基于波形分析的汽車電控系統故障診斷技術(文獻翻譯)-資料下載頁

　　

【正文】 cal inertia, behavior of RPM in one segment is usually very closely tied to the behavior in the immediately previous segment. Thus we create a super feature vector for RPM that contains RPM’s features for the current and previous segment along with the state of the TP signal (increasingacceleration, decreasingdeceleration, stablecruise/idle, etc). A machine learning system can use this information to differentiate between RPM signals that react normally to TP and those that do not.The fifth and final layer consists of a machine learning system. The machine learning system is trained to recognize faults from one type of signal at a time, which results in separate knowledge bases for each signal type. Currently, we use a fuzzy learning system, but this layer can be generalized to neural networks or other suitable machine learning systems. Each of the super vectors produced in Layer 4 is fed into the learning system for training. At this time, the training is supervised so we provide a target output that the system attempts to match. We note that, as with many plex diagnosis problems, we often don’t have many verified “bad” data samples so we mainly train with good samples.The following three sections discuss each of the Layers 24 in more detail.III. SIGNAL SEGMENTATIONThe signal segmentation algorithm we developed partitions a signal into time slices representing different vehicle states. The vehicle states we consider are idle, cruise, acceleration, and deceleration. The TP signal, in particular, is a good candidate for segmentation because its behavior closely mimics these four states of the vehicle. In the TP signal, signal rises occur during periods of acceleration, declines occur during periods of deceleration and the relatively flat TP signal indicates cruise or idle. From TP we can make a good estimate of the vehicle’s state at any given time and partition the signal into segments representing time periods when the vehicle state is consistent. Fig. 3 above (Section II) shows an example of a typical TP signal segmented using our segmentation algorithm.Using TP as the leading signal, we apply the segments obtained from automatically segmenting the TP signal to other signals from the same vehicle recording in order to label the corresponding vehicle states in those signals. In Fig. 5 we show the RPM and SPARKADV (Spark Advance) signals overlaid with the TP segments from Fig. 3, demonstrating how the TP segments map to other signals. The issue then bees how to best segment the TP signal. We propose an approach based on multiresolution analysis (MRA) below that uses wavelet coefficients to help find segment boundaries.A. Using Wavelets and MultiResolution Analysis in Signal SegmentationOur automatic segmentation algorithm is based on MRA and uses the discrete wavelet transform (DWT) [9], [18], [19] to isolate features in multiple scales. The DWT has been used recently for many other applications including image pression [23], pattern recognition [4], speech processing [13], signal detection [17], astronomy [3], and model estimation [5].Finding segment boundaries is a type of edge detection problem. In particular, the states acceleration and deceleration correspond to rising and falling edges in the TP signal, respectively. With the correct choice of mother wavelet function, the wavelet coefficient values associated with a signal can be used to identify these edges [18]. Furthermore, the wavelet coefficients tend to isolate signal features in (such as edges) by scale. This allows us to tune the segmentation to avoid certain edges occurring from random noise or very smooth changes. We implement the DWT using the Fast Wavelet Transformation (FWT) algorithm [18], [19]. The first stage of FWT starts with the original signal. By passing the signal through special low and high pass filters associated with the mother wavelet function, we obtain detail wavelet coefficients corresponding to the highfrequency details of the signal and approximation coefficients corresponding to the signal minus the details. These two sets of coefficients are then down sampled by 2 before the algorithm continues by repeating the process on the approximation coefficients as we show in Time [samples55ms]. Segments from the TP signal in Fig. 3 applied to the RPM and SPARKADV signals.We label the detail coefficients as CD where is the level of deposition (number of iterations of the filtering and downsampling steps). The approximation coefficients are similarly labeled as CA . As bees larger, the corresponding detail coefficients, CD indicate coarser (larger scale) details in the signal. Highfrequency details can be found in the fine detail levels (., 13), while lowfrequency details can be found in the courser detail levels (., 46). Specifically what the detail coefficients indicate is dependent on the mother wavelet function. In our case, we have chosen to use the mother wavelet DB1 (Daubauchies One) because its detail coefficients indicate sharp changes in a signal indicating transition states (acceleration or deceleration). Specifically, using DB1 sharp changes correspond to peaks and valleys in the detail coefficients obtained from the FWT.B. Segmentation AlgorithmThe segmentation has four major steps we describe below. Details of each of these steps follows.Step 1)Create approximate segment boundaries. This step uses the wavelet coefficients from a range, [1,k], of detail coefficient levels to place segment boundaries very close to the correct location in the signal. We use a recursive, multiscale algorithm to divide large segments into multiple smaller segments for further detailed analysis. Step 2) Combine same state segments: After Step 1, some adjacent segments may have the same state and this step joins the segments together. Step 3) Fine tune segment boundaries: This step looks at a small neighborhood, typically one or two samples, around the segm