已知双正态分布参数的多信号变点检测算法及相关工具咨询

Hi there! Let's break down your problem and share some targeted insights since you already have full knowledge of the two normal distributions' parameters (μ₁, μ₂, σ)—that's a big advantage over generic change point detection methods like Matlab's findchangepts.

问题核心回顾 #

You're dealing with multi-change-point detection for a signal generated by two known normal distributions (only differing in mean). The key here is leveraging the exact parameter values to get more precise results than methods that only estimate mean shifts from scratch.

相关文献方向 #

This is a well-studied subfield of change point detection, focused on parametric piecewise constant mean models with known distributions:

• Generalized Likelihood Ratio Test (GLRT) for multiple change points: Since you know μ₁, μ₂, σ, you can construct an exact likelihood function for any segmentation of the signal. The GLRT framework extends naturally to multiple change points, with methods to either fix the number of change points (using your k_est) or automatically select it via penalties (like BIC).
• Dynamic Programming-based methods: These are optimal for finding the best segmentation with a fixed number of change points (or penalty-driven selection). The core idea is to compute the minimal negative log-likelihood for all possible segmentations, which is straightforward here because each segment's likelihood can be calculated exactly using your known parameters.
• Key references to dive into:
  • Nonparametric Methods in Change Point Problems (Brodsky & Darkhovsky, 1993) – despite the title, it has a dedicated section on parametric change point problems with known distributions.
  • Exact and efficient Bayesian inference for multiple change-point problems (Fearnhead, 2006) – this Bayesian approach can be simplified significantly when you fix the distribution parameters instead of estimating them.

可用工具与实现建议 #

Matlab

While findchangepts is generic, you can easily implement a custom solution tailored to your known parameters:

1. Precompute the negative log-likelihood for every possible segment if it were generated by μ₁ or μ₂. For a segment of length n with values z, the negative log-likelihood is:

   neg_log_likelihood = sum((z - mu).^2)/(2*sigma^2) + (n/2)*log(2*pi*sigma^2);
   

2. Use dynamic programming to find the optimal k_est change points that minimize the total negative log-likelihood across all segments.
3. Alternatively, check if the Matlab Changepoint Detection Toolbox has extensions for parametric models (though the core toolbox leans toward generic methods).

Python

The ruptures library is perfect for this scenario because it supports custom cost functions:

• Define a cost function that calculates the minimal negative log-likelihood for a segment (choosing between μ₁ and μ₂, or enforcing that each segment belongs to one of the two distributions).
• Use the library's dynamic programming solver (ruptures.Dynp) with your custom cost function to find the optimal change points, either fixing k_est or using a penalty to select k automatically.

R

• The changepoint package allows custom cost functions—you can define a cost based on your known normal parameters and use it with functions like cpt.meanvar (though you'll fix the variance to σ² here).
• The strucchange package can also be adapted by specifying a parametric model with fixed mean and variance values for each segment.

实践小贴士 #

• If you have a reliable k_est, fixing the number of change points will give you the most precise results via dynamic programming (it guarantees the globally optimal segmentation).
• If k is unknown, using a penalty like BIC (which balances model fit and complexity) will help you automatically select a reasonable number of change points while leveraging your known parameters.

备注：内容来源于stack exchange，提问作者Moha