Multi fault diagnosis based on loading matrix and score matrix of principal component analysis for a centrifugal pump

Centrifugal pumps are one of the rotating machines that are widely used in various industries such as oil and gas, petrochemical, water treatment, power generation, agriculture, and fertilizers. During its operation, it can experience failures which can potentially cause disruption of production processes. Early detection of faults in centrifugal pumps can reduce energy consumption, service and maintenance costs, increase reliability, life-cycle and safety, and therefore significantly reduce through-life-time costs. Application of statistical features combined with principal component analysis (PCA) is one of the techniques that can be used to monitor machine condition. PCA is a widely used procedure which reduces the dimension of an original dataset into lower dimension by converting original possibly correlated variables into new set of uncorre-lated variables. In this paper a method based on loading matrix and score matrix of PCA (Principal Component Analysis) is proposed to build a framework for multi fault diagnosis for a centrifugal pump. The PCA diagnosis model was developed by using historical normal condition, and further fault detection was determined by Hotellings T2 statistic, while fault location was obtained through the combination of loadings and scores of principal components. In this study, normal and faulty conditions of a centrifugal pump were obtained from a Spectra Quest Machinery Fault Simulator and four common fault modes were investigated, impeller fault, blockage, cavitation and bearing fault. The combination of faults were also used, including impeller fault-blockage, impeller fault-cavitation, impeller fault-bearing fault, and impeller fault-cavitation-bearing fault. Vibration signals were measured using two accelerometers mounted on the radial direction of the volute case and the bearing housing respectively. The centrifugal pump was run at the speed of 2095 rpm and 2400 rpm, and acceleration vibration data was collected at a sampling rate of 48192 Hz with sampling duration of 1 second. Five statistical features (kurtosis, RMS, skewness, variance and entropy) were extracted from vibration time waveforms which were previously divided into frequency bands or octaves. The time waveform was firstly transformed to the frequency domain via the FFT and was then divided into octave bands. The frequency separation was based on the impeller shaft speed, which was chosen to be the central frequency for the second octave band, and the central frequencies of the other octave bands corresponded to the shaft speed harmonics. Secondly, the frequency domain data was then transformed back to the time domain (via IFFT) providing many segments corresponding to each of the octave bands. The PCA model for normal condition was then developed based on statistical features extracted from each time waveform segment. The new data set from the faulty condition was tested using the PCA model. The fault detection was achieved through comparing the T2 statistic for each measurement to the T2 control limit. Once the fault had been detected, the next important step was to diagnose the location of the fault which was obtained by analysing the loadings and scores matrices on the first of several principal components. Furthermore, investigations were conducted for all of the time waveform segments for the various fault combinations and by comparison, the most important features for fault diagnosis were obtained