Wireless Sensor Networks (WSNs) are often used for critical applications where trust and security are of paramount importance. Trust evaluation is one of the key mechanisms to ensure the security and reliability of WSNs. Traditional trust evaluation schemes rely on fixed, predetermined thresholds, or rules and static attack models, which may not be suitable for all situations such as dynamic and heterogeneous network environments with new and unknown attack scenarios as well as have several problems such as limited security and scalability, limited accuracy, incomplete coverage, lack of adaptability that can limit their effectiveness. Machine Learning (ML) has been shown to be an effective tool for trust evaluation in WSNs, offering several benefits over existing schemes such as greater adaptability, scalability, and accuracy since ML algorithms can analyze and learn from the data collected in real-time from multiple sources (sensor readings, network traffic, and user behavior) enabling them to dynamically adjust their decision-making criteria based on the current network conditions. Trust-aware ML-based security mechanisms achieve safety and efficient decision-making by reducing uncertainty and risk to accomplish real-world tasks. This paper presents a Machine Learning (ML)-based trust evaluation model in the unattended autonomous WSN environment to achieve reliability, adaptability, scalability, and accuracy by generating quick and reliable trust values dynamically. The proposed machine learning algorithm extracts various trust features such as Co-Location Relationship (CLR), Co-Work Relationship (CWR), Cooperativeness-Frequency-Duration (CFD), and Reward (R) to obtain a robust trust rating of sensor devices and predict future misbehavior. These trust features are combined to generate a final trust rating before making any decision about the reliability of any sensor device. Moreover, the projected trust model (ETDMA) integrate direct communication trust and indirect trust with the help of a logical time window that periodically records the trustworthy and suspicious interactions. Simulation experiments exhibit the effectiveness of the proposed trust evaluation method in terms of change in trust values, malicious nodes detection (94%), FNR (0.9%), F1-Score (0.6), and accuracy (92%) in the presence of 50 malicious nodes.