An intelligent crash recognition method based on 1DResNet-SVM with distributed vibration sensors

Distributed Vibration Sensing (DVS) based on the phase-sensitive optical time-domain reflectometry (φ-OTDR) [1] have many advantages such as easy bending, small size, intrinsic safety, it can realize long-distance, distributed, and real-time monitoring of dynamic strain along optical fibers [2]. At present, it is widely used in the field of road traffic safety monitoring [3], [4]. However, the vibration signals of DVS are highly sensitive to the diverse environmental noise, which will lead to high fault alert rate. In addition, when the fiber is transmitted over long distances, the pulse width of the optical fiber will become wider due to the influence of dispersion effect [5], [6]. When the pulse width increases, the spatial resolution of the system will also increase. This will greatly affect the detection accuracy of the system [7], [8]. This limits the application of DVS in road safety monitoring [9]. Event identification capability has always been the bottleneck restricting its field application performance. Therefore, using effective pattern recognition algorithm to identify the type of vibration signal has become the research topics that have attracted wide attention to researchers [10]. The quality of the identification algorithm in the classification model directly determines the working efficiency of the sensing system [11].

Many scholars are committed to improving the DVS data processing methods to improve the efficiency of the system. At present, the data processing and recognition methods for DVS are mainly divided into machine learning and deep learning. For machine learning methods, in 2020, Chen et al. used Probabilistic Neural Network (PNN) to identify signals and vibration signals, and the false alarm rate could be reduced to 1% [12]. A feature extraction and classification method based on multi-scale Wavelet Packet Shannon Entropy was proposed to improve the identification accuracy of vibration signals in industrial applications [13]. As for deep learning scheme, using CNN to identify DVS signals is still the choice of many researchers. Such as, some scholars have proposed a CNN based classification method, which directly applies the spatio-temporal data matrix to Φ-OTDR system. The scheme has the advantages of small network size and high training speed [14]. At present, how to further improve the DVS data processing method and pattern recognition algorithm is still a difficult problem [15].

Feature extraction and identification algorithm are the two main challenges of pattern recognition algorithm. Among this, feature extraction and selection is the most important which is also the toughest part. Many researchers try to extract features of the signal manually [16]. Time domain characteristics [17], frequency domain characteristics [18] and multiscale wavelet features [19] are their main choices. They usually use some typical classification algorithms such as support vector machine (SVM), AdaBoostClassifier or neural network for classification after feature extraction.

In recent years, with the rapidly development of deep learning and its wide application in the field of image recognition, researchers proposed to use Convolutional Neural Network (CNN) to identify different DVS vibration signals [20]. CNN can adaptively extract features, thus combining feature extraction and classification into one part. They transform one-dimensional vibration signal into spectrum diagram [21] or Mel spectrogram [22] by data processing originally. Finally, the two-dimensional (2D) image is used as the input of the two-dimensional CNN structure [23]. However, these methods not only make the data processing more complicated but may also cause information loss during data processing. Therefore, many fiber vibration identification methods based on 1DCNN are widely used [24]. These methods do not require data transformation and directly input 1D vibration signals for identification. For example, Huijuan Wu et al. used 1DCNN and bidirectional long short-term memory (BiLSTM) network to identify signals of distributed fiber acoustic sensing (DAS) [25]. Wu Jun et al. proposed a multi-scale one-dimensional convolutional neural network (MS 1-D CNN) to identify three events [26]. But in practical applications, vibration signals are often disturbed by various noises, we always need a deep enough network to learn the features of the data better. However, model is prone to degeneration when convolutional layers is simply stacked of. The residual structure is introduced into Residual Neural Network (ResNet) creatively [27], and makes it possible to develop the model to a deeper level through skip connections between layers [28], [29]. In recent years, ResNet has gradually been widely used in the field of pattern recognition [30].

In this paper, in order to correctly identify the accident of vehicle crashing into a guardrail on highway, we combine 1D convolution with residual block for feature extraction, and finally we use SVM for classification to obtain higher recognition accuracy. The one-dimensional raw vibration signal collected by DVS system can achieve end-to-end classification and recognition without any preprocessing. The feature extraction ability of 1DResNet is evaluated by visual analysis. In order to obtain the highest recognition accuracy, e compared two typical machine learning classifiers: SVM, AdaBoostClassifier with the softmax layer in CNN. Finally, it is proved that the proposed method (1DResNet+SVM) can achieve better results in the application of crash recognition between vehicles and guardrail.

留言 (0)

沒有登入
gif