Identification of Noise Sources
This invention relates to an apparatus and a method for detecting and identifying sources of noise in a vehicle.
During use of a vehicle such as a car, the driver or passenger may become aware of various noises, such as buzzes, rattles or squeaks. Such noises may for example be caused by vibration of trim panels, vibration of window glass, or relative movement of components in a seat, and can be distracting. They may indicate a fault, for example in the manufacture or the assembly process, and are preferably prevented; this entails identifying the source of the noise. New cars (in particular at the beginning of a production run) are typically given a test drive immediately after they come off the assembly line, by a tester responsible for identifying any unacceptable noises so they can be rectified. A way of standardising and assisting in this identification process would be desirable.
According to the present invention there is provided an apparatus for detecting and identifying sources of noise in a vehicle, the apparatus comprising at least one microphone and at least one accelerometer for installation in a vehicle, means responsive to signals from the microphone to determine if a noise is above a detectability threshold, a signal analysis unit for determining a plurality of characterising features of signals detected by the accelerometer and/or the microphone, and a pattern recognition unit for identifying, from the characterising features, the nature of the sound source.
The apparatus may comprise more than one microphone,
the time differences between detection of a sound by the different microphones providing geometrical information about the location of the sound source. However it is not easy to obtain precise positional information in this way, particularly for low-frequency sounds.
The accelerometer is preferably attached to the metal bodywork of the vehicle, at a position where there is good transmission of vibrations. For example it may be attached to the pillar in front of the driver's door. An accelerometer attached to the body of the vehicle will readily detect any vibrations, but is insensitive to external sources of noise such as talking, and aircraft flying over.
The signal analysis unit may determine the frequency spectrum of the signals, and its temporal variation. In this case the signal analysis unit must be responsive to brief sound sources, and so must rely on brief samples, for example it may use wavelet packet analysis (as developed by Ronald Coiffman) . Alternatively or additionally the signal analysis unit may firstly categorise signals on the basis of their time variation (for example on the basis of kurtosis) into say two or three categories, and then further subdivide each category on the basis of other features. For example signals of long duration may be analysed using an autoregressive algorithm, while signals that are burstlike may be categorised by parameters such as ring-down count, rise-time, and duration.
Since the system is only intended to identify those - -sounds -Θ-r- noises that a -person -would notice-,- -there is no need to consider sound frequencies outside the range about 20 Hz to say 16 kHz, and in practice a somewhat narrower range of frequencies may prove adequate. Some
sounds will be transmitted well through the bodywork, and so will be most effectively detected by the accelerometer; other sounds will not be transmitted to the bodywork, and so will be most effectively detected by the microphone.
The pattern recognition unit may rely on an algorithm to identify the source of the noise, or alternatively may incorporate a neural network to perform this identification.
The invention also provides a method of identifying sources of noise.
The invention will now be further and more particularly described with reference to, and as shown in, the accompanying drawings in which:
Figure 1 shows a diagrammatic sectional view of a vehicle in which noise identification apparatus is installed;
Figure 2 shows graphically the variation of sound spectrum with time for a rattle in a car door;
Figure 3 shows graphically the variation of sound spectrum with time for a loose screw in the rear trim of a car;
Figure 4 shows graphically the variation of sound spectrum with time for a groan from a car seat; and
Figure .5- shows graphically the -variation of sound spectrum with time for a window rattle in a car.
Referring to figure 1, this shows a diagrammatic
sectional view of a car 10, with the engine and luggage compartments in elevation, with a noise identification apparatus. The apparatus consists of a microphone 14 clipped to the top of the driver's seat 15, and an accelerometer 16 clamped magnetically onto the metal pillar 17 in front of the driver's door. Both the microphone 14 and the accelerometer 16 are connected by leads (not shown) to a portable computer 18 on the rear passenger seat 20. The tester installs the apparatus as shown, and then drives the car 10 along a cobbled test track, listening for rattles and other noises. At the end of the test, the computer 18 provides an output that indicates if any noises were detected above a threshold of pressure corresponding to the threshold of audibility, and if so provides data that characterise the noise, and suggests probable causes. This assists the tester in identifying and locating sources of noise. The microphone 14 and the accelerometer 16, and the computer 18, are then removed.
In an alternative use, the car 10 (with the apparatus installed) is placed on a mechanically driven vibrating support, and is subjected to a range of different vibration frequencies. Again the computer 18 indicates if any noises were detected above a threshold, characterises the noises and suggests probable causes.
There are a range of different mechanisms that can lead to noises. These may be classified as follows:
Buzz
1 . ... Resonant .excitation: -trim- panels,- often the larger ones such as the headlining, may be excited into resonance, typically at a frequency less than 200 Hz.
2. Snubbing resonance: trim panels may be excited to resonate at higher frequencies and at such an amplitude that they contact the vehicle body. There is consequently a transition from a largely unnoticed boom, to a buzz at two or three times the frequency of the boom. The resulting frequency may for example be between 300 Hz and 800 Hz, at which the noise is more likely to be annoying.
Rattle
3. Foreign objects: a small object such as a screw or washer that is lost in the vehicle during production, service or repair may fall into a recess in which it dances around producing a high frequency tinkle, typically in the frequency range 4 kHz to 12 kHz.
4. Body parts: this category consists of components such as window glass, boot lid, and bonnet (rather than the trim, which refers to the softer components) ; these may generate rattles as a consequence of the locking or actuating mechanisms, and may also generate squeaks.
5. Mechanical backlash: power train gearbox rattle is an example.
6. Wiring loom: cables may vibrate against the body or trim panels.
Squeak
7. Inter-surface movement: squeaks arise through two components rubbing together, the frequency and loudness depending on the nature of the two surfaces, and the force holding them together, and no such movement will occur if the frictional force is sufficiently high.
If fixings relax with temperature or age, then squeaks may become evident under certain operating conditions . They may also arise from accident damage repair, fitting of alarms, and other events that disturb or augment the vehicle's body and trim. Such noises are desirably detected and identified by vehicle servicing staff.
It will be appreciated that identification of noise sources is primarily of concern only for those noises that are sufficiently loud to be noticed by the driver or passenger. Consequently the computer 18 may be arranged to analyze signals from the accelerometer 16 only if the signal from the microphone 14 exceeds a threshold.
The signals received from the accelerometer 16 are processed by wavelet packet analysis to determine the frequencies present, and their power. The variation of this frequency spectrum with time enables the noises to be characterised. For example characteristics such as the duration D, the range of frequencies F above a preset sound pressure or power level, and the maximum signal L, while sound is being emitted, may be used as characterizing parameters, as indicated in the following table which shows results obtained by analysis (as described in relation to figures 2 to 5) of noise from four different sources:
Table
Sound source D/ms F/kHz L/dB
door rattle 12 2-13 -44 screw buzz 4 3-20 -35 seat groan 22 1-12 -61 window rattle 5 2-8 -26
The maximum signal values, L, are given in dB relative to a sound pressure of 0.5 Pa (rather than to the more common reference value of 20 μPa, which is the limit of audibility) .
Referring now to figure 2, this shows graphically the time variation of the frequency spectrum, the signal being represented by a gray scale (from white at zero power to black at the highest power) , the axes of the graph representing frequency and time. The noise is from a door rattle of a small car being shaken on a vibrating support, the noise being analyzed over two successive time periods (shown respectively in figure 2a and 2b) . In each case that sound is observed (for example at the occasions marked by the two arrows R) the pattern is similar, forming a column: that is to say the duration of each event (about 12 ms) is similar for all the detected frequencies, the frequencies ranging from about 2 to about 13 kHz. (It will be appreciated that the shape, in such a frequency/time graph, depends upon the scales used for the axes . )
Referring now to figure 3, this represents graphically, in the same way as in figure 2, noise from a buzzing or rattling screw in the rear trim, the noise again being analyzed over two successive periods of time. In each case the pattern is similar, forming a spike: that is to say the duration is short (about 4 ms) , the duration being shorter for higher frequencies, and the frequencies cover a wide range between 3 and 20 kHz (for example at the occasions marked by the arrows B) .
Referring now to figure 4, this represents graphically in the same way as in figure 2, noise from a seat, which may be described as a groan, the noise again being analyzed over two successive periods of time. In
each case (for example the occasions marked by the arrows G) the pattern is similar, forming a wide column, that is to say the duration is quite long (about 22 ms) , and the frequencies cover a wide range of between 1 and 12 kHz for the signals of highest power.
Referring now to figure 5, this represents graphically in the same way as in figure 2, noise from a rattling window. The noise is again analyzed over two successive periods of time. In each case (for example the occasions marked by the arrows W) the pattern is similar, forming a spike, that is to say the duration is quite short (about 5 ms) , and the frequencies cover a narrow range between 2 and 8 kHz for the signals of highest power.
It is evident from the graphs that the noise produced by any one noise source is sufficiently consistent that it may be classified. It is also evident from the graphs and the Table that the characteristics of different noises enable them to be distinguished from each other, so that the computer can provide not only information characterising the noise, but also an indication of a likely source for the noise.
In a modification of the apparatus, two, three or four microphones are installed in the vehicle at different locations, the locations of the microphones either being preset (and programmed into the computer) or being input as data by the tester. By analyzing the times of receipt of sounds at the different microphones the computer 18 can indicate a likely location for the sound source. - ■
The signals may be analysed in real-time, or may be recorded and analysed subsequently. Furthermore the
signals may be analysed in a different way from that described above. For example the signals may first be divided into two categories: continuous events, and burst (short-term) events. This division may be carried out on the basis of the parameter referred to as kurtosis, which is a measure of the shape of the distribution of values in a signal. A Gaussian distribution has kurtosis of zero. 'Sub-Gaussian' distributions have large tails relative to a Gaussian (e.g. uniformly distributed noise) and have kurtosis values of less than zero. 'Super- Gaussian' distributions are sharper, with smaller tails relative to a Gaussian. These have kurtosis greater than zero.
In the context of the acoustic or accelerometer signals, kurtosis gives a rough measure of how close a signal is to a Gaussian distribution. Those signals that consist of bursts of energy separated by sections containing (relatively) little energy have a distribution which is very super-Gaussian. Continuous signals are closer to a Gaussian or sub-Gaussian distribution. Therefore we can use the kurtosis value to separate continuous from burst signals.
Kurtosis K is calculated from the normalised signals as follows:
4 4 t-v,
K = ∑i (x± - μ) /Nσ - 3, where x± is the i value, N
2 is the number of points in the signal, σ is the variance and μ is the signal mean.
Those signals which fall into the continuous category are then analysed using a 5th order auto- regressive algorithm to identify their frequency content. The results from this algorithm can then be processed,
for example (as discussed below) by a hierarchical clustering algorithm, to identify clusters which may be associated with particular sources of noise.
The auto-regressive algorithm of order R (R = 5 in this case) involves modelling a time series yn as a linear combination of R earlier values in the time series, with the addition of an error term xn. For a signal containing n points, the (n+l)th point can be estimated as:
Yn+l = Yτιβ-1 + Yn- _.&2 + • • • • + Yn-pap + xn+l
where P is the number of parameters used, and ai - a are the actual parameters. The error xn+i is the difference between the observed value and the estimated one. In applying this model to observed data, it is possible to adjust the parameters such that the mean squared error over the whole signal is minimised. Values for these parameters can be derived analytically by solving the 'Yule-Walker' equations. Methods exist for estimating the ideal number of parameters, but this may also be done empirically. An estimate of the signal ' s frequency spectrum can be derived from the parameters .
Those signals whose kurtosis value is sufficiently high for them to be classified as burst types are analysed differently, so as to determine acoustic emission parameters. In one analysis procedure, after identifying the start and end points of a burst, we calculate: ring-down count (the number of zero crossings above a certain absolute threshold) , burst-duration, rise-time
(the time between a burst commencing and zero-crossings exceeding the threshold) , and ring-down count / burst- duration (which gives us some frequency information) .
Again these parameters may then be processed to identify clusters which can be associated with particular noise-generating events or sources. Various algorithms are suitable which, being provided with a large number of different signals characterised by a number of different parameters, can categorise the signals into clusters or groups within which the signals have a degree of similarity. As a general rule the accuracy of any such classification increases as the quantity of data increases . The algorithm may be arranged to give a desired level of detail (i.e. a number of different categories) , the level of detail preferably being selected in accordance with the quantity of data available for processing. For example, at one extreme the algorithm might be able to identify a rattle as being produced by a particular, identified, screw (this would require a large quantity of data) , while at the other extreme the algorithm might merely identify a noise as being produced by plastic rubbing on plastic (which would require much less data) . One such algorithm is referred to as a hierarchical clustering algorithm. In practice, the level of detail would generally be preset, to ensure that noises are identified with a consistent degree of particularity, and the apparatus would be operated to ensure sufficient data is obtained to achieve this level of detail.
Preferably both methods of signal analysis are performed. Identification of .a noise source based on wavelet analysis can then be cross-checked against the identification based on auto-regressive parameters or
acoustic emission parameters for continuous or burst-type signals respectively (based on the kurtosis) .