Prediction Method for Decline Cycle Based on Aerial Sensor Sequence Data

Abstract

The present invention discloses a method for predicting the decline cycle based on aviation sensor sequence data, belonging to the fields of machine learning, data analysis technology, and on-site security. This method first processes the original aviation sensor sequence data; generates a sequence data set and an image data set; inputs the image and the sequence data set for decline cycle prediction; the method proposed by the present invention is oriented to the aviation flight test scenario, and this method completes the goal of predicting the sensor decline cycle through the architecture of the cascaded convolutional-recurrent neural network. Due to the adoption of a brand-new network architecture and special algorithm design for the characteristics of the cycle prediction task, a better prediction effect can be finally achieved, improving the intelligence and accuracy of the method in the flight test scenario.

Description

[0001] Field of the Invention

[0002] The present invention belongs to the fields of machine learning, data analysis technology, and on-site security, and specifically relates to a decline cycle prediction method based on aviation sensor sequence data. Background Art

[0003] With the development of the aviation industry, in order to meet the development needs of the aircraft manufacturing industry, improve the level of flight test services, adapt to the characteristics of multiple test indicators, various types, heavy tasks, and high labor costs in the new era of aircraft testing, and conform to the development strategy of intelligent manufacturing, integrating artificial intelligence with aircraft test and measurement, processing and analyzing the sensor sequence data during the aircraft flight test process, researching and developing an intelligent sensor decline cycle prediction system, and conducting aviation early warning and ensuring low-altitude safety are important issues in the intelligent development process of the aviation industry. The decline of aviation sensors means that the data acquisition accuracy of aircraft in the flight test environment will gradually decline with the number of flights until anomalies occur, and calibration is required when the sensors show anomalies. The decline cycle refers to the number of flights required for the sensor to change from the current state to an abnormal state. Cycle prediction refers to predicting the time when the sensor shows anomalies, that is, the number of flights required until the sensor needs to be calibrated. By predicting the decline cycle of the sensor, aviation early warning can be achieved, flight test safety can be ensured, and resources can be saved.

[0004] Sensor sequence data has temporal and spatial correlations, and various types of sensor data involve multimodal cognitive computing. Therefore, its decline cycle prediction belongs to the problem of multi-class time series prediction. Traditional time series prediction methods, such as the ARIMA model and the Holt-Winters seasonal method, have theoretical guarantees. They are mainly used for univariate prediction problems and require the time series to be stationary, which greatly limits their application in complex time series data in the real world. With the increase in data volume and computing power, and the continuous breakthrough of algorithm models, it has also become possible to realize the intelligent prediction of the flight test process by comprehensively applying artificial intelligence technology and data analysis and processing technology, and combining multidisciplinary knowledge such as key aviation data technology and intelligent sensor technology. In the field of deep learning, there are mainly four methods for time series modeling: (1) based on convolutional neural networks; (2) based on recurrent neural networks; (3) based on temporal convolutional networks, such as WaveNet proposed by Oord et al. in the literature “A. Oord, S. Dieleman, H. Zen, K. Simonyan, O. Vinyals, A. Graves, N. Kalchbrenner, A. Senior, and K. Kavukcuoglu, ‘WaveNet: A Generative Model for Raw Audio,’ arXiv preprint, arXiv:1609.03499, 2016.”; (4) based on attention mechanisms, such as Informer proposed by Zhou et al. in the literature “H. Zhou, S. Zhang, J. Peng, S. Zhang, J. Li, H. Xiong, and W. Zhang, ‘Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting,’ AAAI, 2021, pp. 11106-11115.”

[0005] Due to problems such as large data volume, diverse types, and complex structures of aviation sensors, it is difficult to find a suitable prediction method to achieve the decline prediction of all sensors. Although previous research methods are cycle predictions, the research objects are not sequence data based on aviation sensors. Summary of the Invention

[0006] The present invention proposes a method for predicting the decline cycle based on aviation sensor sequence data. For the aviation flight test scenario, this method achieves the goal of predicting the decline cycle of sensors through a cascaded architecture of a convolutional-recurrent neural network. Due to the adoption of a brand-new network architecture and special algorithm design for the characteristics of the cycle prediction task, it can ultimately achieve a good prediction effect, improving the intelligence and accuracy of the method in the flight test scenario.

[0007] The technical solution adopted by the present invention to solve its technical problems includes the following steps:

[0008] A method for predicting the decline cycle based on aviation sensor sequence data, including the following steps:

[0009] Step 1: Process the original aviation sensor sequence data;

[0010] Input the original text sequence data of the aviation sensor, perform data cleaning, eliminate invalid data, and then perform data analysis on the cleaned data to test the correlation between the data, realizing data preprocessing;

[0011] Step 2: Generate a sequence data set and an image data set from the processed data;

[0012] Step 2-1: Perform interval sampling on the text sequence data to obtain sequence data with a fixed length of 1024, and set corresponding decline cycle labels according to the flight characteristics and zero position offset degree of the data, and collect these data into a sequence data set; among them, each text sequence data contains two columns, one column is the flight time of the sensor, and the other column represents the sensor value at the corresponding moment;

[0013] Step 2-2: Perform image conversion on the sequence data with a fixed length; that is, plot the corresponding data points in the sequence data into the corresponding images, where the horizontal axis represents the sequence time and the vertical axis corresponds to the value; the generated images inherit the original labels and are collected into an image data set;

[0014] Step 3: Input the image and the sequence data set for decline cycle prediction;

[0015] Step 3-1: Input the image data. First, perform preprocessing operations of size cropping (adjust the input image size to 224×224) and normalization on the image, and then input it into a simplified VGG (Visual Geometry Group) network for feature extraction to obtain the image data representation;

[0016] Step 3-2: Divide the image representation into 32 equal parts along the horizontal axis, convert it into a sequence data with a corresponding length of 32, input it into a set 5-layer LSTM (Long Short-Term Memory) network for feature extraction to obtain sequence data representation, and perform periodic prediction in the fully connected layer.

[0017] Step 4: Define the loss function;

[0018] Use the mean square error loss, and its formula is as follows:

[0019]

[0020] Where Loss is the mean square error loss, N represents the number of input images, represents the predicted period value, represents the true label value, j represents the jth input image, represents the Euclidean metric.

[0021] Preferably, the simplified VGG network only adopts the first 6 layers of the VGG16 network.

[0022] Advantages of the present invention: The proposed convolutional-recurrent neural network cascade architecture can achieve accurate prediction accuracy on the generated aviation sensor dataset. Brief Description of the Drawings

[0023] In order to more clearly illustrate the technical solutions implemented by the present invention, the following will briefly introduce each module required in the description of the embodiments. Obviously, the drawings in the following description include the flow chart and the cascade network framework diagram of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on this drawing and extended.

[0024] Figure 1 Shows the working flow chart of the decline cycle prediction method based on aviation sensor sequence data of the present invention.

[0025] Figure 2 Shows the specific network architecture of the convolutional-recurrent neural network cascade of the present invention.

[0026] Figure 3 Is the VGG network structure designed in the embodiment.

[0027] Figure 4 Is the multi-layer LSTM network structure in the embodiment. Detailed Embodiments

[0028] The present invention will be further described below in conjunction with the drawings and embodiments.

[0029] The present invention proposes a learning method based on cascaded convolutional-recurrent neural networks, which respectively uses a simplified VGG network to extract features from image data and a multi-layer LSTM network to extract features from sequence data, so as to realize the prediction of the sensor degradation cycle in the aviation flight test scenario.

[0030] A method for predicting the degradation cycle based on aviation sensor sequence data is realized as follows:

[0031] 1. Data processing. Input the text sequence data of flight records, clean the text sequence data, process invalid values and missing values, etc., clean up abnormal data, and conduct review and verification. Then, conduct data analysis on the cleaned data to test the correlation between data.

[0032] 2. Sample the text sequence data to obtain sequence data with a fixed length of 1024, which is aggregated into a sequence data set, and divide the labels according to their data characteristics.

[0033] The specific operation is as follows: First, the text flight record data of a sensors can be expressed as X T See formula (2), where represents the data value of sensor i during the T-th flight, see formula (3), and n is the length of the sensor data. Substitute (3) into (2), and x i,j represents the j-th value of sensor i, see formula (4).

[0034]

[0035]

[0036]

[0037] Secondly, for each flight, let the label be Y T See formula (5), where y i represents the degradation cycle of sensor i after the T-th flight. The degradation cycle presented by each flight record is the same, indicating that the sensor is expected to experience y i flights before an anomaly occurs. Pre in formula (6) is the degradation cycle Y corresponding to different flight records X.

[0038]

[0039]

[0040] Finally, resample the original sequence to a fixed length and encode it into multiple fixed-length sequences, with a length of 1024, that is, sample the X T numerical values at equal intervals to obtain Refer to Equation (7) and transfer the previous mapping relationship labels from X to Y to these sequences, i.e., Pre T Refer to Equation (8).

[0041]

[0042]

[0043] 3. Plot the corresponding data points in the sequence data into the corresponding images, where the horizontal axis represents the sequence time and the vertical axis corresponds to the values. The generated images inherit the original labels and are grouped into an image dataset.

[0044] 4. Input the image data. First, preprocess the image, including size cropping and normalization. The size cropping is to adjust the input image to 224×224. Then input it into the VGG network pre-trained on the ImageNet dataset, and perform feature extraction through convolution to obtain the image data representation. Due to the limited amount of data in the dataset, the VGG network we set only adopts the first 6 layers of the VGG16 network. The VGG16 network was proposed by Simonyan et al. in the literature "K. Simonyan, and A. Zisserman, 'Very Deep Convolutional Networks for Large-Scale Image Recognition,' in Proc. International Conference on Learning Representations, 2015."

[0045] The set VGG network structure is as Figure 3 shown, with a total of 6 layers, manifested as convolution - convolution - max pooling - convolution - convolution - max pooling - convolution - convolution - average pooling, where the convolution kernels of the convolution layers are 3×3 and the number of channels are 64, 128, and 256 respectively.

[0046] Perform equal division on the image data representation, convert it into sequence data, input it into the set multi-layer LSTM network for feature extraction to obtain the sequence data representation, and perform periodic prediction in the fully connected layer (FC). The network structure is as Figure 4As shown, the number of channels is 256, 128, 64 until the output. Among them, the LSTM network was proposed by Hochreiter et al. in the literature "S. Hochreiter and J. Schmidhuber, 'Long Short-Term Memory,' Neural Computation, vol. 9, no. 8, pp. 1735-1780, 1997." The Dropout layer is a regularization method for neural network models proposed by Srivastava et al. in the literature "N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever and R. Salakhutdinov, 'Dropout: A Simple Way to Prevent Neural Networks from Overfitting', J. Mach. Learn. Res, vol. 15, no. 1, pp. 1929-1958, 2014."

[0047] 5. Definition of the loss function. The mean squared error loss is used, and its formula is as follows:

[0048]

[0049] Where Loss is the mean squared error loss, N represents the number of input images, represents the predicted value, represents the true label value, j represents the j-th input image, represents the Euclidean metric.

[0050] The effects of the present invention can be further illustrated by the following experimental results.

[0051] 1. Experimental environment and settings

[0052] The present invention is run on an operation with Intel(R) Core(TM) i7-11700F @ 2.50GHz, 16.0GB RAM.

[0053] The data used in the experiment is the simulation flight test data of a certain civil aircraft model. In this experiment, 557 pieces of data were collected, and an aviation sensor sequence and image dataset were constructed, as shown in Table 1. Each piece of data is the simulation flight data value of the sensor within a certain period of time. The labels of the training set and the test set are calculated based on the zero position offset during the sensor calibration period to judge the decay period of the sensor.

[0054] Table 1 Aviation sensor dataset

[0055]

[0056] 2. Evaluation Metrics

[0057] The evaluation metrics commonly used in regression tasks, MAE and RMSE, are adopted in this experiment.

[0058] MAE (Mean Absolute Error):

[0059]

[0060] RMSE (Root Mean Square Error):

[0061]

[0062] where m represents the number of samples in the test set, y i is the periodic prediction value, and is the true value.

[0063] 3. Experimental Content

[0064] First, train the convolutional-recurrent neural network cascade model on the training set; then, use the cascade model to test on the test set and calculate the degradation period of the sensor. As shown in Table 2, compared with the VGG16 network proposed by Simonyan et al. in the literature "K. Simonyan, and A. Zisserman, 'Very Deep Convolutional Networks for Large-Scale Image Recognition,' in Proc. International Conference on Learning Representations, 2015." and the AlexNet proposed by Alex et al. in the literature "A. Krizhevsky, I. Sutskever, and G. Hinton, 'ImageNet Classification with Deep Convolutional Neural Networks,' NIPS 2012, pp. 1106 - 1114.", the MAE predicted by our network is lower and the effect is better.

[0065] Table 2 Comparison of Experimental Results

[0066]

[0067] To prove the effectiveness of our model, an ablation experiment was conducted on the VGG+LSTM architecture. As shown in Table 3, only the designed VGG network was used to test on this dataset to prove the effect of this model architecture.

[0068] Table 3 Comparison of ablation experiment results

[0069]

Claims

1. A decline cycle prediction method based on aviation sensor sequence data, characterized in that, it includes the following steps: Step 1: Process the original aviation sensor sequence data; Input the original text sequence data of the aviation sensor, perform data cleaning, eliminate invalid data, and then perform data analysis on the cleaned data to test the correlation between the data, realizing data preprocessing; Step 2: Generate a sequence data set and an image data set from the processed data; Step 2-1: Perform interval sampling on the text sequence data to obtain sequence data, and set corresponding decline cycle labels according to the flight characteristics and zero position offset degree of the data, and collect these data into a sequence data set; among them, each text sequence data contains two columns, one column is the flight time of the sensor, and the other column represents the sensor value at the corresponding moment; Step 2-2: Perform image conversion on the sequence data of a fixed length; that is, plot the corresponding data points in the sequence data into the corresponding images, where the horizontal axis represents the sequence time and the vertical axis corresponds to the value; the generated images inherit the original labels and are collected into an image data set; Step 3: Input the image and the sequence data set for decline cycle prediction; Step 3-1: Input the image data, first perform preprocessing operations of size cropping and normalization on the image, and then input it into a simplified VGG (Visual Geometry Group) network for feature extraction to obtain an image data representation; Step 3-2: Divide the image representation into 32 equal parts along the horizontal axis, convert it into sequence data with a corresponding length of 32, input it into a set 5-layer LSTM (Long Short-Term Memory) network for feature extraction to obtain a sequence data representation, and perform cycle prediction in the fully connected layer; Step 4: Define a loss function; Use the mean square error loss, and its formula is as follows: where Loss is the mean squared error loss, N represents the number of input images, represents the predicted periodic value, represents the true label value, j represents the j-th input image, represents the Euclidean metric.

2. A decline cycle prediction method based on aviation sensor sequence data as described in claim 1, characterized in that, the simplified VGG network only adopts the first 6 layers of the VGG16 network.

Images