Kourosh Danai

Tool Wear and Breakage Detection Using a Process Model

On-line sensing of tool wear and breakage In machining has been a long standing goal of the manufacturing community. Wear and breakage detection systems are typically based on force, acoustic emission, current or temperature measurement. They are Important for reliable unmaνed operation, and also for implementation of an adaptive control optimization system. This paper proposes a model-based approach to on-line tool wear and breakage detection under varying cutting conditions based on force measurement. The proposed model is used together with on-line parameter estimation to track flank wear during cutting. The proposed method is illustrated with a simulation study, the results of which confirm the feasibility of the model-based approach.

Flank Wear Estimation Under Varying Cutting Conditions

A model-based methodology, designed to operate under varying cutting conditions, for on-line estimation of flank-wear rate based on cutting force measurements is introduced. The key idea is to employ a model of the relationship between force and flank wear, together with on-line parameter estimation methods. This permits separation of the direct effect of changing cutting conditions on force from the indirect effect where changing cutting conditions affect the wear which, in turn, affects the force. Simulation results confirm the effectiveness of this strategy for turning with varying speed, feed, or depth of cut. Experiments, conducted for turning operations with a varying depth of cut, show good agreement between estimated wear values and the actual values of tool wear measured intermittently during the cut.

AUTONOMOUS SYSTEM FOR MULTISTAGE CYLINDRICAL-GRINDING

An optimization strategy is presented for cylindrical plunge grinding operations. The optimization strategy is designed to minimize cycle time while satisfying production constraints. Monotonicity analysis together with local linearization are used to simplify the non-linear optimization problem and determine the process variables for the optimal cycle. At the end of each cycle, the uncertain parameters of the process are estimated from sensory data so as to provide a more accurate estimation of the optimal process variables for the subsequent cycle. The optimization strategy is validated both in simulation and for actual grinding tests.

Unsupervised Neural Network for Tool Breakage Detection in Turning

An unsupervised neural network is introduced for on-line tool breakage detection in machining using multiple sensors. This neural network performs detection by classifying the measurements either as normal or abnormal. However, it performs classification by relying only on the normal category, so that it does not need to establish the abnormal category requiring samples of measurements taken at tool breakage. This Single Category-Based Classifier (SCEC) also adapts the prototype values on-line so as to continuously update the normal category, and employs the noise suppression techniques: contrast enhancement and ding, in order to cope with different levels of noise in measurements. The performance of the SCBC is evaluated in turning. Extensive tests were performed which produced six tool breakage cases. Four measurements which were clear indicators of tool breakage in these tests were used as inputs to the SCBC and to two other classifiers utilizing Kohonens Feature Mapping and Adaptive Resonance Theory (ART2). The results indicate that the SCBC was the only classifier that could detect all of the tool breakages.

An Adaptive Observer for On-Line Tool Wear Estimation in Turning - Part II: Results

The basic concept and design of an adaptive observer for tool wear estimation in turning, based on force measurement, has been presented in a previous paper (Part I). This paper shows that numerical problems in the estimation of the states of tool wear precludes the use of this method in multi-wear cases where both flank wear and crater wear are present. The method can be applied, however, when one type of wear (either flank wear or crater wear) dominates. The method is applied in turning experiments to a case where flank wear is dominant, and to a second case where crater wear dominates. For the first case the flank wear estimates show excellent agreement with actual wear measurements. For the second case the crater wear estimates are satisfactory, but not as good as in the first case.

Fault Detection of Helicopter Gearboxes Using the Multi-Valued Influence Matrix Method

Abstract : In this paper we investigate the effectiveness of a pattern classifying fault detection system that is designed to cope with the variability of fault signatures inherent in helicopter gearboxes. For detection, the measurements are monitored on-line and flagged upon the detection of abnormalities, so that they can be attributed to a faulty or normal case. As such, the detection system is composed of two components, a quantization matrix to flag the measurements, and a multi-valued influence matrix (MVIM) that represents the behavior of measurements during normal operation and at fault instances. Both the quantization matrix and influence matrix are tuned during a training session so as to minimize the error in detection. To demonstrate the effectiveness of this detection system, it was applied to vibration measurements collected from a helicopter gearbox during normal operation and at various fault instances. The results indicate that the MVIM method provides excellent results when the full range of faults effects on the measurements are included in the training set.

Epsilon-Lexicase Selection for Regression

Lexicase selection is a parent selection method that considers test cases separately, rather than in aggregate, when performing parent selection. It performs well in discrete error spaces but not on the continuous-valued problems that compose most system identification tasks. In this paper, we develop a new form of lexicase selection for symbolic regression, named ε-lexicase selection, that redefines the pass condition for individuals on each test case in a more effective way. We run a series of experiments on real-world and synthetic problems with several treatments of ε and quantify how ε affects parent selection and model performance. ε-lexicase selection is shown to be effective for regression, producing better fit models compared to other techniques such as tournament selection and age-fitness Pareto optimization. We demonstrate that ε can be adapted automatically for individual test cases based on the population performance distribution. Our experiments show that ε-lexicase selection with automatic ε produces the most accurate models across tested problems with negligible computational overhead. We show that behavioral diversity is exceptionally high in lexicase selection treatments, and that ε-lexicase selection makes use of more fitness cases when selecting parents than lexicase selection, which helps explain the performance improvement.

Parameter Estimation by Parameter Signature Isolation in the Time-Scale Domain

It is shown that output sensitivities of dynamic models can be better delineated in the time-scale domain. This enhanced delineation provides the capacity to isolate regions of the time-scale plane, coined as parameter signatures, wherein individual output sensitivities dominate the others. Due to this dominance, the prediction error can be attributed to the error of a single parameter at each parameter signature so as to enable estimation of each model parameter error separately. As a test of fidelity, the estimated parameter errors are evaluated in iterative parameter estimation in this paper. The proposed parameter signature isolation method (PARSIM) that uses the parameter error estimates for parameter estimation is shown to have an estimation precision comparable to that of the Gauss―Newton method. The transparency afforded by the parameter signatures, however, extends PARSIMs features beyond rudimentary parameter estimation. One such potential feature is noise suppression by discounting the parameter error estimates obtained in the finer-scale (higher-frequency) regions of the time-scale plane. Another is the capacity to assess the observability of each output through the quality of parameter signatures it provides.

Cycle-Time Reduction in Machining by Recursive Constraint Bounding

Modeling uncertainty in machining, caused by modeling inaccuracy, noise and process time-variability due to tool wear, hinders application of traditional optimization to minimize cost or production time. Process time-variability can be overcome by adaptive control optimization (ACO) to improve machine settings in reference to process feedback so as to satisfy constraints associated with part quality and machine capability. However, ACO systems rely on process models to define the optimal conditions, so they are still affected by modeling inaccuracy and noise. This paper presents the method of Recursive Constraint Bounding (RCB2) which is designed to cope with modeling uncertainty as well as process time-variability. RCB2 uses a model, similar to other ACO methods. However, it considers confidence levels and noise buffers to account for degrees of inaccuracy and randomness associated with each modeled constraint. RCB 2 assesses optimality by measuring the slack in individual constraints after each part is completed (cycle), and then redefines the constraints to yield more aggressive machine settings for the next cycle. The application of RCB 2 is demonstrated here in reducing cycle-time for internal cylindrical plunge grinding.

Continuous Optimal Infeed Control for Cylindrical Plunge Grinding, Part 1: Methodology

A new methodology is developed for optimal infeed control of cylindrical plunge grinding cycles. Unlike conventional cycles having a few sequential stages with discrete infeed rates, the new methodology allows for continuous variation of the infeed rate to further reduce the cycle time. Distinctive characteristics of optimal grinding cycles with variable infeed rates were investigated by applying dynamic programming to a simulation of the grinding cycle. The simulated optimal cycles were found to consist of distinct segments with predominant constraints. This provided the basis for an optimal control policy whereby the infeed rate is determined according to the active constraint at each segment of the cycle. Accordingly, the controller is designed to identify the state of the cycle at each sampling instant from on-line measurements of power and size, and to then compute the infeed rate according to the optimal policy associated with that state. The optimization policy is described in this paper, and the controller design and its implementation are presented in the following paper [1].