Time series data anomaly detection method and system based on convergence trajectory feature analysis

Abstract

The present application relates to the technical field of data analysis, and discloses a time series data anomaly detection method and system based on convergence trajectory feature analysis, comprising: collecting multi-dimensional time series data of at least one industrial equipment belonging to the same type and preprocessing to obtain a standardized single-equipment feature sequence of each industrial equipment; training a baseline autoencoder model based on the standardized feature tensor to obtain baseline parameters and baseline features of the baseline autoencoder model, and respectively taking the standardized single-equipment feature sequence of each industrial equipment as a single sample to disturb the baseline autoencoder model based on the baseline parameters to obtain the convergence trajectory features of the baseline autoencoder model, thereby obtaining a standard digital portrait library based on all industrial equipment; and performing feature attribution and root cause analysis on the digital portrait. The present application significantly improves the sensitivity and detection accuracy of weak signals, small samples and complex behavior anomalies.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data analysis, and particularly relates to a time series data anomaly detection method and system based on convergence trajectory feature analysis. BACKGROUND

[0002] The anomaly detection and health diagnosis of industrial time series data have become core tasks in intelligent manufacturing, equipment management and production safety guarantee. With the continuous promotion of industrial internet and digital factory, enterprises are increasingly relying on automated means to monitor and warn the running data of massive equipment, workshop production lines and energy systems in real time, in an attempt to maximize equipment availability and production efficiency while reducing operation and maintenance costs.

[0003] Under this background, the industry has proposed various data-driven anomaly detection schemes with statistical analysis, machine learning and deep learning as the core. For example, classic statistical control charts, principal component analysis, isolation forest, LSTM / GRU type time series models and self-encoder neural network methods are all applied to trend identification and anomaly detection of various industrial indicators. These schemes have improved the intelligent level in industrial scenarios to some extent and can respond to some common anomalies or faults.

[0004] However, the actual industrial environment faces complex challenges such as multi-source heterogeneity, serious data missing, significant device differences and dynamic and variable working conditions. Traditional methods generally rely on static thresholds, single-index analysis or global model fitting, and lack systematic processing capabilities for multi-device, multi-dimensional indicators, non-uniform sampling and time series dynamic behavior. Although deep learning models can mine time series and multi-variable features, existing processes often ignore key engineering problems such as standardization of large-scale data, missing data completion, device normalization and stream parallel processing. At the same time, most models only focus on the final prediction or reconstruction error, and do not utilize the dynamic disturbance trajectory of the model convergence process and other high-order features, resulting in insufficient sensitivity and explainability for complex anomalies, small sample hidden dangers and difficult learning conditions.

[0005] More seriously, existing schemes are difficult to adapt to batch deployment, real-time stream analysis and cross-platform system integration in large-scale industrial scenarios, and engineering landing and scalability face bottlenecks.

[0006] Industrial anomaly detection is a core requirement for the Industrial Internet and predictive maintenance. According to market reports, the global predictive maintenance market was worth approximately $5.5 billion in 2022, with a CAGR of nearly 17%. Many industries face high downtime costs (median downtime losses of approximately $125,000 per hour), and accurate fault detection can bring significant economic benefits. Current predictive maintenance solutions generally have low accuracy rates, necessitating more efficient and intelligent algorithm support. Equipment operation and maintenance (O&M) and production management personnel urgently need an intelligent analysis technology that can automatically adapt to equipment differences, operating condition fluctuations, and the diversity of anomaly types, enabling batch, standardized, and interpretable anomaly detection and risk identification, thus helping enterprises move towards higher levels of intelligent manufacturing and digital transformation. Summary of the Invention

[0007] This invention aims to propose a time-series data anomaly detection method and system based on convergence trajectory feature analysis. This method can uniformly process multi-dimensional, non-uniformly sampled, and severely missing production data from large-scale equipment, achieving full lifecycle health behavior modeling and highly sensitive anomaly detection for batch equipment. Based on the time-series indicators of all equipment in the entire work area, this method first achieves automatic alignment and normalization of data from multiple equipment, multiple indicators, and different sampling periods through dynamic baselines and probabilistic standardization, constructing a unified, structured standard input tensor.

[0008] Building upon this foundation, this invention innovatively introduces a convergence perturbation trajectory analysis framework based on a deep autoencoder model. By individually perturbing and training the data for each device, the system collects high-order dynamic features during the neural network convergence process, forming a unique digital profile that comprehensively reflects the behavioral sensitivity spectrum and anomaly response capabilities of the equipment during operation. Through the standardization, normalization, and graphical representation of the digital profile, it achieves automatic identification of ordinary, difficult-to-learn, abnormal, and latent high-risk samples, and supports various intelligent engineering analysis methods such as visual clustering, spatial dimensionality reduction, and anomaly attribution.

[0009] Compared to traditional anomaly detection methods that rely on manual thresholds, single model residuals, or static feature similarity, this invention can adaptively characterize individual equipment differences, long-term dynamic evolution, and various abnormal operating conditions, significantly improving the sensitivity and detection accuracy for weak signals, small samples, and complex behavioral anomalies. The entire process supports batch, streaming, and distributed parallel processing, facilitating automatic deployment and horizontal scaling in industrial internet, cloud platforms, and big data systems.

[0010] This invention requires no additional hardware or reliance on expert rules, and can directly connect to mainstream industrial production databases and data streams to achieve standardized health screening, behavioral profiling, and root cause analysis of equipment groups, significantly reducing engineering implementation and maintenance costs. The methodology is clear and transparent, easily integrated into existing production management and intelligent operation and maintenance platforms, and can serve as a general tool for intelligent diagnosis, risk warning, and decision optimization in industrial enterprises. It significantly improves the intelligence level and inherent safety capabilities of large-scale industrial systems, possessing broad engineering application prospects and promotional value.

[0011] This invention is achieved through the following technical solution:

[0012] A time-series data anomaly detection method based on convergence trajectory feature analysis includes:

[0013] Data preprocessing and standardization: Collect multidimensional time-series data of at least one industrial device of the same type and preprocess it to obtain a standardized single-device feature sequence for each industrial device. ,in, For the standardized single-device feature sequence of the i-th industrial equipment, the standardized single-device feature sequences of all the industrial equipment are... Stacked according to device dimension, forming a standardized feature tensor based on all devices. , ,in, For the number of devices, For lifecycle progress points, For feature dimensions;

[0014] Digital profile construction based on the convergence trajectory of the deep model: based on the standardized feature tensor The baseline parameters of the baseline autoencoder model are obtained by training the baseline autoencoder model. and baseline characteristics Based on the baseline parameters The standardized single-device feature sequence of each of the aforementioned industrial devices is respectively... The baseline autoencoder model is perturbed and trained using a single sample to obtain its convergence trajectory features, thereby generating a digital profile matrix corresponding to the industrial equipment, and thus obtaining a standard digital profile library based on all the industrial equipment. ;

[0015] Anomaly detection and interpretability analysis: Feature extraction and spatial modeling are performed on the digital profile, automated anomaly detection is achieved through anomaly measurement algorithms, and feature attribution and root cause analysis are performed in combination with interpretable AI technology.

[0016] As an optimization, multidimensional time-series data of at least one industrial device of the same type is collected and preprocessed to obtain a standardized single-device feature sequence for each industrial device. The specific process is as follows:

[0017] A1. Collect multidimensional time-series data of at least one industrial device of the same type. And time alignment is performed to obtain the industrial time-series data matrix A, where, This represents the d-th dimension index measurement value of device i at the t-th observation time. , , , This indicates the number of observation data points of the device within the historical period; Indicates the dimension of the indicator;

[0018] A2. Calculate all non- NaN The multidimensional time series data dynamic average baseline and dynamic standard deviation baseline and based on the dynamic average baseline and dynamic standard deviation baseline The deviation of each of the aforementioned industrial devices relative to the dynamic baseline at each observation time was obtained. and the deviation The cumulative deviation score for each of the industrial devices is obtained by accumulating along the lifecycle time dimension. ;

[0019] A3. From the aforementioned multidimensional time-series data Obtaining nonnegative time series indicators For the non-negative time series index The cumulative key indicator sequence is obtained by summing the results. Through quantile standardization and unified lifecycle progress interpolation, multi-device time-series data normalization processing is achieved, generating standardized feature tensors adapted for intelligent analysis. ;

[0020] A4. Design the data interface and convert the standardized feature tensor... Output is made through the data interface.

[0021] As an optimization, the dynamic average baseline The formula is:

[0022] ;

[0023] The dynamic standard deviation baseline The formula is:

[0024] ;

[0025] in, This is an indicator function with the following parameters: The value is 1 if it is true, and 0 otherwise. This represents the dynamic average baseline of the d-dimensional index at time t. This represents the baseline of the dynamic standard deviation of the d-dimensional index at time t;

[0026] The deviation The formula is:

[0027] ;

[0028] in, This represents the deviation of the d-dimensional index of industrial equipment i from the dynamic baseline at time t. It is the cumulative function of the standard normal distribution;

[0029] The cumulative value of the deviation score The formula is:

[0030] ;

[0031] in, denoted as the cumulative deviation score of the d-dimensional index of industrial equipment i at time t, and N represents the number of observation data points.

[0032] As an optimization, the specific process for A3 is as follows:

[0033] A3.1, For each of the aforementioned industrial devices, the non-negative time series indicators... The cumulative key performance indicators are obtained by summing the results. Simultaneously, the non-negative time series index is obtained. During the observation period of the entire work area The cumulative index total value during the global observation period. Then, for each of the aforementioned industrial devices, a cumulative key indicator sequence... Quantile standardization is performed to obtain the quantile standardization progress. Among them, the cumulative key indicator sequence The specific formula is: The cumulative index total value during the global observation period The specific formula is: The progress of the quantile standardization The specific formula is: , This represents the cumulative sum of key performance indicators for device i at time t. This represents the cumulative index total value of device i during the global observation period. This represents the quantile normalization progress of device i at time t. N represents the number of observation data points;

[0034] A3.2. Based on the preset number of lifecycle progress points L, construct a standardized lifecycle progress scale shared by all industrial equipment. Then, for each industrial device i, a multidimensional probabilistic deviation feature sequence... Through interpolation function Map it to a uniform progress scale Generate standardized time series: The standardized deviation characteristics of each dimension of the industrial equipment i are organized according to a unified schedule dimension to form a standardized time series set of industrial equipment i. Finally, the standardized time series sets of all the aforementioned industrial equipment are stacked according to the equipment dimension to form a standardized feature tensor based on all equipment. , .

[0035] As an optimization, based on the standardized feature tensor The baseline parameters of the baseline autoencoder model are obtained by training the baseline autoencoder model. and baseline characteristics Based on the baseline parameters The standardized single-device feature sequence of each of the aforementioned industrial devices is respectively... The baseline autoencoder model is perturbed and trained using a single sample to obtain its convergence trajectory features, thereby generating a digital profile matrix corresponding to the industrial equipment, and thus obtaining a standard digital profile library based on all the industrial equipment. The specific process is as follows:

[0036] Construct a baseline autoencoder model and use the standardized feature tensor The baseline autoencoder model is trained iteratively until the loss function converges, yielding the baseline parameters. and based on the baseline parameters Set baseline features , ,in, Represents the encoder function. Represents the baseline characteristics of industrial equipment i. Indicates the number of feature dimensions. Represents the standardized temporal feature tensor of device i;

[0037] For each of the aforementioned industrial devices, the standardized single-device feature sequence corresponding to that industrial device is used. As a single sample pair based on the baseline parameters The baseline autoencoder model is trained with perturbation K times, and the perturbation features output at each step of the baseline autoencoder model during training are extracted. Then, the perturbation features described in each step are combined with the corresponding baseline features to obtain the feature difference for each step. The feature differences obtained from K perturbations are stacked in order to form a digital profile matrix of industrial equipment i. Digital profile matrix of all the devices The initial digital profile database is obtained by summarizing the data. ;

[0038] For the initial digital portrait library Standardization processing yields a standard digital portrait library. ,in, This represents the perturbation feature output by the baseline autoencoder model for industrial equipment i after the j-th perturbation training. Let be the model parameters of industrial equipment i after the j-th perturbation training.

[0039] As an optimization, the standard digital portrait library The formula is:

[0040] ;

[0041] , ;

[0042] , These represent the smallest and largest digital profile matrices in the initial digital profile database, respectively. A standard digital profile matrix representing industrial equipment i.

[0043] As an optimization, the digital profile undergoes feature extraction and spatial modeling, automated anomaly detection is achieved through anomaly measurement algorithms, and feature attribution and root cause analysis are performed using interpretable AI technology. The specific process is as follows:

[0044] For the standard digital portrait library A feature library is obtained by extracting high-order features from the standard digital profile matrix of each of the aforementioned industrial devices. ,in, , To extract feature dimensions based on the standard digital profile matrix, High-dimensional embedding of the standard digital profile matrix representing industrial equipment i;

[0045] Based on feature library The system calculates the anomaly metric score of the high-order embedding of each industrial device, and determines whether the multidimensional time-series data of the industrial device is abnormal based on the anomaly metric score. Simultaneously, it uses the feature library... The industrial equipment is clustered to obtain clustering results;

[0046] The SHAP contribution of the anomaly metric scores of the standard digital profile matrix of the industrial equipment is calculated respectively, and the SHAP contribution of all the industrial equipment is rendered into a two-dimensional heat map. Feature attribution and root cause analysis are performed by analyzing the two-dimensional heat map.

[0047] As an optimization, the specific formula for the SHAP contribution of the anomaly metric score is as follows:

[0048] ;

[0049] in, Let be the feature attribution value of industrial equipment i in the j-th perturbation dimension and the k-th perturbation; s (⋅) is the anomaly metric score function. To A new profile matrix after single-point intervention; A standard digital profile matrix for industrial equipment i.

[0050] As an optimization, after obtaining the feature library Then, the feature library Projecting onto a low-dimensional space enables the distributed visualization of multi-dimensional time-series data of industrial equipment.

[0051] This invention also discloses a time-series data anomaly detection system based on convergence trajectory feature analysis, used in conjunction with the aforementioned time-series data anomaly detection method based on convergence trajectory feature analysis, comprising:

[0052] The data preprocessing and standardization module is used to collect multidimensional time-series data of at least one industrial device of the same type and preprocess it to obtain a standardized single-device feature sequence for each industrial device. ,in, For the standardized single-device feature sequence of the i-th industrial equipment, the standardized single-device feature sequences of all the industrial equipment are... Stacked according to device dimension, forming a standardized feature tensor based on all devices. , ,in, For the number of devices, For lifecycle progress points, For feature dimensions;

[0053] The deep model convergence trajectory digital profile construction module is used to construct digital profiles based on the standardized feature tensor. The baseline parameters of the baseline autoencoder model are obtained by training the baseline autoencoder model. and baseline characteristics Based on the baseline parameters The standardized single-device feature sequence of each of the aforementioned industrial devices is respectively... The baseline autoencoder model is perturbed and trained using a single sample to obtain its convergence trajectory features, thereby generating a digital profile matrix corresponding to the industrial equipment, and thus obtaining a standard digital profile library based on all the industrial equipment. ;

[0054] The anomaly detection and interpretability analysis module is used to extract features and spatially model the digital profile, achieve automated anomaly detection through anomaly measurement algorithms, and combine interpretable AI technology to perform feature attribution and root cause analysis.

[0055] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0056] 1. A Dynamic Normalization and Probabilistic Standardization Data Processing System for Multi-Source Industrial Time Series: This invention addresses industry pain points such as inconsistent data collection cycles of industrial equipment, heterogeneous multi-indicator data, and data gaps. It proposes a normalization system centered on a dynamic work area baseline (mean / standard deviation), probabilistic z-score standardization, and cumulative quantile normalization techniques. Through dynamic baseline and cumulative key indicator quantile mapping, unified normalization and interpolation of multiple devices, multiple lifecycles, and multi-dimensional indicators are achieved in batches. This effectively eliminates differences in scale and distribution between devices, between periods, and between indicators, laying a highly standardized data foundation for downstream deep model batch training and comparative analysis.

[0057] 2. Digital Profile Features Based on Converging Perturbation Trajectories of Deep Neural Network Models: This invention proposes for the first time to use the fully converged parameters of a deep neural network as a baseline. By training each device (sample) with individual perturbations, the convergence perturbation trajectory of the neural network is collected to form a high-dimensional digital profile. This "converging perturbation trajectory" method innovatively mines the dynamic behavioral features of the neural network parameter space, characterizing the essential differences in device operation and abnormal response capabilities. It far surpasses traditional shallow anomaly detection methods such as reconstruction error and static feature similarity. The obtained digital profile supports direct visualization, clustering, dimensionality reduction, and interpretability analysis, significantly improving the sensitivity and detection accuracy for weak signals, small samples, and complex behavioral anomalies.

[0058] 3. Standardized Digital Profile-Based Batch Anomaly Detection and Interpretable Attribution Mechanism: This invention uses a normalized digital profile matrix as a device behavior fingerprint, supporting automated anomaly detection, clustering, and risk attribution for batch devices. It introduces mainstream interpretable AI methods such as SHAP to perform heatmap attribution on the contribution of digital profiles to anomaly samples, achieving mathematical tracing and physical explanation of the root causes of device anomalies. This feature is widely applicable to various industrial systems requiring equipment health management, behavior monitoring, and root cause tracing.

[0059] 4. End-to-end tensor quantization to support industrial-grade streaming and distributed parallel deployment: The method of this invention supports end-to-end data tensor quantization processing, matrix-based parallel processing, and distributed streaming interfaces, and can be horizontally scaled to ultra-large-scale scenarios with millions of devices and extremely high-frequency data acquisition. Whether in batch offline or real-time streaming scenarios, it can be connected to industrial clouds, big data platforms, and edge computing nodes. The method does not rely on specific hardware or require manual threshold setting, facilitating migration across industries and scenarios, and large-scale engineering implementation.

[0060] Explanation of the method's generality, circumvention, and scalability:

[0061] Universality: All key aspects of this invention (dynamic standardization, convergence trajectory modeling, and digital profile analysis) are universal modules. Data standardization or normalization in accordance with data science principles has been implemented in multiple key steps. It does not depend on specific processes, specific indicators, or industry scenarios and has extremely strong horizontal migration and batch deployment capabilities.

[0062] Scalability: Converging perturbation trajectory analysis is not only applicable to autoencoders, but can also be adapted to any deep learning model that supports parameter perturbations and feature outputs.

[0063] Alternative implementations: If it is necessary to circumvent this invention, a pure reconstruction error or rule engine method that does not involve convergence perturbation trajectory or digital profile normalization can be used, but it is difficult to achieve the high sensitivity, interpretability and batch processing capability of this invention at the same time.

[0064] Scope of application: The technical solution of this invention is not only applicable to typical scenarios such as oil and gas, manufacturing, energy, and power, but can also be extended to other industries that require intelligent anomaly detection of batch time-series data, such as smart cities, building management, intelligent transportation, and medical equipment. Attached Figure Description

[0065] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and form part of this application, do not constitute a limitation thereof. In the drawings:

[0066] Figure 1 A schematic diagram of the three-dimensional spatial trajectory of the normalized cumulative probability fraction of oil, gas and water.

[0067] Figure 2 A schematic diagram of the two-dimensional projection of the normalized cumulative probability fraction of oil and water into a three-dimensional space;

[0068] Figure 3 A schematic diagram of a two-dimensional projection of the three-dimensional space of the cumulative probability fraction of oil and gas;

[0069] Figure 4 A two-dimensional projection diagram of the three-dimensional space of the probability fraction of water vapor accumulation;

[0070] Figure 5 A schematic diagram of standardized time axis (cumulative production percentage) - standardized indicator (oil production);

[0071] Figure 6 A schematic diagram of standardized time axis (cumulative production percentage) - standardized indicator (water production);

[0072] Figure 7 A schematic diagram of standardized time axis (cumulative production percentage) - standardized indicator (GOR, oil-gas ratio);

[0073] Figure 8 A heat map diagram of equipment number (well) - standardized life cycle (cumulative production percentage) - standardized indicators, where (a) is the probability distribution map of cumulative oil production, (b) is the probability distribution map of cumulative water production, and (c) is the probability distribution map of cumulative gas production.

[0074] Figure 9 A pseudo-color diagram illustrating the equipment number (well) - standardized lifecycle (cumulative production percentage) - standardized indicators;

[0075] Figure 10 The diagram shows the digital profiles (standard grayscale) of 16 wells, where (a) is the digital profile of well No. 0 and (b) is the digital profile of well No. 1.

[0076] Figure 11 The diagram shows the digital profiles (Jet color charts) of 16 wells, where (a) is the digital profile of well No. 8 and (b) is the digital profile of well No. 9.

[0077] Figure 12 A schematic diagram showing the comparison of the importance (absolute value) of SHAP in anomaly identification and feature grouping for 16 wells;

[0078] Figure 13 A schematic diagram of the cumulative probability fractions randomly selected from 21 groups of features of sensors from 100 aero engines.

[0079] Figure 14 A schematic diagram of the standardized time axis (cumulative cycle percentage) and standardized indicators (6 randomly selected) for 100 aero engines;

[0080] Figure 15 PCA-reduced feature heatmaps for 100 aero engines;

[0081] Figure 16 A diagram illustrating the importance comparison of SHAP (Shape-Based Analysis and Grouping) for anomaly identification and feature grouping of 100 aero-engines;

[0082] Figure 17 RUL (ground truth) box plots were used to identify 100 aero engines as normal and abnormal. Detailed Implementation

[0083] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0084] This embodiment 1 provides a time-series data anomaly detection method based on convergence trajectory feature analysis, such as... Figure 1 As shown, it includes:

[0085] A. Data Preprocessing and Standardization: Collect multidimensional time-series data of at least one industrial device of the same type and preprocess it to obtain a standardized single-device feature sequence for each industrial device. ,in, For the standardized single-device feature sequence of the i-th industrial equipment, the standardized single-device feature sequences of all the industrial equipment are... Stacked according to device dimension, forming a standardized feature tensor based on all devices. , ,in, For the number of devices, For lifecycle progress points, For feature dimensions;

[0086] B. Construction of Digital Profile of Deep Model Convergence Trajectory: Based on the standardized feature tensor The baseline parameters of the baseline autoencoder model are obtained by training the baseline autoencoder model. and baseline characteristics Based on the baseline parameters Each of the industrial devices is combined with a standardized single-device feature sequence. The baseline autoencoder model is perturbed and trained using single-sample parameters to obtain the convergence trajectory features of the baseline autoencoder model, thereby generating a digital profile matrix corresponding to the industrial equipment, and thus obtaining a standard digital profile library based on all the industrial equipment. ;

[0087] C. Anomaly Detection and Interpretability Analysis: Feature extraction and spatial modeling are performed on the digital profile. Automated anomaly detection is achieved through anomaly measurement algorithms, and feature attribution and root cause analysis are performed in conjunction with interpretable AI technology.

[0088] Next, we will describe the implementation process of each step in detail.

[0089] Step A involves industrial time-series data preprocessing and standardization: This step aims to systematically standardize the preprocessing of multi-source, multi-dimensional, and non-uniformly sampled industrial time-series data collected from industrial sites. This provides a unified and standardized data foundation and structure for subsequent anomaly detection based on deep model convergence trajectory features. The methods employed are applicable to various industrial scenarios (such as oil and gas production, manufacturing monitoring, and energy system operation), demonstrating broad applicability and versatility.

[0090] In some embodiments, the specific steps of A1 are as follows:

[0091] A1. Collect multidimensional time-series data of at least one industrial device of the same type. And time alignment is performed to obtain the industrial time-series data matrix A, where, This represents the d-th dimension index measurement value of device i at the t-th observation time. , , , This indicates the number of observation data points of the device within the historical period; Indicates the dimension of the indicator.

[0092] A1 is actually for data acquisition / definition and alignment, and the specific process is as follows:

[0093] A1.1 Industrial Equipment Data Definition:

[0094] Suppose that the total number of a certain type of equipment in the industrial site is . Unit, each device is numbered The multidimensional indicator data recorded by each device during its historical period are denoted as follows:

[0095] ;

[0096] in:

[0097] : The measured value of the i-th dimension index of the equipment at the i-th observation time;

[0098] : Number of observation data points within the historical period of device i;

[0099] The dimension of the indicator may include, but is not limited to, industrial production-related indicators such as flow rate, power consumption, pressure, and temperature.

[0100] A1.2 Time Alignment and Missing Fill:

[0101] Based on the historical observation time points of all devices, we find the union and sort the sets to obtain a unified observation timeline for all devices. For each device its original observation data Mapped to the full time axis Missing observations are marked with "missing data" ( Fill with (NotaNumber), forming a dimension of Complete industrial time-series data matrix N represents the number of observation data points.

[0102] A2. Calculate all non- NaN The multidimensional time series data dynamic average baseline and dynamic standard deviation baseline and based on the dynamic average baseline and dynamic standard deviation baseline The deviation of each of the aforementioned industrial devices relative to the dynamic baseline at each observation time was obtained. and the deviation The cumulative deviation score for each of the industrial devices is obtained by accumulating along the lifecycle time dimension. .

[0103] A2 is actually a dynamic baseline and probabilistic standardization, and the specific process is as follows:

[0104] A2.1 Definition of Dynamic Baseline:

[0105] Steps and methods:

[0106] For each uniform time t, for each index dimension d, based on all non-NAN device observations, calculate the dynamic mean and standard deviation, and calculate the dynamic mean baseline and dynamic standard deviation baseline:

[0107] Dynamic average baseline:

[0108] ;

[0109] Dynamic standard deviation baseline:

[0110] ;

[0111] in, For the indicator function, the parameter (here is) The value is 1 when it is true, and 0 otherwise. This represents the dynamic average baseline of the d-dimensional index at time t. This represents the dynamic standard deviation baseline of the d-dimensional index at time t. The above processing ensures that the accuracy and robustness of the overall baseline calculation are not affected by factors such as equipment absence or offline status.

[0112] By iterating through all time points and indicator dimensions, the following dimensions can be obtained: Work area dynamic average baseline and dynamic standard deviation baseline .

[0113] Steps to achieve:

[0114] Since A1.2 has already been constructed Complete industrial time-series data matrix This step should directly utilize matrix / vectorized computation tools for industrial-grade batch calculations, avoiding tedious loops and manual statistics. For example, based on... Language Standard Matrix Computation Library ,

[0115] ;

[0116] ;

[0117] This allows for the calculation of the batch mean and standard deviation.

[0118] Generally, other programming languages ​​(such as Matlab, R, Julia, etc.) also have equivalent functionality. To avoid redundancy, the following descriptions of "implementation" in this invention will use Python as an example, and will no longer point out that "it can be implemented in other programming languages".

[0119] A2.2 Calculation of probabilistic deviation scores:

[0120] Steps and methods:

[0121] For each device's deviation relative to the dynamic baseline at each observation time, after standardization and probability mapping, 0.5 is subtracted:

[0122] ;

[0123] in The cumulative function of the standard normal distribution maps the deviation trend of the equipment to... Probability-based standard scores within an interval.

[0124] This step yields the standardized deviations of each device's indicators relative to the dynamic baseline, providing clear, unified, and quantitative indicators adapted to specific field conditions for subsequent anomaly detection. From a physical perspective, Positive / negative indicates that the d index of device i at time t is higher / lower than the average value of the work area.

[0125] Steps to achieve:

[0126] We can continue using industrial-grade matrix calculations to obtain the probabilistic deviation score matrix for the entire work area. Here's an example implementation in Python:

[0127] ;

[0128] in: The cumulative distribution function of the standard normal distribution (can be used) function).

[0129] Generally, if there are any issues during matrix calculations, mainstream scientific computing libraries or programming languages ​​can handle them automatically, without the need for explicit manual processing.

[0130] A.2.3 Cumulative Deviation Trend:

[0131] Steps and methods:

[0132] The probabilistic deviation characteristics are accumulated along the lifecycle time dimension to highlight the long-term deviation trend from the industrial field baseline during equipment operation, and a cumulative deviation score is defined:

[0133] ;

[0134] Steps to achieve:

[0135] Industrial-grade matrix calculations can continue to be used, accumulating along the lifecycle time axis, and the output will still be... This is the matrix, specifically the cumulative deviation trend matrix for the entire work area. For example, using Python code:

[0136] .

[0137] A3. From the aforementioned multidimensional time-series data Obtaining nonnegative time series indicators For the non-negative time series index The cumulative key indicator sequence is obtained by summing the results. Through quantile standardization and unified lifecycle progress interpolation, multi-device time-series data normalization processing is achieved, generating standardized feature tensors adapted for intelligent analysis. .

[0138] In some embodiments, the specific process of A3 is as follows:

[0139] A3.1, For each of the aforementioned industrial devices, the non-negative time series indicators... The cumulative key performance indicators are obtained by summing the results. Simultaneously, the non-negative time series index is obtained. During the observation period of the entire work area The cumulative index total value during the global observation period. Then, for each of the aforementioned industrial devices, the cumulative key performance indicators (KPIs) are... Quantile standardization is performed to obtain the quantile standardization progress. Among them, the cumulative key indicator sequence The specific formula is: The cumulative index total value during the global observation period The specific formula is: The progress of the quantile standardization The specific formula is: , This represents the cumulative sum of key performance indicators for device i at time t. This represents the cumulative index total value of device i during the global observation period. This represents the quantile normalization progress of device i at time t. N represents the actual, maximum observation period with data for device i, and N represents the number of observation data points.

[0140] A3.1 specifically involves cumulative key indicators and quantile standardization, and the steps and methods for implementing it are as follows:

[0141] Depending on the specific industrial application scenario, a non-negative time series indicator (usually output or consumption) with long-term cumulative significance should be selected for this type of equipment. (For example, daily oil production in the oil and gas industry, daily power generation in the power system, daily production in the manufacturing industry, gas consumption of gas turbines, power consumption of electrical equipment, or in some cases, "operation / production / consumption time" itself) as core tracking indicators. It is important to note that for cumulative key indicators in the main sequence... Only for the effective observation period of device i (in Accumulation and maximum value normalization are performed. In the matrix implementation, valid observations are first masked for each device, then normalized / interpolated to avoid... pollute.

[0142] Sum them up cumulatively as a cumulative indicator That is, a cumulative key indicator sequence:

[0143] ;

[0144] In particular, remember for During the equipment observation period across the entire work area The total value above, that is:

[0145] ;

[0146] in, yes The actual, data-rich maximum observation period for device (i.e., device i). and They mean the same thing, representing the cumulative key performance indicator of device i at time k. , These represent the cumulative key indicator sequence of device i at time t and the total cumulative indicator value of device i during the global observation period, respectively.

[0147] Quantile standardization of cumulative key performance indicators for each device:

[0148] ;

[0149] Obviously, The quantile-normalized progress (progress percentage, dimensionless) of device i at time t. , used as a quantile, represents the production / consumption progress ratio of the equipment during the observation period, achieving dimensionless and unified scaling of time-series data for each equipment. N is the total number of moments in the global observation period.

[0150] It is worth emphasizing that the equipment has limitations at certain observation points. When, cumulative sum and quantile standardization Only during the valid observation period (i.e., when the main indicator is not...) (segment) calculation, other times set to This does not affect the uniformity of global interpolation.

[0151] Steps to achieve:

[0152] You can continue to use industrial-grade matrix calculations, taking Python code as an example.

[0153] For the original data matrix of device-time-variable Select a cumulative main indicator The summation is performed directly at the matrix level, row by row (per device):

[0154] ;

[0155] in This represents the cumulative master index matrix for all devices.

[0156] The quantile normalization matrix is ​​then:

[0157] .

[0158] A3.2. Based on the preset number of lifecycle progress points L, construct a standardized lifecycle progress scale shared by all industrial equipment. Then, for each industrial device i, a multidimensional probabilistic deviation feature sequence... Through interpolation function Map it to a uniform progress scale Generate standardized time series:

[0159] The standardized deviation characteristics of each dimension of the industrial equipment i are organized according to a unified schedule dimension to form a standardized time series set of the equipment. Finally, the standardized time series sets of all the aforementioned industrial equipment are stacked according to the equipment dimension to form a standardized feature tensor based on all equipment. , .

[0160] A3.2 is actually for the purpose of standardizing the observation period, and the steps and methods are as follows:

[0161] Obviously, the observation length between devices Inconsistency is highly likely. To ensure the consistency of subsequent neural network model inputs, based on the lifecycle progress percentage obtained in step A.2, the cumulative probability deviation sequence is standardized along the percentile cumulative key indicators and unified to a length of [length missing] through linear or spline interpolation. (like Standardized time series of )

[0162] right The device, in its original timing sequence, forms the following for each variable:

[0163] .

[0164] For equipment i at time t, the cumulative key performance indicators (such as cumulative output, percentage of runtime) are compared with those in A3.1. They mean the same thing.

[0165] Divided into equal parts Using each lifecycle progress point (percentage) as the target, interpolate all variables:

[0166] ;

[0167] ;

[0168] in:

[0169] ; This represents the interpolation function, which can be a linear or spline function interpolation, etc.

[0170] The essence of this formula is to reorganize the original timing characteristics of the i-th device into a set of "unified progress points + multi-dimensional standardized features", so that devices with different runtimes can be aligned and analyzed in the "lifecycle progress" dimension.

[0171] Let i be the i-th device (denoted as i) The standardized feature set after "unified lifecycle progress alignment" This is the l-th unified lifecycle progress point (percentage, for example, when L=5, l=1 corresponds to 0% progress, and l=5 corresponds to 100% progress).

[0172] , represents the standardized deviation cumulative characteristic of the i-th device at the l-th schedule point and in the d-th dimension (a key indicator reflecting the deviation of the device's operation, which has been aligned to the unified schedule).

[0173] L: Preset uniform lifecycle progress points (e.g., divided into 100 progress points, so that all devices can compare according to these 100 progress points).

[0174] The essential function is to calculate the "cumulative progress" based on the device's initial time t. - Deviation characteristics "Correspondence, estimate when the lifecycle progress is..." The deviation characteristic value at that time.

[0175] For interpolation functions (such as linear interpolation and spline interpolation, the core of which is to estimate unknown points based on known points).

[0176] For the i-th device at the original time t, it is a cumulative key indicator (such as cumulative output, percentage of runtime, used to measure the progress of the device's life cycle).

[0177] The original deviation cumulative feature of the i-th device at the original time t and the d-th dimension (the original index before the progress was not aligned).

[0178] The length of the original time-series data for the i-th device (i.e., how many time points were originally run, for different devices) (May be different).

[0179] For example, device i originally ran At that moment, hour, , hour, And so on. Now we need to estimate the progress. The interpolation function will be based on the characteristics of the time. and Calculate the value of . (Example of linear interpolation).

[0180] By using "progress alignment + interpolation," the original characteristics of devices with different runtimes are transformed into a standardized set of comparable features under the same progress dimension. This is a key step in industrial big data preprocessing to "eliminate device differences and unify analysis dimensions."

[0181] Steps to achieve:

[0182] You can continue using industrial-grade matrix calculations. The overall process is as follows:

[0183] Take Python code as an example.

[0184] a. Generation of a unified standard scale

[0185] ;

[0186] in All equipment is shared.

[0187] b.1 For typical industrial scenarios, if D is small, a shallow loop is sufficient. For each variable... :

[0188] ;

[0189] in, This indicates that all devices will be simultaneously and vectorized. Interpolation to Above. Usable. Finally, the data is concatenated into a standardized time series matrix / tensor, corresponding to a standardized lifecycle scale. Shared by all devices:

[0190] ;

[0191] b.2 For certain extremely short scenarios, if D is large, consider introducing... The library further parallelizes processing to avoid loops:

[0192] For each device i and each feature d, in its valid normalized quantile sequence Above, the cumulative probability type bias Align with a unified lifecycle progress point Batch generation of standardized interpolation sequences:

[0193] ;

[0194] The final output is the normalized standard input tensor:

[0195] .

[0196] Each row corresponds to the aligned and normalized curve of all features of a device under a unified lifecycle progress.

[0197] The parallel code can be called directly as follows. Available :

[0198] import numpyasnp

[0199] from numbaimport njit, prange

[0200] @njit(parallel=True)

[0201] defbatch_interp1d_numba(Q_norm,CProbNorm,q_target):

[0202] """

[0203] Q_norm:shape(M,N)

[0204] CProbNorm:shape(M,N,D)

[0205] q_target:shape(L,)

[0206] Returns: S_interp, shape(M, L, D)

[0207] """

[0208] M, N = Q_norm.shape

[0209] _, _, D = CProbNorm.shape

[0210] L = q_target.shape[0]

[0211] S_interp = np.full((M, L, D), np.nan, dtype=np.float32)

[0212] for i in prange(M):

[0213] for d in prange(D):

[0214] x = Q_norm[i, :]

[0215] y = CProbNorm[i, :, d]

[0216] valid = ~np.isnan(x)&~np.isnan(y)

[0217] x_valid = x[valid]

[0218] y_valid = y[valid]

[0219] if x_valid.size<2:

[0220] continue

[0221] idxs = np.argsort(x_valid)

[0222] x_valid = x_valid[idxs]

[0223] y_valid = y_valid[idxs]

[0224] for l in range(L):

[0225] qt = q_target[l]

[0226] if qt<= x_valid[0]:

[0227] S_interp[i, l, d] = y_valid[0]

[0228] elif qt>= x_valid[-1]:

[0229] S_interp[i, l, d] = y_valid[-1]

[0230] else:

[0231] for k in range(x_valid.size - 1):

[0232] if x_valid[k]<= qt<= x_valid[k+1]:

[0233] w = (qt - x_valid[k]) / (x_valid[k+1]- x_valid[k])

[0234] S_interp[i, l, d] = y_valid[k]* (1 - w) + y_valid[k+1] * w

[0235] break

[0236] return S_interp

[0237] The interpolation module supports custom distributed implementations, such as batch processing of all devices / features under frameworks like PySpark, Ray, and Dask, ensuring horizontal scalability and availability even on ultra-large-scale devices and ultra-high-frequency time series (provided sufficient hardware is available).

[0238] Multidimensional time series data after standardization and interpolation ( It can be further processed into a structured tensor, which has the following characteristics:

[0239] Each device corresponds to a length of The lifecycle progress is unified in timing, and each step includes all... Cumulative probability bias of the dimension index.

[0240] Tensors can be directly used as standard inputs for deep learning models (such as PyTorch / TensorFlow), and are suitable for subsequent batch training, inference, and real-time processing of industrial data streams.

[0241] This structured representation not only supports static batch storage, but also facilitates on-demand slicing, batch parallelism, and distributed on-the-fly analysis.

[0242] A4. Design a data interface and output the standardized feature tensor through the data interface.

[0243] The interpolation and standardization steps can be sliced ​​by device or feature dimension to achieve distributed parallel processing, which is compatible with various large-scale industrial internet or industrial big data platforms.

[0244] Regardless of whether batch or streaming implementation is used, the final output normalized tensor can be directly input as a deep model, supporting various industrial system interfaces such as static storage or streaming APIs. In actual industrial system deployments, the above data preprocessing steps (A1–A3) can be flexibly implemented according to the scenario as follows:

[0245] Batch offline analysis scenario: After all steps are completed, tensors can be structured and stored in industrial databases, data lakes, or object storage (such as Parquet / ORC / HDF5 format), providing efficient query and batch loading capabilities for subsequent analysis / backtracking / multi-model experiments.

[0246] Streaming / real-time analytics scenarios: The above processing flow can be integrated into big data pipelines and streaming computing frameworks (such as Spark, Flink, and ETL middleware), supporting on-the-fly batch standardization, interpolation, and tensor quantization, and can be directly connected to downstream deep learning systems without the need for local storage.

[0247] Data interface and format compatibility: final processed data Tensors can be connected through various methods such as DataLoader of mainstream deep learning frameworks (PyTorchDataset, TensorFlowDataset, etc.), RESTful API, message queue or distributed in-memory table, which greatly improves the scalability and engineering applicability of the system.

[0248] Step B is actually the extraction of the digital profile of the convergence trajectory of the depth model.

[0249] This step targets the normalized tensor data output from A.4. This paper proposes a digital profiling feature method based on the convergence of perturbation trajectories using a deep autoencoder (AE) model, applicable to full-scene industrial time-series data modeling and subsequent anomaly detection. This method exhibits broad compatibility with deep autoencoder structures, activation functions, normalization methods, and feature processing methods, and can be adapted to all neural network frameworks that support parameter convergence and autoencoding, such as LSTM, GRU, Transformer, convolutional neural networks (e.g., 1D-CNN), and multilayer perceptrons. It demonstrates high robustness and versatility in industrial scenarios with small sample sizes, heterogeneous structures, and variable lengths.

[0250] The core objective of this step is to fully explore the local perturbation response trajectory of each industrial sample during the deep model convergence process, realize a high-order digital profile representation of the essential behavioral characteristics of the sample, and thereby support anomaly detection, clustering, and multi-scenario industrial intelligent analysis.

[0251] Compared with traditional measurement methods based on Gram matrix, Euclidean distance, or static feature similarity, the "deep autoencoder convergence perturbation trajectory method" proposed in this invention has the following essential advantages:

[0252] Dynamic model space representation: Instead of statically describing the geometric distance or input space similarity between samples, it uses a deep autoencoder model to perturb each sample after the baseline has fully converged, and then describes its response trajectory to perturbations in the model parameter space. In other words, it uses the sample reference "dynamics" as a driving force to reflect its "behavioral sensitivity spectrum" in the complex model space.

[0253] High sensitivity to anomalies and difficult samples: This trajectory can reveal "model-hard" samples, "convergence anomalies" or "hidden high-risk" devices that cannot be captured by ordinary metrics, providing a more essential and discriminative digital fingerprint for tasks such as anomaly detection and fault prediction.

[0254] The method is naturally adapted to small sample sizes and complex industrial data: it actively utilizes overfitting and baseline convergence mechanisms to ensure that the sample perturbation response is mainly dominated by its own behavioral characteristics, effectively avoiding the "random drift" problem under small sample sizes.

[0255] Naturally integrated with deep visual analysis: The final digital profile can be directly expressed as an image, compatible with traditional computer vision (CV) and deep neural network analysis toolkits, greatly enriching the technical system and application scenarios of industrial AI anomaly detection.

[0256] In summary, this step not only realizes the full-process behavioral characterization from "input space" to "deep model parameter space" and then to "convergence trajectory digital profile", but also serves as a link between the previous and subsequent steps. It is highly innovative and has broad engineering adaptability, providing a systematic theoretical and technical foundation for intelligent analysis and anomaly detection of industrial time series data.

[0257] In some embodiments, step B is performed as follows:

[0258] B1. Construct a baseline autoencoder model and use the standardized feature tensor... The baseline autoencoder model is trained iteratively until the loss function converges, yielding the baseline parameters. and based on the baseline parameters Set baseline features , ,in, Represents the encoder function. Represents the baseline characteristics of industrial equipment i. Indicates the number of feature dimensions. Represents the standardized temporal feature tensor of device i.

[0259] The actual process of B1 involves training the baseline autoencoder (AE) model and performing single-sample perturbation, as follows:

[0260] B.1.1 Baseline Autoencoder (AE) Modeling:

[0261] a. Based on the normalized tensor output of step A Construct a unified auto-encoder (AE) model. The output feature dimension of the encoding layer (if the feature is a non-one-dimensional tensor, it is transformed into an ordered one-dimensional feature vector through operations such as flattening / reshaping) is set as . .

[0262] b. Generally, as an autoencoder, the loss function usually considers MSE (MeanSquareError) or MAE (MeanAbsoluteError), etc.

[0263] c. Generally, optimization is performed using typical neural network gradient optimization algorithms such as Adam, AdamW, or SGD.

[0264] d. Iteratively train the autoencoder using full samples from all devices until the training loss is reached. Convergence, obtaining unique baseline parameters Wherein, full convergence can be defined as: satisfying , A preset threshold can be used; or methods such as gradient norm or statistical convergence test can be employed.

[0265] B.1.2 Baseline Feature Acquisition:

[0266] Generally, overfitting is a common phenomenon in deep learning; under small sample conditions, overfitting of neural networks is almost unavoidable. This invention actively causes the autoencoder model to converge or overfit. In this state, the neural network parameters are usually more stable, and the feature descriptions of each sample obtained at this time should gradually converge to the "normal baseline". However, this expectation is not necessary or mandatory. It is only required that the "normal baseline" features are output by the neural network under the same parameter state for each sample, that is, to ensure the same feature mapping method.

[0267] For each device sample i, use the baseline parameters Its standard characteristics are obtained:

[0268] .

[0269] This vector serves as the "normal baseline" feature for each sample, facilitating subsequent comparisons and perturbation response analysis.

[0270] B2. For each of the aforementioned industrial devices, utilize the standardized single-device feature sequence corresponding to that industrial device. As a single sample pair based on the baseline parameters The baseline autoencoder model is trained with perturbation K times, and the perturbation features output at each step of the baseline autoencoder model during training are extracted. Then, the perturbation features described in each step are combined with the corresponding baseline features to obtain the feature difference for each step. The feature differences obtained from K perturbations are stacked in order to form a digital profile matrix of industrial equipment i. Then, a digital profile matrix of all the aforementioned devices. The initial digital profile database is obtained by summarizing the data. .

[0271] B2 is actually the acquisition of single-sample perturbation convergence trajectory and generation of digital profile, and the specific process is as follows:

[0272] B2.1 Baseline Model Replication and Single-Sample Perturbation:

[0273] Baseline parameters of the replicated neural network M copies (one copy for each device or sample, numbered as follows) Using only the data from each device itself (batch_size=1) Perform "single-sample perturbation training" Next time. Here. Typically equal to the dimension of the encoded features This facilitates the formation of a "square digital portrait." It can also be flexibly configured based on the network structure or subsequent analysis objectives, such as computational image processing neural networks. The aspect ratio is not mandatory, so different approaches can be taken. of value.

[0274] Each step is a new epoch, to fully ensure Overfitting to samples (i.e., data from device i each time) The model is trained (batch_size=1, with only 1 sample), and each training iteration is recorded as one epoch; this training is repeated K times (K epochs), forcing the model parameters to be gradually adjusted to specifically adapt to the features of device i (i.e., "overfitting"); the purpose is to adjust the perturbed model parameters. Capture the device's proprietary operating mode (distinguished from the baseline model's general mode).

[0275] Extract the encoder output after each training step:

[0276] ;

[0277] in for Parameters after completing the first perturbation step.

[0278] B2.2 Digital Profile of Converging Perturbation Trajectory (Zhang Quantization):

[0279] For each device The ordered perturbation step feature difference or other corresponding quantities (such as normalization ratio, gradient change) obtained from B2.1 are stacked as follows: The matrix must be kept in order to track the perturbation trajectory and form a comparable "digital profile" based on the convergence:

[0280] With feature difference For example:

[0281] ;

[0282] Let be the feature difference of device i at step j.

[0283] ;

[0284] A digital profile matrix for device i.

[0285] here, Dimensions This yields the initial digital image database:

[0286] .

[0287] B2.3. Standardize the initial digital portrait library to obtain a standard digital portrait library. ,in, This represents the perturbation feature output by the baseline autoencoder model for industrial equipment i after the j-th perturbation training. Let be the model parameters of industrial equipment i after the j-th perturbation training.

[0288] Initial digital portrait database All elements in the feature direction are normalized using the global standard deviation. ) or the whole Normalization (to) (Interval), or other normalization methods can be selected, which can be flexibly adjusted according to the analysis scenario. For example, using Normalization:

[0289] ;

[0290] Ensuring consistent numerical scales among images facilitates subsequent image processing / cluster analysis, resulting in a normalized digital image library, denoted as . :

[0291] ;

[0292] in, It was used The result of global normalization or standardization.

[0293] It is worth emphasizing that, It is essential to ensure that the dimensional order remains consistent. If the full-process matrix-based computation proposed in this invention is adopted, this can be fully guaranteed by ensuring that all tensor indices and their permutation order (sample, perturbation step, feature) uniquely correspond throughout the entire process.

[0294] Feasibility and compatibility description:

[0295] The auto-encoder model used in this method belongs to the neural network paradigm. Its neural network skeleton can not only use LSTM / GRU, but is also compatible with all deep models that support parameter perturbation and feature output (such as 1D-CNN, MLP, Transformer and hybrid neural network skeletons). The neural network architecture that conforms to the auto-encoder model paradigm can be designed according to specific circumstances.

[0296] Step B2 embodies a data-oriented rather than model-oriented approach. Acknowledging the black-box nature of neural networks and the uncertainty of parameter training, the digital profile provided by step B2 is equivalent to a systematic, sample-by-sample scan of the "baseline model's local parameter space." The convergence trajectory of the features of each sample after perturbation of the neural network parameters constitutes its digital profile in the deep model space. This significantly differs from traditional autoencoder models or other temporal feature extraction methods or neural networks (wavelet transform, spectral analysis, LSTM, etc.). This method embodies dynamic, high-order behavioral modeling based on 'perturbation trajectories,' making it adaptable not only to challenging small-sample scenarios but also scalable, meaning it is feasible for large-scale datasets.

[0297] The digital profile generated in this step is not only a "digital profile" in terms of natural semantics, but also typical two-dimensional matrix data, i.e., generalized image data. Subsequent step C can employ various industrial intelligent analysis methods, such as clustering based on computer vision (CV) or traditional matrix analysis, image anomaly detection, and contrastive learning, compatible with multiple AI analysis workflows. This enables downstream automated analysis and decision-making. Furthermore, the digital profile is compatible with the visual feature attribution of interpretable AI tools such as SHAP / LIME.

[0298] As can be seen, step B and all its sub-modules are designed with matrix implementation in mind, supporting distributed / streaming industrial deployment, which greatly improves applicability to large-scale scenarios. The output of the entire step B, like step A, can be sliced ​​and partitioned by device or feature dimension to achieve distributed parallel processing, adapting to various large-scale industrial internet or industrial big data platforms. During system implementation, it is recommended that steps A and B use the same technology stack to reduce technical complexity and improve maintainability.

[0299] Step B1 may employ distributed training methods when necessary (e.g.) Provided or The specific parallel method can be determined based on the actual situation. If the model is small but the amount of data is large, then it is recommended to use parallel methods. Conversely, one can adopt .

[0300] Any processing in step B2 can be computed based on tensor matrix calculations provided by mainstream neural networks. There is no need for manual iteration or loop writing.

[0301] When steps B1 and B2 fully utilize matrix or tensor parallel computing, they can quickly respond to and update the entire AB process when streaming data arrives, facilitating long-term industrial deployment and monitoring.

[0302] The update results can be communicated in real time with DCS, MES, and industrial cloud platforms via API, MQ, industrial data bus, etc.

[0303] Step C is actually anomaly detection and feature interpretability analysis based on digital profiles.

[0304] This step uses the standardized digital image tensor output from step B. Using computer vision (CV), matrix analysis, deep learning, and interpretable AI (XAI) paradigms as core inputs, this system systematically achieves automated anomaly detection, clustering attribution, and feature interpretability analysis of industrial samples. It fully leverages these paradigms, demonstrating high industrial applicability and scalability. This enables end-to-end industrial anomaly detection and interpretability analysis driven by digital profiling.

[0305] In some embodiments, step C is implemented as follows:

[0306] C1. The standard digital portrait library A feature library is obtained by extracting high-order features from the standard digital profile matrix of each of the aforementioned industrial devices. ,in, , To extract feature dimensions based on the standard digital profile matrix, The high-dimensional embedding (i.e., high-dimensional feature) represents the standard digital profile matrix of industrial equipment i.

[0307] Step C1 is actually the modeling and visualization of the digital portrait feature space, and its specific implementation process is as follows:

[0308] C.1.1 Higher-order feature mapping for digital profiling:

[0309] 1) Select one of the following optional image / matrix feature extraction methods to create a digital profile (shape: ...) for each sample. The same processing is applied to the image:

[0310] (a) Classic CV features (such as HOG, SIFT, SURF, LBP, etc.);

[0311] (b) Using pre-trained convolutional neural networks (ResNet, EfficientNet, ViT, etc.) Deep semantic features are extracted, and dimensionality reduction techniques such as Principal Component Analysis (PCA), t-SNE, and UMAP can be optionally used to extract principal feature vectors, resulting in... .

[0312] 2) The results are summarized into a feature library:

[0313] ;in These are the feature dimensions extracted from the digital profile.

[0314] this High-dimensional embeddings of digital profile features can include any CV-compatible representation such as deep networks, PCA, and matrix features.

[0315] C.1.2 Digital Portrait Spatial Visualization:

[0316] Two-dimensional / three-dimensional projection (such as t-SNE / UMAP) can be used to project onto a low-dimensional space to visualize the distribution of industrial equipment / samples, enabling user-friendly human-computer interaction and providing intuitive basis for subsequent clustering, outlier identification and engineering interpretation.

[0317] C2, based on feature library The system calculates the anomaly metric score of the high-order embedding of each industrial device, and determines whether the multidimensional time-series data of the industrial device is abnormal based on the anomaly metric score. Simultaneously, it uses the feature library... The industrial equipment is clustered to obtain clustering results;

[0318] Step C2 involves automated anomaly detection and hard sample mining, and the specific process is as follows:

[0319] The following text will use the same characters. This function represents anomaly detection, clustering criteria, or scores obtained based on digital profiles or their feature extraction. It is also used as the target output of attribution methods such as SHAP.

[0320] 1) Based on feature library Choose an appropriate score or distance metric function. Define an anomaly metric score for each sample, such as:

[0321] a. Distance metrics such as Euclidean distance, cosine similarity, and Mahalanobis distance can be used, as well as cluster boundary scores and depth anomaly scores (such as LOF and IsolationForest).

[0322] b. For any outlier metric score :

[0323] ;

[0324] or

[0325] ;

[0326] Select threshold ,like Then the samples will be automatically labeled. This threshold represents anomalies or difficult samples (i.e., the standard digital profile matrix of device i is anomalous or difficult). This threshold can be adaptively determined using methods such as the 3-sigma principle, quantile method, cluster distance statistics, and Bayesian optimization.

[0327] Here, abnormal or difficult samples can be automatically alerted to the user interface.

[0328] C.2.2 Clustering and Operating Condition Grouping:

[0329] a. to It performs operations including but not limited to DBSCAN, spectral clustering, and hierarchical clustering to achieve automated grouping of industrial equipment / operating conditions.

[0330] b. Clustering results can serve as the basis for engineering hierarchical classification of "healthy operating condition groups", "abnormal operating condition groups", and "latently difficult-to-learn groups".

[0331] C3. Calculate the SHAP contribution of the anomaly metric scores of the standard digital profile matrix of the industrial equipment, and render the SHAP contribution of all the industrial equipment into a two-dimensional heatmap. Analyze the two-dimensional heatmap to perform feature attribution and root cause analysis.

[0332] This step is for digital profiles that have been identified as anomalous / difficult samples through C1, C2 clustering, similarity, or clustering algorithms. or feature vector The system introduces Shapley Additive Explanations, a mainstream interpretable AI theory, to provide mathematical attribution and physical explanation for anomaly detection criteria and sample digital profiling. This is specifically divided into the following sub-steps:

[0333] C.3.1 SHAP Attribution Calculation and Heatmap Interpretation of Digital Profiling:

[0334] a. Mathematical definition of contribution attribution:

[0335] According to digital profiles , If the anomaly detection score or clustering score is defined in C2.1 above, then for any perturbation step... Coding features The SHAP contribution of this component to the final anomaly score is defined using conditional expectation as follows:

[0336] ;

[0337] in, Let be the feature attribution value of industrial equipment i in the j-th perturbation dimension and the k-th perturbation; s (⋅) is the anomaly metric score function. To The new profile matrix after single-point intervention (masking / setting to reference value [e.g., 0 or mean]) is still a two-dimensional array of KxN_feat; A standard digital profile matrix for industrial equipment i.

[0338] b. Heatmaps and interpretability visualization:

[0339] The overall contribution matrix Rendered as a two-dimensional heatmap, it visually reveals which features dominate the abnormal behavior of the samples.

[0340] c. Algorithm Flow and Engineering Implementation:

[0341] In industrial engineering, SHAP value calculation typically employs Monte Carlo sampling, ensemble tree models, or deep network approximation methods, which offer high parallelization and large-scale application capabilities. For ultra-large-scale, high-dimensional scenarios, sampling approximation (such as...) is recommended. , To ensure computational scalability, the specific details of which will not be elaborated here.

[0342] Batch processing is possible Individual samples (devices), all The digital profile components automatically output an attribution matrix for use in subsequent decision-making, alarm, and expert systems.

[0343] C.3.2 Industrial Monitoring Visualization and Expert Interaction:

[0344] All anomaly detection and attribution analysis results can serve as part of the industrial knowledge base's continuous self-learning and adaptive feedback, supporting a closed-loop optimization of production decisions.

[0345] a. Integrated visual interface:

[0346] The SHAP attribution heatmap, digital profile itself, anomaly detection score, and clustering results are integrated and displayed in the industrial production monitoring or intelligent operation and maintenance interface.

[0347] It supports engineers / experts in performing interactive cause analysis, intelligent label correction, feedback optimization, and other operations based on interpretable features.

[0348] b. Production optimization and risk management support:

[0349] Visual attribution reports support data-driven and domain-knowledge-based optimization of production processes, anomaly response classification, and proactive risk management.

[0350] C4 Industrial System Integration and Automation Deployment:

[0351] Similar to steps A and B, the entire analysis process in step C4.1 (feature extraction, clustering, anomaly detection, and interpretability analysis) can be matrixed, batch-streamed, and deployed in parallel on industrial cloud platforms, data lakes, and edge computing nodes.

[0352] Similar to steps A and B, C4.2 supports real-time / batch modes and is compatible with various interfaces such as API, RESTful, message queue, or NoSQL, adapting to various industrial internet and smart manufacturing scenarios.

[0353] C4.3 Although this invention has comprehensively considered numerical instability, anomalies, and adaptability in its design, it is primarily limited to ensuring the robustness and stability of algorithms and processes. When implementing it as industrial software, it is recommended that additional middleware be introduced throughout the system implementation process to support data anomaly tolerance, breakpoint recovery, and encrypted storage, in order to ensure high availability and security for industrial applications.

[0354] The invention will now be illustrated with specific examples.

[0355] Case 1 (Taking the production index of an oil well in an oil field as an example):

[0356] 1. Data preprocessing and standardization:

[0357] The work area has One production facility, namely an oil well. Provided The production data includes daily oil production, daily water production, and daily gas-oil ratio, spanning 1811 days. However, the actual production varies between different equipment (wells).

[0358] Using the standardized time axis and standardized index method described in step A, the cumulative key index (i.e., used for the standardized time axis) is "oil well production" to obtain standardized tensor data. ( ),in .

[0359] The following are the processed results and their preliminary analysis:

[0360] 1.1 Projection Analysis of Cumulative Probability Score Index:

[0361] Each indicator corresponds to a cumulative probability score indicator. Since it is three-dimensional, it can be visualized using certain special methods:

[0362] a. Therefore, it can be mapped to three-dimensional space ( Figure 1 Observing their spatial trajectories, it can be seen that the trajectories of different wells show great differences.

[0363] b. Generally, in The indicators formed On each projection surface, there are also significant differences (such as...). Figures 2-4 ).

[0364] In summary, the analysis Projection between these parameters may be an effective method. This could be indirect evidence of the necessity or effectiveness of data standardization in step A, but performing such semi-quantitative analysis is not the focus or fundamental objective of this invention.

[0365] 1.2 Standardized Time Axis (Cumulative Production Percentage) - Standardized Indicator Analysis:

[0366] a. Generally, a standardized time-standardized indicator curve can be generated for each standardized indicator. For example... Figures 5-7 As shown.

[0367] b. According to Figures 5-7 It can be initially seen that the trends of different devices on different indicators may vary greatly.

[0368] c. Use equipment-standardized time axis-index heatmap ( Figure 8 This enables batch trend comparison and group anomaly observation, allowing direct identification of which wells are different in which life cycle stages.

[0369] Figure 8 The horizontal axis represents the standardized life cycle, and the vertical axis represents the oil well index.

[0370] d. Considering It can be normalized to min-max and mapped to RGB three channels, visualized as a pseudo-color diagram of equipment number (well) - standardized lifecycle (cumulative production percentage) - standardized indicators, such as... Figure 9 As shown; however, on the one hand, this is not applicable to higher-dimensional situations, and on the other hand, it is far less clear and intuitive than step c to determine the differences between devices (wells). Figure 9 The horizontal axis represents the standardized life cycle, and the vertical axis represents the oil well index.

[0371] In summary, the equipment-standardized timeline-indicator heatmap may be an effective means of visualizing equipment indicators, potentially allowing direct observation of abnormal equipment indicators. Similarly, this may be indirect evidence of the necessity or effectiveness of data standardization in step A, but conducting such semi-quantitative analysis is not the focus or fundamental objective of this invention.

[0372] This case demonstrates that the time series data standardization steps employed in this invention effectively enhance the identifiability of abnormal time series data in the oil and gas industry and the stability of normal data.

[0373] 2. Digital profile extraction and analysis of convergence trajectory of deep model:

[0374] After completing the standardized tensor data After construction, this case follows step B, which uses an LSTM autoencoder model to generate a digital profile of the perturbation convergence trajectory for each oil well's time-series standardized index sequence. The specific steps are already clearly stated in the step B text and will not be repeated here.

[0375] Regarding this case:

[0376] 2.1 Convergence Trajectory Digital Profile Formulation Analysis:

[0377] Each matrix represents a "convergence trajectory digital profile" obtained after recursively calculating the perturbation step for that oil well, quantitatively reflecting the global response and characteristic evolution of the model's downhole perturbation behavior. The normalized profile is then presented as a two-dimensional image (black and white grayscale image, Jet color scale heatmap, i.e., ...). Figure 10 , 11 The visualization is performed, with the horizontal axis representing the latent feature dimension and the vertical axis representing the perturbation step. From the resulting image, we can observe:

[0378] a. The digital profiles of most wells show a clear striped structure and regular distribution, indicating good perturbation convergence and stable behavior.

[0379] b. The profiles of some wells show obvious local fluctuation zones or "abnormal areas," which are more prominent under the Jet color scale. This reflects significant anomalies in the model response of these wells, suggesting possible production anomalies or complex operating conditions.

[0380] In fact, the differences in the overall shape, stripe density, and color distribution of digital profiles across different wells provide an intuitive basis for health screening, anomaly detection, and maintenance decisions for equipment groups. This digital profiling method combines high comparability and interpretability, directly displaying differences in equipment behavior in batches, reducing the difficulty of manual analysis, and facilitating subsequent automated processes such as intelligent clustering and alarm discrimination. The image normalization design ensures fair one-to-one comparison between equipment regardless of the size of the model's latent feature space or the number of perturbation rounds, making it suitable for large-scale deployment and application.

[0381] 2.2 Quantitative Analysis of Convergence Trajectory Digital Profile:

[0382] Literature reports that image neural networks have a significant response to contour features in images, a characteristic that makes them particularly suitable for processing images. Figure 10 , Figure 11 This type of digital profile, which may exhibit vertical or horizontal mutations, is used to extract "contour features". Figure 10 , 11 The horizontal axis represents the latent feature dimension, and the vertical axis represents the perturbation step.

[0383] Therefore, as described in step C, this is achieved by using EfficientNet (B0) pre-trained on the ImageNet dataset. The model is mapped to a feature space of dimension 1280, and then reduced to 8 dimensionality using PCA. The results are as follows:

[0384] Anomaly scores (sorted by well number): [0.86555068 0.84843332 2.12561394 5.15019681 3.31919032 0.86555068 2.3860681 25.53061311 0.84843332 8.1375732 6 2.12561394 4.05891974 0.96060261 6.9349321 5 0.8583390 6 5.53061311]

[0385] Automatic alarm (abnormal hash): [3791315]

[0386] Normal well number: [012456810111214]

[0387] It can be seen that the automatic judgment results are largely consistent with the human judgment (2.1.b).

[0388] Based on SHAP analysis, the dominant characteristics of normal / abnormal wells can be identified, such as... Figure 12 As shown.

[0389] It is worth noting that "abnormal" here refers to something that is different from "normal," so it could be either extremely good or extremely bad.

[0390] Case 2. Aircraft engine operating data (C-MAPSS data provided by NASA):

[0391] 1. Data preprocessing and standardization

[0392] Taking FD001.txt from the dataset and parsing it, we can find that the column names are ['unit','cycle','op_set1','op_set2','op_set3','sensor_1','sensor_2','sensor_3','sensor_4','sensor_5','sensor_6','sensor_7','sensor_8','sensor_9','sensor_10','sensor_11','sensor_12','sensor_13','sensor_14','sensor_15','sensor_16','sensor_17','sensor_18','sensor_19','sensor_20','sensor_21'].

[0393] That is, there are a total of Each engine has sensors or indicators. indivual.

[0394] Similarly, the standardized time axis and standardized metric methods described in step A are used, where the cumulative key metric (i.e., used for the standardized time axis) is "cycle" to obtain standardized tensor data. ( ),in .

[0395] The following are the processed results and their preliminary analysis:

[0396] 1.1 Projection Analysis of Cumulative Probability Fractional Index:

[0397] Each indicator can be combined in pairs to form multiple two-dimensional projection surfaces. We selected some typical sensor pairs (such as S9-S16, S2-S14, S10-S8, etc., see...) Figure 13 By plotting the standardized trajectories of all engines in two-dimensional space, the morphological differences between the trajectories of different engines can be clearly seen. The projections of many sensor pairs exhibit strong dispersion or clustering, indicating the similarities and differences in the operating trajectories of different engines.

[0398] This multidimensional-two-dimensional projection visualization reveals the amplifying effect of standardized methods on the differences in group device behavior, and also provides a rich information foundation for subsequent anomaly diagnosis or pattern recognition.

[0399] 1.2 Standardized Time Axis (normalized cycle) - Standardized Indicator Analysis:

[0400] For each sensor channel, a curve of the normalized cycle versus a normalized metric can be plotted (e.g., ...). Figure 14 It can be observed that the trend of the index curves of different engines on different sensor channels varies greatly. Some channels show a high degree of consistency, while others show significant differences. The horizontal axis is the time axis, and the vertical axis is the index.

[0401] Using a standardized timeline ensures consistent data cycles across different devices, facilitating group comparisons and trend clustering.

[0402] Labeling the device number at the end of a two-dimensional curve helps to quickly identify individuals with abnormal trajectories or outside the pattern.

[0403] For all 21 indicator channels, a 7×3 layout is used for batch visualization, allowing direct comparison of the lifecycle evolution of each device for each indicator. Clearly, the data for sensors S18 and S19 are either completely empty or all zeros. However, as emphasized earlier in this invention, such outliers do not affect the overall analysis process.

[0404] When the method of this invention is applied to time series data of the aero-engine industry, it still exhibits a high degree of group differentiation and clarity of individual evolution.

[0405] The standardized cycle alignment and multi-dimensional index probabilistic normalization of tensor representation effectively eliminate external interference such as asynchronous equipment lifecycle and changes in operating conditions.

[0406] By using methods such as projection, curves, and heat maps, we can quantitatively and qualitatively observe the differences between devices and the evolutionary trends of the population.

[0407] For industrial intelligent analysis tasks such as anomaly detection, condition monitoring, and lifespan prediction, the data representation structure of this invention provides a direct and standardized data foundation for subsequent machine learning and intelligent diagnosis.

[0408] 2. Digital profile extraction and analysis of convergence trajectory of deep model:

[0409] Complete the CMAPSS standardized tensor data for aero-engines After the initial setup is completed, following step B, this case study uses an LSTM autoencoder model to generate a digital profile of the perturbation convergence trajectory for each engine's time-series standardized index sequence. The detailed process has been provided in step B and will not be repeated here.

[0410] Regarding this case:

[0411] 2.1 Convergence Trajectory Digital Profile Formulation Analysis:

[0412] Each subgraph is a digital profile of the disturbance convergence of an aero-engine, with the horizontal axis representing the latent feature dimension and the vertical axis representing the number of disturbance steps.

[0413] Overall observation: The digital profiles of most engines show a clear and regular strip structure, indicating that the model has good convergence of perturbations and regular characteristic responses, reflecting the overall consistency and health of the engine group operation.

[0414] Local Anomalies and Differences: A few engine profiles (such as engines 7, 19, and 27) show obvious local anomalies or abrupt changes in color bands, which are more pronounced under the Jet color scale. This reflects significant differences in the characteristic responses of these engines under model perturbation, suggesting possible special cases such as degradation, early failure, or complex operating conditions.

[0415] Comparative Value: The differences in profile morphology, strip density, and color distribution among different engines provide an intuitive and efficient basis for equipment group health screening, anomaly detection, life assessment, and operation and maintenance decisions. The digital profiling method combines high comparability and interpretability, enabling batch visualization of temporal behavioral differences between equipment, greatly improving the efficiency of manual inspections and intelligent diagnostics.

[0416] Normalization ensures fairness: The global normalization design ensures that all device profiles can be compared fairly one-to-one, regardless of the size of the model's latent feature space or the perturbation step size, making it suitable for large-scale group health analysis and automated alarms.

[0417] Overall, the standardized time-series tensor combined with deep model digital profiling method proposed in this invention effectively improves the identifiability of abnormal evolution and the comparability of health trends of aero-engines (and other industrial equipment), and greatly enhances the automation capability and intelligent operation and maintenance level of group equipment analysis.

[0418] 2.2 Quantitative Analysis of Convergence Trajectory Digital Profile:

[0419] Research on image feature extraction based on deep neural networks has shown that convolutional neural networks (such as EfficientNet) have good responsiveness to structural changes and contour features in digital profiles. For this case, following step C, EfficientNet-B0, pre-trained on the ImageNet dataset, was used to extract high-dimensional features (1280 dimensions) from the digital profiles (Jet color scale heatmaps) of each engine. These features were then reduced to 16 dimensions using PCA for subsequent analysis and visualization.

[0420] Anomaly detection and analysis:

[0421] Based on the 32-dimensional principal components after PCA dimensionality reduction, Euclidean distance is used to cluster and detect anomalies in the feature distributions of all 100 engines, obtaining the minimum feature distance for each engine, and automatically identifying abnormal engines accordingly. Specifically:

[0422] The feature matrix heatmap after PCA dimensionality reduction is shown below. Figure 15 As shown, the PCA1 features appear to vary greatly between samples, with the horizontal axis representing the engine number and the vertical axis representing the PCA principal components.

[0423] Automatic alarm (abnormal engine number): A total of 4 abnormal engines were detected, including: [19, 42, 53, 75].

[0424] In this scenario (100 samples), it becomes very difficult to detect anomalies in digital images manually. However, using the automatic alarm process mentioned in the patent, the detected sample lifetime is indeed quite low, such as... Figure 17 It is evident that digital profiling combined with automatic deep feature extraction has proven effective in scenarios such as aircraft engine health status screening and batch anomaly detection.

[0425] SHAP feature attribution and group comparison (e.g.) Figure 16 ):

[0426] a. Distribution of dominant feature dimensions

[0427] Abnormal engine (red):

[0428] The absolute SHAP values ​​of most dimensions are very close to 0, with only a few dimensions having significantly amplified absolute SHAP values ​​(such as PCA_1, PCA_2, PCA_10, PCA_26, and PCA_32).

[0429] Among them, PCA_1, PCA_2, PCA_10, PCA_26 and PCA_32 are the most discriminative features (the red bars protrude to the right and the mean exceeds 0.1, with PCA_10 even approaching +0.26).

[0430] The SHAP mean of a few features is negative (left side), but the largest absolute value is still the red bar on the right.

[0431] Normal engine (blue):

[0432] The mean SHAP of each feature dimension is close to 0, with a maximum of only about 0.03, and all are concentrated in the range of ±0.01 to 0.03, indicating that normal samples do not have obvious dominant features, and the principal components are more evenly and dispersedly distributed.

[0433] b. Feature interpretability conclusions

[0434] PCA_10, PCA_1, PCA_2, PCA_26, and PCA_32 are the most important distinguishing features of abnormal engines.

[0435] These principal components are given greater SHAP importance in the anomalous group, suggesting that the anomalous digital profile (after EfficientNet+PCA mapping) shows significant differences from normal units in the above feature directions; the anomalous engine is likely to be clearly distinguished from most units after its "pattern" in the original sensor time series or digital profile features is mapped to the above principal components.

[0436] The remaining principal components (such as PCA_5~PCA_25) have very small SHAP mean values ​​between groups and do not contribute much.

[0437] c. Intuitive conclusions from SHAP analysis:

[0438] Normal group: None of the PCA features are strongly dominant, indicating that the high-dimensional features of the profile of a healthy engine are relatively "average" and there are no obvious abnormalities.

[0439] Outlier group: The distribution is extremely unbalanced, with a very small number of PCA feature weights being high, showing an abnormal "concentration".

[0440] The group differences are highly concentrated in a very small number of dimensions—this is precisely the basis for the interpretability of intelligent diagnosis in digital profiling.

[0441] In summary, the aero-engine group health screening method based on standardized digital profiling and deep feature analysis has achieved a closed-loop process from qualitative to quantitative analysis and from visualization to automation, providing a paradigm for intelligent diagnosis of large-scale industrial equipment.

[0442] Example 2 also discloses a time-series data anomaly detection system based on convergence trajectory feature analysis, used to execute the time-series data anomaly detection method based on convergence trajectory feature analysis described in Example 1, including:

[0443] The data preprocessing and standardization module is used to collect multidimensional time-series data of at least one industrial device of the same type and preprocess it to obtain a standardized single-device feature sequence for each industrial device. ,in, For the standardized single-device feature sequence of the i-th industrial equipment, the standardized single-device feature sequences of all the industrial equipment are... Stacked according to device dimension, forming a standardized feature tensor based on all devices. , ,in, For the number of devices, For lifecycle progress points, For feature dimensions;

[0444] The deep model convergence trajectory digital profile construction module is used to construct digital profiles based on the standardized feature tensor. The baseline parameters of the baseline autoencoder model are obtained by training the baseline autoencoder model. and baseline characteristics Based on the baseline parameters The standardized single-device feature sequence of each of the aforementioned industrial devices is respectively... The baseline autoencoder model is perturbed and trained using a single sample to obtain its convergence trajectory features, thereby generating a digital profile matrix corresponding to the industrial equipment, and thus obtaining a standard digital profile library based on all the industrial equipment. ;

[0445] The anomaly detection and interpretability analysis module is used to extract features and spatially model the digital profile, achieve automated anomaly detection through anomaly measurement algorithms, and combine interpretable AI technology to perform feature attribution and root cause analysis.

[0446] In summary, the solution of this invention reduces the impact of environmental fluctuations on detection through multi-source dynamic normalization, and achieves macroscopic identification and root cause attribution of abnormal events through autoencoder trajectory image clustering, which helps reduce false alarms and missed alarms and improve detection reliability. Furthermore, the integrated interpretability analysis module facilitates maintenance personnel in quickly locating the cause of faults, meeting the industry's demand for visualized decision-making. The solution supports batch processing, tensor computation, and stream processing, aligning with industrial big data (such as Spark, TensorFlow, etc.) and edge computing architectures, and is easy to deploy in large-scale equipment environments. In conclusion, this invention has significant industrial engineering application value and promotion potential, and is applicable to equipment condition monitoring and fault prediction in multiple fields such as energy, power, and manufacturing.

[0447] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for anomaly detection in time-series data based on convergence trajectory feature analysis, characterized in that, include: Data preprocessing and standardization: Collect multidimensional time-series data of at least one industrial device of the same type and preprocess it to obtain a standardized single-device feature sequence S for each industrial device. i , of which S i For the standardized single-device feature sequence of the i-th industrial equipment, S represents the standardized single-device feature sequence of all the industrial equipment. i Stacked according to device dimension, forming a standardized feature tensor based on all devices. Where M is the number of devices, L is the number of lifecycle progress points, and D is the feature dimension; Digital profile construction based on the convergence trajectory of the deep model: based on the standardized feature tensor The baseline parameters Θ of the baseline autoencoder model are obtained by training the baseline autoencoder model. * and baseline characteristics Based on the baseline parameter Θ * The standardized single-device feature sequence S of each of the aforementioned industrial devices is respectively... i The baseline autoencoder model is perturbed and trained using a single sample to obtain its convergence trajectory features, thereby generating a digital profile matrix corresponding to the industrial equipment, and thus obtaining a standard digital profile library based on all the industrial equipment. The specific process is as follows: Construct a baseline autoencoder model and use the standardized feature tensor The baseline autoencoder model is trained iteratively until the loss function converges, yielding the baseline parameters Θ. * And based on the baseline parameter Θ * Set baseline features Among them, f enc (*) indicates the encoder function. N represents the baseline characteristics of industrial equipment i. feat Indicates the number of feature dimensions. Represents the standardized temporal feature tensor of device i; For each of the aforementioned industrial devices, the standardized single-device feature sequence S corresponding to that industrial device is used. i As a single sample pair based on the baseline parameter Θ * The baseline autoencoder model is trained with perturbation K times, and the perturbation features output at each step of the baseline autoencoder model during training are extracted. Then, the perturbation features described in each step are combined with the corresponding baseline features to obtain the feature difference for each step. The feature differences obtained from K perturbations are stacked sequentially to form a digital profile matrix G of industrial equipment i. i The digital profile matrix G of all the aforementioned devices i The initial digital profile database is obtained by summarizing the data. For the initial digital portrait library Standardization processing yields a standard digital portrait library. in, Θ represents the perturbation feature output by the baseline autoencoder model for industrial equipment i after the j-th perturbation training. (i,j) Let i be the model parameters of industrial equipment after the j-th perturbation training. Anomaly detection and interpretability analysis: Feature extraction and spatial modeling are performed on the digital profile, automated anomaly detection is achieved through anomaly measurement algorithms, and feature attribution and root cause analysis are performed in combination with interpretable AI technology.

2. The time-series data anomaly detection method based on convergence trajectory feature analysis according to claim 1, characterized in that, Multidimensional time-series data of at least one industrial device of the same type are collected and preprocessed to obtain a standardized single-device feature sequence S for each industrial device. i The specific process is as follows: A1. Collect multidimensional time-series data of at least one industrial device of the same type. And time alignment is performed to obtain the industrial time-series data matrix A, where, Let d represent the d-th dimension index measurement value of device i at the t-th observation time, i = 1, ..., M, t = 1, ..., N i , d=1,…,D,N i D represents the number of observation data points for device i within the historical period; D represents the index dimension. A2. Calculate all non-NaN values ​​in the multidimensional time series data. dynamic average baseline and dynamic standard deviation baseline And based on the dynamic average baseline and dynamic standard deviation baseline The deviation of each of the aforementioned industrial devices relative to the dynamic baseline at each observation time was obtained. and the deviation The cumulative deviation score for each of the industrial devices is obtained by accumulating along the lifecycle time dimension. A3. From the aforementioned multidimensional time-series data Obtaining nonnegative time series indicators For the non-negative time series index The cumulative key indicator sequence Q is obtained by summing the results. i,t Through quantile standardization and unified lifecycle progress interpolation, multi-device time-series data normalization processing is achieved, generating standardized feature tensors adapted for intelligent analysis. A4. Design the data interface and convert the standardized feature tensor... Output is made through the data interface.

3. The time-series data anomaly detection method based on convergence trajectory feature analysis according to claim 2, characterized in that, The dynamic average baseline The formula is: The dynamic standard deviation baseline The formula is: Where isNaN(·) is an indicator function, which is 1 when the parameter is NaN, and 0 otherwise; This represents the dynamic average baseline of the d-dimensional index at time t. This represents the baseline of the dynamic standard deviation of the d-dimensional index at time t; The deviation The formula is: in, Φ(·) represents the deviation of the d-dimensional index of industrial equipment i at time t from the relative dynamic baseline, and is the cumulative function of the standard normal distribution. The cumulative value of deviation fraction The formula is: in, denoted as the cumulative deviation score of the d-dimensional index of industrial equipment i at time t, and N represents the number of observation data points.

4. The time-series data anomaly detection method based on convergence trajectory feature analysis according to claim 2, characterized in that, The specific process for A3 is as follows: A3.1, For each of the aforementioned industrial devices, the non-negative time series indicators... The cumulative key indicator Q is obtained by summing the results. i,t Simultaneously, the non-negative time series index is obtained. During the observation period of the entire work area The cumulative index total value during the global observation period. Then, for each of the aforementioned industrial devices, the cumulative key indicator sequence Q is... i,t Quantile standardization is performed to obtain the quantile standardization progress. Among them, the cumulative key indicator sequence Q i,t The specific formula is: The cumulative index total value during the global observation period The specific formula is: Quantile standardization progress The specific formula is: Q i,t This represents the cumulative sum of key performance indicators for device i at time t. This represents the cumulative index total value of device i during the global observation period. This represents the quantile normalization progress of device i at time t. N represents the number of observation data points; A3.

2. Based on the preset number of lifecycle progress points L, construct a standardized lifecycle progress scale shared by all industrial equipment. Then, for each industrial device i, a multidimensional probabilistic deviation feature sequence is generated. It is mapped to a uniform progress scale using the interpolation function Interp(·). Generate standardized time series: The standardized deviation characteristics of each dimension of the industrial equipment i are organized according to a unified schedule dimension to form a standardized time series set of industrial equipment i. Finally, the standardized time series sets of all the industrial equipment are stacked according to the equipment dimension to form a standardized feature tensor based on all equipment.

5. The time-series data anomaly detection method based on convergence trajectory feature analysis according to claim 1, characterized in that, The standard digital portrait library The formula is: These represent the smallest and largest digital profile matrices in the initial digital profile database, respectively. A standard digital profile matrix representing industrial equipment i.

6. The time-series data anomaly detection method based on convergence trajectory feature analysis according to claim 1, characterized in that, The specific process of feature extraction and spatial modeling of the digital profile, automated anomaly detection through anomaly measurement algorithms, and feature attribution and root cause analysis combined with interpretable AI technology is as follows: For the standard digital portrait library A feature library is obtained by extracting high-order features from the standard digital profile matrix of each of the aforementioned industrial devices. in, N CV h is the feature dimension extracted from the standard digital profile matrix. i High-dimensional embedding of the standard digital profile matrix representing industrial equipment i; Based on feature library Calculate the anomaly metric score of the high-dimensional embedding of each industrial device, and determine whether the multi-dimensional time-series data of the industrial device is abnormal based on the anomaly metric score. Simultaneously, based on the feature library... Clustering of the industrial equipment yields the clustering results; The SHAP contribution of the anomaly metric scores of the standard digital profile matrix of the industrial equipment is calculated respectively, and the SHAP contribution of all the industrial equipment is rendered into a two-dimensional heat map. Feature attribution and root cause analysis are performed by analyzing the two-dimensional heat map.

7. The time-series data anomaly detection method based on convergence trajectory feature analysis according to claim 6, characterized in that, The specific formula for the SHAP contribution of the anomaly metric score is as follows: Where, φ i,j,k Let be the feature attribution value of industrial equipment i in the j-th perturbation dimension and the k-th perturbation; s(·) is the anomaly metric score function. To A new profile matrix after single-point intervention; A standard digital profile matrix for industrial equipment i.

8. The time-series data anomaly detection method based on convergence trajectory feature analysis according to claim 6, characterized in that, After obtaining the feature library Then, the feature library Projecting onto a low-dimensional space enables the distributed visualization of multi-dimensional time-series data of industrial equipment.

9. A time-series data anomaly detection system based on convergence trajectory feature analysis, used to execute the time-series data anomaly detection method based on convergence trajectory feature analysis as described in any one of claims 1-8, characterized in that, include: The data preprocessing and standardization module is used to collect multidimensional time-series data of at least one industrial device of the same type and preprocess it to obtain a standardized single-device feature sequence S for each industrial device. i , of which S i For the standardized single-device feature sequence of the i-th industrial equipment, S represents the standardized single-device feature sequence of all the industrial equipment. i Stacked according to device dimension, forming a standardized feature tensor based on all devices. Where M is the number of devices, L is the number of lifecycle progress points, and D is the feature dimension; The deep model convergence trajectory digital profile construction module is used to construct digital profiles based on the standardized feature tensor. The baseline parameters Θ of the baseline autoencoder model are obtained by training the baseline autoencoder model. * and baseline characteristics Based on the baseline parameter Θ * The standardized single-device feature sequence S of each of the aforementioned industrial devices is respectively... i The baseline autoencoder model is perturbed and trained using a single sample to obtain its convergence trajectory features, thereby generating a digital profile matrix corresponding to the industrial equipment, and thus obtaining a standard digital profile library based on all the industrial equipment. The anomaly detection and interpretability analysis module is used to extract features and spatially model the digital profile, achieve automated anomaly detection through anomaly measurement algorithms, and combine interpretable AI technology to perform feature attribution and root cause analysis.

Images