In modern manufacturing systems and industry, a growing number of researchers dedicated to developing efficient machine health monitoring system. In a variety of machine health monitoring methods, data-driven approach due to the advanced sensing and data analysis techniques and the development of more and more popular. However, taking into account the noise behind the sensory data, and irregularly sampled variable length, such sequences can not be directly input to the data classification and regression models. Thus, previous work focused on feature extraction / fusion method requires a lot of manpower and high-quality expertise. In recent years, with the development of deep learning approach, redefining the method to characterize the raw data from the study, we designed a convolution of two-way short and long term memory network (cblstm) to deal with primitive sense data. Firstly cnn cblstm robust local feature extraction and information from a sequence of input. Then, the introduction of two-way lstm time information is encoded. Long - short-term memory network (lstms) able to capture long-term dependency model and sequence data, two-way structure is able to capture the context of past and future. On the basis of two-way lstms, overlays were established, full connectivity layer and a linear correlation to predict the target layer. Here, presented a real tool wear test, CBLSTM we proposed to predict the actual tool wear based on the raw sensory data. Experimental results show that our model can better than several of the latest baseline method.
Time-series data on the sensor data acquired by the sensor is essentially, and is expressed in the form of a sequence. Previous work has focused on a multi-domain feature extraction, including statistics (variance, skewness, kurtosis), the frequency (spectral skewness) and time-frequency (wavelet coefficient) characteristics.
lstm as a neural network model representing the learning and training together, no additional domain knowledge. In addition, this structure can find some invisible structure, improve the generalization ability of the model. In addition to the need for the time information, the original feeling data generally contains noise. Built on the basis of raw sensory data lstm model may not be reliable. Thus, the introduction of a convolutional neural network (CNN) for extracting local feature, the core idea is to extract features by convolution kernel abstract and cooling operations. In cnn, the convolution layer (convolution) of the original data sequence into a plurality of partial convolution filters, generating invariant local feature, then the cell layer was extracted most important characteristics in a sliding window of a fixed length. Here, we first use cnn local features extracted from the range of the original signal.
This article will cnn combined with two-way lstm propose a new machine health monitoring system, called convolution lstm two-way network (cblstms). In cblstms we proposed, cnn robust feature can be extracted locally, but based on two-way lstms cnn can encode time information and learning representation. Different processing a feedforward manner conventional input sequence lstms, lstms bidirectional input sequence of forward and reverse model [22]. The core idea behind the two-way lstm is that each sequence forward and backward presented as two separate lstm, and two-way lstm access to all information on the complete sequence context information before and after the sequence in each time step given. Here, we use an open-source data set: dynamometer, accelerometer and acoustic data acquisition task from the high speed milling machine tool computer numerical control (the CNC) (Visit https://www.phmsociety.org/competition/phm / 10). Tool wear conditions defined based on the respective tasks sensory signal (i.e., tool wear depth) Estimation [23,24]. In our setup, the problem is converted to a data sequence having a regression problem, wherein each of the data sequence, i.e. sensory data indicating wear width corresponding to the actual tool to a particular tool wear conditions. Some of the latest models were compared with cblstms our proposed model.
Compared to the original input sequence, cnn capable of encoding more critical information.
cblstms proposed mainly composed of two parts: one part is the local feature extractor cnn, the other is a bidirectional time encoder lstms. After applying a layer cnn extracts local features and discriminant features on the original input sequence, to establish two bi lstms based on the previous layer cnn time coding mode. Then, the two laminated fully dense in the output communication process lstms together. Finally, the linear regression prediction tool wear layer depth