Invented by Mahmoud Ismail, Saeid HABIBI, Samir Ziada, McMaster University

The market for the enhanced system and method of conducting Principal Component Analysis (PCA) on data signals is witnessing significant growth and is expected to continue expanding in the coming years. PCA is a widely used statistical technique that helps in identifying patterns and reducing the dimensionality of complex data sets. The enhanced system and method of conducting PCA on data signals offer improved accuracy, efficiency, and scalability, making it a valuable tool for various industries and applications. One of the key drivers of the market growth is the increasing adoption of data analytics across industries. With the exponential growth of data generated by businesses, there is a growing need for advanced techniques to analyze and extract meaningful insights from these vast data sets. PCA provides a powerful solution by transforming high-dimensional data into a lower-dimensional representation while preserving the most important information. The enhanced system and method of conducting PCA further enhance the accuracy and speed of this process, making it an attractive choice for businesses looking to gain a competitive edge through data analysis. Another factor contributing to the market growth is the rising demand for anomaly detection and fault diagnosis in various sectors such as manufacturing, healthcare, finance, and telecommunications. PCA is widely used for detecting outliers and anomalies in data signals, enabling businesses to identify potential issues or irregularities in their operations. The enhanced system and method of conducting PCA on data signals offer improved sensitivity and specificity in anomaly detection, enabling businesses to proactively address potential problems and optimize their processes. Furthermore, the increasing focus on machine learning and artificial intelligence (AI) is driving the demand for advanced data analysis techniques. PCA plays a crucial role in dimensionality reduction, a fundamental step in many machine learning algorithms. By reducing the number of variables, PCA helps in improving the efficiency and accuracy of machine learning models. The enhanced system and method of conducting PCA on data signals provide more accurate and reliable results, making it an essential tool for AI and machine learning applications. The market for the enhanced system and method of conducting PCA on data signals is also benefiting from advancements in technology, such as faster computing power and cloud-based solutions. These advancements enable businesses to process and analyze large data sets more efficiently, making PCA a viable option for real-time and big data applications. The enhanced system and method of conducting PCA leverage these technological advancements to deliver faster and more accurate results, further driving the market growth. In conclusion, the market for the enhanced system and method of conducting PCA on data signals is experiencing significant growth due to the increasing adoption of data analytics, the need for anomaly detection and fault diagnosis, the focus on machine learning and AI, and advancements in technology. As businesses continue to generate and collect vast amounts of data, the demand for efficient and accurate data analysis techniques like PCA will continue to rise. The enhanced system and method of conducting PCA on data signals offer improved capabilities and performance, making it a valuable solution for businesses across various industries.

The McMaster University invention works as follows

Systems and Methods relating to Fault Detection and Diagnosis.” The signals received from sensors first need to be filtered out of noise, and then they are analyzed with wavelet packet transform based PCA. The PCA results are automatically classified, allowing for a quick and easy determination of any issues in the finished product or machine under monitoring.

Background for The enhanced system and method of conducting PCA on data signals

Fault Detection and Diagnosis (FDD) is a good tool for End of Line Testing. EOL testers are used to test manufactured products for quality control. If the product passes, it will be processed and ready for shipment. If the product fails to pass the test, it is not shipped. The industry is not only interested in detecting faults, but also troubleshooting the root cause. Industry therefore favours the performance of fault diagnosis/isolation.

Fault detection/isolation” is essential to quickly identify the cause of a problem and fix it. This will reduce manufacturing costs. Electric motors, for example, are frequently used in industry. Since these components are susceptible to manufacturing problems, it is important that robust quality control measures be implemented. Fault Detection and Diagnosis systems are installed on EOL testers by manufacturers to perform fault detection and analysis.

Industrial FDD applications must meet strict requirements. These requirements range from time restrictions to robustness and environmental disturbances. They also include accuracy and ease of operation for operators. In real-time applications, and on production lines, the timing constraint is more pronounced. Manufacturing environments can also contain unwanted disturbances and noise, which affect FDD systems and their testing. This is especially true if sound and vibratory measurements are used to diagnose and detect manufacturing faults. If these disturbances are not removed, they can affect the test results. FDD testers can be operated by workers on the production line. FDD testers are designed to be easy to operate by manufacturers to avoid the need for highly trained and qualified technical personnel.

The concept of FDD can be applied to machine health monitoring. The FDD system in machine health monitoring is continuously run on the same machine, instead of once per product as it was for EOL testers. The FDD tool is the same regardless of the difference.

An example of a practical application of FDD technology is in the field automotive starters and generators. Due to high production rates, automotive starters or alternators are tested for only a few moments. Alternators and starters for automotives can be affected by both high and low frequency faults. FDD systems should be able detect both low and high frequency faults in noisy environments. Any FDD system must also display an easily readable result, which can be used by the operator without any FDD technical training.

The current FDD methods range from simple to complex. In industry, the most basic methods involve simple measurements such as sound or vibration. These methods measure the Root Mean Square values of sound and vibration over a period of time. More complex systems use more statistical measures, such as Peak to Peak levels, Crest Factors, Kurtosis and skewness, in the measured variables. Sawalhi N. et. al. (2007) show an example of kurtosis. Paajarvi P. et. al. show another example of a temporal signal in the patent publication US20130024164, “Method for detecting rolling bearing faults based on increasing statistical asymmetry”. This publication uses a linear filter to separate ball bearing impulses and noise.

All of the above methods use time signals and analyze each signal separately. Signals that measure sound and vibration usually acquire the same events as when testing a component. For example, failures of ball bearings. To maximize fault detection performance and isolation, it is preferable to analyze different signals simultaneously. Multivariate analyses are commonly performed in this area using Principal Components Analysis. PCA recognizes information sharing between signals by modeling the correlation structure. PCA is used for FDD in many references (both academic papers and patent publications) due to FDD?s efficiency and capabilities. As an example, U.S. Pat. No. Purdy M. A. 8,676,538 Adjusting weighting a parameter related to fault detection on the basis of a detected fault PCA is applied in a dynamic-weighting technique to perform fault detection. This reference discusses the application of PCA to semiconductors, and how PCA is used in a feedback system to improve fault detection reliability. In patent publication EP2950177A1, Dutta, P., et al. Asset condition monitoring (? PCA is used in a machine monitoring app that uses data from different sensors. PCA was used in this reference as a feature extraction method, as well as as a dimension reduction method, along with a preprocessor and classifier. The results of the parallel branches are then compared to the known faults.

The above noted methods analyze measurements in a temporal form. However, in many cases it is known that faults can be detected and isolated by the inspection of a signal’s frequency content. In Yang, H., et al., ?Vibration feature extraction techniques for fault diagnosis of rotating machinery: a literature survey?, (2003), Yang shows different types of temporal and frequency domain based FDD systems. Frequency domain systems start from simple spectrum methods, as shown in Ghorbanian, V., et al., ?A survey on time and frequency characteristics of induction motors with broken rotor bars in line-start and inverter-fed modes?, (2015). In this reference, it is shown that broken bar faults in motors cause different peaks in the spectrum of faulty motors. Peak frequencies depend on the slip factor (s), which represents the lag between the magnetic field’s speed and the rotor speed.

Similar to pure temporal methods spectral methods were used to analyze signals in only one domain. To get the best of both worlds, a method can be used that analyses the signal measured in both the spectral and temporal domains. Wavelets can be used to achieve this. The U.S. Patent. No. 6,727,725B2, Devaney, M. J., et al. (?Motor Bearing Damage Detection via Wavelet Analysis of the Starting Current Transient?) Wavelets, and specifically the Discrete Wavelet Transform, were used to detect faults in bearings during motor starting-up transient measurements.

Wavelets are a useful tool, but the problem of maximising fault information in multiple measurements is still present in methods using frequency/time domain, and this includes methods that use wavelets. Bakshi, B. R. ‘Multiscale PCA With Application to Multivariate Statistical Process Monitoring? (1998) combined DWT with PCA into a new method named Multi-Scale PCA. This method uses DWT to break down the signal into different levels of frequency bandwidth and then PCA is applied to each level in order to detect faults, as shown in Figure. 1. Bendjama H. et. al. (2010) show an application of MSPCA. MSPCA and contribution plots are used in this reference to isolate faults that have been detected by MSPCA. MSPCA uses PCA to detect faults and compares the measured signals against a baseline signal (from a control model). It is important to note that a baseline must exist in order for MSPCA be able detect faults. Figure 1 shows the typical steps of MSPCA using Reconstruction Based Contribution plots (RBC) for isolation. 2. These steps are described in detail in Haqshenas, S. R., ?Multiresolution-Multivariate Analysis of Vibration Signals; Application in Fault Diagnosis of Internal Combustion Engines?, (2013).

The above discussion shows a number advancements in the FDD field. These methods are not without their limitations. The resolution of fault frequencies is limited by DWT. It decomposes measured signals into different levels, but the frequency bandwidths are not equal. As shown in FIG. As shown in FIG. PCA detects faults at the highest level of the bandwidth. This can cause problems when faults are detected in different areas. High frequency faults are inseparable due to this. PCA can detect faults on different levels, which is another limitation of current FDD methods. Wavelets and PCA transforms cause the fault sensitivity to vary from one level of PCA to another. This is a bad behaviour, as it reduces the accuracy of the system in diagnosing faults at different frequencies.

The third limitation is that the fault diagnosis and determination by Contribution plots are not very accurate. The same as the second limitation, this lack of accuracy can be attributed to the fact that if there are two faults with similar severity in two measured signals then contribution plots will detect each fault differently, thus showing the severity of one fault over the other. This limits the accuracy of diagnosis.

A fourth limitation is the inefficiency of the existing methods.” Dutta P. et. al. (?Asset Condition Monitoring? PCA is performed up to N times. N is the total number of conditions. This is not an efficient design when performing a complex analysis using PCA and Wavelets.

A final limitation to current methods is the fact that they do not take into account normal meandering movements that occur over a long period of time in a manufacturing environment. These changes can affect measurements, but they are not always related to faults. Humidity is one example. Normal humidity fluctuations can affect sound measurements, and therefore any analysis that is based on these measurements.

From above, it is clear that there is a need for new systems and methods to mitigate or even overcome the limitations and deficiencies of the prior art.

The present invention relates to systems and methods for fault diagnosis and detection. The signals received from sensors first are filtered to remove any noise, and then are analyzed by wavelet packet transform based PCA. The PCA results are automatically classified, allowing for a quick and easy determination of any issues in the finished product or machine under monitoring.

In one aspect, this invention is a system for fault diagnosis and detection based on signals. It’s called Industrial Extended Multi-Scale Components Analyses (IEMSPCA). In one implementation, the invention comprises a filtration, a detection, extraction, and automatic classification block. The filtration blocks removes background noise, while the detection-and-filtration block utilizes wavelets, PCA and statistical indexes to detect and extract faults. The classification block classes faults detected by detection and extraction.

In one implementation, the system uses well-known tools such as wavelets, Principal Components Analysis, and a new statistical indicator to create a robust fault detection and diagnostic solution for industrial applications. A noise filter is used to increase the robustness of the system in noisy environments. The system also has a classifier that automatically provides results that are easily readable. The user interface is simplified and the need for highly-trained technical personnel to operate it has been eliminated. The classifier has an adaptive dynamic feature that can adapt to meandering changes within manufacturing environments.

The systems and methods described in the present invention have numerous advantages and benefits over the traditional and state-of-the-art Fault Detection and Diagnosis methods, such as Multi-Scale Principle Components Analysis. Using Wavelet Packet Transforms (WPT) rather than Discrete Wavelet Transforms (DWT) allows a finer resolution for high frequencies. The faults with high frequencies are better detected and isolated. A new statistical index makes it possible to perform fault detection and separation efficiently. This reduces the processing time for FDD. This statistical index creates a unique signature for each fault type. A unique fault signature can be used to identify the type of problem in a machine.

The new statistical index is not the same as what is used in current statistical methods. This index allows faults to be detected in all frequency bands equally. The current state-of the-art statistical methods are sensitive to faults that occur in certain frequency bands. This leads to uncertainty in fault detection. The new statistical processing method eliminates this uncertainty. The statistical index introduced has another advantage in that it detects faults evenly across all input signals. Other words, faults in different input signal are detected with the exact same sensitivity. This results in a more robust and accurate fault isolating.

In a first aspect of the invention, a system is provided for the analysis of signals from at lease one sensor. The system comprises:

Click here to view the patent on Google Patents.