UAV health status assessment method and system based on transfer learning

Abstract

The present invention relates to the technical field of predictive maintenance of unmanned aerial vehicles, and provides a method and system for evaluating the health status of unmanned aerial vehicles based on transfer learning, the method comprising: S1, determining the type of unmanned aerial vehicle fault, the amount of data in each category, and the input variables of the model; S2, performing a torsion test and a simulation test, obtaining data in each category, and constructing a torsion test data set and a simulation data set; S3, a model for evaluating the health status of unmanned aerial vehicles; S4, training the model; S5, using the model to perform health evaluation of unmanned aerial vehicles. Compared with the health evaluation method based on training models with real historical data, the present invention can identify more types of health status, and the model has stronger generalization ability; compared with the health evaluation method based on training models with torsion test data, the data cost of the training model of the present invention is lower and the data collection cycle is shorter; compared with the health evaluation method based on training models with simulation data, the model trained by the present invention is more in line with the physical laws in actual application scenarios.

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned aerial vehicle predictive maintenance, and in particular to a method and system for evaluating the health status of an unmanned aerial vehicle based on transfer learning. Background Art

[0002] As drone manufacturing technology becomes increasingly mature, drones have been widely used in disaster relief, environmental testing, agricultural plant protection and other fields. These application scenarios mostly involve independent operation of drones, which places high demands on the maintainability of drone equipment. Efficiently and accurately assessing the health status of drone equipment and performing on-demand maintenance on equipment with health risks can ensure the maintainability of drones, reduce redundant maintenance and insufficient maintenance, effectively reduce the risk of catastrophic accidents and associated damage, and greatly reduce unplanned downtime, thereby further improving operational efficiency and quality.

[0003] At present, the commonly used health status assessment methods gradually tend to establish a deep learning model for assessing the health status of drones, and diagnose the data collected by sensors during flight. This method is limited by the data set for training deep learning models. The main sources of data include: real historical data usually have the characteristics of category imbalance, the amount of health data is much larger than the amount of fault data, and the deep learning model cannot be trained to obtain effective knowledge; the copy machine test can obtain a category-balanced data set, but it is often too costly and time-consuming, and the amount of data that can be obtained is not enough to meet the training requirements; the simulation test data has a low cost and can meet the training set data volume requirements, but there will be deviations from the actual physical situation, affecting the accuracy of the model in actual application. In order to solve the limitations of single-source training data, the present invention proposes a method for assessing the health status of drone equipment based on transfer learning, realizes multi-source data fusion modeling, reduces the cost and time of model training, improves the prediction accuracy and generalization ability of the model in practical applications, and ensures the efficiency and effectiveness of drone health status assessment. Summary of the invention

[0004] The purpose of the present invention is to overcome the shortcomings of the prior art: in the current UAV health assessment method, the training data set of the deep learning model from a single source has category imbalance, high collection cost and long time, and deviation from the actual physical situation; and a UAV health status assessment method and system based on transfer learning are provided to improve the accuracy and efficiency of UAV health assessment.

[0005] The present invention adopts the following technical solution:

[0006] On the one hand, the present invention provides a method for evaluating the health status of a drone based on transfer learning, comprising:

[0007] S1. Determine the fault type of the UAV, the amount of data in each category, and the input variables of the UAV health status assessment model;

[0008] S2. According to the data volume of each category of data determined in step S1, a stress test and a simulation test are performed on the UAV to obtain data of each category, and the data is standardized and normalized to be divided into time series data, and a stress test data set and a simulation data set are constructed, and each is divided into a training set and a test set;

[0009] S3, according to the fault type of the drone determined in step S1 and the input variables of the drone health status assessment model, establish a drone health status assessment model using a long short-term memory network LSTM for feature extraction and a fully connected network and a Softmax layer for classification;

[0010] S4. Based on the data set established in step S2, the drone health status assessment model obtained in step S3 is trained and tested. When the accuracy meets the set requirements, the trained drone health status assessment model is obtained.

[0011] S5. Evaluate the health status of the drone based on the trained drone health status evaluation model obtained in step S4.

[0012] As any possible implementation described above, an implementation is further provided, in step S1, the failure type of the drone includes potential minor accidents, precursors to attempted accidents, and potential accidents; the input variables include data collected by drone sensors and drone operating environment data.

[0013] Any possible implementation as described above further provides an implementation method, in which the amount of simulation data should be 10-15 times the input variables, and the amount of stress test data should be 2-5 times the input variables, and the data distribution of each category in the simulation data and stress test data should be as balanced as possible.

[0014] Any possible implementation described above further provides an implementation, wherein the specific method of step S2 includes:

[0015] S2.1 Build a test bench and conduct a stress test. Use sensors to collect time series data generated during the operation of the drone equipment and record the health status of the equipment during the process.

[0016] S2.2 builds simulation models under various health conditions based on the MATLAB Simulink environment, runs the models and collects the time series data generated during the simulation process;

[0017] S2.3 normalizes the time series data collected in step S2.1 and step S2.2 to eliminate the influence of large differences in the numerical scales of different input variables on the modeling accuracy. The specific method is shown in formula (1):

[0018]

[0019] Among them, x′0 is the normalized data, x0 is the original data, and x min is the minimum value of all the data of the input variable to which the data belongs, x max It is the maximum value of all the data of the input variable to which the data belongs;

[0020] S2.4 determines the step length of a single time series data for the time series data processed by step S2.3, slices the stress test data and the simulation data according to the step length, and assigns a health status label to each slice;

[0021] S2.5 divides the copy machine test data slice set into a test data training set and a test set according to a set ratio, and divides the simulation data into a simulation data training set and a test set according to the same ratio.

[0022] Any possible implementation as described above further provides an implementation, in which in step S2.5, the set ratio is 8:2.

[0023] Any possible implementation described above further provides an implementation, in which in step S3, the method for establishing a drone health status assessment model specifically includes:

[0024] S3.1 Establish an LSTM network for extracting flight data time series features and determine the network's hyperparameters, including: the feature dimension of the input data, the dimension of the hidden layer, the number of neural network layers, and the activation function;

[0025] S3.2 establishes a fully connected network and a Softmax layer for classifying the output tensors of the LSTM network, and determines the hyperparameters of the network, including the number of layers of the network, the number of neurons in each layer, and the activation function of each layer;

[0026] S3.3 connects the LSTM network, the fully connected network and the Softmax layer in series as a complete drone health status assessment model;

[0027] S3.4 sets the training hyperparameters of the model, including the number of training rounds, batch size, and loss function.

[0028] Any possible implementation as described above, further provides an implementation, in step S3.4, the loss function includes but is not limited to a cross entropy function, a mean square error, and an exponential loss.

[0029] Any possible implementation described above further provides an implementation, in which in step S4, a specific method for training the drone health status assessment model includes:

[0030] S4.1 The drone health status assessment model established in step S3 is trained based on the simulation data training set, and the prediction accuracy test is performed based on the simulation data test set. The model that meets the accuracy requirements is named a pre-trained model;

[0031] S4.2 Freeze the neurons of the LSTM network in the pre-trained model so that they do not participate in training;

[0032] S4.3 Based on the stress test data training set, the pre-trained model of the frozen LSTM network is retrained based on the transfer learning finetune method, and the model accuracy is tested based on the stress test data test set. The model that meets the accuracy requirements is the UAV equipment health status assessment model finally used for predictive maintenance.

[0033] On the other hand, the present invention also provides a UAV health status assessment system based on transfer learning, comprising:

[0034] The data acquisition and processing module is used to collect the UAV test data and simulation test data, standardize and normalize the data, and construct the test data set and simulation data set;

[0035] The drone health status assessment model establishment and training module, based on the drone fault type and the input variables of the drone health status assessment model, establishes a drone health status assessment model that uses the long short-term memory network LSTM for feature extraction and the fully connected network and Softmax layer for classification;

[0036] The UAV health status assessment model training module trains and tests the UAV health status assessment model based on the data set established by the data acquisition and processing module.

[0037] The drone health status assessment module uses the trained drone health status assessment model to assess the drone health status.

[0038] The beneficial effects of the present invention are as follows: the present invention is based on the transfer learning finetune method, integrates the UAV test data and simulation data for deep learning modeling, and performs UAV equipment health status assessment based on the trained model. Compared with the health assessment method based on a single-source data training model, the present invention has the following advantages: ① Compared with the health assessment method based on the real historical data training model, the present invention can identify more types of health states, and the model generalization ability is stronger; ② Compared with the health assessment method based on the test data training model, the data cost of the health assessment model trained by the present invention is lower and the data collection cycle is shorter; ③ Compared with the health assessment method based on the simulation data training model, the health assessment model trained by the present invention is more in line with the physical laws in the actual application scenario. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 The figure is a flow chart of a method for assessing the health status of a UAV based on transfer learning according to an embodiment of the present invention.

[0040] Figure 2 The figure shows a network architecture diagram of the drone health status assessment model in the embodiment.

[0041] Figure 3 Shown is the LSTM network neuron structure. DETAILED DESCRIPTION

[0042] The specific embodiments of the present invention will be described in detail below in conjunction with the specific drawings. It should be noted that the technical features or combinations of technical features described in the following embodiments should not be considered isolated, and they can be combined with each other to achieve better technical effects.

[0043] like Figure 1 As shown, taking a piston engine of a certain drone as an example, an embodiment of the present invention provides a method for assessing the health status of a drone based on transfer learning, including:

[0044] S1. Determine the fault type of the UAV, the amount of data in each category, and the input variables of the UAV health status assessment model;

[0045] S2. According to the data volume of each category of data determined in step S1, a stress test and a simulation test are performed on the UAV to obtain data of each category, and the data is standardized and normalized to be divided into time series data, and a stress test data set and a simulation data set are constructed, and each is divided into a training set and a test set;

[0046] S3, according to the fault type of the drone determined in step S1 and the input variables of the drone health status assessment model, establish a drone health status assessment model using a long short-term memory network LSTM for feature extraction and a fully connected network and a Softmax layer for classification;

[0047] S4. Based on the data set established in step S2, the drone health status assessment model obtained in step S3 is trained and tested. When the accuracy meets the set requirements, the trained drone health status assessment model is obtained.

[0048] S5. Evaluate the health status of the drone based on the trained drone health status evaluation model obtained in step S4.

[0049] In a specific embodiment, in step S1, the failure types of the drone include potential minor accidents, precursors to unsuccessful accidents, and potential accidents; the input variables include data collected by drone sensors. This embodiment includes 12 design variables such as fuel temperature and lubricating oil temperature, and plans to collect 180 sets of simulation data and 30 sets of test data, with each of the three types of failures accounting for one third.

[0050] In a specific embodiment, the specific method of step S2 includes:

[0051] S2.1 Build a test bench and conduct a stress test. Use sensors to collect time series data generated during the operation of the drone equipment and record the health status of the equipment during the process.

[0052] S2.2 Build simulation models for minor accidents, precursors to accidents, and potential accidents based on the MATLAB Simulink environment, run the models, and collect time series data generated during the simulation process;

[0053] S2.3 normalizes the time series data collected in step S2.1 and step S2.2 to eliminate the influence of large differences in the numerical scales of different input variables on the modeling accuracy. The specific method is shown in formula (1):

[0054]

[0055] Among them, x′0 is the normalized data, x0 is the original data, and x min is the minimum value of all the data of the input variable to which the data belongs, x max It is the maximum value of all the data of the input variable to which the data belongs;

[0056] S2.4 determines the step length of a single time series data for the time series data processed by step S2.3, slices the stress test data and the simulation data according to the step length, and assigns a health status label to each slice;

[0057] S2.5 divides the copy machine test data slice set into a test data training set and a test set according to a set ratio, and divides the simulation data into a simulation data training set and a test set according to the same ratio.

[0058] In a specific embodiment, in step S2.5, the set ratio is 8:2.

[0059] In a specific embodiment, in step S3, the method for establishing a drone health status assessment model specifically includes:

[0060] S3.1 Establish an LSTM network for extracting flight data time series features and determine the network's hyperparameters, including: the feature dimension of the input data, the dimension of the hidden layer, the number of neural network layers, and the activation function;

[0061] S3.2 establishes a fully connected network and a Softmax layer for classifying the output tensors of the LSTM network, and determines the hyperparameters of the network, including the number of layers of the network, the number of neurons in each layer, and the activation function of each layer;

[0062] S3.3 connects the LSTM network, the fully connected network and the Softmax layer in series as a complete drone health status assessment model; the network architecture is as follows Figure 2 As shown;

[0063] S3.4 sets the training hyperparameters of the model, including the number of training rounds, batch size, and loss function.

[0064] In a specific embodiment, in step S3.4, the loss function includes but is not limited to a cross entropy function, a mean square error, and an exponential loss.

[0065] In a specific embodiment, in step S4, the specific method for training the drone health status assessment model includes:

[0066] S4.1 The drone health status assessment model established in step S3 is trained based on the simulation data training set, and the prediction accuracy test is performed based on the simulation data test set. The model that meets the accuracy requirements is named a pre-trained model;

[0067] S4.2 Freeze the neurons of the LSTM network in the pre-trained model so that it does not participate in training; the neuron structure of the LSTM network is as follows Figure 3 As shown;

[0068] S4.3 Based on the stress test data training set, the pre-trained model of the frozen LSTM network is retrained based on the transfer learning finetune method, and the model accuracy is tested based on the stress test data test set. The model that meets the accuracy requirements is the UAV equipment health status assessment model finally used for predictive maintenance.

[0069] Because the amount of simulation data is larger, it is beneficial for the model to fully learn common features and avoid overfitting during the training process; and using the test data to transfer learning the model is to make the model's prediction results closer to the test data rather than the simulation data. If it is the other way around, the first step, pre-training, will overfit, and the second step, transfer learning, will cause the model's prediction results to deviate more from the test data, resulting in reduced prediction accuracy.

[0070] An embodiment of the present invention provides a UAV health status assessment system based on transfer learning, including:

[0071] The data acquisition and processing module is used to collect the UAV test data and simulation test data, standardize and normalize the data, and construct the test data set and simulation data set;

[0072] The drone health status assessment model establishment and training module, based on the drone fault type and the input variables of the drone health status assessment model, establishes a drone health status assessment model that uses the long short-term memory network LSTM for feature extraction and the fully connected network and Softmax layer for classification;

[0073] The UAV health status assessment model training module trains and tests the UAV health status assessment model based on the data set established by the data acquisition and processing module.

[0074] The drone health status assessment module uses the trained drone health status assessment model to assess the drone health status.

[0075] The present invention is compared with a health assessment method using only simulation data and a health assessment method using only stress test data. The results are shown in Table 1:

[0076] Table 1 Comparison of health assessment methods

[0077]

[0078] It can be seen that the present invention realizes multi-source data fusion modeling, reduces the cost and time of model training, improves the prediction accuracy and generalization ability of the model in practical applications, and ensures the efficiency and effectiveness of drone health status assessment.

[0079] Although several embodiments of the present invention have been given herein, those skilled in the art should understand that the embodiments of the present invention may be modified without departing from the spirit of the present invention. The above embodiments are merely exemplary and should not be used as a limitation on the scope of the present invention.

Claims

1. A method for assessing the health status of a drone based on transfer learning, characterized in that: The method comprises: S1. Determine the fault type of the UAV, the amount of data in each category, and the input variables of the UAV health status assessment model; S2. According to the data volume of each category of data determined in step S1, a stress test and a simulation test are performed on the UAV to obtain data of each category, and the data is standardized and normalized to be divided into time series data, and a stress test data set and a simulation data set are constructed, and each is divided into a training set and a test set; S3, according to the fault type of the drone determined in step S1 and the input variables of the drone health status assessment model, establish a drone health status assessment model using a long short-term memory network LSTM for feature extraction and a fully connected network and a Softmax layer for classification; S4. Based on the data set established in step S2, the drone health status assessment model obtained in step S3 is trained and tested. When the accuracy meets the set requirements, the trained drone health status assessment model is obtained; S5, evaluating the health status of the drone based on the trained drone health status evaluation model obtained in step S4; The specific method of step S2 includes: S2.1 Build a test bench and conduct a stress test. Use sensors to collect time series data generated during the operation of the drone equipment and record the health status of the equipment during the process. S2.2 builds simulation models under various health conditions based on the MATLAB Simulink environment, runs the models and collects the time series data generated during the simulation process; S2.3 normalizes the time series data collected in step S2.1 and step S2.

2. The specific method is shown in formula (1): Among them, x′0 is the normalized data of x0, x0 is the original data, and x min is the minimum value of all the input variables to which data x0 belongs, x max It is the maximum value of all the input variables to which data x0 belongs; S2.4 determines the step length of a single time series data for the time series data processed by step S2.3, slices the stress test data and the simulation data according to the step length, and assigns a health status label to each slice; S2.5 divides the copy machine test data slice set into a test data training set and a test set according to a set ratio, and divides the simulation data into a simulation data training set and a test set according to the same ratio; In step S3, the method for establishing the drone health status assessment model specifically includes: S3.1 Establish an LSTM network for extracting flight data time series features and determine the network's hyperparameters, including: the feature dimension of the input data, the dimension of the hidden layer, the number of neural network layers, and the activation function; S3.2 establishes a fully connected network and a Softmax layer for classifying the output tensors of the LSTM network, and determines the hyperparameters of the network, including the number of layers of the network, the number of neurons in each layer, and the activation function of each layer; S3.3 connects the LSTM network, the fully connected network and the Softmax layer in series as a complete drone health status assessment model; S3.4 Set the model's training hyperparameters, including the number of training rounds, batch size, and loss function; In step S4, the specific method for training the drone health status assessment model includes: S4.1 The drone health status assessment model established in step S3 is trained based on the simulation data training set, and the prediction accuracy test is performed based on the simulation data test set. The model that meets the accuracy requirements is named a pre-trained model; S4.2 Freeze the neurons of the LSTM network in the pre-trained model so that they do not participate in training; S4.3 Based on the stress test data training set, the pre-trained model of the frozen LSTM network is retrained based on the transfer learning finetune method, and the model accuracy is tested based on the stress test data test set. The model that meets the accuracy requirements is the UAV equipment health status assessment model finally used for predictive maintenance.

2. The method for assessing the health status of a drone based on transfer learning according to claim 1, characterized in that: In step S1, the failure types of the drone include potential minor accidents, precursors to unsuccessful accidents, and potential accidents; the input variables include data collected by drone sensors and drone operating environment data.

3. The method for assessing the health status of a drone based on transfer learning according to claim 1, characterized in that: In step S2.5, the setting ratio is 8:

2.

4. The method for assessing the health status of a drone based on transfer learning according to claim 1, characterized in that: In step S3.4, the loss function includes but is not limited to a cross entropy function, a mean square error, and an exponential loss.

5. A UAV health status assessment system based on transfer learning, characterized in that: The system is used to implement the method according to any one of claims 1 to 4, and the system includes: The data acquisition and processing module is used to collect the UAV test data and simulation test data, standardize and normalize the data, and construct the test data set and simulation data set; The drone health status assessment model establishment and training module, based on the drone fault type and the input variables of the drone health status assessment model, establishes a drone health status assessment model that uses the long short-term memory network LSTM for feature extraction and the fully connected network and Softmax layer for classification; The UAV health status assessment model training module trains and tests the UAV health status assessment model based on the data set established by the data acquisition and processing module; The drone health status assessment module uses the trained drone health status assessment model to assess the drone health status.

Images