Asset risk tracing maintenance scheduling method and system fusing feature matrix

Abstract

The application relates to the technical field of asset operation and maintenance, and discloses an asset risk tracing maintenance scheduling method and system fusing a feature matrix, which comprises the following steps: acquiring an asset attribute vector set, constructing an asset state feature matrix, performing sliding window aggregation, trend denoising and abnormal point identification, extracting an asset state vector to generate an asset state data stream, associating a whole life cycle tracing code, and generating an asset state time sequence data set; constructing a risk behavior chain template library and an asset state prediction model, matching the risk behavior chain template library to obtain a risk acceleration factor, combining an asset business importance weight and an environmental influence factor to establish a risk assessment function, outputting a risk assessment index sequence, determining an initial maintenance time and triggering a maintenance pre-warning; comprehensively calculating an asset maintenance score, adopting a particle swarm algorithm to perform priority sorting and maintenance resource dynamic scheduling on a maintenance task, forming an asset risk tracing maintenance scheduling report, and improving asset risk prediction precision and asset maintenance scheduling efficiency.

Description

Technical Field

[0001] This invention relates to the field of asset operation and maintenance technology, specifically to an asset risk tracing and maintenance scheduling method and system that integrates feature matrices. Background Technology

[0002] As enterprises advance their digital transformation, the number and types of assets continue to increase, and the scale of multi-source status data generated during asset operation continues to grow. Traditional asset risk tracing and maintenance typically rely on manual ledgers and scattered data statistics, making it difficult to achieve closed-loop tracing and standardized processing of risk data across the entire chain. Existing models suffer from delayed asset ledger updates and the inability to fully trace asset usage and allocation information, easily leading to discrepancies between records and actual assets, and asset loss. Maintenance management relies on manual records and passive reminders, often resulting in missed or over-maintenance, lacking intelligent early warning and proactive intervention mechanisms. Simultaneously, asset data, maintenance records, and usage status are fragmented, forming information silos. During risk tracing, multi-source data dimensions are scattered, feature fusion is low, and a unified feature matrix data modeling mechanism is lacking, resulting in low efficiency in risk tracing data processing and insufficient accuracy in risk correlation tracing. This fails to support full lifecycle tracing and efficient maintenance decision-making, leading to low management efficiency, high compliance risks, and increased operational costs and system risks.

[0003] A patent application with publication number CN121599374A discloses a method and system for full lifecycle data asset management based on BIM, including: encoding components in the BIM model to generate component attribute vectors; generating a method for aligning and completing component states and data logic with structure through a state acquisition template driven by component attributes; constructing a state modeling structure embedding component attribute information; obtaining a predicted risk score sequence based on a state prediction model inputting state data streams; setting recommended maintenance times for time points in the risk score sequence where the risk score is greater than a risk threshold; and designing a component asset package, including a set of visual tags, which, based on an asset scoring mechanism, outputs a structured encapsulation of the component asset package and pushes it through a unified interface. This invention constructs a full lifecycle data asset expression system, transforming building BIM components from information fragments into sustainably manageable data assets.

[0004] Although existing BIM-based full lifecycle data asset management methods and systems have achieved full lifecycle data connectivity for building components, intelligent prediction of operational risks, and closed-loop digital management of assets by uniquely encoding BIM model components, generating component attribute vectors, constructing standardized time-series state data streams, intelligently outputting maintenance suggestions based on state prediction models, encapsulating component asset packages, and completing data push through a unified interface, the above-mentioned BIM asset management solutions cannot be adapted to asset scenarios. Traditional asset risk data processing suffers from problems such as inconsistent feature modeling, incomplete risk traceability links, and low data processing efficiency. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention aims to provide an asset risk traceability and maintenance scheduling method and system that integrates feature matrices. It generates a full lifecycle traceability code through a five-level hierarchical coding rule, enabling end-to-end asset traceability. Using an asset status feature matrix as the core carrier, it completes standardized feature modeling of multi-source status data, generating an asset status data stream. Combined with a risk behavior chain template library and an asset status prediction model, it predicts a risk assessment index and outputs a risk assessment index sequence, completing the full-link traceability calculation of asset risk. Differentiated risk judgment thresholds determine the initial maintenance time and trigger maintenance pre-reminders. Based on asset health, estimated remaining useful life, and maintenance urgency, it comprehensively calculates an asset maintenance score, enabling priority ranking of maintenance tasks and dynamic resource scheduling. This solves the problems of inconsistent feature modeling of asset risk traceability data, incomplete risk traceability links, and low data processing efficiency, improving the data processing accuracy and control efficiency of enterprise asset risk traceability.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] The asset risk tracing and maintenance scheduling method based on the fusion of feature matrices includes:

[0008] After obtaining and preprocessing multi-source status data, a full lifecycle traceability code is generated, and after structured extraction and normalization, an asset attribute vector set is generated.

[0009] Construct an asset status feature matrix, perform sliding window aggregation, trend denoising and outlier identification on feature sequences of each dimension, extract asset status vectors to generate asset status data streams, and associate full lifecycle traceability codes to generate asset status time series datasets.

[0010] Historical risk data is used to construct a risk behavior chain template library, an asset status prediction model is built, risk acceleration factors are obtained by matching the risk behavior chain template library, and a risk assessment function is established by combining the asset business importance weight and environmental impact factors to output a risk assessment index sequence.

[0011] The system iterates through the risk assessment index sequence to determine the initial maintenance time and triggers maintenance pre-reminders. It then comprehensively calculates the asset maintenance score, uses particle swarm optimization to prioritize maintenance tasks and dynamically schedule maintenance resources, and generates an asset risk traceability maintenance scheduling report.

[0012] Specifically, the steps for constructing the asset status feature matrix include:

[0013] We screen feature domains that are strongly correlated with asset health status from the asset attribute vector set, and standardize continuous features and encode dummy variables for categorical features.

[0014] Using asset instances as rows and feature dimensions as columns, an initial feature matrix is ​​constructed to calculate feature correlation coefficients and extract asset business influencing factors.

[0015] Set a balance coefficient, calculate the feature association contribution value by combining the feature correlation coefficient, calculate the business weight contribution value based on the balance coefficient and the asset business impact factor, and sum them to obtain the feature importance weight;

[0016] Principal component analysis dimensionality reduction algorithm is used to scale the initial feature matrix with feature importance weights as variance scaling factors to obtain a weighted covariance matrix;

[0017] The cumulative variance explained rate is calculated by combining the eigenvalues ​​of each principal component in the statistically weighted covariance matrix with the sum of the eigenvalues. The first n principal components whose cumulative variance explained rate is greater than the preset explained rate threshold are retained. The initial feature matrix is ​​then projected onto a low-dimensional subspace to construct the asset status feature matrix.

[0018] Specifically, the steps for generating an asset status data stream include:

[0019] For each asset, state feature vectors are extracted from the asset state feature matrix based on the full lifecycle traceability code to form an initial state feature sequence and an adaptive sliding window is set.

[0020] Within each sliding window, a weighted mean vector and a slope vector are calculated based on the initial state feature sequence, and then concatenated to form an aggregated feature vector to create an aggregated feature sequence.

[0021] The trend feature components are obtained by smoothing the aggregated feature sequence, and the feature residual values ​​are calculated by combining the aggregated feature vector to form the feature residual sequence. An anomaly discrimination criterion based on the sliding interquartile range is established to mark candidate anomalies.

[0022] Acquire historical maintenance event records, perform time-series alignment and comparison with candidate anomalies, retain valid anomalies, and obtain trend feature sequences and a list of valid anomalies through filtering and labeling.

[0023] For the trend feature components, the trend feature sequences within adjacent windows at each time step are used as contextual information to quantify asset health indicators to form an asset state vector and generate an asset state data stream.

[0024] Specifically, the steps for generating a time-series dataset of asset states include:

[0025] Using the full lifecycle traceability code as the primary key, it is associated and mapped with the asset status data stream and the list of valid anomalies, and aggregated according to the asset dimension to generate an asset status time series dataset.

[0026] Specifically, the steps for outputting the risk assessment index series include:

[0027] Acquire historical risk data, perform pattern clustering on the asset status precursor sequence fragments with historical risk events as the endpoint, extract the cluster center sequences of each type and associate them with historical risk type and severity level labels, and construct a risk behavior chain template library;

[0028] An asset state prediction model is constructed based on temporal convolutional networks and long short-term memory networks;

[0029] The asset state time series dataset was used as the training set, the weighted cross-entropy loss function was used as the loss function, and the Adam optimizer was used to train the asset state prediction model.

[0030] After training, the latest real-time state segments and risk behavior chain template library from the asset state time series dataset are input. Through time series feature extraction, state prediction and behavior chain-guided output layer, the risk assessment index sequence is output.

[0031] Specifically, the steps for constructing an asset status prediction model include:

[0032] Construct an asset status prediction model, including an input layer, a time-series feature extraction layer, a status prediction layer, and an output layer;

[0033] The input layer receives the asset state time series dataset and the risk behavior chain template library, and reshapes the asset state time series dataset according to asset instances and time windows to obtain a standardized state vector tensor.

[0034] The temporal feature extraction layer consists of stacked dilated causal convolution modules. The dilated causal convolution modules are used to extract multi-scale local temporal features and fault precursor features from the standardized state vector tensor layer by layer. The receptive field is expanded by increasing the dilation coefficient layer by layer, and residual connection and batch normalization are performed to output a sequence of feature maps.

[0035] Specifically, the steps for constructing an asset status prediction model also include:

[0036] The state prediction layer consists of a two-layer long short-term memory network and an attention decoder. The first LSTM layer receives the feature map sequence, updates the hidden state step by step according to the time step, and outputs the first hidden state sequence. The second LSTM layer receives the first hidden state sequence, further encodes long-term dependencies and suppresses temporal noise, and outputs the second hidden state sequence.

[0037] The attention decoder calculates the context vector based on the weighted summation of the second hidden state sequence, concatenates the context vector with the predicted output state vector of the previous time step, and then iteratively generates a sequence of predicted state vectors.

[0038] The output layer is used to normalize and map the predicted state vector sequence to obtain the baseline risk value, calculate the similarity score between the predicted state vector sequence and the risk behavior chain template library as the risk acceleration factor, obtain the asset business importance weight and determine the environmental impact factor, establish a risk assessment function based on the behavior chain guidance mechanism to predict the risk assessment index, and output the risk assessment index sequence.

[0039] Specifically, the steps for comprehensively calculating the asset maintenance score include:

[0040] Query the threshold benchmark table to obtain the basic threshold, determine the historical reference value and business sensitivity adjustment factor, and form the risk judgment threshold by weighted summation of the basic threshold and historical reference value and interval constraint.

[0041] Traverse the risk assessment index sequence, mark the time step where the first risk assessment index is greater than the risk judgment threshold as the initial maintenance time and trigger a maintenance pre-reminder;

[0042] The performance degradation index, failure frequency and system response time are extracted from the asset status time series dataset, and the weighted sum is used to obtain the health deviation degree and convert it into asset health.

[0043] Determine the estimated remaining useful life by setting a failure threshold and combining it with a degradation trajectory extrapolation method;

[0044] The time urgency is calculated based on the total number of steps predicted by the initial maintenance time statistics. The trend urgency is obtained by combining the slope of the risk assessment index sequence moving average after Sigmoid transformation and then weighted summation to obtain the maintenance urgency.

[0045] The asset maintenance score is calculated by weighting and summing the asset health status, estimated remaining useful life, and maintenance urgency.

[0046] Specifically, the steps for generating an asset risk traceability and maintenance scheduling report include:

[0047] All assets awaiting maintenance are sorted in descending order of their maintenance scores. If the maintenance scores are the same, the asset awaiting maintenance with the earlier initial maintenance date will be prioritized for maintenance.

[0048] Check assets within the sliding window whose risk assessment index is less than or equal to the risk judgment threshold but whose asset health is less than the preset health threshold, mark them as preventive maintenance candidates, assign them the lowest priority, and save them to the scheduling queue.

[0049] For each task in the scheduling queue, the particle swarm optimization algorithm is used to iteratively solve the optimal maintenance task allocation scheme. Resources are dynamically configured and scheduled based on task priority, availability of required spare parts, and skill matching of maintenance personnel.

[0050] The risk assessment index sequence, maintenance priority description, resource allocation status and related information within this assessment period are integrated to form an asset risk traceability maintenance scheduling report.

[0051] The asset risk tracing and maintenance scheduling system integrating feature matrices includes: a multi-source data processing module, a feature matrix construction module, a risk tracing and assessment module, and a maintenance scheduling module;

[0052] The multi-source data processing module is used to acquire multi-source status data, generate full lifecycle traceability codes for each asset after preprocessing, and generate asset attribute vector sets after structured extraction and normalization.

[0053] The feature matrix construction module is used to construct an asset status feature matrix, perform sliding window aggregation, trend denoising and outlier identification on feature sequences of each dimension, extract asset status vectors to generate asset status data streams, and associate full lifecycle traceability codes to generate asset status time series datasets.

[0054] The risk tracing and assessment module is used to acquire historical risk data to construct a risk behavior chain template library, construct an asset status prediction model, match the risk behavior chain template library to obtain risk acceleration factors, combine the asset business importance weights and environmental impact factors to establish a risk assessment function, and output a risk assessment index sequence.

[0055] The maintenance scheduling module is used to traverse the risk assessment index sequence to determine the initial maintenance time and trigger maintenance pre-reminders, comprehensively calculate the asset maintenance score, use the particle swarm algorithm to prioritize maintenance tasks and dynamically schedule maintenance resources, and generate an asset risk traceability maintenance scheduling report.

[0056] The beneficial effects of this invention are:

[0057] 1. This invention acquires multi-source status data and generates a unique full lifecycle traceability code. Using an asset status feature matrix as the core carrier, it completes standardized modeling and dimensionality reduction fusion of multi-source status data, constructing an asset status time-series dataset. This enables real-time tracking and dynamic tracing of the entire process of each asset from warehousing to disposal, overcoming the shortcomings of traditional management such as delayed traceability information and severe data silos. A status prediction model is constructed by integrating temporal convolutional networks and long short-term memory networks, effectively extracting multi-scale local temporal features and fault precursor features. A risk acceleration factor is matched using a risk behavior chain template library, and asset business importance weights and environmental impact factors are introduced to establish a risk assessment function that outputs a risk assessment index sequence. This completes the full-link traceability calculation of asset risk evolution, differentiates risk judgment thresholds, locates the initial maintenance time, and triggers maintenance pre-reminders. This eliminates reliance on fixed cycles or manual experience for maintenance time, fundamentally avoiding missed or over-maintenance, and improving asset operational reliability and risk traceability accuracy.

[0058] 2. This invention comprehensively calculates asset health, estimated remaining useful life, and maintenance urgency to obtain an asset maintenance score. A particle swarm optimization algorithm is used to prioritize all assets awaiting maintenance, ensuring that high-urgency and high-importance assets receive maintenance services first. For candidate assets for preventative maintenance, differentiated priority marking and health threshold checks within a sliding window are used to avoid inefficient or redundant maintenance tasks. At the resource scheduling level, dynamic configuration and scheduling are performed based on task priority, required spare parts availability, and the skill matching of maintenance personnel, generating an asset risk traceability maintenance scheduling report. This maximizes the utilization of limited human and spare parts resources, reduces overall asset lifecycle maintenance costs, and improves operational management efficiency. Attached Figure Description

[0059] Figure 1 A schematic diagram of an asset risk tracing and maintenance scheduling method that integrates feature matrices;

[0060] Figure 2 This is a flowchart illustrating the generation of asset status data streams in this invention;

[0061] Figure 3 This is a flowchart illustrating the output of the risk assessment index sequence in this invention;

[0062] Figure 4 This is a flowchart illustrating the comprehensive calculation of asset maintenance scores in this invention;

[0063] Figure 5 This is a structural diagram of an asset risk traceability and maintenance scheduling system that integrates feature matrices. Detailed Implementation

[0064] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0065] Example 1

[0066] refer to Figures 1 to 4 As shown in the figure, this embodiment introduces an asset risk tracing and maintenance scheduling method based on fused feature matrices, including the following steps:

[0067] Step 1: Acquire multi-source status data of enterprise assets through a unified data acquisition gateway. Preprocess the multi-source status data, and generate a unique full lifecycle traceability code for each asset according to a five-level hierarchical coding rule based on asset type, department, physical location, purchase batch, and unique serial number. Extract and normalize the multi-dimensional attributes corresponding to each asset in a structured manner to generate an asset attribute vector set. Introduce a dynamic attribute increment mechanism to incrementally update and manage the version of the asset attribute vectors. Whenever an asset changes, such as component replacement, system upgrade, or location relocation, automatically add a change tag and update the timestamp field in each asset attribute vector to ensure that the attribute information is synchronized with the actual asset status. The multi-source status data includes, but is not limited to, basic asset information, maintenance history data, location information data, and business-related data. Preprocessing includes data cleaning, format standardization, missing value imputation, outlier removal, and data anonymization.

[0068] Step 2: Based on the asset attribute vector set, an asset status feature matrix is ​​constructed using multi-dimensional feature fusion and principal component analysis dimensionality reduction algorithms. This matrix serves as the unified feature data foundation for asset risk tracing. After performing sliding window aggregation, trend denoising, and outlier identification on the feature sequences of each dimension of the asset status feature matrix, asset status vectors reflecting asset health trends are extracted, such as performance degradation index, system response time, and failure frequency. A continuous asset status data stream is generated according to the timestamp order, and the asset status data stream is associated and mapped with the full lifecycle traceability code of the corresponding asset to generate an asset status time series dataset bound to the full lifecycle traceability code. Among them, trend denoising is performed by using a weighted moving average algorithm to smooth the feature sequences of each dimension, filtering out random fluctuations and instantaneous interference, and extracting the long-term trend component of asset performance evolution. The performance degradation index is calculated based on the asset's rated benchmark performance, and the relative decline of the performance index within the current sliding window after trend denoising is calculated. After normalization, the index is obtained with a value range of [0,1].

[0069] Step 3: Obtain historical risk data from the enterprise asset management database and historical work order system, and construct a risk behavior chain template library; construct an asset status prediction model based on temporal convolutional networks and long short-term memory networks, use the asset status time series dataset and risk behavior chain template library as input to the asset status prediction model, calculate the risk acceleration factor based on the risk behavior chain template library, set the asset business importance weight and determine the environmental impact factor, establish a risk assessment function in combination with the behavior chain guidance mechanism, predict the risk assessment index at each future time step, output the risk assessment index sequence, and complete the full-link traceability calculation of asset risk;

[0070] Step 4: Based on different business levels, asset criticality, and historical fault distribution, set risk assessment thresholds. Traverse the risk assessment index sequence and mark the first time point where the risk assessment index exceeds the risk assessment threshold as the initial maintenance time and trigger a maintenance pre-reminder. This fundamentally solves the problems of missed and over-insured maintenance caused by the lag in traditional risk tracing and incomplete risk tracing links. Construct an asset maintenance mechanism, determine the asset health, estimated remaining useful life, and maintenance urgency, comprehensively calculate the asset maintenance score for each asset, and use particle swarm optimization to prioritize maintenance tasks and dynamically schedule maintenance resources, generating an asset risk tracing maintenance scheduling report.

[0071] In this embodiment, the specific steps in step 2 to generate the asset status time-series dataset bound with the full lifecycle traceability code include:

[0072] Feature domains strongly correlated with asset health status are selected from the asset attribute vector set, including but not limited to hardware operation indicators, maintenance history compliance, location change frequency, and business load pressure index. Continuous features in the feature domains are standardized to ensure a mean of 0 and a standard deviation of 1. Categorical features in the feature domains are dummy-coded to convert them into computable categorical feature values. An initial feature matrix is ​​constructed with asset instances as rows and feature dimensions as columns. During the dummy-coding process, the cascaded mapping relationship between each feature and the full lifecycle traceability code is recorded simultaneously to ensure subsequent traceability.

[0073] Based on the initial feature matrix, the correlation between each feature and the positive and negative labeled samples of historical asset faults is calculated to obtain the feature correlation coefficient. The calculation process includes: traversing the time series data of each dimension of the features in the initial feature matrix, matching the positive and negative labeled samples of historical asset faults at the corresponding time nodes, and using the Pearson correlation coefficient calculation method to calculate the linear correlation between the feature value changes and the fault occurrence event dimension by dimension to obtain the feature correlation coefficient, with a value range of [-1,1]; extracting asset business impact factors from the asset attribute vector set. The asset business impact factors are dimensionless normalized indicators with a value range of 0~1, generated by the asset business level and maintenance sensitivity through a minimum-maximum normalization mapping, used to quantify the impact of asset business importance on feature weight evaluation; and constructing feature importance weights by combining the feature correlation coefficient and the asset business impact factors. The specific construction process includes: for each feature... Features are assigned a balance coefficient based on the asset's business level and maintenance sensitivity to balance the proportion of feature relevance and business importance. The balance coefficient is determined as follows: the higher the asset's business level and maintenance sensitivity, the lower the balance coefficient, ensuring that the business weight contribution dominates the feature importance weight; conversely, the lower the balance coefficient, the higher the balance coefficient, ensuring that the feature association contribution dominates the feature importance weight. The feature association contribution value is obtained by multiplying the absolute value of the feature correlation coefficient by the balance coefficient. The difference between the value 1 and the balance coefficient is calculated and multiplied by the asset business impact factor to obtain the business weight contribution value. The feature importance weight is obtained by summing the feature association contribution value and the business weight contribution value. The positive and negative labels for historical asset faults are obtained by labeling historical maintenance work orders, fault records, and maintenance ledgers according to time nodes.

[0074] Principal component analysis (PCA) dimensionality reduction is employed, using feature importance weights as variance scaling factors for each feature dimension. This scaling method scales each feature dimension of the initial feature matrix, strengthening the dominant role of important features in the principal component directions. The covariance matrix of the scaled matrix is ​​then calculated to obtain a weighted covariance matrix. The eigenvalues ​​of each principal component in the weighted covariance matrix are statistically analyzed, and the sum of these eigenvalues ​​is obtained. The ratio of each principal component eigenvalue to the sum of eigenvalues ​​is calculated to obtain the single variance explained rate. The single variance explained rates of the first m principal components are then summed to obtain the cumulative variance explained rate. An explanation rate threshold of 95% is set based on asset type and business control precision requirements. This threshold is the default recommended value. Specifically, if the asset type is a key core asset and the business control precision requirement is high-precision, the explanation rate threshold can be increased to 98%–99% to retain more data. For subtle feature information, if the asset type is an edge auxiliary asset and the business control precision requirement is at the standard level, the explanatory power threshold can be lowered to 90%–93% to reduce redundant dimensions. For general business assets between high-precision and standard levels, maintain 95% to achieve a balance between dimensionality reduction and state representation fidelity. Retain the first n principal components whose cumulative variance explanatory power exceeds the explanatory power threshold, project the initial feature matrix onto a low-dimensional subspace guided by important business directions, and construct an asset state feature matrix with asset instances as rows and n principal component dimensions as columns. Here, m is the total number of principal components, equal to the total number of feature dimensions of the initial feature matrix; n is the number of principal components corresponding to the first time the cumulative variance explanatory power exceeds the explanatory power threshold; the low-dimensional subspace is a linear space formed by the directions of the first n principal components, with dimension n, which is much smaller than the feature dimension of the initial feature matrix.

[0075] For each asset, based on the corresponding full lifecycle traceability code, state feature vectors corresponding to all historical timestamps are extracted from the asset state feature matrix to form an initial state feature sequence. An adaptive sliding window is set, and the span of the sliding window is dynamically determined according to the recent state change rate of the asset. Specifically, this includes: setting a rate threshold based on the statistical distribution of feature fluctuations during historical stable periods, with the value set at the 90th percentile of the feature volatility rate, to filter out small feature fluctuations, and only using abnormal changes with the highest feature fluctuation magnitude as the criterion for adjusting the window span, balancing window stability and the ability to detect state abrupt changes; calculating the Euclidean distance between the state feature vectors of two adjacent time steps, and dividing the Euclidean distance by the interval between the two adjacent time steps to obtain the feature volatility per unit time; if the feature volatility rate is greater than the rate threshold, the window is automatically shortened by half the current window width to capture short-term sudden trends; if the feature volatility rate is less than or equal to the rate threshold, the current window width is maintained to smooth noise.

[0076] Within each sliding window, a weighted mean vector of the initial state feature sequence is calculated using a feature importance weighted averaging algorithm. The calculation process involves summing the state feature vectors at each time step within the sliding window, using the corresponding feature importance weights as coefficients. The summation is then divided by the sum of the feature importance weights to obtain the weighted average of each dimension, forming the weighted mean vector. A slope vector of the multi-dimensional features is calculated using a least-squares linear fitting method. This involves performing univariate linear regression on the initial state feature sequences of each dimension within the sliding window, with time as the independent variable and the feature value as the dependent variable. The slope coefficients of the regression equations for each dimension are extracted and concatenated to form a slope vector. The weighted mean vector and the slope vector are concatenated to form an aggregated feature vector. This aggregated feature sequence is then sorted sequentially by timestamp within each sliding window, thus preserving both the central trend and direction of change within the current time period, ensuring the removal of momentary fluctuations and highlighting trend information.

[0077] A local weighted regression method is used to smooth and fit the aggregated feature sequence to obtain trend feature components. The difference between the aggregated feature vector and the trend feature components is calculated to obtain the feature residual value. The feature residual values ​​of each time step are sorted according to the timestamp order to obtain the feature residual sequence. An anomaly discrimination criterion based on the sliding interquartile range is established for the feature residual sequence to identify outliers. The specific process of outlier discrimination includes: calculating the first and third quartiles of the feature residual sequence within the sliding window to obtain the interquartile range. If the feature residual value of a feature point within the sliding window exceeds the upper and lower critical intervals of the interquartile range, it is marked as a candidate outlier. The upper and lower critical intervals of the interquartile range are determined by taking the median value of the feature residual sequence within the sliding window as the benchmark, with the upper critical interval being the median value plus three times the interquartile range, and the lower critical interval being the median value minus three times the interquartile range.

[0078] The specific process of the locally weighted regression method includes: for the target feature point to be fitted, selecting W neighboring feature points before and after the target feature point to form a local window; setting Gaussian kernel weights for each feature point within the local window based on the temporal distance between each feature point and the target feature point, with the smaller the temporal distance, the larger the Gaussian kernel weight; performing polynomial fitting on each feature point within the local window using the weighted least squares method; constructing a polynomial design matrix with time as the independent variable; forming a diagonal weight matrix by combining the Gaussian kernel weights of each feature point; solving for the polynomial coefficients by minimizing the weighted sum of squared residuals; and then calculating the value of the polynomial at the target feature point, which is the trend feature component.

[0079] Historical maintenance event records are obtained from the historical maintenance data domain of the asset attribute vector set. The timestamps of candidate anomalies are aligned and compared with the historical maintenance event records. If a candidate anomaly appears within a reasonable warning time window before a known fault or maintenance event occurs, it is determined to be a valid anomaly and retained; otherwise, it is removed as random noise. Through automated filtering and marking of the above determination, the valid anomalies retained in the above determination process are uniformly marked as "valid" and their original feature values ​​and timestamp information are retained. The random noise that is removed is uniformly marked as "invalid". After performing the above marking operation on all time steps in sequence, a purified trend feature sequence and a list of valid anomalies with business confidence are obtained. Among them, the business confidence is the quality score of the historical maintenance event records associated with the valid anomalies. Specifically, it is based on whether the associated historical maintenance event records contain complete fault descriptions, handling measures and result confirmation information. The evaluation result is a value between 0 and 1 as the business confidence score attached to each valid anomaly.

[0080] For the trend feature components at each time step, the trend feature sequence within adjacent windows of each time step is used as context information to quantify asset health indicators with clear physical meaning, forming an asset state vector. The elements of the asset state vector include performance degradation index, system response time, and failure frequency. A continuous asset state data stream is generated in timestamp order. The full lifecycle traceability code is used as the primary key to associate and map with the asset state data stream and the list of valid outliers. The data is then aggregated according to the asset dimension to generate an asset state time series dataset bound to the full lifecycle traceability code.

[0081] In this embodiment, the specific steps for outputting the risk assessment index sequence in step 3 include:

[0082] Historical risk data is obtained from the enterprise asset management database and historical work order system. A hierarchical clustering algorithm with dynamic time warping is used to perform pattern clustering on asset state precursor sequence fragments intercepted from historical risk events. Cluster center sequences are extracted and associated with historical risk types and severity level labels to construct a risk behavior chain template library. The specific process of the hierarchical clustering algorithm includes: intercepting fixed-length asset state precursor sequence fragments from the time each risk event occurs, forming a precursor sequence sample set; determining a fixed length of 32 time steps based on the average duration of precursors of historical risk events and the asset state sampling frequency; using dynamic time warping distance as a sequence similarity measure, calculating the warping distance matrix between all sample pairs in the precursor sequence sample set; performing bottom-up hierarchical clustering based on the warping distance matrix, merging the sequence clusters with the smallest dynamic time warping distance, and calculating the average warping distance and silhouette coefficient within each sequence cluster. The average warping distance is the arithmetic mean of the dynamic time warping distances between all sample pairs within the sequence cluster, and the silhouette coefficient is a measure of the clustering results. The composite index of intra-cluster tightness and inter-cluster separation is calculated as follows: the average distance between each sample and other samples in the same sequence cluster is used as the intra-cluster dissimilarity, and the average distance between each sample and all samples in the nearest neighboring cluster is used as the inter-cluster separation. The difference between the inter-cluster separation and the intra-cluster dissimilarity is divided by the larger of the two to obtain the silhouette coefficient of each sample, with a value range of -1 to 1. Clustering is stopped when the average normalized distance is less than the distance threshold and the silhouette coefficient is greater than the silhouette coefficient threshold. The center sequences of each cluster are then associated with historical risk types and severity level labels to construct a labeled risk behavior chain template. The distance threshold is determined to be 0.15 based on the ratio of the maximum dynamic time normalized distance within the sequence cluster to the sequence length. The silhouette coefficient threshold is determined to be 0.75 based on the business requirements for risk identification accuracy. 0.75 indicates that when the silhouette coefficient of the clustering result reaches the silhouette coefficient threshold, the tightness of samples within the sequence cluster and the separation of samples between clusters are both at an excellent level. The corresponding cluster structure is clear and highly distinguishable, which can meet the business requirements for low false positive rate and high recall rate when matching risk patterns.

[0083] An asset state prediction model is constructed based on temporal convolutional networks and long short-term memory networks, including an input layer, a temporal feature extraction layer, a state prediction layer, and an output layer.

[0084] The input layer receives the asset state time series dataset and the risk behavior chain template library. It reshapes the asset state time series dataset according to asset instances and time windows to obtain a standardized state vector tensor with asset instances as samples and a fixed historical step size as the window. The fixed historical step size is determined to be 32 time steps by the uniform sequence length of the risk behavior chain template and the effective input dimension of the asset state time series dataset.

[0085] The temporal feature extraction layer consists of stacked dilated causal convolution modules. The dilated causal convolution modules are used to extract multi-scale local temporal features and fault precursor features from the standardized state vector tensor layer by layer. The receptive field is expanded by increasing the dilation coefficient layer by layer, and residual connection and batch normalization are performed to output a sequence of feature maps.

[0086] The state prediction layer, consisting of a two-layer Long Short-Term Memory (LSTM) network and an attention decoder, is used to capture long-range dependencies in the feature map sequence. The first LSTM layer receives the feature map sequence output from the temporal feature extraction layer and updates the hidden state step by step according to time steps, outputting the first hidden state sequence for the complete time step. The second LSTM layer receives the first hidden state sequence, further encodes long-term dependencies, and suppresses temporal noise, outputting the second hidden state sequence. The attention decoder consists of an attention mechanism and a single-layer LSTM. At each future prediction time step, the attention mechanism updates the current decoder hidden state and the second hidden state sequence. The dot product operation is performed on the hidden states of each time step, and the relevance score corresponding to each time step is obtained after softmax normalization. The relevance score corresponding to each time step is used as the weight coefficient, and multiplied with the hidden state of the corresponding time step in the second hidden state sequence respectively. The weighted sum is then accumulated element by element to obtain the context vector. The context vector is concatenated with the predicted output state vector of the attention decoder at the previous time step and then input into a single-layer LSTM to generate the asset state vector of the current prediction time step. This vector is used as the input of the next prediction time step for autoregressive prediction. The prediction state vector sequence of the next H time steps is generated in a loop.

[0087] The output layer receives the predicted state vector sequence for the next H time steps, normalizes the predicted state vector sequence to obtain a baseline risk value, and calculates the similarity score between the predicted state vector sequence and the risk behavior chain template library using a dynamic time warping algorithm and a Gaussian kernel function. This similarity score is used as the risk acceleration factor for the current time step. The specific calculation process includes: dynamically warping and aligning the predicted state vector sequence with each template sequence in the risk behavior chain template library to obtain the warped path distance; converting the warped path distance into a similarity score using a Gaussian kernel function; and selecting the highest similarity score as the risk acceleration factor for the current time step. Finally, by searching the key level mapping table of assets in the business system, the risk acceleration factor is obtained... The importance weight of the asset business corresponding to the asset is determined; environmental impact factors are determined based on the load, network fluctuations, and alarm density of the asset's operating environment. The specific determination process includes: collecting four environmental indicators from the environmental monitoring system—CPU load, memory load, network link packet loss rate, and alarm count per unit time—and linearly normalizing each indicator to the [0,1] interval according to its maximum and minimum values; and differentially setting the weights of each environmental indicator based on the contribution of each indicator to the occurrence of faults in historical risk events. CPU load fluctuations directly affect task processing capacity and have the highest contribution to fault triggering among environmental factors; network link packet loss rate affects data transmission integrity. The contribution of CPU load is similar to that of network load; the number of alarms is a passive indicator after a fault occurs, and its contribution is relatively low; memory load usually does not constitute the main cause of a fault, and its contribution is also relatively low. Based on the above contribution ranking, the weights of CPU load, memory load, network link packet loss rate, and alarm count are set to 0.3, 0.2, 0.3, and 0.2 respectively. The four normalized indicators are multiplied by their corresponding weights and then summed to obtain the environmental impact factor for each time step. A risk assessment function is established based on the behavioral chain guidance mechanism. The expression of the risk assessment function is: Risk Assessment Index = Baseline Risk Value × (First Factor Coefficient × Time-Step Risk Acceleration Factor + Second Factor Coefficient × Time-Step Risk Acceleration Factor) The risk assessment index is calculated as follows: (Environmental impact factor) × Asset business importance weight. The first factor coefficient is determined based on the proportion of the risk behavior chain matching degree in driving fault evolution, and is set to 0.6. The sum of the first and second factor coefficients is 1, resulting in a second factor coefficient of 0.4. A first factor coefficient of 0.6 indicates that the risk behavior chain matching degree contributes more to risk assessment than environmental factors. The risk behavior chain reflects the asset's own evolution trend and is a direct precursor to faults, hence it is assigned a higher weight. The risk assessment function is used to predict the risk assessment index for each future time step, outputting a risk assessment index sequence. The key level mapping table is determined comprehensively based on the asset's business system level, service impact scope, and user volume.

[0088] Using the asset state time series dataset as the training set and the weighted cross-entropy loss function adapted to the asset risk imbalance scenario as the loss function, the asset state prediction model is trained in conjunction with the Adam optimizer. After training, the latest real-time state segments and risk behavior chain template library in the asset state time series dataset are input, and the risk assessment index sequence is output through the output layer guided by time series feature extraction, state prediction and behavior chain.

[0089] For example, taking a corporate database server as an asset instance, historical risk data for the past year is obtained from the corporate asset management database and historical work order system. Through a hierarchical clustering algorithm with dynamic time warping, pattern clustering is performed on the asset status precursor sequence fragments intercepted with historical risk events as the endpoint. The cluster center sequences of each type are extracted and associated with historical risk type and severity level labels to construct a risk behavior chain template library.

[0090] An asset state prediction model is constructed based on temporal convolutional networks and long short-term memory networks, including an input layer, a temporal feature extraction layer, a state prediction layer, and an output layer. The input layer receives a temporal dataset of asset states at 32 time steps from the server and a risk behavior chain template library. The temporal dataset of asset states is reshaped according to asset instances and time windows to obtain a standardized state vector tensor with asset instances as samples and 32 time steps as windows. The temporal feature extraction layer consists of stacked dilated causal convolutional modules, which extract multi-scale local temporal features and fault precursor features from the standardized state vector tensor layer by layer. The receptive field is expanded by exponentially increasing the dilation coefficient layer by layer. The dilation coefficients are 1, 2, 4, and 8. Residual connections and batch normalization are performed to output a sequence of feature maps.

[0091] The state prediction layer consists of a two-layer long short-term memory network and an attention decoder. The first layer LSTM receives the feature map sequence and outputs the first hidden state sequence. The second layer LSTM further encodes the long-term dependency and outputs the second hidden state sequence. The attention decoder performs a weighted summation based on the relevance score to obtain the context vector, which is concatenated with the predicted output state vector of the previous time step and then input into a single-layer LSTM to generate a sequence of predicted state vectors for the next 6 time steps in a loop.

[0092] The output layer receives the predicted state vector sequence for the next 6 time steps, normalizes it to obtain the baseline risk value, and arranges the baseline risk values ​​of each time step in the predicted time sequence to form the baseline risk value sequence [0.15, 0.25, 0.48, 0.72, 0.56, 0.38]. The risk acceleration factor is calculated to be 0.82 using a dynamic time warping algorithm and a Gaussian kernel function. The asset business importance weight of the server is obtained by looking up the critical level mapping table, which is 0.9. Data collected from the environmental monitoring system shows CPU load of 78%, memory load of 65%, network link packet loss rate of 2.1%, and 5 alarms per unit time. These are linearly normalized to the [0,1] interval according to their maximum and minimum values. The alarm count was normalized using the historical maximum of 50 as the upper limit. CPU load, memory load, and network link packet loss rate were all processed by directly dividing the percentage by 100, resulting in normalized values ​​of 0.78 for CPU load, 0.65 for memory load, 0.21 for network link packet loss rate, and 0.1 for alarm count. Based on the contribution of various environmental indicators to the occurrence of faults in historical risk events, the weights for CPU load (0.3), memory load (0.2), network link packet loss rate (0.3), and alarm count (0.2) were differentiated. After weighted summation, the environmental impact factor was obtained as 0.78×0.3+0.65×0.2+0.21×0.3+0.1×0.2=0.447.

[0093] A risk assessment function is established based on the behavioral chain guidance mechanism. The expression of the risk assessment function is: Risk Assessment Index = Benchmark Risk Value × (First Factor Coefficient × Time-Step Risk Acceleration Factor + Second Factor Coefficient × Time-Step Environmental Impact Factor) × Asset Business Importance Weight, where the first factor coefficient is 0.6 and the second factor coefficient is 0.4; the risk acceleration factor and environmental impact factor are both time-series variables, and each future time step corresponds to a set of risk acceleration factors and environmental impact factors. The risk acceleration factor sequence is [0.82, 0.80, 0.77, 0.85, 0.83, 0.79], and the environmental impact factor sequence is [0.447, 0.432, 0.451, 0.463, 0.441, 0.428]. The risk assessment index for each future time step is calculated. The risk assessment index for the first time step is approximately 0.15 × (0.6 × 0.82 + 0.4 × 0.447) × 0.9 ≈ 0.0906. Based on the baseline risk value sequence and the risk assessment function expression, the corresponding baseline risk value, risk acceleration factor, and environmental impact factor are substituted into the risk assessment function for each time step. The risk assessment indices for the second to sixth time steps are calculated as follows: 0.1509, 0.2898, 0.4347, 0.3381, and 0.2294, respectively. The final output risk assessment index sequence is [0.0906, 0.1509, 0.2898, 0.4347, 0.3381, 0.2294]. Note that the above values ​​are illustrative examples in this embodiment; in actual applications, the parameter values ​​are dynamically updated based on real-time data.

[0094] Using a time-series dataset of asset states as the training set, and a weighted cross-entropy loss function adapted to scenarios of uneven asset risk as the loss function, an asset state prediction model is trained using the Adam optimizer. The parameters of the Adam optimizer are set to a learning rate of 0.001 and a first-order moment decay coefficient. Second-order moment attenuation coefficient Numerical stability constant Weight decay coefficient During training, a learning rate decay strategy is adopted. The validation loss is monitored every 20 training rounds. If the validation loss does not decrease for 5 consecutive rounds, the current learning rate is multiplied by the decay factor. The decay factor is determined to be 0.5 based on the experimental analysis of the balance between learning rate decay and loss convergence stability and convergence speed in historical training. When the validation loss decreases by less than 0.001 in 10 consecutive training rounds after three learning rate decays, or when the training rounds reach the maximum of 300 rounds, training is stopped and the model parameters with the smallest validation loss are selected as the final asset state prediction model parameters. After training, the latest real-time state segments and risk behavior chain template library in the asset state time series dataset are input. Through time series feature extraction, state prediction and behavior chain-guided output layer, the risk assessment index sequence is output as [0.0906, 0.1509, 0.2898, 0.4347, 0.3381, 0.2294].

[0095] In this embodiment, the specific steps in step 4 of generating the asset risk traceability maintenance scheduling report include:

[0096] For each asset, different business levels and asset criticality pairs are extracted from the asset attribute vector set, and the basic threshold is queried in the threshold benchmark table. Among them, the business level includes core level, important level and general level, and the asset criticality includes critical and general. The threshold benchmark table is constructed based on the basic risk control standards and industry-standard safety threshold specifications for different business levels and asset criticality.

[0097] The risk assessment index series of similar assets within the time window preceding historical failures is retrieved, and the 95th percentile of the risk assessment index series is calculated as a historical reference value to characterize the reference level of actual failure precursors. Risk judgment thresholds are set differentially through an adaptive fusion approach: a business sensitivity adjustment factor is set based on the asset's criticality, and the base threshold is weighted and summed with the historical reference value using this adjustment factor as a weighting coefficient. Specifically, considering the wide impact and high business interruption risk of critical asset failures, the business sensitivity adjustment factor for critical assets is set to 0.3, with the historical reference value having a weight of 0.7, making the risk judgment threshold highly correlated with actual failure symptoms; considering the smaller impact and higher business tolerance of general asset failures... The business sensitivity adjustment factor for general assets is set to 0.7, with the weight of historical reference values ​​at 0.3, making the basic threshold dominant. The initial risk assessment threshold obtained by weighted summation is constrained to the interval (0.3, 0.9). The constraint rules are as follows: if the initial risk assessment threshold is lower than the lower limit of the interval (0.3), then 0.3 is directly taken as the final risk assessment threshold; if the initial risk assessment threshold is higher than the upper limit of the interval (0.9), then 0.9 is directly taken as the final risk assessment threshold; if the initial risk assessment threshold is within the interval (0.3, 0.9), then the initial risk assessment threshold is retained as the final risk assessment threshold, thereby ensuring that the risk assessment threshold is always within the effective boundary of safety control. Finally, the risk assessment threshold for the asset is differentiated.

[0098] Traverse the risk assessment index sequence for the next H time steps, find the first time step in chronological order that satisfies the risk assessment index being greater than the risk judgment threshold, mark it as the initial maintenance time, and push a maintenance pre-reminder message through the unified data acquisition gateway. The reminder content includes the asset's unique full lifecycle traceability code, the expected maintenance timestamp, risk trend snapshot, and other information. If the risk assessment index of all time steps within the prediction window is less than or equal to the risk judgment threshold, then mark the asset as not requiring maintenance in the near future and determine the estimated number of days of safe operation.

[0099] The performance degradation index, normalized fault occurrence frequency, and system response time are extracted from the current state vector of the asset status time series dataset. The performance degradation index weight is set to 0.5 based on the characteristic that performance degradation has the greatest impact on asset health. The fault occurrence frequency weight is set to 0.3 based on the characteristic that fault occurrence frequency has the second greatest impact on asset health. Since the sum of the performance degradation index weight, fault occurrence frequency weight, and system response time weight is 1, the system response time weight is 0.2. The health deviation is obtained by multiplying the performance degradation index, fault occurrence frequency, and system response time by their respective weights and summing them. The difference between the value 1 and the health deviation is calculated to obtain the asset health score.

[0100] The performance degradation index at each time step in the predicted state vector sequence is used. Since the performance degradation index ranges from [0,1], the closer the value is to 1, the more severe the asset performance degradation. 0.85 corresponds to the critical state of asset performance degradation to near failure, so the failure threshold is set to 0.85. The remaining useful life estimate is determined using a degradation trajectory extrapolation method. The specific process includes: starting from the performance degradation index of the current time step, searching by time step within the sliding window. If a time step with a performance degradation index greater than or equal to the failure threshold is found, the sequence number of that time step is defined as the initial remaining useful life step number. If the performance degradation index of all time steps in the sliding window is less than the failure threshold, the performance index degradation slope of the last time step in the sliding window is selected for extrapolation to calculate the time step length required to reach the failure threshold as the initial remaining useful life step number. Dynamic correction is performed using a risk acceleration factor: the correction coefficient is determined to be 0.8 based on the historical degradation acceleration pattern of similar assets. By fitting and statistically analyzing historical failure samples of similar assets, the risk acceleration factor increases the rate of equipment degradation. A correction coefficient of 0.8 is set to eliminate the deviation in remaining life caused by pure theoretical extrapolation, thus conforming to the actual accelerated degradation pattern. The product of the correction coefficient and the risk acceleration factor is calculated to obtain the acceleration correction term. The acceleration correction term is summed with the value 1 to obtain the accelerated degradation coefficient. The ratio of the initial remaining life steps to the accelerated degradation coefficient is calculated to obtain the corrected remaining life steps. The corrected remaining life steps decrease as the risk acceleration factor increases, thereby reflecting the reduction effect of risk factors on remaining life. The remaining life steps are multiplied by the actual time interval between adjacent time steps to convert into an estimated remaining life in hours or days. Among them, the acceleration correction term represents the additional degradation rate increase brought about by the risk factor, and the accelerated degradation coefficient represents the overall degradation ratio after superimposed risk acceleration effect. The larger the accelerated degradation coefficient, the faster the asset degradation speed and the shorter the remaining life.

[0101] If the initial maintenance time does not exist, the maintenance urgency is set to zero; otherwise, the total number of predicted steps is calculated based on the risk assessment index sequence. The total predicted steps are then subtracted from the time sequence number corresponding to the initial maintenance time to obtain the difference in remaining maintenance steps. This difference is then divided by the total predicted steps to obtain the time urgency. Based on asset operation and maintenance standards, the risk assessment index sequences of three time sequences before and after the initial maintenance time are uniformly selected. The moving average slope of the risk assessment index sequences is calculated, and the trend urgency is obtained after applying a Sigmoid transformation to the moving average slope. Maintenance priority is then determined based on the remaining maintenance time. The core impact of priority is assigned a time urgency weight of 0.75. This value indicates that the remaining time factor dominates the maintenance urgency assessment; the shorter the remaining time, the higher the time urgency. The driving effect of the remaining time factor on maintenance priority is stronger than that of the risk change trend factor. The sum of the time urgency weight and the trend urgency weight is 1, so the trend urgency weight is set to 0.25. The trend urgency serves as an auxiliary reference factor to correct the assessment bias of time urgency under the condition of sudden change in risk trend. The maintenance urgency is obtained by weighted summation of the time urgency and the trend urgency.

[0102] An asset maintenance mechanism is constructed. Asset health status is weighted at 0.4, as it is the fundamental criterion for determining the necessity of maintenance. When an asset's health status falls below a threshold, the equipment no longer meets the conditions for safe operation and must be prioritized for maintenance. Asset health has the greatest fundamental impact on the asset maintenance score, hence it is given the highest weight. The remaining useful life is weighted at 0.35, based on its predictive contribution to maintenance urgency. The length of the remaining useful life directly determines the width and urgency of the maintenance time window, and its predictive contribution to maintenance urgency is relatively high, hence it is given the second highest weight. Maintenance urgency is weighted at 0.25, based on its impact on task prioritization. Maintenance urgency is mainly used for prioritizing assets to be maintained and supplements the scheduling results of asset health and remaining useful life assessments. Its overall impact on the asset maintenance score is relatively limited, hence it is given the lowest weight. The asset maintenance score for each asset is calculated by weighted summation, combining asset health status, remaining useful life estimate, and maintenance urgency.

[0103] All assets awaiting maintenance are sorted in descending order of their maintenance scores. If the maintenance scores are the same, the asset with the earlier initial maintenance date is prioritized for maintenance. Assets with a risk assessment index less than or equal to the risk judgment threshold but an asset health score less than the health score threshold within the sliding window are checked, marked as preventative maintenance candidates, assigned the lowest priority, and saved to the scheduling queue. The health score threshold is determined based on the minimum safe operation standard for different business levels and asset criticality, with a value of 0.6. This value is the baseline threshold, and in actual application, it is adjusted differently according to different business levels and asset criticality: for core-level critical assets, the threshold is increased by 0.05 to 0.1, i.e., the health score threshold is 0.65 to 0.7; for general-level general assets, the threshold is decreased by 0.05 to 0.1, i.e., the health score threshold is 0.5 to 0.55; and for other combinations, the baseline threshold of 0.6 remains unchanged.

[0104] For each task in the scheduling queue, a particle swarm optimization algorithm is used to iteratively solve for the optimal maintenance task allocation scheme based on asset maintenance score, maintenance timeliness requirements, and resource matching degree. Resources are dynamically configured and scheduled according to task priority, availability of required spare parts, and skill matching degree of maintenance personnel to ensure efficient utilization of maintenance resources. Specifically, dynamic resource configuration and scheduling includes: determining task priorities as high-priority, medium-priority, and low-priority maintenance tasks based on the threshold range of asset maintenance score; and prioritizing the locking of high-priority maintenance tasks. With matching maintenance personnel with the necessary skills, sufficient spare parts, and maintenance man-hours, maintenance tasks can be initiated immediately. For medium-priority maintenance tasks, corresponding maintenance resources are reserved and saved to the execution queue, and maintenance is carried out in an orderly manner according to the available time slots. For low-priority maintenance tasks, idle resources are coordinated to carry out maintenance operations in a staggered manner without affecting the maintenance of core assets. Spare parts inventory changes and personnel on-duty status are synchronized in real time, and resource allocation plans are dynamically adjusted to avoid resource idleness, resource conflicts, and maintenance delays, thereby achieving dynamic configuration and scheduling of maintenance resources across the entire domain.

[0105] The specific process of the particle swarm optimization algorithm includes: initializing the particle swarm, with each particle encoded as a three-dimensional vector. The first dimension represents the task execution sequence, the second dimension represents the maintenance personnel allocation sequence, and the third dimension represents the spare parts and labor resource allocation sequence. The fitness function is the comprehensive scheduling score. For each candidate allocation scheme, the complete maintenance scheduling execution process is simulated first, and the key indicators after the execution of each candidate allocation scheme are statistically analyzed. The particle velocity and position are iteratively updated until the maximum number of iterations is reached, and the globally optimal particle is output as the optimal maintenance task allocation scheme. The maximum number of iterations is determined by the total number of maintenance tasks in the scheduling queue and the number of available resource types, ranging from 50 to 200. The more tasks and the more complex the resource dimensions, the larger the number of iterations should be selected to ensure scheduling accuracy. The maintenance timeliness requirement is quantified by the on-time completion rate of high-priority maintenance tasks and the total delay time of maintenance tasks. The resource matching degree is comprehensively reflected by the maintenance personnel skill matching degree, the average utilization rate of all maintenance resources, and the number of resource conflicts.

[0106] The comprehensive scheduling score is a weighted composite of five sub-indicators: The first is the on-time completion rate of all high-priority maintenance tasks in the scheduling plan, which has the highest weight (0.3) because high-priority maintenance tasks have the strongest time constraints and directly affect the safe operation of assets; the second is the average skill matching degree of all assigned maintenance personnel, i.e., the degree of match between personnel skill levels and task-required skill levels, which has the second highest weight (0.25) based on the principle of optimal personnel-task fit; the third is the average utilization rate of all maintenance resources, as resource utilization rate is the core evaluation of the economic efficiency of overall maintenance scheduling. The weight of each indicator is set to 0.2; the fourth indicator is the normalized negative value of the total delay time of maintenance tasks. The total delay time of maintenance tasks is the total delay relative to the initial maintenance time of each task. Since the maintenance delay time reflects the timeliness of risk control, it needs to be appropriately constrained and set to a weight of 0.15; the fifth indicator is the normalized negative value of the number of resource conflicts during scheduling. Since resource conflicts only affect the smoothness of scheduling and have the weakest impact on the overall maintenance effect, a weight of 0.1 is set. The five indicators are normalized so that the values ​​of each indicator are unified to the range of 0 to 1. The comprehensive scheduling score is obtained by weighted summation according to the weight of each indicator.

[0107] By integrating the risk assessment index sequence, risk tracing link description, maintenance priority description, resource scheduling status and related information within this assessment period, an asset risk tracing maintenance scheduling report is formed. The asset risk tracing maintenance scheduling report uses the full life cycle traceability code as the link to link the asset ledger, forming a complete traceability record from risk warning to maintenance execution closed loop. It supports multi-terminal distribution and historical comparison analysis, fundamentally solving the problems of inconsistent feature modeling, incomplete traceability links and low data processing efficiency.

[0108] Example 2

[0109] Please see Figure 5 Another embodiment of the present invention provides: an asset risk tracing and maintenance scheduling system integrating feature matrices, comprising: a multi-source data processing module, a feature matrix construction module, a risk tracing assessment module, and a maintenance scheduling module;

[0110] The multi-source data processing module is used to acquire multi-source status data of enterprise assets and preprocess them. After preprocessing, a unique full life cycle traceability code is generated for each asset according to the five-level hierarchical coding rule. The multi-dimensional attributes corresponding to each asset are extracted and normalized in a structured manner to generate an asset attribute vector set. A dynamic attribute increment mechanism is introduced to perform incremental updates and version management of the asset attribute vector.

[0111] The feature matrix construction module is used to construct an asset status feature matrix based on the asset attribute vector set. After performing sliding window aggregation, trend denoising and outlier identification on the feature sequences of each dimension of the asset status feature matrix, the module extracts the asset status vector that reflects the asset health trend, generates an asset status data stream in the order of timestamps, and associates and maps it with the full life cycle traceability code of the corresponding asset to generate an asset status time series dataset.

[0112] The risk tracing and assessment module is used to acquire historical risk data, build a risk behavior chain template library, construct an asset status prediction model based on temporal convolutional networks and long short-term memory networks, calculate risk acceleration factors based on the risk behavior chain template library, set asset business importance weights and determine environmental impact factors, establish a risk assessment function in combination with the behavior chain guidance mechanism, predict the risk assessment index at each future time step, and output the risk assessment index sequence.

[0113] The maintenance scheduling module is used to set risk assessment thresholds differently, traverse the risk assessment index sequence, mark the first time point where the risk assessment index is greater than the risk assessment threshold as the initial maintenance time and trigger a maintenance pre-reminder; construct an asset maintenance mechanism, determine the asset health, estimated remaining useful life, and maintenance urgency, comprehensively calculate the asset maintenance score for each asset, use particle swarm optimization algorithm to prioritize maintenance tasks and dynamically schedule maintenance resources, and generate an asset risk traceability maintenance scheduling report.

[0114] Working principle and effects:

[0115] The multi-source data processing module acquires multi-source status data of enterprise assets. After preprocessing, a unique full lifecycle traceability code is generated for each asset according to the five-level hierarchical coding rule. The multi-dimensional attributes corresponding to each asset are extracted and normalized in a structured manner to generate an asset attribute vector set. A dynamic attribute increment mechanism is introduced to realize the incremental update and version management of attribute vectors, ensuring that the attribute information is synchronized with the actual status of the assets without lag. This eliminates the interference of information silos and data lag on risk traceability decisions and provides a high-quality data foundation for subsequent feature modeling and risk traceability assessment.

[0116] Based on the asset attribute vector set, an asset status feature matrix is ​​constructed through the feature matrix construction module. The feature sequences of each dimension are aggregated by sliding window, denoised by trend, and identified by outliers. The asset status vector reflecting the asset health trend is extracted, and an asset status data stream is generated in the order of timestamps. It is then associated and mapped with the full life cycle traceability code of the corresponding asset to generate an asset status time series dataset bound to the traceability code. This enables continuous perception and traceability of asset health status, breaks through the limitations of static and lagging status assessment in traditional asset management, and provides reliable time series data support for risk traceability assessment.

[0117] Historical risk data is obtained from the enterprise asset management database and historical work order system to construct a risk behavior chain template library. An asset status prediction model integrating temporal convolutional networks and long short-term memory networks is constructed through a risk tracing and assessment module. The asset status time series dataset and risk behavior chain template library are input, and a risk acceleration factor is calculated based on the risk behavior chain template library. The importance weight of asset business is set and the environmental impact factor is determined. A risk assessment function is established in combination with the behavior chain guidance mechanism to predict the risk assessment index at each future time step and output the risk assessment index sequence. This enables proactive tracing and prediction of asset risks, breaks through the lag limitation of traditional threshold alarms, and provides a tracing basis for hierarchical maintenance decisions.

[0118] By setting differentiated risk assessment thresholds through the maintenance scheduling module, and traversing the risk assessment index sequence, the first time point where the risk assessment index exceeds the risk assessment threshold is marked as the initial maintenance time and a maintenance pre-reminder is triggered. This fundamentally solves the problems of missed and over-maintenance caused by the lag in traditional risk tracing and the incomplete risk tracing chain. An asset maintenance mechanism is constructed to determine the asset health, estimated remaining useful life, and maintenance urgency. The asset maintenance score of each asset is calculated comprehensively, and the particle swarm algorithm is used to prioritize maintenance tasks and dynamically schedule maintenance resources, generating an asset risk tracing maintenance scheduling report. This ensures that maintenance decisions are aligned with actual business needs and achieves proactive risk tracing and control that prioritizes prevention while also considering efficiency.

[0119] Overall, through a four-layer modular architecture of multi-source data processing, feature matrix construction, risk tracing and assessment, and maintenance scheduling, a complete asset risk tracing data processing system based on feature matrices has been formed. This system combines data collection, feature modeling, risk tracing, and maintenance decision-making, enabling proactive prediction of risk tracing and maintenance timing throughout the entire asset management process. It solves the problems of inconsistent feature modeling, incomplete risk tracing links, and low data processing efficiency in traditional enterprise asset risk data processing, providing reliable technical support for the safe operation and maintenance of enterprise assets.

[0120] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for asset risk tracing, maintenance, and scheduling based on integrated feature matrices, characterized in that: include: After obtaining and preprocessing multi-source status data, a full lifecycle traceability code is generated, and after structured extraction and normalization, an asset attribute vector set is generated. Construct an asset status feature matrix, perform sliding window aggregation, trend denoising and outlier identification on feature sequences of each dimension, extract asset status vectors to generate asset status data streams, and associate full lifecycle traceability codes to generate asset status time series datasets. Historical risk data is used to construct a risk behavior chain template library, an asset status prediction model is built, risk acceleration factors are obtained by matching the risk behavior chain template library, and a risk assessment function is established by combining the asset business importance weight and environmental impact factors to output a risk assessment index sequence. The system iterates through the risk assessment index sequence to determine the initial maintenance time and triggers maintenance pre-reminders. It then comprehensively calculates the asset maintenance score, uses particle swarm optimization to prioritize maintenance tasks and dynamically schedule maintenance resources, and generates an asset risk traceability maintenance scheduling report.

2. The asset risk tracing and maintenance scheduling method based on the fused feature matrix according to claim 1, characterized in that, The specific steps for constructing the asset state feature matrix include: We screen feature domains that are strongly correlated with asset health status from the asset attribute vector set, and standardize continuous features and encode dummy variables for categorical features. Using asset instances as rows and feature dimensions as columns, an initial feature matrix is ​​constructed to calculate feature correlation coefficients and extract asset business influencing factors. Set a balance coefficient, calculate the feature association contribution value by combining the feature correlation coefficient, calculate the business weight contribution value based on the balance coefficient and the asset business impact factor, and sum them to obtain the feature importance weight; Principal component analysis dimensionality reduction algorithm is used to scale the initial feature matrix with feature importance weights as variance scaling factors to obtain a weighted covariance matrix; The cumulative variance explained rate is calculated by combining the eigenvalues ​​of each principal component in the statistically weighted covariance matrix with the sum of the eigenvalues. The first n principal components whose cumulative variance explained rate is greater than the preset explained rate threshold are retained. The initial feature matrix is ​​then projected onto a low-dimensional subspace to construct the asset status feature matrix.

3. The asset risk tracing and maintenance scheduling method based on the fused feature matrix according to claim 2, characterized in that, The specific steps for generating an asset status data stream include: For each asset, state feature vectors are extracted from the asset state feature matrix based on the full lifecycle traceability code to form an initial state feature sequence and an adaptive sliding window is set. Within each sliding window, a weighted mean vector and a slope vector are calculated based on the initial state feature sequence, and then concatenated to form an aggregated feature vector to create an aggregated feature sequence. The trend feature components are obtained by smoothing the aggregated feature sequence, and the feature residual values ​​are calculated by combining the aggregated feature vector to form the feature residual sequence. An anomaly discrimination criterion based on the sliding interquartile range is established to mark candidate anomalies. Acquire historical maintenance event records, perform time-series alignment and comparison with candidate anomalies, retain valid anomalies, and obtain trend feature sequences and a list of valid anomalies through filtering and labeling. For the trend feature components, the trend feature sequences within adjacent windows at each time step are used as contextual information to quantify asset health indicators to form an asset state vector and generate an asset state data stream.

4. The asset risk tracing and maintenance scheduling method based on the fused feature matrix according to claim 3, characterized in that, The specific steps for generating an asset state time series dataset include: Using the full lifecycle traceability code as the primary key, it is associated and mapped with the asset status data stream and the list of valid anomalies, and aggregated according to the asset dimension to generate an asset status time series dataset.

5. The asset risk tracing and maintenance scheduling method based on the fused feature matrix according to claim 4, characterized in that, The specific steps for outputting the risk assessment index series include: Acquire historical risk data, perform pattern clustering on the asset status precursor sequence fragments with historical risk events as the endpoint, extract the cluster center sequences of each type and associate them with historical risk type and severity level labels, and construct a risk behavior chain template library; An asset state prediction model is constructed based on temporal convolutional networks and long short-term memory networks; The asset state time series dataset was used as the training set, the weighted cross-entropy loss function was used as the loss function, and the Adam optimizer was used to train the asset state prediction model. After training, the latest real-time state segments and risk behavior chain template library from the asset state time series dataset are input. Through time series feature extraction, state prediction and behavior chain-guided output layer, the risk assessment index sequence is output.

6. The asset risk tracing and maintenance scheduling method based on the fused feature matrix according to claim 5, characterized in that, The specific steps for constructing an asset condition prediction model include: Construct an asset status prediction model, including an input layer, a time-series feature extraction layer, a status prediction layer, and an output layer; The input layer receives the asset state time series dataset and the risk behavior chain template library, and reshapes the asset state time series dataset according to asset instances and time windows to obtain a standardized state vector tensor. The temporal feature extraction layer consists of stacked dilated causal convolution modules. The dilated causal convolution modules are used to extract multi-scale local temporal features and fault precursor features from the standardized state vector tensor layer by layer. The receptive field is expanded by increasing the dilation coefficient layer by layer, and residual connection and batch normalization are performed to output a sequence of feature maps.

7. The asset risk tracing and maintenance scheduling method based on the fused feature matrix according to claim 6, characterized in that, The specific steps for constructing an asset condition prediction model also include: The state prediction layer consists of a two-layer long short-term memory network and an attention decoder. The first LSTM layer receives the feature map sequence, updates the hidden state step by step according to the time step, and outputs the first hidden state sequence. The second LSTM layer receives the first hidden state sequence, further encodes long-term dependencies and suppresses temporal noise, and outputs the second hidden state sequence. The attention decoder calculates the context vector based on the weighted summation of the second hidden state sequence, concatenates the context vector with the predicted output state vector of the previous time step, and then iteratively generates a sequence of predicted state vectors. The output layer is used to normalize and map the predicted state vector sequence to obtain the baseline risk value, calculate the similarity score between the predicted state vector sequence and the risk behavior chain template library as the risk acceleration factor, obtain the asset business importance weight and determine the environmental impact factor, establish a risk assessment function based on the behavior chain guidance mechanism to predict the risk assessment index, and output the risk assessment index sequence.

8. The asset risk tracing and maintenance scheduling method based on the fused feature matrix according to claim 7, characterized in that, The specific steps for comprehensively calculating the asset maintenance score include: Query the threshold benchmark table to obtain the basic threshold, determine the historical reference value and business sensitivity adjustment factor, and form the risk judgment threshold by weighted summation of the basic threshold and historical reference value and interval constraint. Traverse the risk assessment index sequence, mark the time step where the first risk assessment index is greater than the risk judgment threshold as the initial maintenance time and trigger a maintenance pre-reminder; The performance degradation index, failure frequency and system response time are extracted from the asset status time series dataset, and the weighted sum is used to obtain the health deviation degree and convert it into asset health. Determine the estimated remaining useful life by setting a failure threshold and combining it with a degradation trajectory extrapolation method; The time urgency is calculated based on the total number of steps predicted by the initial maintenance time statistics. The trend urgency is obtained by combining the slope of the risk assessment index sequence moving average after Sigmoid transformation and then weighted summation to obtain the maintenance urgency. The asset maintenance score is calculated by weighting and summing the asset health status, estimated remaining useful life, and maintenance urgency.

9. The asset risk tracing and maintenance scheduling method based on the fused feature matrix according to claim 8, characterized in that, The specific steps for generating an asset risk traceability and maintenance scheduling report include: All assets awaiting maintenance are sorted in descending order of their maintenance scores. If the maintenance scores are the same, the asset awaiting maintenance with the earlier initial maintenance date will be prioritized for maintenance. Check assets within the sliding window whose risk assessment index is less than or equal to the risk judgment threshold but whose asset health is less than the preset health threshold, mark them as preventive maintenance candidates, assign them the lowest priority, and save them to the scheduling queue. For each task in the scheduling queue, the particle swarm optimization algorithm is used to iteratively solve the optimal maintenance task allocation scheme. Resources are dynamically configured and scheduled based on task priority, availability of required spare parts, and skill matching of maintenance personnel. The risk assessment index sequence, maintenance priority description, resource allocation status and related information within this assessment period are integrated to form an asset risk traceability maintenance scheduling report.

10. An asset risk tracing and maintenance scheduling system based on a fused feature matrix, used to implement the asset risk tracing and maintenance scheduling method based on a fused feature matrix as described in any one of claims 1-9, characterized in that, include: Multi-source data processing module, feature matrix construction module, risk tracing and assessment module, and maintenance scheduling module; The multi-source data processing module is used to acquire multi-source status data, generate full lifecycle traceability codes for each asset after preprocessing, and generate asset attribute vector sets after structured extraction and normalization. The feature matrix construction module is used to construct an asset status feature matrix, perform sliding window aggregation, trend denoising and outlier identification on feature sequences of each dimension, extract asset status vectors to generate asset status data streams, and associate full lifecycle traceability codes to generate asset status time series datasets. The risk tracing and assessment module is used to acquire historical risk data to construct a risk behavior chain template library, construct an asset status prediction model, match the risk behavior chain template library to obtain risk acceleration factors, combine the asset business importance weights and environmental impact factors to establish a risk assessment function, and output a risk assessment index sequence. The maintenance scheduling module is used to traverse the risk assessment index sequence to determine the initial maintenance time and trigger maintenance pre-reminders, comprehensively calculate the asset maintenance score, use the particle swarm algorithm to prioritize maintenance tasks and dynamically schedule maintenance resources, and generate an asset risk traceability maintenance scheduling report.

Images