Remaining useful life prediction based on health index similarity

Highlights

• A novel prognostic method based on health index similarity is proposed.

• Nonlinear degradation evolution is revealed by the health index of cutting tools.

• The distance similarity and the spatial direction similarity are both considered for similarity matching.

• The proposed method shows potential to outperform the LS-SVR method.

Abstract

Condition-based maintenance and the prediction of the remaining useful life (RUL) of cutting tools are of crucial importance to reduce unexpected downtime and ensure quality. Our paper proposes an original RUL prediction model based on health index (HI) similarity, where both distance similarity and spatial direction similarity are considered. Data mining is carried out to the large and messy original monitoring data to construct the HIs of the cutting tools, which are then used to predict the RUL. The novelty of our method is that it can make full use of limited historical datasets to achieve more accurate prediction results. The model is applied to the data obtained from a turbine factory's slotting cutter machining process and is compared to one of the most popular prognostic method - least squares support vector regression. Our proposed approach is also applied to two further case studies –a GaAs-based semiconductor laser and simulated data. The comparative results show the effectiveness and practicability of our proposed method, even when the data fluctuate a lot and show distinctive trends.

Introduction

The gas turbine is one of the main sources of power for many applications, e.g. civil and military aircraft, naval and commercial ships, electricity production, etc. [1]. The gas turbine is highly valuable and its structure is complex. It is manufactured using specialist turbine cutting tools, which are very expensive compared to other cutting tools. Failure of the cutting tool during the manufacturing process could lead to catastrophe: if the gas turbine is damaged it could cause huge financial loss, downtime and accidents. Tool wear/life has a direct effect on both the dimensional precision and the quality of surface finish of the workpiece [2], and it is estimated that 20% of downtime is attributed to tool failure [3]. It is therefore important to find the optimal way to utilize the turbine cutting tools while maintaining their safety and good working order.

Prognostics aims to predict the RUL of an equipment or system, which is an effective technique to avoid risk and financial loss. As such, this area has received greater attention in recent years. Prognostic approaches can be broadly categorized into physics-based, data-driven and hybrid approaches [4]. Physics-based approaches attempt to establish explicit mathematical models of the system for RUL prediction. However, as it is difficult to build accurate models when the real-life system is complex, its applications, especially in practical systems, are limited. Data-driven approaches are black box models that learn equipment behavior directly from condition-monitoring data. This avoids the disadvantages of physics-based approaches, and as such this approach is widely used [5]. A hybrid approach is a combination of physics-based and data-driven approaches that can overcome limitations of each individual method, however, it is not very well developed yet and few studies have been reported so far [4], [6]. Therefore, in this paper, we mainly focus on the data driven approaches.

Data-driven approaches can be mainly classified into two main categories: machine-learning and statistical-learning approaches [5]. Machine-learning approaches mainly include support vector regression (SVR), artificial neural networks, Bayesian networks, hidden Markov models and neuro-fuzzy systems, among others. Nowadays, there are many researches about the RUL prediction of cutting tools based on data-driven approaches. Benkedjouh et al. [7] used SVR to predict both the amount of wear and the RUL of cutting tools. Gokulachandran and Mohandas [8] compared two soft computing techniques – neuro fuzzy logic and SVR – for the assessment of the RUL of cutting tools. Drouillet et al. [9] used artificial neural networks to predict the RUL of milling tools. Tobon-Mejia et al. [10] developed a two-phase RUL prediction method based on dynamic Bayesian networks. Chen et al. [11] and Wang and Wang [12] established the RUL prediction model based on hidden Markov models. Wu et al. [13] proposed a multi-sensor information fusion system for the online RUL prediction of machining tools, based on the adaptive network-based fuzzy inference system. The above approaches mainly focus on establishing the relationship between monitoring data and tool wear during the machining process, and their accuracy is highly dependent on the quantity and quality of the degradation data. However, in industry, tool wear cannot be measured during the middle of a production process, i.e. to guarantee the surface integrity and efficiency of the products, the production process must run in its entirety. Thus, data on the evolution of tool wear is not available. When the available data is insufficient, the distinctive trend and random fluctuations within the degradation data could produce unacceptable errors.

Statistical-learning approaches mainly include the Wiener process, Gamma process, and stochastic filtering methods such as Kalman filters and particle filters [14]. Sun et al. [15] proposed a RUL prediction method based on the Wiener process for cutting tools. However, Winner process modeling is based on the Markov property assumption, and while the Markov property is a valid assumption in many applications, in general it does not always hold. Ling et al. [16] conducted degradation analysis for products with two-phase degradation under a gamma process. Similar to Winner process, modeling degradation with a gamma process implies a special Markov process with a continuous state and time space allowing transitions in one direction. Zheng and Fang [17] and Son et al. [18] developed the RUL prediction model based on Kalman filters. Qian and Yan [19] proposed a particle-filter-based method for the RUL prediction of bearings. However, there are some limitations in stochastic filtering methods. The use of Kalman filters is based on the assumption that both the process and sensor noises are Gaussian distributed. The particle -filter is a more generalized scheme and does not need the distribution assumption, but it has problems of particle degradation and scarcity, which are caused by the resampling. In addition, the prediction accuracy of these methods is dependent on the quantity and quality of the degradation data.

The similarity-based prognostic approach, which belongs to the machine-learning approaches, does not need any hypothesis to build the degradation model, and needs only a small amount of historical data to predict the RUL based on similarity between samples, could be a solution to this problem [20], [21], [22]. Euclidean distance-based similarity is commonly used in previous studies. The similarity in the spatial direction is also very important, but is often neglected due to its complexity.

In this paper, we propose an original RUL prediction method for cutting tools. It includes three main stages: health index (HI) construction, similarity matching, and RUL prediction. The obtained original monitoring data are extracted and the dimension is reduced by principal components analysis (PCA) to build the HIs of the cutting tools. The constructed HIs are then used for similarity matching between the test and reference cutting tools. Finally, the RUL of the current test cutting tool is predicted by applying a weighted average of the RULs of all similar reference cutting tools. The proposed approach is applied to data obtained from a turbine factory's slotting cutter machining process, and two further case studies are used to verify the generalizability and validity of our proposed method. Our paper provides four main contributions:

• As we have no tool wear information during the production process, so we cannot use tool wear to predict the RUL. Instead, we use the original monitoring data to construct the HIs of the cutting tools, which are then used to predict the RUL.

• Since the original monitoring data is noisy and the volume is big for each cutting tool, it is difficult to use the original data in a meaningful way. Therefore, data mining is carried out to extract useful data to construct the HIs.

• Unlike previous research, we consider both the distance similarity and the spatial direction similarity. We then use the constructed HIs for similarity matching between the test and reference cutting tools.

• Our proposed method does not need any hypothesis to build the degradation model and makes full use of limited historical datasets to achieve more accurate prediction results even when the data fluctuate a lot and show distinctive trends.

The remainder of this paper is organized as follows. Section 2 introduces the HI similarity-based methodology. In Section 3, three case studies regarding turbine cutting tools, GaAs-based semiconductor lasers and simulated data are used for validation and comparison. Finally, some conclusions are given in Section 4.