The optical fiber sensing system is widely used in many fields, such as long-distance oil and gas pipelines, tunnel safety detection, large structural safety detection, perimeter security detection and so on. Optical fiber sensing signal identification plays a key role in real-time monitoring, abnormal alarm and other aspects. Its working performance directly determines the performance of the real-time, accuracy and stability of the optical fiber sensing detection system. Therefore, fast and accurate identification and classification is of great significance to ensure the safety of various fields and reduce the cost loss caused by equipment damage.To improve the real-time and accuracy of optical fiber vibration signal pattern recognition, a feature extraction algorithm based on compensation distance estimation is proposed. The algorithm draws on the human auditory perception mechanism and extract the MEL frequency cepstrum coefficients from the optical fiber sensing vibration signals. The algrithm uses the compensation distance estimation technology to formulate the feature selection strategy,and finally realizes feature evaluation and optimization. The MFCC feature extraction algrithm can extract the feature of the vibration signal acquired by the optical fiber sensing system, and then identifies the interference signal according to the modal prediction.However, the extracted feature vector by the MFCC feature extraction algorithm has the problems of high dimension and vector redundancy . When these feature vectors are trained and recognized by the classifier, it will increase the time cost and reduce the recognition accuracy. Therefore, effectively reduce the dimension of the MFCC eigenvector is the key to improving the real-time performance and accuracy of optical fiber sensing vibration signals.This paper proposes a feature extraction method based on compensated distance estimation. The CDET algorithm jointly evaluates the intra-class and inter-class discreteness of eigenvectors. The feature evaluation is performed on different dimensions of the feature matrix, and the redundant vector of low score is deleted from the original feature vector matrix, thereby realizing feature dimension reduction. Solve the influence of redundant vectors on classification, avoid the problem of complex operation caused by too many extracted feature dimensions, and improve real-time performance.The steps of the algorithm are to calculate the average distance between samples of the same condition and different conditions , then average the within-class and between-class distances, and then calculate the within-class variance and between-class variance factors. Then calculate the compensation coefficient between the two variance factors, calculate the ratio of the inter-class distance to the intra-class distance, multiply the compensation coefficient to obtain the distance evaluation standard, and select a better feature dimension according to the distance evaluation standard. The experimental results show that the feature extraction algorithm of vibration signal based on compensation distance estimation technology can effectively reduce the redundant vectors that affect the classification accuracy in optical fiber sensing system. It solves the problems of feature representation and operation complexity for the vibration signal, and further improves the effectiveness and real-time performance of vibration signal pattern recognition of the optical fiber sensor system. Compared with Principal Component Analysis (PCA), our algorithm has the same performance in the low-dimensional case. With the increase of dimension, the performance of the algorithm in this paper is better in recognition accuracy. In terms of anti-noise performance, in the presence of noise, in order to improve the feature identification, the dimension of the feature vector extracted from the MFCC feature increases. The PCA method is difficult to distinguish the features caused by noise, and the algorithm in this paper can reduce the influence of superimposed noise by pruning redundant vectors, and can extract feature vectors with high feature recognition. Therefore, the proposed algorithm also has certain anti-noise performance.