Model based diagnosis of the oxygen sensors

Targets such as complying with legislations on emissions regulated for the protection of the environment and human health, increasing fuel consumption performance, and being competitive in the market, have consistently encouraged the experts to contemplate on air-fuel ratio control for decades. Therefore, many advanced systems, components, and methodologies have been developed and used in the industry. Oxygen sensor is one of these systems. Oxygen sensors in the exhaust layout of the vehicles are used as feedback elements for precise control of lambda calculation in engine cylinders to be sure that the three-way catalyst works in optimum range and to optimize the power output and fuel consumption of the engine. Both OBDII and EOBD regulations oblige the monitoring of oxygen sensors and warning the driver in case of detection of a faulty sensor. The primary purpose of the thesis is to develop a methodology for precise and accurate monitoring and diagnosis of the oxygen sensor to meet the legislations and the performance targets while the required calibration effort is reduced regarding time, cost and human resources. The thesis consists of six different chapters. Chapter 1 explains the role of air-fuel ratio in emission control in detail. The aftertreatment layout is presented for a better understanding. Besides, emission gases and their damages are explained. Finally, different methodologies in the literature about the model-based diagnosis applied to various applications are analyzed. Chapter 2 introduces the oxygen sensors in detail. There are generally two types of oxygen sensors used in vehicles. The first type is a binary sensor and gives information whether the mixture is rich or lean. The type of the oxygen sensor studied in this thesis is the second type called as Universal Exhaust Gas Oxygen Sensor. Therefore, the chapter explains the working principle of Bosch LSU4.9 sensor and importance of the sensor in closed loop control of the air-fuel ratio. Besides, the factors causing oxygen sensor to fail, the defect types and the conclusion of having a malfunctioning sensor in the aftertreatment layout are described in the section. The benefits of modeling the dynamics systems are indicated in Chapter 3. Besides, the advantages of the system identification for both linear and nonlinear systems are detailed for some specific methodologies such as impulse response, state space and ARX models for linear system identification and NARX models for nonlinear system identification. System identification uses a set of experimental data to correlate predefined inputs and the system output. System identification is a helpful approach especially when there is not enough information about the dynamic characteristic of the system, or mathematical equations to represent the physic behind are not known. Chapter 4, firstly, describes the location of the UEGO sensor in the current exhaust layout and the characteristic of experimental data. Then, input selection for system identification methods described in the previous section is realized by using principal component analysis, and fundamentals of combustion and ideal gas law. Using linear system identification follows the input selection step. In this step, three different linear identification model is created by using different combinations of preselected inputs. Since no satisfactory result could be reached, the data is split small portions to analyze the existence of local linearity. Afterward, due to lack of improvement in the performance nonlinear system identification approach is decided to utilize. Nonlinear-autoregressive-exogenous model is designed with several different configurations of the number of hidden layers, the number of the neurons, delays, learning algorithms, etc. to obtain the best suitable model representing the UEGO sensor. In the end, a NARX model with two hidden layers and eight neurons in each hidden layers is decided to use in residual generation phase of sensor diagnosis. After a successful model has been achieved to a sufficient degree, the residual generation and residual evaluation steps are continued in Chapter 5. Residuals are produced by comparing the sensor output to model output in two WLTC data. The first data set is collected in a WLTC with a good sensor on the vehicle whereas a faulty sensor is located on the vehicle during the second WLTC. After that, these two residual sets are analyzed statistically to decide on if the sensor is defective or not. It is decided to use the decision-making approach by using two different threshold values based on the standard deviation and mean values calculated for each residual data set. In future work, some other residual evaluation methods can be tried to see the effectiveness of different approaches. In addition, it is recommended that validations be performed under various environmental conditions.