The monitoring of the turning tool wear process using an artificial neural network G. C. Balan, A. Epureanu Universitatea Dunarea de Jos din Galati, Str. Domneasca nr. 47, Galati–800 008, Romania Abstract The study of machine tool dynamics is performed here as “monitoring”, which involves the checking and improving of machine functioning. Signals collected from certain sensors are processed by a computer. These data then lead to the monitoring decision, which is to associate the current state of operation with one of the classes from a set of known classes. For monitoring in turning, the classes (tool conditions) are shown. The experimental setup, experimental results and data processing are presented. For the monitoring of the tool wear, an artificial neural network (ANN) is used. Keywords: monitoring, turning, ANN (= artificial-neural-network) 1. Introduction where x is the set (vector) of monitoring indices x = [ x 1 , x 2 , … , x m ] , and t – the admissible limit In turning, tool flank wear is one of the major values. factors contributing to the geometric error and thermal The samples can be registered as in table 1, in damage in a machined work-piece. Tool wear not only which m is the number of monitoring indices, n is the directly reduces the part geometry accuracy but also number of classes, and N - the number of samples. So increases the cutting forces drastically. By changing the x k =[x (k,1), x(k,2), . . . , x (k,m)]. represents “the worn tool before or just at the time it fails, the loss vector k”, and c ( x k ) ∈ [c1, c 2,..., c n ] indicates the caused by defect product can be reduced greatly and fact that in this recording the result was one of the thus product quality and reliability is improved. classes: c1 , c 2 ,... , c n . To accomplish these objectives, artificial The function can be introduced Q:c→ x, (2) intelligence methods are the most modern means. which is “obscure“ because on it you can do only The study of machine-tool dynamics is realised indirect measurements which are, or which can be here as “monitoring”, which means the checking and bound to the function. If for Q no theoretical relation improving of machine functioning. Signals collected can be obtained, a two phases data interpretation from certain sensors are processed by a computer. method involving learning and classification can be These data then lead to the monitoring decision, which used. For a set of samples in which both x and c are is to associate the current state of operation with one of known (a part of data from the table 1), in the learning the classes from a set of known classes (process phase an empirical relation between x and c is formed. conditions) c = [ c 1 , c 2 , …. , c n ], according to: In the classification phase the other part of the table 1 if t i n f < x ≤ t s u p , then c = c i , (1) is used to predicting of c, thus testing and adjusting Table 1 The samples Sam - Monitoring indices CLASSES ples x1 x2 ..... xi ...... xm ( Process conditions ) x1 x( 1,1 ) x( 1,2 ) ......... x( 1,i ) ......... x( 1,m ) c ( x 1 ) ∈ [ c 1, c 2,....,c n ] x2 x( 2,1 ) x( 2,2 ) ......... x( 2,i ) ......... x( 2,m ) c ( x 2 ) ∈ [ c 1, c 2,....,c n ] ... ... ... ... ... ... ... ..... xN x( N,1) x( N,2) ......... x( N,i ) ....... x( N,m ) c ( x N ) ∈ [ c 1, c 2,....,c n ] the empirical relation. back edge in B zone is V B B = 0.3 mm if this has a Thus, function Q is reversed: Q -1:x→ c . (3) regular form ; - the maximum breadth of the same Now the empirical relation is able to classify a new wear V B B max = 0.6 mm , if this has an irregular form”. sample x in a certain class c k . So Q can have different C = rε = the top radius of the tool = max. 2 mm. aspects: an analytical one (with a little probability), an The classes c1 , c3 , c4 and c5 in table 2 will be artificial neural network (ANN), a pattern recognition, adapted in accordance with these prescriptions and they a fuzzy system, etc. will be rearranged as in table 3.  shows, in Du, et al.  represent the main source of connection with class c2 in table 2, that “the advanced inspiration for this paper, our contribution being the (catastrophic) deterioration means the intense improvement of the classes and of the monitoring deterioration of the cutting-tool edges after a period of indices. normal cutting, under the combined action of all the factors involved in the processing“. For the quantitative evaluation of this state, an overfulfilment with more 2. Monitoring in turning than 0.1 mm of the wear criteria, and the overtaking till 0.1 mm showing a severe wear are proposed. Table 2  shows and defines the classes (tool Class c7 in table 2 will be eliminated, so that conditions) for monitoring in turning. A chip on the continual cutting (which might occur if keyway exits) tool inserts bigger than 0.05 mm2 identified “Tool does not occur on the lathe, the making of the keyway breakage”. “Chatter” is identified by the high being the last operation in the shaft working process. In frequency noise and the chatter marks on the machined conclusion, the n = 7 classes referring to the working surface. “Transient cutting” (or intermittent cutting) is conditions are those in table 3, where to the first three produced by machining a work piece that has a slot classes the working conditions are normal and the along the feed direction. others are abnormal. The conditions in table 2 will be worked in order In order to obtain the monitoring indices the to coordinate them with [5,6] from which we quote: following will be used: “According to figure 1, the usual criteria for the wear - strain gauges glued on the cutting-tool, which measure the components of the cutting force (Fy - the repelling force , Fz - the main force); - accelerations of cutter holder vibrations (a x , a y , a z). The signals of the sensors are registered simultaneously by means of device SPIDER 8 (H.B.M.). It is 4.8 carrier-frequency technology for S/G (strain gauges) or inductive transducers. Spider 8 is an electronic measuring system for PC, for electric measurement of mechanical variables such as strain, force pressure, path, acceleration and for temperatures. Each channel works with a separate A/D converter which allows measuring rates from 1/s to 9600/s. In  we made “the monitoring simulation”. The Fig. 1 The cutting-tool wear vibrograms representing: - the component variations of of the rapid steel cutting-tools and of the cutting-tools the cutting force; - the relative displacement between with hard-cutting alloy plates are: - the average tool and piece, on the repelling direction; - the power breadth of the wear by the separation from the main furnished by the electric engine, are realized with the Table 2 The classes (tool conditions) Class Tool Identification Identification Number of conditions on cutter on workpiece samples c1 Normal wear < 0 . 1 mm --------- M 1 = 144 c2 Tool breakage chipping > 0.05 mm2 --------- M 2 = 49 c3 Slight wear 0.11 < wear < 0.15 mm --------- M 3 = 114 c4 Medium wear 0.16 < wear < 0.30 mm --------- M 4 = 114 c5 Severe wear 0.31 mm < wear --------- M 5 = 114 c6 Chatter Fresh tool Chatter marks M 6 = 61 c7 Transient cutting Fresh tool An axial slot M 7 = 15 c8 Air cutting --------- --------- M 8 = 13 Table 3 The classes in turning Class Tool Identification Identification conditions on cutter on work-piece c1 Normal V B < 0.1 mm , or --------- V B m a x < 0.2 mm c2 Slight wear 0.11 < V B < 0.2 mm , or --------- 0.21 < V B m a x < 0.4 mm c3 Medium wear 0.21 < V B < 0.3 mm , or --------- 0.41 < V B m a x < 0.6 mm c4 Severe wear 0.31 < V B < 0.4 mm , or --------- 0.61 < VB m a x < 0.7 mm c5 Tool V B > 0.41 mm , or --------- breakage V B m a x > 0.71mm c6 Chatter Fresh tool Chatter marks c7 Air cutting --------- --------- functions RANDN and RAND (from MATLAB). Based looked irregular (like a triangle), so the wear criterion on them, 11 monitoring indices are calculated. The VBmax was used. ANN with 11 inputs (the number of monitoring indices) and 8 outputs (the number of classes) is 3. Experimental results realized with 3 layers. In  the experimental setup and experimental results 191 recordings were made and the parameters of are presented: the Spider device were set on: sampling frequency = - Components of cutting force were calculated on the 9600 /s, no. of periods = 1, samples / period = 4800; i. basis of the experimental study of the lathe cutting-tool e. the device samples the received signals with a bending and with the help of two strain gauges, stuck frequency of 9600 Hz, but it can send to PC a on the lathe cutting-tool and connected to SPIDER. recording with 4800 samples, which corresponds to 0.5 The recordings were made during the longitudinal sec. Each working session lasted nearly 30 sec., and by turning of a OLC 45 cylinder (Φ 113, L = 1000), with a half this time, the Spider device was connected for one lathe cutting-tool with metal carbide P20 and ℜ = 45o. second. It results: The cutting working conditions were: piece F z = 1136 ( ε2 I n r - ε1 I n r ) [daN], (4) diameters D = 113 ÷ 93.4 mm, the cutting depths t = where ε1 I n r and ε2 I n r are the registered relative 0.5÷3mm, rotations n = 63÷500 rot/min., longitudinal deformations of the strain gauges. advances s = 0.024 ÷ 0.5 mm/rot.; cutting speeds v = - Cutter-holder accelerations (3 Bruel&Kjaer 4329 type π Dmed n / 1000 = 22.3÷177.4 m/min. accelerometers were mounted on a plate solidary with On each passing on the whole piece length (L = cutter holder). 1000 mm) the t constant was preserved, while s or n - Tool wear (after each passing, the tool wear was varied. measured with the help of a Brinell lens. The wear spot 12 monitoring indices were calculated: Z1 = v → cutting speed; layer s2 = 27. The input matrix p has the dimensions 12 Z2 = t → cutting depth; (monitoring indices) x 655 (recordings), and the output Z3 = s → longitudinal advance; matrix y has dimensions 7 (classes) x 655. The training functions are: tf1 = purelin, tf2 = tansig, tf3 = logsig, Z4 = Fz → average value of the main cutting force; therefore the output vectors have 7 elements, with Z5 → Fz variation range (recording which has 960 values in domain (0, 1). samples was split into 4 equal parts – 240 samples each The first runs (with the training functions trainrp, – and the maximum and minimum values were trainscg, etc.) showed errors only in the positions calculated for each part; X5 is the difference between corresponding to the recordings in classes c5 and c7, i. the maximum and minimum average values); e. in the case of those classes which have the fewest Z6 → number of intersections of oscillogram Fz with its recordings. The increase of the number of recordings – average value Fz ; without making new experiments – may be performed Z7 → the average of Fz power spectral density in the by adding the same recordings several times to the frequencies range 1 - 2400 Hz ; same class, which may be eventually affected by a Z8 → the average of Fz power spectral density in the noise of an average value of 0.1. frequencies range 2401 - 4800 Hz ; Consequently: - we fourfold the recordings in class c5: an average Z9 → the average of Fz power spectral density in the value noise of 0.1 is attached to the first set, the second frequencies range 4801 - 9600 Hz ; set is identical with the original one, an average value Z10 → the average of azinr power spectral density in the noise of 0.15 will be attached to the third set; frequencies range 1 - 2400 Hz ; - we threefold the recordings in class c7, acting by a Z11 → the average of azinr power spectral density in the way of analogy with the foregoing (only with first and frequencies range 2401 - 4800 Hz ; the second set). Z12 → the average of azinr power spectral density in the Therefore, c5 will consist of 60 recordings, c7 – 75 frequencies range 4801 - 9600 Hz . recordings, and the number of columns in the above matrices grows from 655 to 750. Using the following instructions: 4. Use of ANN on monitoring of the tool wear ind = find(c = = 0); dim = length(ind); er = dim / 750, we find: the “c” elements indices which are null, the The recordings are divided in two sets, “Learning” “ind” number of elements as well as the network error, and “Classification”, the first set having 60% of the respectively. number of recordings (those noted with “a”, “c” and The results of a run are presented in what follows: “e”), and the second set having the recordings noted R = 8 ; Q = 750 with “b” and “d”. There was a redundancy in the data set, since the Columns “az” and “Fz” belonging to the recordings principal component analysis has reduced the size of in the “Learning” set, will be transferred into the input vectors from 12 to 8. MATLAB, where 7 tables corresponding to the 7 TRAINRP, Epoch 0/300, MSE 1.53359/0, Gradient classes will be made up. 0.292509/1e-006 The table analysis shows that – except for two TRAINRP, Epoch 25/300, MSE 0.487294/0, Gradient cases – the recordings in one class have quite similar 0.0150939/1e-006 average values of Fz and, as expected, values which TRAINRP, Epoch 43/300, MSE 0.477332/0, Gradient grow (as the wear grows) from one class to another. 0.0148044/1e-006 The training set will contain the recordings in the TRAINRP, Validation stop. “Learning” set, therefore 60% of the total number of ind = recordings. Recordings “b” in the “Classification” set Columns 1 through 18 will be allotted to the validation set, and recordings “d” 440 441 445 446 447 448 449 450 454 455 will be input into the testing set; therefore each set has 456 457 458 459 463 468 472 473 20 % of the recordings. Columns 19 through 37 ANN consists of 3 layers, with 2 hidden layers, 474 475 476 477 481 482 483 519 520 521 No. of Inputs = 12 (monitoring indices), No. Output 533 545 594 595 596 646 647 658 672 Neurons s3 = 7(classes), No. Neurons in the first dim = 37 ; er = dim / 750 = 0.0493 = 4.93 %. hidden layer s1 = 23, No. Neurons in the second hidden It is a useful diagnostic tool to plot the training, validation and test errors to check the progress of training. The result as shown in the figure 2 is reasonable, since the test set error and the validation set error has similar characteristics. Fig. 3. The experimental setup should transmit a recording to the PC, based on which with table 3). Then “Spider” connects (therefore itself again, and so on. In case the class exceeds 3 ANN will say the class the processing is into (in conformity “abnormal”), PC will produce a sonorous signal, or stop the processing. To detail: the recording has 4800 samples, being a Fig. 2. The progress of training table in EXCEL with 4800 rows and 3 columns (A = In other runs: ε 1inreg, B = ε 2inreg, C = a z). The transfer Spider → PC - carried out under the same circumstances, results is carried out within nearly 1 min. Out of this table a were twice as above, and once as follows: new (smaller) table is selected and it will consists of 51 epochs; dim = 27 ; er = 3.6 %; 960 rows and 4 columns, the first element being - without “init” function (to reinitialize weights and selected at random, at a location higher than A500. biases), in two runs the errors were 4.67% and 13.3%; Function Fz (= 1136*A-1136*B) is calculated in fourth - with the trainscg training function without “init” the column, according to formula (4). This table is error was 5.33%, whereas with “init”- 25.6%. transferred in MATLAB, where 12 monitoring indices will be calculated and then they are presented at the ANN input. 5. Monitoring To see how ANN responds we use some of the recordings; the results are presented in table 4. With Monitoring the tool wear involves that during the the first 4 recordings, after the first run a second run continuous process of cutting (fig. 3) the “Spider” was carried out, and it lasted 2 min., the class given by device should be connected into the system, and it ANN coinciding with the one of the first run. Table 4 The answer of ANN v t s Error Class Time No Rec. Class [m/min] [mm] [mm/rot] Epochs % from [min] ANN 0 1 2 3 4 5 6 7 8 9 1 046 1 69 0.5 0.25 43 4.93 1 10 2 183 5 71 3 0.302 40 4.4 5 8 3 168 4 75.4 2 0.302 76 2.5 4 6 4 174 4 120.6 2 0.416 67 3.47 4 8 5 099 2 83.2 1 0.334 43 4.93 2 6 190 7 0 0 0 47 3.6 7 7 7 095 6 83.2 1 0.353 47 23.87 6 7 The coincidence of the values in columns 2 and 8 To asses the soft efficiency, the runs will be shows that ANN provides correct outputs. resumed, as in table 4, but for one third of the 44 Table 5 The answer of ANN for class 3 No. Recor. v t s L Epochs Error Class [m/min] [mm] [mm/rot] [mm] % in ANN 0 1 2 3 4 5 6 7 8 1 115 130.6 1.2 0.334 245 40 4.4 2 2 118 130.6 1.2 0.416 90 76 2.53 3 3 126 79.3 1.4 0.292 770 43 4.93 3 4 129 79.3 1.4 0.353 630 43 4.93 3 5 133 63.4 1.4 0.353 425 40 4.4 4 6 136 63.4 1.4 0.292 300 76 2.53 3 7 139 63.4 1.4 0.212 200 64 2.67 3 8 142 63.4 1.4 0.167 120 67 3.47 3 9 146 77.3 1.5 0.375 780 43 4.93 3 10 149 77.3 1.5 0.302 615 47 3.6 3 11 152 77.3 1.5 0.25 480 47 23.87 3 12 155 123.7 1.5 0.177 340 40 4.4 3 13 158 123.7 1.5 0.146 230 31 15.2 3 14 161 123.7 1.5 0.118 130 76 2.53 3 recordings of class 3. Table 5 presents the results; “L” possible. is the distance from the centre of the cut zone to universal. It is noticed that only the recordings under no. 1 and no. 5 did not give the correct class (3), but a 7. Acknowledgement neighboring class. Other two runs carried out for each of these recordings – with the modification of the This research was supported through two grants by location of the first element in the table (960 x 4) – Ministry of Education of Romania [1, 7]. showed the correct class (3). Therefore, with 25 runs only two were erroneous, the error amounting to 8 %. References We consider this error to decrease in the future if we  Balan, G, 2002, The monitoring of a lathe using an take several recordings in tables 4 and 5. artificial neural network, Grant type A nr. 33 445, Theme 19, Cod CNCSIS 451  Balan, G., Tarau, C., 2003, The monitoring 6- Conclusions simulation of a lathe, Mathematical & Computational The algorithm to monitor the tool wear making use Applications, an International Journal published by the of ANN proved efficient, the error range being below 5 Association for Scientific Research, Vol. 8, Nr. 3, pp. percent. 279-286  Balan G. and Epureanu A., 2005, The monitoring In the case of real monitoring, when the cutting is of a lathe using an artificial neural network (1-st part), continuous, to avoid “thermal no-compensation” a Annals of DAAAM for 2005 & Proceedings of the 16- cooling of the knife should be provided. However, th International DAAAM Symposium “Intelligent water can cause trouble in the circuits of strain gauges, Manufacturing…”, Croatia, p. 019-020. although they are protected (with Poxipol).  Du, R., Elbestawi, M. A., Wu, S. M., 1995, Consequently, for this step of the experiment, the strain Automated Monitoring of Manufacturing Processes, gauges will be removed and the cutting force Part 1: Monitoring Methods, Part 2: Applications, components will be measured by averages of a ASME Journal of Engineering for Industry, may, vol. KISLER device (Austria). Moreover, column 9 in table 4 shows that the 117, Part 1-pp. 121 - 132, Part 2 − pp .133 - 141. current hard provides delayed information, i. e. we  STAS 12046 / 1 - 81, Cutting life testing. Wear. know that the tool is – for example - in class 4 (Severe General notions wear), when it may have reached class 5 (Breakage), or  STAS 12046 / 2 - 81, Cutting life testing. Tool - 6 (Chatters). Therefore, a highly specialized PC is life testing methods in turning tools. required, to reduce the responding time as much as  A. Epureanu, Contract nr. 22CEEXI03/’05, MEdC.
Pages to are hidden for
"The monitoring of the turning tool wear process using"Please download to view full document