Chip form monitoring in turning based on neural network processing of cutting force sensor data

T. Segreto^a, J. L. Andreasen^b, L. De Chiffre^b, R. Teti^a

^aUniversity of Naples Federico II, Naples, Italy

^bTechnical University of Denmark

Introduction

• Sensor monitoring of chip form in turning was performed through cutting force signal detection and analysis

• Longitudinal turning tests were carried out on carbon steel Ck45

• Decision making on chip form typology was performed using a supervised neural network (NN) approach, using diverse back-propagation feed-forward NN configurations

Material and Experimental Procedures

• Carbon steel Ck45 cylindrical bars were used as work material

• Longitudinal turning tests were carried out with:

• Cutting speed: 180 m/min

• Feed rate: 0.05 - 0.1 - 0.15 - 0.2 - 0.25 - 0.3 - 0.4 - 0.5 - 0.6 - 0.75 mm/rev

• Depth of cut: 0.5 - 1 - 2 - 3 mm

• Four chip form typologies were obtained according to ISO 3685

Snarled (2.3)	Long (2.1)	Short (5.2)	Loose (6.2)

Cutting Force Monitoring System

• For each cutting test, the feed force (FA) component was measured by a Kistler dynamometer

• Force signal acquisition was carried out with sampling frequency 2.5 kHz for a duration of 2 s (number of samples: 2.5 kS)

Sensor Signal Analysis

• The 2.5 kS feed force file from each cutting test was subdivided into 5 sub-files, each representing a 0.5 kS signal specimen
• Signal analysis was performed through a parametric method of spectral estimation: the spectral estimation problem becomes one of estimating the unknown spectrum parameters rather than the spectrum itself
• From the signal specimen, p features or predictor coefficients {a₁, ..., a_p}, characteristic of the spectrum model, are obtained through linear predictive analysis (LPA) with p = 4, 8, 16

Neural Network Data Processing

• Three-layer feed-forward back propagation NN were built with diverse configurations:

• Input layer: p = 4, 8 or 16 nodes for the input feature vector components a₁, ..., a_p

• Hidden layer: q = 4, 8, 16, 32 or 64 nodes depending on the number of input nodes

• Output layer: 1 node for chip form coded value output

Neural Network Pattern Recognition

• Two different pattern recognition procedures were applied

• A. Single chip form classification

• B. Favourable/unfavourable chip form identification

• NN training and testing was carried out using the leave-k-out method (k = 1)

CHIP FORM	CODED VALUE	TRAINING CASES
Snarled	0	45
Long	1	20
Loose	2	65
Short	3	10
Favourable (Loose/Short)	0	75
Unfavourable (Snarled/Long)	1	65

NN SR: Single Chip Form Classification

• The NN performance was evaluated in terms of Success Rate given by SR = (number of correct classifications)/(number of test cases)

• All NN total SR values for single chip form classification are low: the highest value, obtained with the 4-4-1 NN, was only 60%

• This is due to the very small number of training cases (Short is the lowest: only 10 cases over 140 total cases)

NN Total SR for different numbers of input and hidden nodes

• The NN total SR for single chip form classification decreases with increasing number of hidden nodes as well as with increasing number of input features

4-4-1 NN Configuration: Single Chip Form Classification

• A high number of misclassication cases (black symbols) are evidenced, particulary for the Short chip form, showing deficient learning conditions for this type of NN knowledge acquisition

 

NN SR: Favourable/Unfavourable Chip Form Identification

NN Total SR for different numbers of input and hidden nodes

• The NN total SR for favourable/unfavourable chip form classification decreases with increasing number of hidden nodes as well as with increasing number of input features

4-4-1 NN Configuration: Favourable/unfavourable Chip Form Identification

• Only few uncorrect classification cases (black symbols) are evidenced, revealing more exhaustive learning conditions for the favourable/unfavourable NN identification procedure

Conclusion

• Chip form monitoring during turning of carbon steel Ck45 was carried out through cutting force signal detection and analysis

• Decision making on chip form was performed through two NN approaches:

• a) Single chip form classification

• b) Favourable/unfavourable chip form identification

• The single chip form classification procedure presented a low performance, the highest NN total SR being only 60%

• The favourable/unfavourable chip form identification procedure provided for high perfomance, yielding a NN total SR as high as 87%

• On this basis, the favourable/unfavourable identification procedure can be actually implemeted for on-line chip form monitoring in industrial applications