Real-Time Digital Signal Processing of Ionization Current for Engine Diagnostic and Control

Combustion quality diagnostic techniques utilizing flame ionization measurement, with the spark plug as a sensor, have been in production for some time. This acquired “Ionsense” signal represents the changes in the electrical conductivity of the flame during each combustion event. The present analog versions of this sensor are used to detect knock and engine misfire, and can be used for cam phasing. However, current methodology has fallen short of unlocking the wealth of combustion thermodynamics information encrypted in the ion sense signal. Digital Signal Processing incorporating Artificial Neural Networks (ANN) is well suited for handling the statistical fluctuations of combustion. However to obtain acceptable accuracy, traditional ANN implementations can require processing resources beyond the capability of current engine controllers. Using Air/Fuel Ratio and Location of Peak Pressure as examples, this paper explores the practicality of performing real-time digital processing of the Ionsense signal to extract additional combustion information. An assessment of required processor resources is made and alternative preprocessing employing a pattern recognition wavelet filter is proposed. As a result the post-processed signal seems to be immune to some engine combustion fluctuations not included in the ANN training. The concepts discussed were successfully demonstrated throughout the normal operating range, in real-time, on a 6-cylinder engine. Examples of performance data are included. INTRODUCTION Theoretical foundations linking the Ionsense signal with engine thermodynamics and combustion kinetics were laid by a group of scientists at Lund Institute of Technology, Sweden [1, 2]. The attempt, however, fell short of producing a robust theoretical model since the modeling of molecule formation and destruction kinetics in combustion, even if simplified, is extremely complex (see e.g. [3]). Consequently, the numerous attempts to demonstrate Ionsense “advanced functionality”, i.e. the ability to provide information on the location of peak pressure, air to fuel ratio, percentage of mass fraction burned, etc., proved to be successful only in a narrow range of automotive engine operating conditions [1 – 7] (leading papers are quoted here only). As was pointed out in numerous publications, fuel type, fuel additives [8], and fluctuation in early flame development [9] made the interpretation of the Ionsense signal extremely difficult. Remembering that air humidity, spark plug performance, engine aging, etc. may also affect the combustion process; a straightforward interpretation of the Ionsense signal appears to be almost impossible. Consequently, conventional analytical methods surely do not provide the robustness expected for mass production applications. In an attempt to solve this formidable problem, Halmstad University scientists together with Mecel, an independent subsidiary of Delphi Corporation, proposed the application of artificial neural networks (ANN) for Ionsense data interpretation (see e.g.: [10 12]). Indeed, the statistical fluctuations of combustion are well handled by the trained ANN-based Ionsense sensor which was demonstrated in experiments described by the above mentioned research group [11, 12]. Clearly, a very comprehensive ANN training covering a broad range of possible engine operating conditions would assure the correct data interpretation, at least, in the statistical sense, enough to enhance the performance of the engine control system. 2003-01-1119 Real-Time Digital Signal Processing of Ionization Current for Engine Diagnostic and Control Gerard W. Malaczynski and Michael E. Baker Delphi Corporation, Technical Center Brighton Copyright © 2003 SAE International The implementation of an ANN block into the Ionsense signal processing defines the sensor electronics. It becomes Digital Signal Processing (DSP)-based with all the issues associated with this technique. The option of designing and manufacturing a dedicated ANN chip seems to be unacceptable due to the cost and lack of algorithm flexibility. The only option seems to be an ANN algorithm, developed as a result of bench training, executed by an on-board microprocessor in real-time. Therefore, the next step in the introduction of Ionsense advanced features for mass production would be to prove its ability to operate in real-time with the support of currently available or next generation, automotiveapproved, microprocessors. In other words, the assessment of required computation power with a realistically defined signal-processing algorithm becomes a priority in the development activity leading to an advanced version of the Ionsense soft sensor. This paper presents an attempt in this direction. In addition, the troublesome problem of the ANN algorithm formulation that requires extensive, “never guaranteed to be fully complete”, training is addressed. Namely, the proposed approach delivers signal processing results that seem to be immune to some engine combustion fluctuations not included in the ANN training. Although the activity described here is still far from complete, the results presented continue the earlier efforts [11, 12] in implementing statistical methods for Ionsense signal processing. ALTERNATIVE DSP ALGORITHMS As previously stated, the ANN subsystem requires extensive training on pre-production data samples. Yet, once established, it may represent a simple computational algorithm easily executed either by a dedicated DSP chip or by the on-board microprocessor. The ANN algorithm complexity depends on the input vector size to the first, hidden neuron layer. This complexity directly affects its ability to execute in realtime – between adjacent combustion events. If the Location of Peak Pressure (LPP) is to be extracted from the Ionsense signal, the input vector size reflects the eventual system resolution. Specifically, if the required resolution is 1⁄2 degree of crank angle, and the window of interest covers, say, 60 degrees, the signal input vector size would be 120. In addition, as previously demonstrated by others [11, 12], the desired correlation between the ANN output and the actual location of peak pressure is acquired only if Manifold Absolute Pressure (MAP), engine speed (RPM), and spark advance (SA) are also fed to the ANN input. This increases the vector size to 123. This is far too large if multi-layer perceptron architecture is to be emulated in real-time. This realtime operation requires a data acquisition period (window) and, beginning with the closure of this window, processing completion prior to the next cylinder event. One of the options available [11] is the input data reduction achieved with the statistical method called Principal Component Analysis (PCA). Briefly, this method [13] generates a new set of input vector components from a linear combination of the original vector components. All the new components are orthogonal to each other so there is no redundant information. However, it is commonplace for the sum of the variances of the first few new components to almost match the total variance of the components of the original vector. Thus, the new vector can be substantially downsized without a significant loss in the information needed for further processing. In the language of DSP, it means that the algorithm is introduced, in the form of a matrix consisting of 120 input elements (the initial vector size of the Ionsense signal in our example) and, say, an output consisting of 10 elements. Such a size reduction was proven not to affect the outcome of the ANN processing when used for Ionsense data interpretation [11]. Input data reduction from 123 elements to 13 elements (reduced in size to a 10 element Ionsense vector plus MAP, RPM, and SA) makes even a fairly complex ANN emulation feasible in real-time at all engine speeds. The application of the PCA, however, does not come without penalty. Following our example, the PCA method translates to an additional algorithm represented by a matrix consisting of 120 rows (input size), and 10 columns (output size). The reduced output is a result of matrix multiplication performed on the 120-element Ionsense vector input. This matrix multiplication actually consumes more computer power than any ANN algorithm used in our experiments. As is shown below, this defines a limit for the maximum engine speed within which our experimental set-up (modeled in Mathworks’ Simulink) could operate in real-time, given available signal processing speed and resources. To illustrate DSP resource requirements for a typical ANN-based ion current soft sensor supported by a PCA data reduction matrix, an estimate was made for an algorithm emulating a 15-neuron hidden layer and 1-neuron output layer. It was assumed that the hidden layer is fed from a PCA matrix having 10 outputs plus three key engine operating parameter (MAP, rpm, Spark Advance) for a total of 13 elements. The results are presented in Table 1. In the process of calculating number of multiply-accumulate instructions it was found that more than 2/3 must be allocated to serve the PCA data reduction matrix! Motorola 683XX Motorola DSP 56800 Motorola PowerPC500 Clock Speed 32 MHz 150 MHz 50 MHz Execution Time 5.3 ms 50 μs 110 μs A/D Conversion 14 μs 6 μs 8 μs Table 1. An example of DSP Resource Requirements for a Typical Ion Current Soft Sensor (a PCA data reduction matrix providing 13 inputs to the ANN consisting of 15-neuron hidden layer and 1-neuron output layer). Resource requirements are calculated for a sampling rate of 40 kHz. The application of the Principal Component Analysis defines the risk, or better, the robustness of the Ionsense soft sensor. As long as the PCA is formulated with extensive pre-production collected data sets and the ANN is subsequently well trained, the system guarantees valued performance. That is due to the fact that once the algorithm is formulated, it can be executed in real-time with a known,