Oliver Nelles

Fast neural networks for diesel engine control design

Abstract Advanced engine control systems require accurate dynamic models of the combustion process, which are substantially nonlinear. This contribution presents the application of fast neural net models for engine control design purposes. After briefly introducing a special local linear radial basis function network (LOLIMOT) the process of building adequate dynamic engine models is discussed in detail. These neuro-models are then integrated into an upper-level emission optimization tool which calculates a cost function for exhaust versus consumption/torque and determines optimal engine settings. A DSP-based process computer system allows a fast application of the optimization tool at the engine test stand.

Identification with Dynamic Neural Networks - Architectures, Comparisons, Applications

Abstract An overview of presently known neural networks for the identification of nonlinear dynamic systems is given. The suitable networks are classified where two different approaches, the external and internal dynamic networks, have to be distinguished. One alternative is to apply standard static neural networks and incorporate the dynamics by providing the network with information about previous inputs and outputs created by external linear filters. Another possibility is to include dynamic elements within the neural network structure and therefore making the network itself a nonlinear dynamic system. The principles, advantages and drawbacks of both approaches are pointed out. Two special neural network architectures, one with external and one with internal dynamics, are considered in more detail. Their capability to identify complex nonlinear dynamic real-world processes is studied. Experimental results for two multi-variable processes are shown, a combustion engine turbocharger and an industrial scale tubular heat exchanger.

Multi-site damage localization in anisotropic plate-like structures using an active guided wave structural health monitoring system

A new approach for structural health monitoring using guided waves in plate-like structures has been developed. In contrast to previous approaches, which mainly focused on isotropic or quasi-isotropic plates, the proposed algorithm does not assume any simplifications regarding anisotropic wave propagation. Thus, it can be used to improve the probability of detection. In this paper the mathematical background for damage localization in anisotropic plates will be introduced. This is an extension of the widely known ellipse method. The formalism is based on a distributed sensor network, where each piezoelectric sensor acts in turn as an actuator. The automatic extraction of the onset time of the first waveform in the differential signal in combination with a statistical post-processing via a two-dimensional probability density function and the application of the expectation-maximization algorithm allows a completely automatic localization procedure. Thus, multiple damages can be identified at the same time. The present study uses ultrasonic signals provided by the spectral element method. This simulation approach shows good agreement with experimental measurements. A local linear neural network is used to model the nonlinear dispersion curves. The benefit of using a neural network approach is to increase the angular resolution that results from the sparse sensor network. Furthermore, it can be used to shorten the computational time for the damage localization procedure.

Genetic programming for model selection of TSK-fuzzy systems

This paper compares a genetic programming (GP) approach with a greedy partition algorithm (LOLIMOT) for structure identification of local linear neuro-fuzzy models. The crisp linear conclusion part of a Takagi-Sugeno-Kang (TSK) fuzzy rule describes the underlying model in the local region specified in the premise. The objective of structure identification is to identify an optimal partition of the input space into Gaussian, axis-orthogonal fuzzy sets. The linear parameters in the rule consequent are then estimated by means of a local weighted least-squares algorithm. LOLIMOT is an incremental tree-construction algorithm that partitions the input space by axis-orthogonal splits. In each iteration it greedily adds the new model that minimizes the classification error. GP performs a global search for the optimal partition tree and is therefore able to backtrack in case of sub-optimal intermediate split decisions. We compare the performance of both methods for function approximation of a highly non-linear two-dimensional test function and an engine characteristic map.

Integrated control, diagnosis and reconfiguration of a heat exchanger

Heat exchangers play an important role in chemical and process industries. In order to improve reliability and control performance, intelligent concepts for control, supervision and reconfiguration are necessary. In the paper, an approach is presented which integrates model-based adaptive control and reconfiguration based on fault detection/diagnosis applied to a heat exchanger plant. The adaptive controller and the fault detection scheme are based on a fuzzy model of the process (Takagi-Sugeno type). The fault diagnosis is performed using a self-organizing fuzzy structure.

Local Linear Model Trees for On-Line Identification of Time-Variant Nonlinear Dynamic Systems

This paper discusses on-line identification of time-variant nonlinear dynamic systems. A neural network (LOLIMOT, [1]) based on local linear models weighted by basis functions and constructed by a tree algorithm is introduced. Training of this network can be divided into a structure and a parameter optimization part. Since the network is linear in its parameters a recursive least-squares algorithm can be applied for on-line identification. Other advantages of the proposed local approach are robustness and high training and generalisation speed. The simplest recursive version of the algorithm requires only slightly more computations than a recursive linear model identification. The locality of LOLIMOT enables on-line learning in one operating region without forgetting in the others. A drawback of this approach is that systems with large structural changes over time cannot be properly identified, since the model structure is fixed.

Orthonormal Basis Functions for Nonlinear System Identification with Local Linear Model Trees (LOLIMOT)

Abstract A new approach for identification of nonlinear dynamic systems is proposed. It is based on a combination of generalized orthonormal basis functions and local linear model trees (LOLIMOT). The main idea is to approximate an unknown function from data by the interpolation of many local linear models. The number of local linear models and their validity regions are determined by a tree construction algorithm that partitions the input space by axisorthogonal cuts. The parameters of the local linear models are estimated by minimizing a equation error criterion (ARX models). These local parametric models are replaced by linear parameterized output error models based on generalized orthonormal basis functions. Thus, stability of the nonlinear dynamic model can be proven and the low bias property of output error models can be exploited. Furthermore, a subset selection technique may be applied for choosing the complexity of each local linear model. Since different basis functions may be constructed for each local linear model, processes with operating condition dependent dynamic behavior can be modeled efficiently.

Predictive control based on local linear fuzzy models

This paper deals with predictive control based on fuzzy models. A novel algorithm (LOLIMOT) is proposed for the construction of Takagi-Sugeno fuzzy models. The rule consequents are optimized by a local orthogonal least-squares method that selects the significant regressors. The rule premises are optimized by a tree construction algorithm which partitions the input space in hyper-rectangles. A generalized predictive controller (GPC) and a dynamic matrix controller (DMC) are designed. Both controllers require the extraction of a linear model from the Takagi-Sugeno fuzzy model. For the GPC a new technique called local dynamic linearization is proposed that exploits the special structure of the local linear models. The DMC is based on the evaluation of a step response. The effectiveness of both the identification algorithm and the predictive controllers is shown by application to temperature control of an industrial-scale cross-flow heat exchanger.

Adaptive predictive control of a heat exchanger based on a fuzzy model

Abstract In this work, a fuzzy model-based predictive controller is applied to the temperature control of an industrial-scale cross-flow water/air heat exchanger. The process reveals time-variant behaviour due to unmeasurable disturbances. Thus, on-line adaptation of the process model is required. First, a Takagi-Sugeno fuzzy model is identified from measurement data. Guidelines for the generation of proper excitation signals are given. For on-line adaptation of the fuzzy model, a local recursive least-squares algorithm is proposed. Since the nonlinearity of the process retains its structure, it is sufficient to adapt only the linear parameters in the rule consequents.

Supervised Hierarchical Clustering in Fuzzy Model Identification

This paper presents a new, supervised, hierarchical clustering algorithm (SUHICLUST) for fuzzy model identification. The presented algorithm solves the problem of global model accuracy, together with the interpretability of local models as valid linearizations of the modeled nonlinear system. The algorithm combines the advantages of supervised, hierarchical algorithms, which are based on heuristic tree-construction algorithms, together with the advantages of fuzzy product space clustering. The high flexibility of the validity functions that is obtained by fuzzy clustering combined with supervised learning results in an efficient partitioning algorithm, which is independent of initialization and results in a parsimonious fuzzy model. Furthermore, the usability of SUHICLUST is very undemanding, because it delivers, in contrast with many other methods, reproducible results. In order to get reasonable results, the user only has to set either a threshold for the maximum number of local models or a value for the maximum allowed global model error as a termination criterion. For fine-tuning, the interpolation smoothness controls the degree of regularization. The performance is illustrated on both analytical examples and benchmark problems from the literature.