Challenges in Monitoring Industrial Machinery
The development of Industry 4.0 has led to an increasing number of production machines being equipped with sensors capable of continuously recording equipment conditions. These sensors can monitor various indicators, including vibration, temperature, electrical current, and other parameters related to machine health.
Such data are important because machine failures do not always occur suddenly. Small changes in operating patterns can provide early indications of malfunctions or component degradation.
This approach is closely related to predictive maintenance, a maintenance strategy that uses machine-condition data to anticipate failures before they cause production downtime. Unexpected machine failures can disrupt production, increase maintenance costs, and reduce system reliability.
However, cloud-based monitoring also presents its own challenges. Sensor data must be transmitted to a central server for processing. This process can increase communication latency, consume bandwidth, and make the system dependent on network availability.
This is where Edge AI becomes important. Rather than sending all data to a central server, part of the artificial intelligence processing is performed closer to the data source, such as on an edge device located near the machinery. This approach enables faster analysis and reduces dependence on communication with the cloud.
Combining Three Types of Sensor Data
Marsutiyawan Aji developed a monitoring system using three main parameters: vibration, temperature, and electrical current. Each provides different information about machine conditions.
Vibration can indicate mechanical changes, including imbalance or signs of component degradation. Temperature reflects changes in thermal conditions while the machine is operating, whereas electrical current can indicate changes in machine load.
The study analyzed 10,000 sensor observations, consisting of 8,000 normal-condition data points and 2,000 anomaly data points. The data were then processed and used to train and test the anomaly detection model.
The sensor values showed considerable variation. Vibration data ranged from 8.5–95.7 Hz, with an average of 27.4 Hz. Temperature ranged from 28.3–86.5°C, with an average of 43.8°C, while electrical current ranged from 1.2–6.8 A, with an average of 2.9 A.
Autoencoder Learns to Recognize Normal Conditions
To detect anomalies, Aji used an Autoencoder, a deep-learning approach capable of learning patterns associated with normal machine conditions.
Simply put, the model first learns what normal machine operation looks like. When it receives data with a substantially different pattern, the system produces a larger reconstruction error. This difference is then used as an indicator of an abnormal condition.
The model was trained using 7,000 samples, validated with 1,500 samples, and subsequently tested using 1,500 samples. The training process lasted for 100 epochs. The results showed that the training loss decreased from 0.0184 at epoch 20 to 0.0039 at epoch 100. The validation loss also decreased from 0.0197 to 0.0045.
Detection Accuracy Reaches 96.8 Percent
In testing with 1,500 samples, the system demonstrated relatively high performance.
- Accuracy: 96.80%
- Precision: 95.90%
- Recall: 96.30%
- F1-score: 96.10%
Of the 1,500 test samples, 1,452 samples were correctly classified. The model was also able to identify most anomalous conditions in the test dataset.
Further analysis showed that among the 1,200 samples representing actual normal conditions, the system classified 1,186 as normal and 14 as anomalous. Meanwhile, of the 300 samples representing actual anomalous conditions, 266 were correctly identified as anomalous, while 34 were classified as normal.
These results indicate that using multiple sensor parameters simultaneously can provide a more comprehensive picture of machine conditions than relying on a single type of sensor.
System Response Takes Only 0.28 Seconds
In addition to accuracy, speed is an important factor in industrial monitoring. A system that is accurate but slow to issue warnings may still be less effective when rapid responses are required.
In Aji's study, the entire process, from data acquisition to anomaly classification, took an average of 0.28 seconds. This consisted of 0.12 seconds for sensor data acquisition, 0.05 seconds for preprocessing, 0.08 seconds for Autoencoder inference, and 0.03 seconds for anomaly classification.
According to Aji of PT. Karyatama Solusindo, integrating vibration, temperature, and electrical current data enables the system to obtain more comprehensive information about machine behavior. Processing AI at the edge device also allows detection to be performed without relying entirely on cloud-based processing.
Supporting Predictive Maintenance
These findings have practical implications for industries seeking to improve their machine maintenance systems. An Edge AI system can provide early indications when machine operating patterns begin to deviate from normal conditions.
With this information, maintenance teams can obtain additional evidence to determine when inspections or maintenance actions may be necessary. In operational settings, this approach could potentially help reduce unplanned downtime and improve equipment reliability.
However, the findings do not mean that the system has been validated across all types of industrial machinery. The study used a limited number of sensor parameters and was conducted under a specific monitoring scenario. The author recommends testing the system with larger industrial datasets, different types of machinery, and real-world operating environments. Additional parameters, such as acoustic signals, pressure, and energy consumption, could also be considered in future research.
Author Profile
Marsutiyawan Aji is the author of the study Edge AI Based Anomaly Detection for Real Time Monitoring of IoT Enabled Industrial Systems. His affiliation is listed as PT. Karyatama Solusindo. The published research focuses on the application of Edge AI, Internet of Things (IoT), anomaly detection, industrial monitoring, and Autoencoders.
Research Source
Aji, Marsutiyawan. “Edge AI Based Anomaly Detection for Real Time Monitoring of IoT Enabled Industrial Systems.” Formosa Journal of Science and Technology (FJST), Vol. 5, No. 9, 2026, pp. 2523–2538. DOI: 10.55927/fjst.v5i9.162. The article was published as an open-access article.
Journal URL: Formosa Journal of Science and Technology
0 Komentar