Anomaly detection plays an essential role in health monitoring and reliability assurance of complex system. However, previous researches suffer from distraction by outliers in training and extensively relying on empiric-based feature engineering, leading to many limitations in the practical application of detection methods. In this paper, we propose an unsupervised anomaly detection method that combines random convolution kernels with isolation forest to tackle the above problems in equipment state monitoring. The random convolution kernels are applied to generate cross-dimensional and multi-scale features for multi-dimensional time series, with combining the time series decomposing method to select abnormally sensitive features for automatic feature extraction. Then, anomaly detection is performed on the obtained features using isolation forests with low requirements for purity of training sample. The verification and comparison on different types of datasets show the performance of the proposed method surpass the traditional methods in accuracy and applicability.