Wireless sensor networks (WSNs) are exposed to several threats that may lead the nodes to be damaged and provide fault estimations. However, the existing methods have failed to detect anomalies accurately due to the dynamic nature of input and require more data for predicting anomalies. To overcome this problem, a self-adaptive-weighted kernel ridge regression (SA-WKRR) is employed for a prediction model in anomaly detection for wireless sensor networks. The dataset used for anomaly detection was Intel Berkeley Research Lab (IBRL). Initially, the sensor data are fed to data compression using principal component analysis (PCA) technique to reduce dimensions of data. After that, SA-WKRR prediction model predicts anomalies by scaled Euclidean distance and Gaussian kernel functions. Finally, anomalies in WSN are detected by dynamic thresholding technique, which generates threshold dynamically through sliding window technique. The experimental analysis indicates promising results for anomaly detection with an accuracy of 99.90%, a precision of 99.89%, and an F1-score of 99.89%, which is higher than other detection methods like extreme learning machine (ELM), global and local perspectives of supervised learning-graph convolutional network (GLSI-GCN), hybrid model one-class support vector machine and isolation forest (HMOI), and fuzzy method.