A sensor-based method and system for predicting fluid motion data

Abstract

This invention discloses a sensor-based method and system for predicting fluid motion data, specifically including the following steps: pouring fluid into an experimental pool; setting up multiple physical sensor groups in the fluid, recording the first coordinate data of the physical sensor groups, and using the data collected by the physical sensor groups as the physical state information corresponding to the first coordinates; placing a non-powered boat in the fluid, the boat equipped with a position sensor and an acceleration sensor; generating waves in the fluid; using the sensor groups as nodes in a graph structure, obtaining the state field information of the fluid in the experimental pool based on the data collected by the sensor groups and the first coordinate data; calibrating the state field information according to the data collected by the sensors on the boat to obtain the experimental state field information; and predicting the state field information of the experimental pool at future moments based on the experimental state field information. This invention can predict the state field information of the fluid at subsequent moments, facilitating the study of changes in the fluid's motion state.

Description

Technical Field

[0001] This invention belongs to the field of physical data monitoring technology, and in particular relates to a sensor-based method and system for predicting fluid motion data. Background Technology

[0002] In fluid mechanics research, real-time and accurate monitoring of fluid physical data (such as velocity field, temperature field, pressure distribution, concentration distribution, and turbulence intensity) is crucial. However, traditional monitoring methods and monitoring systems based on conventional data processing techniques have a series of inherent defects. Existing technologies generally rely on deploying a limited number of physical sensors at discrete points within the pool. The number of sensors is limited by cost, deployment difficulty, and interference with the fluid field, resulting in the acquisition of isolated data from sparse points, failing to depict the continuous, multi-dimensional spatial distribution of the fluid physical field (such as velocity and temperature). For large-scale pools or areas with complex eddies or stratification, there are numerous monitoring blind spots. Single-point data may not accurately represent the overall state of its area, especially in areas with non-uniform flow fields or abrupt gradient changes. This leads to biases or even errors in condition judgments or control decisions based on point data. Moreover, most existing monitoring systems focus on monitoring the current state rather than predicting future states. Summary of the Invention

[0003] Purpose of the invention: In order to solve the problems existing in the prior art, the present invention provides a sensor-based method and system for predicting fluid motion data.

[0004] Technical solution: This invention discloses a sensor-based method for predicting fluid motion data, specifically including the following steps:

[0005] Pour the fluid into the experimental tank;

[0006] Multiple physical sensor groups are set up in the fluid, and the first coordinate data of the physical sensor groups are recorded. The data collected by the physical sensor groups is used as the real physical state information corresponding to the first coordinate data.

[0007] A non-powered boat, equipped with position and acceleration sensors, is placed in a fluid.

[0008] Creating waves in a fluid;

[0009] Using the physical sensor array as nodes in the graph structure, the state field information of the fluid in the experimental pool is obtained based on the data collected by the physical sensor array and the first coordinate data.

[0010] Based on the data collected by the position and acceleration sensors on board the ship, the state field information is calibrated to obtain the experimental state field information;

[0011] Based on the experimental state field information, the state field information of the experimental pool at a future time is predicted.

[0012] Further, carriers with buoyancy are arranged at multiple positions in the fluid, a physical sensor group is arranged on a corresponding carrier, a lower surface of the carrier is connected to the bottom of the experimental pool through a rope, and a plane coordinate of a connection position of the rope and the bottom of the experimental pool is taken as first coordinate data, and a length of the rope is longer than a vertical distance between the lower surface of the carrier and the bottom of the experimental pool.

[0013] Further, the state field information of the fluid in the experimental pool is specifically obtained by:

[0014] According to the first coordinate data and the data collected by the physical sensor group, a first state vector corresponding to the physical sensor group is obtained.

[0015] The first state vector is processed by using a multi-layer perception layer of a graph neural network model to obtain a first input vector.

[0016] An attention mechanism of the graph neural network model is used to process the first input vector of each node to obtain a connection weight between each node.

[0017] A message passing module of the graph neural network model is used to process the first input vector and the connection weight to obtain a first output vector of the node.

[0018] All the first output vectors are combined into a first output tensor according to the first coordinate data of the physical sensor group, and the first output tensor has a dimension of HxWxC, H is the number of physical sensor groups in each row, W is the number of physical sensor groups in each column, and C is the dimension of the first output vector.

[0019] A multi-layer deconvolution of the graph neural network model is used to perform deconvolution processing on the first output tensor to obtain the state field information.

[0020] Further, the state field information is calibrated based on a sensor carried on the ship to obtain the experimental state field information, and specifically:

[0021] According to the state field information, an acceleration at a position of the ship is estimated.

[0022] An error between the estimated acceleration of the ship and a real acceleration of the ship is calculated.

[0023] An error between a real physical state information at the first coordinate and a physical state information at the first coordinate in the state field information is calculated.

[0024] A first loss function is constructed based on the two errors mentioned above. The graph neural network model is then calibrated based on this first loss function. The first coordinate data and the data collected by the physical sensor group are then input back into the calibrated graph neural network model to obtain the experimental state field information.

[0025] Furthermore, the acceleration at the ship's location is estimated based on the state field information, specifically as follows:

[0026] Estimate the ship's acceleration based on the acceleration at multiple locations in the state field information and the distances between these locations and the ship.

[0027] Furthermore, the first loss function The expression is:

[0028] ;

[0029] in, Let be a vector composed of the ship's actual acceleration data at time t. For the state field information at time t A vector composed of acceleration data at coordinates. For the t-th time Distance data between coordinates and ship position data, , Let be a vector composed of the physical state information at the k-th first coordinate data point in the state field information at time t. This is a vector composed of the actual physical state information at the k-th first coordinate data point at time t. and The preset weights are N, where N is the number of time points, and t, N, k, H, W, i, j All are positive integers. The number of coordinates in each column of the state field information. This represents the number of coordinates in each row of the state field information.

[0030] Further, based on the experimental state field information at multiple time points, the state field information at a future time point is predicted, specifically: a prediction neural network model is set, including F convolution layers connected in turn, a state prediction model, and F deconvolution layers connected in turn, the experimental state field information at the tth time point is input into the prediction neural network model at the tth state, the F convolution layers output the features at the tth time point; the features at the tth time point and the hidden state vector at the t-1th time point are input into the state prediction model to obtain the hidden state vector at the tth time point, the hidden state vector at the tth time point is input into the F deconvolution layers, and the fth deconvolution layer is connected with the fth convolution layer in a skip connection manner; f=1, 2, …, F; the F deconvolution layers output the predicted state field information at the t+1th time point; the prediction neural network model at the t+1th state is obtained by training the prediction neural network model at the tth state according to the experimental state field information at the t+1th time point and the predicted state field information; the experimental state field information at the t+1th time point and the prediction neural network model at the t+1th state are used to predict the predicted state field information at the t+2th time point, until the experimental state field information at all time points is traversed, and the training of the prediction neural network model is completed.

[0031] Further, the expression of the loss function when training the prediction neural network model is

[0032] ;

[0033] Wherein, is the physical state information of each coordinate in the experimental state field information at the t+1th time point, is the physical state information of each coordinate in the predicted state field information at the t+1th time point, is the number of coordinates in each row in the experimental state field information, is the number of coordinates in each column in the experimental state field information, and N is the number of time points.

[0034] Further, the method further comprises evaluating the risk coefficient of the experimental ship by predicting the acceleration data of a specific position in the experimental pool within a future period of time.

[0035] A sensor-based fluid motion data prediction system, comprising

[0036] A setting module is configured to set a plurality of physical sensor groups in the fluid and record first coordinate data of the physical sensor groups;

[0037] A ship module comprises a ship and position sensors and acceleration sensors carried on the ship;

[0038] A wave module is configured to generate waves in the experimental pool.​

[0039] The acquisition module is used to acquire data collected by various physical sensor groups at multiple times, as well as position data collected by the ship's position sensors and acceleration data collected by the ship's acceleration sensors.

[0040] The state field information module is used to obtain the state field information of the fluid in the experimental pool based on the data collected by the physical sensor group and the first coordinate data.

[0041] The experimental state field information module is used to calibrate the state field information based on the data collected by the sensors on board the ship, and obtain the experimental state field information.

[0042] The predicted state field information module is used to determine the predicted state field information for subsequent times based on the experimental state field information at multiple times.

[0043] Beneficial Effects: This invention, based on a limited number of physical sensor groups at multiple locations and sensors in an experimental vessel capable of free movement within the experimental pool, can simulate state field information at more locations within the pool. This information can be calibrated, allowing for the acquisition of fluid motion characteristics at more locations, facilitating the study of fluid motion properties. Furthermore, it can predict state field information at subsequent moments, enabling the study of changes in fluid motion states and improving the comprehensiveness and accuracy of the data. When solving for the state field information, graph structures and graph neural network models can be used to describe the interactions between the locations of each physical sensor group. This allows for a more accurate description of the fluid motion patterns at the location of that physical sensor group and serves as a prediction benchmark, more accurately predicting physical data at more locations surrounding that location. Thus, by using physical data collected from a smaller number of physical sensor groups, the invention predicts physical state information at more locations within the experimental pool, obtaining state field information for a more comprehensive understanding of the fluid motion patterns at various locations within the experimental pool. When training the graph neural network model, the estimated acceleration data can be compared with data collected from the experimental vessel at random locations, and the predicted data can be compared with data collected from a group of physical sensors at fixed locations. This comparison yields a first loss function. Through multiple training iterations, the overall accuracy of the graph neural network model is improved based on the uncertainty of random locations. This ensures that the error is reduced even when the experimental vessel moves to any position. The error at the fixed location is used as an adjustment benchmark to further reduce the overall error, thereby improving the performance and overall prediction accuracy of the graph neural network model and enhancing the accuracy of the experimental state field information. When acquiring the predicted state field information, a recurrent neural network model can be embedded in the U-NET model to solve the prediction problem of multidimensional tensors. Before prediction processing, the model can be trained using experimental state field information from multiple time points to improve model accuracy and thus prediction accuracy. Furthermore, during training, the weights of the second loss function are set based on the distance between the training time and the current time, enhancing the training's relevance and further improving prediction accuracy. Attached Figure Description

[0044] Figure 1 This is a flowchart of the present invention;

[0045] Figure 2 This is a schematic diagram of the process for obtaining state field information according to the present invention;

[0046] Figure 3 This is a schematic diagram of the process for obtaining predicted state field information according to the present invention;

[0047] Figure 4 This is a system block diagram of the present invention. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0049] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0050] Figure 1 An exemplary flowchart illustrates a physical experiment data monitoring method based on physical sensors according to an embodiment of the present invention, the method comprising:

[0051] Step S1: Set up physical sensor groups at multiple locations in the experimental pool and record the first coordinate data of the physical sensor groups;

[0052] Step S2: Place the experimental vessel in the experimental pool. The experimental vessel is equipped with a position sensor and an acceleration sensor. The vessel is a non-powered vessel.

[0053] Step S3: After the experiment begins, create waves in the experimental pool;

[0054] Step S4: At multiple times, acquire physical sensing data from each physical sensor group, as well as position data collected by the ship's position sensor and acceleration data collected by the acceleration sensor.

[0055] Step S5: Based on the first coordinate data and physical sensing data, obtain the state field information of the experimental pool at multiple moments;

[0056] Step S6: Based on the position data and acceleration data, calibrate the state field information to obtain experimental state field information at multiple times;

[0057] Step S7: Determine the predicted state field information for subsequent time steps based on the experimental state field information at multiple time points;

[0058] Step S8: Determine the risk monitoring coefficient of the experimental vessel based on the experimental state field information and the predicted state field information.

[0059] According to an embodiment of the present invention, the physical sensor-based physical experiment data monitoring method can simulate state field information including physical state information of more locations in the experimental pool based on a limited number of physical sensor groups at multiple locations and sensors in an experimental vessel that can move freely in the experimental pool. The state field information can be calibrated, thereby acquiring the motion characteristics of the fluid at more locations to facilitate the study of the fluid motion characteristics. Furthermore, it can predict the state field information at subsequent times to facilitate the study of changes in the fluid motion state, thereby improving the comprehensiveness and accuracy of the data.

[0060] According to one embodiment of the present invention, in step S1, a physical sensor group can be set at multiple locations in the experimental pool. The experimental pool has a rectangular cross-section, and the physical sensor group can be evenly distributed in the experimental pool and set on a buoyant carrier so that it can float on the water surface of the experimental pool. The lower surface of the carrier is connected to the bottom of the experimental pool by a rope. The planar coordinates of the connection position between the rope and the bottom of the experimental pool are the first coordinate data. In a calm state, the length of the rope is slightly longer than the distance between the lower surface of the carrier and the bottom of the experimental pool. For example, the rope length is 1.1 times the distance between the lower surface of the carrier and the bottom of the experimental pool. When waves are generated on the water surface, the carrier and the physical sensor group have a certain degree of freedom of movement, so that they can move in a small range with the waves to obtain physical data during the movement. The physical sensor group includes accelerometers, water pressure sensors, etc., which are not limited in the present invention.

[0061] According to one embodiment of the present invention, in step S2, the experimental vessel may be equipped with a position sensor and an acceleration sensor. The experimental vessel is unpowered and can move with the waves in the experimental pool, but cannot move on its own. It can be used to determine the influence of the fluid on the vessel's motion and can measure various physical data, such as position data and acceleration data, at non-fixed locations. The shape and weight of the experimental vessel may be the same as the aforementioned carrier; that is, when impacted by the same force of water flow, the acceleration of the experimental vessel and the aforementioned carrier is the same.

[0062] According to one embodiment of the present invention, in step S3, after the experiment begins, waves can be generated in the experimental pool using a wave generator or vibration device. Waves can be generated continuously during the experiment, or they can be generated at the beginning of the experiment and then the wave-generating device can be turned off during the experiment to allow the waves to gradually calm down; the present invention does not limit this.

[0063] According to one embodiment of the present invention, in step S4, physical sensing data, such as acceleration data and water pressure data, can be collected at multiple times by a group of physical sensors, and position data can be collected by a position sensor mounted on the experimental vessel, and acceleration data can be collected by an acceleration sensor mounted on the experimental vessel. The time interval between adjacent times can be 1 minute, 3 minutes, etc., and the present invention does not limit this.

[0064] Figure 2 A schematic diagram illustrating the acquisition of state field information according to an embodiment of the present invention is shown as an example.

[0065] According to one embodiment of the present invention, in step S5, physical data at more locations can be predicted by physical sensing data collected at the locations of a smaller number of physical sensor groups, thereby obtaining state field information, which facilitates the analysis of the fluid motion state at each location.

[0066] According to an embodiment of the present invention, in step S5, obtaining state field information of the experimental pool at multiple moments based on the first coordinate data and physical sensing data includes: using the physical sensing data collected by the sensor group as the real physical state information of the corresponding physical sensor group; obtaining a first state vector for each physical sensor group based on the first coordinate data and physical sensing data; processing the first state vector at the first multilayer perceptual network layer of the graph neural network to obtain the first input vector of the node corresponding to each physical sensor group; processing the first input vector of each node through the attention mechanism of the graph neural network to obtain the connection weights between each node; and using the graph neural network model... The message passing module processes the first input vector and the connection weights to obtain the first output vector of the node corresponding to each physical sensor group; it then forms a first output tensor based on the first coordinate data of the physical sensor group, wherein the dimension of the first output tensor is H×W×C, where H is the number of physical sensor groups in each row, W is the number of physical sensor groups in each column, and C is the dimension of the first output vector; the first output tensor is then subjected to multi-layer deconvolution processing to obtain state field information, wherein the state field information includes physical state information at multiple coordinates in the experimental pool, the multiple coordinates including the first coordinate and other coordinates, and the dimension of the state field information is H×W×C. , The number of coordinates per row. This represents the number of coordinates in each column.

[0067] According to an embodiment of the present invention, as described above, the physical sensing data may include a variety of data, such as acceleration data and water pressure data, and the acceleration data and water pressure data in the three coordinate axes, as well as the first coordinate data (two-dimensional plane coordinates), can be combined to form a first state vector.

[0068] According to one embodiment of the present invention, since the physical sensor groups are all located within the experimental pool, these locations influence each other to a certain extent. For example, waves at one location can propagate to other locations, thus affecting the data collected by the physical sensor groups at those locations. Therefore, for physical sensor groups that influence each other, a graph structure can be used to describe the mutual influence. Each physical sensor group can be treated as a node in the graph structure. Furthermore, the first state vector can be increased in dimensionality through the first multilayer perceptual network layer of the graph neural network (e.g., including multiple fully connected layers and activation layers) to obtain the first input vector of each node.

[0069] According to one embodiment of the present invention, although physical sensor groups may influence each other, it is difficult to accurately determine the influence of the location of one sensor group on the location of another. Therefore, the connection weights between two nodes can be determined through the attention mechanism of a graph neural network model, which can be used to describe the mutual influence of the data detected by the two physical sensor groups. In the example, the first input vectors of the two nodes can be multiplied by the weight matrix respectively to obtain processed vectors, and the processed vectors can be concatenated to obtain a concatenated vector. Then, the concatenated vector is processed through a multilayer perceptron layer (e.g., multiple fully connected layers and activation layers, and the parameters of the multilayer perceptron layer can be obtained through training) to obtain the connection weights between the two nodes.

[0070] According to one embodiment of the present invention, after obtaining the connection weights, the first input vector of a node can be weighted and summed with the first input vectors of other nodes based on the connection weights. The summed vector is then processed by the message passing module of the graph neural network model to obtain the first output vector of the node. The first output vector can not only reflect the physical sensing data collected by the physical sensor group corresponding to the node, but also reflect the influence of other locations on the location of the physical sensor group. This allows for a more accurate description of the fluid motion law at the location of the physical sensor group, and can be used as a prediction benchmark in subsequent processes to more accurately predict the physical data (e.g., acceleration data) at more locations around the location.

[0071] According to one embodiment of the present invention, the first output vector can be combined according to the first coordinate data of the physical sensor group to obtain a first output tensor. The first output tensor has a dimension of H×W×C. For example, based on the first coordinate data, the physical sensor groups in each row and column can be determined, and the dimension of the first output vector of the node corresponding to each physical sensor group is C. Therefore, the data of multiple channels in the h-th row and w-th column of the first output tensor is the data of multiple dimensions of the first output vector of the physical sensor group in the h-th row and w-th column. The data of the c-th channel in the h-th row and w-th column of the first output tensor is the data of the c-th dimension of the physical sensor group in the h-th row and w-th column. The data of each channel in the first output tensor is an H×W feature map. For example, the data of the c-th channel in the first output tensor is a feature map of size H×W composed of the c-th dimension data of the first output vectors of the nodes corresponding to H×W physical sensor groups.

[0072] According to one embodiment of the present invention, after obtaining the first output tensor, multiple deconvolution layers can be used to obtain state field information. The state field information includes physical state information at multiple coordinates within the experimental pool, and the dimension of the state field information is... , >H, >W, thus obtaining more physical state information at various locations. In the state field information, the data from multiple channels in the h-th row and w-th column represent various physical data (e.g., acceleration data) at the h-th row and w-th column location, and the data from the c-th channel in the h-th row and w-th column represent the c-th type of physical data at the h-th row and w-th column location. Furthermore, The positions are evenly distributed in the experimental pool, and these positions include those corresponding to the first coordinate data. Furthermore, by performing interpolation on the first coordinate data, we can obtain... By obtaining the coordinates of a location, the physical state information at multiple coordinates can be determined, that is, the state field information can be obtained. In other words, the physical state information at more locations in the experimental pool can be predicted based on the physical data collected by a smaller number of physical sensor groups, thereby obtaining the state field information.

[0073] In this way, the interaction between the locations of each physical sensor group can be described through graph structures and graph neural network models. This allows for a more accurate description of the fluid motion patterns at the location of the physical sensor group and serves as a prediction benchmark to more accurately predict the physical data at more locations around that location. Thus, physical data collected by a smaller number of physical sensor groups can be used to predict the physical state information at more locations in the experimental pool, thereby obtaining state field information and acquiring a more comprehensive understanding of the fluid motion patterns at various locations in the experimental pool.

[0074] According to an embodiment of the present invention, in step S6, the above state field information is predicted information and may contain certain errors. In order to reduce the errors, the graph neural network model can be trained by collecting acceleration data and position data from an experimental vessel that is freely floating in the experimental pool, thereby calibrating the state field information obtained by the graph neural network model, improving the accuracy of the graph neural network model, and reducing the errors.

[0075] According to one embodiment of the present invention, calibrating the state field information based on position data and acceleration data to obtain experimental state field information at multiple times includes: acquiring distance data between multiple coordinates of the state field information in the experimental pool and the ship's position data; determining acceleration data at multiple coordinates based on physical state information at multiple coordinates; obtaining physical state information at a first coordinate data point based on the state field information; obtaining a first loss function for a graph neural network model based on the physical state information, physical sensing data at the first coordinate data point, the distance data, the ship's actual acceleration data, and the acceleration data at multiple coordinates; training the graph neural network model based on the first loss function to obtain a trained graph neural network model; and processing the first coordinate data and physical sensing data based on the trained graph neural network model to obtain experimental state field information at multiple times.

[0076] According to one embodiment of the present invention, at a given moment, state field information, i.e., physical state information at multiple coordinates, can be obtained. The acceleration at the location of the experimental vessel can be estimated using the physical state information at multiple coordinates. Since there are mutual influences between the various locations, with farther locations having less influence and closer locations having greater influence, the distance data between the multiple coordinates in the experimental pool and the position data of the experimental vessel can be determined. Based on the distance data and the physical state information at each coordinate, the acceleration at the location of the experimental vessel can be estimated, and this estimated information can be compared with the data actually collected by the experimental vessel to obtain the error between the estimated acceleration and the measured data. Further, as described above, the locations corresponding to the multiple coordinates include the locations corresponding to the first coordinate data, and at the locations corresponding to the first coordinate data, there is measured data collected by the physical sensor group. Therefore, the physical state information at the location corresponding to the first coordinate data in the state field information can be compared with the measured data collected by the physical sensor group to obtain the error between the two. In this way, the errors at random positions can be obtained through the experimental vessel's free-floating motion. These errors, along with those at the first coordinate data, together constitute the first loss function, which is then used for training. Through multiple training iterations, the errors at different random positions and at multiple fixed positions are minimized, thereby improving the accuracy of the graph neural network model and reducing the overall error in predicting multiple positions. Furthermore, the error at random positions is estimated based on the physical state information of the predictions at multiple positions, allowing for simultaneous training of the prediction results at multiple positions during the training process, further reducing the overall error.

[0077] According to an embodiment of the present invention, the first loss function The expression is as follows:

[0078] (1);

[0079] in, Let be a vector consisting of the ship's acceleration data at time t. For the state field information at time t A vector composed of acceleration data at coordinates. For the t-th time Distance data between coordinates and the ship's position data. , This is a vector composed of the physical state information at the k-th first coordinate data point at time t. It is a vector composed of physical sensing data at the k-th first coordinate data point at time t (that is, a vector composed of the actual physical state information at the k-th first coordinate data point). and The weights are preset, N is the number of time points, t≤N, k≤ , i≤ j≤ , and t, N, k, H, W, i, j All are positive integers.

[0080] According to an embodiment of the present invention, in formula (1), and These are all vectors composed of acceleration data, for example, a three-dimensional vector composed of acceleration along the x-axis, y-axis, and z-axis. To estimate the acceleration data at the location of the experimental vessel using acceleration data at various coordinates, the acceleration data at locations closer to the experimental vessel's location are more similar to the acceleration data at the experimental vessel's location. Therefore, a weighted average of the acceleration data at each location can be used to obtain an estimate of the acceleration data at the experimental vessel's location, where the weights are... This indicates that the closer the distance, the closer the acceleration data at that location is to the acceleration data at the location of the experimental vessel, and the greater the weight. This is the estimated value obtained by weighting the acceleration data at each location using this weight. The error between this estimated value and the actual acceleration data measured by the experimental vessel is... .

[0081] According to one embodiment of the present invention, This refers to the error between the predicted physical state information at the first coordinate data point and the measured physical sensor data. Therefore, This is the average value of this type of error. This involves a weighted summation of the errors at random positions and multiple fixed first coordinate data points corresponding to the freely floating experimental vessel, yielding the acceleration estimation error at time t. Summing the acceleration estimation errors across multiple time points yields the first loss function of the graph neural network model. Training using this first loss function minimizes the errors at different random positions and multiple fixed positions during multiple training iterations, thereby improving the accuracy of the graph neural network model and reducing the overall prediction error.

[0082] According to an embodiment of the present invention, the graph neural network model and the deconvolution layer can be trained using the gradient descent method and the first loss function described above to obtain the trained graph neural network model. The trained graph neural network model can then be used to process the first coordinate data and physical sensing data again to obtain experimental state field information of the experimental pool at multiple times. The accuracy of the experimental state field information is higher than that of the state field information.

[0083] In this way, the data collected by the experimental vessel at random locations is compared with the estimated acceleration data, and the data collected by the physical sensor group at fixed locations is compared with the predicted data to obtain the first loss function. Thus, in multiple training sessions, based on the uncertainty of random locations, the overall accuracy of the graph neural network model is improved, so that the error can be reduced when the experimental vessel moves to any location. The error corresponding to the fixed location is used as the adjustment benchmark to further reduce the overall error, thereby improving the performance of the graph neural network model and the overall prediction accuracy, and improving the accuracy of the experimental state field information.

[0084] According to one embodiment of the present invention, in step S7, the state field information at subsequent time steps can be predicted by a state prediction model.

[0085] Figure 3 An exemplary schematic diagram of obtaining predicted state field information is shown according to an embodiment of the present invention.

[0086] According to one embodiment of the present invention, determining the predicted state field information for subsequent time steps based on the experimental state field information at multiple time steps includes: training a state prediction model, multiple layers of convolutional layers, and multiple layers of deconvolutional layers based on the experimental state field information at multiple time steps to obtain multiple layers of convolutional layers for the Nth state, multiple layers of deconvolutional layers for the Nth state, and a state prediction model for the Nth state, where N is the number of time steps; processing the predicted state field information for the N+s time step through the multiple layers of convolutional layers for the Nth state to obtain the predicted state field information for the N+s time step. The predicted state vector is obtained, where s is an integer (when s=0, the predicted state field information at time N+s is the experimental state field information at time N); the predicted state vector at time N+s and the hidden state vector at time N+s-1 are input into the state prediction model of the Nth state for processing to obtain the hidden state vector at time N+s; the hidden state vector at time N+s and the feature map output by each layer in the multi-layer convolution process are processed through multiple layers of deconvolution layers in the Nth state to obtain the predicted state field information at time N+s+1. In the example, the convolutional and deconvolutional layers can adopt the U-NET model structure, and the state prediction model can be a recurrent neural network model. An LSTM model can be embedded between the convolutional and deconvolutional layers of the U-NET model as the state prediction model. Thus, the predicted state field information can be processed based on the convolutional layers of the U-NET model to obtain the predicted state vector. The predicted state vector is then processed by the LSTM model to predict the future state, obtaining the predicted state vector at the future time. Finally, the predicted state vector at the future time is processed by the deconvolutional layer to obtain the predicted state field information at the future time.

[0087] According to one embodiment of the present invention, before using the state prediction model, multiple layers of convolutional layers, and multiple layers of deconvolutional layers to predict subsequent time steps, the experimental state field information of multiple past time steps can be used to train the state prediction model, multiple layers of convolutional layers, and multiple layers of deconvolutional layers to improve prediction accuracy. Furthermore, the experimental state field information of multiple time steps can be used to update and train the state prediction model, multiple layers of convolutional layers, and multiple layers of deconvolutional layers during each experiment, thereby improving model accuracy.

[0088] According to an embodiment of the present invention, based on the experimental state field information at multiple time points, a state prediction model, multiple levels of convolutional layers, and multiple levels of deconvolutional layers are trained to obtain multiple levels of convolutional layers for the Nth state, multiple levels of deconvolutional layers for the Nth state, and a state prediction model for the Nth state. This includes: processing the experimental state field information at time t using the multiple levels of convolutional layers for the tth state to obtain the experimental state vector at time t; inputting the experimental state vector at time t and the hidden state vector at time t-1 into the state prediction model for the tth state for processing to obtain the hidden state vector at time t, where t is a positive integer, and at t=1, the hidden state vector at time t-1 is a zero vector; and so on. The hidden state vector at time t and the feature maps output by each layer in the multi-layer convolution process are processed through multiple deconvolutional layers at the t-th state to obtain the predicted state field information at time t+1. Based on the experimental state field information and the predicted state field information at time t+1, the state prediction model, the multiple convolutional layers, and the multiple deconvolutional layers at the t-th state are trained to obtain the state prediction model, the multiple convolutional layers, and the multiple deconvolutional layers at the t+1-th state. When t+1=N, the multiple convolutional layers, the multiple deconvolutional layers, and the state prediction model of the N-th state are obtained.

[0089] According to one embodiment of the present invention, after each training, the states of multiple convolutional layers, multiple deconvolutional layers, and the state prediction model can be updated. That is, after training using the experimental state field information at time t, the states of multiple convolutional layers, multiple deconvolutional layers, and the state prediction model can be updated from the t-th state to the (t+1)-th state.

[0090] According to one embodiment of the present invention, the experimental state field information at time t can be processed by multiple convolutional layers of the t-th state, continuously downsampling, increasing the number of channels, and flattening the feature map output by the last convolutional layer into a vector. Then, the vectors of the feature maps of multiple channels are concatenated to obtain the experimental state vector at time t. The experimental state vector is a high-dimensional vector, and the data of each dimension of the experimental state vector can be regarded as the data of one channel. The number of dimensions of the experimental state vector is greater than the number of channels of the experimental state field information.

[0091] According to one embodiment of the present invention, the state prediction model can be a recurrent neural network model, such as an LSTM model, which can use the input information at the current time and the output information at the previous time to obtain the output information at the current time. That is, the state prediction model can process the experimental state vector at time t and the hidden state vector at time t-1 to obtain the hidden state vector at time t.

[0092] According to one embodiment of the present invention, the hidden state vector at time t can be used to predict the training state field information at time t+1. The hidden state vector at time t can be processed by multiple layers of deconvolution layers. The hidden state vector can be converted into a feature map with multiple channels in the reverse manner of the above flattening process, or the hidden state vector can be processed by a fully connected layer to obtain a feature map with multiple channels. Then, deconvolution processing is performed on the feature maps with multiple channels. During the deconvolution process, the feature map of the same size output by the convolution layer can be concatenated with the input feature map of the deconvolution layer to obtain a feature map with more channels, and then deconvolution processing is performed to improve the accuracy of the deconvolution processing. After processing by multiple layers of deconvolution layers, the predicted state field information at time t+1 can be obtained.

[0093] According to one embodiment of the present invention, the experimental state field information and the predicted state field information at time t+1 can be used to train a state prediction model for state t, multiple layers of convolutional layers for state t, and multiple layers of deconvolutional layers for state t. This training yields the state prediction model for state t+1, the multiple layers of convolutional layers for state t+1, and the multiple layers of deconvolutional layers for state t+1. The experimental state field information at time t+1 can then be processed using the state prediction model for state t+1, the multiple layers of convolutional layers for state t+1, and the multiple layers of deconvolutional layers for state t+1 to obtain the predicted state field information at time t+2. Training is performed again to obtain the state prediction model for the (t+2)th state, multiple layers of convolutional layers for the (t+2)th state, and multiple layers of deconvolutional layers for the (t+2)th state, and so on. This process can be repeated multiple times to continuously improve the model's accuracy and the accuracy of predicting the training state field information. Training stops at t+1=N, and the multiple layers of convolutional layers for the Nth state, multiple layers of deconvolutional layers for the Nth state, and the state prediction model for the Nth state are obtained. These are used to obtain the predicted state field information for subsequent time steps. In subsequent prediction processes, training does not continue, that is, the states of the multiple layers of convolutional layers, multiple layers of deconvolutional layers, and the state prediction model are not updated.

[0094] According to one embodiment of the present invention, based on the experimental state field information and the predicted state field information at time t+1, the state prediction model of state t, the multiple layers of convolutional layers of state t, and the multiple layers of deconvolutional layers of state t are trained to obtain the state prediction model of state t+1, the multiple layers of convolutional layers of state t+1, and the multiple layers of deconvolutional layers of state t+1, including: obtaining the second loss function of the state prediction model of state t, the multiple layers of convolutional layers of state t, and the multiple layers of deconvolutional layers of state t according to formula (2).

[0095] (2);

[0096] in, This represents the physical state information of each coordinate in the experimental state field information at time t+1. This represents the predicted state information for each coordinate in the predicted state field information at time t+1. This represents the number of coordinates in each row of the experimental state field information. The number of coordinates in each column of the experimental state field information, i ≤ j≤ ,t≤N-1, and i, j , t, and N are all positive integers; using the second loss function, the state prediction model of the t-th state, the multiple layers of convolutional layers of the t-th state, and the multiple layers of deconvolutional layers of the t-th state are trained to obtain the state prediction model of the (t+1)-th state, the multiple layers of convolutional layers of the (t+1)-th state, and the multiple layers of deconvolutional layers of the (t+1)-th state.

[0097] According to an embodiment of the present invention, in formula (2), and All information is in vector form. The magnitude of the difference vector between the two can be used as the error between the physical state information and the training state information for each coordinate. This represents the error between the experimental state field information and the training state field information at time t+1. This represents the weight at time t. The closer time t is to time N, the closer time t is to the current time, and the greater its impact on subsequent times. Therefore, the prediction accuracy at time t is more important and can be assigned a higher weight.

[0098] According to one embodiment of the present invention, the second loss function can be backpropagated, and the parameters of the state prediction model of the t-th state, the multiple layers of convolutional layers of the t-th state, and the multiple layers of deconvolutional layers of the t-th state can be adjusted by gradient descent to obtain the state prediction model of the (t+1)-th state, the multiple layers of convolutional layers of the (t+1)-th state, and the multiple layers of deconvolutional layers of the (t+1)-th state.

[0099] According to one embodiment of the present invention, when performing prediction processing of the predicted state field information at subsequent time steps, for example, the Nth time step is the current time step, the data obtained from the time steps before the current time step is used for training to obtain multiple layers of convolutional layers, a state prediction model for the Nth state, and multiple layers of deconvolutional layers for the Nth state, and these are used to predict the data at subsequent time steps. That is, the multiple layers of convolutional layers, a state prediction model for the Nth state, and multiple layers of deconvolutional layers for the Nth state are used each time, and the parameters are no longer updated. The predicted state field information at time N+s can be input into multiple layers of convolutional layers in the Nth state to obtain the predicted state vector at time N+s. The predicted state vector at time N+s and the hidden state vector at time N+s-1 are then input into the state prediction model of the Nth state for processing to obtain the hidden state vector at time N+s. Finally, the hidden state vector at time N+s is input into multiple layers of deconvolutional layers in the Nth state to obtain the predicted state field information at time N+s+1. Subsequently, the predicted state field information at time N+s+1 can be input into multiple layers of convolutional layers in the Nth state to obtain the predicted state vector at time N+s+1. The predicted state vector at time N+s+1 and the hidden state vector at time N+s are then input into the state prediction model of the Nth state for processing to obtain the hidden state vector at time N+s+1. The hidden state vector at time N+s+1 is then input into multiple layers of deconvolutional layers in the Nth state to obtain the predicted state field information at time N+s+2, and so on, to obtain the predicted state field information for subsequent time periods.

[0100] In this way, a recurrent neural network model can be embedded in the U-NET model to solve the prediction problem of multidimensional tensors. The model can be trained with experimental state field information at multiple time steps before prediction processing to improve model accuracy and thus improve prediction accuracy. Furthermore, during training, the weight of the second loss function is set according to the distance between the training time and the current time step to improve the targeting of training and further improve prediction accuracy.

[0101] According to an embodiment of the present invention, in step S8, the risk monitoring coefficient of the experimental vessel can be determined based on the experimental state field information and the predicted state field information. The risk monitoring coefficient of the experimental vessel is the danger coefficient of the experimental pool. If the waves in the experimental pool are large, or if the newly generated waves are superimposed with the waves rebounding after the previously generated waves hit the pool wall, the experimental vessel may face risks such as capsizing. Furthermore, the larger the waves, the more uneven the acceleration at different locations. Therefore, the acceleration data of the three axes at each location can be obtained from the experimental state field information or the predicted state field information at a certain moment, and the vector sum of the acceleration data of each axis can be solved. Then, the magnitude of the vector sum can be calculated. Further, the standard deviation of the magnitude of the vector sum at each location can be calculated as the risk monitoring coefficient at that moment. Then, the maximum value of the risk monitoring coefficient at multiple moments can be obtained as the risk monitoring coefficient of the experimental vessel, which is used to describe the magnitude of the acceleration generated by the waves in the experiment, and can also describe the risks that the experimental vessel may encounter.

[0102] According to embodiments of the present invention, a physical sensor-based physical experiment data monitoring method can simulate state field information at more locations within an experimental pool using a limited number of physical sensor groups at multiple locations and sensors on an experimental vessel capable of free movement within the experimental pool. This method can be calibrated to acquire fluid motion characteristics at more locations, facilitating the study of fluid motion properties. Furthermore, it can predict state field information at subsequent moments to study changes in fluid motion states, improving the comprehensiveness and accuracy of the data. When solving for the state field information, graph structures and graph neural network models can be used to describe the interactions between the locations of each physical sensor group. This allows for a more accurate description of the fluid motion patterns at the location of that physical sensor group and serves as a prediction benchmark to more accurately predict physical data at more locations surrounding that location. Thus, by using physical data collected from a smaller number of physical sensor groups, the method can predict physical state information at more locations within the experimental pool, obtaining state field information and comprehensively acquiring the fluid motion patterns at various locations within the experimental pool. When training the graph neural network model, the estimated acceleration data can be compared with data collected from the experimental vessel at random locations, and the predicted data can be compared with data collected from a group of physical sensors at fixed locations. This comparison yields a first loss function. Through multiple training iterations, the overall accuracy of the graph neural network model is improved based on the uncertainty of random locations. This ensures that the error is reduced even when the experimental vessel moves to any position. The error at the fixed location is used as an adjustment benchmark to further reduce the overall error, thereby improving the performance and overall prediction accuracy of the graph neural network model and enhancing the accuracy of the experimental state field information. When acquiring the predicted state field information, a recurrent neural network model can be embedded in the U-NET model to solve the prediction problem of multidimensional tensors. Before prediction processing, the model can be trained using experimental state field information from multiple time points to improve model accuracy and thus prediction accuracy. Furthermore, during training, the weights of the second loss function are set based on the distance between the training time and the current time, enhancing the training's relevance and further improving prediction accuracy.

[0103] Figure 4 An exemplary block diagram of a physical experiment data monitoring system based on physical sensors according to an embodiment of the present invention is shown, the system comprising:

[0104] The setup module is used to set up multiple physical sensor groups in the experimental pool and record the first coordinate data of the physical sensor groups.

[0105] A placement module for placing an experimental vessel in an experimental pool, wherein the experimental vessel is equipped with a position sensor and an acceleration sensor;

[0106] The wave module is used to generate waves in the experimental pool after the experiment has started;

[0107] The acquisition module is used to acquire physical sensing data from each physical sensor group, as well as position data collected by the position sensor and acceleration data collected by the accelerometer at multiple times.

[0108] The state field information module is used to obtain the state field information of the experimental pool at multiple moments based on the first coordinate data and physical sensing data.

[0109] The experimental state field information module is used to calibrate the state field information based on position data and acceleration data to obtain experimental state field information at multiple times.

[0110] The predicted state field information module is used to determine the predicted state field information for subsequent times based on the experimental state field information at multiple times.

[0111] One embodiment of the present invention further includes a risk monitoring coefficient module, used to determine the risk monitoring coefficient of the experimental vessel based on the experimental state field information and the predicted state field information.

[0112] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0113] Those skilled in the art should understand that the embodiments of the present invention described above and shown in the accompanying drawings are merely examples and do not limit the present invention. The objectives of the present invention have been fully and effectively achieved. The functions and structural principles of the present invention have been demonstrated and explained in the embodiments, and any variations or modifications may be made to the implementation of the present invention without departing from the stated principles.

[0114] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A sensor-based method for predicting fluid motion data, characterized in that, Specifically, the steps include the following: Pour the fluid into the experimental tank; Multiple physical sensor groups are set up in the fluid, and the first coordinate data of the physical sensor groups are recorded. The data collected by the physical sensor groups is used as the real physical state information corresponding to the first coordinate data. A non-powered boat, equipped with position and acceleration sensors, is placed in a fluid. Creating waves in a fluid; Using the physical sensor array as nodes in the graph structure, the state field information of the fluid in the experimental pool is obtained based on the data collected by the physical sensor array and the first coordinate data. Based on the data collected by the position and acceleration sensors on board the ship, the state field information is calibrated to obtain the experimental state field information; Based on the experimental state field information, predict the state field information of the experimental pool at future moments; The obtained state field information of the fluid in the experimental pool is specifically as follows: Based on the first coordinate data and the data collected by the physical sensor group, the first state vector of the corresponding physical sensor group is obtained. The first state vector is processed by a multi-layer perceptual layer of a graph neural network model to obtain the first input vector; The attention mechanism of the graph neural network model is used to process the first input vector of each node to obtain the connection weights between the nodes. The message passing module of the graph neural network model is used to process the first input vector and connection weights to obtain the first output vector of the node; All the first output vectors are combined into a first output tensor according to the first coordinate data of the physical sensor group. The dimension of the first output tensor is H×W×C, where H is the number of physical sensor groups in each row, W is the number of physical sensor groups in each column, and C is the dimension of the first output vector. The first output tensor is deconvolved using a multi-layer deconvolution of a graph neural network model to obtain state field information; The experimental state field information is obtained by calibrating the state field information using sensors mounted on the ship, specifically: Estimate the acceleration at the ship's location based on the state field information; Calculate the error between the estimated acceleration of the ship and the actual acceleration of the ship; Calculate the error between the actual physical state information at the first coordinate and the physical state information at the first coordinate in the state field information; A first loss function is constructed based on the two errors mentioned above. The graph neural network model is then calibrated based on this first loss function. The first coordinate data and the data collected by the physical sensor group are then input back into the calibrated graph neural network model to obtain the experimental state field information.

2. The sensor-based fluid motion data prediction method according to claim 1, characterized in that, Buoyant carriers are placed at multiple locations in the fluid, and physical sensor groups are placed on the corresponding carriers. The lower surface of the carrier is connected to the bottom of the experimental pool by ropes. The planar coordinates of the connection position between the rope and the bottom of the experimental pool are used as the first coordinate data. The length of the rope is longer than the vertical distance between the lower surface of the carrier and the bottom of the experimental pool.

3. The sensor-based fluid motion data prediction method according to claim 1, characterized in that, The acceleration at the ship's location is estimated based on the state field information, specifically as follows: Estimate the ship's acceleration based on the acceleration at multiple locations in the state field information and the distances between these locations and the ship.

4. The sensor-based fluid motion data prediction method according to claim 1, characterized in that, The first loss function The expression is: ； in, Let be a vector composed of the ship's actual acceleration data at time t. For the state field information at time t A vector composed of acceleration data at coordinates. For the t-th time Distance data between coordinates and ship position data, , Let be a vector composed of the physical state information at the k-th first coordinate data point in the state field information at time t. This is a vector composed of the actual physical state information at the k-th first coordinate data point at time t. and The preset weights are N, where N is the number of time points, and t, N, k, H, W, i, j All are positive integers. The number of coordinates in each column of the state field information. This represents the number of coordinates in each row of the state field information.

5. The sensor-based fluid motion data prediction method according to claim 1, characterized in that, Based on experimental state field information at multiple time points, the system predicts future state field information. Specifically, it sets up a predictive neural network model, including F convolutional layers connected in sequence, a state prediction model, and F deconvolutional layers connected in sequence. The experimental state field information at time t is input into the predictive neural network model for the t-th state. First, the F convolutional layers output the features at time t. The features at time t and the hidden state vector at time t-1 are input into the state prediction model to obtain the hidden state vector at time t. The hidden state vector at time t is then input into the F deconvolutional layers. Simultaneously, the f-th deconvolutional layer is skipped to the f-th convolutional layer; f = 1, 2, ..., F; the F-level deconvolutional layers output the predicted state field information at time t+1; the predicted neural network model for the t-th state is trained based on the experimental state field information and the predicted state field information at time t+1, and the predicted neural network model for the t+1-th state is obtained; the experimental state field information and the predicted neural network model for the t+1-th state are used to predict the predicted state field information at time t+2, until the experimental state field information at all times is traversed, and the training of the predicted neural network model is completed.

6. The sensor-based fluid motion data prediction method according to claim 5, characterized in that, Loss function when training a predictive neural network model The expression is: ； in, This represents the physical state information of each coordinate in the experimental state field information at time t+1. This represents the physical state information of each coordinate in the predicted state field information at time t+1. This represents the number of coordinates in each row of the experimental state field information. N represents the number of coordinates in each column of the experimental state field information, and N represents the number of time points.

7. The sensor-based fluid motion data prediction method according to claim 1, characterized in that, The method also includes assessing the risk factor of the experimental vessel by predicting acceleration data at specific locations in the experimental pool over a future period.

8. A system applied to the sensor-based fluid motion data prediction method of claim 1, characterized in that, include The setup module is used to set up multiple physical sensor groups in the fluid and record the first coordinate data of the physical sensor groups. The ship module, which includes position sensors and acceleration sensors mounted on the ship; Wave module, used to generate waves in the experimental pool; The acquisition module is used to acquire data collected by various physical sensor groups at multiple times, as well as position data collected by the ship's position sensors and acceleration data collected by the ship's acceleration sensors. The state field information module is used to obtain the state field information of the fluid in the experimental pool based on the data collected by the physical sensor group and the first coordinate data. The experimental state field information module is used to calibrate the state field information based on the data collected by the sensors on board the ship, and obtain the experimental state field information. The predicted state field information module is used to determine the predicted state field information for subsequent times based on the experimental state field information at multiple times.

Images