Condition-Based Monitoring of Biogas Compressor Using Nonintrusive Sensors and Machine Learning

APPLICABILITY The objective of this work is to develop an anomaly detection approach for biogas blowers using nonintrusive vibration sensors for machine conditions and health management. High-frequency vibration signals are collected from a tri-axial accelerometer installed on the exterior of the biogas blower casing. Long short-term memory (LSTM) artificial neural network algorithm is applied for signal classification. The proposed condition-based monitoring approach could provide a low-cost non-intrusive solution for asset management of compressors at water resource recovery facilities. INTRODUCTION Predictive maintenance (PdM) organizes maintenance activities according to the actual health state of machine tools, equipment, and systems [1]. Diagnostic PdM detect detects faults, determines root causes of operation anomalies, and estimates the current health state of a system under investigation to prevent unexpected failures. Condition-based monitoring is an approach for realizing diagnostic PdM as standardized by ISO 17359:2018. Condition-based monitoring can provide significant advantages in relation to product quality, safety, availability, and cost reduction in industrial applications. The recent advent of information and communication technologies and artificial intelligence presents a great opportunity to enhance the practice of data-driven assessment of the operation and maintenance of wastewater infrastructure [2, 3]. The objective of this study is to develop an anomaly detection approach for biogas compressors through vibration analysis for predictive maintenance. To achieve this goal, piezoelectric sensors and edge computing platforms (Raspberry Pi and Arduino board) are deployed on the biogas compressors. Machine learning algorithms are applied for classifying the collected vibration data for anomaly pattern recognition. METHODS The vibration monitoring in this study is implemented for compressor Unit 1 (KAESER) and compressor Unit 2 (SIEMENS) in a biogas facility. A low-power tri-axial accelerometer (model# ADXL345, Analog Devices, Wilmington, MA) is used with serial data communication devices such as Raspberry Pi and Arduino boards to collect vibration data from compressors. The data collection was initiated from the point of time when the compressor operated in normal condition and continued to the occurrence of any failures. The sampling frequency of the accelerometer is 1k Hz for each axis with a measurement interval of 30 minutes. Data are measured and obtained via Inter-Integrated-Circuit (I2C) bus. During the measurement, the sensor collects 10000 data points for each axis, which would take approximately 10 seconds, and is set to idle until the next measurement. The structure of the implemented condition-based monitoring is shown in Figure 1. The raw data includes the time stamp and acceleration data for the x, y, z-axis, and the sampled time for the corresponding sample. Sample x-axis vibration signals are visualized in Figure 2. The machine diagnosis is realized by scoring the anomaly levels of the vibration data, which will increase as failure approaches. Hence, an anomaly scoring model is tested to characterize the different levels of anomalies and provide an informative interpretation of forthcoming failures. The implemented model is inspired by dilated RNN network and Temporal Hierarchical One Class (THOC) network, which contributed to solving complex dependencies and vanishing gradients problems in layers. The THOC network is adopted and modified to develop the anomaly scoring model. The model consists of 3 long-short term memory (LSTM) layers with dilated skip connections. Details of the method implementation could be found elsewhere [6]. The sequential length of the transformed data is 36, and the input batch size is set to 128. In Fig. 3, each node represents a hidden RNN cell. The number of input nodes highlighted as green is determined by the sequential length of input data, which is 36. Similarity score obtained from the score function is translated to loss using the following equation that will be optimized through a training process. Adaptive Moment Estimation (Adam) optimizer is employed as it is known to have good capabilities of reaching global minimum with a high dimensional dataset. The structure of the implemented condition-based monitoring is shown in Figure 1. RESULTS The results of the model are shown in Figure 3. The input data (time-frequency domain) is visualized using power spectrum analysis. It is important to be noted that the Euclidean distance returns the length of a line segment between two vectors. Thus, the test data with a larger distance from the centers in each layer are expected to be assigned higher anomaly scores. As shown in Figure 3, anomaly points are captured when the power spectrum has abnormal patterns or when a peak value pops out from the power spectrum. Also, data points with small anomaly scores are not identified as anomalies, which might be sound as not all anomalies cause machine failures. Thus, a threshold should be carefully determined as it plays a critical role in linking such anomaly information with maintenance strategies. In this study, the threshold value is selected based on the 99.9 percentile in the anomaly scores. Different threshold values are used with an optimal threshold that prevents generating excessive false positives. The results may need to be compared to the actual machine condition such as whether abnormal operations were present while it was monitored. Since Euclidean distance may not be intuitive on a high dimensional dataset, different scoring functions are also tested for comparison. Chebyshev scoring function is adopted and the result is visualized in Figure 4. Chebyshev distance provides the greatest differences between two different vectors along any coordinate dimension. Hence, input data with the highest variance and difference are likely to obtain higher anomaly scores. As shown in Figure 4, the power spectrum with peak values tends to be identified as anomalies. In this case, the warning stage depicted with the orange line was applied with an additional threshold (99 percentile) to demonstrate a warning system that may be applicable to the data-driven maintenance strategy. In this study, the anomaly scoring model has been utilized to detect the abnormal vibration patterns and interpret them into anomaly scores. Based on the results from the proposed anomaly scoring model, a notification strategy can be established for further maintenance scheduling. The operation and maintenance manual or guidelines are provided by manufacturers of the machines. An example of the operation and maintenance manual of a compressor is shown in Table 1. Proper maintenance requires timely tracking of operation hours followed by appropriate actions such as checking oil levels, changing filters, and grease bearings. CONCLUSIONS AND RELEVANCE This study presents a condition-based monitoring approach for biogas compressors as an example of data-driven asset assessment. The test anomaly scores were calculated based on a similarity between unseen data and clusters that are learnable by machine learning algorithms. The study shows that condition-based monitoring using nonintrusive vibration sensors could be a useful asset management tool and providing early failing alarms of equipment embedded in WRRFs. The result of this study suggests that the operator's supervisory control and condition-based monitoring could be integrated for asset management (Figure 5). A combination of preventative maintenance following manufacturer's O&M manual and predictive maintenance guided by condition-based monitoring could lead to adaptive asset management program. Meanwhile, deploying condition-based monitoring solution at WRRFs may add additional responsibilities to the maintenance crew at WRRFs. Future studies are required to understand the operator demand, capital investment, and cybersecurity implications of condition-based monitoring for WRRFs. Attendance at this presentation will benefit operators and design engineers who are interested in data-driven asset management through condition based-monitoring. This study presented a novel application of deploying nonintrusive vibration sensors and edge computing for anomaly detection of biogas compressors. The method developed in this study could be applied in asset monitoring of other rotary equipment in WRRFs.