A TOPSIS-Assisted Feature Selection Scheme and SOM-Based Anomaly Detection for Milling Tools Under Different Operating Conditions

Abstract

Anomaly detection modeled as a one-class classification is an essential task for tool condition monitoring (TCM) when only the normal data are available. To confront with the real-world settings, it is crucial to take the different operating conditions, e.g., rotation speed, into account when approaching TCM solutions. This work mainly addresses the issues associated with multi-operating-condition TCM models, namely the varying discrimination ability of sensory features; the overlap between normal and anomalous data; and the complex structure of input data. A feature selection scheme is proposed in which the Technique for Order Preference by Similarity to Ideal Solution (TOPSIS) is presented as a tool to aid the multi-objective selection of sensory features. In addition, four anomaly detection approaches based on Self-Organizing Map (SOM) are studied. To examine the stability of the four approaches, they are applied on different single-operating-condition models. Further, to examine their robustness when dealing with complex data structures, they are applied on multi-operating-condition models. The experimental results using the NASA Milling Data Set showed that all the studied anomaly detection approaches achieved a higher assessment accuracy with our feature selection scheme as compared to Principal Component Analysis (PCA), Laplacian Score (LS), and extended LS in which we added a final step to the original LS method in order to eliminate redundant features.

On the Stability and Homogeneous Ensemble of Feature Selection for Predictive Maintenance: A Classification Application to Tool Condition Monitoring in Milling

“…In [33], FS techniques based on decision trees, neural fuzzy systems, scatter matrix, and a cross correlation were compared for TCM in milling. Aside from the well-known FS methods, some application-specific FS schemes were proposed in the field of PdM to meet particular needs and scenarios, e.g., an FS scheme was proposed in [4] to tackle the challenges related to anomaly detection of milling tools when the data belong to different operating conditions. In [6], an FS scheme was proposed to reduce the detrimental effect of the data outliers on the accuracy of fault diagnosis.…”

“…To evaluate the effectiveness of the FS schemes proposed in the literature, their performance is usually compared with popular FS methods [4], [17], [27], [34], with other related works [6], [17], and/or with the case when no FS is performed (i.e., all the features are used) [6], [10], [27], [35], [36]. The major performance indicator used to evaluate a given FS or to compare different FSs is the predictive performance of the learning model, e.g., accuracy, which was built using the features selected by the corresponding FS method [4], [17], [18], [20], [27], [34]- [39].…”

“…However, it is quite intricate, if not impossible, to theoretically determine the optimal sensory features for the problem at hand. This is mainly due to the complexity and dynamics of the industrial processes being monitored, and the presence of diverse factors affecting the importance of a given feature, including the operating conditions [4], [9], [10], position of sensors, signal-to-noise ratio of the acquired signals [10], the window size of feature extraction [11], etc. Therefore, in order to ensure that sufficient information is derived, many features are initially extracted from the sensor signals in the time, frequency, and/or time-frequency domain.…”

“…FS methods have been extensively employed, investigated, and developed in the field of PdM, including a wide range of industrial processes, e.g., machining [4], industrial components, e.g., machining tools [4], gears [17], bearings [18], as well as PdM tasks, e.g., diagnosis [6] and prognosis [19]. Regarding the evaluation of FS methods, the majority of these works only consider the predictive power of the FS, expressed as the predictive performance of the learning model whose input are the features selected by the FS method, e.g., accuracy or similar measures [4]. Some works also consider other performance indicators such as the run time of the FS algorithm [20], among others.…”

“…As such, it is not unlikely for the PdM dataset to suffer from a low sample-to-dimension ratio (SDR) (i.e., the ratio between the number of observations and number of features), in which case examining the FS stability would be particularly crucial. • It is common in the literature to assess the candidate sensors based on the relevance of the features extracted from their signals, as performed in [4], [26], [27]. Such an assessment is especially important when there is a constraint on the number of the sensors to be used.…”