The influence of sensor data quality on data analysis
Article 2: Six hints for improving sensor data quality
If you have an error in the measurements from your sensors, you will not be able to obtain appropriate insights from analyzing the data. There are six ways that you can avoid these measurement errors:
Understanding the object to be monitored
You need to know how abnormal conditions develop. For example, a defect in the bearing of a rotating machine such as a pump will generate abnormal vibration and unusual noise. An imperfection in installing a speaker in audio equipment might cause an odd sound called chattering. For reciprocating machines such as presses, fluctuations in the reciprocating cycle can lead to quality variations in the manufactured products. Variations in the temporal change of material injection pressure in injection molding machines may cause the molding to fail.
It is therefore vital to understand why an abnormal condition occurs from a physical point of view. This will tell you what type of data is necessary for analysis. However, this is not an easy process. It is crucial to collect information from skilled operators in the field, who are likely to understand the mechanisms of anomaly development.
Selection of the sensor and the data to be analyzed
One common mistakes is selecting the wrong sensor. This will mean that you acquire the wrong data for analysis. This is often through lack of understanding of the object, or the mechanism by which anomalies occur. For example, a bearing failure in a rotating machine appears as abnormal vibration, so it is best to use an acceleration sensor to acquire vibration data to detect abnormalities. Acoustic measurement using a microphone would be appropriate for detecting the chattering noise caused by the faulty installation of a speaker in an acoustic device.
There are cases where sensor selection is not necessary. For instance, if the control signal of the machine can be output externally throughout the embedded controller such as PLC and industrial PC, you could acquire the data from them.
There may also be issues related to the specialty of the system integrators in charge of the system building. Appropriate sensors may not be selected if the system integrators are not familiar with the sensors. Table 1 shows a list of typical sensors used for condition monitoring. Some sensors need particular support from specialized manufacturers or system integrators, and some require high-performance measurement instruments. For example, in one case, an accelerometer was installed even though an electric current sensor would have been better. This was because of the vendor’s previous experience. A simple change of sensor was sufficient to significantly improve the quality of the analysis.
Table 1．Typical sensors used for condition monitoring
Note. SIers: system integrators.
Mounting position of the sensor
The mounting position of the sensor is also important. For example, the picture on the left side of Figure 1 shows the condition monitoring of bearings using a rolling machine. Three accelerometers are mounted on the X-, Y-, and Z-axis. This is an example of correct installation.
You would not expect a doctor to listen to a heartbeat via the foot. In the same way, you should always ask, “Why are you installing that sensor there?”, and, “How many sensors should be installed and in what direction?” You need clear answers to those questions to be confident that you are mounting the sensors in the correct positions.
Figure 1. Quality control of manufactured products and anomaly detection of manufacturing equipment (roller machines)
Installation method of the sensor
You now have the correct sensor, and in the right place. The next step is to make sure that the installation method is appropriate. The type of sensor determines the installation method. Figure 2 shows the different kinds of installation methods, using piezoelectric accelerometers as an example. The upper limit of the vibration frequency that can be measured varies with the installation method. For example, with a probe, the upper limit is 1 kHz. With a screw mount, it may be possible to measure up to 15 kHz.
Figure 2. Loss of vibration data due to incorrect installation of accelerometers
There may even be simple problems, such as one installation where the user had attached the accelerometers using double-sided tape, and the vibration was absorbed by the tape. Reinstalling the accelerometers was a quick and easy way to solve the problems with the data analysis,
Selection of data acquisition devices
Poor performance of the data acquisition device can also be a problem. Unfortunately, this is a blind spot, and it is difficult to identify. Even with a highly accurate sensor, the accuracy of the data may be degraded if the appropriate data acquisition device is not selected. There are three essential factors: sampling frequency, resolution, and synchronous measurement (Figure 3).
Figure 3. The selection of appropriate measurement equipment is essential
The sampling frequency is usually listed in catalogs as one of the selection criteria for data acquisition devices. In recent years, there have been few data acquisition devices with low sampling frequency, reducing this problem. However, the resolution is important. For example, it is industry standard to apply data acquisition devices with 24-bit resolution for measurements using accelerometers and microphones. However, equipment with 16-bit resolution is sometimes used and oscilloscopes have 8-bit resolution. This reduces the waveform reproducibility and leads small changes to be overlooked, which causes lower sensitivity of anomaly detection (Table 2 ).
Table 2. Causes of measurement errors and influence on data analysis results
This may sound minor, but can be critical. For example, I was involved in the root cause analysis of printing misalignment in a high-speed printing machine. The problem was caused by slight damage to the bearing, which resulted in a misprint. However, the sampling frequency and resolution of the data acquisition device were low, and it could not detect the abnormal vibration. A more sensitive data acquisition device enabled its detection.
Although extremely important, synchronous measurement is seldom considered. If the measurement data from various sensors are not synchronized correctly, this may affect the data analysis. For example, when detecting abnormalities in rotating or reciprocating machinery that has a cyclic nature, the timing of the rising edge of various signals and the cycle of signal generation are critical. It will be difficult to detect abnormalities if the data are not synchronized (Figure 4). It is also necessary to check whether the synchronization accuracy of the measurement device is appropriate for the application. In some cases, a microsecond level of synchronization accuracy is required. In other cases, a millisecond level is sufficient.
Figure 4. Importance of synchronous measurement
Selection of data to be stored in the data lake (feature extraction)
The final area to consider is selecting the data to be stored in the data lake. A data lake is necessary to store data for model development and updating. However, sensors can produce an enormous amount of data. For example, thermal imaging cameras generate more than 1 GB of data per minute, as shown in the table on the right side of Figure 5 . It is essential to reduce the data volume to manage pay-as-you-go service fees and network traffic load. Organisations also often want to reduce the number of expensive sensors. Strategic data selection (feature extraction) is therefore required. This means we have to ask ourselves what kind of data we need and how we could use edge computing to reduce the amount of data.
Figure 5. Strategic selection of relevant data, sensors, and features
The waveform shown in the upper left of Figure 5 is the raw data from the accelerometer, which is 1.8 MB of data per minute. However, if we stored just the average value once per second in a data lake, the data stream would only be a hundred bytes per minute. It is therefore vital to understand what type of data acquisition and time interval is sufficient. In some cases, raw data is unnecessary, and the average value every eight hours is adequate.
These six areas create unique challenges when data scientists analyze industrial data. They need to take advantage of data analysis algorithms and understand the objects being monitored to ensure that they have the right data to answer the most important business questions.
Registration is open! SAS is returning to Vegas for an AI and analytics experience like no other! Whether you're an executive, manager, end user or SAS partner, SAS Innovate is designed for everyone on your team. Register for just $495 by 12/31/2023.
If you are interested in speaking, there is still time to submit a session idea. More details are posted on the website.
Data Literacy is for all, even absolute beginners. Jump on board with this free e-learning and boost your career prospects.