Lightweight unmanned aerial vehicles anomaly detection model based on synaptic evolution mechanism and layer-adaptive neural network

The wide application of unmanned aerial vehicles (UAVs) puts strict requirements on reliable operation, and anomaly detection is a crucial method to ensure the reliability of UAVs. Existing anomaly detection models are highly dependent on time-series log data, and models based on Long Short-Term Memory (LSTM) are widely used due to their effectiveness in processing time-series data. However, the complex internal structure of LSTM involves many learning parameters. In addition, the traditional static parameter pruning methods fail to balance the conflict between performance and parameter scale dynamically. To address the above problems, this paper proposes a lightweight anomaly detection model based on the synaptic evolutionary mechanism and layer-adaptive neural network (LUV-DSA), which can be deployed in resource-constrained UAV application scenarios. Firstly, LUV-DSA simplifies the internal structure of LSTM by optimising the cell state update process with a new linearly weighted computational method. Secondly, inspired by the evolution of biological synapses, a method for intra-layer parameter pruning and inter-layer structured pruning is designed. For intra-layer parameters, LUV-DSA achieves dynamic model parameter competition by simulating the self-optimisation of synapses, minimising parameter scale while ensuring performance. For inter-layer structures, LUV-DSA enables inter-layer adaptation by calculating plasticity factors to assess the contribution of each layer. The experimental results show on seven UAV datasets that the model significantly reduces the number of parameters and inference time while ensuring accuracy. For example, on the ALFA dataset, LUV-DSA achieves 99.51 % accuracy with 96.14 % fewer parameters than MobileNetV4.