Unmanned aerial vehicle fault prediction based on entropy weight fusion and time graph convolution network

Abstract

This invention relates to time series prediction of UAV faults based on entropy weight fusion and a time-graph convolutional network (T-GCN). It acquires data on the UAV's acceleration, angular velocity, and angle in the x, y, and z directions using sensors; combines the data into a matrix and calculates the entropy weight fusion value; determines the threshold value for fault occurrence based on the chi-square distribution; and constructs graph data for input into a T-GCN model based on the relationships between the sensor data. n historical time data points are segmented from the graph data and input into the T-GCN model. After inputting n historical time data points, the T-GCN model can predict data at time T in the future. After model training, it captures the spatiotemporal correlation of the data obtained from the UAV's sensors and predicts the graph data based on the time series. The prediction results are accurate and reliable, facilitating timely maintenance of quadcopter UAVs.

Description

Technical Field

[0001] This invention relates to the field of mechanical engineering technology, specifically to time series prediction of unmanned aerial vehicle (UAV) failures. Background Technology

[0002] With the continuous advancement of science and technology, quadcopter drones are becoming increasingly popular in applications such as aerial photography, logistics delivery, and agricultural environmental protection. However, during the operation of quadcopter drones, various malfunctions may occur, such as motor shutdown, remote control failure, and propeller blade damage. Due to the high precision and complex mechanisms of drone mechanical structures, these malfunctions are often difficult to identify. To address this issue, establishing a drone condition monitoring system can provide an effective solution for drone flight fault diagnosis. Such a system can predict the future operating status of the drone equipment, ensuring safety during drone operation and reducing equipment maintenance costs. This process, known as Predictive Health Management (PHM), is crucial for achieving optimal drone performance.

[0003] Fault diagnosis involves understanding and monitoring the state of a machine during operation, determining whether it is normal or abnormal overall or in parts, and detecting faults and their causes early. Deep learning is now widely used for fault diagnosis. The Temporal Graph Convolutional Network (T-GCN) model can simultaneously capture spatial and temporal dependencies and is a neural network-based traffic volume prediction method. The problem of predicting the spatiotemporal sequence of traffic flow is transformed into learning a mapping function f between a topological graph G and a feature matrix X to calculate traffic information at the next time step T. First, historical time-series data of length n is used as input. The GCN receives the spatial information of the topological structure. Then, the received spatial and temporal information is input into a GRU to obtain dynamic information changes between units, extracting temporal features. Finally, a fully connected layer is used to obtain the prediction result. For a given adjacency matrix A and feature matrix X, the GCN model constructs a filter in the Fourier domain. This filter acts on the nodes of the graph, capturing the spatial features between nodes through their first-order neighborhoods. Then, multiple convolutional layers are stacked to construct the GCN model. Assuming node 1 is the central road, the GCN model can acquire the topological relationship between the central road and its surrounding roads, encoding the road network topology and road attributes to obtain spatial dependencies. Widely used recurrent neural networks (RNNs) are unsuitable for long-term predictions due to gradient explosion; LSTM and GRU, as variants of RNNs, overcome this problem. Their common principle is to use gate-level mechanisms to store as much periodic information as possible. However, because LSTM has a complex structure and long training time, while GRU has a simpler structure and shorter computation time, the GRU model is chosen to extract temporal correlations from traffic data. Summary of the Invention

[0004] The purpose of this invention is to predict the failure status of quadcopter drones. It proposes a drone failure prediction method based on entropy weight fusion and temporal graph convolutional network. Data is obtained by sensors and constructed into graph data. Finally, the failure prediction of quadcopter drones is performed through the T-GCN graph convolutional neural model.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] A method for drone fault prediction based on entropy weight fusion and temporal graph convolutional networks includes the following steps:

[0007] Step 1: Acquire acceleration, angular velocity, and angle signals in the x, y, and z directions using sensors;

[0008] Step 2: Combine the acquired data into a matrix, normalize the elements in the matrix, calculate the weight of the i-th element in each column, and then obtain the entropy value of the j-th column. Finally, obtain the entropy-weighted fusion value.

[0009] Step 3: Calculate the threshold value at which the entropy weight fusion value fails based on the chi-square distribution. This threshold is set as the identifier for the T-GCN time-map convolutional network model to start prediction.

[0010] Step 4: Based on the relationships between the data acquired by the sensors, construct graph data for output to the T-GCN temporal graph convolutional network model;

[0011] Step 5: Extract n historical time data points from the graph data and input them into the T-GCN temporal graph convolutional network model. After inputting n historical time data points, the T-GCN temporal graph convolutional network model can predict the data at the future time T.

[0012] Step 6: The T-GCN temporal graph convolutional network model begins training, and after training, it captures the spatiotemporal correlation of UAV sensor data and predicts graph data based on time series.

[0013] In the above scheme, the entropy weight fusion principle in step two is as follows:

[0014] In information theory, entropy is a measure of uncertainty. The greater the uncertainty, the greater the entropy, and the more information it contains; the smaller the uncertainty, the smaller the entropy, and the less information it contains.

[0015] Based on the properties of entropy, the randomness and disorder of an event can be determined by calculating the entropy value. The entropy value can also be used to determine the dispersion of a certain indicator; the greater the dispersion of the indicator, the greater its impact (weight) on the overall evaluation. For example, if all sample data have the same value for a certain indicator, then the indicator's impact on the overall evaluation is 0, and its weight is 0. The entropy weighting method is an objective weighting method because it relies solely on the dispersion of the data itself.

[0016] The entropy weight fusion method merges multiple sets of different data into a single set. After normalizing each set of data, the entropy value of each set is calculated. The proportion of the entropy value of each set of data in the sum of the entropy values ​​of all data is the entropy weight. The entropy weight fusion value is obtained by weighting and summing multiple sets of data using the entropy weight.

[0017] In the above scheme, the principle of the chi-square distribution threshold division in step three is as follows:

[0018] The chi-square distribution threshold is determined by dividing the entropy weight fusion value of the UAV during normal operation into two parts, calculating the chi-square statistic, and then looking up the chi-square distribution table based on the degrees of freedom and significance level. The resulting critical value is the threshold for the occurrence of a fault.

[0019] In the above scheme, the principle behind step four of constructing the graph data used as input to the T-GCN temporal graph convolutional network model is as follows:

[0020] An adjacency matrix and a degree matrix are created based on the relationships between the data. The rows and columns of the adjacency matrix represent data nodes, and the elements represent the correlation between two nodes. An element is 1 if two nodes are correlated, and 0 otherwise. Assume the fused data is correlated with nine other datasets, and the XYZ axes of acceleration, angular velocity, and angle are interrelated. An identity matrix I is added to the adjacency matrix A to obtain a scaled adjacency matrix A' with self-looping nodes; the degree matrix is ​​used to represent the connecting edges of each node.

[0021] In the above scheme, the operational principle of the T-GCN temporal graph convolutional network model in step five is as follows:

[0022] T-GCN is a neural network-based traffic prediction method that combines GCN and GRU. GCN constructs a filter in the Fourier domain, which acts on the nodes of the graph and their first-order neighbors to capture the spatial features between nodes, then stacks multiple convolutional layers. GRU incorporates two gate structures to selectively retain or forget information. The update gate determines which information needs to be forgotten and which needs to be remembered, while the reset gate determines the degree of forgetting of previous information. GRU uses these gate structures to capture the temporal features of the data.

[0023] Compared with the prior art, the present invention has the following advantages:

[0024] 1. This invention can predict the failure status of quadcopter drones based on historical data, which helps to carry out timely maintenance and thus extend the service life of quadcopter drones.

[0025] 2. This invention first processes the signals collected by the sensors into graph data, and then uses the T-GCN graph neural convolutional network model to predict the faults of quadcopter drones. The prediction effect is accurate and reliable. Attached Figure Description

[0026] Figure 1 This is a picture of a quadcopter drone.

[0027] Figure 2 The actual data collected by the sensor;

[0028] Figure 3 This refers to graph data that is processed from data collected by the sensor and then input into the T-GCN graph neural convolutional network.

[0029] Figure 4 The image shows the prediction results of the T-GCN time-plot convolutional network model compared to other models. Detailed Implementation

[0030] The following will further explain the concept, specific structure, and technical effects of the present invention in conjunction with the accompanying drawings, so as to fully understand the purpose, features, and effects of the present invention.

[0031] It should be noted that in the description of this invention, the terms "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc., which indicate the direction or positional relationship, are based on the direction or positional relationship shown in the drawings. This is only for the convenience of description and is not intended to indicate or imply that the device or element must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this invention. Example

[0032] See Figures 1-4 As shown, the UAV fault prediction based on entropy weight fusion and temporal graph convolutional network provided by this invention includes the following steps:

[0033] Step 1: Obtain data on the drone's acceleration, angular velocity, and angle in the x, y, and z directions using sensors.

[0034] Step 2: Combine the acquired data into a matrix, normalize the elements in the matrix, calculate the weight of the i-th element in each column, and then obtain the entropy value of the j-th column. Finally, obtain the entropy-weighted fusion value. In information theory, entropy is a measure of uncertainty. The greater the uncertainty, the greater the entropy, and the more information it contains; the smaller the uncertainty, the smaller the entropy, and the less information it contains. Based on the characteristics of entropy, the entropy value can be used to determine the randomness and disorder of an event, and it can also be used to determine the dispersion of an indicator. The greater the dispersion of an indicator, the greater its impact (weight) on the overall evaluation. For example, if the sample data all take the same value under a certain indicator, then the influence of that indicator on the overall evaluation is 0, and the weight is 0. The entropy weighting method is an objective weighting method because it relies only on the dispersion of the data itself. The entropy-weighted fusion method merges multiple different sets of data into one set of data. After normalizing each set of data, the entropy value of each set of data is calculated. The proportion of the entropy value of each set of data in the sum of the entropy values ​​of all data is the entropy weight. The entropy weight fusion value is obtained by weighting and summing multiple sets of data using the entropy weight.

[0035] Step 3: Calculate the threshold value at which a fault occurs in the entropy weight fusion value based on the chi-square distribution. This threshold is set as the indicator for when the model begins prediction. The chi-square distribution threshold is determined by dividing the entropy weight fusion value during normal operation of the UAV into two parts, calculating the chi-square statistic, and then looking up the chi-square distribution table based on the degrees of freedom and significance level. The resulting critical value is the threshold value at which a fault occurs.

[0036] Step 4: Based on the relationships between the data acquired by the sensors, construct the graph data to be output to the model. Create an adjacency matrix and a degree matrix based on the relationships between the data. The rows and columns of the adjacency matrix represent data nodes, and the elements represent the correlation between two nodes. If two nodes are correlated, the element is 1; otherwise, it is 0. Assume that the fused data is related to nine other datasets, and the XYZ axes of acceleration, angular velocity, and angle are interrelated. Add the identity matrix I to the adjacency matrix A to obtain a scaled adjacency matrix A' with self-looping nodes. The degree matrix is ​​used to represent the connecting edges of each node.

[0037] Step 5: Extract n historical time data points from the graph structure data and input them into the T-GCN temporal graph convolutional network model. After inputting n historical time data points, the T-GCN temporal graph convolutional network model can predict the data at time T in the future. The T-GCN temporal graph convolutional network model combines GCN and GRU. On one hand, GCN is used to capture the topological structure of the graph to obtain spatial correlation; on the other hand, the GRU model is used to capture the dynamic changes of node attributes to obtain temporal correlation. The T-GCN temporal graph convolutional network model can obtain the best prediction results under different prediction ranges.

[0038] Step Six: The T-GCN temporal graph convolutional network model begins training. It learns and trains using graph data constructed from real-time sensor data, and after training, captures the spatiotemporal correlation of UAV sensor data and predicts graph data based on time series. The GCN (Graph Convolutional Neural Network) in the T-GCN temporal graph convolutional network model captures the spatial features between nodes by constructing a filter in the Fourier domain, which acts on the nodes of the graph and their first-order neighborhoods. This is then achieved by stacking multiple convolutional layers. The GRU incorporates two gate structures to select whether to retain or forget information. The update gate determines the information to be forgotten and remembered, while the reset gate determines the degree of forgetting of previous information. The GRU captures the temporal features of the data through these gate structures.

[0039] Figure 4 As shown in the figure, T-GCN is compared with five other baseline models. The prediction results are shown in the figure. T-GCN shows good performance in prediction accuracy, spatiotemporal prediction capability and long-term prediction capability under all evaluation indicators. Therefore, using T-GCN graph convolutional neural network to predict the faults of quadcopter UAVs can effectively detect the faults of UAVs, such as the damage changes of propellers, and thus enable timely maintenance of UAVs, which greatly extends the service life of quadcopter UAVs.

[0040] This invention can predict the malfunctions of quadcopter drones based on historical data, facilitating timely maintenance and extending the service life of quadcopter drones. First, the signals collected by the sensors are processed into graph data, and then a T-GCN graph neural convolutional network model is used to predict the malfunctions of the quadcopter drone, resulting in accurate and reliable predictions.

[0041] While preferred embodiments of the present invention have been described above in conjunction with the accompanying drawings, the present invention should not be limited to structures and operations that are exactly the same as those described above and shown in the drawings. Those skilled in the art can make many equivalent improvements and variations to the above embodiments through logical analysis, reasoning, or limited experiments without departing from the concept and scope of the present invention, but all such improvements and variations should fall within the scope of protection claimed by the present invention.

Claims

1. An unmanned aerial vehicle fault prediction based on entropy weight fusion and time graph convolution network, characterized in that, The method comprises the following steps: Step one, through the sensor to obtain the acceleration, angular velocity and angle of the unmanned aerial vehicle respectively in x, y, z direction data; Step two, the data obtained is combined into a matrix, the elements in the matrix are normalized, the weight of the i-th element in each column is calculated to obtain the entropy value of the j-th column, and finally the entropy weight fusion value is obtained; Step three, the threshold size of the entropy weight fusion value when a fault occurs is calculated based on the chi-square distribution, and the threshold is set as the identification of the T-GCN time graph convolutional neural network model for prediction; Step four, according to the relationship between the data obtained by the sensor, the graph data used for input into the T-GCN time graph convolutional neural network model is constructed; Step five, n historical time data is segmented from the graph data and input into the T-GCN time graph convolutional neural network model, after inputting n historical time data, the T-GCN time graph convolutional neural network model can predict the data at T time in the future; Step six, the T-GCN time graph convolutional neural network model starts training, and after training, the spatiotemporal correlation of the data obtained by the sensor of the unmanned aerial vehicle is captured, and the time series prediction graph data is obtained.

2. The unmanned aerial vehicle fault prediction based on entropy weight fusion and temporal graph convolution network according to claim 1, characterized in that, The principle of chi-square distribution threshold in step three is as follows: The chi-square distribution threshold is divided into two parts by dividing the entropy weight fusion value when the unmanned aerial vehicle is running normally, and the chi-square statistic is calculated, the critical value obtained by looking up the chi-square distribution table according to the degrees of freedom and the significance level is the threshold value of the fault.

3. The UAV fault prediction based on entropy weight fusion and temporal graph convolution network according to claim 1, characterized in that, The principle of constructing the graph data used for input into the T-GCN time graph convolutional neural network model in step four is as follows: Based on the relationship between the data, an adjacency matrix and a degree matrix are created, the rows and columns of the adjacency matrix represent data nodes, and the elements represent the correlation between two nodes; If two nodes are related, the element is 1; otherwise, it is 0; suppose that the fusion data is related to other nine data sets, and the XYZ axes of acceleration, angular velocity and angle are related to each other; Add the unit matrix I to the adjacency matrix A to obtain the scaled adjacency matrix A' with self-loop on the node, and the degree matrix is used to represent the connection edges of each node.

4. The unmanned aerial vehicle fault prediction based on entropy weight fusion and temporal graph convolution network according to claim 1, characterized in that, The operation principle of T-GCN time graph convolutional neural network model in step five is as follows: T-GCN is a neural network based traffic prediction method, which combines GCN and GRU; GCN constructs a filter in the Fourier domain, which acts on the nodes and their first-order neighborhood of the graph, captures the spatial features between nodes, and then multiple convolution layers are stacked; two gate structures are set in GRU, which are used to select to save information or forget information; among them, the update gate decides the information that needs to be forgotten and the information that needs to be remembered, and the reset gate decides the forgetting degree of the previous information; GRU realizes the capture of the time characteristics of the data through the gate structure.

Images