User:Woodrow.Gonsalves/sandbox

As internal combustion (IC) engines remain vital to numerous industries, from automotive to aerospace, ensuring their smooth and reliable operation is critical. Over time, engines are subject to wear and tear, which can lead to malfunctions, reduced efficiency, or even complete failure. Traditionally, engine maintenance and fault detection relied on regular manual inspections by skilled technicians, which, while effective, often require considerable time and are prone to human error. The classic monitoring methods for detecting faults in automotive vehicles based on onboard diagnostics (OBD) are insufficient when diagnosing several mechanical failures. To address these limitations, audio signature-based condition monitoring has emerged as an innovative approach that leverages sound analysis to detect faults in IC engines automatically and more accurately.

A labelled Internal Combustion Engine

Need for Audio-based Fault Detection

IC engines produce unique sound patterns during operation, with each component contributing to the overall audio signature. When a fault occurs, it alters the sound frequency, tone, or amplitude, creating a specific audio “signature” associated with that defect. Using audio signatures for fault detection offers several advantages:

1. Non-invasive Monitoring: Audio sensors can capture sounds without physically interacting with the engine, making this method highly practical and cost-effective.
2. Early Detection: Changes in audio signatures often appear before major failures occur, allowing early intervention and prevention of costly repairs.
3. Reduced Downtime: Automated audio-based monitoring can occur in real-time, enabling prompt fault diagnosis and minimizing operational delays.

Given these benefits, audio signature-based condition monitoring can significantly enhance the reliability and efficiency of IC engine maintenance programs. This article delves in the application of using FFT and correlation for audio signature-based condition monitoring to transform raw audio data into actionable insights for engine maintenance and fault prevention.

Algorithm for Identifying IC Engine faults using Audio:

1. Fault Diagnosis Using Sound

This method aims to use audio signatures (specific sounds generated by engine parts) to detect faults in Internal Combustion (IC) engines. Traditional techniques for engine condition monitoring relied heavily on physical inspections and vibration monitoring, but these approaches are limited in their ability to isolate specific issues at an early stage. Recent advancements have shown that condition-based monitoring (CBM) methods, particularly those based on audio signatures, can be effective.

To effectively capture engine sounds for analysis, the researchers employed robust, the following system:

A piezoelectric microphone

1. Microphone: The PCB 130D20 piezoelectric microphone was chosen for its reliability in industrial settings, where high temperature and harsh conditions can affect other microphone types. Piezoelectric microphones are known for their durability and wide frequency range, making them ideal for capturing the broad spectrum of sounds in engine operations.
2. Data AcquiDT-9837 (a data acquisition board) was used to convert audio signals into digital data at a high sampling rate of 50 kHz (or 50,000 Hz). This rate satisfies the Nyquist criterion, which states that the sampling rate should be at least twice the maximum frequency in the signal. This ensures accurate frequency analysis up to 25 kHz (or 25,000 Hz), allowing the system to capture detailed information across the engine’s operational range.

2. Pre-processing the Sound (Filtering and Down sampling)

Before analysing the data, it must be pre-processed to reduce noise and focus on relevant frequency ranges. Pre-processing includes band-pass filtering and down sampling:

Microstrip Bandpass Filter

• Band-Pass Filtering: The initial audio signal often contains irrelevant low and high frequencies. A band-pass filter focuses on frequencies between 800 Hz and 12,500 Hz, where most engine sounds that indicate faults are found. Filtering reduces interference from environmental noises and highlights relevant sounds associated with engine defects, improving diagnostic accuracy
• Down sampling: To simplify data processing down sampling is performed, retaining every n^th sample (e.g., every 2nd or 4th sample). This reduces the data size, allowing for faster processing while preserving important signal characteristics.

3. FFT Analysis (Breaking Down the Sound into Frequencies)

The Fast Fourier Transform (FFT) is applied to convert the sound signal from the time domain (how sound varies over time) to the frequency domain (specific frequencies present in the sound).

FFT is widely used in audio analysis for fault detection as it allows the identification of distinct frequency components that characterize various engine issues. This step involves:

• Dividing the sound signal into windows of 2,000 samples each.
• Application of FFT to each window, and splitting the frequencies into 120 bins (small frequency ranges).
• The average values of these frequencies are calculated to create a frequency spectrum for analysis.

4. Sumpeak Algorithm (Initial Fault Detection)

The Sumpeak Algorithm is used as an initial test to determine whether the engine is healthy or faulty:

Hilbert Transform using a bandpass filter

• Hilbert Transform: This transform generates an envelope around the audio signal, which highlights variations in amplitude over time. Peaks in this envelope represent sudden changes in sound, often associated with mechanical issues.
• The Sumpeak Algorithm then calculates the sum of these peaks. If this sum exceeds a specific threshold, it suggests that the engine may have a fault.

This method provides a quick preliminary indication, allowing the system to filter out healthy engines before selecting to work on the potentially faulty ones.

5. Correlation Analysis (Comparing with Known Faulty Engines)

Once the Sumpeak Algorithm flags an engine as faulty, correlation analysis is used to determine the type of fault. The idea is to compare the sound of the faulty engine with “prototype” faulty engines that have known issues.

Correlation

• A correlation coefficient is calculated between the unknown engine sound and the sound profiles of various faulty engines.
• If the coefficient is high (above a specified threshold), it means that the unknown engine has a similar fault to one of the known engines.
• The system computes these coefficients for all types of faults (e.g., cam chain noise, cylinder head noise, etc.) and assigns the fault based on the closest match.

6. Prototype Engine Database

For accurate correlation analysis, a reference database of prototype faulty engines is essential. This database is built by:

• Recording sounds from engines with known issues, like cam chain noise or primary gear damage.
• Creating a correlation coefficient matrix, where values indicate the similarity between different faults, enabling quick comparisons with new engine data. This reference matrix forms the basis of the diagnostic system’s decision-making process.

7. Unknown Engine Classification

If the unknown engine’s sound is classified as faulty, the system then compares it with the database of prototype engines:

• The sound is segmented in the frequency domain using FFT.
• A correlation coefficient is calculated between the unknown sound and each prototype fault type.
• Based on the closest match, the system categorizes the engine fault, improving diagnostic accuracy in identifying specific issues like cylinder head noise or magneto rotor noise.

8. Voting System for Multiple Faults

To accommodate the possibility of multiple faults within the same engine, a voting system is applied, if two faults are detected with similar correlation coefficient, both faults are flagged.

This voting system prevents premature diagnosis, as it combines results from multiple analyses to increase confidence in fault identification. By setting a threshold for similarity, the system decides whether one or multiple faults are present, providing a robust classification even in complex cases.

Conclusion

This entire setup is designed to work in real-time, allowing for the quick detection of engine faults in an industrial setting. The proposed method has been tested on hundreds of engine samples, showing accuracy rates between 80% and 93%, depending on the fault type. The system is capable of analysing an engine in just 1.5 to 2 seconds, compared to traditional methods that could take up to 2 minutes.

By combining FFT analysis and correlation-based fault detection, the proposed method offers a fast, reliable way to monitor engine health and identify specific faults.