Lubricating oil is a critical component in machinery, and accurately assessing its condition is vital for equipment health monitoring. However, when parameters approach threshold values, traditional anomaly detection methods often misinterpret normal fluctuations as anomalies or fail to identify early abnormal trends, reducing detection accuracy. To address this, we propose a multi-level anomaly detection model, BayesLSTM-VAE, based on Bayesian inference, Long Short-Term Memory (LSTM) networks, and Variational Autoencoders (VAE), using wear data from online monitoring of diesel engine oil. The model first employs LSTM to predict multi-parameter time series and quantifies uncertainty in the predictions using variational Bayesian methods to generate confidence intervals for each parameter. Two types of test sets are then processed by VAE: one with predicted values, and the other incorporating confidence intervals as enhanced features. By analyzing reconstruction error differences between the two sets, the model effectively detects trends in equipment health and refines status classification. Experimental results show that the trend prediction model excels in accuracy and loss control, with anomaly detection results aligning closely with actual monitoring data. Compared to traditional LSTM-VAE and BayesLSTM-Mahalanobis Distance methods, the BayesLSTM-VAE model achieves a high alarm rate (93.33%) while maintaining a low false positive rate (0.07%). It also identifies normal fluctuations and detects potential abnormal trends 85   minutes in advance. In conclusion, the BayesLSTM-VAE model can accurately identify abnormal lubricating oil conditions, detect potential abnormal trends in advance, improve anomaly detection reliability, and provide new insights for industrial equipment health assessment