An Enhanced Adaptive Neural Fuzzy Tool Condition Monitoring for ...

An Enhanced Adaptive Neural Fuzzy Tool Condition Monitoring for Turning Process Department of automatic manufacturing Engineering, École de technologie supérieure, University of Quebec Montreal, QC, Canada Canada [email protected]

Krzysztof Jemielniak Faculty of Production Engineering, Warsaw University of Technology, Warsaw, Poland [email protected]

Sofiane Achiche Department of Mechanical Engineering, École Polytechnique de Montréal, University of Montreal Montreal, QC, Canada [email protected]

Department of automatic manufacturing Engineering, École de technologie supérieure, University of Quebec Montreal, QC, Canada [email protected]

Qun Ren

Abstract—The research work carried out in this paper presents a novel intelligent tool condition monitoring solution for the turning process using an enhanced adaptive neural fuzzy inference system based on extended subtractive clustering. The hybrid system is constructed from training a takagi-sugeno-kang fuzzy logic system by integrating machining parameters– feed, cutting force, feed force and cutting tool wear dataset using the extended subtractive clustering. The parametric search for clustering parameters in extended subtractive clustering ensures the high precision of the fuzzy system identification. Further on, neural network automatically acquires the knowledge by the back propagation algorithm and trains the connectionist structure to refine the fuzzy logic rules and find optimal input/output membership functions for the enhanced adaptive neural fuzzy tool condition monitoring. The experimental results show its effectiveness and competitiveness in comparison with five other artificial intelligence methods applied on the same data sets. Keywords—tool condition monitoring, fuzzy subtractive clustering, hybrid intelligent system,

I.

logic,

INTRODUCTION

Effective monitoring of a manufacturing process is essential for ensuring product quality and reducing production costs. Analysis, implementation and evaluation of machining processes present significant challenges to the manufacturing industry. Machining is a multivariate dynamic process that varies considerably depending on the workpiece material, temperature, cutting fluids, chip formation, the tool material, chatter and vibration. The information obtained during the machining process generally is neither complete nor precise because of the high number of influencing parameters and also because of the cutting forces that change periodically, according to the load of the tool. It is, therefore, hard to establish a theoretical analytical approach or a mechanistic model to tool wear by using this kind of uncertain information.

c 978-1-5090-0626-7/16/$31.00 2016 IEEE

Pascal Bigras

Advanced signal processing techniques and artificial intelligence (AI) play a key role in the development of modern tool condition monitoring (TCM) systems [1]. The most frequently chosen methods are neural network (NN) [2], Mamdani fuzzy logic (FL) [3], takagi-sugenokang (TSK) FL [4, 5], or a combination of Mamdani FL and an automatic generating method, i.e., genetic algorithm (GA) [6]. All these methods have a similar objective – matching the estimate of average cutting tool wear with the directly measured wear value without building an explicit mathematical model describing the exact physics of the phenomenon. FLs allow decision making with estimated values under incomplete or uncertain information from complex machining processes. But they usually can’t automatically generate the fuzzy rules properly and knowledge acquisition is difficult and tedious. NNs are as simplified mathematical models of brain-like systems and they function as parallel distributed computing networks. But most NNs must be provided a set of connection strengths (weights) which allow the network to carry out the desired computation to train the parameters of a monitoring system. These limitations in both systems have stimulated the creation of intelligent hybrid systems, such as adaptive network-based fuzzy inference systems (ANFISs) [7]. ANFISs overcome the limitations of the individual techniques and integrates the best features of FLSs and NNs. ANFISs not only represent fuzzy prior knowledge into a set of constraints (network topology) to reduce the optimization search space, but also adapt back propagation to structured network to automate fuzzy logic system (FLS) parametric tuning. ANFIS is considered as one of the best trade-offs between NN and FLS even if it present strong computational complexity restrictions. ANFISs have been applied effectively to a wide variety of cutting tool condition monitoring in different machining process [8-12]. But for most of these ANFISs, the pre-

1976

identified TSK FLSs are without a parametric search for clustering parameters, therefore the accuracies of the system identification are limited. The aim of this paper is to present an enhanced adaptive neural fuzzy TCM (EANF-TCM) using ANFIS incorporating an extended subtractive clustering (ESC) method [13] to accomplish the integration of multi-sensor information - feed ( f ), feed force ( F f ) and cutting force ( Fc ) and tool wear (VB) information. As the same of other intelligent monitoring systems, this kind of TCM considers the information directly from the machining process, without the need to explicitly understand the exact physics of the machining process. The advantage of this method is that the parametric search for clustering parameter in ESC method ensures high accuracy and high reliability of the tool wear prediction over a range of cutting conditions. The hybrid ANFIS system is constructed from pre-identifying a TSK FLS by integrating machining parameters ( f , F f , Fc ) and cutting tool wear (VB) dataset using the ESC. Then NN's weights values are initialized by this precise TSK FLS. Furthermore, NN automatically acquires the knowledge using the back propagation algorithm and train the connectionist structure to develop FL rules and find optimal input/output membership functions (MFs) for the EANF-TCM. The experimental results are used to compare, with the same case study, with those from five different intelligent modelling techniques to show its advantage and effectiveness. This following paper includes other four sections. The initial theoretical foundation: TSK FLS, subtractive clustering method, back-propagation NN and hybrid ANFIS learning procedure are recalled briefly in Section II. ANFIS identification algorithm based on ESC and proposed EANF-TCM are presented in Section III. A specific turning case study on EANF-TCM, along with the comparison with other different AI techniques, is in Section IV. Finally, the concluding remarks and future research recommendations are given in Section V. II.

non-linear approximator that can approximate every continuous mapping, and on the other hand being a piecewise linear model that is relatively easy to interpret and whose linear sub-models can be exploited for control and fault detection. A generalized type-1 TSK model can be described by fuzzy IF-THEN rules which represent input-output relations of a system. For a Multiple Inputs Single Output (MISO) first–order type-1 TSK model, its kth rule can be expressed as: IF x1 is Q1k and x 2 is Q2k and … and x n is Qnk , THEN Z is w k = p0k + p1k x1 + p 2k x 2 + ... + p nk x n (1) where x1 , x 2 …, x n and Z are linguistic variables; Q1k , Q2k , …, and Qnk are the fuzzy sets on universe of discourses X 1 , X 2 …, and X n , and p0k , p1k , p 2k , …, p nk are regression parameters. B. Extended Subtractive Clustering Method The aim of subtractive clustering identification algorithm [17] is to estimate both the number and initial location of cluster centers and extract the TSK fuzzy rules from input/output data. Subtractive clustering operates by finding the optimal data point to define a cluster center based on the density of surrounding data points. Subtractive clustering algorithm has various parameters to be set. Not knowing the best parameters to be used for a given data, even a parameter search is performed to identify a better model. In his paper [13], Demirli described in detail the influences of the four parameters to clusters and fuzzy models, and proposed an ESC method with parametric search on various clustering parameters to identify the best model. As a result of parametric search, ranges of clustering parameters that provide best models are also identified. The algorithm of TSK FLS based on ESC with parametric search [18] is given in Fig. 1. The parameters of clustering algorithm initialised after completion of each session of learning progress.

THEORETICAL FOUNDATION

A. TSK Fuzzy Logic System TSK FLS was proposed in an effort to develop a systematic approach to generate fuzzy rules from a given input-output data set. This model consists of rules with fuzzy antecedents and a mathematical function in the consequent part. Usually the conclusion function is in the form of a dynamic linear equation [14, 15]. The antecedents divide the input space into a set of fuzzy regions, while the consequents describe behaviours of the system in those regions. The main difference with more traditional [16] (Mamdani FL) fuzzy rules is that the consequents of the rules are a function of the values of the input variables. TSK FLSs are widely used for modelbased control and model-based fault diagnosis. This is due to the model’s properties; on one hand being a general

Fig 1. Fuzzy modelling based on ESC algorithm

2016 IEEE International Conference on Fuzzy Systems (FUZZ)

1977

For the fuzzy system identification in this paper, ESC is used for the fuzzy system structure identification to determine of the number of rules and the variables involved in the rule premises. Least square estimation is used for parameter identification for the estimation of the membership function parameters and the estimation of the consequent regression coefficients. C. Back-propagation Neural Network The back-propagation NN was developed by Rumelhart et al. [19] as a solution to the problem of training multi-layer perceptron. Each node in the NN layers is connected to each of the nodes in the next layer. All connections between nodes are directed, and there are no connections between the nodes within a particular layer. Each connection between nodes has a weighting factor associated with it. Then the back-propagation algorithm is used to modify weights values that generate NN outputs that most closely match the output values in the training data. Training with back-propagation is an iterative process. In ANFIS, the NN's weights values are initialized by pre-identified TSK FLS. Each rule in the FLS can be interpreted to a training pattern for the multilayer NN, which generally have 6 layers: input layer, IF part layer or input membership layer, fuzzy rule layer, normalization layer, THEN part layer or output membership layer and the output layer. The premise part of fuzzy rule is the NN input and the consequent part of the fuzzy rule is the desired output of the NN. Relatively, the kth training sets of the NN derived from the fuzzy rule (1) can be written as follows:

{(Q1k , Q2 k ..., Qnk ), wk } , while 1 ≤ k ≤ m

TABLE I

TWO PASSES IN ANFIS HYBRID LEARNING PROCEDURE

Premise parameters (nonlinear) Consequent parameters (linear) Overall output

III.

Forward pass

Backward pass

Fixed

Back propagation gradient descent Fixed

Least square estimation (offline learning) Node output

Error rates

ENHANCED ADAPTIVE NEURAL FUZZY TOOL CONDITION MONITORING

Figure 2 shows the structure of a TCM system. In such system, real-time data are first acquired from sensors (e.g. dynamometers on Fig. 4) located as different locations of the workpiece, tool and machine-tool, then signal processing technique (amplifiers) are used to extract valid data, and decision making system (computers) is used to analyze the data and classify the results in order to make a more reliable estimation of the state of the tool and consequently of the machined parts themselves.

(2)

D. Hybrid ANFIS Pocedure ANFIS identification algorithm defined by Jang in 1993 [7], uses a hybrid learning algorithm to discriminates itself from TSK FLS by the adaptive parameters, i.e., both the premise and consequent parameters are adjustable. It uses a two pass learning cycle, which is a combination of the least-squares method (forward pass) and the back propagation gradient descent method, for training FLS membership function parameters to emulate a given training data set as shown in TABLE I. The first learning cycle is the consequent parameters training. The premise parameters are fixed at this step. Least-squares method is used because the output of the ANFIS is a linear combination of the consequent parameters, which is the same as the TSK FLS. The adaptation procedure in second learning cycle is based on the back-propagation gradient descent method, which is the same as in the training of the backpropagation NN. After the consequent parameters have been adjusted, the approximation error is backpropagation through every layer to update the premise parameters. When the designated epoch number, set 1978

manually based on concerns such as precision of the prediction or the learning time, is reached or the training error goal is achieved, the training process stops.

Fig. 2 Structure of tool condition monitoring system

For the proposed EANF-TCM, the process modelling uses the ANFIS learning procedure based on ESC which is summarized by the flow chart illustrated in Fig. 3. ESC is used for pre-identify a precise TSK FLS. While FLS provides an inference mechanism under cognitive uncertainty, computational NN offers learning, adaptation, fault-tolerance, and generalization. IV.

CASE STUDY

A. Experimetnal Setup In TCM systems, real-time data are first acquired from sensors (e.g. dynamometers on Fig. 4) put at different locations of the workpiece, tool and machinetool, then signal processing techniques (amplifiers) are used to extract valid data, and a decision making system (computers) is used to analyze the data and classify the results in order to make a more reliable estimation of the state of the tool and consequently of the machined parts themselves.

2016 IEEE International Conference on Fuzzy Systems (FUZZ)

1.

F f is independent of f , but rather depends on

VB and the depth of cut, denoted a p . 2.

Fc depends on a p and f , while being only

weakly dependent on VB. So, in this paper f and the measurement Fc are used to determine a p , and the measurement F f is used to determine VB without requiring a p as an input variable. Cutting speed of each cut was selected to ensure approximately the same share in the tool wear. VB was measured after carrying out each sequence. The value for F f and Fc were measured corresponding to a single cut Fig. 3. Flow chart of ANFIS algorithm based on ESC

Fig 4. Experimental setup

The data used in this paper, was collected from the experiments conducted on a conventional lathe TUD-50. A CSRPR 2525 tool holder equipped with a TiN-Al2O3TiCN coated sintered carbide insert SNUN 120408 was used in the test. The whole TCM experimental setup is shown in Fig. 4.

using a Kistlter 9263 dynamometer during 5-s intervals while the cut was executed. Recent research has attempted to investigate the application of multiple sensors with complementing characteristics to provide a robust estimate of tool wear condition. Since the inserts used in the experiments had a soft, cobalt-enriched layer of substrate under the coating, the tool life had a tendency to end suddenly after this coating wore through. The experiments were carried out until a tool failure occurred. The experiments were carried out until a tool failure occurred. Ten cycles were performed until a sudden rise of the flank wear VB occurred, reaching approximately 0.5 mm. Fig. 6 presents the cutting force components Fc and F f versus VB obtained in the experiments.

B. Data Acquisition To simulate factory floor conditions, six sets of cutting parameters were selected and applied in sequence as presented in Fig. 5.

Fig. 5.Cutting parameters used in experiments, where f is feed rate and t is time.

Fig. 6.Cutting force components Fc and F versus tool wear VB f obtained in the experiments with six sets of cutting parameters

During machining, F f and Fc were recorded while the tool wear was manually measured after each test. For our purposes VB was estimated from three input sources: f , F f and Fc .

C. Signal Processing By the ESC based ANFIS fuzzy modelling algorithm, a 10 rule neural fuzzy inference system can be used to describe the VB with f , F f and Fc as input variables.

The choice of input variables was based on the following two experimental based observations:

TABLE II lists the 10 cluster centers obtained by ESC method from the experiment.

2016 IEEE International Conference on Fuzzy Systems (FUZZ)

1979

Cluster

TABLE II CLUSTER CENTER Feed Cutting Feed force (mm) force (N) (N)

Tool wear (mm)

1

0.33

1044

405

0.158

2

0.47

1413

431

0.156

3

0.17

660

417

0.163

4

0.47

2656

816

0.167

5

0.33

1993

751

0.17

6

0.24

865

568

0.287

7

0.47

1474

627

0.302

8

0.24

749

273

0.08

9

0.33

1120

941

0.455

10

0.47

2708

1000

0.307

The related parameters are listed in TABLE III. Hypersphere cluster radius, squash factor, accept ratio and reject ratio are subtractive clustering parameters. Their optimal values are searched from the intervals [0.15, 1], [0.05, 2], [0, 0.9] and [0, 1] in order to get the best TCM model from the cutting parameter data sets ( f , F f , Fc and VB).

Fig 8. ANFIS model structure

EANF-TCM output surfaces are shown in Fig. 9. They are 3D representations of neural fuzzy inference system with two input variables and one output variable. Here Fig. 9(a) is with f, Fc , and VB. Fig. 9(b) is with f, F f and VB. And Fig. 9(c) is with Fc , F f and VB.

TABLE III. PARAMETERS FOR SUBTRACTIVE CLUSTERING AND ANFIS Parameter

Subtractive clustering

ANFIS

Value

hypersphere cluster radius

0.65

squash factor

1

accept ratio

1

reject ratio

0.9

Number of nodes

86

Number of linear parameters

40

Number of nonlinear parameters

60

Number of total parameters

100

Number of training data pair

71

(a) Output surface with f, Fc , and VB

The hybrid ANFIS is the key of proposed EANF-TCM. Its 10 fuzzy rules are shown in Fig. 7.

(b) Output surface with f, F f , and VB

Fig.7. 10 fuzzy rules for the EANF-TCM

ANFIS model structure is shown in Fig. 8. ANFIS has 86 neural nodes with 100 parameters in which 40 are linear and 60 are nonlinear. 71 datasets in the experiment are used to train the hybrid neural fuzzy system. There are totally 10 rules inside the ANFIS. 1980

(c) Output surface with Fc , F f , and VB Fig 9. ANFIS output surfaces

2016 IEEE International Conference on Fuzzy Systems (FUZZ)

D. Comparisons For these considered AI methods - NN, Mamdani FL, NF, TSK FL, ANFIS, the quality of the tool wear estimation was evaluated using root-mean-square-error (rmse): 2 (3) rmse = ¦ (VBm − VBe ) N and maximum error (max):

max = max(VBm − VBe )

(4)

NF are based on expert knowledge but fuzzy rules in TSK fuzzy system are generated directly using the experimental data, so the TSK modeling has better performance. ANFIS combines the advantage of the NN and FLS, its two-pass learning cycle, the least-square method and the back-propagation gradient descent method, ensures the accuracy of the systematic approach. Therefore it provides the best performance with the least monitoring errors. V.

where VBm , VBe are measured and estimated flank wear respectively, and N is the number of patterns (71 for this experiment). Figure 10 summarizes the results of EANF-TCM and compares them with the five different AI methods described in [1-4, 6] applied to the same experimental arrangements.

CONCLUSIONS

This paper introduces an intelligent TCM system – EANF-TCM. ANFIS overcomes the limitations of the individual modelling techniques and integrates the best features of FLS and NN. While FLS provides an inference mechanism under cognitive uncertainty, computational NN offers learning, adaptation, faulttolerance, and generalization. The ESC ensures the high precision of the fuzzy system identification. Hence, in our case ANFIS provides high accuracy and high reliability of the tool wear prediction over a range of cutting conditions from the input-output data (f, Fc , F f and VB) acquired from sensors. The experimental results show its effectiveness and a higher performance by achieving lower rms and max prediction errors in comparison with four other AI methods tested on the same data. Furthermore, the rules linking the inputs to the outputs are explicitly built into the prediction model, making them available for the operators and engineers for a possible understanding of the tool wear behavior. However, one major limitation of this work is the limited data samples. In order to get better representative results, more accurate continuous experimental data sets need to be acquired during the machining process. REFERENCES [1]

Fig. 10. Tool wear monitoring using different AI methods: Mamdani FL, NN, NN based fuzzy system (NF), TSK FL and ANFIS

From TABLE IV, the proposed TCM system with ANFIS has the lowest rmse and the smallest maximum error, hence, the best overall performances in terms of tool wear prediction.

[2]

[3]

[4] TABLE IV. SUMMARY OF RMSE AND MAXIMUM ERROR(MAX) FROM THE EXPERIMENTAL RESULTS WITH DIFFERENT AI METHODS AI methods rmse (mm) max(mm) [5] NN

0.0150

0.0360

0.0240

0.0680

NF

0.0140

0.0300

TSK FL

0.0110

0.0230

ANFIS

0.0086

0.0167

Mamdani FL

The results can be explained using the characters of those intelligent systems. Rules in NN, Mamdani FL and

[6]

[7]

[8]

M. Balazinski, E. Czogala, K. Jemielniak, J. Leski J, “Tool condition monitoring using artificial intelligence methods”, Eng. App. of Art. Int., vol. 15, pp. 73–80, 2002. K. Jemielniak, L. Kwiatkowski, P. Wrzosek, “Diagnosis of tool wear based on cutting forces and acoustic emission measures as inputs to neural network”, J. of Int. Man., vol. 9, pp. 447–455, 1998. L. Baron, S. Archiche, M. Balazinski, “Fuzzy decisions system knowledge base generation using a genetic algorithm”, Int. J. of App. Reasoning, vol. 28, no. 2-3, pp. 125-148, 2001. Q. Ren, L.Baron, M. Balazinski and K. Jemielniak, “TSK Fuzzy Modeling for Tool Wear Condition in Turning Processes: An Experimental Study”, Eng. App. of Art. Int. vol. 24, no. 2, pp. 260265, 2011. Q Ren, M Balazinski, L Baron, K Jemielniak, R Botez, “Type-2 fuzzy tool condition monitoring system based on acoustic emission in micromilling”, Information science, vol. 255, pp. 121– 134, 2015 S. Achiche, M. Balazinski, L. Baron, K. Jemielniak K, “Tool wear monitoring using genetically-generated fuzzy knowledge bases”, Eng. App. of Art. Int., vol. 15, no. 3-4, pp. 303-314, 2002. J.-S. R. Jang, “ANFIS: Adaptive-Network-based Fuzzy Inference Systems”, IEEE Transactions on Systems, Man, and Cybernetics, vol.23, no. 3, pp. 665-685, 1993. A. Gajate , R. Haber, R. Toro, P. Vega, A. Bustillo, “Tool wear monitoring using neuro-fuzzy techniques: a comparative study in a

2016 IEEE International Conference on Fuzzy Systems (FUZZ)

1981

turning process”, Journal of Intelligent Manufacturing, vol. 23, no. 3, pp 869-882, 2012. [9] X. Li, A. Djordjevich, P. K. Venuvinod, “Current-sensor-based feed cutting force intelligent estimation and tool wear condition monitoring”, IEEE Transactions on Industrial Electronics, vol 47, no. 3, pp. 697 - 702, 2000. [10] X. Li, H. X. Li, X.P. Guan, R.Du, “Fuzzy estimation of feedcutting force from current measurement-a case study on intelligenttool wear condition monitoring”, IEEE Transactions on Systems, Man, and Cybernetics, Part C: Applications and Reviews, vol. 34, no. 4, pp. 506-512, 2004. [11] M. Rizal, J. Ghani, M. Nuawi, C. Haron, “Online tool wear prediction system in the turning process using an adaptive neurofuzzyinference system”, Applied Soft Computing, vol. 13, no. 4, pp. 1960–1968, 2013. [12] A. Azmi, “Monitoring of tool wear using measured machining forces and neuro-fuzzy modelling approaches during machining of GFRP composites”, Advances in Engineering Software, vol. 82, pp. 53–64, 2015.

1982

[13] K. Demirli, S. X. Cheng, P. Muthukumaran, “Subtractive clustering based modeling of job sequencing with parametric search”. Fuzzy Sets and Systems. vol. 137, no.2, pp.235-270, 2003 [14] M. Sugeno, G. Kang, “Structure identification of fuzzy model”. Fuzzy Sets and Systems, vol. 28 , pp. 1, pp.15-33, 1988. [15] T. Takagi, M. Sugeno, M., “Fuzzy identification of systems and its applications to modelling and control”. IEEE Transaction on Systems, Man, and Cybernetics, vol. 15, no. 1, pp. 116-132, 1985. [16] E. H. Mamdani, S. Assilian, “Applications of Fuzzy Algorithms for Control of Simple Dynamic Plant”, Proc. Inst. Elec. Eng., vol. 121, pp. 1585-1588, 1974. [17] S. L. Chiu, “Fuzzy model identification based on cluster estimation”. International Journal on Fuzzy Systems, vol. 2, pp. 267-278, 1994 [18] Q. Ren, M. Balazinski, K. Jemielniak, L. Baron and S. Achiche, “Experimental and Fuzzy Modeling Analysis on Dynamic Cutting Force in Micro-milling”, Soft Computing, vol. 17, pp. 1687-1697, 2013. [19] D. Rumelhart, G. Hinton, and R. Williams, “Learning representations by back-propagating errors”, Nature, vol. 323, pp. 533-536, 1986.

2016 IEEE International Conference on Fuzzy Systems (FUZZ)