Vibro-acoustic assessment at the end-of-line test of automotive components, like engines, transmissions steering components and electric motors, has become a fixed part of today’s quality assurance concepts. The driving force is the requirement of an objective test against Noise/Vibration/Harshness (NVH) related issues on 100% of the manufactured products.
The measurement chain using vibration and noise assessment in production is described in this article. Attention is paid to how the vibro-acoustic test is managed for the production line, and what type of functions can be extracted from the data to determine product quality.
This article has the following sections: 1. Objectives 2. Test Setup, Regime, and Sensors 2.1 Test Setup 2.2 Test Regime 2.3 Sensor Concept 3. Functions for Noise and Vibration Assessment 3.1 Overall Levels 3.2 Octaves 3.3 Frequency Spectrum and Envelope 3.4 Psychoacoustic Parameters 4. Additional Functions for Fault Identification 4.1 Signal Analysis of Rotating Machinery 4.2 Angle Synchronous Digital Resampling 4.3 Interpolation on the Amplitude Axis 4.4 Order Spectral Analysis 4.5 Analysis in Angle Domain 4.6 Extraction of Metrics 4. Automatic Definition of Limits 5. Conclusions
The vibro-acoustic quality control of components today is usually introduced with the aim to sort out specimens which do not meet the quality requirements regarding their noise and vibration behavior.
The definition of quality requirements is often based on the subjective noise or vibration perception of test subjects. Based on the requirements of the vehicle manufacturer's specifications and quality standards, acceptable values for measurable vibration or noise components are specified and tested as shown in Figure 1.
Figure 1: End of line test for an automatic transmission uses noise criterion to assess quality.
On the supplier side, the introduction of the vibro-acoustic testing means additional expenses. However, using vibro-acoustic screening reduces the expense of reworking and increases the overall quality of the production process.
2. Test Setup, Regime, and Sensors
Meeting requirements for vibration and sound quality are the responsibility of the production line. Whether the requirements were defined by the end customer or the factory, the production test facility must be able to quickly and easily measure the production samples with a high degree of accuracy. Some considerations for a successful production testing are discussed next.
2.1 Test setup
Usually a functional test rig is existing, which must be extended to vibro-acoustic testing.
For vibration and noise measurements, an existing test rig is rarely the optimum for measurement purposes. The rig must be isolated from other noise and vibration that normally occurs in a factory setting in order to assess the actual production specimen and avoid false positives. Even compared to the mounting position in the end application (i.e., a vehicle), there are differences that affect the specified vibration and noise characteristics.
In general, a noise and vibration test bench must meet the demands of gauge capability. To ensure a repeatable screening process, vibration and noise isolation can be guaranteed by performing detailed Measurement Systems Analysis (MSA). Testing of components with low noise emissions are possible today at the production line. For example, vibration acceleration levels of about 70 dB(m/s^2 ) are measured reliably at steering components for a given noise level of about 50 dB(m/s^2 ).
Sometimes economic reasons do not allow installation of a completely optimized vibration or acoustic test benches at the production line. The limits of quality measures that have been specified under laboratory conditions or at the vehicle have to be adapted to the changed environment conditions in the factory test bench.
2.2 Test regime
For the evaluation of noise and vibration phenomena the testing regime should simulate the end application of product. Some examples include:
Operational Conditions - Product should be tested under the same loads and speeds as expected during actual operation.
Boundary Conditions - The boundary conditions should match the expected real world usage conditions. For example, if the product is constrain in three locations in real life, the same should be done in the production test.
Deviations from this principle are possible if all factors are known that significantly affect the vibration and noise characteristics.
2.3 Sensor concept
At end-of-line test bench vibration (via accelerometer) and noise (via microphone) can both be measured production safely as shown in Figure 2. Often the design effort on the test bed for vibration measurement is less than for a reliable measurement of airborne noise.
Figure 2: Sensors used in end of line testing for measuring vibration (accelerometer, left) and noise (microphone, right).
Since the emitted noise is caused by the vibrations of the specimen, a noise measurement often can be replaced by a vibration measurement.
3. Functions for Noise and Vibration Assessment
In this section, different functions will be used on the same data to ascertain how well functions can highlight faults. An example data set of 150 production samples containing a few faults will be used.
Both measures assess the overall level of the signal. Due to the current quality of vehicle components, both parameters are useful only to a limited extent as shown in Figure 3.
Figure 3: Distribution of 150 overall level values (blue) compared to limit line (red). The distribution on the right shows a few samples outside of the production limits (dashed blue line) but no faults are found (red line).
Despite there being known faulty production samples, the faults have little effect on the overall level, and limits are difficult to define.
Using the same data, third-octave spectra allow a frequency selective evaluation, while providing easy limit definition based on the vehicle requirement as shown in Figure 4.
Figure 4: Third Octave spectra - Dark blue is baseline production, red are octave measurements of product containing a fault.
In the third octave spectrum the effects of the faults can be seen. The frequency and amplitude characteristics of the abnormal noise can be described. However, determining the exact root cause is not possible because each octave is only a range of frequencies, but the actual underlying specific frequencies are not known.
3.3 Frequency Spectrum and Envelope
Frequency and envelope spectra give a more detailed root cause analysis. An example of frequency spectra from the same data is shown in Figure 5,
Figure 5: Frequency spectrum sideband lines of gear mesh orders (in the 3000 Hertz range) in the event of an error (red) compared to production (blue) baseline allow assignment to a faulty component.
With some prior knowledge of the operating frequencies of the production samples, the frequency spectra allow the assignment to the appropriate component (a gear pair) and a description of the error (improperly produced gear teeth).
See the knowledge article "Interpreting colormaps" for more information on sidebands and their causes in geared systems.
Envelope analysis of the same data even more clearly shows the fault as seen in Figure 6.
Figure 6: Envelope spectrum shows a measurement containing a fault (red line) very clearly versus measurements without fault (blue lines).
If an evaluation based on signal amplitudes and amplitude ratios is not possible, psycho-acoustic parameters can be used.
3.4 Psychoacoustic Parameters
It should be noted that these characteristics require a very precise definition of test conditions. Psychoacoustic parameters are defined only for airborne sound. The modulation metric roughness for this data is plotted in Figure 7 below.
Figure 7: Distribution of the psychoacoustic parameter roughness from150 different samples.
These pictures above show an example of how an error manifests itself at a test specimen in the acceleration overall level, the third-octave, frequency spectrum, envelope, as well as in psycho-acoustic parameters of roughness.
4. Additional Functions for Fault Identification
Causes of noise are always traceable to vibrations. These are caused by transient excitations or periodically varying forces that occur during normal operation and lead to the typical operating noise. Deviations from the normal noise only occur if atypical oscillations occur caused by changes in the dynamics. This applies to all phenomena, even if they are detectable only, for example, by psycho-acoustic parameters.
If the assignment of an abnormal noise to a certain part of the device under test is possible, the cause is identifiable.
Depending on their cause, the oscillations have typical signal characteristics. These allow a distinction from other vibration components, assuming a suitable method for the isolation provided. Consider the gear pair shown in Figure 8.
Figure 8: The kinematics of meshing determines the resulting oscillations. When the mesh contact is misaligned, the single mesh tone (top right) becomes several tones, including sidebands (bottom left).
The choice of an appropriate signal processing method for fault identification is directly related to the characteristics of the signal:
In contrast to the pure noise assessment methods, algorithms and features usable for error identification provide a high selectivity in the frequency, order, or angle-time domain. This is necessary for separating signal components of many different parts of the specimen.
A useful test system allows the flexible configuration of different signal processing methods in parallel to each other.
4.1 Signal Analysis of Rotating Machinery
As explained the previous section, the selection of suitable signal processing methods depends from the characteristic of the noise and vibration signals emitted from the component under test. A main characteristic of signals emitted from rotating machinery is its synchronicity to the rotation angle. This synchronicity can be used to provide stable input signals for the assessment metrics. This section explains the processing of angle equidistant signal processing methods as basic analysis methods for rotating machinery.
The graphic shows a typical application of digital resampling within an order analyzer:
Figure 9: Flow chart of digital resampling into the angle domain.
4.2 Angle Synchronous Digital Resampling
Goal of this processing step is the determination of the time instances that should be used as new sampling points for the angle equidistant sampled signal.
The input signal is a tachometer signal, i.e. a series of encoder pulses (circle with k values below) acquired from an appropriate tachometer sensor. The usage of a model with a constant acceleration between two synchronization pulses bases on the constantly accelerated rotary motion. The figure illustrates the determination of the sampling graphically:
Figure 10: An encoder (left) with a specific number of pulses produces an increasing frequency, i.e., rotational speed, versus time (right side). The higher the number of k values over one revolution of the encoder, the more precisely the speed of the rotating system is known.
The algorithm works correctly for linear speed changes between two known angles or between two synchronization pulses. With rapid changes in speed high standards of accuracy and resolution of the speed measurement are necessary. Therefore, a reasonable number of high precision angle measurements are needed and a high quality tachometer acquisition hardware as well. Angle-related errors have a greater impact in this regard to the higher order harmonics at high order resolutions. To provide reliable angle information at high revolution speed, typical signal sampling rates of 100 to 500 kHz are not enough. The tachometer signal should be sampled at 20 MHz or higher.
4.3 Interpolation on the Amplitude Axis
If the time instant of a certain angle is determined, the amplitude value of the vibration signal at this time instant must be calculated. This is possible by using a FIR-low pass filter with two independent time-axes, one time-equidistant for the input and one angle-equidistant for the output as shown in Figure 11.
Figure 11: Interpolation of angle synchronous signals
For optimizing calculation expense, a FIR-filter with a variable sample-count is used. The impulse response of the lowpass will be stretched or pushed together in proportion to the actual rpm-value.
4.4 Order Spectral Analysis
With the data in the angle domain, an order spectral analysis is often performed. This transfers the signal from the angle domain to the order domain. The result could be an averaged order spectrum or an order-related signal versus time or revolution speed as shown in Figure 12.
Figure 12: Averaged order spectrum - Amplitude of vibration of an accelerometer is plotted against order number rather than frequency.
The averaged order spectrum is calculated over a defined period or test step. It represents the average level of each individual order. The averaged order spectrum is well-suited for issues that occur over the whole test step independent from other parameters, like changes in speed.
If changes over time or speed are of interest, a level track should be used as shown in Figure 13.
Figure 13: Order amplitude level tracked versus rotational speed from 1200 rpm to about 2300 rpm.
This level track also can take harmonics and side band lines into account.
4.5 Analysis in Angle Domain
Based on the angle synchronous resampled signal, other signal processing methods are used to analyze the sensor signal. The goal of these analysis functions is the separation of signal components coming from the device under test against other signal components originated by environment or test bench itself.
For analysis of transient signals with a certain angular period the angle synchronous averaged envelope is a suitable analysis function as shown in Figure 14.
Figure 14: Vibration data is plotted against one revolution of rotating system, from 0 to 360 degrees. The fault data (red) stands out from the other production samples (blue).
Here the production sample with a fault stands out clearly.
4.6 Extraction of Metrics
Finding the right analysis function to extract signals originated from a certain component is only the first step. To identify a certain issue or a certain faulty component, metrics have to be defined. The metrics make use of the high selectivity of the analysis functions to assess only signal components that are generated by a certain component or caused by a certain issue. Typical metrics are:
Single order or frequency levels
Harmonic or sideband levels of orders, frequencies, qfrencies or 1/orders.
Relations between spectral lines, like for example sideband energy measures
The position of a transient signal component in time or angle domain representations
The position of a spectral line in the angle synchronous order cepstrum
The following example shows, how a sideband energy measure can help to detect a gear geometry issue at the timing drive of a commercial engine.
Figure 15 below shows a part of the averaged order spectrum around the 3rd harmonic of the timing drive’s gear mesh order.
Figure 15: Averaged order spectrum
The gear mesh harmonic, the 96th order and the sideband lines caused by the meshing gears are clearly visible. They also can be assigned to the single gear wheels.
The blue lines show the production reference at the end-of-line test bench. The red line show engines with gear noise issue caused by the gear geometry.
There is no line in the order spectrum where the red lines are outside of the scatter of the blue lines. Therefore, a direct metric based on an order line cannot be defined.
But the red lines have a different behavior regarding the distribution of order levels then the blue lines: All measurements of faulty engines have a low level of the gear mesh order and relatively high levels of the sideband lines. This is caused by the poor geometry of the tooth gears. While the gear mesh order basically shows the overall behavior of the tooth gear, the sideband lines are caused by modulation and show the deviation of single teeth from the averaged tooth geometry.
Based on this information a new metric can be defined. In this case a so-called sideband energy measure is used, where the relation of the gear mesh line level and the averaged level of the sideband lines is calculated. The following picture shows the statistics of the sideband energy measure for the engines showed in the previous picture.
Figure 16: Sideband energy clearly finds the production samples that contain faults.
It is clearly visible that the results from the engines with gear issue are outside of the production scatter, so that a reliable limit can be defined.
5. Automatic Definition of Limits
Developed for non-destructive material testing, statistical classification methods also are used at the end- of-line test. There are several reasons for using automatic methods for limit definition:
Reduction of manual configuration work for setting and maintaining limits and limit curves.
Setting up automatic limits to be able to detect new issues at their first occurrence. This is important to achieve a “zero fault production”
Automatic methods can do this work continuously, while the quality engineer can use hit expertise to improve the classification results directly.
Using statistical method for automatic limit adaptation has some advantages comparing to pure KI-based methods:
No training, but adaptation
No labeled data required
Easy to handle, easy to understand
Quality metrics are calculated from a measurement signal and evaluated using limit curves. The algorithm calculates the limit curves from series of measurements by statistical evaluation of the quality characteristics.
The statistical evaluation depends on the distribution density functions of the respective characteristic. For the simple case of one a Gaussian-distributed characteristic it is sufficient to use mean and standard deviation. With increasing number of measurements used for the calculation of statistical moments, their fidelity increases. It may be useful to already have a few test objects, e.g. for the pilot series to support an initial assessment.
Once the limit curves are determined, the statistical algorithm can adjust the limit curves accordingly to adapt the process development within predefined tolerance ranges. This automatic “drift adaptation” not only supports the quality engineer. It provides rather the chance that the current production is checked with narrower limits as it can be realistically feasible with manual setup.
Advantage of the operator is that fine errors can be reliably detected, which are otherwise within the tolerance band. Precise automatic adaptation helps in this case to avoid pseudo errors.
Of course, statistical methods work with scalar (e.g. correlation, level) as well vectoral quality metrics (spectrum, time signal). Tools for visualizing the adaptation history are available.
There are different ways to use this approach. So it can make sense to only allow adaptation when special events are fulfilled, for example when a quality characteristic reaches a predetermined warning limit. Another variant is the support of non-Gaussian features.
Ideally, the vibration and noise testing of components meet the requirements of the customers in terms of noise level. At the same time a supplier is able to optimize production processes and increase the overall quality.
Based on these two fundamental requirements, specified test systems have been established and are in use in several production environments. The required technology is available, modular and adaptable to different application scenarios, thus, without sacrificing usability. In the specification of the test benches, vibration and noise testing will gain an increasingly important role. In the future, a wider distribution of the methodology will also play a role by means of sophisticated automatic functions in the definition of new features for fault identification.