Industrial systems are becoming more digital and connected through IoT technologies and subsequently the need for accurate and scalable solutions for finding anomalies is growing. A major issue in the field of anomaly diagnosis is that there isn’t enough labeled fault data, which makes it difficult to develop reliable machine learning models. This research gap is addressed by the generation of synthetic data, producing physically plausible time-series signals that replicate normal and faulting operating conditions of a system. A variational autoencoder is trained to learn how a normal system works and to find anomalies by looking at reconstruction errors. This paper presents an anomaly detection framework which is trained on synthetic sensor data (temperature, vibration, pressure, and power metrics) to encompass a diverse range of industrial systems. The performance metrics result shows that the model achieved its goal in detection with an accuracy of 88.75%, a precision score of 81.39%. The sensitivity and F1 scores were 94.68% and 87.54%, respectively. The findings conclusively showed the benefits of implementing synthetic data with VAE models for anomaly identification in situations where fault labeled data is not easily accessible particularly in IoT settings.