A new criterion to validate and improve the classification process of LAMDA algorithm applied to diesel engines

This work proposes a new criterion to validate and improve the classification efficiency of the Learning Algorithm Multivariable and Data Analysis (LAMDA) fuzzy algorithm, which is an algorithm that combines the concepts of neural networks architecture and fuzzy clustering. LAMDA is based on finding the Global Adequacy Degree (GAD) of one data (individual) to a class (functional state), considering the contributions of each descriptor or variable. LAMDA is capable of generating new classes after the training stage and it uses probability density functions (PDF) for the estimation of similarity analysis between classes in order to determine the grouping criterion. The LAMDA algorithm was used here to identify new functional states that were not included during the training stage. However, this algorithm induced significant uncertainties when the recognized classes, corresponding to engine operating modes, exhibited similar membership degree values (MDV). To solve this, a new criterion to validate functional states after recognition (LAMDA-FAR), based on the minimum and maximum distances among MDV was developed. Both LAMDA and LAMDA-FAR algorithms were used in supervised learning mode to classify a historical database obtained from an experimental mapping methodology of an automotive diesel engine operating under several steady state conditions. For each engine operating mode the engine speed (rpm), exhaust gas temperature (°C) and accelerator pedal position (%) were measured as the representative variables to carry out the classification. Both algorithms were trained with 70% of the historical database. The remaining 30% of the data, as well as new engine operating modes (not taken into account during the training stage), were used to validate classifier results. It was found that the LAMDA algorithm alone was unable to properly classify similar engine operating modes, while the LAMDA-FAR algorithm showed 100% efficiency for both known and unknown operating modes. This high efficiency and low computational cost tool can be used to improve engine control strategies based on experimental mapping methods, as well as to monitoring and controlling on-line vehicle performance.