The invention discloses a log parameter anomaly detection method based on word embedding, and the method comprises the following steps: 1, analyzing all parameters in a log, and independently dividingdiscrete parameters in all the parameters; 2, converting the discrete parameters into continuous parameter word vectors; 3, training a parameter word vector by using a long-term and short-term memoryneural network model, and predicting the parameter word vector at a subsequent target moment by using the trained parameter word vector; 4, determining the association degree of the prediction parameter word vector and the target parameter word vector by using cosine similarity, calculating a loss value through the association degree, feeding back the loss value to the network, and updating and optimizing the model until convergence; and 5, acquiring a log to perform parameter anomaly detection, calculating cosine similarity between the prediction parameter and the target parameter, and if the cosine similarity is lower than a threshold value, determining that the log parameter is abnormal. The detection bottleneck caused by parameter dynamics and difference can be effectively solved, andthe overall accuracy of log detection is improved.