Change Point Detection in the Frequency Domain with Statistical Reliability

Authors: Akifumi Yamada, Tomohiro Shiraishi, Shuichi Nishino, Teruyuki Katsuoka, Kouichi Taji, Ichiro Takeuchi

TMLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results demonstrate that the proposed method reliably identifies genuine CPs with strong statistical guarantees, enabling more accurate root-cause analysis in the frequency domain of complex systems. ... In Section 5, we demonstrate the effectiveness of our proposed method through comprehensive numerical experiments using both synthetic and real-world data.
Researcher Affiliation Academia Akifumi Yamada EMAIL Nagoya University Tomohiro Shiraishi EMAIL Nagoya University RIKEN Shuichi Nishino EMAIL Nagoya University RIKEN Teruyuki Katsuoka EMAIL Nagoya University Kouichi Taji EMAIL Nagoya University Ichiro Takeuchi EMAIL Nagoya University RIKEN
Pseudocode Yes Algorithm 1 metropolis_algorithm Input: Difference of objective function values E, temperature c 1: Uniformly sample ξ from [0, 1) 2: if E + c ln ξ < 0 then 3: status Acceptance 4: else 5: status Rejection 6: end if Output: status Algorithm 2 CP Candidate Selection by Simulated Annealing Input: Time sequence x 1: Obtain spectral sequences f (d) for d {0, . . . , D 1} in (3) by applying STFT to x 2: Initialize the CP candidates T as T init in (7) using dynamic programming 3: Determine the initial temperature c in a preliminary experiment 4: while true do 5: while The number of local search is less than |Dinit| T do 6: Frequency d is uniformly sampled from Dinit 7: The operation is uniformly sampled from adding, removing, and moving a CP for d 8: Obtain T by applying the operation to T 9: status metropolis_algorithm( E(T , T , x), c) 10: if status is Acceptance then 11: T T 12: end if 13: end while 14: Obtain T by merging uniformly sampled two adjacent CP locations in T 15: status metropolis_algorithm( E(T , T , x), c) 16: if status is Acceptance then 17: T T 18: end if 19: if no transition to neighborhoods occured at c then 20: break 21: end if 22: Update the temperature using a constant factor λ as c λc 23: end while Output: CP candidates T
Open Source Code Yes For reproducibility, our implementation is available at https://github.com/Takeuchi-Lab-SI-Group/si_for_frequency-domain_change_ point_detection.
Open Datasets Yes We used the set No.2 of the IMS bearing dataset, which is provided by the Center for Intelligent Maintenance Systems (IMS), University of Cincinnati (Qiu et al., 2006) and is available from the Prognostic data repository of NASA (Lee et al., 2007).
Dataset Splits Yes We conducted CP candidate selection for the sensor data of bearing 1 in the frequency domain for two periods: 0.25–2.25 days when no significant changes in the spectra existed and 2.25–4.25 days when the BPFO harmonics exhibited the spectral amplification. For each period, we computed the DFT of M = 1024 consecutive points in the 20480 samples and repeated the procedure T = 2 days/10 min = 288 times. ... The variance σ2 for testing was estimated from the data on 0–0.25 days that did not contain significant spectral changes and was not used in any of the experiments.
Hardware Specification Yes All numerical experiments were conducted on a computer with a 96-core 3.60GHz CPU and 512GB of memory.
Software Dependencies No The paper does not explicitly list specific software dependencies with version numbers.
Experiment Setup Yes In all synthetic experiments, we set window size M {512, 1024}, the number of frequencies D = M / 2 + 1, the length of sequence N = M T, where T was specified for each experiment, the sampling rate fs = 20480... We used BIC for the choice of penalty parameters β and γ as indicated in Appendix A, and set the parameters of simulated annealing as c+ 0 = 1000, λ+ = 1.5, η(c0) = 0.5 and λ = 0.8 in Section 3.4. After detecting CP candidates, a CP candidate location τ det k randomly selected from τ det was tested at the significance level α = 0.05. ... For other parameters related to the CP candidate selection algorithm, please refer to the experimental setup in Section 5.2.