Karam, Sara (2014) Machining Process Monitoring via Cognitive Sensor Fusion. [Tesi di dottorato]

Anteprima

Testo Thesis Sara Karam.pdf Download (4MB) | Anteprima

Abstract

In recent years, manufacturing industry pressures for better productivity, better quality and cost savings. Surviving in today’s highly competitive world, industries need to work on optimizing the machining processes through adopting, updating, and using new machining technologies. Aiming at improving the quality of the worked product, several characteristics of the machined part should be taken into consideration. Some major characteristics that have been studied during the course of the 3-year PhD studies and that are closely related to the quality of the final product are: chip form and residual stress in turning process, and surface roughness in polishing process. The chip form produced, the level of residual stress, and the value of the surface roughness can be related to sensorial measurements that can be done during the machining process. For this purpose, sensor systems are mounted on the machine tool and are employed to monitor the machine process. The sensor monitoring systems can consist of several type of sensors: triaxial force sensor, triaxial vibration sensor, and acoustic emission sensor. At the end of the machining process, the detected sensor signals are processed and analyzed using several techniques (conventional and advanced techniques) in order to relate them to the required product output. Moreover, features were extracted from the detected sensor signals and were used to develop a decision-making system that can relate the extracted features to the required final product characteristic. Two signal processing and feature extraction methodologies were used: the conventional feature extraction and the wavelet packet transform. Conventional Feature Extraction The conventional features to be extracted from each dataset are statistical features: mean, variance, skewness, kurtosis, and energy. Wavelet Packet Transform A wavelet packet transform (WPT) was used for feature extraction. The WPT of a sensor signal generates packets of coefficients calculated by scaling and shifting a chosen mother wavelet, which is a prototype function. In this way, at the 1st level of WPT the original sensor signal S is split into two frequency band packets, called approximation A1 and detail D1. At the 2nd level, each approximation and detail packet are again split into further approximations, AA2 and AD2, and details, DA2 and DD2, and the process is repeated generating a "tree" of decomposition packets. The original signal S can be represented by the summation of output packets that cover the full signal decomposition "tree". The employed mother wavelet is a Daubechies 3 denoted by “db3”. The decomposition was performed up to the 3rd level, yielding 14 packets. For each packet, 5 statistical features were calculated: mean, variance, skewness, kurtosis, and energy. Two different approaches may be used: sensor fusion of different types of detected sensor signals processing approach and combining extracted features in a fused approach for decision-making Principle Component Analysis (PCA) is an example of a sensor signal processing procedure where sensor signals of different types are joined together to extract relevant features. The decision-making system was based on neural network (NN) for pattern recognition. NNs are inspired by the nervous system and are composed of nodes that operate in parallel. The connections between the nodes determine the function of the network. By adjusting the weights of the connections, a NN is trained to perform a particular function. The NN is usually trained and adjusted so that a particular input gives a specific output. Comparison is done between the obtained output and the target output. Based on this comparison, the NN is adjusted. Several input/target pairs are required to train a network. One main function that NNs are trained for is pattern recognition. The NN is trained to distinguish a pattern among the input features and relate them to the output. During the 3-year PhD study course, three major industrial sensor monitoring applications were studied: chip form recognition, residual stress level identification, and surface roughness in polishing. Chip form recognition Cutting force sensor monitoring and wavelet decomposition signal processing were implemented for feature extraction and pattern recognition of chip form typology during turning of 1045 carbon steel. The wavelet packet transform was applied for the analysis of the detected cutting force signals by representing them in a time-frequency domain and providing for the extraction of wavelet packet statistical features. The latter were used to construct wavelet packet feature vectors, ranked according to the number of overlapping elements related to favourable or unfavourable chip forms that cause noise in the pattern recognition procedure (lower number, lower noise, higher rank). The eight highest ranked wavelet packet feature vectors were selected as inputs to a neural network decision-making system on chip form acceptability. Subsequently, a data refinement procedure was employed to improve the neural network performance in the chip form identification process. Residual stress level identification On-line residual stress assessment in turning of Inconel 718 was carried out through multiple sensor monitoring based on cutting force, acoustic emission and vibration signals acquisition and analysis. The detected sensor signals were processed by the wavelet packet transform technique to extract statistical features from the packet coefficients for the construction of wavelet feature vectors. The latter were used for sensor fusion pattern recognition through neural network data processing grounded on X-ray diffraction residual stress measurements on the turned part surface. The scope of the sensory data fusion approach was to achieve a robust scheme for multi-sensor monitoring decision making on machined surface integrity in terms of residual stress level acceptability. Surface roughness in polishing Surface roughness assessment on Unimax roll 1 cylinders was carried through multiple sensor monitoring of robot assisted polishing. Several sensors have been mounted on the robot assisted polishing machine. The signals detected during the polishing experiments were: acoustic emission, strain, voltage, and current. Furthermore, the surface roughness was measured after a number of passes defined during the performance of the experiment. Relevant features were extracted from the detected signals and were related to the relevant surface roughness of the polished part. The two types of feature extraction were used: conventional statistical features and the wavelet packet transform. Through the online analysis of the sensor signals, the polishing process should be stopped when the required surface roughness is reached, i.e. the features extracted from the sensor signals were be used to make a decision on when the required surface roughness has been achieved and when to stop the polishing process.

Downloads

Actions (login required)

Modifica documento