【正文】 y, gear changes and airflow changes further modifying system behavior to keep performance at a maximum and exhaust pollution at a minimum. Outside factors such as road quality, road gradient, vehicle weight, active accessories, etc. provide physical feedback to the system further altering behavior. In our system, we capture these physical events and dependencies through the power train signals. For example, Fig. 1(a) demonstrates a simplification of the relationship between the throttle position (TP) and revolution per minute (RPM) signals. TP makes a sudden rise and fall while RPM mimics this behavior but more smoothly. This simplification is not pletely accurate but demonstrates the key point that important physical relationships can be seen through the vehicle signals. Fig. 1(b) shows a more typical set of relationships between four different signals. Each circle is a signal and each edge indicates a feature that the tail signal influences in the head signal. These relationships are often plex, include five to ten different important signals, and have many cyclic dependencies between signals. We note several important issues related to using signals to diagnose a vehicle. First, we must differentiate between a bad signal and bad vehicle behavior reflected in the signal. A bad signal is generally caused by a bad PCM or a bad sensor. Bad vehicle behavior can be caused by any of a number of factors, physical or electronic. Our system detects signal features that indicate bad vehicle behavior, whether it is caused by bad electronic parts or physical faults. Second, we note that not all of the physical dependencies present in the actual vehicle can be modeled with corresponding signals. For instance, there is no signal to indicate road bumpiness, a physical factor that can effect vehicle and, therefore, signal behavior. To handle these unknown conditions we train with vehicle data in several conditions while avoiding extreme driving conditions (., offroad racing). Finally, the same signals are not available from all vehicles. When considering behavior that depends on signal relationships, this can lead to an inability diagnose certain faults that depend on information present in the missing signal.In this paper, we focus on developing techniques of deposing multiple signals, diagnostic feature extraction, and intelligent diagnosis. The paper is organized as follows. In Section II we briefly introduce the diagnostic system. In Section III, an automatic segmentation algorithm based on wavelet multiresolution analysis is introduced. In Section IV we discuss how we can process and bine feature and segment information to form feature vectors suitable for input into a machine learning system. Section V describes how a fuzzybased machine learning system can be used to learn good and bad signal behavior. Section VI describes the implemented diagnostic system and the encouraging experimental results we’ve obtained. Finally, Section VII discusses the impact of our work thus far and our future goals for this research.II. SYSTEM OVERVIEWThe system we have developed is a multilayered diagnostic system (see Fig. 2). Here, we present a brief overview of each layer and its goals. Layers 2, 3, and 4 are discussed in more detail in later sections. Layers 1 and 5 are discussed only briefly. The first layer translates the data into a format suitable for processing. This layer is relatively simple and is not discussed further.The second layer automatically partitions the signal into segments using either wavelet features from that signal or the segments of another signal. shows a TP signal that has been segmented using this module. These segments have three purposes. First, they divide the signal into regions that relate to some physical vehicle state, ., acceleration or idle. If we know the general physical state of the vehicle we can eliminate many possible faults and behaviors that we know cannot occur in the given physical state. Second, segmentation leads to a natural clustering of the signal data. Signal behavior within a given segment is generally very similar to behavior in other segments of the same state. This leads to more consistent training and test data. Third, using the segments we can isolate fault location within a signal. This can lead to easier fault identification. Finally, the segmentation strikes a nice balance between analyzing the original signal as a whole, which would result in enormous amounts of superfluous data, and analyzing the signal in one piece, which would result in a very plex feature vector. Segmentation allows us to examine important details of the signals without overloading the system with data.External Vehicle data1. Data Translation layer↓︱ ︱ ︱ TP signal RPM signal SPARKAVD signal↓ ↓ ↓2. segmentation layer ︱ ︱ ∣ TP segmentation RPM segmentation SPARKAVD segmentation3. Feature vector construction layer↓ ↓ ↓∣ ︱ ︱ ∣ TP Feature RPM Feature SPARKAVD Feature4. “super”Feature vector construction layer↓ ↓ ↓. Diagram of the proposed diagnostic system. Each layer handles different signals individually thus enabling the diagnosis of each signal to be customized. Time[ samples55ms]. Example of segmentation on the TP signal. In ADSAS, the segment lines are colorcoded to indicate the beginning of acceleration (rising), cruise (steady high), deceleration (falling), and idle (steady low) segments. The segmentation was done automatically by our system, ADSAS (see experiments).. Example of a few of the features extracted from the TP signal shown in Fig. 3. Start and end indicate the beginning and ending sample of the segment. States are 03 indica