A rain monitoring system and method based on a heterogeneous sensing path graph space-time network

Abstract

The present application relates to the technical field of intelligent meteorological monitoring, and more particularly to a rain monitoring system and method based on a heterogeneous sensing path graph space-time network. The method comprises the following steps: acquiring multi-source monitoring data; setting a bottom physical node, a top physical node and a plurality of virtual intermediate nodes in a measuring cylinder to construct a heterogeneous sensing path graph; inputting the multi-source monitoring data and the heterogeneous sensing path graph into a pre-trained heterogeneous sensing path graph space-time network model to obtain fused water level estimates and rain estimates through regression decoding; and outputting the health degree diagnosis results of each component of the device through the analysis of the global state of the heterogeneous sensing path graph by the heterogeneous sensing path graph space-time network model. The present application realizes the deep fusion of double sensor data and the health diagnosis of the device by constructing a heterogeneous sensing path graph model containing virtual nodes, thereby solving the problem of inaccurate rain estimation under dynamic working conditions.

Description

Technical Field

[0001] This invention relates to the field of intelligent meteorological monitoring technology, and in particular to a rainfall monitoring system and method based on a heterogeneous sensor path graph spatiotemporal network. Background Technology

[0002] In actual operation, rainfall monitoring equipment is affected by physical factors such as water surface fluctuations caused by changes in rainfall intensity, water density changes due to impurities, and ambient temperature gradients. It also faces equipment-related interference such as momentary sensor signal blockage and gradual drift. These factors combined result in complex nonlinear deviations in the water level values ​​calculated by sensors based on different principles (such as pressure and laser), ultimately affecting the accuracy and reliability of the fused rainfall estimate. However, existing technologies struggle to effectively fuse heterogeneous data with nonlinear errors when dealing with complex conditions such as water surface fluctuations and density changes, leading to large rainfall estimation errors. Furthermore, existing equipment status monitoring suffers from coarse granularity and delayed response, only able to determine if the sensor is offline, lacking the ability to identify "sub-healthy" states of the pathway, such as contamination of the measuring cylinder wall or slight blockage of the guide tube. This results in long delays in fault diagnosis and difficulty in dynamically adjusting data reliability under interference (such as temporary sensor blockage), failing to provide strong support for accurate meteorological and hydrological monitoring. Summary of the Invention

[0003] Based on this, the present invention provides a rainfall monitoring system and method based on a heterogeneous sensing path graph spatiotemporal network to solve at least one of the above-mentioned technical problems.

[0004] To achieve the above objectives, a rainfall monitoring method based on a heterogeneous sensor path graph spatiotemporal network is applied to a graduated cylinder. The graduated cylinder is equipped with a pressure sensor, a laser rangefinder, and a temperature sensor. The rainfall monitoring method based on the heterogeneous sensor path graph spatiotemporal network includes the following steps:

[0005] Step S1: Periodically acquire time-series data monitored by pressure sensors, laser rangefinders, and temperature sensors as multi-source monitoring data;

[0006] Step S2: Set the bottom physical node, top physical node and multiple virtual intermediate nodes in the measuring cylinder, and establish bidirectional connection relationships between the nodes to construct a heterogeneous sensing path diagram;

[0007] Step S3: Input the multi-source monitoring data and heterogeneous sensing pathway map into the pre-trained heterogeneous sensing pathway map spatiotemporal network model. Based on the heterogeneous graph attention mechanism, perform graph attention information propagation and bidirectional feature fusion between nodes of the heterogeneous sensing pathway map, and obtain the fused water level estimate and rainfall estimate through regression decoding.

[0008] Step S4: Analyze the global state of the heterogeneous sensing path diagram using the spatiotemporal network model of the heterogeneous sensing path diagram, and output the health diagnosis results of each component of the device.

[0009] This invention also provides a rainfall monitoring system based on a heterogeneous sensor path graph spatiotemporal network, which executes the rainfall monitoring method based on a heterogeneous sensor path graph spatiotemporal network as described above. The rainfall monitoring system based on the heterogeneous sensor path graph spatiotemporal network includes:

[0010] The multi-source data acquisition module is used to periodically acquire time-series data monitored by pressure sensors, laser rangefinders, and temperature sensors as multi-source monitoring data.

[0011] The heterogeneous path diagram construction module is used to set the bottom physical node, the top physical node and multiple virtual intermediate nodes in the graduated cylinder, and establish bidirectional connection relationships between the nodes to construct a heterogeneous sensing path diagram.

[0012] The rainfall inference module is used to input multi-source monitoring data and heterogeneous sensor path maps into a pre-trained heterogeneous sensor path map spatiotemporal network model. Based on the heterogeneous graph attention mechanism, graph attention information propagation and bidirectional feature fusion are performed between nodes of the heterogeneous sensor path map to obtain fused water level estimates and rainfall estimates through regression decoding.

[0013] The equipment health diagnosis module is used to analyze the global state of the heterogeneous sensor path diagram through the spatiotemporal network model of the heterogeneous sensor path diagram, and output the health diagnosis results of each component of the equipment.

[0014] The beneficial effects of this invention are as follows:

[0015] By constructing a heterogeneous sensing pathway graph within the logical space of a graduated cylinder, comprising two physical nodes (bottom pressure and top laser) and multiple virtual intermediate nodes, and utilizing a heterogeneous graph attention spatiotemporal network model for bidirectional information propagation and feature fusion, high-precision fusion estimation of rainfall under dynamic conditions is achieved. The virtual intermediate nodes elevate the one-dimensional measurement problem to a graph space, enabling the model to explicitly learn and characterize the implicit physical states at different heights within the graduated cylinder. This method deeply explores and models the complex nonlinear relationships between sensors caused by factors such as water surface fluctuations and density changes, avoiding the large rainfall estimation bias caused by the inability of traditional fusion algorithms to characterize physical pathway connections, thus providing a reliable data foundation for accurate rainfall monitoring.

[0016] By analyzing the global state of the heterogeneous sensing pathway graph and inputting the fused features of all nodes into a classifier, a refined and proactive diagnosis of the health of each component of the equipment is achieved. The model can identify "sub-healthy" states, such as contamination of the inner wall of the measuring cylinder or slight blockage of the guide tube, from the changes in the interaction patterns of pressure and laser data across the entire pathway graph. This graph-based global state diagnostic approach avoids the problem of long fault diagnosis delays caused by coarse monitoring granularity and response lag in existing technologies. It provides scientific decision support for preventative maintenance and long-term stable operation of equipment, effectively reducing data quality risks caused by latent equipment faults.

[0017] By introducing a heterogeneous graph attention mechanism, weights are dynamically calculated during information propagation between nodes, efficiently integrating multi-source heterogeneous monitoring data into an adaptive response to changes in operating conditions. When faced with strong interference, such as momentary blockage of laser sensors, the model automatically reduces the credibility of information from affected nodes, focusing on inference based on information from other nodes, thus ensuring the stability and robustness of the output results. This efficient adaptive fusion and inference approach improves the system's survivability in complex environments, solving the problem of traditional methods struggling to dynamically adjust data credibility under interference. Furthermore, the model's lightweight structure makes it easy to deploy at edge environments, providing strong technical support for high-precision, high-reliability rainfall monitoring in unattended environments. Attached Figure Description

[0018] Figure 1 This is a schematic flowchart of the rainfall monitoring method based on a heterogeneous sensing path graph spatiotemporal network according to the present invention.

[0019] Figure 2 This is a schematic diagram of the module of the rainfall monitoring system based on the heterogeneous sensing path graph spatiotemporal network of the present invention.

[0020] Figure 3 This is a schematic diagram of the hardware structure of the rainfall monitoring system in this invention;

[0021] Figure 4 This is a schematic diagram of the node construction of the heterogeneous sensing path diagram in this invention;

[0022] Figure 5 This is a line graph comparing the measurement accuracy of the present invention under different rainfall intensities;

[0023] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0024] The technical method of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0025] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor methods and / or microcontroller methods.

[0026] It should be understood that although the terms "first," "second," etc., may be used herein to describe various units, these units should not be limited by these terms. These terms are used merely to distinguish one unit from another. For example, without departing from the scope of the exemplary embodiments, a first unit may be referred to as a second unit, and similarly, a second unit may be referred to as a first unit. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0027] To achieve the above objectives, please refer to Figures 1 to 5 This invention provides a rainfall monitoring method based on a heterogeneous sensor path graph spatiotemporal network, applied to a graduated cylinder. The graduated cylinder is equipped with a pressure sensor, a laser rangefinder, and a temperature sensor. The rainfall monitoring method based on the heterogeneous sensor path graph spatiotemporal network includes the following steps:

[0028] Preferably, step S1: periodically acquire time-series data monitored by the pressure sensor, laser rangefinder, and temperature sensor as multi-source monitoring data;

[0029] Optionally, step S1 includes the following steps:

[0030] Step S11: Set the rainfall monitoring time window, and backtrack the duration of the rainfall monitoring time window from the current time as the endpoint, and extract the original pressure signal sequence, original laser distance signal sequence and temperature signal sequence collected by the pressure sensor, laser distance sensor and temperature sensor within the time period.

[0031] Step S12: Perform time-series data preprocessing on the original pressure signal sequence, the original laser distance signal sequence, and the temperature signal sequence respectively to obtain multi-source monitoring data; wherein, the multi-source monitoring data includes pressure filter value, laser filter value, pressure-converted water level value, laser-converted water level value, and temperature compensation parameters.

[0032] In one implementation of this invention, the rainfall monitoring time window in step S11 is preset to 60 seconds, and the edge module controller performs periodic data acquisition at a frequency of 1 Hz. Specifically, at the current time t, the controller extracts data from the time period from t-60 seconds to t seconds, thereby forming a raw pressure signal sequence, a raw laser distance signal sequence, and a temperature signal sequence containing 60 data points, each data point having a precise timestamp.

[0033] In one implementation of this invention, the time-series data preprocessing in step S12 first performs a moving average filter on the original pressure signal sequence and the original laser distance signal sequence. Specifically, a moving window of size 3 is set, and for each point in the sequence, the value of itself and the values ​​of the two preceding points are taken as the arithmetic mean, which is used as the filtered value for that point. For example, if the values ​​of the 3rd, 4th, and 5th points of the original pressure signal sequence are 101.32 kPa, 101.35 kPa, and 101.34 kPa, respectively, then the pressure filtered value of the 5th point is (101.32 + 101.35 + 101.34) / 3 = 101.337 kPa.

[0034] In one implementation of this invention, water level conversion and temperature compensation are performed.

[0035] It should be noted that the water level conversion based on the pressure sensor follows the following formula: .in, To convert pressure to water level value, This is the current pressure filter value. The external atmospheric pressure (which can be measured by an independent barometer or obtained from a meteorological service; in this embodiment, it is set to standard atmospheric pressure of 101.325 kPa). It is the acceleration due to gravity. This is the density of water after temperature compensation.

[0036] Specifically, the density of water The compensation calculation is performed using the current temperature signal, and the formula is an empirical polynomial: ,in The temperature value is in Celsius. to This is an internationally recognized standard coefficient for fitting the density of pure water at different temperatures. Finally, the multi-source monitoring data set was integrated to include pressure-filtered values, laser-filtered values, pressure-converted water level values, laser-converted water level values, and real-time water density values ​​as temperature compensation parameters, which were then used as inputs for subsequent model processing.

[0037] Of particular importance, time-series data preprocessing was performed on the original pressure signal sequence, the original laser distance signal sequence, and the temperature signal sequence to obtain multi-source monitoring data, including:

[0038] The original pressure signal sequence is subjected to median filtering, and the median value of all sampling points in the sequence is calculated as the pressure filter value.

[0039] The average value of the temperature signal sequence within the time period is extracted as the temperature compensation parameter;

[0040] Jump detection is performed on the original laser distance signal sequence, and the absolute value of the difference between adjacent sampling points is calculated. When the absolute value of the difference exceeds the preset jump threshold, it is marked as an abnormal point. After removing all abnormal points, the arithmetic mean of the remaining valid sampling points is calculated as the laser filter value.

[0041] Based on the temperature compensation parameters, find the corresponding water density correction factor and laser refraction correction factor from the preset temperature compensation factor table;

[0042] The pressure-filtered value is divided by the product of the water density correction factor and the gravitational acceleration constant to obtain the pressure-converted water level value.

[0043] Obtain the total height parameter of the measuring cylinder, and subtract the product of the laser filter value and the laser refraction correction coefficient from the total height parameter of the measuring cylinder to obtain the laser-converted water level value;

[0044] The pressure filter value, laser filter value, pressure-converted water level value, laser-converted water level value, and temperature compensation parameter are packaged into multi-source monitoring data.

[0045] In one implementation of this invention, the original pressure signal sequence (here referring to the gauge pressure signal after deducting atmospheric pressure) undergoes median filtering. Specifically, the 60 original pressure signal sampling points collected within the time window are arranged in ascending order of their numerical values, and the arithmetic mean of the values ​​of the 30th and 31st sampling points after sorting is taken as the pressure filtering value for this calculation. For example, if the value of the 30th sampling point after sorting is 9800 Pa and the value of the 31st sampling point is 9802 Pa, then the pressure filtering value is (9800 + 9802) / 2 = 9801 Pa.

[0046] In one implementation of this invention, the arithmetic mean of 60 temperature signal sampling points within a time period is extracted as a temperature compensation parameter. It should be noted that this temperature compensation parameter is directly used to find the subsequent correction coefficient. For example, if the sum of the 60 temperature sampling points is 1212 degrees Celsius, then the temperature compensation parameter is 1212 / 60 = 20.2 degrees Celsius.

[0047] In one operation step of this embodiment of the invention, jump detection is performed on the original laser distance signal sequence. Specifically, the preset jump threshold is determined according to the water inflow rate of the measuring cylinder, and in this embodiment, it is set to 0.05 meters. The absolute value of the difference between two adjacent sampling points in the sequence is calculated one by one. If the value is greater than 0.05 meters, the latter sampling point is marked as an outlier. For example, for a short sequence [0.481, 0.482, 0.215, 0.483], |0.482-0.481|=0.001 meters (less than the threshold) and |0.215-0.482|=0.267 meters (greater than the threshold). The sampling point at 0.215 meters is then marked as an outlier and removed. After removing all outliers within the time window, the arithmetic mean of all remaining valid sampling points is calculated as the final laser filtering value.

[0048] In one implementation of this invention, a lookup operation is performed from a preset temperature compensation coefficient table built into the edge module controller based on the temperature compensation parameters obtained in the previous step. It should be noted that this table is generated based on a large amount of experimental data and stored in key-value pair format. For example, when the temperature compensation parameter is 20.2 degrees Celsius, the corresponding water density correction coefficient is found to be 998.2 kg / m³, and the laser refraction correction coefficient is 1.0001.

[0049] In one operation step of this embodiment of the invention, the pressure-to-water-level value is calculated based on the obtained parameters, and the pressure filter value is divided by the product of the water density correction coefficient and the gravitational acceleration constant (9.8 in this embodiment).

[0050] In one operation step of this embodiment of the invention, the laser-converted water level value is calculated to obtain a preset total height parameter of the measuring cylinder, which is 1.5 meters. Then, the product of the laser filter value and the laser refraction correction coefficient is subtracted from the total height parameter of the measuring cylinder. For example, if the laser filter value is 0.481 meters, then the laser-converted water level value = 1.5 meters - (0.481 meters × 1.0001) ≈ 1.0188 meters. All the calculated results, namely the pressure filter value, laser filter value, pressure-converted water level value, laser-converted water level value, and temperature compensation parameter, are encapsulated into a structured data packet as the multi-source monitoring data at that moment.

[0051] Please see Figure 3This is a schematic diagram of the hardware structure of the rainfall monitoring system in this invention. The rain collection port is a circular funnel-shaped structure, located at the top of the device, used to collect rainfall and guide rainwater into the inlet pipe. The inlet pipe is a cylindrical water guide, connecting the rain collection port and the lower cylindrical main body, used to transport the rainwater collected by the rain collection port into the interior of the main body. The pressure sensor (A) includes a strain gauge pressure detection unit and a pressure signal amplification circuit. The strain gauge pressure detection unit is fixed to the force-bearing area at the bottom of the cylindrical main body, used to sense the static pressure generated by the internal water column. When the water level changes, causing a change in bottom pressure, the strain gauge pressure detection unit outputs a resistance change signal. The pressure signal amplification circuit is connected to the strain gauge pressure detection unit, receives the resistance change signal, performs differential amplification processing, and outputs a signal corresponding to the water level. The original pressure signal corresponding to the column pressure; the laser range sensor (B) includes a laser pulse emitting unit and a laser pulse receiving unit. The laser pulse emitting unit is installed at the center of the top of the cylindrical body and emits a laser pulse beam downward along the vertical axis of the body; the laser pulse receiving unit is located on the side of the laser pulse emitting unit, receives the laser pulse beam reflected back from the water surface, calculates the distance from the laser head to the water surface based on the round-trip time of the laser pulse, and outputs the original laser distance signal; the edge module controller is installed on the outer side of the cylindrical body and is used to receive the original signals output by the pressure sensor and the laser range sensor, and to perform data processing and logic control; the drain outlet is a pipe structure with bends, located on the bottom side of the cylindrical body, and is used to discharge internal rainwater.

[0052] Preferably, step S2: set the bottom physical node, the top physical node and multiple virtual intermediate nodes in the measuring cylinder, and establish bidirectional connection relationships between the nodes to construct a heterogeneous sensing path diagram;

[0053] Optionally, step S2 includes the following steps:

[0054] Step S21: Set the number of virtual nodes; wherein the number of virtual nodes is an integer value greater than or equal to 6 and less than or equal to 16;

[0055] Step S22: Divide the obtained total height parameter of the graduated cylinder by the number of virtual nodes plus one, and use the result as the node spacing parameter;

[0056] Step S23: Starting from the bottom physical node of the graduated cylinder, generate virtual intermediate nodes one by one along the height direction of the graduated cylinder according to the node spacing parameters; wherein, the bottom physical node of the graduated cylinder is the pressure measurement point of the pressure sensor in the graduated cylinder, and the top physical node of the graduated cylinder is the laser measurement point of the laser rangefinder.

[0057] Step S24: Set the height position of the bottom physical node to zero, and set the height position of the top physical node to the total height parameter of the measuring cylinder;

[0058] Step S25: Establish bidirectional connections between the bottom physical node, all virtual intermediate nodes, and the top physical node in order of their height position in the measuring cylinder from low to high, and construct a heterogeneous sensing path diagram.

[0059] In one implementation of this invention, for example, the number of virtual nodes is set to 8. It should be noted that this number is a hyperparameter preset during the model design stage, and its value is between 6 and 16, taking into account both the model's ability to express the internal state of the measuring cylinder and the computing power limitations of the edge computing device.

[0060] In one implementation of this invention, a preset total height parameter of the graduated cylinder, which is 1.5 meters, is obtained, and this total height parameter is divided by the number of virtual nodes plus one. Specifically, the node spacing parameter = 1.5 meters / (8 + 1) = 0.1667 meters.

[0061] In one operation step of this embodiment of the invention, virtual intermediate nodes are generated at the computational level. It should be noted that the bottom physical node logically corresponds to the pressure sensor installed at the bottom of the measuring cylinder, and the top physical node corresponds to the laser rangefinder sensor installed at the top. Starting from the bottom physical node, eight virtual intermediate nodes are generated sequentially along the height direction with a node spacing parameter of 0.1667 meters, and are respectively labeled as follows: .

[0062] In one implementation of this invention, the spatial height position of all nodes is assigned, and the bottom physical node (marked as...) is... Assign a height of 0 meters to the top physical node (marked as...). The height of the virtual intermediate node is assigned a value of 1.5 meters. The height of the virtual intermediate node is calculated based on its index and node spacing parameters, using the following formula: Node spacing parameters, where This refers to the sequence number of the virtual node (from 1 to 8). For example, a virtual intermediate node. The altitude is 1 × 0.1667 = 0.1667 meters. The height is 2 × 0.1667 = 0.3334 meters, and so on, until... The altitude is 8 × 0.1667 = 1.3336 meters.

[0063] In one operation step of this embodiment of the invention, a heterogeneous sensing path graph is constructed according to step S25. Specifically, all 10 nodes (1 bottom physical node, 8 virtual intermediate nodes, and 1 top physical node) are sorted from low to high according to their height position, and bidirectional connections are established between adjacent nodes in sequence to form a chain-like graph structure. The connection relationship is as follows: It should be noted that this connection relationship is represented as a 10×10 adjacency matrix in the implementation. The matrix has a value of 1 only at the positions corresponding to the above adjacent nodes, and a value of 0 at the other positions, thus solidifying the physical path structure between sensors in the calculation model.

[0064] Please see Figure 4 This is a schematic diagram of the node construction of the heterogeneous sensing path diagram in this invention. The top physical node is a laser rangefinder sensor, located at the center of the top of the device, used to emit laser pulses and receive laser pulses reflected from the water surface to measure the distance from the laser head to the water surface; the bottom physical node is a pressure sensor, located in the force-bearing area at the bottom of the measuring cylinder, used to sense the static pressure generated by the internal water column and output the original pressure signal corresponding to the water column pressure; the temperature sensor is installed on the outer side of the measuring cylinder to collect the ambient temperature and output a temperature signal to provide temperature compensation parameters; the number of virtual nodes... The recommended range of values ​​is: Its spatial location has a linear correspondence with its physical height, i.e., a node. The corresponding height is ,in For example, when When the total height of the measuring cylinder is... The length is 1.5m, which is logically discretized into 9 levels. Each virtual node represents a learnable "hidden physical state layer". The bottom physical node, all virtual intermediate nodes and the top physical node are connected bidirectionally from low to high according to their height position in the measuring cylinder, thereby constructing a heterogeneous sensing path graph and realizing the transformation of the one-dimensional scalar measurement problem into a rich graph structure learning problem.

[0065] Preferably, step S3: input the multi-source monitoring data and heterogeneous sensing pathway map into the pre-trained heterogeneous sensing pathway map spatiotemporal network model, and perform graph attention information propagation and bidirectional feature fusion between nodes of the heterogeneous sensing pathway map based on the heterogeneous graph attention mechanism, so as to obtain the fused water level estimate and rainfall estimate through regression decoding.

[0066] Optionally, the internal structure of the pre-trained heterogeneous sensing pathway spatiotemporal network model in step S3 includes:

[0067] The spatiotemporal feature embedding module is used to generate an initial feature vector for each node of the heterogeneous sensing pathway map based on multi-source monitoring data.

[0068] The heterogeneous graph attention spatiotemporal convolution module is used to receive the initial feature vector, perform bidirectional message propagation from bottom to top and top to bottom, and update the state of each node using a gated recurrent unit to generate node temporal features.

[0069] The multi-task output module is used to receive the temporal features of nodes, decode and generate model output results, and complete rainfall regression and equipment health classification.

[0070] In one implementation of this invention, the spatiotemporal feature embedding module first receives multi-source monitoring data within a time window (e.g., 60 seconds). Specifically, for the bottom physical nodes... The corresponding pressure filter value sequence and temperature compensation parameter sequence are input into a dedicated multilayer perceptron with two hidden layers for encoding, and the output is a 24-dimensional initial feature vector. For the top physical node Then, the corresponding laser filter value sequence is input into another structurally independent multilayer perceptron, which is also encoded into a 24-dimensional initial feature vector. .

[0071] It should be noted that for the 8 virtual intermediate nodes Its initial features consist of two parts: one part is determined by a sequence number associated with the node's spatial location. The relevant learnable parameter matrix is ​​initialized; another part is injected with its normalized spatial location code, specifically, its normalized elevation position (e.g., using sine and cosine functions). The height of the virtual node (0.1667 meters) is mapped to a high-dimensional vector and added to the vector, thus giving each virtual node a sense of physical space hierarchy at the feature level.

[0072] In one implementation of this invention, the heterogeneous graph attention spatiotemporal convolution module performs three information propagation and node state updates, and the node... to its neighboring nodes Information dissemination weight It is dynamically calculated using the heterogeneous graph attention mechanism. Specifically, the calculation formula is as follows: ,in and They are nodes and nodes The 24-dimensional feature vector, A shared, learnable weight matrix is ​​used to linearly transform the input features from 24 dimensions to 16 dimensions. The symbol represents the concatenation operation, forming a 32-dimensional concatenated vector. Given a learnable 32-dimensional attention parameter vector, This is the activation function for linear units with leakage correction. The calculated... The attention weights are then normalized using the Softmax function. For example, when laser sensor data experiences drastic changes, the model can learn to automatically calculate the changes from the top physical node. Attention weights propagated to downstream nodes The value decreased significantly, thus weakening the impact of the anomalous information.

[0073] In one operation step of an embodiment of the present invention, a virtual intermediate node... The state update is performed using a gated recurrent unit (GRU) with a hidden layer dimension of 24. Specifically, the node will update the features it receives from all its neighboring nodes (above and below). First, it undergoes a weight matrix transformation (W), then is multiplied by the corresponding attention weights. Then, these weighted pieces of information are summed to form an aggregated message vector. This aggregated message vector is then compared with the virtual node. The state features from the previous time step are used as input to the gated recurrent unit (ROU) to calculate the latest state features of the node at the current time step. It should be noted that the introduction of the gated ROU gives each virtual node temporal memory, enabling it to learn the implicit dynamic representation of the water level at that depth. After three graph convolutional updates, the feature sequence obtained from each node is fed into a lightweight temporal convolutional network with a kernel size of 3 and 16 channels to further capture the evolution pattern of each node's features over time.

[0074] In one operation step of this embodiment of the invention, the multi-task output module is responsible for decoding and generating the final result. For the rainfall regression task, the bottom physical node at the final moment is used... and top physical node The feature vectors are concatenated to form a 48-dimensional fused feature vector. This fused feature vector is then passed through a regression network with two fully connected layers (the network structure has 48-dimensional input, 16-dimensional hidden layers, and a final 1-dimensional output) to obtain the final estimated water level height. For the device health classification task, the feature vectors of all 10 nodes in the graph at the final time are subjected to global average pooling. This involves summing the feature vectors of all nodes element-wise and then averaging the sums to obtain a feature vector representing the global state of the entire graph system. The global feature vector is fed into a classifier consisting of a fully connected layer and a Softmax activation function, which outputs a health probability vector containing three elements that sum to 1, corresponding to the health probabilities of the pressure sensor path, the laser sensor path, and the measuring cylinder pipe, respectively.

[0075] Optionally, the spatiotemporal feature embedding module is used to generate an initial feature vector for each node of the heterogeneous sensing path graph, including:

[0076] For each node in the heterogeneous sensing path diagram, the height position of the node is divided by the total height parameter of the measuring cylinder to obtain the normalized height coordinate. The normalized height coordinate is then subjected to sine and cosine transformations, and the transformation results are spliced ​​together to form a spatial position encoding vector.

[0077] For the bottom physical nodes, the pressure filter values ​​and temperature compensation parameters in the multi-source monitoring data are processed by MLP encoding, thereby mapping them into pressure time series feature vectors;

[0078] For the top physical node, the laser filter values ​​in the multi-source monitoring data are processed by MLP encoding, thereby mapping them into laser temporal feature vectors;

[0079] The pressure time-series feature vector and the spatial location encoding vector are concatenated to form the initial feature vector of the bottom physical node; the laser time-series feature vector and the spatial location encoding vector are concatenated to form the initial feature vector of the top physical node.

[0080] For virtual intermediate nodes, a virtual water level interpolation is obtained by linear interpolation between pressure-converted water level values ​​and laser-converted water level values ​​based on the node's height position. The virtual water level interpolation is then mapped to the spatial position encoding vector as the initial feature vector of the virtual intermediate node.

[0081] In one implementation of this invention, spatial position encoding vectors are generated for all 10 nodes in the heterogeneous sensing path diagram. Specifically, for any node, its height position is first divided by a preset total cylinder height parameter of 1.5 meters to obtain a normalized height coordinate between 0 and 1. Then, a set of sine and cosine functions of different frequencies are used to transform this normalized coordinate, generating an 8-dimensional vector as the spatial position encoding vector for that node. For example, for a virtual intermediate node with a height position of 0.3334 meters... Its normalized height coordinates are 0.3334 / 1.5≈0.2223, and after transformation, its unique 8-dimensional spatial location encoding vector is obtained.

[0082] In one implementation of this invention, the bottom physical node Generate the stress time-series feature vector. It should be noted that this step uses a pre-trained multilayer perceptron model containing one input layer, one 64-neuron hidden layer, and one 16-neuron output layer. Specifically, the stress filter values ​​and temperature compensation parameters for 60 time steps within the current time window are concatenated into a 120-dimensional input vector, which is then input into the multilayer perceptron. After forward propagation, the model outputs a 16-dimensional vector, which is the stress time-series feature vector.

[0083] In one operation step of an embodiment of the present invention, the top physical node... Generate the laser temporal feature vector. Specifically, this step uses a separate multilayer perceptron model with a structure of 60 neurons in the input layer, 32 neurons in the hidden layer, and 16 neurons in the output layer. The laser filter values ​​at 60 time points within the current time window are fed into this model as a 60-dimensional input vector. After calculation, the model outputs a 16-dimensional vector, which serves as the laser temporal feature vector.

[0084] In one operation step of this embodiment of the invention, feature splicing is performed on the physical nodes. Specifically, the bottom physical nodes are... The 16-dimensional pressure temporal feature vector is concatenated with its 8-dimensional spatial location encoding vector to form a 24-dimensional vector, which serves as the initial feature vector for the bottom physical node. Similarly, the top physical node... The 16-dimensional laser temporal feature vector is concatenated with its 8-dimensional spatial location encoding vector to form a 24-dimensional vector, which serves as the initial feature vector for the top physical node.

[0085] In one implementation of this invention, an initial feature vector is generated for all virtual intermediate nodes. Based on the pressure-converted water level value and the laser-converted water level value, a virtual water level interpolation is calculated for each virtual node. The calculation formula is: Virtual Water Level Interpolation. ,in To convert pressure to water level value, For laser-based water level conversion, This represents the current height position of the virtual node. This represents the total height parameter of the graduated cylinder. For example, if the pressure-converted water level is 1.0019 meters and the laser-converted water level is 1.0188 meters, for a virtual node V_C4 with a height of 0.6668 meters, its virtual water level interpolation is 1.0019 + (1.0188 - 1.0019) × (0.6668 / 1.5) ≈ 1.0094 meters. Subsequently, this scalar value of the virtual water level interpolation is mapped to a 16-dimensional feature vector through a linear transformation layer, and then concatenated with the node's 8-dimensional spatial location encoding vector to finally form a 24-dimensional vector, which serves as the initial feature vector for this virtual intermediate node.

[0086] Optionally, the heterogeneous graph attention spatiotemporal convolution module includes the following by performing bidirectional message propagation, both bottom-up and top-down:

[0087] For each node in the heterogeneous sensing path diagram, the initial feature vector of the node is extracted and linearly transformed using preset shared weight parameters;

[0088] For each pair of adjacent nodes, the features of the two nodes after linear transformation are concatenated, the concatenation result is multiplied by the preset attention parameter vector, and then processed by the preset nonlinear activation function to obtain the original attention score of the node pair.

[0089] Extract the raw attention scores between a node and all its neighboring nodes, calculate the exponential function value of each raw attention score, and divide each exponential function value by the sum of all exponential function values ​​to obtain the normalized attention weights from the node to each neighboring node.

[0090] The message aggregation vector of a node is obtained by multiplying the features of each adjacent node after linear transformation with the corresponding attention normalization weights and summing the results.

[0091] In one implementation of this invention, firstly, for each node in the heterogeneous sensing path graph, its 24-dimensional initial feature vector is extracted, and then linearly transformed using a preset weight parameter matrix shared among all nodes. It should be noted that this shared weight parameter is a learnable matrix W of size 24 rows and 16 columns, which maps the 24-dimensional initial feature vector of each node to a 16-dimensional intermediate feature vector to facilitate subsequent attention score calculation.

[0092] In one implementation of this invention, for any pair of adjacent nodes in the graph (e.g., virtual intermediate nodes) Its adjacent nodes above Calculate the raw attention score. Specifically, and The 16-dimensional intermediate feature vectors obtained after linear transformations are concatenated to form a 32-dimensional concatenated vector. This concatenated vector is then appended to a pre-defined 32-dimensional learnable attention parameter vector. Perform a dot product operation to obtain a scalar value. Finally, input this scalar value into a pre-defined nonlinear activation function with leaky ReLU (LeakyReLU) for processing, and the output is the result. and This is the original score for attention between nodes.

[0093] In one operation step of an embodiment of the present invention, a certain node (e.g.) is calculated. The attention normalization weights are applied to all its neighboring nodes. It is important to note that... In a chain graph, there are two adjacent nodes: the one below. and above Calculate them separately. and and The raw attention score, denoted as and Then, the normalized weights are calculated using the Softmax function, and the formula is as follows: ,in Representative node The set of all neighboring nodes. For example, if the calculated raw attention score... It is 0.8. If the value is 0.5, the corresponding exponential function values ​​are exp(0.8)≈2.226 and exp(0.5)≈1.649, respectively. The sum of all exponential function values ​​is 2.226 + 1.649 = 3.875. Therefore, arrive The attention normalized weight is 2.226 / 3.875≈0.574, to The attention normalization weight is 1.649 / 3.875≈0.426.

[0094] In one operation step of an embodiment of the present invention, a node Generate its message aggregation vector. Specifically, The 16-dimensional intermediate feature vector after linear transformation is multiplied by its corresponding attention normalization weight of 0.574 to obtain a weighted feature vector. Simultaneously, The 16-dimensional intermediate feature vector after linear transformation is multiplied by its corresponding attention normalization weight of 0.426 to obtain another weighted feature vector. Finally, these two weighted feature vectors are added element-wise to obtain the 16-dimensional vector. The message aggregation vector received during this message propagation is a vector that incorporates weighted information from all its neighboring nodes.

[0095] Optionally, the heterogeneous graph attention spatiotemporal convolution module further includes, by performing bidirectional message propagation in both bottom-up and top-down directions:

[0096] The process propagates upwards layer by layer along the heterogeneous sensing path graph, starting from the bottom physical node. Each node uses the aggregated vector of messages from its adjacent nodes below as the feature for upward propagation.

[0097] Starting from the top physical node, the process propagates downwards along the heterogeneous sensing path graph, with each node using the aggregated message vector from its adjacent node above as a feature for downward propagation.

[0098] For each node, the upward and downward propagation features of that node are concatenated and linearly transformed using preset feature fusion parameters to output the bidirectional propagation features of the node.

[0099] In one implementation of this invention, a bottom-up propagation is performed, propagating from the bottom physical node. Initially, since it has no adjacent nodes below it, its upward propagation features are initialized as a 16-dimensional vector of all zeros. For its adjacent nodes above it... It will The message aggregation vector (which actually represents) Its own state information serves as its upward propagation characteristic. Similarly, virtual intermediate nodes... Will The message aggregation vector serves as its upward propagation feature, and this process proceeds layer by layer upwards until it reaches the top physical node. Received from The message aggregation vector serves as its upward propagation feature, and its feature (containing "water pressure" information) propagates along the graph edges to upstream nodes.

[0100] In one implementation of this invention, a top-down propagation is performed in parallel, a process symmetrical to the upward propagation. Specifically, the propagation begins from the top physical node. Initially, since it has no adjacent nodes above it, its downward propagation features are also initialized as a 16-dimensional vector of all zeros. For its adjacent nodes below it... It will The message aggregation vector serves as its downward propagation feature. This process proceeds layer by layer downwards along the graph's connectivity until it reaches the bottom physical node. Received from The message aggregation vector serves as its downward propagation feature, and its features (including "water surface distance" information) propagate downstream along the graph edges.

[0101] In one operational step of this embodiment of the invention, bidirectional propagation features are fused for each node. It should be noted that, at this point, except for boundary nodes, each node has obtained a 16-dimensional upward propagation feature and a 16-dimensional downward propagation feature. Specifically, virtual intermediate nodes are used... For example, it will propagate the received 16-dimensional features upwards (from... ) and 16-dimensional downpropagation features (from The features are concatenated to form a 32-dimensional concatenated feature vector. This 32-dimensional vector is then processed through a pre-defined feature fusion linear transformation layer with learnable parameters (its weights are a 32x24 matrix), ultimately outputting a 24-dimensional vector, which is the final output. The node exhibits bidirectional propagation characteristics, simultaneously fusing contextual information from both the pressure sensor and laser sensor directions. For boundary nodes... It concatenates a 16-dimensional all-zero vector and a vector from... The 16-dimensional downward propagation feature is also transformed by the above linear transformation to obtain its 24-dimensional bidirectional propagation feature.

[0102] Optionally, the heterogeneous graph attention spatiotemporal convolution module uses gated recurrent units to update the state of each node and generate node temporal features, including:

[0103] For each node, read the historical state vector of the node in the previous rainfall monitoring time window. If it is the first run, initialize the historical state vector of the node to a zero vector.

[0104] The updated state vector is obtained by updating the features based on the historical state vector and bidirectional propagation features through the update gate and reset gate inside the gated loop unit.

[0105] A one-dimensional convolution kernel is applied to the updated state vector in the time dimension. The convolution operation extracts the temporal evolution pattern of the updated state vector, and the node temporal features are output after convolution processing.

[0106] In one implementation of this invention, for each node in the heterogeneous sensing path graph, its historical state vector calculated in the previous rainfall monitoring time window (i.e., the previous second) is first read from memory. It should be noted that this historical state vector is a 24-dimensional vector that encodes the spatiotemporal information of the node up to the previous moment. If it is the first run or has just been restarted, there is no historical state vector; in this case, the historical state vector of the node is initialized to a 24-dimensional vector of all zeros.

[0107] In one implementation of this invention, each node utilizes a gated recurrent unit (GRU) with a hidden layer dimension of 24 for state updates. Specifically, the 24-dimensional historical state vector from the previous time step and the 24-dimensional bidirectional propagation feature calculated at the current time step are simultaneously used as inputs to the GRU. It is important to note that the update and reset gates within the GRU automatically learn and determine, based on these two input vectors and through a series of internal matrix operations and activation functions, how much historical information should be retained and how much new information from the current time step should be incorporated. After calculation by the GRU, a final 24-dimensional updated state vector is output, representing the node's latest state at the current time step, incorporating historical memory and current observations.

[0108] In one operation step of this embodiment of the invention, a temporal convolutional network is applied to the updated state vector sequence for final feature extraction. Specifically, at each time point within the current time window (e.g., 60 seconds in length), each node generates an updated state vector, forming an updated state vector sequence of length 60 and 24 dimensions per element. A preset one-dimensional convolutional kernel with a kernel size of 3 and 16 output channels is applied to this sequence in the time dimension. For example, to calculate the output feature at second 58, the convolutional kernel will simultaneously operate on the three 24-dimensional updated state vectors at seconds 58, 59, and 60, extracting the state evolution pattern at these three consecutive time points through convolution operations and outputting a 16-dimensional feature. This convolutional operation slides across the entire time series, ultimately outputting a sequence of length 60 and 16 dimensions per element. This sequence is the final temporal feature of the node, used for subsequent multi-task output.

[0109] Optionally, the multi-task output module for completing rainfall regression includes:

[0110] The node temporal features of the bottom physical node and the top physical node are extracted from the heterogeneous sensing path diagram, and the two node temporal features are concatenated in the feature dimension to obtain a joint feature vector.

[0111] The joint feature vector is input into a preset rainfall regression network, and the joint feature vector is mapped into a single scalar output through layer-by-layer dimensional compression to obtain the fused water level estimate.

[0112] Obtain the diameter parameters of the rain collector and calculate the area parameters of the rain collector;

[0113] The water level estimate is used as the water level height, and the rainfall estimate is calculated based on the water level height and the rain collector area parameter.

[0114] In one implementation of this invention, the node timing features at the final moment are first extracted from the heterogeneous sensing path graph. Specifically, the bottom physical nodes are obtained respectively. and top physical node The 16-dimensional node temporal feature vector corresponding to the end of the current time window. Then, these two 16-dimensional feature vectors are concatenated along their feature dimensions to form a 32-dimensional joint feature vector, which integrates the top-level semantic information from the two physical sensing pathways of pressure and laser.

[0115] In one implementation of this invention, the 32-dimensional joint feature vector generated in the previous step is input into a preset rainfall regression network. It should be noted that this network is a fully connected neural network, its structure including a 32-neuron input layer, a 16-neuron hidden layer, and a 1-neuron output layer. The joint feature vector undergoes forward propagation and layer-by-layer dimensionality compression within this network, ultimately being mapped to a single scalar output. This output value is the model's predicted final fused water level estimate, which integrates information from all sensors. For example, if the network output scalar value is 1.015, it indicates that the current fused water level estimate is 1.015 meters.

[0116] In one operation step of this embodiment of the invention, a preset rain collector diameter parameter is obtained, which is 200 mm in this embodiment. The cross-sectional area of ​​the rain collector, i.e., the rain collector area parameter, is calculated based on this diameter parameter. Specifically, the calculation formula is as follows: ,in For area, For caliber, Take 3.14159.

[0117] In another operational step of this embodiment of the invention, the final cumulative rainfall estimate is calculated based on the merged water level estimate. It should be noted that in the hardware design of this embodiment, the area of ​​the rain collector is equal to the cross-sectional area of ​​the measuring cylinder; therefore, the water level height inside the measuring cylinder is directly equivalent to the cumulative rainfall depth. Specifically, the merged water level estimate obtained in the second step, 1.015 meters, is converted to millimeters, i.e., 1015 millimeters. This value is directly output as the final cumulative rainfall estimate. If the areas of the rain collector and the measuring cylinder are not equal, the rainfall estimate needs to be converted using the formula: Rainfall estimate = (Merged water level estimate × Measuring cylinder area) / Rain collector area.

[0118] Preferably, step S4: Analyze the global state of the heterogeneous sensing path diagram using the spatiotemporal network model of the heterogeneous sensing path diagram, and output the health diagnosis results of each component of the device.

[0119] Of particular importance, the analysis of the global state of the heterogeneous sensing path graph through the spatiotemporal network model includes:

[0120] Global average pooling is performed on the temporal features of all nodes in the heterogeneous sensing path graph to calculate the average value of the temporal features of all nodes in the feature dimension, thus obtaining the global state vector.

[0121] The heterogeneous sensing path diagram spatiotemporal network model maps the global state vector to the number of output units of the health category through a classification fully connected network. The normalized exponential function is applied to the output units to output a health probability vector containing the health probability of the pressure sensor, the health probability of the laser sensor, and the health probability of the graduated cylinder path.

[0122] Set a health threshold and a consecutive anomaly count threshold. When any probability value in the health probability vector is lower than the health threshold, the corresponding anomaly counter is started. When the anomaly counter reaches the consecutive anomaly count threshold, the device health alarm signal of the corresponding component is triggered.

[0123] In one implementation of this invention, a global average pooling operation is first performed on the 16-dimensional node temporal features corresponding to all 10 nodes (including 2 physical nodes and 8 virtual nodes) in the heterogeneous sensing path graph at the end of the current time window. Specifically, these 10 16-dimensional feature vectors are added element-wise along their feature dimensions, and then each element in the resulting sum vector is divided by the total number of nodes, 10, to calculate a 16-dimensional global state vector, which encapsulates the overall state information of the entire sensing path.

[0124] In another implementation of this invention, the heterogeneous sensing pathway spatiotemporal network model processes the 16-dimensional global state vector obtained in the previous step through a pre-defined classification fully connected network. It should be noted that this network includes a 16-neuron input layer and a 3-neuron output layer, where the three output units correspond to the health status of the pressure sensor, laser sensor, and graduated cylinder pathway, respectively. A normalized exponential function is applied to the raw output values ​​of the three output units. Specifically, the exponential function value of the raw value of each output unit is calculated, and then each exponential function value is divided by the sum of these three exponential function values, ultimately outputting a health probability vector containing three probability values, for example, [0.95, 0.92, 0.98]. The sum of these three values ​​is always 1, representing the health probability of the pressure sensor, the laser sensor, and the graduated cylinder pathway, respectively.

[0125] In one operation step of this embodiment of the invention, a health threshold of 0.7 and a consecutive abnormality count threshold of 5 are preset. When any probability value in the health probability vector first falls below 0.7, the abnormality counter corresponding to that component is activated and counted to 1. If the probability value remains below 0.7 in subsequent continuous monitoring cycles, the abnormality counter is incremented by 1 after each cycle. It should be noted that once the abnormality counter of a component accumulates to 5, a device health alarm signal corresponding to that component will be triggered immediately. For example, if the health probability of the measuring cylinder path is below 0.7 for 5 consecutive cycles, an alarm of "low measuring cylinder patency, cleaning recommended" will be triggered. If the health probability value of the component recovers to 0.7 or above before the counter reaches 5, its corresponding abnormality counter will be reset to zero, and monitoring will restart.

[0126] Please see Figure 2 The present invention also provides a rainfall monitoring system 100 based on a heterogeneous sensor path graph spatiotemporal network, which executes the rainfall monitoring method based on a heterogeneous sensor path graph spatiotemporal network as described above. The rainfall monitoring system based on the heterogeneous sensor path graph spatiotemporal network includes:

[0127] The multi-source data acquisition module 101 is used to periodically acquire time-series data monitored by the pressure sensor, laser rangefinder, and temperature sensor as multi-source monitoring data.

[0128] The heterogeneous path diagram construction module 102 is used to set the bottom physical node, the top physical node and multiple virtual intermediate nodes in the measuring cylinder, and to establish bidirectional connection relationships between the nodes in order to construct a heterogeneous sensing path diagram.

[0129] The rainfall inference module 103 is used to input multi-source monitoring data and heterogeneous sensing pathway graphs into a pre-trained heterogeneous sensing pathway graph spatiotemporal network model. Based on the heterogeneous graph attention mechanism, graph attention information propagation and bidirectional feature fusion are performed between nodes of the heterogeneous sensing pathway graph to obtain fused water level estimates and rainfall estimates through regression decoding.

[0130] The equipment health diagnosis module 104 is used to analyze the global state of the heterogeneous sensing path diagram through the spatiotemporal network model of the heterogeneous sensing path diagram, and output the health diagnosis results of each component of the equipment.

[0131] Please see Figure 5The graph shows a line graph comparing the accuracy of different rainfall monitoring methods. The horizontal axis represents rainfall intensity, divided into four levels: light rain, moderate rain, heavy rain, and torrential rain. The vertical axis represents measurement accuracy, ranging from 50% to 100%. In this case, the heterogeneous fusion method (represented by solid dots and solid lines) maintains a measurement accuracy above 90% across all rainfall intensities (light, moderate, heavy, and torrential rain), with a slow decreasing trend as rainfall intensity increases. The traditional single-sensor method (represented by hollow dots and dashed lines) achieves an accuracy close to 80% in light rain, but its accuracy decreases significantly with increasing rainfall intensity, dropping to around 60% in torrential rain. This graph clearly demonstrates the measurement accuracy advantage of the heterogeneous fusion method in this patent under different rainfall intensities, showing a significantly superior performance compared to the traditional single-sensor method.

[0132] Therefore, the embodiments should be considered as exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of the equivalents of the application are intended to be included within the invention.

[0133] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.

Claims

1. A rainfall monitoring method based on a heterogeneous sensor path graph spatiotemporal network, characterized in that, Applied to a graduated cylinder, which is equipped with a pressure sensor, a laser rangefinder, and a temperature sensor, the rainfall monitoring method based on a heterogeneous sensor path graph spatiotemporal network includes the following steps: Step S1: Periodically acquire time-series data monitored by pressure sensors, laser rangefinders, and temperature sensors as multi-source monitoring data; Step S2: Set the bottom physical node, top physical node, and multiple virtual intermediate nodes in the graduated cylinder, and establish bidirectional connection relationships between the nodes to construct a heterogeneous sensing path diagram; Step S2 includes: Step S21: Set the number of virtual nodes; Step S22: Divide the obtained total height parameter of the graduated cylinder by the number of virtual nodes plus one, and use the result as the node spacing parameter; Step S23: Starting from the bottom physical node of the graduated cylinder, generate virtual intermediate nodes one by one along the height direction of the graduated cylinder according to the node spacing parameters; wherein, the bottom physical node of the graduated cylinder is the pressure measurement point of the pressure sensor in the graduated cylinder, and the top physical node of the graduated cylinder is the laser measurement point of the laser rangefinder. Step S24: Set the height position of the bottom physical node to zero, and set the height position of the top physical node to the total height parameter of the measuring cylinder; Step S25: Establish bidirectional connections between the bottom physical node, all virtual intermediate nodes, and the top physical node in order of their height position in the measuring cylinder from low to high to construct a heterogeneous sensing path diagram. Step S3: Input the multi-source monitoring data and heterogeneous sensor path graph into the pre-trained heterogeneous sensor path graph spatiotemporal network model. Based on the heterogeneous graph attention mechanism, perform graph attention information propagation and bidirectional feature fusion between nodes in the heterogeneous sensor path graph, and obtain the fused water level estimate and rainfall estimate through regression decoding. The internal structure of the pre-trained heterogeneous sensor path graph spatiotemporal network model includes: The spatiotemporal feature embedding module is used to generate an initial feature vector for each node of the heterogeneous sensing pathway map based on multi-source monitoring data. The heterogeneous graph attention spatiotemporal convolution module is used to receive the initial feature vector, perform bidirectional message propagation from bottom to top and top to bottom, and update the state of each node using a gated recurrent unit to generate node temporal features. The multi-task output module is used to receive the temporal features of nodes, decode and generate model output results, and complete rainfall regression and equipment health classification. Step S4: Analyze the global state of the heterogeneous sensing path diagram using the spatiotemporal network model of the heterogeneous sensing path diagram, and output the health diagnosis results of each component of the device.

2. The rainfall monitoring method based on a heterogeneous sensor path graph spatiotemporal network according to claim 1, characterized in that, Step S1 includes the following steps: Step S11: Set the rainfall monitoring time window, and backtrack the duration of the rainfall monitoring time window from the current time as the endpoint, and extract the original pressure signal sequence, original laser distance signal sequence and temperature signal sequence collected by the pressure sensor, laser distance sensor and temperature sensor within the time period. Step S12: Perform time-series data preprocessing on the original pressure signal sequence, the original laser distance signal sequence, and the temperature signal sequence respectively to obtain multi-source monitoring data; wherein, the multi-source monitoring data includes pressure filter value, laser filter value, pressure-converted water level value, laser-converted water level value, and temperature compensation parameters.

3. The rainfall monitoring method based on a heterogeneous sensor path graph spatiotemporal network according to claim 1, characterized in that, The number of virtual nodes in step S21 is an integer value that is greater than or equal to 6 and less than or equal to 16.

4. The rainfall monitoring method based on a heterogeneous sensor path graph spatiotemporal network according to claim 1, characterized in that, The spatiotemporal feature embedding module is used to generate initial feature vectors for each node of the heterogeneous sensing path graph, including: For each node in the heterogeneous sensing path diagram, the height position of the node is divided by the total height parameter of the measuring cylinder to obtain the normalized height coordinate. The normalized height coordinate is then subjected to sine and cosine transformations, and the transformation results are spliced ​​together to form a spatial position encoding vector. For the bottom physical nodes, the pressure filter values ​​and temperature compensation parameters in the multi-source monitoring data are processed by MLP encoding, thereby mapping them into pressure time series feature vectors; For the top physical node, the laser filter values ​​in the multi-source monitoring data are processed by MLP encoding, thereby mapping them into laser temporal feature vectors; The pressure time-series feature vector and the spatial location encoding vector are concatenated to form the initial feature vector of the bottom physical node; the laser time-series feature vector and the spatial location encoding vector are concatenated to form the initial feature vector of the top physical node. For virtual intermediate nodes, a virtual water level interpolation is obtained by linear interpolation between pressure-converted water level values ​​and laser-converted water level values ​​based on the node's height position. The virtual water level interpolation is then mapped to the spatial position encoding vector as the initial feature vector of the virtual intermediate node.

5. The rainfall monitoring method based on a heterogeneous sensor path graph spatiotemporal network according to claim 4, characterized in that, The heterogeneous graph attention spatiotemporal convolution module includes the following through bidirectional message propagation: bottom-up and top-down. For each node in the heterogeneous sensing path diagram, the initial feature vector of the node is extracted and linearly transformed using preset shared weight parameters; For each pair of adjacent nodes, the features of the two nodes after linear transformation are concatenated, the concatenation result is multiplied by the preset attention parameter vector, and then processed by the preset nonlinear activation function to obtain the original attention score of the node pair. Extract the raw attention scores between a node and all its neighboring nodes, calculate the exponential function value of each raw attention score, and divide each exponential function value by the sum of all exponential function values ​​to obtain the normalized attention weights from the node to each neighboring node. The message aggregation vector of a node is obtained by multiplying the features of each adjacent node after linear transformation with the corresponding attention normalization weights and summing the results.

6. The rainfall monitoring method based on a heterogeneous sensor path graph spatiotemporal network according to claim 5, characterized in that, The heterogeneous graph attention spatiotemporal convolution module also includes, through performing bidirectional message propagation of bottom-up and top-down: The process propagates upwards layer by layer along the heterogeneous sensing path graph, starting from the bottom physical node. Each node uses the aggregated vector of messages from its adjacent nodes below as the feature for upward propagation. Starting from the top physical node, the process propagates downwards along the heterogeneous sensing path graph, with each node using the aggregated message vector from its adjacent node above as a feature for downward propagation. For each node, the upward and downward propagation features of that node are concatenated and linearly transformed using preset feature fusion parameters to output the bidirectional propagation features of the node.

7. The rainfall monitoring method based on a heterogeneous sensor path graph spatiotemporal network according to claim 6, characterized in that, The heterogeneous graph attention spatiotemporal convolution module uses gated recurrent units to update the state of each node and generate temporal features of the nodes, including: For each node, read the historical state vector of the node in the previous rainfall monitoring time window. If it is the first run, initialize the historical state vector of the node to a zero vector. The updated state vector is obtained by updating the features based on the historical state vector and bidirectional propagation features through the update gate and reset gate inside the gated loop unit. A one-dimensional convolution kernel is applied to the updated state vector in the time dimension. The convolution operation extracts the temporal evolution pattern of the updated state vector, and the node temporal features are output after convolution processing.

8. The rainfall monitoring method based on a heterogeneous sensor path graph spatiotemporal network according to claim 4, characterized in that, The multi-task output module is used to complete rainfall regression, including: The node temporal features of the bottom physical node and the top physical node are extracted from the heterogeneous sensing path diagram, and the two node temporal features are concatenated in the feature dimension to obtain a joint feature vector. The joint feature vector is input into a preset rainfall regression network, and the joint feature vector is mapped into a single scalar output through layer-by-layer dimensional compression to obtain the fused water level estimate. Obtain the diameter parameters of the rain collector and calculate the area parameters of the rain collector; The water level estimate is used as the water level height, and the rainfall estimate is calculated based on the water level height and the rain collector area parameter.

9. A rainfall monitoring system based on a heterogeneous sensor path graph spatiotemporal network, characterized in that, For performing the rainfall monitoring method based on a heterogeneous sensor path graph spatiotemporal network as described in claim 1, the rainfall monitoring system based on the heterogeneous sensor path graph spatiotemporal network includes: The multi-source data acquisition module is used to periodically acquire time-series data monitored by pressure sensors, laser rangefinders, and temperature sensors as multi-source monitoring data. The heterogeneous path diagram construction module is used to set the bottom physical node, the top physical node and multiple virtual intermediate nodes in the graduated cylinder, and establish bidirectional connection relationships between the nodes to construct a heterogeneous sensing path diagram. The rainfall inference module is used to input multi-source monitoring data and heterogeneous sensor path maps into a pre-trained heterogeneous sensor path map spatiotemporal network model. Based on the heterogeneous graph attention mechanism, graph attention information propagation and bidirectional feature fusion are performed between nodes of the heterogeneous sensor path map to obtain fused water level estimates and rainfall estimates through regression decoding. The equipment health diagnosis module is used to analyze the global state of the heterogeneous sensor path diagram through the spatiotemporal network model of the heterogeneous sensor path diagram, and output the health diagnosis results of each component of the equipment.

Images