Improved machine learning approaches designed and validated in different engineering application fields

Machine learning (ML) provides machines the ability to automatically learn from data and past experience to identify patterns and make predictions for new data with minimal human intervention. Comparing with other statistical technologies, ML can improve calculation continuously, making decisions automatically, and identifying trends and patterns automatically. Based on the fact more datasets open to public, improved computer calculation capability, and internet speed acceleration, machine learning algorithms develop fast in the past decades. Large number of new ML approaches are raised and updated in the last years. However, most of these designed approaches target on one specific dataset, therefore, generalization of these approaches are not verified. Besides on the specific-designed approach to one dataset, a great deal of pre-trained models (which are proposed by a large number of data) are open to public. However, most of these pre-trained ML models are also concerning to one special field and their performance is not ideal when applied to datasets in other fields. Consequently, improved ML approaches that can be applied in various fields are required.

To design ML approaches available in various fields, four datasets from three application fields are employed to verify approaches designed in this study. Inner speech (IS) dataset belongs to biology field. The frequency bands of IS dataset are very low (from 0.5 Hz to 100 Hz). In this dataset, various signs are acted by inner speech during the experiment. Electroencephalogram (EEG) signals acquired during IS procedure and data are analyzed and classified. Case Western Reserve University (CWRU) bearing dataset is conducted by the mechanical school in CWRU. It is often used for bearing diagnosis approaches verification. Vibration signals are obtained when various faulty bearings are employed during the experiment. Vibration signals has a wide frequency bands from dozen Hz to thousand Hz. Another application field employed in this study is metalworking fluids (MWF). Acoustic Emission (AE) signals when a variety of MWF are applied during
thread forming process are analyzed. Acoustic Emission signals are transient stress waves generated by the rapid release of energy from solid sources with frequency bands up to million Hz. According to the MWF types, two experiments are conducted and relevant AE signals are acquired. In the first experiment (dataset MWF19) only emulsion-based MWF are applied. Both emulsion-based and oil-based MWF are used in the second experiment (dataset MWF16).

According to the different highlights and steps in machine learning, three categories
with 5 sub approaches are designed in this study. Approach 1 contains few steps on data processing - data selection, segmentation, and classification. In this approach, signals are analyzed in time domain. Concentration of approach 1 is on classifier (convolution neural network (CNN)) structure and hyperparameters referring to training algorithms designing and tuning. Convolution neural network is designed according to samples length and functionality of each layer. Beside the CNN structure and hyperparameters in CNN, a new data processing method is designed for various phases differentiation in one measurement. As the boundaries among different phases are not clear in time domain in MWF datasets, continuous wavelet transform (CWT) is applied as a tool to find boundaries among different parts in one measurement. Furthermore, segment’s length is defined by rotating tool speed in the data segmentation step.


Comparing with approach 1, more data processing methods are included in approach 2. Although the same data selection method is used in this approach as approach 1, besides data selection and segmentation in approach 1, data are transformed from time domain to time-frequency domain by Short-time Fourier Transform (STFT) and spectrograms are obtained. Segments and spectrograms are normalized before they are put into the classifier. Unlike measurement are segmented into fixed size segments according to tool speed, segment length is not identical in this approach, they are considered as one adjustable parameter. Besides, overlap among different class are not identical when samples numbers in each class are not the same. Additionally, parameters in data processing and hyperparameters in CNN are optimized together and automatically in one step. Furthermore, based on approach 2, a transfer learning (TL) approach (approach 2.0) is raised between MWF19 and MWF16. Parameters in data processing and hyperparameters in CNN trained from MWF19 are transformed to MWF16 in approach 2.0.

Unlike one method or algorithm is applied in one step in approach 1 and 2, varied data processing methods and ML algorithms in each step are tried in approach 3. From this point of view, approach 3 is a integration of various approaches instead of single approach. Fixed data are used in approach 1 and approach 2, on the contrary, various parts data are tried in this approach. Besides, many similar but not identical function data processing methods and ML algorithms are employed in each step. Based on the destination difference, approach 3 is divided into supervised learning (approach 3.1) and unsupervised learning (approach 3.2). Although the process of approach 3.1 and 3.2 is the same, methods used in each step in approach 3.1 and 3.2 are varied. In approach 3.1, different training-test data split ways are tried. Besides, raw measurement, Savitzky–Golay (SG) filter, and empirical mode decomposition (EMD) are applied in data processing step. Linear, polynomial, and Gaussian kernels SVM are employed as classifier. In approach 3.2, STFT, CWT, and Hilbert–Huang transform (HHT) are used for signals transformation from time domain to time-frequency domain. Autoencoder as an alternative is employed as feature extraction method. K-mean and Gaussian mixture model (GMM) are used for features clustering.

Approach 1, 2, and 3.2 are applied to CWRU and MWF datasets. As the IS dataset is very different from other datasets, so approach 3.1 is specific to it. Besides, considering the similarity between MWF19 and MWF16, the approach 2.0 is applied between them. For IS dataset, when training-test data are split in different ways, the results are significantly different. Best results of CWRU beaing dataset from supervised learning approach (approach 2) are F-score and accuracy is 100% for 29 bearing states, which denotes that all bearing states in CWRU bearing dataset can be classified perfectly. Best results of CWRU bearing dataset from unsupervised learning approach (approach 3.2) is all metrics are 1.0 in fault-free and faulty bearing distinction. For MWF19 dataset, the best classification results are from approach 2 - F-score and accuracy are 98.61 % and 98.58 %. Beside, reference and other fluid are totally distinguished by approach 3.2. For MWF16 dataset, the best classification results arrive to accuracy is 98.11 % from approach 1. Although clustering results are still open for MWF16, some conclusions can still be drawn from calculation process.




Citation style:
Could not load citation form.


Use and reproduction:
All rights reserved