An evolution updating method and system of an industrial vertical large model

Abstract

The application relates to an evolution updating method and system of an industrial vertical large model, and relates to the technical field of industrial large model evolution updating, and comprises the following steps: collecting tool cutting vibration signals collected by a micro-electro-mechanical system vibration sensor and spindle current and feed shaft torque data collected by a programmable logic controller, and packaging the data into an incremental sample set; using a prediction error mean and an uncertainty distribution standard deviation to construct an adaptive performance attenuation coefficient, combining a tool processing stage mutation mark to generate a model evolution start instruction; generating a parameter importance weight vector through momentum accumulation of a Fisher information matrix; combining knowledge distillation loss and tool life cycle comparative learning loss to perform joint fine-tuning, so as to obtain an evolved industrial vertical large model; and after verification, realizing atomic switching of a reasoning pointer through double-buffering and request dyeing marking, and completing zero-stop online replacement. The application fundamentally solves the technical contradiction between continuous evolution of an industrial field model and production continuity guarantee.

Images