Parameter space change point detection method and device, terminal and storage medium

Abstract

The invention provides a parameter space change point detection method and device, a terminal and a storage medium. The method comprises the following steps: determining parameter priori, and calling a simulator to generate an offline simulation observation sequence containing various change point change forms; constructing a training sample pair based on the sequence, and training a window-level neural posteriori estimator to obtain a parameter inference model; then, performing sliding window inference on the online actual observation data by using a fixed complexity forward inference mode after spreading, and calculating parameter posteriori distribution of each window by using a parameter inference model to obtain a parameter posteriori sequence and a statistical magnitude thereof; and performing uncertainty perception change point detection based on the sequence and the statistical magnitude thereof to obtain a change point set and attribution information of each target change point. According to the method, sliding window level parameter posterior distribution is obtained through simulation inference, change points are directly detected in a parameter space in the online stage, attribution information is output, real-time detection can be achieved, delay is reduced, and interpretability is improved.

Description

Technical Field

[0001] This invention relates to the field of computer and information processing technology, and in particular to a method, apparatus, terminal and storage medium for detecting spatial variable points of parameters. Background Technology

[0002] Change point detection (CPD) in time series is used to identify locations in which statistical characteristics or generation mechanisms change abruptly. It is widely used in fields such as industrial equipment monitoring, financial risk control, network operation and maintenance, climate and life sciences.

[0003] Currently, existing change point detection technologies often suffer from high false positive rates, high false negative rates, large detection delays, and difficulty in explaining "why the change occurs" in nonlinear, strongly correlated, non-Gaussian noise, or chaotic systems. In complex models where likelihoods are unavailable but simulations are possible, online Bayesian methods based on explicit likelihoods have high modeling and computation costs and are difficult to deploy on the edge. While deep end-to-end detection has accuracy advantages, it relies on large-scale annotation, lacks cross-domain generalization, and has insufficient interpretability.

[0004] Therefore, existing technologies have shortcomings and need to be improved and developed. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a method, device, terminal and storage medium for detecting parameter space change points, in order to address the above-mentioned defects of the prior art, and to solve the problems of large detection delay and insufficient interpretability of change points in the existing change point detection technology.

[0006] The technical solution adopted by this invention to solve the technical problem is as follows: In a first aspect, the present invention discloses a method for detecting variable points in parameter space, wherein the method includes: Determine the parameter priors and call the simulator to generate an offline simulated observation sequence containing multiple variable point change forms based on the parameter priors; Training sample pairs are constructed based on the offline simulated observation sequences, and the training sample pairs are used to train a window-level neural posterior estimator to obtain a parameter inference model. Using a fixed-complexity forward inference method after amortization, sliding window inference is performed on the online actual observation data and input into the parameter inference model to calculate the parameter posterior distribution of each window, so as to obtain the parameter posterior sequence and its statistics that evolve over time; Uncertainty-aware change point detection is performed based on the parametric posterior distribution sequence and its statistics to obtain the corresponding change point set and the attribution information of each target change point in the change point set.

[0007] Optionally, the step of determining the parameter prior and invoking the simulator to generate an offline simulated observation sequence containing multiple change point variations based on the parameter prior includes: The parameters are sampled from the prior range of the predefined parameter vector to obtain the parameter prior, and a pre-constructed set of variable point transformation forms is obtained. A simulator that can be simulated but whose likelihood function cannot be explicitly obtained is invoked to generate time series data based on the prior parameters of the simulator; According to the set of variable point change forms, observation noise, missing segments, sampling rate changes, and domain randomization perturbations are added to the time series data to obtain an offline simulated observation sequence containing multiple variable point change forms.

[0008] Optionally, constructing training sample pairs based on the offline simulated observation sequence includes: The offline simulated observation sequence is processed by sliding window according to the preset window length and the first preset step size, and the temporal features of each window are extracted to construct training sample pairs that pair windows with parameters.

[0009] Optionally, the method of using amortized fixed-complexity forward inference to perform sliding window inference on online actual observation data includes: The forward inference method with amortized fixed complexity is used, and sliding window inference is performed on the online actual observation data according to the second preset step size.

[0010] Optionally, the uncertainty-aware change point detection based on the parametric posterior distribution sequence and its statistics, to obtain the corresponding change point set and the attribution information of each target change point in the change point set, includes: Generate parameter trajectories and uncertainty trajectories based on the posterior sequence of the parameters and their statistics; Based on the parameter trajectory and the uncertainty trajectory, calculate the change point statistics of uncertainty perception and determine the current calibration threshold; Based on the change point statistics, and according to the current calibration threshold, candidate change points are generated; Perform time persistence criterion operation and nonmaximum suppression merging operation on the candidate change points that are temporally adjacent or repetitive to obtain the corresponding change point set and the attribution information of each target change point in the change point set.

[0011] Optionally, determining the current calibration threshold includes: Based on the obtained threshold baseline information, the current calibration threshold is determined according to the preset false alarm rate; wherein, the threshold baseline information is information used for threshold setting obtained by performing posterior consistency checks and calibrations on the trained neural posterior estimator.

[0012] Optionally, the change point statistics include covariance normalized distance, posterior entropy / divergence transition, and confidence interval crossing test; The step of generating candidate change points based on the change point statistics and according to the current calibration threshold includes: Based on the multi-point statistic fusion strategy, at least two of the point statistics in the covariance normalized distance, posterior entropy / divergence transition and confidence interval crossing test are merged or weighted to obtain the processed point statistics. Based on the processed change point statistics, candidate change points are generated according to the current calibration threshold.

[0013] Secondly, the present invention also discloses a parameter space change point detection device, wherein the device comprises: The simulated observation sequence generation module is used to determine the parameter priors and call the simulator to generate an offline simulated observation sequence containing multiple variable point change forms based on the parameter priors. The training module is used to construct training sample pairs based on the offline simulated observation sequence and use the training sample pairs to train a window-level neural posterior estimator to obtain a parameter inference model. The online inference module is used to perform sliding window inference on the online actual observation data using a forward inference method with amortized fixed complexity and input it into the parameter inference model to calculate the parameter posterior distribution of each window, so as to obtain the parameter posterior sequence and its statistics that evolve over time; The change point detection module is used to detect change points based on the parameter posterior distribution sequence and its statistics to perceive uncertainty, and to obtain the corresponding change point set and the attribution information of each target change point in the change point set.

[0014] Thirdly, the present invention discloses a terminal, comprising: a memory, a processor, and a parameter space change point detection program stored in the memory and executable on the processor, wherein the parameter space change point detection program, when executed by the processor, implements the steps of the parameter space change point detection method as described above.

[0015] Fourthly, the present invention discloses a computer-readable storage medium storing a computer program that can be executed to implement the steps of the parameter space change point detection method as described above.

[0016] This invention provides a method, apparatus, terminal, and storage medium for detecting parameter spatial change points. The method includes: determining parameter priors and calling a simulator to generate an offline simulated observation sequence containing multiple change point variation forms based on the parameter priors; constructing training sample pairs based on the offline simulated observation sequences and using the training sample pairs to train a window-level neural posterior estimator to obtain a parameter inference model; performing sliding window inference on online actual observation data using an amortized fixed-complexity forward inference method and inputting it into the parameter inference model to calculate the parameter posterior distribution of each window, obtaining the parameter posterior sequence and its statistics that evolve over time; and performing uncertainty-aware change point detection based on the parameter posterior distribution sequence and its statistics to obtain a corresponding change point set and attribution information for each target change point in the change point set. Therefore, this invention utilizes simulation inference to obtain the posterior distribution of parameters at the sliding window level, and performs uncertainty-aware change point detection on the parameter posterior sequence in the online stage. Since the online stage uses a forward inference method with amortized fixed complexity, it can achieve real-time detection of change points, thereby reducing detection latency. Furthermore, it can directly detect change points in the parameter space and output the attribution information of change points, which can clarify why the change points change and improve the interpretability of change points. Attached Figure Description

[0017] Figure 1 This is a flowchart of a preferred embodiment of the parameter space variable point detection method in this invention; Figure 2 This is a flowchart of a specific parameter space variable point detection method disclosed in this invention; Figure 3 This is a schematic diagram of the logical architecture of the parameter space variable point detection method in this invention; Figure 4 This is a functional principle block diagram of a preferred embodiment of the parameter space variable point detection device in this invention; Figure 5 This is a functional principle block diagram of a preferred embodiment of the terminal in this invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer and more explicit, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0019] Currently, existing change point detection techniques include: statistical and optimization methods for the observation space, online detection using parameter space and Bayesian methods, simulation inference-related research, and machine learning and deep learning methods. Overall, these methods perform well in cases of independent and identically distributed or weakly correlated change points with low noise. However, in nonlinear, strongly correlated, non-Gaussian noise, or chaotic systems, they are prone to false positives, false negatives, and insufficient interpretability.

[0020] (I) Statistical and Optimization Methods for Observation Space At the observation sequence level, typical methods include CUSUM, GLR, control charts, kernel methods, and cost function-based piecewise algorithms such as PELT (Pruned Exact Linear Time). These methods typically assume that the distribution within a segment is relatively stable or approximately stationary, and are sensitive to abrupt changes in low-order statistics such as mean and variance. However, under non-stationary, noisy, or complex dynamic conditions, natural fluctuations at the observation level may mask structural change signals, leading to false alarms and missed alarms; furthermore, change points often lack interpretability in terms of physical meaning.

[0021] (II) Parameter Space and Bayesian Online Detection Parameter space approaches treat transitions as jumps in generative model parameters. A representative method is Bayesian online transition detection, which achieves online detection capability by recursively estimating the run length or switching probability. There are also works based on state space or switching models, using Kalman / particle filtering to extract latent variables such as trend and variance, and then identifying structural breakpoints. These methods offer some interpretability when the model is well-defined, but typically rely on measurable likelihoods and strong priors. In high-dimensional, multimodal, or complex systems where likelihoods are not readily available, modeling and computational costs are high, and a trade-off exists between real-time performance and robustness.

[0022] (III) Simulation Inference Related Research Simulation-Based Inference (SBI) is a class of inference frameworks suitable for scenarios that are "simulatable but whose likelihoods are difficult to explicitly express." It uses simulation-driven neural methods (such as neural posterior / ratio / likelihood estimation) to characterize parameter uncertainty. Existing research mainly focuses on parameter estimation and uncertainty assessment of complex models. General procedures and evaluation systems for directly using such posterior information for change point detection are still being explored. In particular, in long-term, multi-variable, and noisy scenarios, there is a lack of consistent approaches to balancing linearity, threshold stability, and cross-scenario generalization.

[0023] (iv) Machine Learning and Deep Learning Methods Data-driven methods learn change point discrimination rules from raw sequences through supervised, semi-supervised, or unsupervised approaches. Supervised models, such as CNNs, RNNs, and Transformers, perform well when annotations are sufficient or synthetic. In engineering practice, there are also workflows involving "unsupervised candidate mining + minimal annotation correction + offline training and online deployment" to reduce annotation costs. Multi-scale signal processing (such as wavelets) is also used for noise reduction and feature extraction to improve localization sensitivity. The challenges of these methods lie in annotation and data coverage dependencies, cross-domain transfer stability, and decision interpretability.

[0024] (v) Applicable Scenarios and Existing Problems In scenarios such as finance, climate, industry, and environmental monitoring, data often exhibit characteristics such as strong correlation, non-Gaussian noise, long trends, or chaos. Existing technologies still fall short in achieving both low false alarm rates and high detection rates, sensitivity to weak / gradual changes, and attributability of the causes of these changes. For complex systems where likelihoods are unavailable but simulations are possible, effectively utilizing simulation resources and uncertainty information to form robust detection workflows that can be deployed online remains a common challenge in the industry.

[0025] In summary, existing change point detection technologies often suffer from high false positive rates, high false negative rates, large detection delays, and difficulty in explaining the reasons for changes in nonlinear, strongly correlated, non-Gaussian noise, or chaotic systems. For complex models where likelihoods are unavailable but simulations are possible, online Bayesian methods based on explicit likelihood suffer from high modeling and computational costs, and are difficult to deploy on edge devices. While deep end-to-end detection offers accuracy advantages, it relies on large-scale annotation, lacks cross-domain generalization capabilities, and suffers from insufficient interpretability. Therefore, this application provides a parameter space change point detection scheme that enables real-time detection of changes, thereby reducing detection delays, outputting attribution information for changes, clarifying the reasons for changes, and improving the interpretability of changes.

[0026] Please see Figure 1 , Figure 1 This is a flowchart of the parameter space variable point detection method in this invention. For example... Figure 1 As shown, the parameter space variable point detection method described in this embodiment of the invention includes: Step S11: Determine the parameter priors and call the simulator to generate an offline simulated observation sequence containing multiple change point variations based on the parameter priors.

[0027] Understandably, in the offline phase, parameter priors are obtained by sampling parameters from the prior distribution, and a simulator is invoked to generate an offline simulated observation sequence containing various change point variation forms. Specifically, parameter priors are obtained by sampling parameters from the prior range of a predefined parameter vector, and a pre-constructed set of change point variation forms is acquired; a simulator that can be simulated but whose likelihood function cannot be explicitly obtained is invoked to generate time series data based on the parameter priors; observation noise, missing segments, sampling rate changes, and domain randomization perturbations are added to the time series data according to the set of change point variation forms to obtain an offline simulated observation sequence containing various change point variation forms.

[0028] For example, calling a simulator that can be simulated but whose likelihood function cannot be explicitly obtained. The time series data is generated, and the set of change point patterns includes the expected change point patterns, such as step changes, step changes, impulse changes, slow drift, compound multiple change points, and single change points. According to the change point patterns in the set of change point patterns, observation noise, missing segments, sampling rate changes, and domain randomization perturbations can be added to the time series data to obtain a data sequence covering multiple change patterns and noise perturbations. This data can be used for subsequent model training and model validation to improve the robustness and generalization ability of the subsequent model.

[0029] It should be noted that when the emulator When there are missing or high costs, a differentiable proxy model can be trained first. approximate Based on this, SBI training and online detection can be completed, or a generative model (such as a diffusion / flow model) can be fitted based on historical data as an implicit simulator to maintain the sampling capability of data-parameter pairs. Hierarchical or Latin hypercube sampling, active simulation (sampling based on uncertainty or error), or sequential proposal distribution can also be used to improve sample efficiency and parameter space coverage.

[0030] Step S12: Construct training sample pairs based on the offline simulated observation sequence, and use the training sample pairs to train a window-level neural posterior estimator to obtain a parameter inference model.

[0031] The training of the window-level neural posterior estimator is conducted offline, which can be executed on CPU / GPU clusters. Furthermore, a lightweight posterior estimator can be obtained through model distillation and quantization to adapt to embedded devices. Since offline training relies on controllable simulation data, and combines domain randomization with posterior fine-tuning / feature alignment using a small amount of real data, it maintains stable performance under domain shifts such as equipment replacement, changes in operating conditions, or sampling rate drift. This significantly reduces the reliance on extensive manual annotation and enables cross-scenario transferability.

[0032] Specifically, the offline simulated observation sequence is processed using a sliding window method according to a preset window length and a first preset step size, and the temporal features of each window are extracted to construct training sample pairs that match windows and parameters. It is understood that the window length w, sliding step size s, and maximum allowable detection delay are preset. With equal constraints, training sample pairs are constructed according to the preset window length w and the first preset step size s (i.e., the sliding step size). .

[0033] For example, for an observation window of length w Extracting temporal features ,in, For a sliding window of length w, As a window-level feature extractor, temporal feature extraction can employ structures such as one-dimensional convolutional networks, gated recurrent units, or Transformers. A large number of "window-parameter" paired samples are used as training samples to train a density-modeled head approximating the posterior. This is a window-level neural posterior estimator, which obtains an amortized posterior approximation of the parameters, i.e., a parameter inference model. The window-level neural posterior estimator can be implemented using NRE / NLE (Neural Ratio / Likelihood Estimation), normalized flow, hybrid density networks or neural ratio / likelihood estimation models, neural spline flow, variational autoregressive flow, and density estimation based on diffusion models, etc. Furthermore, when only point estimation is required, discriminative regression can be used in conjunction with uncertainty estimation, such as deep ensembles, McLeod Dropout, and quantile regression, to approximate the posterior interval. Training can be performed using maximum likelihood or contrastive objectives, or sequential SBI, such as SNPE, can be used to focus on hard-sample regions.

[0034] Step S13: Using the amortized fixed-complexity forward inference method, perform sliding window inference on the online actual observation data and input it into the parameter inference model to calculate the parameter posterior distribution of each window, thereby obtaining the parameter posterior sequence and its statistics that evolve over time.

[0035] In this embodiment, a sliding window inference is performed on the online actual observation data to obtain the posterior sequence of parameters and its statistics that evolve over time. Specifically, a forward inference method with amortized fixed complexity is used, and a sliding window inference is performed on the online actual observation data according to a second preset step size. The second preset step size can be the same as the first preset step size mentioned above.

[0036] For example, for online actual observation data (such as real-time or streaming data), a sliding window inference is performed with a step size s and input into the parameter inference model, where the window is... To calculate each window by inferring the model through parameters posterior distribution of parameters ,in, The posterior parameters are output by the neural posterior estimator, and statistics, including the posterior mean vector, are extracted from this posterior distribution. Covariance matrix Posterior entropy This is used to form parametric trajectories and uncertain trajectories. To reduce jitter, the posterior mean vector can be... Perform exponential smoothing or median filtering. Since the online phase uses amortized fixed-complexity forward inference, combined with a short window and... Adaptive search aims to balance sensitivity and response speed, among which, The interval between adjacent windows is defined. Therefore, real-time detection can be achieved within a given delay budget, and the detection delay, such as the MAE index measured in steps, is lower than that of direct detection in the observation space or pure end-to-end detection.

[0037] Step S14: Based on the parametric posterior distribution sequence and its statistics, perform uncertainty-aware change point detection to obtain the corresponding change point set and the attribution information of each target change point in the change point set.

[0038] In this embodiment, since the discrimination is performed directly in the parameter space and the attribution information corresponding to the change point is output, it is possible to clearly identify why the change occurred, thus achieving explainability and attributability. This facilitates industrial fault location and scientific mechanism analysis, and improves the auditability of decisions.

[0039] Specifically, parameter trajectories and uncertainty trajectories are generated based on the parameter posterior sequence and its statistics; uncertainty-aware change point statistics are calculated based on the parameter trajectories and uncertainty trajectories, and the current calibration threshold is determined; candidate change points are generated based on the change point statistics and the current calibration threshold; time persistence criterion operation and non-maximum suppression merging operation are performed on temporally adjacent or repeated candidate change points to obtain the corresponding change point set and the attribution information of each target change point in the change point set. It can be understood that change point statistics are calculated on the parameter trajectory or posterior distribution, and candidate change points are generated based on the calibration threshold. Time persistence and non-maximum suppression merging are performed on temporally adjacent or repeated candidate change points to output the final change point set and attribution information. In the online phase, the inference and change point statistics are amortized calculations, with a single-window complexity equivalent to that of a first-order network forward pass, making it suitable for real-time deployment on the edge.

[0040] For example, the posterior distribution of parameters is calculated for each window, and statistics are extracted from this posterior distribution to form the parameter trajectory and uncertainty trajectory that evolve over time. Based on the parameter trajectory and uncertainty trajectory, one or more types of change point statistics are calculated, and the current calibration threshold is selected according to the offline calibration distribution and a preset false alarm rate. In the online phase, rolling quantiles or hazard priors can be used for fine-tuning, and then combined with the current calibration threshold to generate change point candidates. Next, time persistence criterion operations and non-maximum suppression merging operations are performed on adjacent or repeated candidates to merge duplicate detections, obtaining the final change point determination result. For each finally confirmed change point, the attribution information of the change point is output, such as the time and location of the change point, the dimension k* of the dominant parameter change, and the estimated magnitude of the parameter change. The system provides information on confidence level, direction of change, and relative magnitude, facilitating diagnosis and decision-making. Furthermore, the detection and attribution results can be provided to higher-level alarm or control systems via standardized interfaces, thus offering interpretable information.

[0041] Among them, in the parameter trajectory The above can employ piecewise algorithms such as PELT and binary dynamic programming; it can use kernel two-sample criterion or energy distance to measure the difference in distribution between adjacent windows; it can use a Transformer-based change point detector; and it can also... Considered as observation, Treating it as noise / weight, construct a state-space model of the parameter space and perform Bayesian online detection or switching model inference under Kalman / particle filtering.

[0042] In this embodiment, determining the current calibration threshold may specifically include: determining the current calibration threshold based on the acquired threshold baseline information and a preset false alarm rate; wherein, the threshold baseline information is information used for threshold setting obtained by performing posterior consistency checks and calibrations on the trained neural posterior estimator. It can be understood that the threshold controls the false alarm rate based on the calibration distribution. The threshold is selected according to the offline calibration distribution and the preset false alarm rate, and fine-tuning can be performed in the online phase using rolling quantiles or hazard priors.

[0043] In this embodiment, when the change point statistics include covariance-normalized distance, posterior entropy / divergence transition, and confidence interval crossing test, candidate change points are generated based on the change point statistics and the current calibration threshold. Specifically, this can include: merging or weighting at least two types of change point statistics from covariance-normalized distance, posterior entropy / divergence transition, and confidence interval crossing test using a multi-change point statistics fusion strategy to obtain processed change point statistics; and generating candidate change points based on the processed change point statistics and the current calibration threshold. For example, weighting or merging / intersecting at least two types of statistics from distance, divergence, and interval crossing, and combining them with time persistence criteria and non-maximum suppression merging to complete the decision, reduces the false alarm rate. By introducing statistics such as covariance normalized distance, posterior entropy / divergence transition and confidence interval crossing test into the parameter space, and by weighting the mean difference with posterior covariance, fluctuations in high uncertainty dimensions are suppressed. Therefore, false alarms are significantly reduced under strong noise, occlusion and non-stationary disturbances. The false alarm rate is more controllable under a given threshold strategy, and low false alarm and high robustness can be achieved.

[0044] Among them, (1) the covariance normalized distance can be the covariance normalized Mahalanobis distance, that is:

[0045] in, An adaptive search is performed within the interval [s, δ], where ε is the numerically stable term, I is the identity matrix, and ε > 0. Select from the grid {s,2s,…,δ} that maximizes D(t,Δ). Candidate change points are generated by comparing them with a threshold.

[0046] That is, the difference between the posterior means at two time points is used as the change signal, and normalization is performed according to the sum of the covariances of the two to suppress false alarms caused by high uncertainty dimensions.

[0047] (2) The confidence interval crossing test indicates that if the overlap of the confidence intervals of a certain parameter dimension between adjacent windows is less than a threshold, it is marked as a candidate change point. For example, for the k-th dimension of the parameter, the posterior quantile interval is: ; The degree of overlap is: ; when If a point is found to be a candidate variable, it is marked as such.

[0048] (3) Posterior entropy / divergence transition (i.e., information gain or divergence transition of the posterior distribution of adjacent windows), the relevant calculation formula is as follows: or Significant sudden increases are considered signs of change, and significant transitions are marked as candidate change points.

[0049] Furthermore, sequential detection on the parameter trajectory can be used to apply at least one of the following statistics: CUSUM, GLR, BOCPD, or PELT. Multiple statistics can be combined with union / intersection or weighted fusion, and combined with minimum interval suppression and persistence criteria to reduce false alarms.

[0050] It should be noted that, due to the adoption of threshold setting driven by posterior calibration and multi-statistic fusion decision-making (union / intersection or weighting), under a unified evaluation protocol, such as F1, detection delay MAE, and false alarms per thousand steps FP / 1000, higher detection rates and lower positioning errors are usually obtained under the same false alarm constraints.

[0051] As can be seen, in this embodiment of the invention, offline amortized inference combined with online uncertainty-aware detection is used to perform change point detection in the parameter space of a nonlinear system with unobtainable likelihood. That is, in the offline stage, simulation inference is used to obtain the sliding window-level posterior distribution of parameters, and in the online stage, uncertainty-aware change point detection is performed on the posterior sequence of parameters. Since the online stage uses a forward inference method with fixed complexity after amortization, real-time detection of change points can be achieved, thereby reducing the detection latency. Furthermore, change point detection is performed directly in the parameter space, and the attribution information of change points is output, which can clarify why the change points change and improve the interpretability of change points.

[0052] The technical solution of this application can employ multi-scale parallel windows or pyramid-shaped adaptive windows to simultaneously capture short-term abrupt changes and long-term gradual changes. It can also perform adaptive search on Δ or introduce Bayesian updates based on runtime. Combined with online smoothing / detrending, such as exponential smoothing or local regression, it enhances detection capabilities in gradual change scenarios. Posterior fine-tuning or feature alignment is performed on a small amount of real data, such as alignment based on MMD or CORAL. Simulation-to-real transfer and domain randomization / adversarial perturbation enhancement are employed. Semi-supervised consistency training using weak labels / pseudo-labels improves cross-scenario generalization. It simultaneously supports multi-step changes, step changes, impulse changes, slow drifts, composite multi-variable points, and single-variable points, and can parallelize multi-scale windows and various statistics as needed. Therefore, it possesses good scalability and robustness in complex scenarios and can be smoothly integrated with existing CPD components. Training data can be constructed through domain randomization and posterior fine-tuning or feature alignment can be performed on a small amount of real data to achieve domain adaptation.

[0053] See Figure 2 As shown, this embodiment of the invention discloses a specific method for detecting variable points in parameter space. Compared with the previous embodiment, this embodiment further explains and optimizes the technical solution.

[0054] Step S21: Determine the parameter priors and call the simulator to generate an offline simulated observation sequence containing multiple change point variations based on the parameter priors.

[0055] Step S22: Construct training sample pairs based on the offline simulated observation sequence, and use the training sample pairs to train a window-level neural posterior estimator to obtain a parameter inference model.

[0056] Step S23: Perform a posterior consistency check and calibration on the parameter inference model to obtain calibration results and threshold baseline information for threshold setting in the online detection stage.

[0057] Understandably, the trained window-level neural posterior estimator (i.e., the parameter inference model) undergoes posterior consistency checks and calibration to match the posterior uncertainty output by the parameter inference model with the true coverage. The calibration results and threshold baseline information are recorded, and this baseline information is used for threshold setting in subsequent online detection phases. For example, simulation-based calibration (SBC), posterior predictive check (PPC), temperature scaling, coverage assessment, or quantile calibration can be employed. If simulation-based consistency calibration and coverage assessment are performed on the validation set, and overconfidence or underconfidence occurs, temperature scaling or quantile calibration is used to align the posterior uncertainty with the true coverage. The empirical distribution of key statistics obtained from calibration is recorded for adaptive threshold setting in the online phase; that is, the online threshold is set based on the quantile rules obtained from offline calibration to meet the preset false alarm rate constraint. Because the threshold-false alarm rate correspondence is established through simulation-based consistency calibration, posterior predictive check, and quantile threshold, the threshold configuration has interpretable risk control significance, facilitating deployment and maintenance in industrial and compliance scenarios.

[0058] Step S24: Using the amortized fixed-complexity forward inference method, perform sliding window inference on the online actual observation data and input it into the parameter inference model to calculate the parameter posterior distribution of each window, thereby obtaining the parameter posterior sequence and its statistics that evolve over time.

[0059] Step S25: Based on the parametric posterior distribution sequence and its statistics, perform uncertainty-aware change point detection to obtain the corresponding change point set and the attribution information of each target change point in the change point set.

[0060] For details regarding steps S21 to S22 and steps S24 to S25, please refer to the corresponding content disclosed in the foregoing embodiments, which will not be repeated here.

[0061] As can be seen, in this embodiment of the invention, the posterior distribution of parameters at the sliding window level is obtained by simulation inference, and uncertainty-aware change point detection is performed on the parameter posterior sequence in the online stage. Since the fixed-complexity forward inference method after amortization is used in the online stage, the change point can be detected in real time, thereby reducing the detection latency. Furthermore, the change point is detected directly in the parameter space, and the attribution information of the change point is output, which can clarify why the change point changes and improve the interpretability of the change point.

[0062] like Figure 3 As shown, the left side represents the offline stage, where prior distribution sampling is used to generate offline simulated observation sequences containing various change point variations by calling the simulator. The right side represents the online stage, where the actual online observation data is processed by a sliding window and input into the posterior estimator to obtain the posterior parameters that evolve over time. Then, the change point detection module completes the uncertainty-aware change point determination and attribution output within the parameter space. Specifically, prior distribution sampling parameters are used to obtain parameter priors, and a simulator is called to generate offline simulated observation sequences containing various change point forms. Training sample pairs are constructed according to window length and step size. Based on the training sample pairs, a window-level neural posterior estimator is trained and calibrated. Sliding window inference is performed on online actual observation data, and the data is input into the parameter inference model to calculate the parameter posterior distribution for each window, obtaining the parameter posterior sequence and its statistics (posterior mean vector, covariance matrix, posterior entropy) that evolve over time. Parameter trajectories and uncertainty trajectories are generated based on the parameter posterior sequence and its statistics. Change point statistics for uncertainty perception are calculated based on the parameter trajectory and uncertainty trajectory, and candidate change points are generated according to the calibration threshold. Subsequently, time persistence and non-maximum suppression are performed on temporally adjacent or repeated candidate change points to obtain the final set of change points and attribution information. The technical solution of this application can be applied to scenarios such as nonlinear dynamic system parameter identification and online monitoring, industrial equipment fault early warning, financial risk control and anomaly detection.

[0063] For example, the Lorenz-63 system is selected as a representative nonlinear dynamical system that can be simulated but whose likelihood cannot be explicitly obtained, with parameter vector θ=(σ,ρ,β). The prior range is set as follows: σ∈[8,12], ρ∈[24,32], β∈[2,3]. First, in the training data generation stage, the integration method can be the fourth-order Runge-Kutta; the time step dt can be 0.01 (adjustable according to numerical stability), and the change point transformation can be of three types: single change point, multi-change point step, and slow drift, with at least 100 simulated observation long sequences generated for each type. The window length w=100, and the step size s=1; sample pairs are formed. , This represents the representative value of the parameter within the window, such as taking the middle value within the window for slow drift. Zero-mean Gaussian noise is injected into the sequence (standard deviation is jittered at 1%–5% of the original dimension), the random missing rate is 0%–10%, the amplitude scaling factor is 0.8–1.2, and the sampling rate perturbation is ±5%.

[0064] During the model training phase, the window-level feature extractor is determined. Its structure can be a one-dimensional convolution (such as 3×1 convolution + 2 layers of ReLU + 2 layers of pooling) followed by a GRU (Gated Recurrent Unit) (with a hidden dimension of 64). The density modeling head approximates the posterior. It can be a neural spline flow with 5 layers, 128 hidden dimensions, and 8 spline segments. Optimization settings during training include a batch size of 256 and an initial learning rate of 1e. 3. Cosine annealing 1e 5. Training epochs: 100. Early stopping is based on validation log-likelihood. The final training yields the amortized [result]. That is, the parameter inference model.

[0065] In the posterior calibration phase, the trained window-level neural posterior estimator (i.e., the parameter inference model) undergoes posterior consistency checks and calibration to ensure that the posterior uncertainty output by the parameter inference model matches the true coverage. Calibration results and threshold baseline information are recorded, which is used for threshold setting in subsequent online detection phases, such as performing consistency calibration and coverage evaluation on an independent validation set. If the coverage deviates from the target, temperature scaling is used for calibration. The empirical distribution of the covariance normalized distance and posterior entropy / divergence transitions is recorded to set the online threshold.

[0066] During the online detection phase, the actual observed long sequence is slidably processed with s=1, and the posterior distribution of parameters for each window x(t) is calculated. The statistics are extracted from the posterior distribution, including the posterior mean vector μ(t), covariance matrix Σ(t), and posterior entropy H(t), thus forming a parametric trajectory and an uncertainty trajectory that evolve over time. Based on the parametric trajectory and the uncertainty trajectory, one or more types of change point statistics are calculated, where D(t) adopts Δ∈[1,δ], and Δ adaptively searches for δ=10 within the interval [1,δ]. The posterior entropy / divergence transition is calculated using the following formula: and The confidence interval crossing test is set with a 95% confidence interval overlap threshold, and a quantile threshold is selected based on a preset false alarm rate on the validation set, such as FP / 1000=0.5, and fine-tuned online using rolling quantiles. Quantile thresholds for D(t) and KL (Kullback–Leibler divergence) are selected; fine-tuned online using rolling quantiles. A minimum interval m=δ is set, and non-maximum suppression is applied to candidate change points based on their statistical scores; candidate change points with a duration of less than two windows are discarded.

[0067] Finally, for each final change point, the system outputs the time index t*, the parameter dimension of the dominant change, the estimated parameter change magnitude Δμ, and its 95% confidence interval; and outputs a JSON structure, such as {time,param,delta,ci,score}, to the upper-level system. Under the same false alarm constraint (e.g., FP / 1000=0.5), compared to direct detection in the observation space or pure end-to-end detection, parameter space detection improves the F1 score and reduces the localization error and detection latency (MAE). Furthermore, the detection statistics can be applied to PELT, binary dynamic programming, MMD, energy distance, or to construct a state-space parallel BOCPD (Bayesian Online Change Point Detection) in the parameter space.

[0068] In one embodiment, such as Figure 4 As shown, based on the above-described parameter space change point detection method, the present invention also provides a parameter space change point detection device, comprising: The simulated observation sequence generation module 11 is used to determine the parameter prior and call the simulator to generate an offline simulated observation sequence containing multiple variable point change forms based on the parameter prior. Training module 12 is used to construct training sample pairs based on the offline simulated observation sequence and use the training sample pairs to train a window-level neural posterior estimator to obtain a parameter inference model. Online inference module 13 is used to perform sliding window inference on online actual observation data using a forward inference method with amortized fixed complexity and input it into the parameter inference model to calculate the parameter posterior distribution of each window, so as to obtain the parameter posterior sequence and its statistics that evolve over time; The change point detection module 14 is used to detect change points based on the parameter posterior distribution sequence and its statistics to perceive uncertainty, and to obtain the corresponding change point set and the attribution information of each target change point in the change point set.

[0069] The modules can be merged or split according to resource conditions, supporting single-machine deployment, edge-cloud collaboration, or microservice architecture. The inference engine can run on CPU, GPU, or NPU (Neural Processing Unit). Since the computational load in the online phase is equivalent to only one forward inference and constant-order statistical calculation, it is suitable for real-time deployment on edge or end-device devices. Furthermore, online deployment performs model distillation, INT8 (a model quantization method that stores and computes models with 8-bit integer precision), operator fusion, and heterogeneous acceleration on the posterior estimator to meet the real-time and energy consumption constraints of edge devices and reduce computational and storage overhead. Therefore, it can be stably deployed even in resource-constrained environments on the edge / end-device side, with controllable computational and storage overhead.

[0070] Specifically, the parameter spatial change point detection device may include the following functional units: a data acquisition interface for receiving time series data and performing synchronization and buffering; a simulation and data generation unit for calling the simulator and constructing the dataset; and a feature extraction and posterior estimation unit for running the window-level feature extractor. With window-level neural posterior estimator The system comprises the following components: a calibration and diagnostic unit for simulating consistency calibration, coverage assessment, and threshold baseline maintenance; an online detection unit for calculating statistics such as covariance normalized distance, KL, and posterior entropy / divergence transitions, and generating candidate change points; a result output and alarm unit for attribution information of change points, such as the time and location of the change, the dimension of the dominant parameter, the estimated magnitude of parameter change, confidence level, direction of change, and relative magnitude; a storage and model management unit for managing models, thresholds, logs, and auditing; and a computation and acceleration unit running on CPU / GPU / NPU, with optional quantization and distillation capabilities. The functional units in this parameter space change point detection device communicate via a system bus or high-speed interconnect, and can be integrated within the same device or deployed collaboratively in a cloud-edge environment.

[0071] Furthermore, it is worth noting that the working process of the parameter space change point detection device provided in this embodiment is the same as that of the parameter space change point detection method described above, and will not be repeated here. For details, please refer to the working process of the parameter space change point detection method described above.

[0072] Figure 5 A schematic diagram of the structure of a terminal provided in an embodiment of this application. The terminal may include: The memory 501, the processor 502, and the computer program stored on the memory 501 and capable of running on the processor 502.

[0073] When the processor 502 executes the program, it implements the parameter space change point detection method provided in the above embodiments.

[0074] Furthermore, the terminal also includes: Communication interface 503 is used for communication between memory 501 and processor 502.

[0075] The memory 501 is used to store computer programs that can run on the processor 502.

[0076] Memory 501 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0077] If the memory 501, processor 502, and communication interface 503 are implemented independently, they can be interconnected via a bus to communicate with each other. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of representation, only one line is used in the diagram, but this does not imply that there is only one bus or one type of bus.

[0078] Optionally, in a specific implementation, if the memory 501, processor 502, and communication interface 503 are integrated on a single chip, then the memory 501, processor 502, and communication interface 503 can communicate with each other through an internal interface.

[0079] Processor 502 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of this application.

[0080] This embodiment also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described parameter space change point detection method.

[0081] Other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein.

[0082] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0083] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus or device (such as a computer-based system, a processor-included system or other system that can read and execute instructions from and from an instruction execution system, apparatus or device).

[0084] It should be understood that the various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, the N steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. If implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0085] It should be understood that the application of the present invention is not limited to the examples above. Those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A method for detecting variable points in parameter space, characterized in that, The method includes: Determine the parameter priors and call the simulator to generate an offline simulated observation sequence containing multiple variable point change forms based on the parameter priors; Training sample pairs are constructed based on the offline simulated observation sequences, and the training sample pairs are used to train a window-level neural posterior estimator to obtain a parameter inference model. Using a fixed-complexity forward inference method after amortization, sliding window inference is performed on the online actual observation data and input into the parameter inference model to calculate the parameter posterior distribution of each window, so as to obtain the parameter posterior sequence and its statistics that evolve over time; Uncertainty-aware change point detection is performed based on the parametric posterior distribution sequence and its statistics to obtain the corresponding change point set and the attribution information of each target change point in the change point set.

2. The parameter spatial change point detection method according to claim 1, characterized in that, The process of determining prior parameters and invoking a simulator to generate an offline simulated observation sequence containing various change point variations based on the prior parameters includes: The parameters are sampled from the prior range of the predefined parameter vector to obtain the parameter prior, and a pre-constructed set of variable point transformation forms is obtained. A simulator that can be simulated but whose likelihood function cannot be explicitly obtained is invoked to generate time series data based on the prior parameters of the simulator; According to the set of variable point change forms, observation noise, missing segments, sampling rate changes, and domain randomization perturbations are added to the time series data to obtain an offline simulated observation sequence containing multiple variable point change forms.

3. The parameter spatial change point detection method according to claim 1, characterized in that, The construction of training sample pairs based on the offline simulated observation sequence includes: The offline simulated observation sequence is processed by sliding window according to the preset window length and the first preset step size, and the temporal features of each window are extracted to construct training sample pairs that pair windows with parameters.

4. The parameter space change point detection method according to claim 1, characterized in that, The method of using amortized fixed-complexity forward inference to perform sliding window inference on online actual observation data includes: The forward inference method with amortized fixed complexity is used, and sliding window inference is performed on the online actual observation data according to the second preset step size.

5. The parameter space change point detection method according to any one of claims 1 to 4, characterized in that, The uncertainty-aware change point detection based on the parametric posterior distribution sequence and its statistics yields a corresponding set of change points and attribution information for each target change point in the set, including: Generate parameter trajectories and uncertainty trajectories based on the posterior sequence of the parameters and their statistics; Based on the parameter trajectory and the uncertainty trajectory, calculate the change point statistics of uncertainty perception and determine the current calibration threshold; Based on the change point statistics, and according to the current calibration threshold, candidate change points are generated; Perform time persistence criterion operation and nonmaximum suppression merging operation on the candidate change points that are temporally adjacent or repetitive to obtain the corresponding change point set and the attribution information of each target change point in the change point set.

6. The parameter spatial change point detection method according to claim 5, characterized in that, Determining the current calibration threshold includes: Based on the obtained threshold baseline information, the current calibration threshold is determined according to the preset false alarm rate; wherein, the threshold baseline information is information used for threshold setting obtained by performing posterior consistency checks and calibrations on the trained neural posterior estimator.

7. The parameter spatial change point detection method according to claim 5, characterized in that, The change point statistics include covariance normalized distance, posterior entropy / divergence transition, and confidence interval crossing test. The step of generating candidate change points based on the change point statistics and according to the current calibration threshold includes: Based on the multi-point statistic fusion strategy, at least two of the point statistics in the covariance normalized distance, posterior entropy / divergence transition and confidence interval crossing test are merged or weighted to obtain the processed point statistics. Based on the processed change point statistics, candidate change points are generated according to the current calibration threshold.

8. A parameter spatial change point detection device, characterized in that, The device includes: The simulated observation sequence generation module is used to determine the parameter priors and call the simulator to generate an offline simulated observation sequence containing multiple variable point change forms based on the parameter priors. The training module is used to construct training sample pairs based on the offline simulated observation sequence and use the training sample pairs to train a window-level neural posterior estimator to obtain a parameter inference model. The online inference module is used to perform sliding window inference on the online actual observation data using a forward inference method with amortized fixed complexity and input it into the parameter inference model to calculate the parameter posterior distribution of each window, so as to obtain the parameter posterior sequence and its statistics that evolve over time. The change point detection module is used to detect change points based on the parameter posterior distribution sequence and its statistics to perceive uncertainty, and to obtain the corresponding change point set and the attribution information of each target change point in the change point set.

9. A terminal, characterized in that, include: The device includes a memory, a processor, and a parameter space change point detection program stored in the memory and executable on the processor, wherein the parameter space change point detection program, when executed by the processor, implements the steps of the parameter space change point detection method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that can be executed to implement the steps of the parameter space change point detection method as described in any one of claims 1 to 7.

Images