A Hybrid automatic fault trend prediction system based on synchrosqueezed wavelet transform and density peak cluster

Faults during operation of a system can occur at any time, especially in the modern industry where numerous sensors and automatic machines are involved. One tiny error in the manufacturing process could cause catastrophic damage to the entire system. The aftermath is unbearable. Unfortunately, the more advanced the system is, the longer downtime and higher cost will be required to fix the system. However, tiny errors in the system are hard to detect at its early stage, not only because when the error happens, the system is still functional, but also because it is challenging to technically detect a minor condition change. This brings challenges to fault detection. How to detect the fault before it happens and avoid undesirable maintenance of the system or fix the system when the fault occurs is a problem that needs to solve.  Recently, the concept of predictive maintenance and machine condition monitoring are proposed to solve the fault avoidance problem in modern industry. Preventive maintenance aims to regularly maintain the machine so that the system is always at a good working condition, but this method can be expensive. Machine condition monitoring aims to find deterioration of machine condition through the application of different analysis models with the support of a large amount of collected data, but most solutions are specific to applications. Hence, problems still exist. One problem is contemporary fault diagnosis systems are designed to focus on identifying the fault. However, when a fault occurs, the system has already incurred losses, either as rejects or machine damage. Another problem is when the data provided is limited, the analysis system, which relies heavily on the quality of the data, may not produce correct answers. The fault detection process relies on certain signal processing method to extract features, but the output can be restricted by the limited data sets. The method proposed under this circumstance aims to recognise the upcoming trend of condition changes and predict expected state of the system based on the fault trend in reasonably advanced time frame. In this thesis, after studying numerous data processing methods in different domains, a signal processing method called continuous wavelet transform (CWT) is selected as the initial signal processing method. Compared with other studied methods, it has the advantage of capturing signal information in both the time and frequency domain which is helpful in this research. Moreover, to observe the signal change more strictly, CWT scalogram, which is the CWT's derivation, is applied in this research. However, the method still needs adjustment to find minor errors or transient changes in the signal even before the fault happens. Therefore, a new signal processing method is proposed by dividing the collected data sets into overlapping segments and extracting features from each CWT scalogram segment. Experiments are then set to apply the proposed method. The experiments have proved that the proposed method can extract features with time that are not found in other signal processing methods. From the segmented scalogram, the state change of the signal can be found. To improve the accuracy of the signal analysis, synchrosqueezing methods are applied to enhance the readability of the segmented scalogram and testify the observation of state change in the signal. The proposed signal processing methodology gives useful feedback. However, more reliable, and stable indicators are needed to reflect the transient change in the signal. The most significant innovation in this research is to divide the continuous signal stream into overlapping data segments so that time domain information can be retained with lower resolution. From the time segments and using synchrosqueezed CWT digital signal processing technique, a newly defined parameter designated as standard deviation of frequency (SDF) can be computed. SDF exhibits consistent behaviour over the data collection period, i.e., the trend. Unfortunately, individual SDF values can vary substantially making it difficult to visualise. Instead, adapting from the concept of moving average, the sum standard deviation of frequency (SSDF) has been developed. From the experiment, SSDF proves its ability to reflect the change of inconspicuous conditions in the signal in different data segments over time.  The newly defined indicators further demonstrate the changes in the signal in a more stable way. Three states of the signal, “normal”, “marginal”, and “abnormal”, are then defined by SSDF values. The goal is to research and develop a system which can predict the upcoming fault trend and identify the type of fault trend. The discovery of SSDF as a trend indicator is critical in this development. Several trend prediction methods reported in literature have been examined but their predictions are not accurate. A new trend prediction method is required to predict how the signal stream will perform in the next minute based on the pattern comparison of the SSDF. Subsequently, a new computational method based on “moving SSDF deep learning” (MSDL) concept has been proven to have a better prediction of future capability than four other prediction methods. Although the proposed method is validated on a 3D printer, the analysed signal, in this case, the vibrational signal of the printer head is independent of the context of application. Other types of signals are also suitable for the proposed methodology, in particular, latest industry environment under Industry 4.0.