Investigation of Data Compression Methods for Intelligent Machine Condition Monitoring

Condition monitoring (CM) deliveries significant benefits to the industry by minimising breakdown losses and enhancing the safety and high-performance operation of machinery. However, the use of data acquisition systems with multiple sensors and high sampling rates leads to massive data and causes considerably high cost for purchasing and deploying hardware for data transmission, storage and processing. Hence, data compression is crucial and important to reduce the data size and speed up the calculation for the development of intelligent machine CM systems. Although data compression has received high attention in many fields, few researchers have focused on their research in the field of machine CM. Therefore, this PhD research concentrates on investigating novel and high-performance data compression algorithms according to the characteristics of one-dimensional (1D) and two-dimensional (2D) signals to solve the bottleneck of the massive data transmission, and hence improve the performance of remote and real-time machine CM systems.

The research is carried out according to a compound experimental and analytic route based on a wireless senor network. To demonstrate the effectiveness of data compression based techniques for CM, the prototype of an intelligent wireless sensing system is developed using cost-effective micro-electromechanical systems (MEMS) accelerometers and the Bluetooth low energy (BLE) communication module. Moreover, various waveform parameters with low cost computing in time and frequency domains are investigated and identified that RMS is the most effective parameter to give good indication for the leakage in a piping system, showing that data compression via statistics is effective and thus indicates that the performance of data compression for CM highly depends on applications.

Subsequently, high-performance but high-complexity methods are proposed base on dimension reduction, sparse representation, feature extraction and advanced compressive sensing (CS) for fault diagnosis of rotating machinery with 1D or 2D signals, which have the potentials to be implemented on MEMS modules in a wireless sensor network (WSN) in future. Firstly, a compression scheme based on dimension reduction is proposed to extract the periodic characteristics of the 1D vibration signal. Recurrence plot (RP) of vibration phase space trajectory and its quantification indicators, as well as principal component analysis (PCA), are combined to realize feature extraction, compression and fault classification for a tapered roller bearing system.

Secondly, a two-step compression method is performed on 1D vibration signals based on frequency shift, adaptive sparse representation and CS is explored to overcome the problem of the large quantity of data storage for ball bearing fault diagnosis. Simultaneously, this compression method has the capability to reconstruct envelope signals with noise elimination.Then, for 2D thermal images captured from a two-stage reciprocating compressor, the dense scale-invariant feature transform (SIFT) features indicating edge information are extracted and represented as a sparse matrix by sparse coding. The compressed features are used for the classification of six different types of faults with the support vector machine (SVM).

Finally, the advanced CS technique is exploited on pre-processing the 2D thermal images of gearboxes to realise intelligent fault classification with high accuracy of more than 99.81% by a typical deep learning algorithm, namely convolutional neural network (CNN). The CNN calculation speed is dramatically accelerated with compressed images. All these proposed approaches are evaluated by simulations and experiments, which verifies that they can reliably detect the fault types or classify different fault types with very high accuracy. Besides, the proposed data compression based intelligent CM approaches provide theoretical bases for maintenance-free CM systems because data compression can save the transmission bandwidth and power consumption for remote and real-time machine CM systems.