Infrasound event identification method based on self-attention fusion

Abstract

The invention relates to the technical field of infrasound signal data processing, and discloses an infrasound event identification method based on self-attention fusion, and the method comprises the following steps: building a first layer of a model, and employing a wide kernel convolution layer for the first layer of the model; after preliminary feature extraction, data are input into a dense connection convolutional layer composed of a third layer and a fourth layer, and information contained in signals is deeply mined; setting a fifth layer as a BiLSTM layer, and fully extracting forward and reverse sequence features of the signal; a self-attention mechanism is introduced to capture key features of the signals; the infrasound signals are classified and identified by the full connection layer, and the output size is the category number. According to the infrasound event identification method based on self-attention fusion, the provided Dense-BiLSTM-SA fusion model carries out four classifications on a classic infrasound database LOTIS, so that the classification precision is higher than that of a traditional single neural network, and under the interference of high noise, the good classification precision can still be achieved, and the higher anti-noise capability is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of infrasound signal data processing, and in particular to an infrasound event recognition method based on self-attention fusion. Background Art

[0002] In recent years, natural disasters have occurred frequently, posing a huge threat to the safety of life and property of the people. Therefore, it is of great significance to improve the early warning and monitoring capabilities of natural disaster events. Infrasound signals are widely present in the natural environment and human society. Many natural phenomena and human activities are accompanied by the generation of low-frequency infrasound signals. Unlike audible sound, the frequency of infrasound signals is between 0.01Hz-20Hz, which is a signal that cannot be heard by the human ear. Infrasound has the characteristics of long wavelength, low frequency, easy diffraction, and low attenuation in the medium. It is very suitable for long-distance event detection, which makes infrasound sensing technology a new type of long-distance detection method. Therefore, through the monitoring and analysis of infrasound signals, it plays an important role in geological disaster early warning, chemical explosion monitoring, atmospheric environment governance, mudslides, typhoons, etc.

[0003] In terms of target monitoring, since aircraft engines generate broadband noise, which contains a large part of infrasound, the collection and analysis of infrasound has become a new means of long-distance target detection, positioning, and identification of aircraft. Given the key role and development potential of infrasound in many fields, conducting research in related fields and in-depth analysis of infrasound events have also become important development directions in recent years, which require researchers to focus on research and further promote related applications. Among them, the key to infrasound research has always been to use technical means to judge and extract ideal characteristic parameters and distinguish infrasound events. However, infrasound is easily affected by the environment during atmospheric propagation, and the collected signals contain a lot of noise. Existing methods cannot use network structures to reduce the impact of noise on the model. Summary of the invention

[0004] The purpose of the present invention is to provide a method for infrasound event recognition based on self-attention fusion to solve the problems raised in the above background technology.

[0005] To achieve the above object, the present invention provides the following technical solution: a method for infrasound event recognition based on self-attention fusion, the method comprising the following steps:

[0006] S1. First, the first layer of the model is established. The first layer of the model uses a wide kernel convolution layer. The feature dimension is reduced by the maximum pooling layer to remove redundant information and obtain more information.

[0007] S2. After preliminary feature extraction, the data is input into the densely connected convolutional layer consisting of the third and fourth layers to deeply mine the information contained in the signal, learn high-level features and strengthen information flow to achieve noise suppression;

[0008] S3. Since the convolution layer is not sensitive to the temporal characteristics of the data, the fifth layer is set as the BiLSTM layer to fully extract the forward and reverse sequence characteristics of the signal, and at the same time have the past and future characteristic information of the infrasound signal;

[0009] S4, introduce the self-attention mechanism to capture the key features of the signal. At the same time, the combination of the self-attention mechanism and BiLSTM solves the problem that the latter cannot be calculated in parallel, thereby further improving the performance of the model. After this step, the data output size is 1×512;

[0010] S5. Finally, the fully connected layer classifies and recognizes the infrasound signal, and the output size is the number of categories.

[0011] Optionally, when the model uses LOTIS raw data as model input, the proposed Dense-BiLSTM-SA, CNN, AlexNet, and BiLSTM all obtain classification accuracy.

[0012] Optionally, noise is added to the original data so that the signal-to-noise ratio of the infrasound signal is 10 dB, 5 dB, 0 dB and -5 dB respectively, and the classification performance of the above various networks is compared and analyzed using the same training set and test set.

[0013] Optionally, the convolution kernel size of the wide-core convolution layer is 32, the step size is 1, and ReLU is used as the activation function.

[0014] Optionally, the kernel size of the maximum pooling layer is 2, and the maximum pooling layer convolves 1×6000 data input to 16×2984 to increase the receptive field.

[0015] Optionally, the densely connected convolution layer adopts three layers of convolution, the kernel size of each convolution layer is 3, the step size is 1, the transition block adopts ReLU as the activation function, the kernel size is 1 convolution layer, the average pooling layer kernel size is 2, and the data output size is 128×1492.

[0016] Optionally, the number of neurons in the BiLSTM layer is set to 256, and the data needs to be dimensionally transformed before being input into the BiLSTM, and tanh is used for activation after output, and the data output size is 1492×512.

[0017] Compared with the prior art, the present invention provides an infrasound event recognition method based on self-attention fusion, which has the following beneficial effects:

[0018] 1. The proposed infrasound event recognition method based on self-attention fusion and the proposed Dense-BiLSTM-SA fusion model can achieve higher classification accuracy than the traditional single neural network in four-classification on the classic infrasound database LOTIS. It can still achieve good classification accuracy under high noise interference and has higher noise resistance.

[0019] 2. The infrasound event recognition method based on self-attention fusion is used to identify three types of infrasound signals on actual array data. The fusion model proposed in this paper has higher classification accuracy than AlexNet, which performs well on the LOTIS database, CNN-BiLSTM, which combines traditional convolutional layers and bidirectional long short-term memory networks, and CNN-BiLSTM-SA with added self-attention mechanism. It verifies the effectiveness of feature extraction of signals by each part and further improves the classification accuracy.

[0020] 3. This infrasound event recognition method based on self-attention fusion collects infrasound signals generated by aircraft on the spot, performs simple preprocessing on the signals, and uses the proposed model to identify events with the actual data of other events, which can obtain a higher classification accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 It is a schematic diagram of the process of the present invention;

[0022] Figure 2 Schematic diagram of the original signal and the noise-added signal of the present invention;

[0023] Figure 3 This is a schematic diagram of the classification results of the LOTIS raw data of the present invention;

[0024] Figure 4 This is a schematic diagram of the classification results of LOTIS noisy data of the present invention;

[0025] Figure 5 This is a schematic diagram of the classification results of the actual array data of the present invention;

[0026] Figure 6 This is a schematic diagram of the classification results obtained by the present invention. DETAILED DESCRIPTION

[0027] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0028] like Figure 1-Figure 6As shown, the present invention provides a technical solution: a method for infrasound event recognition based on self-attention fusion, the method comprising the following steps:

[0029] S1. The first layer of the model is established. The first layer of the model uses a wide kernel convolution layer. The feature dimension is reduced by the maximum pooling layer to remove redundant information and obtain more information. When the model uses LOTIS original data as the model input, the proposed Dense-BiLSTM-SA and CNN, AlexNet, and BiLSTM all obtain classification accuracy. Noise is added to the original data so that the signal-to-noise ratio of the infrasound signal is 10dB, 5dB, 0dB, and -5dB respectively. The classification performance of the above-mentioned various networks is compared and analyzed using the same training set and test set;

[0030] The convolution kernel size of the wide kernel convolution layer is 32, the step size is 1, and ReLU is used as the activation function.

[0031] S2. After preliminary feature extraction, the data is input into the densely connected convolutional layer composed of the third and fourth layers to deeply mine the information contained in the signal, learn high-level features and strengthen information flow to suppress noise. The kernel size of the maximum pooling layer is 2. The maximum pooling layer convolves the 1×6000 data input to 16×2984 to increase the receptive field.

[0032] S3. Since the convolution layer is insensitive to the temporal characteristics of the data, the fifth layer is set as a BiLSTM layer to fully extract the forward and reverse sequence characteristics of the signal, and at the same time have the past and future feature information of the infrasound signal. The densely connected convolution layer adopts three layers of convolution, and the kernel size of each convolution layer is 3, the step size is 1, and the transition block uses ReLU as the activation function, the kernel size is 1 convolution layer, the average pooling layer kernel size is 2, and the data output size is 128×1492.

[0033] S4. The self-attention mechanism is introduced to capture the key features of the signal. At the same time, the self-attention mechanism is combined with BiLSTM to solve the problem that the latter cannot be calculated in parallel, thereby further improving the performance of the model. After this step, the data output size is 1×512, and the number of neurons in the BiLSTM layer is set to 256. The data needs to be transformed before being input into BiLSTM, and tanh is used for activation after output. The data output size is 1492×512.

[0034] S5. Finally, the fully connected layer classifies and recognizes the infrasound signal, and the output size is the number of categories.

[0035] The model effect was verified by using the library of classic infrasound signals (LOTIS) for four-classification. Since the signal-to-noise ratio in the database is very high, the classification effect of the model before and after data noise addition was compared to verify the denoising effect of the model, and finally the network was further verified using the actual collected infrasound signal.

[0036] When using LOTIS original data as model input, the proposed Dense-BiLSTM-SA can obtain good classification accuracy with CNN, AlexNet, and BiLSTM, and the proposed network is close to CNN in accuracy. Due to the small number of test data sets, it can be approximately considered that they have the same classification accuracy. After that, we added noise to the original data so that the signal-to-noise ratio of the infrasound signal was 10dB, 5dB, 0dB, and -5dB, respectively. The classification performance of the above-mentioned various networks was compared and analyzed using the same training set and test set. As can be seen from the figure, when the noise mixed in the data set continues to increase, the classification accuracy of all models has decreased, proving that noise has an impact on the accuracy of infrasound signal classification. By comparing the four different network models, it can be seen that when the signal-to-noise ratio is 10dB, the proposed network structure shows strong noise resistance, and the classification accuracy can still reach 94.17%, while the CNN, AlexNet, and BiLSTM networks have poor noise resistance. However, under this signal-to-noise ratio, the four types of networks can still achieve relatively accurate classification, and the classification accuracy is above 85%. As the signal-to-noise ratio gradually decreases, the accuracy of the comparison method begins to decline significantly, and is only about 70% at signal-to-noise ratios of 5dB and 0dB, while the accuracy of the proposed network structure at the two signal-to-noise ratios can reach 89.32% and 85.44% respectively, and the classification accuracy gap between the models is widened. When the signal-to-noise ratio is reduced to -5dB, the classification results of each model show a serious decline, indicating that strong noise interference has seriously affected the classification accuracy of the model. Compared with the other three types of networks, the proposed network model still has a higher classification accuracy. This result shows that the proposed model is superior in classifying infrasound signals under noise interference, and has strong noise resistance and stability.

[0037] In order to verify the performance of the proposed Dense-BiLSTM-SA network in the actual infrasound event classification work, the classification results of the proposed network are compared with other networks by using the same training set and test set. The comparison network uses AlexNet, CNN-BiLSTM fusion model and CNN-BiLSTM-SA fusion model that perform well on LOTIS. By comparing the classification results of each model in the figure, it can be seen that the proposed network model has the highest classification accuracy, reaching 88.67%. Compared with the other three types of fusion models (a, c, d), the classification accuracy obtained by AlexNet model (b) is the worst, indicating that the fusion neural network of convolutional neural network and BiLSTM network has advantages over a single neural network in classifying infrasound events. The classification results of CNN-BiLSTM (c) and CNN-BiLSTM-SA (d) show the effectiveness of the self-attention mechanism in the classification of infrasound events. By comparing Dense-BiLSTM-SA (a) and CNN-BiLSTM-SA (d), we can see that the dense convolutional layer in the proposed network model can more effectively extract the features of the infrasound signal than the traditional convolutional layer and achieve more accurate classification.

[0038] Since infrasound monitoring has many advantages, and aircraft engines generate a lot of broadband noise due to various factors, it is suitable to use infrasound equipment to collect the infrasound part of the noise and perform target detection on it. This experiment introduces the field collection of civil aviation engine and a certain type of military engine noise experiment. Through the experiment, the infrasound signal of the aircraft engine is obtained, and on this basis, the target characteristic identification research is carried out to form a preliminary understanding of the characteristics of the aircraft engine.

[0039] Intercept and process the aircraft infrasound signals collected on site to ensure the integrity of the effective parts;

[0040] The collected original infrasound signals are preprocessed by detrending, filtering, etc. to preliminarily eliminate the interference of some noises;

[0041] The intercepted data is divided into training set and test set in a ratio of 7:3;

[0042] Use the training set to perform classification and recognition training on the algorithm model;

[0043] Use the training set to perform classification and recognition training on the algorithm model;

[0044] The data is shown in the following table:

[0045]

[0046]

[0047] From the confusion matrix of the classification results, it can be seen that the infrasound signals emitted by aircraft can be effectively distinguished from other types of infrasound signals, and the classification accuracy can reach 93%. Therefore, the infrasound emitted by aircraft can be used as a new means of target monitoring and identification of aircraft, and the proposed identification method can achieve more effective event identification compared with traditional methods.

[0048] The above generally describes the present invention in detail, but it is obvious to a person skilled in the art that some modifications or improvements can be made to the present invention. Therefore, modifications or improvements that do not depart from the spirit of the present invention are within the scope of protection of the present invention.

Claims

1. A method for infrasound event recognition based on self-attention fusion, characterized in that: The method comprises the following steps: S1. First, the first layer of the model is established. The first layer of the model uses a wide kernel convolution layer. The feature dimension is reduced by the maximum pooling layer to remove redundant information and obtain more information. S2. After preliminary feature extraction, the data is input into the densely connected convolutional layer consisting of the third and fourth layers to deeply mine the information contained in the signal, learn high-level features and strengthen information flow to achieve noise suppression; S3. Since the convolution layer is not sensitive to the temporal characteristics of the data, the fifth layer is set as the BiLSTM layer to fully extract the forward and reverse sequence characteristics of the signal, and at the same time have the past and future characteristic information of the infrasound signal; S4, introduce the self-attention mechanism to capture the key features of the signal. At the same time, the combination of the self-attention mechanism and BiLSTM solves the problem that the latter cannot be calculated in parallel, thereby further improving the performance of the model. After this step, the data output size is 1×512; S5. Finally, the fully connected layer classifies and recognizes the infrasound signal, and the output size is the number of categories.

2. The infrasound event recognition method based on self-attention fusion according to claim 1, characterized in that: When the model uses LOTIS raw data as model input, the proposed Dense-BiLSTM-SA, CNN, AlexNet, and BiLSTM all obtain classification accuracy.

3. The infrasound event recognition method based on self-attention fusion according to claim 1, characterized in that: Noise was added to the original data so that the signal-to-noise ratios of the infrasound signal were 10 dB, 5 dB, 0 dB and -5 dB respectively. The classification performances of the above-mentioned various networks were compared and analyzed using the same training set and test set.

4. The infrasound event recognition method based on self-attention fusion according to claim 1, characterized in that: The convolution kernel size of the wide-core convolution layer is 32, the step size is 1, and ReLU is used as the activation function.

5. The infrasound event recognition method based on self-attention fusion according to claim 1, characterized in that: The kernel size of the maximum pooling layer is 2, and the maximum pooling layer convolves the 1×6000 data input to 16×2984 to increase the receptive field.

6. The infrasound event recognition method based on self-attention fusion according to claim 1, characterized in that: The densely connected convolution layer adopts three layers of convolution, the kernel size of each convolution layer is 3, the step size is 1, the transition block adopts ReLU as the activation function, the kernel size is 1 convolution layer, the average pooling layer kernel size is 2, and the data output size is 128×1492.

7. The infrasound event recognition method based on self-attention fusion according to claim 1, characterized in that: The number of neurons in the BiLSTM layer is set to 256. The data needs to be dimensionally transformed before being input into the BiLSTM, and tanh activation is used after output, and the data output size is 1492×512.

Images