An asset process intelligent monitoring method and system based on artificial intelligence technology

Abstract

The present application belongs to the technical field of asset process monitoring, and particularly relates to an asset process intelligent monitoring method and system based on artificial intelligence technology. The method extracts standard asset feature vectors in a normal state by analyzing historical asset data using a pre-trained support vector machine model, extracts real-time asset feature vectors of real-time asset data based on a pre-set gradient boosting tree model, judges whether there is an asset process abnormal node in the current real-time asset data by high-dimensional mapping and abnormal analysis of the two kinds of feature vectors, marks and corrects abnormal data when there is an abnormal node, and adopts a blockchain data processing step and chains to a pre-constructed traceability chain when there is no abnormal node. The present application can not only discover abnormal nodes in the asset process in time and correct them, but also help enterprises discover potential risk factors in advance, and guarantee the efficiency and accuracy of asset management.

Description

Technical Field

[0001] This invention belongs to the field of asset process monitoring technology, specifically relating to an intelligent asset process monitoring method and system based on artificial intelligence technology. Background Technology

[0002] In enterprise and institutional operations management, effective monitoring of asset processes is a key link in ensuring asset security, improving utilization efficiency, and reducing operating costs. Asset processes span the entire lifecycle of assets, from procurement, warehousing, requisition, allocation, maintenance to disposal, involving the processing and analysis of large amounts of data. Therefore, ensuring the integrity, accuracy, and security of asset process data is of paramount importance.

[0003] Currently, many enterprises and institutions still rely on traditional asset process monitoring methods that combine manual record-keeping with periodic inspections. Manual record-keeping is not only inefficient and time-consuming, but also prone to data entry errors and untimely information updates. During asset requisition and transfer, manual registration may result in omissions or errors, leading to unclear asset whereabouts and difficulty in accurate tracking. For example, during asset transfers, untimely or inaccurate information transmission may cause assets to go missing. In asset maintenance, it's impossible to accurately predict maintenance needs and timelines, resulting in increased maintenance costs and reduced asset utilization efficiency. Furthermore, periodic asset inspections cannot provide real-time monitoring; problems are often only discovered after they occur, failing to prevent potential risks in a timely manner and failing to meet the demands of efficient modern enterprise operations. Summary of the Invention

[0004] The purpose of this invention is to address the aforementioned problems in the existing technology by providing an intelligent asset process monitoring method and system based on artificial intelligence technology, which can monitor abnormal nodes in the asset process in real time and improve asset management efficiency.

[0005] To achieve the above objectives, the technical solution of the present invention is as follows:

[0006] In a first aspect, the present invention provides an intelligent asset process monitoring method based on artificial intelligence technology, the intelligent asset process monitoring method comprising:

[0007] Obtain historical and real-time asset data;

[0008] Extract standard asset feature vectors under normal conditions based on historical asset data and using a pre-trained support vector machine model;

[0009] Real-time asset feature vectors are extracted from real-time asset data based on a pre-defined gradient boosting tree model.

[0010] Perform high-dimensional mapping and anomaly analysis on standard asset feature vectors and real-time asset feature vectors to determine whether there are abnormal nodes in the current real-time asset data;

[0011] If there are abnormal nodes in the asset process of real-time asset data, then correct the abnormal data of the asset data process nodes.

[0012] If no abnormal nodes are found in the real-time asset data, the real-time asset data will be processed using blockchain data processing steps, and the processed real-time asset data will be uploaded to the pre-built traceability chain.

[0013] The extraction of real-time asset feature vectors from real-time asset data based on a preset gradient boosting tree model includes:

[0014] Iterate through each tree model in the preset gradient boosting tree model and input real-time asset data into the tree model;

[0015] In each tree model, real-time asset data is allocated to the leaf nodes of the tree model according to the preset decision rules, and the code of each leaf node in the tree model is recorded.

[0016] Extract the feature subvector corresponding to the encoding of each leaf node;

[0017] Calculate the feature importance of the feature subvector corresponding to the encoding of each leaf node based on historical asset data;

[0018] Feature vectors with a feature importance greater than a first preset importance threshold are concatenated into a high-dimensional sparse vector to represent the real-time asset feature vector, where the dimension of the real-time asset feature vector is equal to the total number of leaf nodes in the preset tree model.

[0019] The method for extracting real-time asset feature vectors from real-time asset data based on a preset gradient boosting tree model also includes:

[0020] The false positive feature vector is calculated based on the vector fuzzy importance of the high-dimensional sparse vector, and the first preset importance threshold of the gradient boosting tree model is dynamically adjusted based on the false positive feature vector and the dominant profile type of the false positive feature vector.

[0021] The step of calculating the false positive feature vector based on the vector fuzzy importance of high-dimensional sparse vectors, and dynamically adjusting the first preset importance threshold of the gradient boosting tree model based on the false positive feature vector and the dominant profile type of the false positive feature vector, includes:

[0022] The importance of the high-dimensional sparse vectors extracted by the preset gradient boosting tree model is evaluated by fuzzy evaluation to obtain the fuzzy importance of the high-dimensional sparse vectors.

[0023] High-dimensional sparse vectors with a fuzzy importance less than the second preset fuzzy importance are used as false alarm feature vectors.

[0024] Based on graph theory community detection, the dominant profile type of each false alarm feature vector is identified by analyzing the false alarm feature vector.

[0025] The current false alarm rate of the gradient boosting tree model is calculated based on the false alarm feature vector and the dominant profile type.

[0026] Target false alarm rates were set based on receiver operating characteristic curves and corresponding to the dominant profile type.

[0027] The error change rate is calculated based on the error between the current false alarm rate and the target false alarm rate. The nonlinear threshold adjustment amount is then calculated by combining the error change rate and the false alarm rate using a fuzzy logic rule base associated with the dominant profile type.

[0028] The first preset importance threshold of the gradient boosting tree model is dynamically adjusted by combining a non-linear threshold adjustment amount.

[0029] The extraction of standard asset feature vectors under normal conditions based on historical asset data and using a pre-trained support vector machine model includes:

[0030] Linear discriminant analysis algorithm is used to enhance the class distinction of real-time asset data, and clustering algorithm is used to determine the normal status data of historical asset data.

[0031] Fuzzy asset feature vectors are extracted from normal asset status data using an empirical pattern decomposition algorithm.

[0032] A pre-defined support vector machine model is trained using fuzzy asset feature vectors to obtain the trained support vector machine model;

[0033] Historical asset data is input into the trained support vector machine model to extract standard asset feature vectors under normal conditions.

[0034] The extraction of fuzzy asset feature vectors from normal asset status data using the empirical pattern decomposition algorithm includes:

[0035] The empirical mode decomposition algorithm is used to perform empirical mode decomposition on the asset normal state data to obtain multiple intrinsic mode function components of the asset normal state data, and the component energy of each intrinsic mode function component is calculated.

[0036] The component energy is used as the input to the membership function and mapped to a fuzzy set;

[0037] In a fuzzy set, the fuzzy asset feature vector corresponding to each component energy is determined according to pre-constructed fuzzy rules.

[0038] The process of performing high-dimensional mapping and anomaly analysis on standard asset feature vectors and real-time asset feature vectors to determine whether there are abnormal nodes in the current real-time asset data includes:

[0039] The kernel function of the support vector machine is used to map the standard asset feature vector and the real-time asset feature vector to a high-dimensional data space;

[0040] Cluster analysis is performed on the real-time asset feature vectors in the high-dimensional data space, and the discrete asset process state clusters in the real-time asset feature vectors are determined based on the cluster analysis results.

[0041] Calculate the spatial distance between the standard asset feature vector and the discrete asset process state cluster;

[0042] Calculate the local density and neighbor sample distribution characteristics of real-time asset feature vectors in a high-dimensional data space;

[0043] If at least two of the spatial distance, local density, and neighboring sample distribution characteristics do not meet the preset standards, it is determined that there are abnormal nodes in the real-time asset data process.

[0044] If there are abnormal nodes in the real-time asset data flow, the abnormal data in the asset data flow nodes will be corrected, including:

[0045] When spatial distance anomalies exist, the K-nearest neighbor algorithm is used to traverse each discrete asset process state cluster and determine the minimum distance between each discrete asset process state cluster and the nearest normal state cluster.

[0046] The weighted average of discrete asset process state clusters is calculated based on the minimum distance, and the coordinates of discrete asset process state clusters in high-dimensional data space are corrected based on the weighted average.

[0047] When local density anomalies exist, the local anomaly factor algorithm is used to mark low-density regions.

[0048] Oversampling techniques are used to generate composite data for composite points from discrete asset process state clusters in low-density areas.

[0049] Based on the synthetic data, the synthetic points are inserted into the low-density region in chronological order.

[0050] When there are abnormal distribution characteristics of neighboring samples, principal component analysis is performed on the feature values ​​of the normal state clusters adjacent to the discrete asset process state clusters to determine the principal components of the normal state clusters.

[0051] Map the feature values ​​of discrete asset process state clusters to the principal component space and compare whether the principal component distribution of discrete asset process state clusters deviates from a preset distance threshold.

[0052] If the principal component distribution deviates from the distance threshold, the principal components of the normal state cluster are used to correct the discrete asset process state cluster.

[0053] If no abnormal nodes are found in the real-time asset data, then the real-time asset data is processed using blockchain data processing steps, and the processed real-time asset data is uploaded to the pre-built traceability chain, including:

[0054] Obtain information on all asset process nodes in the pre-built traceability chain, including publicly available asset information and private asset information;

[0055] The asset privacy information is encrypted using a preset encryption algorithm to obtain encrypted asset data;

[0056] Based on all asset process nodes of the traceability chain and according to the blockchain consensus mechanism preset by the traceability chain, consensus verification is carried out on public asset information and private asset information.

[0057] If the consensus verification is successful, the public asset information and the private asset information will be broadcast to all asset process nodes.

[0058] Public and private asset information is stored in the blockchain ledger of the traceability chain in the form of blocks.

[0059] Secondly, the present invention provides an intelligent asset process monitoring system based on artificial intelligence technology. The intelligent asset process monitoring system is based on the aforementioned intelligent asset process monitoring method and includes:

[0060] The data acquisition module is used to acquire historical and real-time asset data.

[0061] The first feature extraction module is used to extract standard asset feature vectors under normal conditions based on historical asset data and using a pre-trained support vector machine model.

[0062] The second feature extraction module is used to extract real-time asset feature vectors from real-time asset data based on a preset gradient boosting tree model.

[0063] The anomaly analysis module is used to perform high-dimensional mapping and anomaly analysis on standard asset feature vectors and real-time asset feature vectors to determine whether there are abnormal nodes in the current real-time asset data.

[0064] The data correction module is used to correct abnormal data in the asset data process nodes when there are abnormal nodes in the real-time asset data.

[0065] The data storage module is used to process real-time asset data using blockchain data processing steps when there are no abnormal nodes in the asset process, and then upload the processed real-time asset data to the pre-built traceability chain.

[0066] The second feature extraction module is used to extract real-time asset feature vectors from real-time asset data based on a preset gradient boosting tree model according to the following steps:

[0067] Iterate through each tree model in the preset gradient boosting tree model and input real-time asset data into the tree model;

[0068] In each tree model, real-time asset data is allocated to the leaf nodes of the tree model according to the preset decision rules, and the code of each leaf node in the tree model is recorded.

[0069] Extract the feature subvector corresponding to the encoding of each leaf node;

[0070] Calculate the feature importance of the feature subvector corresponding to the encoding of each leaf node based on historical asset data;

[0071] Feature vectors with a feature importance greater than a first preset importance threshold are concatenated into a high-dimensional sparse vector to represent the real-time asset feature vector, where the dimension of the real-time asset feature vector is equal to the total number of leaf nodes in the preset tree model.

[0072] The second feature extraction module is used to calculate the false alarm feature vector based on the vector fuzzy importance of the high-dimensional sparse vector, and dynamically adjust the first preset importance threshold of the gradient boosting tree model based on the false alarm feature vector and the dominant profile type of the false alarm feature vector.

[0073] The second feature extraction module is used to dynamically adjust the first preset importance threshold of the gradient boosting tree model according to the following steps:

[0074] The importance of the high-dimensional sparse vectors extracted by the preset gradient boosting tree model is evaluated by fuzzy evaluation to obtain the fuzzy importance of the high-dimensional sparse vectors.

[0075] High-dimensional sparse vectors with a fuzzy importance less than the second preset fuzzy importance are used as false alarm feature vectors.

[0076] Based on graph theory community detection, the dominant profile type of each false alarm feature vector is identified by analyzing the false alarm feature vector.

[0077] The current false alarm rate of the gradient boosting tree model is calculated based on the false alarm feature vector and the dominant profile type.

[0078] Target false alarm rates were set based on receiver operating characteristic curves and corresponding to the dominant profile type.

[0079] The error change rate is calculated based on the error between the current false alarm rate and the target false alarm rate. The nonlinear threshold adjustment amount is then calculated by combining the error change rate and the false alarm rate using a fuzzy logic rule base associated with the dominant profile type.

[0080] The first preset importance threshold of the gradient boosting tree model is dynamically adjusted by combining a non-linear threshold adjustment amount.

[0081] The first feature extraction module is used to extract standard asset feature vectors under normal conditions based on historical asset data and using a pre-trained support vector machine model according to the following steps:

[0082] Linear discriminant analysis algorithm is used to enhance the class distinction of real-time asset data, and clustering algorithm is used to determine the normal status data of historical asset data.

[0083] Fuzzy asset feature vectors are extracted from normal asset status data using an empirical pattern decomposition algorithm.

[0084] A pre-defined support vector machine model is trained using fuzzy asset feature vectors to obtain the trained support vector machine model;

[0085] Historical asset data is input into the trained support vector machine model to extract standard asset feature vectors under normal conditions.

[0086] The first feature extraction module is used to extract fuzzy asset feature vectors from the normal asset state data using an empirical pattern decomposition algorithm according to the following steps:

[0087] The empirical mode decomposition algorithm is used to perform empirical mode decomposition on the asset normal state data to obtain multiple intrinsic mode function components of the asset normal state data, and the component energy of each intrinsic mode function component is calculated.

[0088] The component energy is used as the input to the membership function and mapped to a fuzzy set;

[0089] In a fuzzy set, the fuzzy asset feature vector corresponding to each component energy is determined according to pre-constructed fuzzy rules.

[0090] The anomaly analysis module is used to perform high-dimensional mapping and anomaly analysis on the standard asset feature vector and the real-time asset feature vector according to the following steps, to determine whether there are any abnormal nodes in the current real-time asset data, including:

[0091] The kernel function of the support vector machine is used to map the standard asset feature vector and the real-time asset feature vector to a high-dimensional data space;

[0092] Cluster analysis is performed on the real-time asset feature vectors in the high-dimensional data space, and the discrete asset process state clusters in the real-time asset feature vectors are determined based on the cluster analysis results.

[0093] Calculate the spatial distance between the standard asset feature vector and the discrete asset process state cluster;

[0094] Calculate the local density and neighbor sample distribution characteristics of real-time asset feature vectors in a high-dimensional data space;

[0095] If at least two of the spatial distance, local density, and neighboring sample distribution characteristics do not meet the preset standards, it is determined that there are abnormal nodes in the real-time asset data process.

[0096] The data correction module is used to correct abnormal data in the asset data flow nodes according to the following steps:

[0097] When spatial distance anomalies exist, the K-nearest neighbor algorithm is used to traverse each discrete asset process state cluster and determine the minimum distance between each discrete asset process state cluster and the nearest normal state cluster.

[0098] The weighted average of discrete asset process state clusters is calculated based on the minimum distance, and the coordinates of discrete asset process state clusters in high-dimensional data space are corrected based on the weighted average.

[0099] When local density anomalies exist, the local anomaly factor algorithm is used to mark low-density regions.

[0100] Oversampling techniques are used to generate composite data for composite points from discrete asset process state clusters in low-density areas.

[0101] Based on the synthetic data, the synthetic points are inserted into the low-density region in chronological order.

[0102] When there are abnormal distribution characteristics of neighboring samples, principal component analysis is performed on the feature values ​​of the normal state clusters adjacent to the discrete asset process state clusters to determine the principal components of the normal state clusters.

[0103] Map the feature values ​​of discrete asset process state clusters to the principal component space and compare whether the principal component distribution of discrete asset process state clusters deviates from a preset distance threshold.

[0104] If the principal component distribution deviates from the distance threshold, the principal components of the normal state cluster are used to correct the discrete asset process state cluster.

[0105] The data storage module is used to process real-time asset data according to the following steps, and upload the processed real-time asset data to the pre-built traceability chain:

[0106] Obtain information on all asset process nodes in the pre-built traceability chain, including publicly available asset information and private asset information;

[0107] The asset privacy information is encrypted using a preset encryption algorithm to obtain encrypted asset data;

[0108] Based on all asset process nodes of the traceability chain and according to the blockchain consensus mechanism preset by the traceability chain, consensus verification is carried out on public asset information and private asset information.

[0109] If the consensus verification is successful, the public asset information and the private asset information will be broadcast to all asset process nodes.

[0110] Public and private asset information is stored in the blockchain ledger of the traceability chain in the form of blocks.

[0111] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0112] 1. The intelligent asset process monitoring method of this invention utilizes a pre-trained support vector machine model to analyze historical asset data and extract standard asset feature vectors under normal conditions, establishing a reliable reference standard for judging the normality of asset data. Based on a preset gradient boosting tree model, it extracts feature vectors from real-time asset data, effectively capturing feature information in the real-time data. Combining these two feature vectors for judgment can accurately identify whether there are abnormal asset process nodes in the current real-time asset data, improving the accuracy and reliability of anomaly detection and timely identifying potential asset risks. When abnormal asset process nodes are detected in real-time asset data, the abnormal nodes are marked and the abnormal data is corrected, helping to ensure the accuracy of asset data. Ensuring data integrity and completeness prevents erroneous data from further spreading and impacting the asset process, improving data quality and providing a more reliable data foundation for subsequent asset analysis and decision-making. For real-time asset data without any abnormal nodes in the asset process, blockchain data processing steps are used and the data is uploaded to a pre-built traceability chain, making the asset data immutable and traceable, enhancing data security and credibility. When problems occur, the traceability chain can quickly and accurately trace the source and flow of asset data, facilitating auditing and supervision. It can also promptly identify and address abnormal nodes in the asset process, helping enterprises to identify potential risk factors in advance, take corresponding measures for prevention and control, and ensure the efficiency and accuracy of asset management.

[0113] 2. The intelligent asset process monitoring method of the present invention extracts real-time asset feature vectors from real-time asset data based on a preset gradient boosting tree model. Then, it calculates false alarm feature vectors based on the vector fuzzy importance of high-dimensional sparse vectors. Based on the false alarm feature vectors and the dominant profile type of the false alarm feature vectors, it dynamically adjusts the first preset importance threshold of the gradient boosting tree model. This allows the threshold adjustment to adaptively reflect the comprehensive characteristics of the current false alarm rate, error change rate, and system dynamic mode. This enables fine-tuning of the model's false alarm behavior under different dominant profile types, making the false alarm rate close to the target false alarm rate corresponding to the dominant profile type. Ultimately, this improves the adaptive performance and stability of the gradient boosting tree model under complex working conditions. Attached Figure Description

[0114] Figure 1 This is a flowchart of the intelligent asset process monitoring method described in this invention.

[0115] Figure 2 This is a structural block diagram of the intelligent asset process monitoring system described in this invention. Detailed Implementation

[0116] The present invention will now be described in further detail with reference to specific embodiments and accompanying drawings.

[0117] Example 1:

[0118] See Figure 1 An intelligent monitoring method for asset processes based on artificial intelligence technology is implemented in the following steps:

[0119] S1. Obtain historical asset data and real-time asset data.

[0120] Specifically, historical asset data can be obtained through an enterprise information query platform, while real-time asset data can be obtained from the enterprise's internal financial system. Asset data can include procurement data, operation and maintenance data, and replacement data. Procurement data refers to information generated during the asset acquisition phase, covering the entire process from demand to procurement completion. Operation and maintenance data refers to information generated during asset use, such as maintenance, upkeep, and repair, reflecting the asset's operational status and health condition. Replacement data refers to information generated when assets are scrapped, sold, updated, or transferred, reflecting the asset's exit or transfer process. These three types of data together constitute the full lifecycle management information of the asset process, providing enterprises with a comprehensive view of asset status and supporting decision-making, cost control, and risk management.

[0121] S2. Extract standard asset feature vectors under normal conditions based on historical asset data and using a pre-trained support vector machine model.

[0122] Specifically, historical asset data refers to the performance data of an asset over a certain period of time, including price, trading volume, market capitalization, and financial data. This data reflects the asset's performance under different market conditions and is an important basis for analyzing asset characteristics and predicting future trends. The Support Vector Machine (SVM) model is a supervised machine learning algorithm primarily used for classification and regression tasks. It finds an optimal hyperplane to classify data into different categories or fits data for regression prediction. In machine learning, a feature vector is a set of numerical features used to describe a data point. For asset data, feature vectors may include key indicators such as price, trading volume, and market capitalization. A standard feature vector refers to a set of features that accurately describes the characteristics of an asset under normal conditions and can be used for subsequent asset management tasks, such as condition monitoring, fault diagnosis, and performance evaluation.

[0123] Specifically, first, historical asset data is collected and preprocessed, such as removing noise, handling missing values, and normalizing or standardizing, to improve the model's performance and accuracy. A subset of the historical asset data (usually labeled data, i.e., data known to be in normal or abnormal states) is used to train the support vector machine (SVM) model. During training, the SVM model attempts to find an optimal hyperplane to maximize the margin between positive and negative samples (i.e., normal and abnormal states). By adjusting the SVM parameters, the model's performance is optimized, and standard asset feature vectors in normal states are extracted. Next, the pre-trained SVM is used to perform classification or regression prediction on the remaining historical asset data. For data points classified as normal, their corresponding feature vectors are extracted. These feature vectors represent the standard characteristics of assets in normal states and can serve as a benchmark for subsequent analysis or prediction.

[0124] S3. Extract real-time asset feature vectors from real-time asset data based on a preset gradient boosting tree model.

[0125] Specifically, extracting real-time asset feature vectors from real-time asset data based on a pre-trained gradient boosting tree model refers to using a pre-trained gradient boosting tree model to extract features from real-time asset data, thereby obtaining a set of feature representations that reflect the current state of the asset. Gradient boosting trees are an ensemble learning method that constructs a strong learner by combining multiple weak learners (usually decision trees). It uses an iterative approach, training a new decision tree at each step based on the residual of the previous model (i.e., the difference between the predicted and actual values), thus gradually reducing prediction errors. Real-time asset data refers to data such as the operational status and performance indicators collected from assets in real time. The pre-processed real-time asset data is input into the pre-trained gradient boosting tree model. In the gradient boosting tree model, each leaf node of a decision tree can be considered a feature. For real-time data, the model traverses all decision trees, eventually landing on a leaf node, thus obtaining a set of leaf node indices (or encoded values). These indices or encoded values ​​can be combined into a feature vector, reflecting the feature representation of the data in the model. The gradient boosting tree model can also provide feature importance information. When extracting feature vectors, the original features can be weighted according to their importance to obtain weighted feature vectors, thereby emphasizing features that contribute more to the model's prediction. In this invention, the gradient boosting tree model is used as a feature extractor, and its output (such as prediction probability, residuals, etc.) or the representation of intermediate layers (such as the encoding of decision paths) is used as feature vectors for real-time asset data. These feature vectors are implicit representations of the internal structure of the data by the model, capable of capturing complex patterns and relationships in the data. The extracted feature vectors are output for subsequent analysis, decision-making, or model training.

[0126] S4. Perform high-dimensional mapping and anomaly analysis on the standard asset feature vector and the real-time asset feature vector to determine whether there are abnormal nodes in the current real-time asset data.

[0127] Specifically, the standard asset feature vector is extracted from historical asset data and represents a set of numerical characteristics of an asset under normal conditions. This may include key indicators such as price, trading volume, market capitalization, and financial data, reflecting the asset's performance characteristics under normal market conditions. The real-time asset feature vector, on the other hand, represents the numerical characteristics of the asset data at the current moment, reflecting the asset's current real-time state. Comparing the real-time asset feature vector with the standard asset feature vector reveals differences, which may manifest as significant changes in certain feature values, such as sudden large price fluctuations or abnormal increases or decreases in trading volume. Furthermore, by setting certain thresholds or using statistical methods, it can be determined whether the difference between the real-time asset feature vector and the standard feature vector exceeds the normal range. If it does, it may indicate an abnormal node in the asset process. When there is a significant difference between the real-time asset feature vector and the standard feature vector, further analysis can be conducted to identify the specific factors causing these differences, such as market fluctuations, internal business problems, and policy changes. By analyzing these factors, potential abnormal nodes in the asset process can be identified. After identifying abnormal nodes, risk assessments can be performed to determine the potential impact of these anomalies on asset value and business operations.

[0128] S5. If there are abnormal nodes in the asset process, correct the abnormal data in the asset data process nodes.

[0129] Specifically, anomaly nodes in the asset process refer to data points or segments that significantly deviate from the normal state, potentially indicating errors, system failures, or other abnormal situations. Anomalies are detected by real-time monitoring of asset data, utilizing preset rules, thresholds, or machine learning models (such as support vector machines and anomaly detection algorithms). When real-time asset data differs significantly from standard asset feature vectors (extracted from historical data), an anomaly detection mechanism may be triggered. This mechanism includes anomaly labeling and correcting abnormal data. Anomaly labeling clearly identifies which data points or segments are abnormal, facilitating further analysis and processing. Anomaly labeling can be achieved by adding anomaly label fields to the data system to mark abnormal data points or segments, or by generating anomaly reports detailing the time, location, and anomaly type of the abnormal data points. Correcting abnormal data restores the accuracy and integrity of the data, ensuring the normal operation of the asset data process. Correction methods can include data correction and data interpolation. Data correction refers to the direct correction of obvious erroneous data (such as entry errors, transmission errors, etc.). Data interpolation refers to the estimation and filling of reasonable data values ​​for missing or abnormal data points using interpolation methods (such as mean interpolation, regression interpolation, etc.).

[0130] S6. If there are no abnormal nodes in the asset process, the real-time asset data will be processed using blockchain data processing steps, and the processed real-time asset data will be uploaded to the pre-built traceability chain.

[0131] Specifically, if real-time asset data does not trigger any anomaly monitoring mechanisms, it is confirmed that there are no abnormal nodes in the asset flow. Next, the real-time asset data is processed using blockchain data processing steps and then uploaded to the traceability chain. The traceability chain is a pre-built blockchain network specifically designed to record and track the source and flow of asset data. It possesses characteristics such as immutability and traceability, providing reliable security for asset data. The constructed transaction is broadcast to the traceability chain network, where nodes verify the transaction to ensure its legality and validity. Once the transaction is verified, the real-time asset data is officially recorded on the traceability chain. After the data is on the chain, its source, flow, and other information become transparent and tamper-proof. Through the traceability chain, the source and flow of real-time asset data can be easily traced, helping to identify potential problems or risks and providing strong support for subsequent audits and compliance checks.

[0132] This invention utilizes a pre-trained support vector machine model to analyze historical asset data and extract standard asset feature vectors under normal conditions, establishing a reliable reference standard for judging the normality of asset data. Simultaneously, it extracts feature vectors from real-time asset data based on a pre-defined gradient boosting tree model, effectively capturing characteristic information in real-time data. Combining these two feature vectors allows for accurate identification of any abnormal nodes in the current real-time asset data process, improving the accuracy and reliability of anomaly detection and enabling timely discovery of potential asset risks. When an abnormal node is detected in the real-time asset data process, it is marked and corrected, helping to ensure the accuracy and integrity of the asset data. This prevents erroneous data from further propagating and impacting the asset process, improving data quality and providing a more reliable data foundation for subsequent asset analysis and decision-making. For real-time asset data without abnormal nodes, blockchain data processing steps are employed, and the data is uploaded to a pre-built traceability chain. The application of blockchain technology gives asset data immutability and traceability, enhancing data security and credibility. The traceability chain enables quick and accurate tracking of the source and flow of asset data, facilitating auditing and supervision. It also allows for the timely detection and handling of abnormal nodes in the asset process, helping enterprises to identify potential risk factors in advance, take corresponding measures for prevention and control, and ensure the efficiency and accuracy of asset management.

[0133] In another implementation, the step of extracting standard asset feature vectors under normal conditions based on historical asset data and utilizing a pre-trained support vector machine model includes:

[0134] S21. Enhance the class distinction of real-time asset data through linear discriminant analysis algorithm, and use clustering algorithm to determine the normal status data of historical asset data;

[0135] S22. Extract fuzzy asset feature vectors from normal asset state data using an empirical pattern decomposition algorithm;

[0136] S23. Use fuzzy asset feature vectors to train a pre-set support vector machine model to obtain the trained support vector machine model;

[0137] S24. Input historical asset data into the trained support vector machine model to extract standard asset feature vectors under normal conditions.

[0138] Specifically, the linear discriminant analysis (LDA) algorithm is a supervised learning algorithm that projects high-dimensional data into a low-dimensional space, separating data of different categories as much as possible and clustering data of the same category as much as possible, thereby enhancing category distinguishability and improving the separability between different asset categories or states. The clustering algorithm is an unsupervised learning algorithm that aims to divide data into multiple clusters, ensuring high similarity among data within the same cluster and low similarity between data in different clusters. Since real-time asset data in an asset management system may contain various types of asset status information (such as normal, warning, fault, etc.), using the LDA algorithm can enhance the distinguishability between these different status categories, making subsequent classification or anomaly detection tasks more accurate and efficient. Specifically, S21 includes: collecting real-time asset data and labeling each data point with its category (e.g., normal, warning, fault); training a linear discriminant analysis (LDA) model using the labeled data to find the optimal projection direction; projecting new real-time asset data onto the trained LDA direction to obtain a low-dimensional feature representation. In the low-dimensional space, data points of different categories will be more dispersed, facilitating subsequent classification or anomaly detection; next, using a clustering algorithm to determine the normal asset status data from historical asset data. Historical asset data may contain a large number of asset operation records, most of which reflect the normal status of the assets. The clustering algorithm can group these normal status data points together to form normal status data clusters. Once the normal status data clusters of the assets are determined, the new real-time asset data can be compared with these clusters. If the new data points have a low similarity to the normal status clusters, it may indicate that the assets have an anomaly. By collecting historical asset data and performing necessary preprocessing, selecting appropriate clustering algorithms (such as K-means, DBSCAN, etc.) and setting suitable parameters according to the characteristics of the data, the clustering algorithm is run to divide the historical asset data into several clusters. By analyzing the clustering results, the data clusters representing the normal state of the assets can be identified, thereby obtaining the normal state data of the assets.

[0139] Specifically, the Empirical Mode Decomposition (EMD) algorithm is an adaptive signal processing method used to decompose nonlinear, non-stationary signals into several intrinsic mode functions (IMFs) and a residual term. Asset normal state data refers to data collected under normal asset operation conditions, reflecting the asset's behavior patterns in a healthy state. Since asset state data in asset management systems is often nonlinear and non-stationary, the EMD algorithm can effectively extract feature information from these signals. After obtaining the IMFs, the energy of each IMF component is calculated. The energy of each IMF can be calculated using a sum of squares or integral method. Next, the energy is mapped to a fuzzy set, with the energy of each IMF used as the input to a membership function. The pre-constructed rule refers to determining fuzzy features based on membership degrees. For example, if IMF1 has "high" energy and IMF2 has "medium" energy, the feature vector is [high, medium]. Based on the energy membership degrees of all IMFs, a fuzzy feature vector is formed, resulting in a fuzzy feature vector describing the asset's normal state. Specifically, S22 includes: First, obtaining fuzzy asset feature vectors can be achieved by Empirical Pattern Decomposition (EMD) and fuzzy processing, extracting fuzzy feature vectors from normal asset state data, where each sample corresponds to a fuzzy feature vector, for example: [low (0.5), medium (0.8), medium (0.2), low (1.0), low (0.6)]; Second, multiple fuzzy asset feature vectors are used to form a training sample set and set corresponding asset state labels, such as normal (1) or abnormal (0), for example, sample 1: feature vector [Low (0.5), Medium (0.8), Medium (0.2), Low (1.0), Low (0.6)], Label 1 (Normal); Sample 2: Feature vector [High (0.9), Low (0.3), Medium (0.5), Low (0.7), Low (0.4)], Label 0 (Abnormal); Then, the fuzzy feature vector and the corresponding label are input into the preset support vector machine model. The model uses an optimization algorithm (such as Sequence Minimum Optimization Algorithm SMO) to find the optimal hyperplane, maximize the margin between different categories, and finally obtain the trained SVM model.

[0140] Historical asset data is input into the trained support vector machine model. Based on business needs or model performance, the extracted features are filtered to retain the features that most accurately describe the normal state. The filtered features are combined into a standard feature vector for subsequent asset management tasks. This can promptly detect abnormal asset states and issue early warnings, thereby improving the accuracy of asset process monitoring.

[0141] In another implementation, the step of extracting the fuzzy asset feature vector from the asset normal state data using the empirical pattern decomposition algorithm includes:

[0142] S221. Use the empirical mode decomposition algorithm to perform empirical mode decomposition on the asset normal state data to obtain multiple intrinsic mode function components of the asset normal state data, and calculate the component energy of each intrinsic mode function component.

[0143] S222. Map the component energy as the input of the membership function to the fuzzy set;

[0144] S223. In a fuzzy set, determine the fuzzy asset feature vector corresponding to the energy of each component according to the pre-constructed fuzzy rules.

[0145] Specifically, the empirical mode decomposition is an adaptive signal processing method that can decompose complex nonlinear and non-stationary signals into multiple intrinsic mode functions (IMFs). S221 specifically includes: after obtaining multiple IMF components, the square of each IMF can be calculated first, and then the squared value can be integrated to obtain the energy value of each IMF. On the one hand, the energy value quantifies the contribution of each IMF component to the original signal; on the other hand, since the energy distribution of each IMF has a certain regularity under normal conditions, and abnormal conditions may lead to changes in the energy distribution, the energy value can also be used as a feature for subsequent pattern recognition or state monitoring.

[0146] Specifically, the membership function is a core concept in fuzzy set theory, used to describe the degree to which an element belongs to a certain fuzzy set. The membership function ranges from [0, 1], with a larger value indicating a higher degree of belonging to the set. S222 specifically includes: defining fuzzy sets related to IMF energy values ​​according to application requirements; for example, low energy indicates low energy of the IMF component; medium energy indicates medium energy of the IMF component; and high energy indicates high energy of the IMF component. A corresponding membership function is designed for each fuzzy set. The membership function type can be triangular, trapezoidal, Gaussian, etc. The energy Ei of each IMF component is substituted into the corresponding membership function to calculate its membership value to each fuzzy set. Through the membership function, the specific numerical value of the component energy can be mapped to the corresponding fuzzy set. For example, the fuzzy set can be designated as the "high energy" set, and a membership function can be designed for it. When the component energy value is input into the membership function, the function outputs a value between 0 and 1, indicating the degree to which the component energy belongs to the "high-energy" fuzzy set. In this embodiment, fuzzy sets are an extension of classical sets, allowing elements to belong to the set to a certain degree.

[0147] Specifically, the fuzzy rule is a reasoning rule based on fuzzy logic, used to map inputs to outputs. Here, the fuzzy rule is used to map the energy of IMF components to fuzzy asset feature vectors. Specifically, based on practical application requirements, a set of fuzzy rules is constructed. These rules define the mapping relationship between IMF component energy and fuzzy asset feature vectors. For example, when the IMF component energy is low, the corresponding fuzzy asset feature vector can be defined as "low risk"; when the energy is high, the corresponding fuzzy asset feature vector is defined as "high risk". The energy of each IMF component is used as input and substituted into the pre-constructed fuzzy rules. Through fuzzy reasoning, the fuzzy asset feature vector corresponding to each IMF component energy is obtained. The fuzzy asset feature vectors corresponding to all IMF component energies are combined to form a complete fuzzy asset feature vector, which describes the distribution and characteristics of the original signal in the fuzzy asset feature space.

[0148] By extracting fuzzy asset feature vectors from normal asset status data using empirical pattern decomposition algorithms, we can reflect the stability and degree of abnormality of asset operation status, which helps to identify potential faults or anomalies and improve the accuracy of status monitoring and fault diagnosis.

[0149] In another implementation, the extraction of real-time asset feature vectors from real-time asset data based on a preset gradient boosting tree model includes:

[0150] S31. Traverse each preset tree model in the preset gradient boosting tree model and input the real-time asset data into the preset tree model;

[0151] S32. In each preset tree model, real-time asset data is allocated to the leaf nodes of the preset tree model according to the preset decision rules, and the code of each leaf node in the preset tree model is recorded.

[0152] S33. Extract the feature sub-vector corresponding to the encoding of each leaf node;

[0153] S34. Calculate the feature importance of the feature sub-vector corresponding to the encoding of each leaf node based on historical asset data;

[0154] S35. Concatenate feature vectors with a feature importance greater than the first preset importance threshold into a high-dimensional sparse vector. The high-dimensional sparse vector is used to represent the real-time asset feature vector, where the dimension of the real-time asset feature vector is equal to the total number of leaf nodes in the preset tree model.

[0155] Specifically, S31 includes: traversing each preset tree model in the preset gradient boosting tree model and inputting real-time asset data into each tree of the preset gradient boosting tree model to prepare for subsequent feature extraction. By traversing all tree models, the information in the model can be fully utilized to capture the performance of asset data in different feature spaces.

[0156] Specifically, S32 includes: Since the gradient boosting tree consists of multiple decision trees, the preset tree model refers to each decision tree in the gradient boosting tree model, and each decision tree is trained based on the residual of the previous tree. Real-time asset data is sequentially input into each tree, and each tree independently processes and makes decisions on the data. The preset decision rules refer to a series of decision rules that each decision tree learns based on the training data during the training process. These rules are usually expressed as node splitting conditions in the tree structure. For example, if the value of a certain feature is less than a certain threshold, the left subtree is traversed; otherwise, the right subtree is traversed. Next, the real-time asset data starts from the root node of the tree and sequentially passes through each internal node according to the decision rules. At each internal node, the data is assigned to the left or right subtree according to the conditions satisfied by the feature value. Finally, the data reaches a leaf node. The leaf node is the end of the tree structure, has no child nodes, and each leaf node represents a specific decision result or feature combination. Leaf node encoding is used for convenient representation and subsequent processing; each leaf node is usually assigned a unique code (e.g., an integer number). When real-time asset data is assigned to a leaf node, the system records the encoding of that leaf node. The leaf node encoding reflects the decision path and final position of the data within the tree model. Different leaf node encodings correspond to different feature representations or decision outcomes of the data. Subsequently, by assigning real-time asset data to leaf nodes, the data is effectively mapped to a discrete feature space. Each leaf node can be viewed as a feature bucket, representing a feature pattern of the data within the tree model. The leaf node encoding provides a degree of explanation for the model's decision-making process. By examining which leaf node the data falls into, we can understand how the model classifies or regresses the data based on features. The recorded leaf node encodings can be used for subsequent steps such as feature extraction, feature importance calculation, and high-dimensional sparse vector construction, supporting the feature representation and analysis of real-time assets.

[0157] Specifically, extracting the feature vector corresponding to the encoding of each leaf node refers to obtaining the feature information associated with each leaf node. S33 includes: each leaf node is associated with certain feature vectors during training (these feature vectors may be combinations or transformations of the original features), and extracting these feature vectors corresponding to the encoding of the leaf node. The feature vectors reflect the feature patterns of the data at a specific leaf node, which helps to understand the characteristics of the data at that node.

[0158] Specifically, calculating the feature importance of each leaf node's encoded feature vector based on historical asset data refers to assessing the importance of each feature vector to the model's predictions. S34 includes: using historical asset data to calculate the contribution or importance of each feature vector in the model, which can be achieved through various methods, such as gain-based feature importance or SHAP value-based interpretation methods. Feature importance reflects the degree of influence of feature vectors on the model's prediction results and helps to identify key features.

[0159] Specifically, S35 includes: First, setting an importance threshold and selecting feature vectors with importance higher than the threshold. These feature vectors are then concatenated into a high-dimensional vector, the dimension of which is equal to the total number of leaf nodes in the gradient boosting tree model. Since not all leaf nodes contribute to the current real-time asset data, this vector is sparse. The high dimensionality indicates a comprehensive representation of the asset data's feature information in the model; the sparsity reduces computational and storage requirements while highlighting key features.

[0160] Steps S31-S35 effectively extract highly informative feature vectors from real-time asset data. These vectors not only reflect the asset's performance under different conditions but also improve feature reliability and processing efficiency through feature importance filtering and sparse vector construction.

[0161] In another implementation, the extraction of real-time asset feature vectors from real-time asset data based on a preset gradient boosting tree model further includes:

[0162] S36. Calculate the false alarm feature vector based on the vector fuzzy importance of the high-dimensional sparse vector, and dynamically adjust the first preset importance threshold of the gradient boosting tree model based on the false alarm feature vector and the dominant profile type of the false alarm feature vector.

[0163] Specifically, S36 includes:

[0164] S361. Perform fuzzy importance evaluation on the high-dimensional sparse vectors extracted from the preset gradient boosting tree model to obtain the vector fuzzy importance of the high-dimensional sparse vectors.

[0165] S362. Use high-dimensional sparse vectors with vector fuzzy importance less than the second preset fuzzy importance as false alarm feature vectors.

[0166] S363. Based on graph theory community detection, analyze the false alarm feature vectors to identify the dominant profile type of each false alarm feature vector;

[0167] S364. Calculate the current false alarm rate of the gradient boosting tree model based on the false alarm feature vector and the dominant profile type;

[0168] S365. Set the target false alarm rate corresponding to the dominant profile type by combining the receiver operating characteristic curve;

[0169] S366. The error change rate is calculated based on the error between the current false alarm rate and the target false alarm rate. The nonlinear threshold adjustment amount is calculated by combining the error change rate and the false alarm rate using a fuzzy logic rule base associated with the dominant profile type.

[0170] S367. Dynamically adjust the first preset importance threshold of the gradient boosting tree model by combining non-linear threshold adjustment amount.

[0171] Specifically, the importance assessment refers to using fuzzy logic to evaluate the importance of each feature (or dimension) in a high-dimensional sparse vector; the vector fuzzy importance refers to the degree of fuzzy importance of each feature in a high-dimensional sparse vector obtained through the fuzzy assessment method; S361 specifically includes: for high-dimensional sparse vector features extracted from a preset gradient boosting tree model, first extract their importance values, such as indicators based on split point gain, number of covered samples, etc.; then, input the extracted importance values ​​into a membership function to obtain the degree to which each feature belongs to different fuzzy sets, combine the membership degrees of all fuzzy sets to generate a fuzzy importance vector for each feature, the fuzzy importance vector represents a fuzzy set or vector representing the importance of each feature of each high-dimensional sparse vector, so as to quantify the contribution of each feature in the high-dimensional sparse vector to the model prediction, and provide a basis for subsequent steps.

[0172] Specifically, using high-dimensional sparse vectors with a fuzzy importance less than a second preset fuzzy importance as false positive feature vectors is to identify feature vectors that contribute little to the model's prediction, i.e., false positive features. In this embodiment, the second preset fuzzy importance can be set quantitatively by statistically analyzing the fuzzy importance of all high-dimensional sparse vectors extracted by the gradient boosting tree model to obtain their numerical distribution (e.g., sorting all fuzzy importance values ​​from smallest to largest). Based on the business's accuracy requirements for false positives, the corresponding quantile is selected as the second preset fuzzy importance. For example, if the business wants to select the "lowest 20% of features" as false positive features, then the 20th quantile of fuzzy importance (P20) is taken as the threshold. False positive feature vectors refer to feature vectors that are incorrectly identified as important by the model but actually have low contributions. By comparing the fuzzy importance with the threshold, these "false positive" features are selected. By setting a threshold, features that may have a negative impact on the model are selected, providing a basis for subsequent false positive rate calculation and threshold adjustment.

[0173] Specifically, after determining the false positive feature vectors, graph-based community detection is used to analyze and identify the dominant profile type of each false positive feature vector. In this embodiment, the false positive feature vectors are constructed as a false positive feature vector graph, where each feature vector corresponds to a node in the graph, and the edge weights between nodes are determined based on the similarity between feature vectors, thus constructing the false positive feature vector graph. Subsequently, a community detection algorithm is used to partition the graph, clustering false positive feature vectors with high similarity into the same community, with each community corresponding to a typical false positive feature pattern. Finally, by statistically analyzing the distribution of false positive feature vectors within each community, representative feature vectors of each community are extracted and used as the dominant profile type of that community. The dominant profile type reflects the core feature distribution and typical behavior of the false positive vectors, and each dominant profile type corresponds to the main feature combination of the false positive feature vectors within the community, which can be used to summarize the commonalities and representativeness of this type of false positive. In specific implementation, the dominant profile types include, but are not limited to: intensity type, i.e., false alarm signals with large amplitudes that may be caused by noise or sporadic anomalies; pattern type, i.e., signals with specific waveform, spectrum, or sequence characteristics that are abnormal but whose overall trend conforms to normal patterns; local anomaly type, i.e., signals with anomalies in only some indicators or local areas; persistent deviation type, i.e., signals that continuously deviate from the normal range without abrupt changes; and composite type, i.e., complex false alarm types that simultaneously exhibit multi-dimensional characteristic anomalies. In this embodiment, by analyzing its amplitude, local anomalies, and time series deviation trends, its corresponding predefined dominant profile type can be determined, including intensity type, local anomaly type, persistent deviation type, and composite type. Specifically, the mean, maximum value, and standard deviation of each component of the representative feature vector are calculated and compared with the historical normal average value obtained from the system. If the maximum value or mean is significantly higher than the normal level, the vector is intensity type; that is, if only the amplitude is abnormal while other dimensions are normal, the community determines it to be intensity type. Anomaly detection is performed on each component of the vector, such as determining whether it exceeds the historical normal average value obtained by the system, and the proportion of abnormal components to the total components is calculated. If the proportion is low (e.g., <50%, this is just an example) and concentrated in a specific dimension or local area, it is judged as a local anomaly. Trend analysis is performed on the sequence changes of the feature vector over time; if the vector as a whole deviates from the normal range in one direction for a long period of time (without abrupt changes), it is judged as a persistent shift, which can be quantified as the cumulative time or magnitude of the sliding window mean deviating from the normal range. Composite type determination is when amplitude anomaly, local anomaly, and persistent shift features coexist. This invention can effectively distinguish different types of false alarms by identifying each dominant profile type, providing a basis for false alarm correction, strategy optimization, and system reliability improvement.

[0174] Specifically, after determining the set of false positive feature vectors, the false positive feature vectors are first classified and statistically analyzed based on the dominant profile type to obtain the number of false positive samples corresponding to each dominant profile type. Then, using the total number of true negative samples as a benchmark, the categorical false positive rate for each dominant profile type is calculated, thus forming a false positive rate structure reflecting the distribution characteristics of false positives among different types. Based on the degree of influence of different dominant profile types on the model's discrimination performance, type weights are preset, which can be set based on quantifiable indicators such as the contribution of the dominant profile type to the model's false positive risk, the sensitivity of the business scenario, and the importance of the consequences of false positives. The false positive rates of each type are then weighted and fused with their corresponding weights to obtain a weighted false positive rate that comprehensively reflects the current false positive level of the model. This calculation method enables refined quantification of the false positive performance of the gradient boosting tree model, allowing the model to not only identify the overall false positive level but also distinguish the contribution of different false positive patterns, providing an accurate error quantification basis for subsequent adaptive threshold adjustment.

[0175] Specifically, after determining the dominant profile type of the false alarm feature vector, the receiver operating characteristic (ROC) curve is calculated to set corresponding target false alarm rates for different profile types. First, based on historical sample data, the true positive rate and false positive rate of each dominant profile type at different discrimination thresholds are statistically analyzed to construct the ROC curve for that profile type. Then, the acceptable false alarm range for that profile type is determined based on the performance of different operating points on the ROC curve. Subsequently, based on factors such as the overall system accuracy requirements, the risk level of the profile type in actual business scenarios, and the cost of false alarms, the operating point on the ROC curve that meets the preset detection rate requirements and has the optimal false alarm rate is selected. The false positive rate corresponding to this operating point is used as the target false alarm rate for that dominant profile type. Through this method, differentiated and quantifiable target false alarm rates can be obtained for different profile types, thereby achieving a more refined configuration of false alarm control strategies and improving the stability and reliability of the overall identification system.

[0176] Specifically, the error change rate is calculated based on the error between the current false alarm rate and the target false alarm rate. Then, combining the error change rate with the current false alarm rate, a fuzzy logic rule base associated with the dominant profile type is invoked to output a nonlinear threshold adjustment amount. The purpose of these steps is to introduce a nonlinear adjustment mechanism, enabling the threshold adjustment process to adaptively optimize based on the dynamic characteristics of different false alarm profile types, thereby avoiding the adjustment lag, overcorrection, or oscillation problems caused by single linear feedback control.

[0177] Specifically, the error change rate is calculated based on the error between the current false alarm rate and the target false alarm rate. Then, combining the error change rate and the false alarm rate, a nonlinear threshold adjustment is calculated using a fuzzy logic rule base associated with the dominant profile type. This process includes the following steps:

[0178] A1. Construct a continuous error sequence from the current false alarm rate and error change rate;

[0179] A2. Generate error state vectors from continuous error sequences using phase space reconstruction methods, and generate error phase space trajectories from error state vectors in chronological order.

[0180] A3. Analyze and identify the attractor type corresponding to the error phase space trajectory. The attractor types include periodic, decaying, divergent, and chaotic.

[0181] A4. Call the corresponding fuzzy logic rule base according to the dominant profile type, and input the continuous error sequence into the fuzzy inference system of the fuzzy logic rule base to output the threshold adjustment amount;

[0182] A5. By adjusting the step size of the threshold adjustment amount according to the attractor type, the threshold adjustment strategy corresponding to different attractor types can be obtained.

[0183] A6. Determine the nonlinear threshold adjustment amount through a threshold adjustment strategy.

[0184] Specifically, A1 includes: recording and updating the false alarm rate obtained in the current detection cycle and the error change rate calculated based on the difference between the false alarm rates of two adjacent detection cycles in chronological order to construct a continuous error sequence. The continuous error sequence is used to characterize the evolution of the false alarm rate in the time dimension, where each time node corresponds to a set of error state data containing the false alarm rate and the error change rate.

[0185] Specifically, A2 includes: mapping the false alarm rate and error change rate at different time points in the continuous error sequence to a multi-dimensional state space according to the phase space reconstruction principle, forming an error state vector to characterize the transient features of the error system. Then, constrained by time order, each error state vector is arranged sequentially according to its corresponding time index, ensuring temporal continuity between adjacent error state vectors. Based on this, by sequentially connecting the error state vectors corresponding to adjacent time points in the phase space in a series fashion, a continuous state evolution curve is formed, thus constructing the evolution path of the error system in the multi-dimensional phase space, i.e., the error phase space trajectory. This error phase space trajectory clearly demonstrates the stable region, oscillation region, and abrupt change trend of the error system, making the model threshold adjustment process more stable, accurate, and adaptive.

[0186] Specifically, A3 includes: based on the geometric characteristics, convergence, and long-term evolution patterns of the error phase space trajectory, the trajectory is classified into four attractor types: periodic, decaying, divergent, and chaotic. After constructing the error phase space trajectory, geometric features are extracted, including but not limited to trajectory amplitude features, trajectory density distribution features, and trajectory cycle features. The trajectory amplitude feature calculates the Euclidean distance from each state vector in the error phase space trajectory to the origin or reference equilibrium point of the phase space, obtaining the trajectory radius sequence, which reflects the trend of the state amplitude of the error system changing over time. The trajectory density distribution feature statistically analyzes the distribution density of the trajectory in each region of the phase space, used to determine whether the trajectory is concentrated in a finite region or shows a diffusion trend. The trajectory cycle feature determines whether the trajectory has obvious closed loops or repeated paths, and periodic structures can be identified by calculating the similarity (cosine similarity) of adjacent trajectory segments. Subsequently, through the joint analysis of the trajectory amplitude feature, trajectory density distribution feature, and trajectory cycle feature, the trend of trajectory radius changing over time, the concentration of trajectory distribution in the phase space, and the similarity between trajectory segments, the attractor type classification of the error phase space trajectory is performed. For example, if the trajectory amplitude and density distribution characteristics show a significant decrease in amplitude and a high concentration of density within a finite region, it is classified as an attenuating type. If the trajectory amplitude, density distribution, and cyclic characteristics show periodic fluctuations in amplitude and a density distribution within a circular region, it is classified as a periodic type. When the trajectory is confined to a finite region but exhibits non-periodic and irregular distribution characteristics, it is classified as a chaotic attractor. A periodic attractor indicates that the error system forms a closed loop in phase space, and the false alarm rate fluctuates regularly over time. An attenuating attractor indicates that the trajectory gradually converges to a stable region, and the false alarm rate gradually decreases and tends to stabilize over time. A divergent attractor indicates that the trajectory continuously moves away from the initial region, and the false alarm rate continuously increases or fluctuates abnormally. A chaotic attractor indicates that the trajectory exhibits a complex and disordered pattern in phase space, and the false alarm rate changes nonlinearly over time and is unpredictable in the short term. By identifying the type of attractor in the error phase space trajectory, the system can accurately characterize the dynamic behavior of the false alarm rate, thereby achieving refined control and optimization of the false alarm rate under different dynamic modes.

[0187] Specifically, A4 includes: calling the fuzzy logic rule library corresponding to the identified dominant profile type, and inputting the aforementioned continuous error sequence as input into the fuzzy logic inference system for fuzzy inference. The fuzzy logic rule library pre-sets input-output relationships associated with each dominant profile type. The inputs include the current false alarm rate and error change rate. After fuzzification, rule matching, and inference processing, defuzzification is performed to obtain a nonlinear threshold adjustment amount. In this embodiment, the fuzzy logic rule library is not a general rule set, but a multi-rule sub-library structure constructed according to the dominant profile type. Each dominant profile type corresponds to a dedicated fuzzy logic rule sub-library. First, based on historical data and model operation characteristics, the typical behavior patterns of false alarm rates under different dominant profile types are summarized. For example, intensity profile: the false alarm rate is highly sensitive to threshold changes, and the error change rate fluctuates greatly; local anomaly profile: the false alarm rate changes slowly, and the error change rate is small but persistent; continuous offset profile: the false alarm rate has a directional offset, and the error change rate shows a continuous positive or negative trend. For each dominant profile type, the input and output variables of the fuzzy logic inference system are uniformly defined. The input variables are the current false alarm rate and the error change rate. The current false alarm rate is used to characterize the false alarm level of the current output of the model; the error change rate is used to characterize the trend and adjustment direction of the false alarm rate over time. The output variable is a non-linear threshold adjustment amount, which is used to incrementally or decrementally correct the category discrimination threshold of the gradient boosting tree model. Subsequently, fuzzy linguistic variables and membership functions are constructed for the input and output variables. For example, the current false alarm rate: {low, medium, high}; the error change rate: {decreasing, stable, increasing}; the threshold adjustment amount: {significantly decreasing, slightly decreasing, no adjustment, slightly increasing, significantly increasing}. Each linguistic variable corresponds to at least one membership function (triangular membership function or trapezoidal membership function) to map precise values ​​to fuzzy sets. For each dominant profile type, a set of targeted fuzzy logic rules is constructed, exemplified by the "IF–THEN" format. An example of an intensity-based profile rule is: IF the current false alarm rate is "high" AND the error change rate is "increasing", THEN the threshold adjustment is "slightly increase"; an example of a local anomaly-based profile rule is: IF the current false alarm rate is "medium" AND the error change rate is "stable", THEN the threshold adjustment is "fine-tuning". The threshold adjustment is used to dynamically correct the class discrimination threshold of the gradient boosting tree model, enabling the model's false alarm rate to approach the target false alarm rate corresponding to the dominant profile type, thereby achieving adaptive control and optimization under different false alarm modes.

[0188] Specifically, A5 includes: adjusting the step size of the threshold adjustment amount output by the fuzzy logic inference system based on the type of attractor in the identified error phase space trajectory, to obtain threshold adjustment strategies corresponding to different attractor types. For periodic attractors, since the false alarm system is dynamically stable, the step size of the threshold adjustment amount can be set to a small value or remain unchanged; for decaying attractors, since the system error gradually converges, the step size of the threshold adjustment amount can be appropriately reduced; for divergent attractors, since the system is prone to runaway, the step size of the threshold adjustment amount should be appropriately increased to accelerate the threshold correction speed; for chaotic attractors, since the system exhibits complex nonlinear fluctuations, the step size of the threshold adjustment amount can be dynamically adjusted according to the error trend. Through the above steps, a differentiated threshold adjustment strategy based on attractor type can be realized, thereby enhancing the false alarm control capability of the gradient boosting tree model under different dynamic behaviors.

[0189] Specifically, A6 includes: dynamically adjusting the category discrimination threshold of the gradient boosting tree model based on the threshold adjustment strategy determined by the aforementioned attractor types. The system calculates a nonlinear threshold adjustment amount based on the initial threshold adjustment amount output by fuzzy logic inference, combined with the step size adjustment strategy corresponding to different attractor types. This nonlinear threshold adjustment amount adaptively reflects the comprehensive characteristics of the current false alarm rate, error change rate, and system dynamic mode, thereby achieving fine-tuning of the model's false alarm behavior under different dominant profile types. This makes the false alarm rate close to the target false alarm rate corresponding to the dominant profile type, improving the adaptive performance and stability of the gradient boosting tree model under complex conditions.

[0190] In another embodiment, A5 specifically includes the following steps:

[0191] A51. Calculate the attractor strength index corresponding to each attractor type;

[0192] A52. Call the fuzzy logic rule base corresponding to the dominant profile type and input the continuous error sequence into the fuzzy logic inference system to obtain the preliminary threshold adjustment amount;

[0193] A53. Set the step size adjustment coefficient according to the attractor strength index corresponding to the attractor type. The first adjustment coefficient is for the periodic type, the second adjustment coefficient is for the decay type, the third adjustment coefficient is for the divergent type, and the fourth adjustment coefficient is for the chaotic type. The first adjustment coefficient is less than or equal to the second adjustment coefficient, the second adjustment coefficient is less than or equal to the third adjustment coefficient, and the third adjustment coefficient is less than or equal to the fourth adjustment coefficient.

[0194] A54. Adjust the initial threshold adjustment amount by step size adjustment coefficient corresponding to the attractor type to obtain the adjusted threshold amount.

[0195] Specifically, A51 includes: For periodic attractors, calculating the amplitude, width, or standard deviation of their error trajectory to quantify error fluctuation and obtain a strength index for the periodic attractor. The strength index for periodic attractors is relatively small, used to characterize the stable and limited fluctuation characteristics of system errors. Secondly, for decaying attractors, calculating their error convergence rate, local Lyapunov exponent, or trajectory volatility to obtain a strength index for decaying attractors. This index is moderately low, used to reflect the gradual convergence and moderate activity of system errors. Thirdly, for divergent attractors, calculating their error growth rate, trajectory radius, or root-mean-square amplitude to obtain a strength index for divergent attractors. This index is relatively large, used to represent the gradually increasing and highly active characteristics of system errors. Finally, for chaotic attractors, calculating their local Lyapunov exponent, trajectory complexity, root-mean-square amplitude, or volatility, and combining these with other weighted indices to obtain a strength index for chaotic attractors. This index is moderately large, used to characterize the complex and dynamically active characteristics of system errors.

[0196] Specifically, A52 includes: selecting a corresponding rule set from a pre-defined fuzzy logic rule base based on the dominant profile type of the current sample. Each dominant profile type corresponds to a set of dedicated rules used to describe the fuzzy relationship between the error state and the threshold adjustment amount. Next, the fuzzy logic inference system infers the fuzzified error sequence according to the selected rules to obtain the fuzzy output corresponding to the error, i.e., the fuzzy set of preliminary threshold adjustment amounts. Finally, the fuzzy output is converted into specific values ​​using a defuzzification method (such as the centroid method) to obtain the preliminary threshold adjustment amount that can be used for subsequent step size adjustment.

[0197] Specifically, A53 includes the following steps: For periodic attractors, the error fluctuation amplitude is small, and the system is in a relatively stable state. Therefore, a first adjustment coefficient is set to make small-scale step adjustments to the initial threshold adjustment amount to maintain system stability. Secondly, for decaying attractors, the error gradually decreases, and the system shows a convergence trend. Therefore, a second adjustment coefficient, slightly larger than the first adjustment coefficient, is set to moderately accelerate the convergence speed of the threshold adjustment and promote rapid error reduction. Then, for divergent attractors, the error gradually increases, and the system has high dynamic activity. Therefore, a third adjustment coefficient, significantly larger than the first and second adjustment coefficients, is set to quickly suppress the increase in error and enhance the system's responsiveness. Finally, for chaotic attractors, the error dynamics are complex and unpredictable. Therefore, a fourth adjustment coefficient, moderately large, is set to effectively cope with the complexity of error fluctuations while ensuring system stability. Through these steps, the dynamic characteristics of each attractor type are matched with the corresponding step adjustment coefficient, achieving fine-tuning of the initial threshold adjustment amount, thus providing a reliable basis for the smooth adaptive update of the first preset importance threshold of the gradient boosting tree model.

[0198] This invention sets a step size adjustment coefficient for each type of attractor based on its corresponding attractor strength index. The different step size adjustment coefficients are designed to reflect the dynamic activity level of each attractor type, allowing for appropriate amplification or reduction of the threshold adjustment amount. The first adjustment coefficient corresponds to periodic attractors, where error fluctuations are small and the system is stable, therefore the step size adjustment range is the smallest or relatively small. The second adjustment coefficient corresponds to decaying attractors, where errors gradually decrease and the system converges quickly, so the step size is slightly larger than that of periodic attractors to accelerate convergence. The third adjustment coefficient corresponds to divergent attractors, where errors increase and the system's dynamic activity is high, so the step size is larger to quickly suppress errors. The fourth adjustment coefficient corresponds to chaotic attractors, where error dynamics are complex and unpredictable, so the step size is the largest or relatively large to cope with complex fluctuations and achieve effective control. In other words, as the dynamic activity of the attractor type increases, the step size adjustment coefficient gradually increases.

[0199] Specifically, A54 includes combining the initial threshold adjustment amount calculated by the fuzzy logic inference system with the step size adjustment coefficient of the corresponding attractor type. Specifically, the initial threshold adjustment amount is multiplied by the corresponding step size adjustment coefficient to obtain the adjusted threshold amount for each attractor type. This adjusted threshold amount considers both the magnitude and rate of change of the error, as well as the dynamic characteristics of the attractor type, achieving fine-grained control over the threshold adjustment.

[0200] This invention combines the aforementioned nonlinear threshold adjustment amount to dynamically correct the first preset importance threshold of the gradient boosting tree model. The nonlinear threshold adjustment amount is output by the fuzzy logic inference system and is adjusted in conjunction with the attractor type of the error phase space trajectory to reflect the current false alarm rate and error change trend. After applying the nonlinear threshold adjustment amount to the first preset importance threshold, an updated importance threshold is obtained, which is used to determine the weight of features in the gradient boosting tree model. By applying this nonlinear adjustment amount to the first preset importance threshold, the model can adaptively adjust the feature importance selection criteria during operation, thereby avoiding the decision lag or over-correction problems caused by a fixed threshold, and improving the stability and false alarm suppression effect of the model under complex operating conditions.

[0201] This invention forms a closed-loop model optimization system through steps such as fuzzy evaluation, false alarm feature vector identification, ROC curve analysis, target false alarm rate setting, feedback control logic calculation, and dynamic threshold adjustment. It can dynamically adjust the feature screening threshold according to the real-time performance of the model, thereby optimizing the model's prediction performance, especially reducing the false alarm rate.

[0202] In another implementation, the step of performing high-dimensional mapping and anomaly analysis on the standard asset feature vector and the real-time asset feature vector to determine whether there are abnormal nodes in the current real-time asset data includes:

[0203] S41. Use the kernel function of support vector machine to map standard asset feature vectors and real-time asset feature vectors to a high-dimensional data space;

[0204] S42. Perform cluster analysis on the real-time asset feature vectors in the high-dimensional data space, and determine the discrete asset process state clusters in the real-time asset feature vectors based on the cluster analysis results.

[0205] S43. Calculate the spatial distance between the standard asset feature vector and the discrete asset process state cluster;

[0206] S44. Calculate the local density and neighbor sample distribution characteristics of real-time asset feature vectors in high-dimensional data space;

[0207] S45. If at least two of the spatial distance, local density, and neighboring sample distribution characteristics do not meet the preset standards, it is determined that there are abnormal nodes in the real-time asset data.

[0208] Specifically, S41 includes: mapping the standard asset feature vector and the real-time asset feature vector to a high-dimensional data space using the kernel function of a support vector machine (SVM). The SVM kernel function maps data from the original space to a high-dimensional (or even infinite-dimensional) feature space, making it easier to find the optimal classification hyperplane in the high-dimensional space. Mapping to a high-dimensional data space means mapping the standard asset feature vector and the real-time asset feature vector to this high-dimensional space through the kernel function, making features that were difficult to distinguish in the original space distinguishable in the high-dimensional space. In this embodiment, the standard asset feature vector is a baseline feature constructed based on historical data; the real-time asset feature vector is the asset state feature at the current moment. The kernel function can be a linear kernel function; by mapping to the high-dimensional space, the SVM kernel function can identify the non-linear deviation between the real-time features and the standard features.

[0209] Specifically, S42 includes: performing cluster analysis on real-time asset feature vectors in a high-dimensional data space; determining discrete asset process state clusters within the real-time asset feature vectors based on the cluster analysis results; and using clustering algorithms (such as K-Means, DBSCAN, hierarchical clustering, etc.) to divide the real-time asset feature vectors mapped to the high-dimensional space into several clusters, each cluster representing a discrete asset process state (such as "normal," "abnormal," "transitional state," etc.), thereby achieving dynamic monitoring and classification of asset states. Cluster analysis is an unsupervised learning method used to divide objects in a dataset into several groups or "clusters," ensuring high similarity among objects within the same cluster and low similarity among objects in different clusters. Discrete asset process state clusters refer to the clusters obtained after clustering real-time asset feature vectors in a high-dimensional data space; each cluster may represent a specific state or stage in the asset process.

[0210] Specifically, S43 includes: calculating the distance between the standard asset feature vector and the center of each discrete asset process state cluster in a high-dimensional data space. This distance reflects the degree of difference between the standard asset state and different states of real-time assets. The spatial distance can be calculated by calculating the cosine distance between the standard asset feature vector and the center of each discrete asset process state cluster (suitable for high-dimensional sparse data). The cosine distance measures the degree of difference between the two; a smaller distance indicates that the standard vector is closer to the typical state of the cluster; a larger distance indicates that the standard vector deviates further from the cluster, possibly belonging to an abnormal state or another cluster.

[0211] Next, we calculate the local density and neighbor sample distribution characteristics of the real-time asset feature vector in the high-dimensional data space. Local density describes the density of samples surrounding the real-time asset feature vector in the high-dimensional data space, reflecting the "crowding" of the area where the point is located. Neighbor sample distribution characteristics analyze the distribution pattern of samples around the real-time asset feature vector, including distance, direction, and category. For example, whether the samples around a point are uniformly distributed or concentrated in certain directions. Points with low local density may be outliers, and uneven distribution of neighbor samples may indicate abnormal patterns (such as data drift). This can be achieved based on local density estimation using K-nearest neighbors. The neighbor sample distribution characteristics can be obtained by plotting the distance distribution of neighbor samples and observing whether they follow a certain distribution (such as a normal distribution).

[0212] Based on pre-defined normal ranges using business experience or historical data, such as the maximum allowable spatial distance, the minimum threshold for local density, and the compactness index of neighboring sample distribution (e.g., the proportion of samples within a cluster must be >80%), if at least two of the spatial distance, local density, and neighboring sample distribution characteristics are abnormal, then the real-time asset data is determined to have an abnormal node in the asset process. Based on these three indicators—spatial distance, local density, and neighboring sample distribution characteristics—if at least two of them indicate a significant difference between the real-time asset feature vector and the normal state, then the asset process corresponding to that feature vector is considered abnormal. Based on the above judgment, the specific node or stage in the real-time asset data where the asset process is abnormal is determined.

[0213] This invention combines multiple technologies, such as kernel function mapping, cluster analysis, spatial distance calculation, local density and neighbor sample distribution characteristic analysis of SVM, to form a comprehensive asset process anomaly detection framework. It can comprehensively and accurately assess the differences between real-time asset data and standard asset data, thereby timely discovering and locating abnormal nodes in the asset process and ensuring the stable operation of the asset process.

[0214] In another implementation, the step of correcting the abnormal data at the asset data flow nodes if abnormal nodes exist in the real-time asset data includes:

[0215] S51. When there is an abnormal spatial distance, use the K-nearest neighbor algorithm to traverse each discrete asset process state cluster and determine the minimum distance between each discrete asset process state cluster and the nearest normal state cluster.

[0216] S52. Calculate the weighted average of the discrete asset process state clusters based on the minimum distance, and correct the coordinate values ​​of the discrete asset process state clusters in the high-dimensional data space based on the weighted average.

[0217] S53. When local density anomalies exist, use the local anomaly factor algorithm to mark low-density regions;

[0218] S54. Use oversampling technology to generate composite data of composite points for discrete asset process state clusters in low-density areas;

[0219] S55. Insert the synthesis points into the low-density region according to the time sequence of the synthesis data;

[0220] S56. When there are abnormal distribution characteristics of neighboring samples, perform principal component analysis on the feature values ​​of the normal state clusters adjacent to the discrete asset process state clusters to determine the principal components of the normal state clusters.

[0221] S57. Map the feature values ​​of the discrete asset process state cluster to the principal component space and compare whether the principal component distribution of the discrete asset process state cluster deviates from the preset distance threshold.

[0222] S58. If the principal component distribution deviates from the distance threshold, the principal components of the normal state cluster are used to correct the discrete asset process state cluster.

[0223] When spatial distance anomalies exist, the K-nearest neighbor algorithm is used to traverse each discrete asset process state cluster to determine the minimum distance between each discrete asset process state cluster and its nearest normal state cluster. Spatial distance anomalies refer to the calculated spatial distance between the standard asset feature vector and the discrete asset process state cluster exceeding the normal range, indicating that the asset state deviates from the known pattern. Discrete asset process state clusters refer to anomalous state clusters; the K-nearest neighbor algorithm is a simple machine learning algorithm used to find the K nearest points to a given point. For each discrete asset process state cluster, the K-nearest neighbor algorithm is used to find its nearest normal state cluster, and the distance between them is calculated; that is, for each anomalous cluster, the minimum distance among all normal clusters is found.

[0224] Subsequently, a weighted average of the discrete asset process state clusters is calculated based on the minimum distance. This weighted average is then used to correct the coordinates of the discrete asset process state clusters in the high-dimensional data space. In other words, the coordinates of the abnormal and normal clusters are weighted and averaged based on the minimum distance to generate corrected coordinates. The positions of the abnormal clusters in the high-dimensional data space are adjusted to be closer to the normal pattern, reducing deviations. Weighting coefficients can be defined based on the minimum distance to control the weight ratio between abnormal and normal clusters. A linear weighting method is used to define the weighting coefficients. Then, the weighted average is calculated, which is essentially a weighted average of the coordinates of the abnormal clusters, generating corrected coordinates. The corrected coordinates replace the original coordinates of the abnormal clusters. Based on the distance between the discrete asset process state cluster and the nearest normal state cluster, a weight is assigned to each cluster, and then the weighted average of these clusters is calculated. The weighted average is used to adjust the coordinates of the discrete asset process state clusters to be closer to the normal state clusters.

[0225] Next, when local density anomalies exist, the Local Anomaly Factor (LOF) algorithm is used to mark low-density regions. Local density refers to the density measured by the distance between a sample point and its neighboring points; points in low-density regions are more likely to be anomalous. The LEF is the ratio of the local anomaly factor value of a sample point to the mean local reachability density of its neighboring points and the local reachability density of that point, measuring the degree of anomalousness of that point relative to its neighborhood. The principle of the LEF algorithm is to quantify the degree of anomaly by comparing the local density of a data point with the local density of its neighbors. A LEF value > 1 indicates a low-density region (anomaly); a LEF value ≈ 1 indicates that the density is consistent with the surrounding area (normal); and a LEF value less than 1 indicates a high-density region (possibly a core point). First, the k-distance is calculated, which is the distance from the data point to its k-th nearest neighbor. Then, the reachability distance is calculated, which is the maximum distance from the neighboring points to the current point. Next, the local reachability density is calculated, which reflects the local density. Finally, the LEF value is calculated as the ratio of the local reachability density of the current point to the local reachability density of its neighbors. Local density anomalies refer to the calculated real-time asset feature vector having a local density in the high-dimensional data space that is lower than the normal range. The Local Anomaly Factor (LAF) algorithm is used to detect local outliers in a dataset. It identifies outliers by comparing the local density of a point with the local density of its neighbors. The LAF algorithm is used to identify and mark regions with abnormally low local density.

[0226] Oversampling is a technique used to increase the number of minority class samples in a dataset, typically achieved by generating synthetic samples. Oversampling is applied to discrete asset process state clusters in low-density regions to generate new synthetic points and their corresponding synthetic data. First, low-density clusters are identified, which can be done using clustering (such as K-Means) or density detection to label them. Then, an oversampling algorithm is selected, such as SMOTE (Synthetic Minority Oversampling) or ADASYN (Adaptive Synthetic Sampling). Finally, synthetic data is generated by interpolating existing samples within the low-density clusters to create new samples. By generating synthetic data points through oversampling, the number of samples in low-density clusters is increased, improving data balance and the model's ability to identify rare states.

[0227] Next, inserting composite points into low-density regions in chronological order based on the composite data means inserting composite points into low-density regions according to the chronological order of the composite data to maintain the temporal continuity of the data. First, extract the timestamps of the original samples in the low-density clusters, calculate the time interval distribution (e.g., average time interval, variance of time interval), and generate the time intervals of the composite points based on the time interval distribution (e.g., normal distribution, Poisson distribution) and recursively calculate the timestamps. Then, find the time interval of the low-density region in the original time series (e.g., detect sparse sample periods using a sliding window), and insert composite points within this interval in the order of the generated timestamps.

[0228] When there are anomalies in the distribution characteristics of neighboring samples, principal component analysis (PCA) is performed on the eigenvalues ​​of adjacent normal state clusters of discrete asset process state clusters to determine the principal components of the normal state clusters. Anomalies in the distribution characteristics of neighboring samples refer to significant differences between the distribution characteristics of the analyzed real-time asset feature vectors and those of the neighboring samples and those of the normal state. PCA is a technique used for dimensionality reduction and feature extraction, simplifying data representation by identifying the main directions of change (principal components) in the data. PCA is performed on the eigenvalues ​​of adjacent normal state clusters to determine their principal components. Specifically, this involves extracting the main directions of change (principal components) from the high-dimensional features of the normal state clusters, compressing the data dimensionality while retaining key information. First, samples of normal state clusters adjacent to the abnormal region (such as the "normal operation" cluster data) are selected to form a feature matrix, where each row represents a sample and each column represents a feature. Next, standardization is performed to normalize or normalize the feature matrix to avoid the influence of dimensions, and the covariance matrix is ​​calculated. Subsequently, the eigenvalues ​​and eigenvectors of the covariance matrix are solved, sorted by eigenvalue size, and the top k eigenvectors with a cumulative variance contribution of 80%–95% are selected as principal components.

[0229] Mapping the feature values ​​of discrete asset process state clusters to the principal component space and comparing whether the principal component distribution of the discrete asset process state clusters deviates from a preset distance threshold involves mapping the feature values ​​of the discrete asset process state clusters to the principal component space determined by the normal state clusters, analyzing the distribution of the discrete asset process state clusters in the principal component space, calculating the distance between the discrete asset process state clusters and the principal component distribution of the normal state clusters, and determining whether this distance exceeds the preset threshold. If more than a certain proportion of samples within a cluster are more than a preset internal threshold away from the cluster center, the internal distribution of that cluster is determined to be abnormal (e.g., excessively high sample dispersion). If the inter-cluster distance between a cluster and the "normal operation" cluster exceeds a preset separation threshold, the cluster is determined to deviate from the normal state (e.g., the principal component distribution of the faulty cluster differs too much from that of the normal cluster).

[0230] If the principal component distribution deviates from the distance threshold, then the principal components of the normal state cluster are used to correct the discrete asset process state cluster. This means the calculated distance exceeds the preset threshold. If the principal component distribution of the discrete asset process state cluster deviates from the principal component distribution of the normal state cluster, then the principal components of the normal state cluster are used to adjust the eigenvalues ​​of the discrete asset process state cluster to make them closer to the normal state. The eigenvalues ​​of the discrete asset process state cluster are mapped to a space composed of the principal components of the normal state cluster (i.e., a principal component coordinate system based on the normal state). The distance between the distribution of this cluster in the principal component space and the normal state cluster (e.g., the difference in coordinates after projection) is calculated to determine if it exceeds the threshold. Using the principal component vectors of the normal state cluster, the eigenvalues ​​of the discrete cluster are linearly combined to reconstruct its eigendistribution, correcting it towards the direction of the principal components of the normal state. Within the range of the principal component distribution of the normal state cluster, new eigenvalues ​​can be generated through interpolation (e.g., linear interpolation, Gaussian mixture model sampling) to replace or supplement abnormal samples in the discrete cluster.

[0231] This invention combines multiple anomaly detection indicators such as spatial distance, local density, and neighbor sample distribution characteristics, and utilizes various techniques such as the K-nearest neighbor algorithm, local anomaly factor algorithm, oversampling technology, and principal component analysis to mark and correct abnormal data in asset data flow nodes. This enables comprehensive and accurate identification and correction of anomalies in asset data, improving the accuracy and reliability of the data.

[0232] In another implementation, if no abnormal nodes exist in the real-time asset data, the real-time asset data is processed using blockchain data processing steps, and the processed real-time asset data is uploaded to the pre-built traceability chain, including:

[0233] S61. Obtain all asset process node information on the pre-built traceability chain. The asset process node information includes publicly available asset information and private asset information.

[0234] S62. Encrypt the asset privacy information using a preset encryption algorithm to obtain encrypted asset data;

[0235] S63. Based on all asset process nodes of the traceability chain and according to the blockchain consensus mechanism preset by the traceability chain, consensus verification is performed on public asset information and private asset information.

[0236] S64. If consensus verification is successful, the public asset information and the private asset information will be broadcast to all asset process nodes.

[0237] S65. Store publicly available asset information and private asset information in the blockchain ledger of the traceability chain in the form of blocks.

[0238] The traceability chain is a distributed ledger based on blockchain technology, used to record and track the flow of assets throughout their entire lifecycle. Through its immutable and traceable characteristics, it ensures the transparency and credibility of asset information. Asset process nodes refer to the key links or steps an asset undergoes during its flow, such as production, processing, transportation, and sales. When building the traceability chain, these asset process nodes need to be treated as transactions or blocks on the blockchain, linked according to chronological order and logical relationships. The blockchain's consensus mechanism can be used to ensure the authenticity and consistency of node information, and smart contracts and other technologies can be used to automate the recording and verification of asset process nodes. Public asset information refers to asset information that can be publicly disclosed, such as the asset's name, specifications, production date, and production location. Private asset information refers to sensitive information involving trade secrets and personal privacy, such as the asset's cost price, supplier information, and customer data. Public information of asset process nodes can be queried and obtained through the traceability chain's public interface or a browser. Access to private information requires authorization from the relevant parties. This is typically achieved through encryption technology, access control lists, and other means. Subsequently, the acquired asset process node information is integrated to form a complete asset traceability report.

[0239] The preset encryption algorithm refers to one or more encryption methods that are selected and configured in advance, such as symmetric encryption algorithms (AES) or asymmetric encryption algorithms (RSA). A suitable encryption algorithm can be selected based on factors such as the sensitivity of the asset privacy information, encryption requirements (such as speed and security), and system compatibility. Next, the asset privacy information that needs to be protected is used as input data. The selected encryption algorithm and key are then used to encrypt the input data, generating encrypted asset data, i.e., ciphertext.

[0240] A blockchain consensus mechanism is an algorithm or protocol used by nodes in a blockchain network to reach a consensus, ensuring the consistency and security of the blockchain. It determines which blocks can be added to the blockchain and how conflicts and disagreements between nodes are resolved. Asset process nodes refer to key stages or steps an asset undergoes during its circulation, such as production, processing, transportation, and sales. On the traceability chain, each asset process node corresponds to a transaction or block on the blockchain, recording detailed information about the asset at that node. Public asset information refers to publicly disclosed asset information, such as the asset's name, specifications, production date, and production location. This information is public during the consensus verification process, and all nodes can view and verify it. Private asset information refers to sensitive information involving trade secrets and personal privacy, such as the asset's cost price, supplier information, and customer data. This information is encrypted during the consensus verification process, and only authorized nodes can view and verify it. After collecting public and private asset information from all asset process nodes on the traceability chain, the nodes prepare to verify the collected information according to preset consensus mechanism rules, including data integrity, authenticity, and consistency. Once all nodes have received the verification results, they reach a consensus through a consensus mechanism algorithm. If a majority of nodes consider the information valid, then the information is officially recorded on the blockchain.

[0241] Consensus verification signifies that both public and private asset information have been verified by all participating nodes and confirmed as valid and consistent data. Broadcasting refers to sending information to all nodes in the network or within a specific group. Next, through the network protocols and communication mechanisms of the blockchain or traceability chain, the verified asset information is broadcast to all nodes. Upon receiving the broadcast information, each node verifies and stores it to ensure the accuracy and consistency of the information. Furthermore, when broadcasting private asset information, it is essential to ensure that the information has been encrypted and that only authorized nodes can decrypt and access it.

[0242] A blockchain consists of a series of blocks arranged chronologically, each block containing transaction data or information records within a specific timeframe. Storing asset information in block form means packaging information into a block and adding it to the blockchain ledger. The Traceability Chain is a distributed ledger based on blockchain technology used to record and track the flow of assets throughout their entire lifecycle. The specific storage process first requires collecting and organizing publicly available and private asset information, encrypting the private information to ensure security. Then, the publicly available and encrypted private information are packaged into a block, which includes a block header (recording metadata such as timestamps and hash values) and a block body (recording the specific asset information). Using all asset process nodes in the Traceability Chain and a pre-defined blockchain consensus mechanism, the information in the block is verified through consensus, including verification of the information's completeness, authenticity, and consistency. If the block passes consensus verification, it is added to the Traceability Chain blockchain ledger.

[0243] This invention enhances the transparency and traceability of asset processes by applying blockchain data processing steps to real-time asset data and uploading it to a pre-built traceability chain. This helps to identify potential risks and problems, improves the controllability and stability of asset processes, ensures the stable operation of asset processes, and improves the accuracy of asset management.

[0244] Example 2:

[0245] See Figure 2An intelligent asset process monitoring system based on artificial intelligence technology includes a data acquisition module, a first feature extraction module, a second feature extraction module, an anomaly analysis module, a data correction module, and a data storage module. The data acquisition module is used to acquire historical asset data and real-time asset data; specifically, the data acquisition module performs S1 in Embodiment 1, which will not be elaborated here. The first feature extraction module is used to extract standard asset feature vectors under normal conditions based on historical asset data and using a pre-trained support vector machine model; specifically, the first feature extraction module performs S2 in Embodiment 1, which will not be elaborated here. The second feature extraction module is used to extract real-time asset feature vectors from real-time asset data based on a preset gradient boosting tree model; specifically, the second feature extraction module performs S3 in Embodiment 1, which... The details are omitted here; the anomaly analysis module is used to perform high-dimensional mapping and anomaly analysis on the standard asset feature vector and the real-time asset feature vector to determine whether there are abnormal nodes in the current real-time asset data; specifically, the anomaly analysis module is used to execute S4 in embodiment 1, which is omitted here; the data correction module is used to correct the abnormal data of the asset data process node when there are abnormal nodes in the real-time asset data; specifically, the data correction module is used to execute S5 in embodiment 1, which is omitted here; the data storage module is used to process the real-time asset data using blockchain data processing steps when there are no abnormal nodes in the real-time asset data, and upload the processed real-time asset data to the pre-built traceability chain; specifically, the data storage module is used to execute S6 in embodiment 1, which is omitted here.

Claims

1. An intelligent monitoring method for asset processes based on artificial intelligence technology, characterized in that: The intelligent monitoring method for asset processes includes: Obtain historical and real-time asset data; Extract standard asset feature vectors under normal conditions based on historical asset data and using a pre-trained support vector machine model; Real-time asset feature vectors are extracted from real-time asset data based on a pre-defined gradient boosting tree model. Perform high-dimensional mapping and anomaly analysis on standard asset feature vectors and real-time asset feature vectors to determine whether there are abnormal nodes in the current real-time asset data; If there are abnormal nodes in the asset process in the real-time asset data, then correct the abnormal data in the asset data process nodes. If no abnormal nodes are found in the real-time asset data, the real-time asset data will be processed using blockchain data processing steps, and the processed real-time asset data will be uploaded to the pre-built traceability chain.

2. The intelligent asset process monitoring method based on artificial intelligence technology according to claim 1, characterized in that: The method for extracting real-time asset feature vectors from real-time asset data based on a preset gradient boosting tree model includes: Iterate through each tree model in the preset gradient boosting tree model and input real-time asset data into the tree model; In each tree model, real-time asset data is allocated to the leaf nodes of the tree model according to the preset decision rules, and the code of each leaf node in the tree model is recorded. Extract the feature subvector corresponding to the encoding of each leaf node; Calculate the feature importance of the feature subvector corresponding to the encoding of each leaf node based on historical asset data; Feature vectors with a feature importance greater than a first preset importance threshold are concatenated into a high-dimensional sparse vector to represent the real-time asset feature vector, where the dimension of the real-time asset feature vector is equal to the total number of leaf nodes in the preset tree model.

3. The intelligent asset process monitoring method based on artificial intelligence technology according to claim 2, characterized in that: The method for extracting real-time asset feature vectors from real-time asset data based on a preset gradient boosting tree model also includes: The false positive feature vector is calculated based on the vector fuzzy importance of the high-dimensional sparse vector, and the first preset importance threshold of the gradient boosting tree model is dynamically adjusted based on the false positive feature vector and the dominant profile type of the false positive feature vector.

4. The intelligent asset process monitoring method based on artificial intelligence technology according to claim 3, characterized in that: The step of calculating the false positive feature vector based on the vector fuzzy importance of high-dimensional sparse vectors, and dynamically adjusting the first preset importance threshold of the gradient boosting tree model based on the false positive feature vector and the dominant profile type of the false positive feature vector, includes: The importance of the high-dimensional sparse vectors extracted by the preset gradient boosting tree model is evaluated by fuzzy evaluation to obtain the fuzzy importance of the high-dimensional sparse vectors. High-dimensional sparse vectors with a fuzzy importance less than the second preset fuzzy importance are used as false alarm feature vectors. Based on graph theory community detection, the dominant profile type of each false alarm feature vector is identified by analyzing the false alarm feature vector. The current false alarm rate of the gradient boosting tree model is calculated based on the false alarm feature vector and the dominant profile type. Target false alarm rates were set based on receiver operating characteristic curves and corresponding to the dominant profile type. The error change rate is calculated based on the error between the current false alarm rate and the target false alarm rate. The nonlinear threshold adjustment amount is then calculated by combining the error change rate and the false alarm rate using a fuzzy logic rule base associated with the dominant profile type. The first preset importance threshold of the gradient boosting tree model is dynamically adjusted by combining a non-linear threshold adjustment amount.

5. The intelligent asset process monitoring method based on artificial intelligence technology according to claim 1, characterized in that: The extraction of standard asset feature vectors under normal conditions based on historical asset data and using a pre-trained support vector machine model includes: Linear discriminant analysis algorithm is used to enhance the class distinction of real-time asset data, and clustering algorithm is used to determine the normal status data of historical asset data. Fuzzy asset feature vectors are extracted from normal asset status data using an empirical pattern decomposition algorithm. A pre-defined support vector machine model is trained using fuzzy asset feature vectors to obtain the trained support vector machine model; Historical asset data is input into the trained support vector machine model to extract standard asset feature vectors under normal conditions.

6. The intelligent asset process monitoring method based on artificial intelligence technology according to claim 5, characterized in that: The extraction of fuzzy asset feature vectors from normal asset status data using the empirical pattern decomposition algorithm includes: The empirical mode decomposition algorithm is used to perform empirical mode decomposition on the asset normal state data to obtain multiple intrinsic mode function components of the asset normal state data, and the component energy of each intrinsic mode function component is calculated. The component energy is used as the input to the membership function and mapped to a fuzzy set; In a fuzzy set, the fuzzy asset feature vector corresponding to each component energy is determined according to pre-constructed fuzzy rules.

7. The intelligent asset process monitoring method based on artificial intelligence technology according to claim 1, characterized in that: The process of performing high-dimensional mapping and anomaly analysis on standard asset feature vectors and real-time asset feature vectors to determine whether there are abnormal nodes in the current real-time asset data includes: The kernel function of the support vector machine is used to map the standard asset feature vector and the real-time asset feature vector to a high-dimensional data space; Cluster analysis is performed on the real-time asset feature vectors in the high-dimensional data space, and the discrete asset process state clusters in the real-time asset feature vectors are determined based on the cluster analysis results. Calculate the spatial distance between the standard asset feature vector and the discrete asset process state cluster; Calculate the local density and neighbor sample distribution characteristics of real-time asset feature vectors in a high-dimensional data space; If at least two of the spatial distance, local density, and neighboring sample distribution characteristics do not meet the preset standards, it is determined that there are abnormal nodes in the real-time asset data process.

8. The intelligent asset process monitoring method based on artificial intelligence technology according to claim 1, characterized in that: If there are abnormal nodes in the real-time asset data flow, the abnormal data in the asset data flow nodes will be corrected, including: When spatial distance anomalies exist, the K-nearest neighbor algorithm is used to traverse each discrete asset process state cluster and determine the minimum distance between each discrete asset process state cluster and the nearest normal state cluster. The weighted average of discrete asset process state clusters is calculated based on the minimum distance, and the coordinates of discrete asset process state clusters in high-dimensional data space are corrected based on the weighted average. When local density anomalies exist, the local anomaly factor algorithm is used to mark low-density regions. Oversampling techniques are used to generate composite data for composite points from discrete asset process state clusters in low-density areas. Based on the synthetic data, the synthetic points are inserted into the low-density region in chronological order. When there are abnormal distribution characteristics of neighboring samples, principal component analysis is performed on the feature values ​​of the normal state clusters adjacent to the discrete asset process state clusters to determine the principal components of the normal state clusters. Map the feature values ​​of discrete asset process state clusters to the principal component space and compare whether the principal component distribution of discrete asset process state clusters deviates from a preset distance threshold. If the principal component distribution deviates from the distance threshold, the principal components of the normal state cluster are used to correct the discrete asset process state cluster.

9. The intelligent asset process monitoring method based on artificial intelligence technology according to claim 1, characterized in that: If no abnormal nodes are found in the real-time asset data, then the real-time asset data is processed using blockchain data processing steps, and the processed real-time asset data is uploaded to the pre-built traceability chain, including: Obtain information on all asset process nodes in the pre-built traceability chain, including publicly available asset information and private asset information; The asset privacy information is encrypted using a preset encryption algorithm to obtain encrypted asset data; Based on all asset process nodes of the traceability chain and according to the blockchain consensus mechanism preset by the traceability chain, consensus verification is carried out on public asset information and private asset information. If the consensus verification is successful, the public asset information and the private asset information will be broadcast to all asset process nodes. Public and private asset information is stored in the blockchain ledger of the traceability chain in the form of blocks.

10. An intelligent monitoring system for asset processes based on artificial intelligence technology, characterized in that, include: The asset process intelligent monitoring system is based on the asset process intelligent monitoring method of claim 1, and includes: The data acquisition module is used to acquire historical and real-time asset data. The first feature extraction module is used to extract standard asset feature vectors under normal conditions based on historical asset data and using a pre-trained support vector machine model. The second feature extraction module is used to extract real-time asset feature vectors from real-time asset data based on a preset gradient boosting tree model. The anomaly analysis module is used to perform high-dimensional mapping and anomaly analysis on standard asset feature vectors and real-time asset feature vectors to determine whether there are abnormal nodes in the current real-time asset data. The data correction module is used to correct abnormal data in the asset data process nodes when there are abnormal nodes in the real-time asset data. The data storage module is used to process real-time asset data using blockchain data processing steps when there are no abnormal nodes in the asset process, and then upload the processed real-time asset data to the pre-built traceability chain.

Images