<p>Real-time and high-quality acquisition of key sensing data is an essential prerequisite for accurate monitoring of machining performance of computer numerical control machine tools. However, unfavorable factors such as coolant spray and magnetic interference can lead to abnormal sensing data, posing significant challenges to machine tool performance monitoring. This paper presents an anomaly recognition method for high- and low-frequency sensing data in complex machining environment of machine tools. Using the long short-term graph neural network (LSTGNN) model, the low-frequency temperature sensor data is accurately predicted, and a low-frequency anomaly sensing mode evaluation system is constructed. By comparing the measured values with the predicted values, the causes of anomalies in low-frequency sensors are accurately identified. Simultaneously, continuous wavelet transform and the LSTGNN model are employed to mine the temporal, frequency and spatial characteristics of high-frequency data. These features are subsequently used to accurately locate abnormal sensors and recognize abnormal moments. The effectiveness of this method is validated on a spindle test bench and a grinding machine. The results demonstrate that the prediction error of the LSTGNN model is reduced by more than 94.2% compared to temporal convolutional network, long-short term memory, and other time series analysis methods. Furthermore, compared to conventional methods such as the convolutional neural network, graph neural network and Transformer, the accuracy of high-frequency data anomaly recognition is improved by more than 10.72%.</p>