مقایسه شبکه های بیزی و شبکه های عصبی مصنوعی برای تشخیص کیفیت در یک فرآیند ماشینکاری

Machine tool automation is an important aspect for manufacturing companies facing the growing demand of profitability and high quality products as a key for competitiveness. The purpose of supervising machining processes is to detect interferences that would have a negative effect on the process but mainly on the product quality and production time. In a manufacturing environment, the prediction of surface roughness is of significant importance to achieve this objective. This paper shows the efficacy of two different machine learning classification methods, Bayesian networks and artificial neural networks, for predicting surface roughness in high-speed machining. Experimental tests are conducted using the same data set collected in our own milling process for each classifier. Various measures of merit of the models and statistical tests demonstrate the superiority of Bayesian networks in this field. Bayesian networks are also easier to interpret that artificial neural networks.

Quality is defined as the extent to which a product conforms to the design specifications and how it complies with the requirements of component functionality. For some industries, such as automotive and aeronautical sectors, the quality of their parts is very important given the high requirements to which they are subject. However, difficulties arise from the fact that a measure of quality can only be evaluated “out-of-process”, resulting in losses because there is no alternative to removing defective parts from the production line. Therefore, it is necessary to incorporate machine learning methods that provide in-process solutions to predict quality from some measured variables. Nowadays, many papers have been published about modelling the machining process and, more specifically, about the prediction of surface quality in machining processes. Researchers have approached the problem from different points of view and using different techniques. The most frequently used are artificial neural networks (ANNs) (Huang and Chen, 2003, Samson and Chen, 2003 and Tsai et al., 1999) and linear and multiple regression (Aboulatta and Mádl, 2001, Feng and Wang, 2003 and Kirby et al., 2004). However, their models focus on very reduced environments and with limited experimentation. In Correa, Bielza, Ramírez, and Alique (in press), we have recently proven the advantages of using Bayesian networks (BNs) as a successful solution for predicting surface quality in high-speed milling. As an important added value, the current research includes the influence of the geometry of the workpiece and the hardness of the material to be machined as key variables in the model construction aimed at a particular subdomain that contains a range of aluminium hardnesses used in automotive and aeronautical pieces. This is a landmark in this application domain, since it extends and generalizes the scope of the experimentation, which is no longer confined to a single test profile. BN models were learnt from data. These data were collected in our laboratory using experimental design to guarantee statistical validity. Besides BNs, we constructed ANNs, known to be a strong competitor widely used in this field, to make a comparison and to demonstrate the superiority of BNs. As far as we know, there have been no comparisons of how well these two techniques solve this kind of problem. Obviously, these two models have already been compared in other contexts like e.g. modelling manufacturing processes (Perzyk, Biernacki, & Kochański, 2005), discriminating plants, weeds and soil in color images (Marchant & Onyango, 2003), and modelling the response time of service-oriented systems (Zhang & Bivens, 2007). The two have been used in a combined fashion (Antal, Fannes, Timmerman, Moreau, & de Moor, 2003). Our paper aims to compare the two approaches (BNs and ANNs) in the context of a practical industrial problem, the prediction of surface roughness in high-speed milling. The proposed models are target the automotive and aeronautical industry using some typical geometric features and a number of aluminium alloys giving a wide range of hardness. The remainder of this paper is structured as follows. Section 2 presents the difficulties for measuring quality in a high-speed milling process, how surface roughness is inspected and which techniques and data we have used for surface roughness monitoring, since this is a variable that is very difficult to measure in-process. Sections 3 and 4 summarise the main principles of BNs and ANNs, respectively, and introduce the models on which our comparison is based. Section 5 focuses on the quantitative comparison of the two models and, in particular, on the knowledge engineering aspects of both. Finally, Section 6 concludes with our most important findings.

An ANN is a model often used to predict surface quality in machining processes. In this paper, we propose using BNs instead, showing a number of advantages over ANNs and extending the application domain to include features, not easily found in the experimental studies, that influence surface roughness, like the geometry of the workpiece and the hardness of material to be machined. After validating both models with the same data and technique (K-fold cross-validation), BNs achieve the best results from the point of view of classifier goodness applied to the problem of quality prediction in high-speed milling processes. The results have been confirmed by several hypothesis tests. As for the time it takes to build the model, BNs also outperform ANNs, requiring 0.08 CPU seconds and 12.69 CPU seconds, respectively, on a 3 GHz, 1.5 GB Dell Dimension PC. The ANN optimization procedure does not guarantee the convergence to a global minimum. There are no principled methods of choosing the network parameters (number of hidden layers, number of nodes in the hidden layer(s), form of activation functions). On the other hand, BNs have an easy and fast construction procedure without tuning parameters. Note in favor of ANNs, that the memory requirements, represented by parameters in an analytical form, are smaller than for BNs, represented by tabular conditional probability tables. However, this is not such a relevant question nowadays where computer memory is cheap and extensive. Relative speeds of operation follow the same pattern. Thus, the BN can be easily implemented as a simple table look-up, and it is intrinsically fast. However, the ANN requires a number of multiplications and additions at evaluation time, rendering it comparatively slow if high intensity predictions are required. Both classifier models are simple to use, but BNs can be more easily understood by humans. ANN models work like a black box. In contrast, friendly and intuitive BNs help users to update models and increase the confidence in the correctness of the model to be finally adopted. Finding factors that determine surface roughness can help to optimize high-speed milling, which is extensible to other industrial applications. Moreover, BNs support inference in any direction providing responses to any kind of query, not only about surface roughness but also, given some evidence, about different predictor variables. Because they capture these different types of reasoning to infer knowledge, BNs are useful models with significant representation power. From the comparison performed here, BNs are preferred over ANNs.