A multi-industry adaptive technical basis detection method and system

Abstract

This invention relates to the field of anomaly detection technology, specifically to a multi-industry-adaptive technical foundation detection method and system. It is used to perform basic state detection on multivariate operational data generated by equipment, components, process units, or production line links in multiple industry scenarios. The method involves: offline, acquiring historical datasets with state annotations, establishing a candidate detection algorithm library, and training, validating, and evaluating the performance of each candidate algorithm to form associated metadata; extracting univariate statistical features, multivariate correlation features, anomaly distribution features, and overall structural features from each historical dataset, and obtaining a unified meta-feature representation through screening, aggregation, and embedding transformation; training a multi-output performance prediction model using the unified meta-feature representation as input and the candidate algorithm performance vector as output; online, extracting the meta-features to be detected from the industry object to be detected and inputting them into the model to obtain the predictive detection performance of each candidate algorithm, selecting a single target detection component or determining multiple algorithms and their integrated weights, and outputting at least one of anomaly score and state label to obtain the anomaly state detection result. This invention can solve the problem that anomaly detection algorithms are difficult to adapt quickly and accurately based on human experience when there are diverse types of industry objects, large differences in operational data structures, and frequent changes in operating conditions in multi-industry scenarios.

Description

Technical Field

[0001] This invention relates to the field of anomaly detection technology, and in particular to a technically fundamental detection method and system adaptable to multiple industries. Background Technology

[0002] With the advancement of industrial digitalization and real-time data acquisition capabilities, numerous devices and processes in manufacturing, energy, transportation, and process industries continuously generate sensing, control, and operational condition-related data. Anomaly detection based on machine learning has become a crucial technical means for identifying equipment degradation, process deviations, and potential faults. Anomaly detection models can output anomaly scores or status labels to characterize the degree to which data points deviate from normal operating behavior. However, different algorithms exhibit significantly different detection performance on different datasets; no single algorithm is consistently optimal across various tasks. Algorithm selection and parameter adaptation typically rely on collaboration between data experts and domain experts, involving data understanding, feature construction, algorithm evaluation, and model validation, resulting in high computational and labor costs. Especially in multivariate time-series data scenarios, different variables may exhibit lagged correlations, collaborative changes, or information transmission relationships, and anomalies may present as point-like, contextual, or collectively distributed characteristics. These data properties affect the applicability of detection algorithms. Furthermore, the adaptability of existing detection models may decrease when the industry, operating load, process procedures, or environmental conditions change. Therefore, how to automatically determine a suitable basic detection algorithm based on the characteristics of the data itself without conducting a large number of manual experiments for each new scenario has become an urgent technical problem to be solved in the detection of abnormal states in multiple industries. Summary of the Invention

[0003] In view of the above technical problems, the present invention provides a technical basis detection method and system for multi-industry adaptation, which aims to solve the problem that anomaly detection algorithms are difficult to adapt quickly and accurately based on human experience when there are diverse types of industrial objects, large differences in operating data structures, and frequent changes in operating conditions in multi-industry scenarios.

[0004] Other features and advantages of the invention will become apparent from the following detailed description, or may be learned in part by practice of the invention.

[0005] According to one aspect of the present invention, a technology foundation detection method for multi-industry adaptation is proposed. The method is used to perform basic state detection on multivariate operational data generated by industrial objects in multiple industrial scenarios. The industrial objects include at least one of equipment, components, process units, and production line links. The multivariate operational data includes at least one of sensor data, control data, operating condition data, and production context data. The method includes: During the offline construction phase, a set of historical datasets with state annotations from different industry objects is obtained, a candidate detection algorithm library is established, and the candidate detection algorithms in the candidate detection algorithm library are trained, validated, and evaluated on the set of historical datasets. The data feature descriptions, candidate detection algorithm identifiers, candidate detection algorithm parameter configurations, and candidate detection algorithm detection performance corresponding to each historical dataset are associated to form metadata. In the meta-feature extraction stage, meta-features are extracted for each historical dataset. The meta-features include at least features used to characterize the statistical properties of the univariate sequence itself, features used to characterize the correlation between multivariate sequences, features used to characterize the abnormal distribution pattern, and features used to characterize the overall structure of the dataset. The meta-features are then filtered, aggregated, and embedded to transform historical datasets from different sources into a unified meta-feature representation. During the meta-model training phase, a unified meta-feature representation is used as input, and a performance vector composed of the detection performance of each candidate detection algorithm on the same historical dataset is used as the output target. The multi-output performance prediction model is trained so that it learns the mapping relationship between meta-features and the detection performance of multiple candidate detection algorithms. During the training process, the model parameters are adjusted according to the deviation between the predicted performance vector and the true performance vector. In the online application phase, the multivariate running data of the industry object to be detected is obtained, and the meta-features to be detected are extracted and transformed in the same way as in the offline construction phase. The meta-features to be detected are input into the multi-output performance prediction model to obtain the predicted detection performance of each candidate detection algorithm in the candidate detection algorithm library for the industry object to be detected. In the algorithm determination stage, an algorithm selection strategy is executed based on the predictive detection performance of each candidate detection algorithm to determine the target detection component from the candidate detection algorithm library; when there is a candidate detection algorithm whose predictive detection performance meets the advantage condition, the candidate detection algorithm is determined as the target detection component; when multiple candidate detection algorithms meet the integration condition, the integration weight is determined for the corresponding multiple candidate detection algorithms based on the predictive detection performance, and the multiple candidate detection algorithms with integration weights are determined as the target detection component. In the detection output stage, the target detection component is applied to the multivariate operating data of the industry object to be detected, and at least one of anomaly score and status label is generated for the data points in the multivariate operating data. The abnormal status detection result of the industry object to be detected is output according to at least one of the anomaly score and status label.

[0006] Furthermore, during the offline construction phase, when training, validating, and evaluating the performance of the candidate detection algorithms, a training data portion and a validation data portion are established for each historical dataset. This allows each candidate detection algorithm to complete model construction based on the training data portion under the corresponding candidate detection algorithm parameter configuration, and obtain detection results based on the validation data portion. The detection performance is determined based on the consistency between the detection results and the state labels. The detection performance is a comprehensive evaluation value that simultaneously reflects precision and recall, so that the impact of false positives and false negatives are jointly included in the performance vector. In addition, the metadata stores the performance vectors corresponding to different historical datasets in the same candidate detection algorithm arrangement order, enabling the multi-output performance prediction model to simultaneously learn the relative fit relationship of multiple candidate detection algorithms on the same dataset.

[0007] Furthermore, the candidate detection algorithm library includes multiple anomaly detection algorithms of different types capable of handling multivariate operational data. These anomaly detection algorithms at least cover histogram-based outlier scoring algorithms, principal component projection-based detection algorithms, clustering-based local outlier factor detection algorithms, lightweight online anomaly detection algorithms, connection-based distribution relationship-based detection algorithms, local density neighborhood-based detection algorithms, single-class boundary learning-based detection algorithms, random isolation tree-based detection algorithms, nearest neighbor distance-based detection algorithms, feature bagging-based detection algorithms, and connection structure outlier factor-based detection algorithms. When the candidate detection algorithm library is established, it retains information on the input data type compatibility, parameter configuration range, whether it supports multivariate sequence processing, and whether it is suitable for real-time deployment for each candidate detection algorithm, ensuring that the algorithm selection strategy has a transferable candidate basis across different industry objects.

[0008] Furthermore, the features used to characterize the correlations between multivariate sequences include autoregressive features, cross-correlation features, co-variation features, transitive information features, and dominant frequency features. Specifically, the autoregressive features are obtained by describing the dependence between the current sequence value and historical lagged sequence values; the cross-correlation features are obtained by describing the similarity of different variable sequences under lagged shifts; the co-variation features are obtained by describing the common changing trends of different variable sequences; the transitive information features are obtained by describing the uncertainty reduction effect of the historical state of one variable sequence on the future state of another variable sequence; and the dominant frequency features are obtained by identifying the dominant oscillation frequency of the energy distribution of the variable sequence. Moreover, the correlation features obtained from different variable sequences and combinations of different variable sequences are statistically aggregated for mean, extreme values, dispersion, and distribution patterns to form meta-features capable of describing the overall temporal relationships of multivariate operational data.

[0009] Furthermore, the features characterizing the abnormal distribution pattern are generated from a historical dataset with state annotations. These features include total anomaly percentage, point-like anomaly percentage, collective anomaly percentage, anomaly duration, and normal duration. Specifically, the total anomaly percentage characterizes the overall occurrence of anomalous data points in the historical dataset; the point-like anomaly percentage characterizes the degree to which a single data point deviates from its normal state; the collective anomaly percentage characterizes the degree to which consecutively related data points deviate from their normal state; the anomaly duration is obtained by identifying consecutive anomaly-annotated segments and calculating their lengths; and the normal duration is obtained by identifying consecutive normal-annotated segments and calculating their lengths. In the online application phase, the multivariate operational data of the target industry object does not require pre-obtained state annotations as necessary input, and the candidate detection algorithm adaptation judgment is completed through the multi-output performance prediction model.

[0010] Furthermore, the features used to characterize the overall structure of the dataset include clustering quality description features, data complexity description features, information theory description features, benchmark learner description features, model structure description features, and statistical distribution description features. Specifically, the clustering quality description features reflect the separability of samples in the feature space; the data complexity description features reflect class imbalance, boundary samples, distance between neighboring samples, and feature overlap; the information theory description features reflect entropy, mutual information, joint entropy, and attribute noise; the benchmark learner description features are obtained through the performance of a lightweight base model on historical datasets; the model structure description features are obtained through the homogeneity of tree structure nodes and leaf nodes, branch morphology, and feature importance; and the statistical distribution description features are obtained through mean, variance, skewness, kurtosis, correlation, covariance, and sparsity. Thus, the meta-features simultaneously contain information about the internal structure of the data, attribute relationships, and the basic learning difficulty.

[0011] Furthermore, the meta-feature extraction stage also includes a meta-feature preprocessing step, which processes missing meta-features, meta-features with inconsistent dimensions, meta-features with abnormal ranges, and redundant meta-features, and inputs the processed meta-features into the embedding conversion module. The embedding conversion module uses at least one of autoencoder representation learning, principal component projection, and random neighborhood embedding to convert high-dimensional meta-features into low-dimensional representations suitable for the learning of the multi-output performance prediction model. In the online application stage, the meta-features to be detected are subjected to the same preprocessing rules and embedding conversion rules as in the offline construction stage, so that the mapping relationship obtained by offline training can be directly used for new industry objects.

[0012] Furthermore, the multi-output performance prediction model adopts a multi-output regression structure, which uses the detection performance of each candidate detection algorithm as an output dimension and establishes a nonlinear mapping between meta-feature representations and performance vectors through an ensemble regression learner. The ensemble regression learner consists of multiple base learners, each trained on different sample subsets and different meta-feature subsets. During prediction, the outputs of each base learner are aggregated to obtain the predicted detection performance of each candidate detection algorithm. The training objective of the multi-output performance prediction model is to reduce the overall deviation of the predicted performance vector from the true performance vector, so that the multi-output performance prediction model not only predicts the absolute performance of a single candidate detection algorithm, but also predicts the relative superiority or inferiority relationship between different candidate detection algorithms.

[0013] Furthermore, the algorithm selection strategy includes a sorting and filtering step, a threshold exclusion step, an advantage judgment step, and an integration and construction step. In the sorting and filtering step, candidate detection algorithms are sorted from high to low according to their predictive detection performance. In the threshold exclusion step, candidate detection algorithms with predictive detection performance below a preset lower limit are excluded. In the advantage judgment step, when the top-ranked candidate detection algorithm meets a preset advantage condition relative to other candidate detection algorithms, a single candidate detection algorithm is determined as the target detection component. In the integration and construction step, when multiple candidate detection algorithms jointly meet a preset integration condition, they are sequentially included until the cumulative performance meets a preset upper limit condition, and the integration weight is obtained by normalizing the predictive detection performance of each candidate detection algorithm. When the target detection component is composed of multiple candidate detection algorithms, the member anomaly scores output by each candidate detection algorithm are weighted and fused, and the status label is generated based on the fused anomaly scores. When the workload, process, component status, and environmental conditions of the industry object change and cause the meta-feature to be detected to meet the reselection trigger condition, the online application stage and the algorithm determination stage are re-executed to update the target detection component.

[0014] According to another aspect of the present invention, a multi-industry-adaptive technology foundation detection system is provided. The system is used to perform basic state detection on multivariate operational data generated by industrial objects in multiple industrial scenarios. The industrial objects include at least one of equipment, components, process units, and production line links. The multivariate operational data includes at least one of sensor data, control data, operating condition data, and production context data. The system includes: The offline construction module is used to acquire a set of historical datasets with state annotations from different industry objects, establish a candidate detection algorithm library, train, verify and evaluate the performance of the candidate detection algorithms in the candidate detection algorithm library on the set of historical datasets, and associate the data feature descriptions, candidate detection algorithm identifiers, candidate detection algorithm parameter configurations and candidate detection algorithm detection performance corresponding to each historical dataset to form metadata; The meta-feature extraction module is used to extract meta-features for each historical dataset. The meta-features include at least features that characterize the statistical properties of the univariate sequence itself, features that characterize the correlation between multivariate sequences, features that characterize the abnormal distribution pattern, and features that characterize the overall structure of the dataset. The module also filters, aggregates, and embeds the meta-features to convert historical datasets from different sources into a unified meta-feature representation. The meta-model training module is used to train a multi-output performance prediction model by taking a unified meta-feature representation as input and a performance vector composed of the detection performance of each candidate detection algorithm on the same historical dataset as the output target. The multi-output performance prediction model learns the mapping relationship between meta-features and the detection performance of multiple candidate detection algorithms, and adjusts the model parameters according to the deviation between the predicted performance vector and the true performance vector during the training process. The online application module is used to acquire the multivariate running data of the industry object to be detected, extract and transform the meta-features to be detected in the same way as the offline construction module, input the meta-features to be detected into the multi-output performance prediction model, and obtain the prediction detection performance of each candidate detection algorithm in the candidate detection algorithm library for the industry object to be detected. The algorithm determination module is used to execute an algorithm selection strategy based on the predictive detection performance of each candidate detection algorithm, and determine the target detection component from the candidate detection algorithm library; when there is a candidate detection algorithm whose predictive detection performance meets the advantage condition, the candidate detection algorithm is determined as the target detection component; when multiple candidate detection algorithms meet the integration condition, the integration weight is determined for the corresponding multiple candidate detection algorithms based on the predictive detection performance, and the multiple candidate detection algorithms with integration weights are determined as the target detection component. The detection output module is used to apply the target detection component to the multivariate operating data of the industry object to be detected, generate at least one of anomaly score and status label for the data points in the multivariate operating data, and output the abnormal status detection result of the industry object to be detected based on at least one of the anomaly score and status label.

[0015] The technical solution of the present invention has the following beneficial effects: This invention transforms the algorithm selection process from manual experience-based screening to automatic adaptation judgment based on historical task mapping relationships by offline construction of a candidate detection algorithm library, extraction and unified representation of historical dataset metadata features, training a multi-output performance prediction model, and online prediction of the performance of each candidate detection algorithm based on the metadata features of the data to be detected. This approach simultaneously considers factors affecting detection performance, such as univariate statistical properties, multivariate correlations, abnormal distribution patterns, and overall data structure, improving the matching degree between the target detection component and the operational data of specific industry objects. When a candidate algorithm has a clear advantage, a single detection algorithm can be directly selected; when multiple algorithms have adaptation value, a weighted ensemble can be constructed based on predictive performance, thus balancing detection accuracy and robustness. Compared to methods that rely on a fixed general algorithm or select an algorithm based solely on local verification results, this invention can reduce the cost of repetitive parameter tuning and model evaluation when deploying across industries, improve the ability to adapt to different equipment, components, process units, and production line links, and support the reselection of detection components when operating conditions or production contexts change. This helps to improve the stability of abnormal state identification, reduce the risk of false alarms and false negatives, and provide more reliable basic state detection results for equipment maintenance, process monitoring, and production operation management. Attached Figure Description

[0016] Figure 1 This is a flowchart of a multi-industry-adaptive technical basis testing method in an embodiment of this specification. Figure 2 This is a structural block diagram of a multi-industry-adaptive technical foundation testing system as described in the embodiments of this specification. Detailed Implementation

[0017] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided to make the invention more comprehensive and complete, and to fully convey the concept of the exemplary embodiments to those skilled in the art. The described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. In the following description, numerous specific details are provided to give a full understanding of embodiments of the invention. However, those skilled in the art will recognize that the technical solutions of the invention may be practiced with one or more of these specific details omitted, or other methods, components, systems, steps, etc., may be employed. In other instances, well-known technical solutions are not shown or described in detail to avoid obscuring various aspects of the invention.

[0018] Furthermore, the accompanying drawings are merely illustrative of the invention. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor systems and / or microcontroller systems.

[0019] This invention provides a technical foundation detection method adaptable to multiple industries. (Refer to...) Figure 1 The diagram shown is a flowchart illustrating a multi-industry adaptation technology foundation detection method provided by an embodiment of the present invention. The method is used to perform basic state detection on multivariate operational data generated by industry objects in multiple industry scenarios. The industry objects include at least one of equipment, components, process units, and production line links. The multivariate operational data includes at least one of sensor data, control data, operating condition data, and production context data. Specifically, the method may include the following steps S101-S105: In step S101, during the offline construction phase, a set of historical datasets with state annotations from different industry objects is obtained, a candidate detection algorithm library is established, and the candidate detection algorithms in the candidate detection algorithm library are trained, validated, and evaluated on the set of historical datasets. The data feature descriptions, candidate detection algorithm identifiers, candidate detection algorithm parameter configurations, and candidate detection algorithm detection performance corresponding to each historical dataset are associated to form metadata.

[0020] Specifically, during the offline construction phase, when training, validating, and evaluating the performance of the candidate detection algorithms, a training data portion and a validation data portion are established for each historical dataset. This allows each candidate detection algorithm to complete model construction based on the training data portion under the corresponding candidate detection algorithm parameter configuration, and obtain detection results based on the validation data portion. The detection performance is determined based on the consistency between the detection results and the state labels. The detection performance is a comprehensive evaluation value that simultaneously reflects precision and recall, so that the impact of false positives and false negatives are jointly included in the performance vector. Furthermore, the metadata stores the performance vectors corresponding to different historical datasets in the same candidate detection algorithm arrangement order, enabling the multi-output performance prediction model to simultaneously learn the relative fit relationships of multiple candidate detection algorithms on the same dataset.

[0021] The candidate detection algorithm library includes multiple anomaly detection algorithms of different types capable of handling multivariate operational data. These anomaly detection algorithms cover at least the following: histogram-based outlier scoring algorithms, principal component projection-based detection algorithms, clustering-based local outlier factor detection algorithms, lightweight online anomaly detection algorithms, connection distribution-based detection algorithms, local density neighborhood-based detection algorithms, single-class boundary learning-based detection algorithms, random isolation tree-based detection algorithms, nearest neighbor distance-based detection algorithms, feature bagging-based detection algorithms, and connection structure outlier factor detection algorithms. When the candidate detection algorithm library is established, it retains information on the input data type compatibility, parameter configuration range, support for multivariate sequence processing, and suitability for real-time deployment for each candidate detection algorithm, ensuring that the algorithm selection strategy has a transferable candidate basis across different industry objects.

[0022] As an explanation, the offline construction phase is used to form the foundational data and candidate detection basis for subsequent meta-learning. The historical dataset collection comes from different industry objects, which can correspond to different equipment, components, process units, or production line stages. Each historical dataset contains multivariate operational data that characterizes the object's operational state and has state labels corresponding to the data points. State labels are used to distinguish between normal and abnormal states, where a normal state is represented by 0 and an abnormal state by 1. Therefore, for any industry object m, its state at time t can be represented as: ; in, This represents the state of industry object m at time t, with 0 indicating normal and 1 indicating abnormal. The multivariate runtime data in the historical dataset implicitly reflects this state change, and therefore can serve as the data foundation for training and validating candidate detection algorithms.

[0023] This candidate detection algorithm library provides a variety of anomaly detection algorithms capable of handling multivariate operational data. These algorithms possess different data processing mechanisms and applicable characteristics, thus covering detection needs under different data structures and anomaly patterns. The library may include: HBOS, an outlier scoring algorithm based on histogram distribution, which scores outliers based on the rarity of variable values ​​within the histogram; PCA, a detection algorithm based on principal component projection, which identifies anomalies based on the degree of projection deviation in the principal component space; CBLOF, a detection algorithm based on clustering local outlier factor, which evaluates sample anomaly by combining clustering structure and local outlier degree; LODA, a lightweight online anomaly detection algorithm suitable for online anomaly detection with low computational cost; COPOD, a detection algorithm based on connection distribution relationships, which generates outlier judgments using variable distribution relationships; and LOF, a detection algorithm based on local density neighborhood. The algorithm compares the local density differences between a sample and its neighboring samples; the OCSVM detection algorithm, based on single-class boundary learning, learns the boundaries of normal data and identifies samples outside the boundaries as anomalies; the IFOrest detection algorithm, based on random isolation trees, determines anomalies by the ease with which samples are randomly isolated; the KNN detection algorithm, based on nearest neighbor distance, detects based on the distance relationship between a sample and its nearest neighbors; the FeatureBagging detection algorithm, based on feature bagging, enhances stability by using detection results across multiple feature subspaces; and the COF detection algorithm, based on connection structure outlier factor, evaluates the degree of outlier status using the connection structure between samples. When establishing the candidate detection algorithm library, information such as the input data type compatibility, parameter configuration range, support for multivariate sequence processing, and suitability for real-time deployment is also recorded for each candidate algorithm, providing a unified candidate basis for different industry targets in subsequent algorithm determination.

[0024] For each historical dataset, training data and validation data are established. Candidate detection algorithms, with corresponding parameter configurations, use the training data to build their models and then use the validation data to generate detection results. Detection results can be represented as anomaly scores for data points or as normal or anomalous state labels. Anomaly scores quantify the degree of deviation of data points from normal operating behavior, while state labels provide a more direct anomaly assessment. For candidate detection algorithms that can directly output state labels, the output can be directly compared with the state labels; for candidate detection algorithms that output anomaly scores, the anomaly scores can be converted into state labels based on a threshold or the algorithm's own conversion rules before comparison.

[0025] Detection performance is determined based on the consistency between the detection results and the status labels, using a comprehensive evaluation value that reflects both precision and recall. Precision reflects the proportion of actual anomalous data points among those judged as anomalous, while recall reflects the proportion of correctly identified actual anomalous data points. To ensure that the impact of false alarms and false negatives are both considered in the performance evaluation, detection performance can be represented by the F1 score, i.e.: ; Here, Precision represents the level of accuracy, and Recall represents the level of recall. The F1 score evaluates the detection performance of candidate detection algorithms by harmonic relationship between accuracy and recall. It is suitable for situations where the number of abnormal and normal samples is unbalanced, and also suitable for basic state detection scenarios where both false positive and false negative costs need to be considered.

[0026] In a historical dataset Above, candidate detection algorithm With parameter configuration The verification loss can be expressed as: ; Where A represents the set of candidate detection algorithms, This represents the i-th candidate detection algorithm. Candidate detection algorithm The corresponding parameter configuration space, This indicates the specific parameter configuration. Indicates the number of verification partitions. This represents the training data portion in the j-th partition. Let F1 represent the validation data portion in the j-th partition, and L represent the loss of the candidate detection algorithm on the validation data portion. This expression illustrates that the candidate detection algorithm can achieve corresponding detection performance under different parameter configurations and different data partitions. In this embodiment, the F1 score obtained from the evaluation can be used as a component of the subsequent performance vector.

[0027] After evaluating candidate detection algorithms for each historical dataset, metadata is formed by associating the dataset's data feature description, candidate detection algorithm identifier, candidate detection algorithm parameter configuration, and candidate detection algorithm performance. The data feature description records the basic data attributes and operational data structure of the historical dataset. The candidate detection algorithm identifier distinguishes different anomaly detection algorithms. The candidate detection algorithm parameter configuration records the parameters used during algorithm training and validation. The candidate detection algorithm performance records the evaluation results of the algorithm on that historical dataset. For the same historical dataset, the detection performance of each candidate detection algorithm is arranged into a performance vector according to a preset and consistent candidate detection algorithm arrangement. For different historical datasets, the performance vectors are stored in the same arrangement order, ensuring that the same output dimension always corresponds to the same candidate detection algorithm.

[0028] In this way, the offline construction phase not only obtains the absolute detection performance of each candidate detection algorithm on different historical datasets, but also retains the relative fit relationships of multiple candidate detection algorithms on the same historical dataset. When the subsequent multi-output performance prediction model is trained based on this metadata, it can simultaneously learn the mapping relationship between different data features and the detection performance of multiple candidate detection algorithms. Furthermore, when new industry objects enter the market, it can determine the fit of each candidate detection algorithm based on the features of the data to be detected, thus providing a performance basis for single algorithm selection or multi-algorithm integration.

[0029] In step S102, during the meta-feature extraction stage, meta-features are extracted for each historical dataset. The meta-features include at least features used to characterize the statistical properties of a univariate sequence itself, features used to characterize the correlation between multivariate sequences, features used to characterize abnormal distribution patterns, and features used to characterize the overall structure of the dataset. The meta-features are then filtered, aggregated, and embedded to transform historical datasets from different sources into a unified meta-feature representation.

[0030] The features used to characterize the relationships between multivariate sequences include autoregressive relationship features, cross-correlation relationship features, co-variation relationship features, transitive information relationship features, and dominant frequency relationship features. Specifically, the autoregressive relationship feature is obtained by describing the dependence between the current sequence value and historical lagged sequence values; the cross-correlation relationship feature is obtained by describing the similarity of different variable sequences under lagged shifts; the co-variation relationship feature is obtained by describing the common changing trends of different variable sequences; the transitive information relationship feature is obtained by describing the uncertainty reduction effect of the historical state of one variable sequence on the future state of another variable sequence; and the dominant frequency relationship feature is obtained by identifying the dominant oscillation frequency of the energy distribution of the variable sequence. Furthermore, the relationship features obtained from different variable sequences and combinations of different variable sequences are statistically aggregated for mean, extreme values, dispersion, and distribution patterns to form meta-features that can describe the overall temporal relationship of multivariate operational data.

[0031] The features characterizing the abnormal distribution pattern are generated from a historical dataset with status annotations. These features include total anomaly percentage, point-like anomaly percentage, collective anomaly percentage, anomaly duration, and normal duration. Specifically, the total anomaly percentage characterizes the overall occurrence of anomalous data points in the historical dataset; the point-like anomaly percentage characterizes the degree to which a single data point deviates from its normal state; the collective anomaly percentage characterizes the degree to which consecutively related data points deviate from their normal state; the anomaly duration is obtained by identifying consecutive anomaly-annotated segments and calculating their lengths; and the normal duration is obtained by identifying consecutive normal-annotated segments and calculating their lengths. In the online application phase, the multivariate operational data of the target industry object does not require pre-obtained status annotations as necessary input, and the candidate detection algorithm adaptation judgment is completed through the multi-output performance prediction model.

[0032] Features used to characterize the overall structure of a dataset include clustering quality descriptive features, data complexity descriptive features, information theory descriptive features, benchmark learner descriptive features, model structure descriptive features, and statistical distribution descriptive features. Specifically, the clustering quality descriptive features reflect the separability of samples in the feature space; the data complexity descriptive features reflect class imbalance, boundary samples, distance between neighboring samples, and feature overlap; the information theory descriptive features reflect entropy, mutual information, joint entropy, and attribute noise; the benchmark learner descriptive features are obtained through the performance of a lightweight base model on historical datasets; the model structure descriptive features are obtained through the homogeneity of tree structure nodes and leaf nodes, branch morphology, and feature importance; and the statistical distribution descriptive features are obtained through mean, variance, skewness, kurtosis, correlation, covariance, and sparsity. Thus, these meta-features simultaneously contain information about the internal structure of the data, attribute relationships, and the basic learning difficulty.

[0033] The meta-feature extraction stage further includes a meta-feature preprocessing step, which processes missing meta-features, meta-features with inconsistent dimensions, meta-features with abnormal ranges, and redundant meta-features, and inputs the processed meta-features into the embedding conversion module. The embedding conversion module uses at least one of autoencoder representation learning, principal component projection, and random neighborhood embedding to convert high-dimensional meta-features into low-dimensional representations suitable for learning by the multi-output performance prediction model. In the online application stage, the meta-features to be detected are subjected to the same preprocessing and embedding conversion rules as in the offline construction stage, so that the mapping relationship obtained by offline training can be directly used for new industry objects.

[0034] As an explanation, the meta-feature extraction stage is used to transform historical datasets from different sources, of different sizes, and with different variable structures into meta-feature representations that can be uniformly compared and learned. For each historical dataset, information that describes the data's own properties, relationships between variables, anomaly distribution patterns, and the overall structure of the dataset is extracted. Since the multivariate operational data of various industry objects may have different numbers of variables, sampling lengths, operating condition distributions, and anomaly occurrence patterns, directly using the raw data as the basis for similarity between different historical datasets is not conducive to model learning. Therefore, meta-features are used to abstractly describe the datasets, enabling different historical datasets to be expressed in a unified feature space.

[0035] Features used to characterize the statistical properties of a univariate sequence can be calculated separately for each variable sequence and then aggregated. This allows for the reflection of the numerical distribution, volatility, and morphological characteristics of a single variable. These features may include mean, variance, skewness, kurtosis, sparsity, correlation statistics, and other statistics that describe the distribution of sequence values. For multiple variable sequences in a historical dataset, the statistical features corresponding to each variable can be calculated separately, and then aggregated using metrics such as mean, maximum, minimum, standard deviation, or quantiles. This allows datasets with varying numbers of variables to form descriptive results with fixed dimensions.

[0036] Features used to characterize the relationships between multivariate sequences are used to describe the temporal dependencies, co-variations, and information flow relationships between different variable sequences in the same historical dataset. Multivariate operational data is often used to characterize the state of technical systems. Different variables may have causal correlations, lagged correlations, or co-variations, and these relationships may affect the applicability of candidate detection algorithms, so they need to be expressed in meta-features.

[0037] Autoregressive features are used to quantify the degree of dependence between current sequence values ​​and historical lagged sequence values. For any variable sequence, its autoregressive relationship can be expressed as: ; in, Let c represent the sequence value at time t, and c represent the constant term. Represents the autoregressive coefficient. This represents the historical sequence values ​​under different lag orders. This represents the error term at time t. The autoregressive coefficients reflect the strength of the sequence's dependence on its historical states, thus forming meta-features that describe the temporal stability and inertia of individual variables.

[0038] Cross-correlation characteristics are used to describe the similarity of two variable sequences under lag conditions, thus reflecting whether there is a synchronous change or a lag response relationship between the variables. For two variable sequences x and y, their cross-correlation coefficient under lag tau can be expressed as: ; in, Indicates lag quantity The cross-correlation coefficients under the following conditions and Let represent the values ​​of the two variable sequences at time t, respectively. and These represent the means of the two variable sequences, and N represents the sequence length. The value of this coefficient reflects the positive, negative, or weak correlation between the two variable sequences.

[0039] Covariance characteristics are used to describe the degree to which two variable sequences change together. For variable sequences X and Y, the covariance can be expressed as: ; in, This represents the covariance between variable sequences X and Y. and Indicates the corresponding observation value. and Let represent the means of the two variable sequences, and n represent the number of observations. A positive covariance indicates that the two variables have the same trend, a negative covariance indicates that the two variables have opposite trends, and a covariance close to zero indicates that the linear covariance is relatively weak.

[0040] The transitive entropy characteristic describes the effect of the historical state of one variable sequence on the uncertainty of the future state of another variable sequence. For two discrete-time processes X and Y, the transitive entropy from X to Y at a lag order k can be expressed as: ; in, This represents the transfer entropy from X to Y. Represents conditional entropy. Represents the future value of process Y. and Let represent the historical states of processes Y and X at lag order k, respectively. The first term represents the uncertainty of the future state of Y when both the historical states of Y and X are given simultaneously; the second term represents the uncertainty of the future state of Y when only the historical state of Y itself is given. The difference between the two terms is used to quantify the additional information provided by the historical state of X to the future state of Y.

[0041] Dominant frequency relationship features are used to identify the dominant oscillation frequencies in the energy distribution of a variable sequence, thereby characterizing periodic or oscillatory patterns in the running data. For continuous signals... Its power spectral density is The maximum power frequency can be expressed as: ; in, Indicates the maximum power frequency. This represents the power spectral density of the signal at frequency f. This feature reflects the dominant frequency component in the variable sequence and can serve as a meta-feature for identifying periodic operating patterns and anomalous oscillations.

[0042] Statistical aggregation is performed on autoregressive, cross-correlation, co-variation, information transmission, and dominant frequency characteristics obtained from different variable sequences and combinations thereof. Aggregation methods may include mean, extreme values, standard deviation, quantiles, skewness, and kurtosis, thereby transforming an indefinite number of variable relationships into fixed-dimensional meta-features that can describe the overall temporal relationships of multivariate operating data.

[0043] Features characterizing the anomaly distribution pattern are generated from a historical dataset with state annotations, describing the proportion, form, and duration of anomalies in the historical dataset. Since different anomaly detection algorithms have varying adaptability to point anomalies, continuous anomalies, and collective anomalies, the anomaly distribution pattern serves as an important basis for judging algorithm adaptability. For the state-annotated sequence in the historical dataset, let... This indicates that the i-th data point is in an abnormal state. This indicates that the i-th data point is in a normal state, and N represents the total number of data points. The total percentage of abnormal data points can be expressed as: ; Here, TAR represents the overall prevalence of outlier data points in the historical dataset. The proportion of point-like outliers can be expressed as: ; Here, PAR represents the degree to which a single data point deviates from the normal state. The percentage of collective anomalies can be expressed as: ; Here, CAR represents the degree to which consecutively associated data points deviate from the normal state. The duration of anomalies is obtained by identifying consecutively labeled anomaly segments with a value of 1 and calculating the length of each segment. For example, if an anomaly segment lasts from time i to time j, its duration is j-i+1. The duration of normality is obtained by identifying consecutively labeled normality segments with a value of 0 and calculating the length of each segment. For example, if a normality segment lasts from time i to time j, its duration is j-i+1. These anomaly distribution morphological characteristics are generated only from historical datasets with state annotations during the offline construction phase. During the online application phase, the multivariate runtime data to be detected does not require pre-obtained state annotations; instead, the candidate detection algorithm is adapted and judged using a pre-trained multi-output performance prediction model.

[0044] Features characterizing the overall structure of the dataset are used to describe historical datasets in terms of sample space structure, attribute relationships, and basic learning difficulty. Clustering quality features reflect the separability of samples in the feature space and may include measures related to cluster compactness, inter-cluster separation, and silhouette values. Data complexity features reflect information such as class imbalance, boundary samples, distance to neighboring samples, feature overlap, network density, and sample neighborhood structure. Information theory features reflect information such as entropy, mutual information, joint entropy, attribute noise level, and the number of equivalent attributes. Benchmark learner features are obtained through the performance of a lightweight base model on historical datasets and can be used to characterize the initial learnability of the base model on that dataset. Model structure features are obtained through information such as tree structure nodes, leaf node homogeneity, branch morphology, tree depth, and feature importance. Statistical distribution features are obtained through information such as mean, variance, skewness, kurtosis, correlation, covariance, eigenvalues, interquartile range, and sparsity. Through these structural features, meta-features not only describe the statistical distribution of the original variables but also the internal structure of the data, attribute relationships, and the difficulty of the detection task.

[0045] After extracting meta-features, they undergo preprocessing. For missing meta-features, pre-defined imputation rules are applied to ensure that the same feature dimension has valid values ​​across different historical datasets. For meta-features with inconsistent scales, standardization or normalization is performed to prevent model bias due to differences in numerical scales. For meta-features with outlier ranges, truncation, replacement, or smoothing is performed to reduce the impact of extreme values ​​on subsequent learning. For redundant meta-features, correlation analysis, importance filtering, or dimension selection can be used to reduce repetitive representations, making the input meta-features more stable.

[0046] The preprocessed meta-features are input into the embedding transformation module. The embedding transformation module can employ at least one of autoencoder representation learning, principal component projection, and random neighborhood embedding to convert high-dimensional meta-features into low-dimensional representations. Autoencoder representation learning learns a compact representation of the meta-features through an encoder and constrains this representation to retain key information through the decoding process; principal component projection extracts directions that explain the main variance through linear projection; random neighborhood embedding forms a low-dimensional representation by maintaining the similarity of sample neighborhoods. After embedding transformation, historical datasets from different sources form a unified meta-feature representation, which serves as input to the subsequent multi-output performance prediction model. In the online application phase, the meta-features to be detected are subjected to the same preprocessing and embedding transformation rules as in the offline construction phase, enabling the mapping relationship between the meta-features obtained from offline training and the detection performance to be directly applied to new industry objects.

[0047] In step S103, during the meta-model training phase, a unified meta-feature representation is used as input, and a performance vector composed of the detection performance of each candidate detection algorithm on the same historical dataset is used as the output target to train a multi-output performance prediction model. This enables the multi-output performance prediction model to learn the mapping relationship between meta-features and the detection performance of multiple candidate detection algorithms, and adjusts the model parameters based on the deviation between the predicted performance vector and the true performance vector during the training process.

[0048] The multi-output performance prediction model employs a multi-output regression structure, which uses the detection performance of each candidate detection algorithm as an output dimension and establishes a nonlinear mapping between meta-feature representations and performance vectors through an ensemble regression learner. The ensemble regression learner consists of multiple base learners, each trained on different sample subsets and different meta-feature subsets. During prediction, the outputs of each base learner are aggregated to obtain the predicted detection performance of each candidate detection algorithm. The training objective of the multi-output performance prediction model is to reduce the overall deviation between the predicted performance vector and the true performance vector, so that the multi-output performance prediction model not only predicts the absolute performance of a single candidate detection algorithm, but also predicts the relative superiority or inferiority relationship between different candidate detection algorithms.

[0049] The meta-model training phase is used to establish performance prediction relationships based on a unified meta-feature representation and corresponding performance vectors. Each historical dataset, after processing in step S102, yields a unified meta-feature representation, which serves as the input to the training samples. The detection performance of each candidate detection algorithm on the same historical dataset is arranged in a fixed order to form a performance vector, which serves as the output target for the training samples. Since each dimension of the performance vector corresponds to a candidate detection algorithm in the candidate detection algorithm library, the multi-output performance prediction model can simultaneously learn the performance differences of multiple candidate detection algorithms on the same historical dataset during training, rather than independently judging a single candidate detection algorithm.

[0050] The multi-output performance prediction model employs a multi-output regression structure. Assume there are N historical datasets in the historical dataset set, each represented by M meta-features. Then the input can be represented as matrix X, where... Let Y represent the j-th meta-feature of the i-th historical dataset. The output target can be represented as matrix Y, where This represents the true detection performance of the j-th candidate detection algorithm on the i-th historical dataset. The multi-output performance prediction model learns a mapping function f, enabling it to predict the detection performance of each candidate detection algorithm based on the meta-feature representation, thus obtaining a prediction performance vector. .

[0051] During training, the overall deviation between the predicted performance vector and the true performance vector is used as the basis for optimization. This overall deviation can be expressed as the mean squared error: ; in, This represents the overall deviation of the predicted performance vector from the true performance vector. Indicates actual detection performance. The loss function represents the predicted detection performance, where D represents the number of output dimensions (i.e., the number of candidate detection algorithms), and N represents the number of historical datasets used in training. This loss function constrains the model parameter tuning by the prediction errors of multiple candidate detection algorithms simultaneously, enabling the model to learn the overall performance vector rather than just the local performance of a single candidate detection algorithm.

[0052] The training objective of a multi-output performance prediction model can be expressed as: ; in, Let f(X) represent the optimal performance prediction model obtained through training, and let f(X) represent the prediction performance matrix output by the model based on the meta-feature matrix X. By reducing... The model gradually establishes a nonlinear mapping relationship between meta-feature representation and the detection performance of candidate detection algorithms.

[0053] In one implementation, the multi-output performance prediction model can be achieved through an ensemble regression learner. The ensemble regression learner consists of multiple base learners, each trained on different subsets of historical datasets and different subsets of meta-features. This allows each base learner to learn performance prediction patterns from different data fragments and feature combinations. During prediction, each base learner outputs its performance estimate for a candidate detection algorithm. The outputs of each base learner are then aggregated to obtain the predicted detection performance for each candidate algorithm. This structure reduces prediction fluctuations caused by the influence of local data distribution on a single regression model, thus improving the stability of performance prediction.

[0054] Since each output dimension corresponds to a fixed candidate detection algorithm, the prediction performance vector output by the multi-output performance prediction model not only represents the absolute detection performance of each candidate detection algorithm under the features of the data to be processed, but also includes the relative advantages and disadvantages among different candidate detection algorithms. In other words, when the prediction detection performance of a certain candidate detection algorithm is higher than that of other candidate detection algorithms, the model can reflect its relative suitability advantage; when the prediction detection performance of multiple candidate detection algorithms is similar, the model can also provide a basis for determining weights for subsequent integration and construction. Therefore, the multi-output performance prediction model generated in step S103 can serve as the core model for judging the suitability of candidate detection algorithms in the online application stage.

[0055] In step S104, during the online application phase, the multivariate running data of the industry object to be detected is obtained, and the meta-features to be detected are extracted and transformed in the same manner as in the offline construction phase. The meta-features to be detected are then input into the multi-output performance prediction model to obtain the predicted detection performance of each candidate detection algorithm in the candidate detection algorithm library for the industry object to be detected.

[0056] The online application phase involves applying the trained multi-output performance prediction model to new industry objects. The industry object to be tested can be equipment, components, process units, or production line segments, and its multivariate operational data can include sensor data, control data, operating condition data, and production context data. This multivariate operational data characterizes the state characteristics of the industry object under current operating conditions. The data can contain continuous time series data or contextual attributes related to production procedures, workload, material type, product model, or environmental conditions.

[0057] After acquiring the multivariate operational data of the industry object to be detected, the meta-features to be detected are extracted in a manner consistent with the offline construction phase. This consistency includes the same rules for calculating variable statistical features, the same rules for calculating multivariate sequence association relationships, the same rules for describing the overall structure of the dataset, the same rules for preprocessing meta-features, and the same rules for embedding transformation. Through this processing method, even though the industry object to be detected originates from a new industry scenario or a new operating condition, its operational data can still be transformed into the same meta-feature representation space as the historical dataset, thereby ensuring that the multi-output performance prediction model can directly process the meta-features to be detected.

[0058] In the online application phase, the meta-features to be detected are mainly used to describe the structure and relational characteristics of the current multivariate operational data of the industry object to be detected, without requiring prior state labeling of the data to be detected. For scenarios where it is impossible to confirm anomaly labels in advance during online deployment, univariate statistical properties, multivariate temporal relationships, and overall data structure information can be extracted from the sensor data and control data themselves, and converted into low-dimensional meta-feature representations. For features related to anomaly distribution patterns, when there are no pre-defined state labels for the data to be detected, they are not required inputs. They can be processed by default, by substitution, or by using only meta-feature dimensions that do not depend on labels, according to the input processing rules determined during offline training, so that online prediction can still maintain the same feature input format as the offline training phase.

[0059] After the meta-features to be detected are input into the multi-output performance prediction model, the model outputs the predicted detection performance of each candidate detection algorithm in the candidate detection algorithm library for the target industry object. Since the output dimensions of the multi-output performance prediction model correspond to the candidate detection algorithms in the library in a fixed order, the output results form a prediction performance vector for the target industry object. Each element in this prediction performance vector represents the expected detection effect of the corresponding candidate detection algorithm on the current multivariate running data, reflecting the degree to which each candidate detection algorithm is adapted to the current data structure, variable relationships, and operating conditions of the target industry object.

[0060] This online application process eliminates the need to run candidate detection algorithms one by one and perform manual verification for each industry object to obtain the predicted detection performance of the candidate algorithms. The process utilizes mapping relationships learned from different historical datasets and candidate algorithm evaluation results in the offline phase to map the meta-feature representations of new data to the performance estimates of each candidate detection algorithm. Therefore, the subsequent algorithm determination phase can perform sorting, filtering, advantage judgment, or integrated construction based on the predicted performance vectors. This avoids relying entirely on manual experience to select anomaly detection algorithms for new industry objects and enables the detection components to adapt to changes in industry object type, operating load, process procedures, and environmental conditions.

[0061] In step S105, during the algorithm determination stage, an algorithm selection strategy is executed based on the predictive detection performance of each candidate detection algorithm to determine the target detection component from the candidate detection algorithm library; when there is a candidate detection algorithm whose predictive detection performance meets the advantage condition, the candidate detection algorithm is determined as the target detection component; when multiple candidate detection algorithms meet the integration condition, integration weights are determined for the corresponding multiple candidate detection algorithms based on the predictive detection performance, and the multiple candidate detection algorithms with integration weights are determined as the target detection component.

[0062] The algorithm selection strategy includes a sorting and filtering step, a threshold exclusion step, an advantage judgment step, and an integration and construction step. In the sorting and filtering step, candidate detection algorithms are sorted from high to low according to their predictive detection performance. In the threshold exclusion step, candidate detection algorithms with predictive detection performance below a preset lower limit are excluded. In the advantage judgment step, when the top-ranked candidate detection algorithm meets a preset advantage condition relative to other candidate detection algorithms, a single candidate detection algorithm is determined as the target detection component. In the integration and construction step, when multiple candidate detection algorithms meet a preset integration condition, they are sequentially included until the cumulative performance meets a preset upper limit condition, and the integration weight is obtained by normalizing the predictive detection performance of each candidate detection algorithm. When the target detection component consists of multiple candidate detection algorithms, the member anomaly scores output by each candidate detection algorithm are weighted and fused, and the status label is generated based on the fused anomaly scores. When the workload, process, component status, and environmental conditions of the industry object change and cause the meta-features to be detected to meet the reselection trigger condition, the online application stage and the algorithm determination stage are re-executed to update the target detection component.

[0063] As an explanation, the algorithm determination stage is used to determine a suitable target detection component from the candidate detection algorithm library based on the prediction performance vector output in step S104. Each element in the prediction performance vector corresponds to a candidate detection algorithm in a fixed order, so the algorithm can be directly sorted and selected based on its prediction performance. The purpose of this stage is to use a single candidate detection algorithm as the target detection component when it has a significant advantage in adapting to the current target industry; and to construct a target detection component composed of multiple candidate detection algorithms when multiple candidate algorithms have some adaptability and no algorithm has a clear advantage, thereby reducing the impact of the uncertainty of a single algorithm's prediction on the detection results.

[0064] Algorithm selection strategies can be performed by ranking and filtering based on predictive detection performance. Candidate detection algorithms are ranked from highest to lowest according to their predictive detection performance, and the ranking results are used to determine the relative merits of the candidate algorithms. Since the multi-output performance prediction model simultaneously outputs performance estimates for multiple candidate detection algorithms, the ranking results not only reflect the expected detection performance of each candidate algorithm but also the differences in suitability between different candidate detection algorithms for the same target industry object.

[0065] In the threshold exclusion process, candidate detection algorithms whose predictive detection performance is below a preset lower limit are excluded. The preset lower limit can be denoted as: This process is used to eliminate candidate detection algorithms with low predictive detection performance and insufficient adaptability to the current industry objects being detected. By doing so, poorly performing candidate detection algorithms can be prevented from entering the target detection component, especially when building the integrated component later, thus reducing the adverse impact of poorly adapted algorithms on the fusion anomaly score.

[0066] In the advantage determination process, if the top-ranked candidate detection algorithm meets a preset advantage condition compared to other candidate detection algorithms, then that top-ranked candidate detection algorithm is determined as the target detection component. The advantage condition can be determined based on the predictive detection performance of the top-ranked candidate detection algorithm, its performance gap with subsequent candidate detection algorithms, or its leading position relative to the overall predictive performance of the candidate set. When a clearly superior algorithm exists, forming a single target detection component is more suitable, because combining a significantly superior algorithm with other algorithms with lower predictive performance may reduce the overall detection effect and increase the computational and deployment burden.

[0067] In the ensemble building process, when multiple candidate detection algorithms jointly meet the preset ensemble conditions, and no single candidate detection algorithm meets the dominance condition, they are included in the candidate detection algorithms in descending order of predicted detection performance, until the cumulative performance meets the preset upper limit condition. The preset upper limit condition can be denoted as... This is used to control the number of candidate detection algorithms included in the target detection component and the overall performance coverage. The inclusion of candidate detection algorithms is still subject to a preset lower limit condition. Constraints, predictive detection performance is lower than Candidate detection algorithms are not included in the integration process. This approach prioritizes candidate detection algorithms with high predictive performance in the object detection component and avoids including candidate detection algorithms with insufficient adaptability in the fusion process.

[0068] For a target detection component consisting of multiple candidate detection algorithms, the ensemble weights are determined based on the predictive detection performance of each algorithm. Candidate detection algorithms with higher predictive detection performance receive larger ensemble weights in the target detection component; conversely, candidate detection algorithms with relatively lower predictive detection performance but still meeting the ensemble criteria receive smaller ensemble weights. The ensemble weights can be obtained by normalizing the predictive detection performance of the included candidate algorithms, ensuring that the contribution of each member algorithm aligns with its predictive fit. Through this dynamic weight allocation method, the target detection component can leverage the complementary detection capabilities of multiple candidate detection algorithms for complex data structures while also highlighting candidate detection algorithms with higher predictive fit.

[0069] When the target detection component consists of multiple candidate detection algorithms, each candidate algorithm outputs a member anomaly score based on the multivariate runtime data of the target industry object. The member anomaly score represents the corresponding candidate detection algorithm's judgment of the degree of anomaly of the data point. The target detection component performs weighted fusion of the multiple member anomaly scores according to the fusion weights to obtain a fused anomaly score, and generates a status label based on the fused anomaly score. The status label indicates whether the corresponding data point or time segment is in a normal or abnormal state. This process can form a comprehensive detection result through weighted fusion when there are differences in the judgments of multiple candidate detection algorithms, improving the adaptability of the detection component to different anomaly forms and different data structures.

[0070] When the workload, process procedures, component status, or environmental conditions of an industrial object change, the statistical distribution, variable correlations, and overall data structure of multivariate operational data may change accordingly, potentially reducing the adaptability of the original target detection component. In such cases, the reselection trigger condition can be determined based on the changes in the detected meta-features. The reselection trigger condition can be determined by the degree of difference between the detected meta-feature and the previous meta-feature representation, the magnitude of change in key meta-features, or the change in the operational context. When the reselection trigger condition is met, the online application phase and algorithm determination phase are re-executed. The new detected meta-features are input into the multi-output performance prediction model, and the predictive detection performance of each candidate detection algorithm is re-obtained, updating the target detection component accordingly. Through this mechanism, the target detection component can adjust to changes in the industrial object's operational status and production conditions, preventing the continued use of fixed detection algorithms after a change in operating conditions, which could lead to an increase in false alarms or false negatives.

[0071] In step S106, during the detection output stage, the target detection component is applied to the multivariate operating data of the industry object to be detected, and at least one of anomaly score and status label is generated for the data points in the multivariate operating data. The abnormal status detection result of the industry object to be detected is output according to at least one of the anomaly score and status label.

[0072] The detection output stage applies the target detection component determined in step S105 to the multivariate operational data of the industry object to be detected, and generates anomaly detection results. The target detection component can be a single candidate detection algorithm or multiple candidate detection algorithms with integrated weights. After the target detection component is determined, the sensor data, control data, operating condition data, and production context data of the industry object to be detected are input into the target detection component, which then performs anomaly judgment on the data points or continuous data segments.

[0073] When the target detection component is a single candidate detection algorithm, it directly processes the multivariate runtime data and generates at least one of anomaly scores and state labels for the data points. Anomaly scores quantify the degree of deviation of a data point from its expected behavior; a higher score indicates that the data point is more likely to be in an abnormal state. State labels provide a more direct detection conclusion, classifying the data point into a normal or abnormal state. For candidate detection algorithms that can directly output state labels, their output can be used as the anomaly detection result; for candidate detection algorithms that output anomaly scores, the anomaly scores can be converted into state labels according to a preset threshold, quantile rules, or the algorithm's own conversion rules.

[0074] When the target detection component consists of multiple candidate detection algorithms, each candidate algorithm detects the same multivariate data separately, obtaining corresponding member anomaly scores. These member anomaly scores are then fused according to the ensemble weights determined in step S105 to obtain a fused anomaly score. The fused anomaly score comprehensively reflects the judgments of multiple candidate detection algorithms regarding the degree of anomaly at the same data point. Based on the fused anomaly score, state labels can be further generated, enabling the detection results of multiple candidate detection algorithms to form a unified output. This method can, when different candidate detection algorithms have different judgments regarding the same data point, assign greater influence to candidate detection algorithms with higher predictive detection performance, thereby improving the stability of anomaly state detection results.

[0075] Anomaly detection results can include anomaly scores at the data point level, status labels at the data point level, anomaly intervals within consecutive time segments, anomaly occurrence time, anomaly duration, and anomaly alerts corresponding to the industry object being monitored. For production line links or process units, status labels from multiple consecutive data points can be aggregated to obtain an operational status judgment within the corresponding time window. For equipment or components, anomaly detection results for maintenance decisions can be output based on a continuous increase in the anomaly score, the continuous appearance of anomaly labels, or the length of anomaly segments.

[0076] Anomaly detection results are used to indicate whether the monitored industrial object deviates from normal operating behavior. The results can be in the form of anomaly scores or binary status labels. Anomaly scores retain information about the degree of deviation of data points, facilitating the sorting of data points according to the probability of anomaly; status labels provide concise judgment results, easily used directly by production monitoring systems, alarm systems, or maintenance systems. Therefore, the detection output stage can provide continuous anomaly levels for subsequent status assessments and also provide clear normal or abnormal judgments for on-site handling.

[0077] Based on the same line of thought, such as Figure 2As shown, a multi-industry-adaptive technology foundation detection system is provided. This system is used to perform basic state detection on multivariate operational data generated by industrial objects in multiple industry scenarios. The industrial objects include at least one of equipment, components, process units, and production line links. The multivariate operational data includes at least one of sensor data, control data, operating condition data, and production context data. The system includes: The offline construction module 201 is used to acquire a set of historical datasets with state annotations from different industry objects, establish a candidate detection algorithm library, train, verify and evaluate the performance of the candidate detection algorithms in the candidate detection algorithm library on the set of historical datasets, and associate the data feature descriptions, candidate detection algorithm identifiers, candidate detection algorithm parameter configurations and candidate detection algorithm detection performance corresponding to each historical dataset to form metadata; The meta-feature extraction module 202 is used to extract meta-features for each historical dataset. The meta-features include at least features that characterize the statistical properties of the univariate sequence itself, features that characterize the correlation between multivariate sequences, features that characterize the abnormal distribution pattern, and features that characterize the overall structure of the dataset. The meta-features are then filtered, aggregated, and embedded to transform the historical datasets from different sources into a unified meta-feature representation. The meta-model training module 203 is used to train a multi-output performance prediction model with a unified meta-feature representation as input and a performance vector composed of the detection performance of each candidate detection algorithm on the same historical dataset as the output target. The multi-output performance prediction model learns the mapping relationship between meta-features and the detection performance of multiple candidate detection algorithms, and adjusts the model parameters according to the deviation between the predicted performance vector and the true performance vector during the training process. The online application module 204 is used to acquire the multivariate running data of the industry object to be detected, extract and transform the meta-features to be detected in the same way as the offline construction module, input the meta-features to be detected into the multi-output performance prediction model, and obtain the prediction detection performance of each candidate detection algorithm in the candidate detection algorithm library for the industry object to be detected. The algorithm determination module 205 is used to execute an algorithm selection strategy based on the predictive detection performance of each candidate detection algorithm to determine a target detection component from the candidate detection algorithm library; when there is a candidate detection algorithm whose predictive detection performance meets the advantage condition, the candidate detection algorithm is determined as the target detection component; when multiple candidate detection algorithms meet the integration condition, the integration weight is determined for the corresponding multiple candidate detection algorithms based on the predictive detection performance, and the multiple candidate detection algorithms with integration weights are determined as the target detection component. The detection output module 206 is used to apply the target detection component to the multivariate operating data of the industry object to be detected, generate at least one of anomaly score and status label for the data points in the multivariate operating data, and output the abnormal status detection result of the industry object to be detected based on at least one of the anomaly score and status label.

[0078] The technical solution of the present invention has the following beneficial effects: This invention transforms the algorithm selection process from manual experience-based screening to automatic adaptation judgment based on historical task mapping relationships by offline construction of a candidate detection algorithm library, extraction and unified representation of historical dataset metadata features, training a multi-output performance prediction model, and online prediction of the performance of each candidate detection algorithm based on the metadata features of the data to be detected. This approach simultaneously considers factors affecting detection performance, such as univariate statistical properties, multivariate correlations, abnormal distribution patterns, and overall data structure, improving the matching degree between the target detection component and the operational data of specific industry objects. When a candidate algorithm has a clear advantage, a single detection algorithm can be directly selected; when multiple algorithms have adaptation value, a weighted ensemble can be constructed based on predictive performance, thus balancing detection accuracy and robustness. Compared to methods that rely on a fixed general algorithm or select an algorithm based solely on local verification results, this invention can reduce the cost of repetitive parameter tuning and model evaluation when deploying across industries, improve the ability to adapt to different equipment, components, process units, and production line links, and support the reselection of detection components when operating conditions or production contexts change. This helps to improve the stability of abnormal state identification, reduce the risk of false alarms and false negatives, and provide more reliable basic state detection results for equipment maintenance, process monitoring, and production operation management.

[0079] This system performs time-frequency analysis on mixed microphone signals, using the amplitude spectrum as input to a neural network. It utilizes a bidirectional long short-term memory network to model the correlation between consecutive time frames, improving the model's ability to represent the temporal differences between human voices and musical sounds. By embedding representations to cluster and form binary masks corresponding to different sound source categories, the mixed amplitude spectrum can be divided into human voice and musical sound amplitude spectra. Combined with the preserved mixed phase spectrum, the corresponding time-domain signal is reconstructed, thus achieving separation of human voices and musical sounds under overlapping sound source conditions. This invention does not rely on directly enumerating all possible multi-source combinations for classification, reducing the computational burden caused by the expansion of training sample size in multi-source overlapping scenarios, and is applicable to enclosed acoustic environments where room impulse response is a factor.

[0080] It should be noted that although several modules or units of the device for performing actions have been mentioned in the detailed description above, this division is not mandatory. In fact, according to exemplary embodiments of the present invention, the features and functions of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided and embodied by multiple modules or units.

[0081] Other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only, and the true scope and spirit of the invention are indicated by the claims.

[0082] It should be understood that the present invention is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A technology-based testing method adaptable to multiple industries, characterized in that, The method is used to perform basic state detection on multivariate operational data generated by industrial objects in multiple industrial scenarios. The industrial objects include at least one of equipment, components, process units, and production line links. The multivariate operational data includes at least one of sensor data, control data, operating condition data, and production context data. The method includes: During the offline construction phase, a set of historical datasets with state annotations from different industry objects is obtained, a candidate detection algorithm library is established, and the candidate detection algorithms in the candidate detection algorithm library are trained, validated, and evaluated on the set of historical datasets. The data feature descriptions, candidate detection algorithm identifiers, candidate detection algorithm parameter configurations, and candidate detection algorithm detection performance corresponding to each historical dataset are associated to form metadata. In the meta-feature extraction stage, meta-features are extracted for each historical dataset. The meta-features include at least features used to characterize the statistical properties of the univariate sequence itself, features used to characterize the correlation between multivariate sequences, features used to characterize the abnormal distribution pattern, and features used to characterize the overall structure of the dataset. The meta-features are then filtered, aggregated, and embedded to transform historical datasets from different sources into a unified meta-feature representation. During the meta-model training phase, a unified meta-feature representation is used as input, and a performance vector composed of the detection performance of each candidate detection algorithm on the same historical dataset is used as the output target. The multi-output performance prediction model is trained so that it learns the mapping relationship between meta-features and the detection performance of multiple candidate detection algorithms. During the training process, the model parameters are adjusted according to the deviation between the predicted performance vector and the true performance vector. In the online application phase, the multivariate running data of the industry object to be detected is obtained, and the meta-features to be detected are extracted and transformed in the same way as in the offline construction phase. The meta-features to be detected are input into the multi-output performance prediction model to obtain the predicted detection performance of each candidate detection algorithm in the candidate detection algorithm library for the industry object to be detected. In the algorithm determination stage, an algorithm selection strategy is executed based on the predictive detection performance of each candidate detection algorithm to determine the target detection component from the candidate detection algorithm library; when there is a candidate detection algorithm whose predictive detection performance meets the advantage condition, the candidate detection algorithm is determined as the target detection component; when multiple candidate detection algorithms meet the integration condition, the integration weight is determined for the corresponding multiple candidate detection algorithms based on the predictive detection performance, and the multiple candidate detection algorithms with integration weights are determined as the target detection component. In the detection output stage, the target detection component is applied to the multivariate operating data of the industry object to be detected, and at least one of anomaly score and status label is generated for the data points in the multivariate operating data. The abnormal status detection result of the industry object to be detected is output according to at least one of the anomaly score and status label.

2. The multi-industry adaptability technical basis detection method according to claim 1, characterized in that, During the offline construction phase, when training, validating, and evaluating the performance of the candidate detection algorithms, a training data part and a validation data part are established for each historical dataset. This allows each candidate detection algorithm to complete model construction based on the training data part under the corresponding candidate detection algorithm parameter configuration, and obtain detection results based on the validation data part. The detection performance is determined based on the consistency between the detection results and the state labels. The detection performance is a comprehensive evaluation value that reflects both precision and recall, so that the impact of false positives and false negatives are included in the performance vector. Furthermore, the metadata stores the performance vectors corresponding to different historical datasets in the same candidate detection algorithm order, so that the multi-output performance prediction model can learn the relative fit relationship of multiple candidate detection algorithms on the same dataset at the same time.

3. The multi-industry adaptability technical basis detection method according to claim 1, characterized in that, The candidate detection algorithm library includes multiple anomaly detection algorithms of different types capable of handling multivariate operational data. These anomaly detection algorithms cover at least the following: histogram-based outlier scoring algorithms, principal component projection-based detection algorithms, clustering-based local outlier factor detection algorithms, lightweight online anomaly detection algorithms, connection distribution-based detection algorithms, local density neighborhood-based detection algorithms, single-class boundary learning-based detection algorithms, random isolation tree-based detection algorithms, nearest neighbor distance-based detection algorithms, feature bagging-based detection algorithms, and connection structure outlier factor detection algorithms. When the candidate detection algorithm library is established, it retains information on the input data type compatibility, parameter configuration range, support for multivariate sequence processing, and suitability for real-time deployment for each candidate detection algorithm, ensuring that the algorithm selection strategy has a transferable candidate basis across different industry objects.

4. The multi-industry adaptability technical basis detection method according to claim 1, characterized in that, Features used to characterize the relationships between multivariate sequences include autoregressive features, cross-correlation features, co-variation features, transitive features, and dominant frequency features. Specifically, the autoregressive feature is obtained by describing the dependence between the current sequence value and historical lagged sequence values; the cross-correlation feature is obtained by describing the similarity of different variable sequences under lagged shifts; the co-variation feature is obtained by describing the common changing trends of different variable sequences; the transitive feature is obtained by describing the uncertainty reduction effect of the historical state of one variable sequence on the future state of another variable sequence; and the dominant frequency feature is obtained by identifying the dominant oscillation frequency of the variable sequence's energy distribution. Furthermore, the features representing the relationships between different variable sequences and combinations of different variable sequences are statistically aggregated for mean, extreme values, dispersion, and distribution patterns to form meta-features capable of describing the overall temporal relationships of multivariate operational data.

5. The multi-industry adaptability technical basis detection method according to claim 1, characterized in that, The features characterizing the abnormal distribution pattern are generated from a historical dataset with status annotations. These features include total anomaly percentage, point-like anomaly percentage, collective anomaly percentage, anomaly duration, and normal duration. Specifically, the total anomaly percentage characterizes the overall occurrence of anomalous data points in the historical dataset; the point-like anomaly percentage characterizes the degree to which a single data point deviates from its normal state; the collective anomaly percentage characterizes the degree to which consecutively related data points deviate from their normal state; the anomaly duration is obtained by identifying consecutive anomaly-annotated segments and calculating their lengths; and the normal duration is obtained by identifying consecutive normal-annotated segments and calculating their lengths. In the online application phase, the multivariate operational data of the target industry object does not require pre-obtained status annotations as necessary input, and the candidate detection algorithm adaptation judgment is completed through the multi-output performance prediction model.

6. The multi-industry adaptability technical basis detection method according to claim 1, characterized in that, Features used to characterize the overall structure of a dataset include clustering quality descriptive features, data complexity descriptive features, information theory descriptive features, benchmark learner descriptive features, model structure descriptive features, and statistical distribution descriptive features. Specifically, the clustering quality descriptive features reflect the separability of samples in the feature space; the data complexity descriptive features reflect class imbalance, boundary samples, distance between neighboring samples, and feature overlap; the information theory descriptive features reflect entropy, mutual information, joint entropy, and attribute noise; the benchmark learner descriptive features are obtained through the performance of a lightweight base model on historical datasets; the model structure descriptive features are obtained through the homogeneity of tree structure nodes and leaf nodes, branch morphology, and feature importance; and the statistical distribution descriptive features are obtained through mean, variance, skewness, kurtosis, correlation, covariance, and sparsity. Thus, these meta-features simultaneously contain information about the internal structure of the data, attribute relationships, and the basic learning difficulty.

7. The multi-industry adaptability technical basis detection method according to claim 1, characterized in that, The meta-feature extraction stage further includes a meta-feature preprocessing step, which processes missing meta-features, meta-features with inconsistent dimensions, meta-features with abnormal ranges, and redundant meta-features, and inputs the processed meta-features into the embedding conversion module. The embedding conversion module uses at least one of autoencoder representation learning, principal component projection, and random neighborhood embedding to convert high-dimensional meta-features into low-dimensional representations suitable for learning by the multi-output performance prediction model. In the online application stage, the meta-features to be detected are subjected to the same preprocessing and embedding conversion rules as in the offline construction stage, so that the mapping relationship obtained by offline training can be directly used for new industry objects.

8. The multi-industry adaptability technical basis detection method according to claim 1, characterized in that, The multi-output performance prediction model employs a multi-output regression structure, which uses the detection performance of each candidate detection algorithm as an output dimension and establishes a nonlinear mapping between meta-feature representations and performance vectors through an ensemble regression learner. The ensemble regression learner consists of multiple base learners, each trained on different sample subsets and different meta-feature subsets. During prediction, the outputs of each base learner are aggregated to obtain the predicted detection performance of each candidate detection algorithm. The training objective of the multi-output performance prediction model is to reduce the overall deviation between the predicted performance vector and the true performance vector, so that the multi-output performance prediction model not only predicts the absolute performance of a single candidate detection algorithm, but also predicts the relative superiority or inferiority relationship between different candidate detection algorithms.

9. The multi-industry adaptability technical basis detection method according to claim 1, characterized in that, The algorithm selection strategy includes a sorting and filtering step, a threshold exclusion step, an advantage judgment step, and an integration and construction step. In the sorting and filtering step, candidate detection algorithms are sorted from high to low according to their predictive detection performance. In the threshold exclusion step, candidate detection algorithms with predictive detection performance below a preset lower limit are excluded. In the advantage judgment step, when the top-ranked candidate detection algorithm meets a preset advantage condition relative to other candidate detection algorithms, a single candidate detection algorithm is determined as the target detection component. In the integration and construction step, when multiple candidate detection algorithms jointly meet a preset integration condition, they are sequentially included in the candidate detection algorithms until the cumulative performance meets a preset upper limit condition, and the integration weight is obtained by normalizing the predictive detection performance of each candidate detection algorithm. When the target detection component consists of multiple candidate detection algorithms, the member anomaly scores output by each candidate detection algorithm are weighted and fused, and the status label is generated based on the fused anomaly scores. When the workload, process, component status and environmental conditions of the industrial object change and cause the meta-feature to be detected to meet the reselection trigger condition, the online application stage and the algorithm determination stage are re-executed to update the target detection component.

10. A technology foundation testing system adaptable to multiple industries, characterized in that, The system is used to perform basic state detection on multivariate operational data generated by industrial objects in multiple industrial scenarios. The industrial objects include at least one of equipment, components, process units, and production line links. The multivariate operational data includes at least one of sensor data, control data, operating condition data, and production context data. The system includes: The offline construction module is used to acquire a set of historical datasets with state annotations from different industry objects, establish a candidate detection algorithm library, train, verify and evaluate the performance of the candidate detection algorithms in the candidate detection algorithm library on the set of historical datasets, and associate the data feature descriptions, candidate detection algorithm identifiers, candidate detection algorithm parameter configurations and candidate detection algorithm detection performance corresponding to each historical dataset to form metadata; The meta-feature extraction module is used to extract meta-features for each historical dataset. The meta-features include at least features that characterize the statistical properties of the univariate sequence itself, features that characterize the correlation between multivariate sequences, features that characterize the abnormal distribution pattern, and features that characterize the overall structure of the dataset. The module also filters, aggregates, and embeds the meta-features to convert historical datasets from different sources into a unified meta-feature representation. The meta-model training module is used to train a multi-output performance prediction model by taking a unified meta-feature representation as input and a performance vector composed of the detection performance of each candidate detection algorithm on the same historical dataset as the output target. The multi-output performance prediction model learns the mapping relationship between meta-features and the detection performance of multiple candidate detection algorithms, and adjusts the model parameters according to the deviation between the predicted performance vector and the true performance vector during the training process. The online application module is used to acquire the multivariate running data of the industry object to be detected, extract and transform the meta-features to be detected in the same way as the offline construction module, input the meta-features to be detected into the multi-output performance prediction model, and obtain the prediction detection performance of each candidate detection algorithm in the candidate detection algorithm library for the industry object to be detected. The algorithm determination module is used to execute an algorithm selection strategy based on the predictive detection performance of each candidate detection algorithm, and determine the target detection component from the candidate detection algorithm library; when there is a candidate detection algorithm whose predictive detection performance meets the advantage condition, the candidate detection algorithm is determined as the target detection component; when multiple candidate detection algorithms meet the integration condition, the integration weight is determined for the corresponding multiple candidate detection algorithms based on the predictive detection performance, and the multiple candidate detection algorithms with integration weights are determined as the target detection component. The detection output module is used to apply the target detection component to the multivariate operating data of the industry object to be detected, generate at least one of anomaly score and status label for the data points in the multivariate operating data, and output the abnormal status detection result of the industry object to be detected based on at least one of the anomaly score and status label.

Images