Pattern recognition method based on multi-scale and multi-task learning

Abstract

The invention discloses a pattern recognition method based on multi-scale and multi-task learning. The method comprises the following steps: representing a complex pattern of original information on multiple time scales through multi-scale feature extraction; training pattern recognition tasks under multiple scales at the same time by constructing a multi-task learning framework; introducing a parameter sharing unit between tasks of different scales, sharing cross-scale feature information, expanding the breadth and depth of the feature learning process, and finally improving the performance of pattern recognition. The method can be stably and reliably used for intelligent information processing in a complex environment state, an effective method is provided for timely and effectively achieving a mode recognition task, and the method has high precision and generalization performance.

Description

technical field [0001] The invention relates to information acquisition and processing methods; in particular, it relates to a pattern recognition method based on multi-scale and multi-task learning. Background technique [0002] Non-stationary time-varying characteristics are the main characteristics of obtaining information under objective natural environment and actual working conditions. A large amount of background noise and interference signals are mixed in this type of information, and the information usually has high complexity, coupling and uncertainty. At the same time, the observed objects also have different modes and states. This allows information to contain complex patterns on multiple time scales. Conventional global or single-scale feature extraction and analysis methods are often difficult to fully characterize and classify pattern information. In recent years, advanced technologies have attempted to improve the ability to represent multi-class patterns ...

AI Technical Summary

Problems solved by technology

[0003] Taking rolling bearing fault diagnosis as an example: ① Vibration signals usually exhibit multi-scale characteristics and contain complex patterns on multiple time scales. This inherent multi-scale feature is often ignored by traditional methods due to the lack of effective multi-scale feature extraction methods. The diagnostic model is ignored; ②Because the operating conditions of rolling bearings are complex and changeable and contain a lot of background noise, the original time domain signal often contains a lot of redundant information; ③In the process of feature learning, traditional methods mostly focus on a single task , ignoring other information that may help optimize the metrics, making feature learning incomplete, and ultimately affecting the generalization performance of the model

Images