frequency of liquid splashing are controlled to record its impact on sensors at different points. These experimental data effectively supplement the special situations that may not be covered in the natural environment data. All the collected temperature and humidity data are transmitted to the parameter selection unit for processing through the sensor network. In the parameter selection unit, the system first performs preliminary correction and denoising on the collected data to filter out abnormal values ​​caused by sensor hardware drift or sudden interference signals. Subsequently, according to the environmental characteristics under different acquisition conditions, the system uses a dynamic weight allocation algorithm to weight the data. For example, under typhoon conditions, the top area sensor is given a higher weight, while under flood conditions, the data of the bottom area sensor is more emphasized. This dynamic weighting method ensures that the feature matrix can accurately reflect the comprehensive importance of sensors under different conditions. After weighting processing, the system stores the processed temperature and humidity data in a matrix, generating a temperature feature matrix and a humidity feature matrix respectively. Each row of the matrix corresponds to a sensor, and each column corresponds to a data point under a specific condition. For example, the temperature data of the top sensor under typhoon conditions will occupy a higher weight in the temperature feature matrix, and the relevant rows of the humidity feature matrix will highlight the fluctuation range of its humidity values.

[0026] Furthermore, the construction of a Gaussian process regression prediction model based on a multidimensional feature vector comprises the following steps: extracting a multi-dimensional feature vector from an environment vector generating unit, and performing normalization processing on the multi-dimensional feature vector; Based on Gaussian process regression, the covariance function is constructed. The radial basis kernel function is selected in combination with the bias term. The preprocessed multidimensional feature vector is used as the training data of the Gaussian process regression prediction model. The covariance matrix between any two points in the training data is calculated using the covariance function, and the Gaussian process distribution is constructed in combination with the mean of the sampled historical data. Optimize hyperparameters in Gaussian process regression prediction model using gradient descent method; The optimized Gaussian process regression prediction model is used to predict the input real-time multidimensional feature vector to generate the predicted values ​​of temperature parameters and humidity parameters at future time points and their corresponding abnormal probability distributions.

[0027] First, the multidimensional feature vectors generated by the graph attention network are extracted from the environmental vector generation unit. These vectors integrate the temperature and humidity parameters, spatial position relationship and dynamic weight characteristics of the sensor, and can fully reflect the global state of the environment around the distribution cabinet. Since different feature components may have different numerical ranges, the multidimensional feature vectors need to be normalized to eliminate the influence of different dimensions on the model calculation. The normalization adopts the maximum and minimum normalization method to linearly map the value of each component to the preset interval, and the normalized feature vector is input into the Gaussian process regression model. After the feature vector is normalized, the system models the data through the Gaussian process regression model. The core of the model is the design of the covariance function. The system adopts the form of radial basis kernel function (RBF kernel) combined with bias term. The radial basis kernel function measures the similarity between data points by calculating the Euclidean distance between multidimensional feature vectors. The bias term is used to capture the noise in the data and enhance the model's tolerance to uncertainty. The parameters of the covariance function such as signal variance, noise variance and length scale can be flexibly adjusted to adapt to different environmental monitoring scenarios. After constructing the covariance function, the system uses historical temperature and humidity data to calculate the covariance matrix. Each element of the covariance matrix represents the correlation between historical data points. At the same time, combined with the mean of these data, the Gaussian distribution of multidimensional feature vectors is established to describe the probability model of future trends. In the training phase, the system optimizes the model parameters by minimizing the negative logarithmic marginal likelihood function. The optimization process uses the gradient descent algorithm to dynamically adjust the hyperparameters of the covariance function so that the model can more accurately capture the distribution law of the data. After completing the model training, the system uses the multidimensional feature vectors obtained in real time for prediction. The correlation between real-time data and historical data is calculated through the covariance function, and the predicted values ​​of temperature parameters and humidity parameters at future time points are generated by combining the covariance matrix and mean obtained through training. At the same time, Gaussian process regression also outputs the probability distribution of the predicted value. This probability distribution not only provides the predicted value, but also quantifies the uncertainty of the predicted result through the variance of the distribution. Taking a sudden weather event as an example, the system may predict that the humidity value of a sensor in the next hour will be 80%, but at the same time points out that the variance of the prediction is large, reflecting the potential risk that the humidity may fluctuate violently.

[0028] Furthermore, the formula of the Gaussian process regression prediction model is as follows: ; in, Forecast point The predicted value of and are the mean functions of the prediction points and training points respectively; is the covariance matrix between the prediction point and the training data; is the covariance matrix between training data; is the prediction distribution variance; is the identity matrix; is the observed value of the training data.

[0029] Specifically, first, the value of the prediction point is calculated through the following process. The model assumes that the changes in environmental parameters follow a Gaussian distribution, and the value of the prediction point consists of the prior mean and the contribution of historical observations to its correction. The prior mean is a preliminary estimate of the value of the prediction point, which is usually assumed to be zero mean to simplify the calculation. Subsequently, the system corrects the prior mean of the prediction point through the covariance matrix. These covariance matrices reflect the correlation between the prediction point and the historical data. Specifically, the covariance matrix is ​​calculated by the radial basis kernel function, which measures the similarity based on the distance between the prediction point and the training data. The closer the distance, the stronger the correlation. The magnitude of the correction is determined by the deviation between the observed value and the mean of the historical data. The covariance matrix distributes the impact of this deviation to the prediction point according to the weight, and finally obtains the corrected prediction value. This calculation method based on statistical characteristics ensures the reliability of the prediction result while reducing the impact of single-point data errors. In addition, the uncertainty of the prediction result is quantified by the variance of the prediction distribution. The system calculates the propagation effect of the correlation between the prediction point and the historical data in the covariance matrix. The size of the prediction variance reflects the model's confidence in the predicted point value. For example, when there is less data around the prediction point or the data varies greatly, the variance will increase, indicating that the uncertainty of the prediction result is high. In the application, the model generates a covariance matrix through the multi-dimensional feature vectors collected in real time, and combines the observed values ​​of historical data to calculate the predicted values ​​of future environmental parameters. For example, under typhoon conditions, the model may predict the humidity value of the top sensor within the next hour, and attach a prediction variance to remind maintenance personnel that the humidity may fluctuate drastically.

[0030] Furthermore, the generation of the alarm level comprises the following steps: Obtain the predicted value from the Gaussian process regression prediction model, compare it with the real-time monitoring value, and calculate the deviation between the two; According to the calculated deviation value and the preset deviation range, the deviation value is divided into normal, slight deviation, moderate deviation and severe deviation; Using the variance of the prediction distribution provided by the Gaussian process regression model, the cumulative probability within the deviation interval is calculated; The cumulative probability and parameter change rate within the deviation interval are weighted to obtain the alarm score; Match the alarm level according to the alarm score.

[0031] In some embodiments, first, the system extracts the predicted values ​​of environmental parameters, such as the future change trend of temperature and humidity, from the Gaussian process regression prediction model. These predicted values ​​are generated by the model's learning of historical data and the input of real-time multidimensional feature vectors, and are estimates of future environmental states. At the same time, the real-time monitoring values ​​are collected in real time by sensors deployed at different points in the distribution cabinet as a reflection of the actual environmental state. Subsequently, the system compares the predicted value with the real-time monitoring value and calculates the deviation between the two. The system further uses the prediction distribution variance provided by the Gaussian process regression model to calculate the cumulative probability of the predicted value and the actual value within the deviation interval. The prediction distribution variance reflects the uncertainty of the model in the predicted value, and the cumulative probability of the deviation value within a specific interval can be obtained by integrating the probability density function. This probability measures the confidence level of the deviation, that is, the probability of the deviation value in the current interval. Next, the system combines the cumulative probability with the rate of change of the environmental parameters. The rate of change is calculated by the continuous data of the sensor through the sliding window difference algorithm, reflecting the dynamic nature of the change of the environmental state. For example, when the humidity change rate is high, even if the deviation is small, it may mean potential risk. By weighted calculation of the cumulative probability and the rate of change, the system generates an alarm score. The weighting coefficient is dynamically adjusted according to the actual application scenario to adapt to different monitoring needs. Finally, the corresponding alarm level is matched according to the alarm score. The alarm level classification includes: Normal: the score is lower than the preset threshold, and no measures are required; Slightly abnormal: the score is close to the threshold, indicating that attention is needed; Moderately abnormal: the score significantly exceeds the threshold, and inspection is recommended; Severely abnormal: the score significantly exceeds the threshold, triggering an emergency response. For example, in the case of a southerly wind, when the predicted value of the humidity sensor is 95%, the real-time value is 98%, the deviation value is 3%, and the cumulative probability is high, the system may generate an alarm of a slight abnormality level. If the humidity changes rapidly and the deviation value increases further, the alarm level may be upgraded to a moderate abnormality or even a severe abnormality.

[0032] Furthermore, sending maintenance information to the maintenance personnel according to the alarm level includes the following steps: Acquire the generated alarm level from the predictive maintenance unit, where the alarm level includes normal, slightly abnormal, moderately abnormal, and severely abnormal; If the alarm level is normal, no maintenance information will be sent to the maintenance personnel; If the alarm level is a minor abnormality, a moderate abnormality, or a serious abnormality, a maintenance message is sent to the maintenance personnel.

[0033] Specifically, the system first obtains the generated alarm level from the predictive maintenance unit, which is derived from a comprehensive analysis of the results of the Gaussian process regression prediction model, real-time monitoring values, deviation values, and parameter change rates. The alarm levels are divided into normal, slightly abnormal, moderately abnormal, and severely abnormal, corresponding to different environmental conditions and response requirements.

[0034] After receiving the alarm level, the system performs hierarchical processing according to the preset logic. For the alarm level of "normal", the system determines that the current environmental status is within the safe range and no intervention measures are required, so no maintenance information will be sent to the maintenance personnel. This design can avoid unnecessary notifications, reduce information interference and labor costs. If the alarm level is "slightly abnormal", the system will send a maintenance message of a reminder nature to the maintenance personnel. This information mainly includes: the specific location of the sensor, the monitored environmental parameters, the predicted deviation value and the alarm level. For example, "Sensor A (location: top of the distribution cabinet) detects a slight abnormality in humidity. The current humidity is 85% and the predicted deviation value is 3%." This type of information is intended to remind maintenance personnel to pay attention to environmental changes at this point, but no immediate action is required. For the alarm level of "moderately abnormal", the maintenance information generated by the system will be more detailed and include recommended measures. For example, the system may prompt maintenance personnel to check certain equipment components that are susceptible to environmental influences, or provide short-term monitoring suggestions to further track changes in environmental parameters. For example, "Sensor B (location: north side of the distribution cabinet) detected moderate humidity anomalies. The current humidity is 90% and the deviation is 7%. It is recommended to check the sealing status of the north side of the distribution cabinet to avoid further deterioration." When the alarm level reaches "serious anomalies", the system will generate emergency maintenance information and notify maintenance personnel in real time through multiple channels (such as SMS, email, or dedicated maintenance applications). Such information not only includes abnormality details and recommended measures, but may also include a brief analysis of estimated risks. For example, "Sensor C (location: inside the distribution cabinet) detected severe temperature anomalies. The current temperature is 60°C and the deviation is 15%. There is a risk of equipment overheating. Please check the cooling device inside the distribution cabinet and repair potential problems." The above implementation modes are merely descriptions of the preferred implementation modes of the present invention, and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary engineering and technical personnel in the field shall fall within the protection scope determined by the claims of the present invention.

Claims

1. A power distribution cabinet for realizing environmental monitoring based on a sensor network, characterized in that: include: A power distribution cabinet body, a sensor network, and a cloud monitoring module, wherein the cloud monitoring module includes a plurality of temperature sensors and humidity sensors disposed at different points outside the power distribution cabinet body, and the sensor network is communicatively connected with the cloud monitoring module; The cloud monitoring module includes a sensor management unit, a parameter selection unit, an environmental vector generation unit and a predictive maintenance unit; The sensor management unit is used to store the type of each sensor in the sensor network, the position relative to the power distribution cabinet body, and the temperature parameter or humidity parameter obtained by the corresponding sensor; The parameter selection unit is used to perform weighted fusion of the relative position of the power distribution cabinet body, the historical abnormal frequency and the current environmental change rate based on a dynamic weight allocation algorithm to obtain a comprehensive weight, and weight the temperature parameter and the humidity parameter based on the comprehensive weight to generate a temperature feature matrix and a humidity feature matrix; The environmental vector generation unit is used to use the temperature feature matrix and the humidity feature matrix as inputs of the graph attention network at the same time, learn the correlation and weight changes between temperature and humidity parameters at specific positions through the node attention mechanism, model the global characteristics of the environmental state around the distribution cabinet, and generate a multi-dimensional feature vector for describing the environmental state; The predictive maintenance unit is used to construct a Gaussian process regression prediction model based on a multidimensional feature vector, calculate the abnormal probability distribution by comparing the deviation interval between the predicted value generated by the Gaussian process regression prediction model and the real-time value, and generate an alarm level in combination with the parameter change rate, and send maintenance information to the maintenance personnel according to the alarm level.

2. A power distribution cabinet for realizing environmental monitoring based on a sensor network according to claim 1, characterized in that: The weighted fusion of the relative position of the power distribution cabinet body, the historical abnormal frequency and the current environmental change rate based on the dynamic weight allocation algorithm includes the following steps: Based on the orientation of the sensor relative to the power distribution cabinet body stored in the sensor management unit, combined with the key area of ​​the power distribution cabinet and the preset weight area influence factor, the initial position weight of each sensor is calculated and stored in the position weight set; Count the abnormal frequency and type of sensors in historical data, calculate the abnormal impact factor of each sensor based on the weighted time decay function, and generate a historical abnormal frequency weight set; Call the historical temperature and humidity parameters, use the sliding window difference algorithm to calculate the change rate of continuous data, and map the change rate to a rate weight to generate a current environment change rate weight set; Based on the location weight set, historical abnormal frequency weight set and rate weight set, the weights are fused through a dynamic weight allocation algorithm to generate a comprehensive weight for each sensor.

3. A power distribution cabinet for realizing environmental monitoring based on a sensor network according to claim 1, characterized in that: The step of weighting the temperature parameters and the humidity parameters based on the comprehensive weights to generate the temperature characteristic matrix and the humidity characteristic matrix comprises: Extract the orientation information of each sensor relative to the power distribution cabinet body from the sensor management unit, map each sensor position into a coordinate point using a preset two-dimensional plane rectangular coordinate system, and generate an initial coordinate matrix containing the sensor position; Match the real-time collected temperature parameters and humidity parameters with the initial coordinate matrix respectively, and generate an initial temperature matrix and an initial humidity matrix containing humidity data; The comprehensive weights are called and multiplied point by point with the corresponding positions of the initial temperature matrix and the initial humidity matrix according to the position index to generate the temperature feature matrix and the humidity feature matrix.

4. A power distribution cabinet for realizing environmental monitoring based on a sensor network according to claim 1, characterized in that: The graph attention network is trained by the following steps: Extract the historical temperature feature matrix and humidity feature matrix as node feature input data, combine the relative position and environmental correlation between sensors, build the graph structure of the sensor network, and initialize the node set and edge set; Initialize the node features based on the parameter values ​​of each node in the temperature feature matrix and the humidity feature matrix, and initialize the edge weight matrix according to the preset edge weight calculation function; Through the node attention mechanism of the graph attention network, the attention score is calculated based on the node features and the edge weight matrix, and the weight values ​​of the adjacent nodes in the graph structure are updated; Use attention scores to perform weighted calculations on adjacent node features to generate multi-dimensional feature vectors; Based on the mean square error loss function, the actual environmental parameter change trend in the training data is taken as the target to optimize the model output prediction value, and the model parameters of the graph attention network are updated through the back propagation algorithm.

5. A power distribution cabinet for realizing environmental monitoring based on a sensor network according to claim 4, characterized in that: The historical temperature feature matrix and humidity feature matrix are obtained by the following steps: The temperature parameters and humidity parameters under different seasons, weather, natural events and man-made events are collected and a temperature characteristic matrix and a humidity characteristic matrix are generated through a parameter selection unit. The natural events include typhoons, floods and return of the south wind, and the man-made events include pouring water on the distribution cabinet body from different directions.

6. A power distribution cabinet for realizing environmental monitoring based on a sensor network according to claim 1, characterized in that: The Gaussian process regression prediction model based on the multidimensional feature vector is constructed, comprising the following steps: extracting a multi-dimensional feature vector from an environment vector generating unit, and performing normalization processing on the multi-dimensional feature vector; Based on Gaussian process regression, the covariance function is constructed. The radial basis kernel function is selected in combination with the bias term. The preprocessed multidimensional feature vector is used as the training data of the Gaussian process regression prediction model. The covariance matrix between any two points in the training data is calculated using the covariance function, and the Gaussian process distribution is constructed in combination with the mean of the sampled historical data. Optimize hyperparameters in Gaussian process regression prediction model using gradient descent method; The optimized Gaussian process regression prediction model is used to predict the input real-time multidimensional feature vector to generate the predicted values ​​of temperature parameters and humidity parameters at future time points and their corresponding abnormal probability distributions.

7. A power distribution cabinet for realizing environmental monitoring based on a sensor network according to claim 6, characterized in that: The formula of the Gaussian process regression prediction model is as follows: ； in, Forecast point The predicted value of and are the mean functions of the prediction points and training points respectively; is the covariance matrix between the prediction point and the training data; is the covariance matrix between training data; is the prediction distribution variance; is the identity matrix; is the observed value of the training data.

8. A power distribution cabinet for realizing environmental monitoring based on a sensor network according to claim 7, characterized in that: The generation of the alarm level comprises the following steps: Obtain the predicted value from the Gaussian process regression prediction model, compare it with the real-time monitoring value, and calculate the deviation between the two; According to the calculated deviation value and the preset deviation range, the deviation value is divided into normal, slight deviation, moderate deviation and severe deviation; Using the variance of the prediction distribution provided by the Gaussian process regression model, the cumulative probability within the deviation interval is calculated; The cumulative probability and parameter change rate within the deviation interval are weighted to obtain the alarm score; Match the alarm level according to the alarm score.

9. A power distribution cabinet for realizing environmental monitoring based on a sensor network according to claim 1, characterized in that: The sending of maintenance information to the maintenance personnel according to the alarm level comprises the following steps: Acquire the generated alarm level from the predictive maintenance unit, where the alarm level includes normal, slightly abnormal, moderately abnormal, and severely abnormal; If the alarm level is normal, no maintenance information will be sent to the maintenance personnel; If the alarm level is a minor abnormality, a moderate abnormality, or a serious abnormality, a maintenance message is sent to the maintenance personnel.

Images