A time series anomaly detection method, electronic device and medium

Abstract

The application discloses a time series anomaly detection method, an electronic device and a medium, and comprises the following steps: S1, based on the signal-to-noise ratio parameterization noise injection and removal process; S2, receiving input time series data and based on the signal-to-noise ratio characteristics of the data, adaptively estimating the damage factor and the de-noising step; S3, based on the damage factor and the de-noising step, partially damaging the input time series data, and starting from the partially damaged data to execute the reverse de-noising process to generate the reconstructed time series; S4, calculating the deviation between the input time series data and the reconstructed time series, and outputting the anomaly detection result based on the deviation, which enables the model to start from the partially damaged state matching the input data characteristics to determine the stronger reconstruction, effectively inhibiting the randomness deviation in the traditional diffusion model generation process, and significantly improving the precision and recall rate of anomaly detection.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence and data detection technology, in particular to a time series data anomaly detection method, and specifically to a time series anomaly detection method, an electronic device and a computer readable storage medium. BACKGROUND

[0002] Time series anomaly detection is a method for detecting abnormal signals generated by machines or devices during operation. Its core is to model the normal signals of the device, and when the device has abnormal changes, the system can identify and mark it as an abnormal event. This method is widely used in industrial manufacturing, energy equipment, transportation and robot automation fields, aiming to improve the reliability and safety of the device.

[0003] In recent years, with the development of deep learning technology, deep neural networks have been gradually applied to time series anomaly detection. Thanks to the powerful modeling capability of diffusion models, diffusion model-based anomaly detection methods for time series reconstruction have gradually attracted attention. The idea of diffusion model method is to model the normal time series features using diffusion model, and then to normalize the reconstruction of input sequence. Since diffusion model only learns the distribution of normal data, when the abnormal segment of the input sequence is reconstructed, its reconstruction error will be greater than that of the normal segment, thereby achieving anomaly detection.

[0004] Currently, there are several patents of diffusion model methods disclosed, for example, patent 202510324397.8 discloses an anomaly detection method for fixed pollution source monitoring data, which can effectively improve the accuracy and efficiency of anomaly detection in fixed pollution source monitoring data. Patent 202510296913.0 discloses a multi-dimensional time series anomaly detection method, device, equipment and medium, which can improve the accuracy and scope of multi-dimensional time series anomaly detection, and enhance the robustness of the detection process. However, the current diffusion model method has a serious problem, namely the uncertainty reconstruction of diffusion model. As a generative model, the generation result of diffusion model has strong randomness, while the anomaly detection task requires the model to perform deterministic reconstruction on the input sequence. The existing method (such as the above application) fails to solve this problem, resulting in matching errors between the reconstructed sequence and the original sequence, so that the reconstruction error can no longer reliably represent the degree of abnormality, ultimately seriously affecting the anomaly detection performance of the model, causing misjudgment and omission of abnormal points. SUMMARY

[0005] The present application aims to solve the problem of reconstruction uncertainty caused by the randomness of the generation process in the existing time series anomaly detection method, and proposes a time series anomaly detection method, an electronic device and a computer readable storage medium.

[0006] The technical problem of the present application is solved by the following technical solutions:

[0007] In a first aspect, the present application provides a time series anomaly detection method based on an implicit conditional diffusion model, comprising the following steps:

[0008] S1, receiving input time series data;

[0009] S2, adaptively estimating a destruction factor and a denoising step based on the characteristics of the input time series data;

[0010] S3, based on the destruction factor and the denoising step, partially destroying the input time series data, and starting from the partially destroyed data to perform a reverse denoising process to generate a reconstructed time series;

[0011] S4, calculating the deviation between the input time series data and the reconstructed time series, and outputting an anomaly detection result based on the deviation.

[0012] In some embodiments, the following technical features are further included:

[0013] The adaptive estimation of step S2 includes the following steps:

[0014] The injection and removal process of the noise parameterized based on the signal-to-noise ratio defines the noise level of each step in the diffusion process through the target signal-to-noise ratio and the energy recursive relationship, to ensure the consistency of information preservation of noise scheduling in the training and inference process.

[0015] In some embodiments, the energy recursive relationship ensures that the total energy dissipated in all diffusion steps is equal to the predefined terminal signal energy.

[0016] In some embodiments, the estimation method of the destruction factor in step S2 includes:

[0017] The input time series data is band-decomposed to obtain its low-frequency part and high-frequency residual part, and the inference signal-to-noise ratio is calculated based on the energy of the low-frequency part and the high-frequency residual part, and then the destruction factor is derived from the inference signal-to-noise ratio according to the statistical characteristics of the input time series data.

[0018] In some embodiments, the band decomposition is performed by a zero-phase Gaussian low-pass filter.

[0019] In some embodiments, the determination method of the denoising step in step S3 includes:

[0020] The estimated destruction factor is mapped to the nearest feasible step in the predefined diffusion time step sequence.

[0021] In some embodiments, the operation of partially corrupting the input time series data in step S3 is expressed as: wherein denotes the implicit input of the test instance in the reverse process, is the test instance, is the corruption factor, denotes the data at the denoising step closest to the optimal corruption factor, is the Gaussian noise.

[0022] In some embodiments, in step S4, the bias is the L2 norm between the input time series data and the reconstructed time series.

[0023] In a second aspect, the present application provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the above method when executing the program.

[0024] In a third aspect, the present application provides a computer readable storage medium, having stored thereon a computer program, which program, when executed by a processor, implements the steps of the above method.

[0025] The beneficial effects of the present application compared with the prior art include:

[0026] The application adaptively estimates the damage factor and the denoising step based on the characteristics of the input time series data, thereby constructing an implicit conditional mechanism highly matched with the data characteristics, for guiding the reverse denoising process of the diffusion model. The core innovation of the mechanism lies in: first, the damage factor is derived according to the statistical characteristics of the input data, ensuring that the noise damage process is dynamically aligned with the inherent noise scale of the data; second, the input sequence is partially damaged based on the damage factor, so that the model starts to perform deterministic reverse reconstruction from a partially damaged state (rather than completely random noise) that is adapted to the characteristics of the input data. This "adaptive estimation + partial damage start" cooperative mechanism effectively solves the reconstruction uncertainty problem caused by the completely random initialization of traditional diffusion models. Without this adaptive estimation and the cooperative mechanism of starting from a partially damaged state, relying solely on fixed parameterization or other single features (such as pre-defined noise scheduling) will not effectively cope with the diversity of noise scales and trend patterns in different time series data. Fixed parameters are difficult to adapt to the signal-to-noise ratio characteristics of different data, and are prone to cause over-damage or insufficient damage, thereby introducing significant bias in the reconstruction process. In particular, the traditional method starts reconstruction from pure random noise, and the randomness can mask the normal patterns of the input data and amplify the reconstruction uncertainty of abnormal components, leading to increased abnormal detection error. Through the above cooperative mechanism, the model can better preserve the normal trend patterns in the input data while attenuating abnormal high-frequency components during the reconstruction process. Finally, the deviation between the input time series and the reconstructed sequence can more reliably and accurately represent the degree of abnormality, significantly improving the precision and recall rate of abnormal detection.

[0027] In addition, in some embodiments, the following beneficial effects are also achieved:

[0028] The noise level is defined by the target signal-to-noise ratio and the energy recursive relationship, ensuring the consistency of information preservation of the noise scheduling in the training and inference processes, and enhancing the robustness of the model under different noise levels.

[0029] The damage factor is estimated by frequency band decomposition and obtaining low-frequency and high-frequency residual parts, so that the noise damage strategy can more specifically focus on the high-frequency components that may contain abnormalities, improving the sensitivity of abnormal detection.

[0030] The damage factor is mapped to a pre-defined diffusion time step sequence, ensuring that the reverse denoising process can be performed on the feasible trajectory trained by the model, improving the stability and reliability of the reconstruction process.

[0031] Other beneficial effects in the embodiments of the application will be further described below. BRIEF DESCRIPTION OF DRAWINGS

[0032] Figure 1is a flow chart of a time series anomaly detection method based on an implicit conditional diffusion model according to an embodiment of the present application.

[0033] Figure 2 is a network overall structure diagram of a time series anomaly detection method based on an implicit conditional diffusion model according to an embodiment of the present application. DETAILED DESCRIPTION

[0034] The application will be further described below with reference to the drawings and in conjunction with preferred embodiments. It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict.

[0035] It should be noted that the left, right, up, down, top, bottom and other orientation terms in the present embodiment are only relative concepts or are referenced to the normal use state of the product, and should not be considered as limiting.

[0036] Abbreviations and key terms definitions:

[0037] Time series anomaly detection: A method for detecting abnormal signals generated by machines or devices during operation. The normal signals of the device are modeled, and when the device has an abnormal change, the system can identify and mark it as an abnormal event. This method is widely used in industrial manufacturing, energy equipment, transportation and robot automation, etc., aiming to improve the reliability and safety of the device.

[0038] Denoising diffusion probabilistic model (DDPM): A generative model based on a step-by-step denoising process. In DDPM, data is destroyed through a series of processes that gradually add noise, and the generative model recovers the data through an inverse process of step-by-step denoising. This method performs well in generating images and time series, but its computational complexity is high.

[0039] Specifically, let the input data distribution be The forward process gradually disturbs by injecting noise scaled by , where The joint distribution of the diffusion model is

[0040]

[0041] where, is the noisy t-step data, which can be represented by recursively applying the reparameterization trick:

[0042]

[0043] where

[0044] Signal-to-Noise Ratio (SNR): A measure of the strength of a signal relative to the strength of noise, widely used in communication, image processing, audio engineering, machine learning, etc. to evaluate the quality of the signal.

[0045] Noise Scheduler: A mechanism for controlling the way DDPM introduces or removes noise during the forward process and reverse process. It directly determines the degree of adding / removing noise at each step of the model, and has a great impact on the performance, stability and learning ability of the model.

[0046] The idea of this embodiment is introduced as follows:

[0047] The existing time series anomaly detection method based on diffusion model has the problem of uncertainty in the reconstruction process, which is contrary to the starting point of using reconstruction error to represent the degree of anomaly, resulting in misjudgment and missed judgment of abnormal points, seriously affecting the anomaly detection performance of the model. To solve this problem, the embodiment of the present invention proposes an anomaly detection method based on implicit diffusion model. The SNR-based noise scheduler dynamically adjusts the noise level during the training process to enable the model to adaptively learn robust reconstruction under different damage intensities in complex and heterogeneous time series. Secondly, the SNR-based implicit conditional mechanism adaptively estimates the noise level and denoising step of each input sequence, implicitly guides the reverse process using the information of the input sequence itself, rather than explicit external labels, which is more suitable for time series anomaly detection tasks. This method solves the uncertainty reconstruction problem of diffusion model in time series anomaly detection, making the model have more reliable and excellent anomaly detection performance, providing a more reliable and economical solution for industrial equipment anomaly detection.

[0048] The core of the present invention is to solve the reconstruction uncertainty problem of diffusion model in time series anomaly detection. By reparameterizing the noise injection process through the SNR scheduler, and using the SNR implicit conditional mechanism to adaptively estimate the optimal damage factor and denoising step of each input sequence, the model is guided to start from partially damaged data for more deterministic reconstruction, and finally realizes accurate anomaly detection by calculating the reconstruction error.

[0049] The embodiment of the present invention proposes a time series anomaly detection method based on implicit conditional diffusion model, which overcomes the uncertainty reconstruction problem of diffusion model in anomaly detection task through SNR scheduler and SNR-based implicit conditional mechanism, and realizes more reliable and accurate anomaly detection. As shown in Figure 1 The method specifically includes the following steps:

[0050] S1. Receive input time series data;

[0051] S2. Based on the signal-to-noise ratio characteristics of the input time series data, adaptively estimate the destruction factor and denoising steps;

[0052] S3. Based on the destruction factor and the denoising step, the input time series data is partially destroyed, and the reverse denoising process is performed starting from the partially destroyed data to generate a reconstructed time series.

[0053] S4. Calculate the deviation between the input time series data and the reconstructed time series, and output the anomaly detection result based on the deviation.

[0054] The specific explanation is as follows:

[0055] I. Time Series Data Preprocessing

[0056] Step S1: Receive input time series data. To capture the dependencies between variables while maintaining temporal coherence, sensor data collected from industrial equipment or systems is cascaded through channels to generate multivariate time series data. (i.e., the original data time length is) The characteristic number is Represented as a 2D matrix , Represent a The space of real matrix numbers.

[0057]

[0058] in, Indicates a point in time Input data at the location, This indicates concatenation along the feature dimension. This indicates the original data at a specific time point. The first Dimensional features.

[0059] II. Overall Network Architecture

[0060] The overall network structure is as follows Figure 2 As shown, this method addresses two systemic limitations of traditional diffusion-based time series anomaly detection methods.

[0061] This invention proposes two innovations to jointly address these problems. These include the following steps:

[0062] Step S2, receiving input time series data and adaptively estimating a corruption factor and a denoising step based on signal-to-noise ratio characteristics of the data. SNR is a measurable and intuitive metric for reparameterizing the noise schedule. The SNR scheduler innovatively modulates the noise injection rate as a function of the signal-to-noise ratio, redefining the diffusion noise schedule to ensure optimal information preservation during training and inference.

[0063] The implicit conditioning mechanism estimates two key parameters: a corruption factor , which quantifies the optimal noise intensity for a given input; and a denoising step , which determines the optimal reverse denoising trajectory, where T denotes the total number of noise-adding steps.

[0064] Step S3, based on the corruption factor and the denoising step, partially corrupting the input time series data and performing a reverse denoising process starting from the partially corrupted data to generate a reconstructed time series. By generating from a controlled corrupted input rather than Gaussian noise, an implicit guided reconstruction is obtained to coordinate the stochastic generation of the diffusion model. Here denotes the probability of generating a certain value under the model weights , is a Markov process, denotes the input , and the output .

[0065] Step S4, calculating the deviation between the input time series data and the reconstructed time series, and outputting an anomaly detection result based on the deviation. The deviation is the L2 norm between the input time series data and the reconstructed time series. The final anomaly detection calculates the norm between the original time series and its implicitly guided reconstruction generated by the learned diffusion process , which measures is used as an anomaly score, with a larger value indicating a higher likelihood of temporal anomaly.

[0066] III. SNR Scheduler

[0067] In step S2, the adaptive estimation includes parameterizing the injection and removal of noise based on the signal-to-noise ratio, defining the noise level at each step in the diffusion process through a target signal-to-noise ratio and an energy recursion relationship to ensure consistency of information preservation in the noise schedule during training and inference. The energy recursion relationship ensures that the total energy dissipated in all diffusion steps is equal to the predefined terminal signal energy.

[0068] A conventional noise schedule is defined by , and Parameterization, where , These refer to the maximum and minimum values ​​of the noise timeline, which independently define different but related physical dynamics. For example, the larger... and smaller This leads to a more refined noise reduction process. However, larger... This may lead to longer reasoning time and smaller... This could lead to insufficient disruption of the diffusion process for training the model. Therefore, we reparameterized these two parameters using the signal-to-noise ratio (SNR), an intuitive, simple, and relevant metric for time-series data. Let... The target signal-to-noise ratio is

[0069]

[0070] in It is the terminal signal energy. The initial energy is represented by the expectation of the F(Frobenius) norm of the input data, where... This represents the expectation operation. express The F-norm. Doing so will ensure that the final signal-to-noise ratio of the noise scheduler is exactly 1. Then, we define ,in , Let represent the energy at step t. The recursive relation is:

[0071]

[0072] ( Through The definition obtained This recursive relation ensures that the total energy dissipated in all diffusion steps is equal to a predefined value. This ensures that the end-point noise level matches the target. To ensure monotonic energy decay... We introduced a time-corruption scheduler. (Linear scheduler) (etc.) Among them, This indicates that the domain is positive, and the overall noise trend is explicitly adjusted during the noise addition process.

[0073]

[0074] in, This represents the natural exponential function. This represents the value of the time-damaged scheduler at time t. (Through...) right The parametric implementation of the two key advances compared to the conventional approach. First, we derive the heuristic adjustment of the parameters by means of a physically interpretable estimate of the signal-to-noise ratio

[0075]

[0076] where denotes the cumulative signal preservation up to step

[0077] Four. SNR implicit conditioning mechanism

[0078] In step S2, the way of estimating the destruction factor comprises performing band decomposition on the input time series data to obtain a low frequency part and a high frequency residual part, and calculating an inferred signal-to-noise ratio based on energies of the low frequency part and the high frequency residual part, and deriving the destruction factor from the inferred signal-to-noise ratio according to statistical characteristics of the input time series data. The band decomposition is performed by a zero-phase Gaussian low-pass filter. In step S3, the way of determining the denoising step comprises mapping the estimated destruction factor to a nearest feasible step in a predefined sequence of diffusion time steps.

[0079] The SNR implicit conditioning mechanism operates under the premise that time series anomalies manifest as high-frequency deviations from the normative pattern. The SNR provides a physically interpretable measure of signal fidelity across frequency bands. By modulating the signal-to-noise ratio, low-frequency trends (dominant in normal patterns) are preserved, while high-frequency anomalies or noise are attenuated. Let denote a test instance with time steps and variables. We first decompose :

[0080]

[0081] where denotes the low-frequency part in , denotes the residual, is a zero-phase Gaussian low-pass filter with bandwidth . The residual captures the high-frequency content assumed to contain potential anomalies. From this decomposition, we derive the inferred signal-to-noise ratio , a case-specific measure that quantifies the optimal noise destruction of the implicit conditioning​​​​

[0082]

[0083] where and denote the F-norm of and respectively, is a constant, prevents division by zero in the noiseless edge case. The F-norm aggregates the energy of all K variables. The adaptive initialization of the backpropagation process is guided by two estimation stages: from the optimal corruption factor is estimated; and is aligned with the nearest diffusion step.

[0084] Stage 1: Optimal corruption factor estimation

[0085] According to the instantaneous formula, we exploit the statistical properties of to derive the instance optimal corruption factor

[0086]

[0087] where and denote the mean and variance of the elements in the test instance

[0088] Stage 2: Diffusion step alignment

[0089] Since may not exist in the pre-trained schedule , we project it to the nearest denoising step :

[0090]

[0091] Finally, the test instance is corrupted to , i.e., the operation formulation of partially corrupting the input time series data in step S3, where denotes the implicit input of the test instance in the backward process, denotes the noisy data at the nearest denoising step to the optimal corruption factor, the Gaussian noise strategyically reduces the unusually high frequency components while preserving the original data correlation. This controlled initialization enables the conditional backward process while ensuring the consistency of the training inference SNR and the abnormal decay.

[0092] ​As Figure 2 shown in FIG. 1, during inference, the input data is processed by the (b) SNR implicit conditioning mechanism module, which dynamically estimates (1) the optimal corruption factor and (2) the denoising time step , while this process is guided by the (a) SNR scheduler. These parameters drive the partially corrupted input denoising process, enabling implicit guided generation to reconstruct the target signal. The denoising UNet, which is a classic UNet structure (a specific neural network architecture) in diffusion models, is responsible for estimating the noise in the denoising process.

[0093] V. Implementation details

[0094] The Adam optimizer is used in the embodiments of the present application, with an initial learning rate of 1e-4. The sliding window size is fixed at 1024, the diffusion time step number is set to 200, the SNR scheduler parameter is set to -10, and the function g(t) is defined as a linear function. For the optimal corruption factor estimation in the SIC mechanism, we empirically set the Gaussian kernel size s and the Gaussian kernel standard deviation σ to 128 and 16 for each dataset.

[0095] VI. Experimental results

[0096] The embodiments of the present application are tested on five commonly used time series anomaly detection datasets, including Mars Science Laboratory rover (MSL), SoilMoisture Active Passive satellite (SMAP), Server Machine Dataset (SMD), Pooled Server Metrics (PSM) and Secure Water Treatment (SWaT). The embodiments of the present application are compared with various baseline models from four main categories, including traditional machine learning methods such as Isolation Forest (IF), Empirical Cumulative Distribution Function-based Unsupervised Outlier Detection (ECOD), Clustering-based Local Outlier Factor (CBLOF), Principal Component Analysis (PCA), representation-based methods (these models learn the representation of normal patterns, and identify anomalies through representation distance), such as Deep SVDD, Dual Attention Contrastive Representation Learning for Time Series Anomaly Detection (DCDetector), reconstruction-based methods (understand normal behavior by encoding sub-sequences of normal time series, and determine anomalies according to reconstruction error), such as Adversarial Time-Frequency Reconstruction Network for Unsupervised Anomaly Detection (ATF-UAD), Drift Doesn't Matter: Dynamic Decomposition for Unstable Multivariate Time Series Anomaly Detection Based on Diffusion Reconstruction (D3R), Anomaly Transformer: Time Series Anomaly Detection Based on Correlation Difference (AT), Multivariate Time Series Anomaly Detection Based on Generative Adversarial Network (FGAnomaly), TimesNet for General Time Series Analysis, Spatial Association Perception Anomaly Detection and Diagnosis for Multivariate Time Series (SARAD), Diffusion Model-based Methods (generate normal time series corresponding to input sequences using diffusion model, and identify anomalies through reconstruction error), such as ImDiffusion for Multivariate Time Series Anomaly Detection, Time Series Anomaly Detection Based on Interpolation Conditional Weighted Incremental Diffusion Model (DiffAD). The precision (P), recall (R) and F1 score (F1) are used as evaluation indicators, and the experimental results of the five time series anomaly detection datasets are shown in Table 1.

[0097] Table 1

[0098]

[0099] The embodiments of the present application achieve significant improvements on five benchmark datasets, demonstrating two key advances. First, the embodiments of the present application have a significant advantage over previous diffusion model methods, with an F1 of +20.2% over DiffAD and +25.51% over ImDiffusion, demonstrating that our SNR-guided semi-deterministic generation (SNR implicit conditioning mechanism) addresses the reconstruction instability of diffusion models in time series anomaly detection. The catastrophic failure of ImDiffusion (-15.14% F1 vs. DiffAD) further emphasizes the necessity of our SNR-guided implicit conditioning approach. Second, our (Ours) method (embodiments of the present application) outperforms the previous state-of-the-art reconstruction-based technique of SARAD, with an average F1 of +5.1%, despite a slight gap in SMD, with absolute improvements of 8.27% (-0.42%) and SWaT (-1.58%) on noisy PSM datasets. This highlights the effectiveness of SNR in handling temporal heterogeneity.

[0100] Compared with the prior art, the embodiments of the present application have the following advantages:

[0101] The embodiments of the present application meet the specific needs of reconstruction for time series anomaly detection: by adjusting generation on the potential state inferred by the signal-to-noise ratio, random bias is suppressed while anomalies are attenuated. The embodiments of the present application mainly include two innovations:

[0102] The SNR scheduler is a reparameterized training-inference scheme that unifies noise scaling under quantifiable SNR spectra. Unlike traditional fixed or linear scheduling, this new scheduler enables the model to learn normal patterns at different signal-to-noise ratio calibrated levels of corruption, enhancing robustness to noise perturbations and temporal heterogeneity.

[0103] The SNR implicit conditioning mechanism is a corruption factor estimator that jointly predicts (i) the optimal corruption factor for a given input and (ii) the denoising step (the estimation formula is given in the SNR implicit conditioning mechanism). By dynamically aligning the inverse process with the inherent noise scale of the input, the SNR implicit conditioning mechanism ensures reconstruction balanced manifold fidelity (avoids unrealistic outputs) and input fidelity (maintains non-anomalous trends), addressing the true fidelity trade-off of diffusion models in time series anomaly detection. In addition, when estimating the optimal corruption factor, statistical features such as autocorrelation length, entropy, etc. or frequency domain-based complexity such as power spectrum entropy, high frequency energy ratio, etc. can also be used as ISNR to estimate the optimal corruption factor.

[0104] The above is further detailed description of the present application in combination with specific preferred embodiments, and cannot be deemed as limitation of the specific implementation of the present application to these descriptions. For those skilled in the art to which the present application belongs, without departing from the concept of the present application, a number of equivalent substitutions or obvious variations can be made, and the performance or use is the same, which should be deemed as falling within the protection scope of the present application.

Claims

1. A method for detecting anomalies in time series data, characterized in that, Includes the following steps: S1. Receive input time series data; S2. Based on the characteristics of the input time series data, adaptively estimate the destruction factor and denoising steps, thereby constructing an implicit conditional mechanism that is highly matched with the data characteristics. S3. Based on the destruction factor and the denoising step, the input time series data is partially destroyed, and a reverse denoising process is performed starting from the partially destroyed data to generate a reconstructed time series; implicit guided reconstruction is obtained from the controlled destroyed input. To coordinate the random generation of the diffusion model; where Indicates in model weights The probability of generating a certain value. It is a Markov process. Indicates that the input is The output is , This represents the denoising step that is closest to the optimal destruction factor. Add noise to the data; S4. Calculate the deviation between the input time series data and the reconstructed time series, and output the anomaly detection result based on the deviation; The estimation method for the destruction factor in step S2 includes: The input time series data is decomposed into frequency bands to obtain its low-frequency and high-frequency residual components. The signal-to-noise ratio is inferred based on the energy of the low-frequency and high-frequency residual components. Then, the destruction factor is derived from the inferred signal-to-noise ratio based on the statistical characteristics of the input time series data to improve the sensitivity of anomaly detection.

2. The method according to claim 1, characterized in that, The adaptive estimation in step S2 includes the following steps: The noise injection and removal process is based on signal-to-noise ratio parameterization. The noise level at each step of the diffusion process is defined by the target signal-to-noise ratio and the energy recursion relationship to ensure the consistency of noise scheduling information retention during training and inference.

3. The method according to claim 2, characterized in that, The energy recursion relationship ensures that the total energy dissipated in all diffusion steps is equal to the predefined terminal signal energy.

4. The method according to claim 1, characterized in that, The frequency band decomposition is performed using a zero-phase Gaussian low-pass filter.

5. The method according to claim 1, characterized in that, The method for determining the denoising step in step S3 includes: The estimated disruption factor is mapped to the nearest feasible step in a predefined diffusion time step sequence.

6. The method according to claim 1, characterized in that, The operation of partially destroying the input time series data in step S3 is described as follows: ,in This represents the implicit input of the test instance during the reverse process. For test examples, As a destructive factor, This represents the denoising step that is closest to the optimal destruction factor. Add noise to the data. It is Gaussian noise.

7. The method according to claim 1, characterized in that, In step S4, the deviation is the L2 norm between the input time series data and the reconstructed time series.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method as described in any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the method as described in any one of claims 1 to 7.

Images