DEVELOPMENT OF ANOMALY DETECTOR FOR MOTOR BEARING CONDITION MONITORING USING FAST FOURIER TRANSFORM AND LONG SHORT TERM MEMORY (LSTM)-AUTOENCODER

Abstract

Motor bearings have over the years been one of the components which aids efficiency in industries as it helps smooth running of rotary machines in industries. However, it is important to note that the rate of degradation of motor bearings vary from one machine to another. This phenomenon is unavoidable but vary as a result of operational and environmental factors such as the time of operation of machines where installed, ambient temperature, load factor and maintenance ethics like constant oiling. For sustainability of machines, safe points of these factors have to be considered to avoid anomaly of bearings that could lead to high maintenance cost such as bearing faults, fatigue and accelerated aging or even complete breakdown which accounts for 30% to 40% failures of machines. With this effect, production in industries could halt as a result of prolonged downtime due to anomaly. Furthermore, it is important to note that complete breakdown could be catastrophic especially in automobiles and heavy duty production machines which could as a result of sudden failure during operation, jet parts from the machine leading to accidents and sometimes death. This therefore suggests that consistent monitoring of the health status of bearings is important so as to ensure efficiency and avert complete breakdown. Aside that, it is also important to detect this anomaly much earlier via sensitive methods so as to be fore warned. However, to avoid downtime, it is a good practice to constantly monitor the component so as to aid early detection of anomaly before a break down. Over the years, a lot of researches has been done in the detection of anomaly in motor bearing using time sequence data which has evolved via the use of Artificial intelligence (AI) these days. However, since bearings can run for months or years without anomaly, there is a need to use an unsupervised AI model which could be trained with data characterized with normal bearing operation and can flag an anomaly when an outlier is detected. Despite the use of AI techniques, often times, anomaly is predicted at a high threshold signifying low sensitivity. This is because of the sequential data generated is often laced with noise from sensor and therefore characterized by low signal to noise ratio (SNR). Furthermore, for better accuracy, the noise within the data generated as a result of the sensors used has to be taken into consideration and worked upon using a digital signal processing technique such as Fast Fourier Transform (FFT) so as to aid fast computation and anomaly detection at low signal to noise ratio. To improve detection, this research work presents the Development of an Anomaly Detector for Motor Bearing Condition Monitoring using Fast Fourier Transform (FFT) and Long Short Term Memory (LSTM)-Autoencoder (AE). This was achieved via the use of Fast Fourier Transform (FFT) and Long Short Term Memory (LSTM)-Autoencoder which helped to detect the anomaly at a threshold less than 0.2 and also attain an accuracy of 91.3%.