The Research Of Motor Abnormal Sound Detection Based On Machine Learning

Modern industrial production and household appliances are inseparable from all kinds of motor. People attach importance to the motor performance and hope to reduce the noise generated by the motor rotation. At present, a large number of small motors in production line are inspected by trained workers for abnormal sound detection. A lot of dull, repetitive listening tasks causes auditory fatigue and influence subjective judgment, lead to abnormal sound motor with normal flow into the market, and the company’s economy and reputation will cause irreparable damage. Therefore, the development of motor industry has great significance to achieve automatic detection of abnormal sound motor in production line.For the statistical characteristics of the motor sound signal and artificial detection, use acoustic sensor technology instead of the human ear to collecting motor sound signal. This way of non-contact measurement is conform to the requirements of the production line’s simple and efficient test equipment. According to the human ear hearing characteristics to acquisition motor’s acoustic signal and analyze the spectrum of the motor. Because the human ear is not sensitive to the phase, only the amplitude spectrum is required to analyze. In order to highlight the characteristics of absolute difference, using principal component analysis of motor’s acoustic signal to data compression and reduce the dimensions. Consider a possible non-stationary signal in running motor, introducing the wavelet transform to motor sound detector and in a more accurate analysis of the time-frequency characteristics. Using the wavelet packet decomposition to get signal spectrum coefficient, according to the eigenvector of singular value decomposition extracting its singular value and mapped it to the feature vector generated by the state space, to achieve the motor signal feature extraction. Meanwhile, make the sound signal to perform wavelet packet decomposition to the mutually orthogonal frequency band. It is no loss of energy and contains a wealth of characteristic information. The motor characteristics are mapped to the subspace energy distribution and make the normalized energy to build feature matrix to achieve different sound feature extraction. Taking into account the less abnormal sound samples in production line, access it difficulties, have individual differences and other issues caused the abnormal sound difficult to analyze, and the formation of motor abnormal sound are abnormal complex, so a new machine learning methods like support vector machine for one class learning just enough to achieve the detection of different sound. In this method, a sample of normal motor foundation established discriminant function quality does not require different sound samples, avoiding the other classification algorithms require training class comprehensive and cover a wide range of sample conditions. Finally, through a lot of normal motor samples training, and then with abnormal sound motor samples to validation, experiments on the recognition rate of abnormal sound of the motor to meet the needs of the factory and achieved the desired goals.