CN-121980476-A - Cross-working-condition multivariable time sequence anomaly detection method based on stage perception migration diffusion

Abstract

The invention provides a cross-working condition multivariable time sequence anomaly detection method based on stage perception migration diffusion, which comprises the steps of obtaining historical monitoring multivariable time sequence data, obtaining a source domain training sample set, a target domain adaptation sample set and a sample set to be detected through preprocessing and segmentation, carrying out time local standardization on each sample set, constructing a dynamic graph structure through BallTree, carrying out space neighborhood weighted standardization, extracting space-time joint characteristics by utilizing a time convolution network and a graph annotation force network, obtaining stage probability representation by a training graph variation self-encoder and a stage encoder, obtaining a source domain pre-training model by utilizing stage condition diffusion model training, constructing a stage normal prototype library and a source domain priori threshold, carrying out cross-working condition migration adaptation through stage perception statistics alignment, prototype driving constraint and parameter efficient fine adjustment, obtaining a final detection model and outputting a detection result. Stable cross-working condition abnormal detection under the condition of few samples is realized, and the accuracy and the robustness of detection are improved.

Inventors

• SHU MINGYANG
• WANG CHENG
• CHEN LIYAO

Assignees

• 华侨大学

Dates

Publication Date

20260505

Application Date

20260407

Claims (10)

1. 1. A cross-working condition multivariable time sequence anomaly detection method based on phase perception migration diffusion is characterized by comprising the following steps: Acquiring historical monitoring multivariable time series data acquired by a preset sensor group, and sequentially performing data preprocessing and segmentation on the historical monitoring multivariable time series data to obtain a source domain training window sample set, a target domain adaptation window sample set and a target domain window sample set to be detected; performing time local standardization processing on a source domain training window sample set, a target domain adaptation window sample set and a target domain window sample set to be detected to obtain time standardized window data corresponding to each sample set; According to window data after time standardization corresponding to each sample set, k neighbor searching is conducted by combining BallTree data structures, and a dynamic graph structure corresponding to window samples of each sample set is constructed; Carrying out space neighborhood weighted standardization on window data corresponding to each sample set based on a dynamic graph structure, and carrying out space-time joint feature extraction by combining a preset time convolution network and a graph annotation force network to obtain window-level space-time feature representation corresponding to each sample set; training a graph variable self-encoder and a phase encoder by using window level space-time characteristic representations of source domain training window sample sets, and sequentially inputting window level space-time characteristic representations corresponding to each sample set into the trained graph variable self-encoder and the trained phase encoder to obtain latent variable representations and phase probability representations corresponding to each sample set; Training a phase condition diffusion model by using a latent variable representation and a phase probability representation corresponding to a source domain training window sample set to obtain a source domain pre-training abnormal detection model, and constructing a phased normal prototype library and a source domain priori threshold; Inputting a target domain adaptation window sample set into a source domain pre-training anomaly detection model, and performing cross-working condition migration adaptation processing through stage perception statistics alignment, prototype driving constraint and parameter efficient fine adjustment to obtain a final anomaly detection model; and acquiring time sequence data to be detected, and inputting the time sequence data to be detected into a final abnormality detection model to obtain an abnormality detection result.
2. 2. The method for detecting the cross-working condition multivariable time sequence abnormality based on the phase-aware migration diffusion according to claim 1, wherein the method is characterized by obtaining historical monitoring multivariable time sequence data acquired by a preset sensor group, sequentially performing data preprocessing and segmentation processing on the historical monitoring multivariable time sequence data to obtain a source domain training window sample set, a target domain adaptation window sample set and a target domain window sample set to be detected, and specifically comprises the following steps: acquiring historical monitoring multivariable time series data of industrial equipment under source domain working condition and target domain working condition, which are acquired by a preset sensor group , , , wherein, To at the first The state observation vectors commonly collected by N sensors at sampling moments, T is the total length of the time sequence, For the univariate time series acquired by sensor 1, For the univariate time series acquired by sensor 2, For the univariate time series acquired by the ith sensor, Is a real number, N is the number of sensors, For the ith sensor at the ith Observation values acquired at each sampling moment; Dividing the historical monitoring multivariable time sequence data into a source domain original sequence according to working conditions And the original sequence of the target domain , For the length of the original sequence of the source domain, The length of the original sequence of the target domain; Carrying out data preprocessing on the source domain original sequence and the target domain original sequence, wherein the data preprocessing comprises time stamp alignment processing, resampling processing, missing value detection processing, completion processing, duplicate removal processing and sensor channel validity screening processing; Segmenting the preprocessed source domain original sequence and the preprocessed target domain original sequence by adopting a sliding window mode to obtain a source domain window sample set And target domain window sample set ; Taking a normal window sample in the source domain window sample set as a source domain training window sample to obtain a source domain training window sample set, taking a part of normal window samples in the target domain window sample set as target domain adaptation window samples, and taking the rest normal window samples and abnormal window samples in the target domain window sample set as target domain window samples to be detected to obtain a target domain adaptation window sample set and a target domain window sample set to be detected.
3. 3. The method for detecting the cross-working condition multivariable time sequence anomaly based on the phase-aware migration diffusion according to claim 2, wherein the method is characterized in that a source domain training window sample set, a target domain adapting window sample set and a target domain window sample set to be detected are subjected to time local standardization processing to obtain window data corresponding to each sample set after time standardization, and specifically comprises the following steps: Splitting window samples of a source domain training window sample set, a target domain adaptation window sample set and a target domain window sample set to be detected according to the dimension of the univariate time sequence to obtain an observation value corresponding to each window sample; based on the observed value corresponding to each window sample, extracting the window sample corresponding to the ith sensor As a time local normalization processing object, the formula is: w is the window length, V is the number of the segmentation moments obtained after the segmentation of the univariate time sequence according to the window length W, For the observation of the 1 st slicing moment in the current window for the univariate time series of the ith sensor, For the observation of the time series of univariate of sensor 2 at the 2 nd slicing instant in the current window, The observation value of the ith segmentation moment in the current window is the univariate time sequence of the ith sensor; Performing recursive calculation on a window sample corresponding to the ith sensor in an exponential moving average mode to obtain a local mean value of the window sample at the t cutting moment And local variance The formula is as follows: , , , is a smoothing coefficient; According to local mean And local variance Performing time-by-time normalization processing on a window sample corresponding to the ith sensor to obtain a time normalization value of the ith sensor at the t cutting time , A numerical stability constant greater than 0; Recombining time standardization values corresponding to all sensors according to an original time sequence to obtain time standardized window data corresponding to each sample set, wherein a formula of the time standardized window data is as follows: , in order for the number of sensors to be effective, Is the first The time normalized value of each sensor at the t-th slicing instant.
4. 4. The method for detecting the multi-variable time sequence anomaly based on the stage-aware migration diffusion according to the claim 3, wherein k neighbor searching is performed by combining BallTree data structures according to window data after time standardization corresponding to each sample set, and a dynamic graph structure corresponding to window samples of each sample set is constructed, specifically: According to window data after time standardization corresponding to each sample set and dimension of univariate time sequence, extracting time standardization sequence corresponding to the ith sensor in the current window ; Time-based normalization sequence Extracting multiple statistics reflecting local dynamic characteristics of the data, including mean value Standard deviation of Root mean square value Peak-to-peak value Trend coefficient , , Representing the mean of the time index within the window; combining all statistics to form a graph node statistical feature vector corresponding to the ith sensor , For the purpose of the transposition, The dimension is the statistical feature dimension; Summarizing the graph node statistical feature vectors corresponding to all the effective sensors to obtain a graph node statistical feature set corresponding to the current window sample ; Graph node based statistical feature set Carrying out k neighbor search on the point statistical feature vector by using a preset BallTree data structure, and generating a k neighbor search result; constructing an adjacency matrix of the current window sample according to the k neighbor search result , , For the index set of k nearest neighbor graph nodes of the graph node corresponding to the ith sensor, Is a graph node pair Is a contiguous relationship of (2); symmetric processing is carried out on the adjacent matrix to obtain an undirected dynamic graph structure And after determining the adjacency relation, calculating the edge weight according to the feature distance between the feature vectors of the graph node statistics, wherein for satisfying the following conditions Is a graph node pair of (a) Its side weight is defined as , Is a graph node pair Is used in the adjacent relationship of (a), As a function of the index of the values, Is the first Graph node and the first The feature distance between the nodes of the individual graph, Is the first Graph node and the first The feature distance between the nodes of the individual graph, A scale parameter greater than 0; aiming at graph node pairs which do not meet the adjacency relationship, the edge weight of the graph node pairs is set to be 0, and an edge weight matrix of a current window sample is obtained ; And forming a dynamic graph structure corresponding to the current window sample according to the adjacent matrix and the side weight matrix, and repeatedly executing the steps to obtain the dynamic graph structure corresponding to the window sample of each sample set.
5. 5. The method for detecting the cross-working condition multivariable time sequence abnormality based on stage-aware migration diffusion according to claim 4, wherein the method is based on a graph node statistical feature set Carrying out k neighbor search on the point statistical feature vector by using a preset BallTree data structure to generate k neighbor search results, wherein the k neighbor search results are specifically as follows: Statistical feature vector by graph node As query point, statistics feature set at graph node In the constructed feature space, k neighbor searching is carried out on the rest graph node statistical feature vectors by utilizing BallTree data structures, and k neighbor graph node index sets closest to the graph node corresponding to the ith sensor are obtained K is the number of neighbor graph node indexes, The 1 st adjacent graph node with local similarity relationship exists for the graph node corresponding to the i-th sensor, The 2 nd adjacent graph nodes with local similarity relationship exist for the graph nodes corresponding to the i-th sensor, A kth adjacent graph node with a local similarity relationship exists for the graph node corresponding to the ith sensor; calculating characteristic distance between graph node corresponding to ith sensor and graph node corresponding to jth sensor , The feature vector is counted for the graph node corresponding to the jth sensor, Is the L2 norm; When judging to When the image node corresponding to the jth sensor and the image node corresponding to the ith sensor are judged to have a local spatial association relationship under the current window; And determining the feature distance between all the graph nodes, and generating a k neighbor search result.
6. 6. The cross-working condition multivariable time sequence anomaly detection method based on phase perception migration diffusion according to claim 5, wherein the method is characterized in that window data corresponding to each sample set after time normalization is subjected to space neighborhood weighted normalization based on a dynamic graph structure, and space-time joint feature extraction is performed by combining a preset time convolution network and a graph annotation force network to obtain window-level space-time feature representation corresponding to each sample set, and specifically comprises the following steps: according to the ith sensor Dynamic graph structure of individual graph nodes and determination of neighborhood set thereof And calculate the t time Neighborhood weighted mean of individual graph nodes And neighborhood weighted variance , The j-th sensor corresponds to A time normalized value of the graph node at the t-th moment in the window; Based on the neighborhood weighted mean value and the neighborhood weighted variance, spatial neighborhood weighted normalization processing is carried out on the window data after corresponding time normalization, and the spatial normalization result of the ith sensor at the t moment is given by the following formula: ; reorganizing all spatial normalization results corresponding to each sample set according to the original time sequence to obtain window data after spatial neighborhood weighting normalization , Weighting normalized window data for the spatial neighborhood of the ith sensor at the t moment; based on window data after spatial neighborhood weighting normalization, determining a univariate sequence corresponding to an ith sensor And sequence single variable Inputting the time-dependent characteristic into a preset one-dimensional time convolution network, extracting the time-dependent characteristic through multi-layer causal convolution and expansion convolution, and obtaining a time characteristic vector corresponding to an ith sensor , For the time-convolved network mapping function, Is a time feature dimension; summarizing all the time feature vectors to obtain a node-level time feature matrix of the current window sample , Is the first Time feature vectors corresponding to the sensors; Inputting a dynamic graph structure corresponding to the node-level time feature matrix into a preset graph annotation force network, and modeling a spatial interaction relation between univariate time sequences, wherein graph nodes are calculated Node of graph Attention relevance score of (a) , For a linear rectification activation function, In order to be able to take the vector of attention parameters, For a characteristic linear transformation matrix, For the vector concatenation operation, The ith sensor corresponds to the ith The temporal feature vectors of the individual graph nodes, Corresponds to the j-th sensor Time feature vectors of the individual graph nodes; Carrying out normalization processing on neighborhood attention coefficients by combining with a dynamic graph structure to obtain graph nodes Opposite graph node Is a normalized attention weight of (2) , Is a graph node Node of graph Is provided with an attention relevance score of (a), Is a graph node pair Is a side weight of (2); Based on normalized attention weights Performing weighted aggregation processing to obtain node-level space interaction characteristics corresponding to the ith sensor , Is a nonlinear activation function; Splicing the node-level space interaction features output by a plurality of attention heads by adopting a multi-head diagram attention mechanism to obtain a final node-level space-time joint feature representation , Is the first The node-level spatial interaction characteristics corresponding to the individual sensors, The node level space-time joint feature dimension; Aggregating node-level space-time joint feature representation to obtain window-level space-time feature representation corresponding to the current window sample , And repeatedly executing the steps to obtain window-level space-time characteristic representations corresponding to the window samples of each sample set.
7. 7. The method for detecting the cross-working condition multivariable time sequence anomaly based on the phase perception migration diffusion according to claim 6, wherein the window level space-time characteristic representation of the source domain training window sample set is used for training the graph variation self-encoder and the phase encoder, and window level space-time characteristic representations corresponding to the sample sets are sequentially input into the trained graph variation self-encoder and the phase encoder to obtain latent variable representations and phase probability representations corresponding to the sample sets, specifically comprising the following steps: representing window-level space-time characteristics corresponding to source domain training window sample set Inputting the latent variable posterior distribution into a coding network of a preset graph variation self-coder to obtain a mean vector of corresponding latent variable posterior distribution And logarithmic variance vector , , For a mapping network used to generate the latent variable mean, To generate a mapping network of latent variable log variances, As a dimension of the latent space, Is a logarithmic variance vector; Let a priori noise Satisfy the following requirements Obtaining training latent variable representation corresponding to window samples of a source domain training window sample set The method comprises the following steps: , For the element-by-element multiplication, The standard multi-element Gaussian distribution is characterized in that the mean value is zero, and the covariance matrix is the identity matrix I; representing training latent variables Inputting the window level space-time characteristic representation into a decoding network of a graph variation self-encoder, and reconstructing the corresponding window level space-time characteristic representation to obtain a reconstructed window level space-time characteristic representation corresponding to window samples of a source domain training window sample set , A decoding network separating the graph variations from the encoder; training loss to construct a graph variation self-encoder , And by minimizing training loss Joint iteration updating is carried out on the coding network and decoding network parameters of the graph variation self-encoder to obtain the trained graph variation self-encoder, wherein, , , The reconstruction is lost to the process, To minimize the weight coefficients of the training loss, For the loss of the KL divergence, For the Kullback-Leibler divergence, For encoding the posterior distribution obtained by the network learning, Is a standard Gaussian prior distribution; representing window-level space-time characteristics corresponding to source domain training window sample set Inputting the obtained image variation into a coding network of a trained image variation self-coder to obtain a corresponding latent variable representation Summarizing all latent variable representations Obtaining a latent variable set ; Clustering is carried out on the latent variable set, the distribution of window samples of the source domain training window sample set in the latent space is divided into K clusters, and phase pseudo labels corresponding to the source domain training window sample set are obtained K is the total number of stages; Training phase pseudo tags corresponding to a window sample set using a source domain Training a preset phase encoder to obtain a phase probability representation corresponding to a source domain training window sample set, wherein the latent variable representation Inputting the phase encoder to obtain a phase probability vector corresponding to the source domain training window sample set , , For the normalization of the exponential function, In the case of a phase encoder, Window samples of the training window sample set for the source domain are at the first Probability of each stage; pseudo-label of stage corresponding to source domain training window sample set Construction phase encoder training loss as supervisory information , In order to indicate the function, Training the window sample number of a window sample set for a source domain and training the loss by minimizing the phase encoder Iteratively updating parameters of the phase encoder to obtain a trained phase encoder; And sequentially inputting the target domain adaptation window sample set and the target domain window sample set to be detected into a trained coding network of the graph variation self-coder, extracting latent variable representations corresponding to the sample sets, sequentially inputting the latent variable representations corresponding to the sample sets into a trained phase coder, and obtaining phase probability representations corresponding to the sample sets.
8. 8. The cross-working condition multivariable time sequence anomaly detection method based on phase perception migration diffusion according to claim 7, wherein the method is characterized in that a phase condition diffusion model is trained by using a latent variable representation and a phase probability representation corresponding to a source domain training window sample set to obtain a source domain pre-training anomaly detection model, and a phased normal prototype library and a source domain priori threshold are constructed, specifically comprising the following steps: stage probability vectors corresponding to source domain training window sample sets As a condition control signal, embedding the condition control signal into a preset mapping module to obtain stage condition embedding , For conditional embedding of the mapping function, Embedding dimensions for conditions; Representing latent variables As an initial latent variable Setting the total number of diffusion steps as The noise scheduling sequence is Performing forward diffusion noise adding processing in the latent space to obtain noise-added latent variables under different diffusion moments, wherein each diffusion step meets the following requirements , Is the first The noise intensity of the diffusion step numbers injected into the latent variable, wherein as the diffusion step numbers increase, the original structure information in the latent variable representation is gradually disturbed by random noise; Wherein, setting , , Is the first The number of spread steps preserves the original signal proportion, From the 1 st diffusion step number to the 1 st Accumulating the ratio of the retained original signals by the number of diffusion steps; In the first place Initial latent variable at several diffusion steps Corresponding noisy latent variable The formula of (2) is: , , Random noise that is subject to a standard gaussian distribution; for the first The number of diffusion steps is the noise-added latent variable Diffusion time coding Stage condition embedding The noise-spreading and denoising network is input together to obtain the prediction result of the network on the noise item , Is as the parameter of A diffuse denoising network of (2); calculating a prediction result With truly added random noise The difference between them, construct the diffusion denoising loss function , In order to obtain a trained stage condition diffusion model and a source domain pre-training anomaly detection model, iteratively updating diffusion denoising network parameters by minimizing diffusion denoising loss; Pseudo-labels of the stage corresponding to the sample set of the training window according to the source domain Dividing a source domain training window sample set to obtain Normal latent variable subset corresponding to each stage And for normal latent variable subset Clustering to obtain Prototype set under stage , Is the first The m-th normal prototype center at the stage, The number of prototypes at the stage; summarizing prototype sets corresponding to all stages to form a staged normal prototype library ; Representing latent variables And stage condition embedding Inputting a trained stage condition diffusion model, and performing J independent diffusion reverse sampling on window samples of each source domain training window sample set to obtain J groups of reconstruction results, wherein the J groups of reconstruction results are respectively expressed as: , For the reconstruction result of group 1, For the reconstruction result of group 2, Reconstructing the result for group J; And calculating a reconstruction error between each set of reconstruction results and the original latent variable representation, wherein the formula is as follows: , For the first set of reconstruction errors, For the first set of reconstruction results, calculating an average reconstruction error for the window samples And reconstruction uncertainty Constructing an anomaly score for the window sample , For the weight coefficients of the average reconstruction error, Weight coefficients that are anomaly scores; Pseudo tag according to stage Dividing all abnormal scores corresponding to the source domain training window sample set according to stages to obtain Normal sample anomaly score set at each stage And according to preset dividing points Calculation of Source domain prior threshold value corresponding to each stage Obtaining a source domain priori threshold value set corresponding to all phases 。
9. 9. The method for detecting the cross-working condition multivariable time sequence anomaly based on the phase-aware migration diffusion according to claim 8, wherein a target domain adaptation window sample set is input into a source domain pre-training anomaly detection model, and the cross-working condition migration adaptation processing is carried out through phase-aware statistics alignment, prototype driving constraint and efficient fine adjustment of parameters to obtain a final anomaly detection model, which is specifically as follows: Freezing coding trunk parameters in a source domain pre-training anomaly detection model, and updating only low-rank adaptation parameters and partial normalization layer parameters in a diffusion denoising network, wherein the source domain pre-training anomaly detection model comprises a trained time local normalization module, a dynamic diagram construction module, a spatial neighborhood weighting normalization module, a space-time joint feature extraction module, a diagram variation self-encoder and a phase encoder; adapting window samples of a window sample set according to a target domain Probability of individual phases Latent variable representation corresponding to window samples of a target domain adaptation window sample set Calculating a weighted average of the target domain at the stage And a weighted covariance matrix ; Window samples of the training window sample set according to the source domain are at the first Probability of individual phases Latent variable representation corresponding to window samples of a source domain training window sample set Calculating a weighted average of the source domain at this stage And a weighted covariance matrix ; Comparing the mean and covariance differences of the source domain and the target domain in each stage, and constructing stage perception statistics alignment loss , Is the first The weight coefficient of each stage is set to be equal to the weight coefficient of each stage, Is the Frobenius norm; Prototype-driving constraints are applied to window samples of a target domain adaptation window sample set according to a phased normal prototype library Maintaining the stability of the multi-mode structure in the stage of the normal sample of the target domain in the adapting process while completing stage-level statistics alignment; Performing condition fine adjustment on the stage condition diffusion denoising network by adopting a target domain adaptation window sample set, and combining with target domain diffusion denoising loss Stage aware statistical alignment loss And prototype driven constraint loss Building target domain adaptation total loss , Diffusion denoising penalty for target domain Is used for the weight coefficient of the (c), Statistical alignment loss for stage awareness Is used for the weight coefficient of the (c), Constraint loss for prototype drive Weight coefficient of (2); Performing iterative optimization on parameters in the source domain pre-training anomaly detection model by minimizing the total loss of target domain adaptation to obtain a migration adapted target domain anomaly detection model; Performing multiple condition diffusion reverse sampling reconstruction by using normal window samples in the target domain adaptation window sample set to obtain abnormal score distribution of the target domain normal samples, calculating a target domain stage calibration threshold according to stage division, and further combining a source domain prior threshold and the target domain stage calibration threshold to obtain a fusion stage threshold; Inputting a target domain window sample set to be detected into a target domain abnormality detection model, and carrying out abnormality judgment by combining the fusion phase threshold value to obtain a test abnormality detection result; when the comparison result reaches a preset index, the target domain abnormality detection model and a corresponding fusion stage threshold value are used as a final abnormality detection scheme; and when the comparison result does not reach the preset index, returning to the target domain migration adaptation step, and carrying out parameter optimization, threshold calibration and result evaluation again until the comparison result reaches the preset index.
10. 10. The method for detecting the cross-working condition multivariable time sequence abnormality based on the phase-aware migration diffusion according to claim 9, wherein the preset indexes comprise accuracy, precision, recall, F1 fraction, cross-working condition performance improvement rate and adaptation time.

Description

Cross-working-condition multivariable time sequence anomaly detection method based on stage perception migration diffusion Technical Field The invention relates to the technical field of anomaly detection, in particular to a cross-working condition multivariable time sequence anomaly detection method based on phase perception migration diffusion. Background With the development of industrial equipment to large-scale, complicated and intelligent directions, the requirements of fields such as manufacturing, energy, process industry, intelligent operation and maintenance and the like on the real-time monitoring and reliability of the running state of the equipment are increasingly improved. In the long-term operation process of industrial equipment, the industrial equipment is influenced by a plurality of factors such as mechanical abrasion, environmental interference, load fluctuation, component aging, working condition switching and the like, the problems of performance degradation, abnormal operation, even fault shutdown and the like are easy to occur, and safety accidents or great economic losses are possibly caused when the industrial equipment is serious. Therefore, a multi-sensor monitoring system is commonly deployed in an industrial field, continuously collects various sensor data including temperature, pressure, flow, vibration and the like, forms massive multivariable time series data, and provides a basic condition for abnormality detection based on data driving. In the field of industrial multivariable time sequence anomaly detection, the existing methods are mainly divided into three types, namely a supervised method, an unsupervised method and a semi-supervised or mixed method. The supervised method relies on a large number of high-quality labeling samples to realize abnormal recognition through classification or regression models, but in an actual industrial scene, the abnormal samples are rare, the labeling cost is high, and the training requirement of supervised learning is difficult to meet. The unsupervised method generally only uses normal data to perform distributed modeling, and identifies abnormality through reconstruction error or density estimation, so that the problem of labeling dependence is avoided, but the characterization capability of the unsupervised method on a normal mode directly influences detection performance. Semi-supervised or hybrid approaches attempt to compromise known anomaly identification and unknown anomaly discovery capabilities, attempting to improve generalization performance under limited supervision information. However, the prior art still has obvious defects in the industrial equipment multivariable time sequence anomaly detection scene. Firstly, industrial multivariable monitoring data has the characteristics of high dimension, strong coupling, non-stability and multiple working conditions, and anomalies are often represented by over-limit of single variable amplitude, and are further represented by changes of association structures among variables and dynamic response modes. The traditional method is generally difficult to simultaneously describe complex space-time dependency relationship among variables and normal multi-mode structures in different operation stages, and false alarm is easy to generate in the normal mode switching process. Secondly, in actual deployment, the model often needs to be migrated from a source working condition to a target working condition, distribution offset exists between a source domain and a target domain generally, only a small amount of normal samples are available in the target domain, and stable and effective cross-working condition adaptation is difficult to realize under the condition of few samples by the existing method. In addition, in the existing method, the abnormality judgment is mostly carried out based on a single reconstruction error and a fixed threshold value, the reconstruction uncertainty information is difficult to fully utilize, and the difference of the error distribution of the normal sample in different operation stages is difficult to adapt, so that the accuracy, the robustness and the practicability of the abnormality detection still need to be improved. In view of this, the present application has been proposed. Disclosure of Invention The invention provides a cross-working condition multivariable time sequence anomaly detection method based on stage perception migration diffusion, which can at least partially improve the problems. In order to achieve the above purpose, the present invention adopts the following technical scheme: a cross-working condition multivariable time sequence anomaly detection method based on phase perception migration diffusion comprises the following steps: Acquiring historical monitoring multivariable time series data acquired by a preset sensor group, and sequentially performing data preprocessing and segmentation on the historical monitoring multivariable time ser