Investigation of a multisensor-based breather valve fault diagnosis method

To meet the requirements of new engineering while developing students' abilities and practical skills, this study develops a multisensor fusion-based fault diagnosis method for oil storage tank breather valves. The breather valve is a critical mechanical component for reducing the volatilization loss of oil and ensuring tank safety. Due to its long-term working time, some failure phenomena are common, such as leakage, rust, or seizure. Therefore, it is important to investigate the real-time monitoring and fault diagnosis method, reducing the probability of damage to the storage tank caused by breather valve failure and ensuring production safety.

A monitoring system for the breather valve based on a single-chip microcomputer is established, the functions of which include real-time monitoring, data acquisition, data storage, data preprocessing, and threshold alarms. Four typical failure phenomena of the breather valve—nonfaulty, leakage, rust, and seizure—are assessed. The characteristic signal analysis methods for breather valve disc movement under failure conditions are also discussed. Based on the measured displacement signals, the vertical acceleration signals of the valve disc are further used for feature extraction in the time, frequency, and time–frequency domains. Five dimensional parameter indicators of the acceleration signal (maximum value, minimum value, variance value, peak-to-peak value, and root mean square value) in the time domain are discussed, and three dimensionless parameter indicators (kurtosis, impulse factor, and margin factor) are also investigated.

The maximum values of the acceleration signals for nonfaulty, leaking, rusty, and seized valves are most obviously. The minimum values of the valve disc acceleration signals are smallest for the seized valves and largest for the rusty valve. The average values of the valve disc acceleration under the four typical failure phenomena are quite similar. The variance values of nonfaulty valves and seized valves are small. The peak-to-peak values of the rusty valves are highest, while those of the seized valves are smallest. The root mean square (RMS) values are highest for the rusty valves and lowest for the seized valves. The RMS of the leaky valves is slightly larger than that of the nonfaulty valves. Compared with the dimensional parameter indicators, the dimensionless parameters are less affected by the environment. All the dimensionless parameters of the rusty valves are large, while the dimensionless parameters of the leaky, rusty, and seized valves are slightly different. The fault feature extracted in the frequency domain is the frequency standard deviation. In the time–frequency domain, complementary empirical ensemble mode decomposition and wavelet packet transform are used to extract the fault signals. The wavelet packet transform has a better signal decomposition effect and faster decomposition efficiency. Therefore, the fault feature extracted in the time–frequency domain is the third-layer band energy value after wavelet packet decomposition. Finally, the multisensor-based breather valve fault diagnosis method is assessed. The feature-level and data-level fusion methods are used to combine the eigenvalues of the multisensor signals into eigenvectors. Multiple acceleration signals are processed through feature-level fusion, extracting parameter indicators and then consolidating them into a feature vector. Meanwhile, the displacement signals using the data-level fusion are used to calculate the maximum displacement values and the valve disc angles during breather valve movement, which are then incorporated into the feature vector. Taking this merged new feature vector as the input, a breather valve fault diagnosis model was built using the least-squares support vector machine optimized by the particle swarm algorithm, and the fault state of the breather valve was accurately identified.

The classification of 60 breather valve samples with typical failure phenomena was conducted. The recognition accuracy of the fault diagnosis model is 96.66%. This engineering case could help students understand sensor technology applications in real projects, improve their grasp of sensor operation and signal processing, and strengthen their capabilities in practical projects.