Fade performance prediction of automotive friction ...

5 downloads 39021 Views 1MB Size Report
Available online 10 October 2006 ... neural networks have been trained by the 5 training algorithms. ... Keywords: Automotive friction material; Fade performance; Prediction; Artificial neural ...... content 9%, and free phenol content 2%.
Wear 262 (2007) 778–790

Fade performance prediction of automotive friction materials by means of artificial neural networks ˇ Duboka D. Aleksendri´c ∗ , C. Automotive Department, Faculty of Mechanical Engineering, University of Belgrade, Kraljice Marije 16, 11120 Belgrade 35, Serbia Received 10 May 2005; received in revised form 28 August 2006; accepted 31 August 2006 Available online 10 October 2006

Abstract Temperature sensitivity of friction materials has always been a critical aspect while ensuring their smooth and reliable functioning, and that sensitivity need to be constantly optimized. The performance of friction materials at elevated temperatures is defined by their fading performance. In this paper, possibilities for predicting the fading performance of the friction materials, regarding their formulation and manufacturing conditions, have been investigated by means of artificial neural networks. The neural modelling of the friction materials behaviour at elevated temperatures has been based on the two different training data sets regarding the number, type, and distribution of the stored data. The first training data set is consisted by 360 data related to cold, fading, and recovery performance. These data have been used for developing of the neural model for predicting not only the fading performance but also cold and recovery performance. The second training data set, consisted by 120 data, has been used for developing the neural model that is going to be only used for predicting the fading performance of the friction materials. In this paper, 18 neural networks have been trained by the 5 training algorithms. These networks have been tested by the testing data set formed using the parameters of formulating, manufacturing, and testing of the two friction materials which input parameters were completely unknown for the networks. © 2006 Elsevier B.V. All rights reserved. Keywords: Automotive friction material; Fade performance; Prediction; Artificial neural networks; Neural models

1. Introduction Competitive advantages in the industry of friction materials need to be reached by appropriate management of friction material formulation and manufacturing conditions changing and their skilful implementation in a cost effective manner. Management of these changes can have decisive influence on whether a new friction material can be launched into the market. It is important because the industry of friction material manufacturing has been always focused on greater customer satisfaction, i.e. (i) improved friction stability, (ii) improved life, (iii) no judder, (iv) no noise, (v) improved rotor compatibility [1]. The basic requirements imposed on friction materials behaviour are related to the level of the friction coefficient and its stability versus different operating regimes. Thus, the brake performance is mostly determined by the friction material behaviour. Fric-



Corresponding author. Tel.: +381 11 3370358; fax: +381 11 3370364. E-mail address: [email protected] (D. Aleksendri´c).

0043-1648/$ – see front matter © 2006 Elsevier B.V. All rights reserved. doi:10.1016/j.wear.2006.08.013

tion material performance need to be characterized by relatively high and stable friction force, reliable strength, and good wear resistance irrespective of temperature, humidity, age, degree of wear and corrosion, presence of dirt and water spraying from the road, etc. Accordingly, friction materials performance becoming more demanding and many formulations have to be evaluated. It is well known that sensitivity of the friction coefficient to the operating conditions, i.e. sliding speed and/or applied pressure and/or temperature are substantially influenced by friction pair contact situation. The contact situation between a cast iron brake disc and an organic brake pad is complicated and can be dramatically changed for different material combinations. It is a result of the wide diversity in properties (mechanical and chemical) of the pad’s constituents. To understand the development of friction materials, it is necessary to understand the interfacial relations between rubbing surface. There are many asperities between the rubbing surfaces because the formation of many micro-contact points, these points are dynamically changing from place to place in fraction of seconds during rubbing [2–5]. The total contact area

ˇ Duboka / Wear 262 (2007) 778–790 D. Aleksendri´c, C.

is unknown and depends on friction pair interaction. Furthermore, it is known that when two bodies slide against each other with a relative speed and a contact pressure, frictional heat is generated at their sliding interface [3]. The subsequent thermo-mechanical deformation of the bodies modifies the contact profile and the pressure distribution, altering the frictional heat. The friction heat generated during braking application easily raises the temperature at the friction interface beyond the glass transition temperature of the binder resin and often rises above decomposition temperature. The gas evolution at the braking interfaces because of pyrolysis and thermal degradation of the material results in the friction force decreasing at elevated temperatures. The thermal load in the contact of brake pads is not built up homogeneously and hot rings can be often seen on the brake disc changing their radius [4]. The instability of the friction coefficient after a certain number of brake applications is common and depends on contact interface situation. In general, the change of friction coefficient during sliding depends on the changes of the real area of contacts at the friction interface, the strength of the binder resin, the frictional properties of ingredients at elevated temperatures, and friction film formation. The friction film or third body layer is produced incessantly by maintaining a certain thickness and it is composed of carbonaceous reaction products, unreacted constituents, oxide from metallic ingredients, etc. [6–8]. Friction characteristics of the transfer layer between a brake disc and pads developed during subsequent braking influence the braking effectiveness at elevated temperature. This important friction material characteristic is referred to fade and has decisive importance in the friction materials performance evaluation. The friction materials performance can be generally classified into five different groups: (i) cold performance 1 (before braking at elevated temperatures), (ii) fading performance (friction material behaviour under subsequent braking, i.e. at elevated temperatures), (iii) recovery performance (braking at elevated temperature after period of cooling), (iv) cold performance 2 (after braking at elevated temperatures), and (v) wear behaviour under different temperature regimes. It is obvious that friction materials are supposed to satisfy very complex and mutually opposed requirements. That is why prediction of the friction materials performance versus influences of their formulations, manufacturing, and operating conditions can help in friction materials developing. Therefore, there is no doubt that engineering changes is a fact [9]. To support the future decision-making related to the mentioned influencing factors, new development policy has to be introduced in order to change the moment for decision-making about the friction material performance. This paper is going to investigate how artificial intelligence can help in the friction material development using capabilities of artificial neural networks. In this paper, possibilities for predicting the friction materials fading performance, regarding influences of their formulation and manufacturing conditions, have been investigated by means of artificial neural networks. The neural modelling of the friction materials performance at elevated temperatures has been investigated by introducing the two different training data sets of the neural networks. The difference between these training data sets is related to the size of the training sets, data dis-

779

tribution, and type of the stored data. The first training data set is consisted by 360, randomly distributed, data related to cold, fading, and recovery performance. The main goal with this training data set is to investigate possibilities for developing the neural model of the friction materials behaviour capable to predict the fading performance of the friction materials but at the same time to be able to predict the friction materials behaviour under testing conditions related to the cold and recovery performance. The second training data set have been obtained by formulating, manufacturing, and testing of the eight different friction materials (120 data). This training data set has been used for developing the neural model that is going to be only used for predicting the fading performance of the friction materials. 2. Experiment Temperature sensitivity of the friction materials has always been a critical aspect while ensuring their smooth and reliable functioning. It is particularly related to front brakes that absorb a major amount (up to 80%) of the vehicle total kinetic energy. This kinetic energy is converted into heat causing the generation of high temperature up to 400 ◦ C on the disc. The severity of such temperature rise is further manifested in the form of a very high flash temperature at the contacting asperities (600–800 ◦ C). At such high temperatures, friction force suffers from a loss of effectiveness called ␮-fade [10,11]. This loss of effectiveness cannot be easily predicted due to different ingredients used for friction material’s mixture, its different formulations, and different setting of manufacturing parameters. It is known that organic constituents, fibrous materials, and solid lubricants play important role in establishing the transfer layer at the friction interface [12,13]. The clear connection between component mixture, used manufacturing conditions and friction layer and the connection between friction layer and friction behaviour of the system is not known yet [4]. There is no doubt that measures towards optimization of the friction materials fade performance have to be taken but trial and error method needs to be eliminated due to cost and time consuming. On the other hand, brake pads usually contain more than 20 different ingredients and existence of highly complex tribological surface layer, the frictional behaviour of the system can be differently modulated. It was mentioned that the sensitivity of friction material’s fade performance to different operating conditions due to instantaneously established contact situation depends on the type of selected ingredients, friction material formulations and manufacturing conditions. It should be obvious that achievement of the stable friction coefficient under severe thermal load for different braking systems characteristics impose engineering changes that need to be done regarding ingredients selection, their proper formulations, and setting of the manufacturing conditions. The prediction of friction material sensitivity to temperature changing is critically important because the main constituent, such as the phenolic resin, changes its properties above glass transition temperature. The thermal decomposition of binder resin provokes changing of reinforcing capabilities of the used fibers and that is why the reduction of friction coefficient (fading) happens. These changes have to be subject of

780

ˇ Duboka / Wear 262 (2007) 778–790 D. Aleksendri´c, C.

the systems engineering investigation, and that is why artificial neural network technology has been employed for predicting friction material behaviour under high temperature load. Due to complex composition of friction materials and stochastic nature of braking phenomena, artificial intelligence abilities should be embedded into the process of friction materials development. Generally, intelligence can be defined as the ability to learn and understand, to solve problems and to make decisions [14]. That is why, brain-like information processing is needed. The field of neural networks, as artificial intelligence technology, could be introduced necessity changes in friction materials development. When the rules are not known in the situation with fuzzy or incomplete information, neural network technology can be applied. The neural networks abilities have been applied for searching numerical data for fuzzy rules, i.e. for “transformation” of known data and experience into rules. It is done because the neural networks are capable of handling a large set of complicated information and has the reputation of dealing with noisy data. That is why, artificial neural networks are good candidates for that due to their capabilities of non-linear behaviour, learning from experimental data and generalization. In order to learn complex input/output functional approximation, artificial neural networks need to be trained so that a particular input leads to a specific target output [15–17]. It is clear that experimental data are needed in order that input/output relationship may be generalized. The process of neural models developing is not trivial and involves many critical issues: (i) select data generator, (ii) data generation (define the range and distribution of the training data in the model input parameter space where training data needs to be generated), (iii) perform data generation, (iv) data preprocessing, (v) selection the architectures of neural networks, (vi) selection the training algorithms, (vii) neural network training, (viii) validation accuracy evaluation, and (ix) testing of the artificial neural networks [18]. Therefore, the neural model for predicting the friction materials performance has to be learned about influences of the inputs data on the outputs data. The preliminary step in a neural model development is identification of the model’s inputs and outputs. Input/output identification depends on model objectives and choice of the data generator. According to objectives of this paper to develop the neural model capable to predict fade performance of the friction materials versus their composition, manufacturing, and testing conditions, the input parameters are determined by formulations, manufacturing, and testing parameters versus the brake factor C changing (one-dimensional output parameter). The type of data generator depends on application and the availability. In this case, as a data generator, single-end full-scale inertia dynamometer has been used (see Fig. 1), developed at the laboratory for friction mechanism and braking systems—FRIMEKS (Automotive Department) Faculty of Mechanical Engineering, University of Belgrade. The brake performance resulting from the complex interrelated tribo-processes occurring during braking. These triboprocesses are mostly affected by the physicochemical properties of the friction materials’ ingredients under specified testing conditions. In order to establish relationships between input and

Fig. 1. Single-end full-scale inertia dynamometer.

output parameters space, the role of data generator is important from the point of view repeatability of the testing conditions. That is why, it has been decided to perform testing of the friction materials under strictly controlled conditions related to changes of pressure application, initial speed, initial temperature, and inertia of revolving masses. These testing conditions are chosen in order to simulate the real operating regimes by full-scale inertia dynamometer shown in Fig. 1. The dc motor (1) drives, via coupling (2), a set of six flywheels (3) providing in such way different inertia from 10 to 200 kg m2 independently mounted on the driving shaft (4). The flange (5) firmly jointed to the shaft (4), bears rotating part of the tested brake (disc) while immobile flange (6), being firmly connected to the foundation (7) is used for mounting stationary parts of the tested brake (calliper). The full-scale inertia dynamometer is equipped by PC-based automatic control and data acquisition system of pressure, speed, temperature, and braking torque at a sampling rate of 50 Hz. Thus, the braking torque and pressure application can be measured during the braking cycle, the average value of brake factor C is calculated regarding the changes of the braking torque and pressure, for known values of the piston diameter and effective brake disc radius, in the range of speed changing between 0.8 and 0.1 Vinitial . Importance of the input/output range and distribution strategy has already mentioned, once the range of input parameters is decided, the next step is to choose a sampling strategy. However, taking into consideration that the input parameters have been consisted by three different groups of data (composition of the friction material, manufacturing conditions, and testing conditions) for each group of these data need to be determined ranges and sampling strategy. The most complicated problem is related to selection of the ranges for the group of data related to the composition of friction materials. The problem related to the range and distribution of composition’s parameters is complicated by the fact that 150 different raw materials have been used in the friction material industry today. Furthermore, ingredients selection and their ranges depend on the type of friction material (metalic, semi-metalic, NAO). Similarly, the selection of input parameters and ranges for manufacturing conditions depending also on the type of friction material which performance are going to be modelled. Regarding the group of testing conditions, that task is much easier. Therefore, by inertia dynamometer as a data generator shown in Fig. 1, the friction material mounted

ˇ Duboka / Wear 262 (2007) 778–790 D. Aleksendri´c, C.

781

Table 1 Testing methodology Test number

Test

Pressure line application (bar)

Initial speed (km/h)

Temperature (◦ C)

Braking number

1. 2. 3.

Burnishing Cold performance 1 Fading

90 20, 40, 60, 80, 100 90

Suggest Documents