Analysis of longitudinal phase space evolution of fresh injected beam in electron storage ring based on image matching

Abstract This paper proposes an image-matching-based longitudinal phase space evolution analysis framework to address the difficulty of reverse calibration of initial beam parameters during electron storage ring injection. First, using custom-developed single beam injection simulation software and the High-speed Oscilloscope-based Three-dimensional bunch Charge And Position (HOTCAP) measurement, a turn-by-turn evolution of the longitudinal charge distribution dataset covering different noise levels (10–60 dB) was constructed across varied parameter spaces, providing a basis for within-group ( K = 30) benchmark evaluation of the proposed image matching algorithms. Building on this dataset, this paper systematically implements and compares three efficient image matching methods: a pixel-by-pixel matching method directly quantifies image differences by calculating the average of pixel grayscale differences; a Fourier transform matching method maps spatial-domain images to the frequency domain, combining amplitude and phase spectrum differences to enhance robustness against periodic structures; and an eigenvector matching method uses a pre-trained MobileNetV2 network to extract low-dimensional eigenvectors and measures the similarity of global statistical features using cosine similarity. We conducted a theoretical analysis of the time complexity and physical applicability of the three methods and statistically analyzed their matching accuracy in 20 independent experimental trials. The experimental results show that under low signal-to-noise ratio and narrow parameter space conditions, the pixel-by-pixel matching method performs most robustly; the Fourier transform matching method can achieve a matching accuracy of ≥95% in medium to high signal-to-noise ratio environments; and the eigenvector matching method achieves 100% accuracy in high signal-to-noise ratio and wide parameter range conditions. Overall, each method has its advantages, and they can be dynamically fused and weighted based on actual diagnostic requirements to balance noise robustness, periodicity sensitivity, and computational efficiency. The findings of this work demonstrate a shift from iterative parameter optimization to image-based intelligent matching, providing practical tools for optimizing synchrotron radiation source injection. Moreover, the approach illustrates how machine learning can be integrated with accelerator physics to support high-precision beam control. Looking forward, adaptive noise-resilient hybrid matching models and physics-informed neural networks, coupled with hardware acceleration, could enable online diagnosis and real-time control.