Lung Sound Signal Feature Extraction And Pattern Recognition

Lung sounds is a kind of acoustic signal in the ventilation process of human respiratory system with external environment, and the correlation study with pathology and physiology has become an important theme in lung sounds study and clinic medicine. However, because of its randomicity, the non-integrality of recording equipment and the diversity of analysis methods, the lung sounds study results are greatly different. With the development of computer and signal processing technique, lung sounds recognition has become a new focus and it is worth paying more and more attention. Feature extraction is the crucial work for lung sounds pattern recognition. The extracting process is to throw off the useless information and look for the most efficient signal feature to form a pattern feature vectors for classification with mathematics tools. For a steady lung sounds analysis system, its transfer function means unchange and the corresponding amplitude-frequency and phase-frequency characteristics are fixed. So the sampled data can reflect the physical characteristic of signal to some degree.It is found that lung sound signals are composed of multi-frequency components and take on obvious periodic wave. Normal and abnormal lung sounds are not the same in energy distribution. Different energy distribution means different signal characteristics. So the differences among every frequency components may be used to distinguish every distinct signals. But it is hard to extract these characteristics with conventional methods (STFT, for example). So a novel method of non-stationary lung sounds recognition based on signal energy is put forward in this paper.Multi-resolution analysis (MRA) of wavelet is a suitable technique for analyzing the non-stationary lung sounds. Wavelet analysis is filtering of original signal in nature (a series of constant-Q band-pass filters). The signal is mapped to a series of subspaces formed of mutual orthogonal wavelet functions and is stretched out at different scales. In the view of frequency spectrum analysis, wavelet transform(WT) is to decompose the signal to low and high frequency components; next, the low frequency part isdecomposed to lower frequency and higher frequency part and then further similar decompositions. It is chiefly due to the "adaptive feature" and "mathematical microscope feature", WT is becoming a focus of many sciences. It plays an important role in the signal & information processing.After preprocessing of collected lung sounds data firstly, four kinds of typical lung sound signals (normal, tracheitis, pneumonia and asthma) are sampled from various subjects. By studying the time-frequency distribution characteristic of lung sounds, we select the wavelet that has the trait of arbitrary distinction and decomposition. After space partition of wavelet, the best decomposition level for feature extraction is affirmed. Then we do fast multi-scale WT and obtain each high-dimensional wavelet coefficients matrix. And the equivalence relation in time domain between wavelet coefficients and signal energy is founded. The energy is used as eigenvalue and feature vectors of artificial neural network (ANN) for classification are formed. Then the original high-dimensional wavelet coefficients space is transferred to energy feature space, which greatly decreases the number of input vectors of ANN and simplifies the structure of classification neural network, and the recognition rate reaches 89.75%.In order to produce a stable and effective pattern library, this paper adopts learning mechanism, which makes further improvement on the recognition effect.What's more, the software and hardware system of "Analysis and Recognition of Lung Sounds" are generated. It is proved that the feature extraction method proposed in this paper has the advantage of robust, anti-disturbance and anti-noise. It is highly efficient and has good stability and reliability to the experiment.