The application discloses a hardware-aware
dynamic neural network architecture search method based on industrial
edge computing, and comprises the following steps: constructing a
neural network architecture search space including a dynamic adjustment node, the dynamic adjustment node corresponding to multiple execution branches, the execution
branch corresponding to multiple neural network sub-paths, the dynamic adjustment node being internally provided with a hardware-aware
router, the hardware-aware
router inputting a hardware
state vector of a target industrial
edge device and outputting a probability distribution of each execution
branch, deploying a hardware monitoring agent on the target industrial
edge device to access an evaluation loop of the architecture search, building a dynamic evaluation environment, initializing a candidate architecture
population, dynamically distributing the code of the candidate architecture and performing reasoning, collecting reasoning-related data, selecting an execution
branch under different combination conditions, recording a dynamic
adaptation score of the architecture search, screening candidate architectures in the candidate architecture
population that meet search conditions, generating a new generation of candidate architecture
population, and outputting an optimal
dynamic neural network architecture when a convergence condition is reached.