Operational decision-making auxiliary analysis method and system based on multi-source heterogeneous big data fusion

Abstract

The application discloses an operation decision auxiliary analysis method and system based on multi-source heterogeneous big data fusion, and belongs to the technical field of big data processing. The method firstly constructs a data tensor and a mask, separates clean tensors and abnormal tensors through adaptive robust tensor decomposition; compresses the characteristics of the clean tensors and estimates the cross-source time offset prior; fuses multi-source features by using deformable cross-attention, outputs feature sequences; constructs an entity relationship graph, combines multi-scale time coding and graph convolution gate recurrent units for space-time prediction, obtains future indicators and risk scores, and finally generates graded early warning and recommended strategies and visualizes them. Through adaptive robust tensor decomposition under the observation mask constraint, deformable cross-attention fusion with time offset prior constraint and space-time prediction based on dynamic graphs, the application realizes missing completion, abnormal separation, asynchronous alignment and entity relationship dynamic modeling of multi-source operation data, and can be used for operation indicator prediction, risk scoring and graded early warning.

Description

Technical Field

[0001] This invention relates to the field of big data processing technology, specifically to an operational decision support analysis method and system based on the fusion of multi-source heterogeneous big data. Background Technology

[0002] During enterprise digital transformation, operational decisions increasingly rely on multi-source time-series data from various business systems, including internal operational records, external market intelligence, supply chain sensor monitoring, and user feedback text. This data exhibits significant heterogeneity in terms of collection frequency, storage format, time stamping, and quality level. For example, some data is updated daily, while others are sampled at high frequencies; the reporting chains and transmission delays of different data sources vary, causing data reflecting the same operational events to be out of sync on the timeline. Furthermore, the data often contains random missing data, abnormal fluctuations, and noise, affecting the reliability of subsequent analysis.

[0003] Existing technologies for processing such multi-source heterogeneous data often employ simple concatenation after independent modeling or feature fusion based on fixed timestamp alignment. These approaches struggle to effectively compensate for time skew between data sources, potentially leading to the misinterpretation of misaligned features as corresponding ones, thus reducing the effectiveness of the fused representation. Furthermore, in modeling relationships between operational entities, many methods rely on static association graphs or ignore dynamic influences between entities, failing to capture interaction patterns that evolve over time, thus limiting the accuracy of operational metric predictions and risk warnings. Therefore, an operational decision support solution is needed that can balance data quality restoration, asynchronous alignment, adaptive fusion, and dynamic relationship modeling. Summary of the Invention

[0004] The purpose of this invention is to provide an operational decision-making support analysis method and system based on the fusion of multi-source heterogeneous big data, which can improve the effect of multi-source data fusion from aspects such as noise reduction, asynchronous alignment and dynamic relationship modeling, and improve the reliability of operational forecasting and risk warning.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: An operational decision support analysis method based on multi-source heterogeneous big data fusion includes the following steps: a data preprocessing and tensor construction stage, which performs time resampling, missing data labeling, anomaly detection, and normalization on multi-source operational time-series data, and constructs a data tensor and an observation mask tensor with a unified time granularity; an adaptive robust tensor decomposition stage, which separates the data tensor into a clean tensor and a sparse anomaly tensor under the constraint of the observation mask by minimizing the weighted nuclear norm and sparsity constraints; and a feature compression and shift estimation stage, which reduces the dimensionality of the clean tensor and extracts compressed features with unified dimensionality, while utilizing cross-correlation analysis. The process involves several stages: First, estimating the temporal offset prior of each source relative to the reference source. Second, a deformable cross-attention fusion stage takes compressed features and offset priors as input, outputting a fusion feature sequence at the entity level through a deformable cross-attention network. Third, a dynamic graph construction and spatiotemporal prediction stage dynamically constructs a directed adjacency graph between entities based on the fusion features, and uses multi-scale temporal coding and graph convolutional gated recurrent units for prediction, obtaining predicted values ​​for future operational indicators and risk scores. Fourth, an early warning and strategy generation stage generates tiered early warnings based on the comparison results between predicted values ​​and thresholds, as well as the risk score, and matches recommended adjustment strategies and a visualization interface.

[0006] Further, in step 2, the adaptive robust tensor decomposition specifically includes: An optimization objective is constructed by combining the weighted sum of the nuclear norms of the expansion matrices of each module with the L1 norm of the sparse outlier tensor. Under the constraint of the observation mask, the alternating direction multiplier method is used to solve the problem. In each iteration, the nuclear norm weight of each module is dynamically updated based on the spectral norm of the expansion matrices of the clean tensor obtained in the previous iteration. This results in modules with larger spectral norms receiving smaller weights to retain information, and modules with smaller spectral norms receiving larger weights to suppress noise.

[0007] Furthermore, in step 2, the weights assigned to each tensor modulus in the weighted nuclear norm are adaptively adjusted based on the spectral norm of the current expanded matrix of that modulus during the decomposition process.

[0008] Further, in step 3, the feature compression specifically includes: For each clean tensor from a data source, perform high-order singular value decomposition, truncate the factor matrix corresponding to the feature dimension, retain the column vectors that make the singular value cumulative energy ratio reach a preset ratio, and then use the truncated factor matrix to project the clean tensor along the feature dimension to a low-dimensional core tensor. The core tensor or its feature matrix after being split by entity is used as the compressed feature.

[0009] Furthermore, in step 4, the deformable cross-attention network generates learnable offsets for each time step through the offset prediction sub-network, performs deformable temporal sampling based on the offsets to obtain features of each source at the expected alignment position, and then uses a multi-head cross-attention mechanism to output a fused feature sequence at the entity granularity.

[0010] Furthermore, the offset prediction subnetwork consists of a one-dimensional causal convolutional layer, which receives the joint input of the reference source features and the source features to be aligned concatenated along the channel, outputs an initial offset estimate, and then obtains a learnable offset within the constraint range by mapping an activation function centered on the time offset prior and bounded by a preset relaxation range.

[0011] Furthermore, in step 5, an adjacency graph representing the strength of associations between entities is dynamically constructed, specifically as follows: Temporal context-enhanced encoding is performed on the fused features at each time step. The association score matrix between entities is calculated using a self-attention mechanism. The top k largest connections are retained based on the score value of each row and normalized to obtain a directed weighted adjacency matrix, where the value of k is logarithmically related to the total number of entities.

[0012] Further, in step 5, the spatiotemporal prediction network includes: a multi-scale temporal convolutional front-end, a graph convolutional gated recurrent unit, a prediction head, and a risk assessment head. The multi-scale temporal convolutional front-end has multiple parallel one-dimensional convolutional branches with different dilation rates and kernel sizes to capture temporal patterns of different granularities, and the outputs of each branch are fused and input into the stacked graph convolutional gated recurrent unit. The graph convolutional gated recurrent unit integrates neighbor information in the adjacency graph into the node hidden state update through Chebyshev graph convolution. Finally, the prediction head outputs the predicted values ​​of future multi-step operational indicators, and the risk assessment head outputs the risk score.

[0013] Furthermore, in step 6, tiered early warning information is generated, specifically as follows: Multiple threshold levels are set based on the historical statistics of the target indicators in the clean tensor. The predicted value is compared with the threshold to obtain the initial alarm level. The alarm level is then upgraded or downgraded based on the risk score. The recommended adjustment strategy is matched and output by the rule engine based on the change trend of the abnormal feature labels and offsets extracted from the sparse abnormal tensor and the preset knowledge base.

[0014] This invention also provides an operational decision support analysis system based on the fusion of multi-source heterogeneous big data, applied to the aforementioned operational decision support analysis method based on the fusion of multi-source heterogeneous big data. The system includes a data acquisition module, a data preprocessing module, a tensor anomaly detection and feature extraction module, a deformable cross-attention fusion module, a dynamic graph construction and prediction module, and a decision support output module connected in sequence. Each module is used to implement the corresponding steps in the above method and shares a distributed storage system to exchange intermediate data and model parameters.

[0015] In summary, the technical solution provided by this invention can improve the fusion quality and prediction reliability of multi-source heterogeneous operational data to a certain extent. Adaptive robust tensor decomposition, while preserving the data structure, fills in missing values ​​and separates outliers, helping to improve the input signal-to-noise ratio for subsequent analysis. A deformable cross-attention network utilizing cross-source time-shift prior constraints can learn reasonable time shifts for each data source, mitigating feature misalignment caused by acquisition and transmission delays. The dynamic graph construction mechanism and the spatiotemporal prediction network based on graph convolution and gated recurrent units can capture the time-varying characteristics of relationships between operational entities, improving the accuracy of future operational indicator predictions and risk assessments. The final generated early warning classification and adjustment strategy suggestions can provide relatively effective auxiliary information for the operational decision-making process, reducing the risk of false alarms and missed alarms. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a diagram showing the comparison between the projected sales curve and the actual value of a typical store over the next 7 days. Figure 3 A heatmap showing how the estimated time delay of each data source relative to the reference source changes over time. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0018] In this specification, the following terms have the following meanings: 1. Operating Entity: This refers to the basic unit of analysis, such as stores, product lines, regional business units, etc.

[0019] 2. Data Tensor: A three-dimensional array constructed after preprocessing each data source, with dimensions of time step, feature, and entity.

[0020] 3. Observation mask tensor: A binary tensor with the same shape as the data tensor, indicating whether the corresponding position is a valid observation.

[0021] 4. Clean Tensor: The low-rank component tensor obtained after adaptive robust tensor decomposition, representing the real operational signals.

[0022] 5. Sparse outlier tensor: The sparse component tensor obtained after adaptive robust tensor decomposition captures outliers and noise.

[0023] 6. Graph Convolution Gated Recurrent Unit (GC-GRU): A spatiotemporal prediction unit that integrates graph convolution operations into a gated recurrent unit, used to fuse node historical states and neighbor information.

[0024] 7. Risk Score: The 0-1 value output by the model represents the probability that the entity will fall into operational abnormality in the future.

[0025] 8. Warning Level: The warning level is determined based on a comprehensive assessment of the predicted value, threshold, and risk score.

[0026] like Figure 1 As shown, this invention provides an operational decision support analysis method based on the fusion of multi-source heterogeneous big data, including: Step 1: Obtain operation-related time-series data from multiple data sources, perform time resampling, missing value marking, anomaly candidate marking and normalization on the operation-related time-series data, and construct a data tensor and corresponding observation mask tensor with unified time granularity according to the time dimension, feature dimension and operation entity dimension; Step 2: For each data tensor from a data source, at the effective observation position defined by the observation mask tensor, make the sum of the clean tensor and the sparse outlier tensor match the data tensor at the effective observation position, and separate the data tensor into a clean tensor and a sparse outlier tensor by minimizing the weighted nuclear norm and sparsity constraints. Step 3: Perform feature compression on the clean tensors of each data source, extract compressed features with uniform dimensions and retain entity information, and use the cross-correlation of multi-source time series to estimate the time offset prior of each data source relative to the selected reference source. Step 4: Using compressed features and time-shifted priors as input, multi-source feature fusion is performed through a deformable cross-attention network to output a fused feature sequence at the entity level; Step 5: Based on the fused feature sequence, dynamically construct an adjacency graph representing the strength of association between entities at each time step, and input the adjacency graph and the fused feature sequence together into a spatiotemporal prediction network consisting of multi-scale temporal coding and graph convolutional gated recurrent units, and output the predicted values ​​of operational indicators and risk scores of each entity in future time steps. Step 6: Based on the comparison results of the predicted values ​​of operational indicators with the preset thresholds and the risk score, generate graded early warning information, combine the abnormal patterns recorded in the sparse anomaly tensor to output recommended adjustment strategies, and generate a visual decision support interface.

[0027] Next, this invention will provide a detailed explanation of the above method with specific parameter settings: In step 1, operation-related time-series data from multiple data sources are acquired. This data undergoes time resampling, missing value marking, anomaly candidate marking, and normalization. A data tensor with a unified time granularity and a corresponding observation mask tensor are constructed according to the time dimension, feature dimension, and operation entity dimension. Specifically: Step 1 involves acquiring and organizing raw data from multiple business processes to provide structurally consistent and quality-controlled input for subsequent steps. This step includes three sub-steps: data acquisition, source-specific preprocessing, and tensor construction. Step 101: Data Acquisition Through standard interfaces with enterprise information systems and external platforms, four types of operational data are acquired in real time or periodically, including: The first category is internal operational data, sourced from the Enterprise Resource Planning (ERP) system and stored in structured forms. Typical fields include sales revenue, operating costs, inventory levels, number of employees on duty, and order delivery cycles for each operating entity on a specific date. This type of data is usually updated daily and naturally includes timestamps and entity identifiers.

[0028] The second category is external market data, obtained through the application programming interfaces (APIs) of partner data providers, and is semi-structured data. This includes competitor price change records, industry prosperity indices, regional market capacity reports, etc. The update frequency of this type of data fluctuates; some indicators are released daily, while others are released weekly or monthly, requiring date alignment after acquisition.

[0029] The third category is supply chain data, which comes from IoT sensors deployed in warehousing environments and production equipment. These sensors continuously collect environmental physical quantities, such as temperature and humidity in the warehousing area, vibration amplitude of equipment operation, and power consumption. The raw sampling frequency can reach the second or minute level, which is high-density time-series data.

[0030] The fourth category is user feedback data, which is obtained from public platforms through web scraping and is unstructured text. It mainly includes user comments about products, news headlines, social media mentions, etc., with each piece of content accompanied by a timestamp.

[0031] Of the data mentioned above, the first three categories can be directly obtained as numerical or categorical fields; the fourth category exists in natural language form and needs to be embedded and extracted to be converted into numerical features.

[0032] Step 102: Source-Specific Preprocessing Different preprocessing operations are performed according to the characteristics of different data sources, but all are based on a unified time window and entity segmentation standard.

[0033] First, determine the time window granularity as daily, and define the total number of time steps as follows. The data is arranged in calendar order from the earliest available date to the current analysis cutoff date. For all data sources, the raw data is split by daily boundaries, with records falling on the same calendar day grouped into the same time step.

[0034] For internal enterprise operational data, which is inherently a structured record at the daily granularity, the values ​​of each field are directly extracted and organized by operational entity. If a field is missing for a particular entity on a certain day, the missing location is filled with... Fill in and mark the corresponding observation mask element as... This indicates that the observation at that location is invalid. The observation mask element for a valid data location is denoted as... Numerical fields are outlier detected based on their distribution across all entities and time steps: the interquartile range (IMR) of the field is calculated, and values ​​below the lower quartile minus 1.5 times the IMR or above the upper quartile plus 1.5 times the IMR are considered outliers and treated as missing values, with imputation performed accordingly. And record the mask as Then, min-max normalization is performed on each feature dimension to map the values ​​to... Interval.

[0035] For external market data, which is mostly numerical indicators accompanied by descriptive text fields, the numerical values ​​are extracted first and retained as a numerical sequence. If the text descriptions exist, they are merged with user feedback text. The numerical indicators are also aligned daily, and the handling of missing and outlier data is consistent with that of internal operational data. Extreme values ​​for each dimension are calculated independently during normalization.

[0036] For supply chain sensor data, the raw sampling density is much higher than the daily granularity, therefore aggregation is necessary within a daily window. For each sensor channel, the mean and standard deviation of all samples within the daily window are calculated as the daily features. For example, a temperature sensor generates [data] within a day. If a reading is given, then the characteristics of that day include the mean. and standard deviation Based on this, the resulting aggregated features are subjected to the same missing data handling, anomaly detection, and max-min normalization as the internal enterprise data.

[0037] For user feedback text data, a pre-trained language model is used for semantic embedding. Each piece of text is input into the pre-trained language model, and the sentence embedding vector from the last layer of the model is taken as the text representation. This vector is usually of fixed dimensions, such as... Dimension. To ensure that the embedding vector is compatible with other features in terms of numerical range, the embedding is... Normalize it so that its modulus is If multiple texts exist within a single calendar day, the average of all embedding vectors for that day is taken to obtain a text feature vector representing that day. For missing data handling, if no text data exists on a given day, the text feature vector for that day is filled in across all dimensions. And the observation mask is set at the corresponding position. .

[0038] After the source-specific preprocessing described above, each data source forms a feature set organized by time step and entity, and the numerical units are unified, thus preparing for tensor construction.

[0039] Step 103: Tensor and Mask Construction Within the defined analysis range, there are a total of The first operating entity can be a store, product line, or regional business unit, depending on the application scenario. For the first... There are data sources, and their feature dimensions are denoted as . The preprocessed data from the source is then organized into a third-order data tensor. The three dimensions correspond to time step, feature, and entity, respectively. For indicators in external market data that do not distinguish between entities (i.e., global indicators), the feature is copied and filled in the entity dimension to ensure that the tensor shape is consistent.

[0040] Simultaneous construction of observation mask tensor Its elements The rule for determining the value is: if Corresponding position If the data consists of valid observations that have been normalized or aggregated and have not been filled with zero values ​​due to missing or outliers, then... If the location was originally missing or was converted to missing fill after anomaly detection, then Thus, the mask tensor explicitly indicates whether each element in the data tensor is reliable or not.

[0041] The output of step 1 is Each data source's respective data tensor With observation mask tensor These two types of tensors will serve as direct inputs to the adaptive robust tensor decomposition in step 2, guiding the recovery and denoising of missing and outlier locations during modeling.

[0042] In step 2, for each data tensor from a data source, at the effective observation positions defined by the observation mask tensor, the sum of the clean tensor and the sparse outlier tensor is matched with the data tensor at the effective observation positions. The data tensor is then separated into a clean tensor and a sparse outlier tensor through weighted nuclear norm minimization and sparsity constraints. Specifically: Step 2 takes over the source data tensors output from Step 1. With observation mask tensor Its core task is to perform structured decomposition on each data source, separating the raw observation data into clean tensors carrying true information and sparse anomaly tensors capturing anomalies and noise. The output of this step is the clean tensor of each source. It is directly used as the input for multi-source feature extraction in step 3.

[0043] Next, this invention will be described in detail from five aspects: input and problem modeling, objective function construction, adaptive weighting mechanism, alternating direction multiplier method for solving, and output.

[0044] 1. Input and Preprocessing For the Data sources ( Enter two items: Data tensors ,in To ensure a uniform total number of time steps, For the feature dimension of this source, The number of operating entities.

[0045] Observation mask tensor ,element This indicates that the entry is a valid observation. This indicates an invalid item that is missing or has been set to empty after an anomaly detection.

[0046] Since the source data has been normalized in the previous step and the numerical dimensions of each feature dimension have been unified, it can be directly used for tensor decomposition modeling without further scale adjustment.

[0047] 2. Problem Modeling and Objective Function Construction Real-world operational data typically exhibits multimodal, low-rank characteristics. For example, operational metrics show periodicity and trends over time, some metrics exhibit collinearity over features, and similar stores or products show convergent change patterns over entities. This characteristic implies that clean data tensors... The structure should exhibit low-rank properties on each module expansion matrix. Meanwhile, anomalous fluctuations, acquisition noise, and input errors do not possess low-rank properties; therefore, sparse anomaly tensors are used. To depict.

[0048] Based on the above understanding, this step models data recovery as a constrained optimization problem: under the constraint of the observation mask, finding the one that best fits the low-rank prior. and sparse prior This ensures that the sum of the two data points equals the original data at the effective observation location. The problem can be expressed as: ; ; In the objective function: Represents a clean tensor Along the first mold( The matrices obtained by expanding the corresponding time module, feature module, and entity module have the following shapes in order: , , .

[0049] Let be the matrix nuclear norm, which is the sum of the singular values ​​of the matrix. It is a convex relaxation of the rank function. Minimizing this term promotes the low-rank property of the expanded matrix.

[0050] For tensor Norm, which is the sum of the absolute values ​​of all elements, ensures that It has a sparse structure.

[0051] This is the sparsity penalty coefficient, used to balance the strength between low-rank components and sparse components. It is typically determined through cross-validation based on the data noise level; in this embodiment, it is taken as... .

[0052] For the first The weighting coefficient of the modulus is the core mechanism of this improvement and innovation, which is used to adaptively adjust the low-rank penalty intensity for different expansion directions.

[0053] Constraints are expressed using the Hadamard product. Only valid observation locations are allowed to participate in the fitting; tensor elements at missing locations are naturally filled in by the low-rank structure.

[0054] 3. Adaptive weighting mechanism Standard nuclear norm minimization assigns equal weights to the three expansion modules, but the degree of low rank varies significantly across different data on different modules. For example, when the number of entities is small but the feature dimensions are high, the feature modules may contain more redundancy, and the singular values ​​of their expansion matrices decay rapidly. In this case, the nuclear norm penalty of the feature modules should be increased to encourage strong low rank. Conversely, the time modules may contain rich trend information, and excessive penalty can lead to signal distortion. Fixed weights cannot accommodate this diversity.

[0055] This invention designs an adaptive weighting strategy based on iterative spectral norm feedback. In the first... In the alternating direction multiplier method iteration, the weights Instead of setting it as a constant, it is based on the clean tensor obtained from the previous iteration. The spectral norm of each modulus expansion matrix is ​​dynamically updated: ; in It is the matrix spectral norm (i.e., the largest singular value). To prevent the denominator from being zero, take a small positive number. In the initial iteration, there is no previous result, so... Set as unified .

[0056] The underlying logic is as follows: if the spectral norm of a certain module expansion matrix is ​​large, it indicates that the currently recovered tensor has significant energy or structural changes in that module direction. Therefore, its weight should be appropriately reduced to alleviate the compression of that direction by the kernel norm and protect effective information. Conversely, if the spectral norm is small, it means that there is almost no significant structure in that direction. The weight automatically increases to strengthen the low-rank constraint and eliminate residual noise. This mechanism allows the weights to automatically adapt to the internal structure of each data source as the iteration progresses, without the need for manual parameter tuning. It achieves a more refined balance between denoising and structure preservation than fixed regularization.

[0057] 4. Model Solution: Alternating Direction Multiplier Method Introducing the Lagrange multiplier tensor (Initialized to an all-zero tensor) and penalty coefficient Construct the augmented Lagrangian function: + + in For tensor inner product, For the Frobenius norm, Initially set to And after each iteration, multiply by a factor slightly greater than 1 (e.g.) The constraints are gradually increased. The solution is achieved by alternately updating the variables: renew :fixed and This subproblem degenerates into soft-threshold shrinkage for each element. Define the residual. ,but Each element is updated to ,in for The element value.

[0058] renew :fixed and The subproblem is equivalent to the tensor under missing mask data. We perform weighted nuclear norm minimization recovery. Using the proximal gradient method, we first calculate the current target tensor: ; in Indicates and All-1 tensors of the same shape This is the clean tensor from the previous iteration. The meaning of this expression is: valid observation locations use the current residual, and missing locations use the recovered values ​​from the previous iteration. Then, for... The three moduli are sequentially subjected to singular value thresholding. Specifically, for the first... Model, will Expanding along this modulus into a matrix, performing singular value decomposition, and separating the singular values... Soft thresholding The matrix is ​​reconstructed and then folded back into a tensor. After processing the three modules sequentially, the current round is obtained. Because the weights are adaptive, the reduction of singular values ​​varies for each module, thus enabling selective preservation of the structure of each module.

[0059] renew Standard dual ascent: ; Before the end of each iteration, update the weights of each module according to the aforementioned formula. The convergence condition is as follows: iteration stops when the relative residuals satisfy the following condition: ; Or the number of iterations reaches its limit (e.g., 200 times). Introduce a formula into the denominator. Prevent division by zero errors in extreme cases.

[0060] After decomposition, each data source outputs a clean tensor. This tensor fills in missing values, eliminates most of the anomalous fluctuations, and maintains the multimodal low-rank structure of the data. Sparse Anomalous Tensor The separated strong noise components are recorded and can be selectively used for anomaly tracing, but are no longer passed on in the regular process.

[0061] These clean tensors are then used in step 3 for multi-source feature extraction and alignment. Since step 2 has already independently cleaned the data for each source, the cross-correlation analysis and feature compression in step 3 will be based on cleaner data, effectively avoiding the interference of outliers on the accuracy of cross-source alignment.

[0062] In step 3, the clean tensors of each data source are compressed to extract compressed features with uniform dimensionality and preserved entity information. The cross-correlation of multi-source time series is then used to estimate the prior time offset of each data source relative to a selected reference source. Specifically: Step 3 uses the clean source tensors output in Step 2. Using this as input, two tasks are performed: First, high-dimensional features from each source are compressed into a compact low-dimensional representation using multilinear mapping to obtain the time series feature matrix. Second, the time offset interval is initially estimated through cross-source cross-correlation analysis, providing initialization constraints for the deformable cross-attention in step 4. The output is the feature matrix of each source (and the corresponding entity dimension information) and the initial offset of each source relative to the reference source. Both are directly used in step 4 for the construction of the fusion network.

[0063] The following sections will elaborate on four parts: input analysis, feature compression, offset estimation, and output and concatenation.

[0064] 1. Input Analysis and Processing Objectives Step 2 is the first Data sources ( This provides clean tensors. ,in To standardize the number of time steps, This represents the original feature dimension of the source. The number of operating entities. Due to... Missing values ​​have been filled in and outliers have been filtered out. Its internal structure can more accurately reflect the real operational status under different times and entities.

[0065] However, there are two obstacles to directly using these clean tensors in cross-source fusion. First, the feature dimensions of each source... The differences are significant. For example, internal enterprise operational data may include dozens of fields, sensor aggregated features may only have a few channels, and text embeddings can reach hundreds of dimensions. Directly feeding high-dimensional heterogeneous features into an attention network can lead to uncontrollable parameter scale and the risk of overfitting. Second, although the source data are all segmented by date, there are inherent time asynchronies in the actual collection, aggregation, and reporting processes. For example, market data release lags behind operational data, and sensor data uploads are delayed. If attention is calculated by forcibly aligning time steps, features with misaligned times will be mistakenly matched, introducing semantic confusion.

[0066] Therefore, step 3 requires feature compression of each source clean tensor to eliminate the original dimensional differences and project the data into a compact and semantically unified low-dimensional space; at the same time, temporal correlation is used to estimate the time lag between each source and the reference source, providing physically meaningful initial values ​​and search range for the learnable temporal offsets in the subsequent attention network.

[0067] 2. Feature Compression: Multi-mode Dimensionality Reduction Based on Tucker Decomposition Clean Tensor The three modules correspond to time, features, and entities, respectively, with the feature module (second module) having the following dimensions. This is the main target for eliminating redundancy. This invention uses Tucker decomposition for feature compression, which expresses the original tensor as the product of a core tensor and three orthogonal factor matrices, and achieves feature magnitude reduction by truncating the factor matrices.

[0068] The specific operation is divided into the following steps: (1) Higher-order singular value decomposition. For Perform full-order Tucker decomposition to obtain factor matrices of the three moduli. , , and full-order core tensor .in The column vectors form an orthogonal basis for the feature space, arranged in descending order of their corresponding singular values.

[0069] (2) Feature truncation. Set the target dimensionality reduction dimension. The selection of this factor requires a balance between information preservation and compression. For each source, the singular value cumulative energy ratio is calculated, and the minimum ratio satisfying the following conditions is selected. : ; That is, at least 95% of the characteristic mode energy is retained, of which for The second modulus expansion matrix of the first One singular value. Only take The former The columns form the cutoff factor matrix The clean tensor is then projected along its eigenmode onto this low-dimensional subspace to obtain the compressed core tensor: ; in This represents the second modular multiplication of the tensor (i.e., contraction along the feature dimension). The core tensor is now located in the compressed feature space, with the feature dimension reduced to [value missing]. .

[0070] (3) Entity dimension processing and feature matrix organization. Two organization methods can be selected based on backend analysis requirements: Entity retention method: If downstream tasks require differentiation between different operational entities (such as the entity-level dynamic graph construction in step 5), then the entity dimension is directly retained. Define entities. The characteristic matrix is , its first The core tensor is the action. In time and entity The eigenvectors below The subsequent step 4 will perform deformable cross-attention fusion on each entity, and the fusion output will still retain the entity dimension, which will facilitate the construction of a dynamic graph in step 5 using the relationships between entities.

[0071] Entity aggregation method: If the analysis object is a global operational indicator and there is no need to distinguish between entity differences, then average pooling is performed on the core tensor along the entity module to obtain the aggregated feature matrix. Its row vector is: ; This method removes entity granularity and condenses information from each source into a single time series, making it suitable for the scenario in step 4 where cross-source fusion is performed using aggregated features.

[0072] In both approaches, the core tensor carries the main information of the clean tensor, and the feature dimension is significantly reduced, which is beneficial to the efficiency and numerical stability of subsequent attention calculation.

[0073] 3. Cross-source time migration estimation Different data sources may have different time lags in the actual acquisition process. For example, supply chain sensor data may be uploaded after being aggregated by edge devices, which may have a delay of several hours; market data may be released after being processed by suppliers, which may have a delay of several days. In step 4, deformable cross-attention compensates for this asynchrony by learning offset parameters, but if these parameters are not properly initialized, they are very likely to get stuck in local optima or produce unreasonable large offsets.

[0074] This step utilizes cross-correlation analysis to provide initial values ​​and a reasonable search range for the learnable offset. The source with the best data integrity and feature stability is selected as the reference source, denoted as Source 1 without loss of generality. Other sources... The time cross-correlation function between the source and the reference source is calculated to determine the relative delay.

[0075] To eliminate feature scale differences, the feature matrices of each source are first aggregated. (If entity retention is used, the aggregated sequence is first obtained by averaging along the entity modulo.) Standardize each dimension to make its mean equal to... Standard deviation is Then, calculate the multivariate cross-correlation measure between the two sources. Define the reference source aggregation sequence. With source Delay The aggregation sequence after the step The mean cosine similarity between them is used as a correlation function: ; in In order to make and All are valid time steps within the index range. To prevent small constants with a denominator of zero. The range of values ​​is set to , Determined based on actual business needs, for example, taking The coverage should be within a reasonable delay range.

[0076] The delay at which the correlation function reaches its maximum value is used as the initial offset estimate: When the maximum correlation value is lower than the preset threshold (e.g.) When there is no definite time correspondence between the two sources, it is assumed that there is no definite time correspondence between them. Set as And in the subsequent step 4, the search range of learnable offsets is appropriately widened.

[0077] This offset estimation provides two key priors for the deformable attention in step 4: first, an initial offset value, which serves as an initial reference for the bias term in the offset prediction network, accelerating training convergence; and second, an offset search range, constraining the learnable offset parameters within a certain range. Within the interval, For relaxation margin, it is usually taken as To prevent attention from seeking alignment within an unreasonable timeframe.

[0078] 4. Output and Backlink Step 3 outputs the following content, which is then passed to Step 4: The compressed feature representations of each source. If entity preservation is used, the output is the core tensor of each source. The feature matrix corresponding to each entity If entity aggregation is used, the output is an aggregated feature matrix. .

[0079] Initial time offset of each source relative to the reference source and search margin This is used to initialize the offset parameters and constraint range of the deformable cross attention network.

[0080] These outputs enable step 4 to perform cross-source fusion within an information-compact and dimensionally compatible feature space, and to guide attention alignment with accurate offset priors, thereby improving the overall accuracy and robustness of multi-source heterogeneous data fusion.

[0081] In step 4, compressed features and temporal offset priors are used as inputs, and multi-source features are fused through a deformable cross-attention network to output a fused feature sequence at the entity level, specifically: Step 4, based on the compressed features and time offset priors provided in Step 3, completes cross-source alignment and deep fusion of multi-source temporal features. Its input is the compressed feature representations of each source, and its output is a unified fused feature sequence, which will serve as the core input for constructing the entity dynamic graph in Step 5.

[0082] The core of this step is to construct and train a deformable cross-attention network. This network compensates for the temporal asynchrony between data sources through learning and integrates multi-source information into a compact joint representation at the entity granularity. The following details the process from four aspects: input and problem analysis, network structure design, training scheme, and output and concatenation.

[0083] 1. Input and Problem Analysis Step 3 provides two items to this step: Compression characteristics of each source: An entity retention approach is adopted, for each operational entity... and each data source Provide feature matrix ,in To standardize the number of time steps, This represents the feature dimension after compression of the source.

[0084] Cross-source time offset prior: Taking source 1 as the reference source, for each other source... Provides initial offset estimation and slack margin The offset of source 1 itself is always 0.

[0085] Directly Point-by-point stitching or ordinary attention fusion with time-step alignment can lead to feature misalignment due to inherent acquisition and publication delays between sources. For example, market sentiment data for a particular day may actually reflect operational conditions from several days prior. Forcibly treating features under the same date index as corresponding data introduces semantic noise and degrades the quality of the fused representation. Therefore, a fusion mechanism that can adaptively adjust cross-source time correspondence is needed.

[0086] The mechanism must meet three requirements: First, the offset adjustment should be learnable to adapt to complex delay patterns of different entities and time periods; second, the learning process should be subject to prior constraints to prevent the offset from diverging or falling into meaningless ranges; and third, the fusion result should maintain a high-fidelity reconstruction capability of the original multi-source information to ensure that the fusion does not lose key operational signals.

[0087] 2. Deformable cross-attention network structure This invention designs a deformable cross-attention network. Based on the standard cross-attention mechanism, this network introduces a deformable sampling module and a shift prediction subnetwork, thereby giving the network the freedom to actively adjust cross-source temporal alignment. The overall network structure consists of three components: a shift prediction subnetwork, a deformable temporal sampling layer, and a cross-attention fusion module.

[0088] (1) Migration prediction subnetwork For the source to be merged ( (Source 1 is used as a reference and no offset prediction is needed). This subnetwork is based on the reference source and source... The local temporal characteristics of each time step Predict a differentiable continuous offset .

[0089] The sub-network structure is a lightweight one-dimensional convolutional module. Specifically, it references source features. With source feature Concatenated along the time dimension by channel, forming a dimension of The joint input is passed through two one-dimensional causal convolutional layers (utilizing only past and current information), both with a kernel size of 3, a stride of 1, and a ReLU activation function. The first convolutional layer reduces the number of channels from... Mapped to 32, the second convolutional layer maps the number of channels from 32 to 1, outputting the initial prediction value for a single channel. .

[0090] To constrain the offset to a physically meaningful range, use an activation function with upper and lower bounds: ; because The range of values ​​is , Strictly bound to Within the interval. This design makes full use of the offset prior provided in step 3, which not only ensures the stability of training with small samples, but also preserves the flexibility of fine-tuning the offset for the network.

[0091] For source 1 itself, no offset prediction module is set, and its corresponding time index is always 1. .

[0092] (2) Deformable time sampling layer In standard cross-attention, source The keys and values ​​are taken from a fixed time step. This network employs deformable time sampling, based on the predicted offset. Features are extracted from non-integer time locations. Because... For continuous real numbers, linear interpolation is used to achieve differentiable sampling.

[0093] Source The feature in the time dimension is represented as For a given query time step Sampling location is . use The two nearest integer time indices and and its corresponding eigenvectors , The sampling features are obtained through linear interpolation: ; like Exceeding the valid time range The nearest valid boundary value is then used as the sampling result. This interpolation operation is fully differentiable and supports end-to-end gradient backpropagation.

[0094] (3) Cross-attention fusion module The aligned multi-source features are fused using multi-head cross-attention. (Based on entities...) At time step For example, the integration of: Query: Retrieves the current time step of the reference source (source 1). The features are used to generate a query vector through linear transformation. ,in It is a learnable matrix.

[0095] Key and Value: For each source The aligned feature vector is obtained using the aforementioned deformable temporal sampling layer. They are then transformed into key vectors using linear transformations. Sum value vector ,in It is a source-specific linear projection matrix.

[0096] Attention calculation: The similarity between the query and each source key is evaluated using scaled dot products, and the attention weights are obtained using softmax normalization. ; Fusion Output: Weighted aggregation of value vectors from each source: ; To ensure sufficient information retention, this module employs multi-head attention. There are 3 parallel attention heads, each with a dimension of 1. The results from each header are spliced ​​together and then output via a projection matrix. Obtain the final fusion vector .

[0097] Repeat the above operations along the timeline to obtain the entity. fusion feature matrix .

[0098] 3. Training Program Since multi-source fusion tasks often lack explicitly labeled ground truth values, this invention employs a self-supervised learning framework for training. The training set is directly derived from the entity feature matrices output in step 3, requiring no additional manual annotation.

[0099] (1) Training set construction 80% of the operational entities were randomly selected as training entities, and the remainder were used as validation and test entities. For each entity, its time series was divided into consecutive time periods: [The text abruptly ends here, so the translation stops as well.] Steps are used for training, and then The step is used for verification. This partitioning ensures the causality of the time dimension and the generalization of the entity dimension's distribution.

[0100] (2) Loss function design The training objective consists of three parts that work together to optimize the parameters of the offset prediction subnetwork and the cross-attention module.

[0101] Multi-source reconstruction loss: A high-quality fusion representation should be able to reconstruct the original aligned features of each source well from itself. Specifically, a reconstruction decoder (composed of a multilayer perceptron) symmetrical to the fusion module structure is used to reconstruct the original aligned features from the fusion vector. Mapping back to the feature space of each source, recovering the features of each source at the expected alignment position. The reconstruction loss is the mean squared error: ; This loss urges the fusion process to retain the main information from each source, while requiring that the features after offset sampling be consistent with the original semantics.

[0102] Future prediction loss: To enhance the ability of fused features to represent operational status, a single-layer gated recurrent unit predictor is added, based on the fused sequence. Predicting reference source features for the next time step The loss prediction uses mean squared error: ; This loss-driven fusion approach focuses on cross-source patterns that have predictive value for future operational changes.

[0103] Offset smoothing regularization: To prevent the offset from fluctuating drastically between adjacent time steps, a time smoothing regularization is introduced. ; The total loss is the weighted sum of the three: ; in , The balance coefficient is set to 0.1 by default.

[0104] (3) Training process and parameter settings The network is trained end-to-end using the Adam optimizer with an initial learning rate of The validation loss is halved if it doesn't decrease for 5 consecutive epochs, and a total of 50 epochs are trained with a batch size of 32 entities. The parameters of the offset prediction subnetwork are initialized using Xavier normal initialization, but the bias term is initialized to... (Right now This ensures that the initial offset output is the one provided in step 3. This accelerates convergence and avoids early misalignment.

[0105] After training, for all Individual entities throughout the entire time range to Perform forward reasoning to obtain each entity. fusion feature sequence This sequence occupies a unified timeline and dynamically compensates for the latency of each source, containing rich operational information after denoising, completion, and cross-source alignment.

[0106] Fusion features This information will be directly used as input to step 5 to calculate the strength of dynamic relationships between entities. Since the fused features have eliminated asynchronicity and noise interference, the entity associations extracted in step 5 will be more accurate, thereby improving the reliability of subsequent graph neural network inference. Simultaneously, the trained offset prediction network itself can also be used in step 6 to trace delay patterns, providing a reference for business adjustments.

[0107] In step 5, based on the fused feature sequence, an adjacency graph representing the strength of associations between entities is dynamically constructed at each time step. This adjacency graph, along with the fused feature sequence, is then input into a spatiotemporal prediction network consisting of multi-scale temporal encoding and graph convolutional gated recurrent units. The network outputs predicted operational indicators and risk scores for each entity at future time steps. Specifically: Step 5 builds upon the entity fusion feature sequences output from Step 4, completing two core tasks: first, dynamically capturing the implicit relationships between operational entities from the fusion features to construct a topology graph that evolves over time; second, based on this dynamic graph and node temporal features, predicting future key operational indicators and generating decision support information based on the prediction results. The output is the predicted indicator values ​​and risk level assessments for each entity at future time steps, directly supporting the anomaly tracing and visualization analysis in the subsequent Step 6.

[0108] The following section provides a detailed explanation from five aspects: input and problem modeling, dynamic graph construction mechanism, spatiotemporal graph convolution-gated recurrent unit network structure, training scheme, output and connection.

[0109] 1. Input and Problem Modeling Step 4 provides each operating entity to this step. fusion feature sequence The sequence has eliminated measurement noise, missing values, and temporal asynchrony from each data source, and contains low-dimensional embeddings that are cross-source aligned and represent the true operational state.

[0110] Entities do not exist in isolation; factors such as upstream and downstream supply chains, regional market influences, and capital linkages lead to close interrelationships in their operations. Traditional time series forecasting often models entities independently, neglecting this dynamic transmission effect between them. This step constructs a time-varying weighted directed graph. vertex set For all entities, edge set and adjacency matrix At each time step All of these change, characterizing the instantaneous influence relationships between entities. Based on this, a spatiotemporal graph neural network is used to mine spatiotemporal dependency patterns and make time series predictions, overcoming the information loss caused by the static graph assumption.

[0111] The prediction task is set as follows: given an observation window Using dynamic graphs and node features within the graph to predict the future Key operational metrics for step This metric can be obtained from the clean tensor in step 2. The selection of reference source feature dimensions reflects the core quantitative value of the company's operational health.

[0112] 2. Dynamic graph construction mechanism The graph structure is not directly imported from an external knowledge graph, but is adaptively generated from the fused feature sequence and updated independently at each time step. The specific construction process is as follows: (1) Entity feature enhancement in asynchronous perception Considering the potential time lag in the influence between entities, and given that the deformable attention in step 4 has already preliminarily aligned the multi-source temporal sequence, this step further embeds asynchronous propagation awareness at the entity feature level. For the time step... entity fuse them with feature vectors Rather than the past The features at each time step are subjected to one-dimensional convolutional temporal encoding to obtain the enhanced representation. The encoder uses dilated convolution, with dilation rates of... The kernel size is 3 for all entities to cover different hysteresis scales. The enhancement matrix for all entities at this time step is denoted as... .

[0113] (2) Self-attention map generation Based on the enhanced features, scaled dot product self-attention is used to calculate the original association score between entities: ; in The projection matrix is ​​learnable. For the attention dimension. Score matrix. It measures the correlation between entities at the current time step.

[0114] (3) Sparse adjacency matrix Fully connected graphs are computationally expensive and prone to weak correlation noise. Execute top- line by line Sparsification: Only retain the row with the highest score. Set one edge to 0, and the rest to 0. Then apply softmax normalization to make the sum of the weights of the retained edges equal to 1, thus obtaining the directed adjacency matrix. . Pick This significantly reduces the number of parameters and computational overhead of subsequent graph convolutions while maintaining graph connectivity.

[0115] The above dynamic graph construction is entirely driven by fused features and jointly optimized with downstream prediction tasks in an end-to-end manner, enabling the graph structure to automatically focus on the inter-entity interaction paths that are most explanatory to the prediction target.

[0116] 3. Spatiotemporal Graph Convolutional-Gated Recurrent Unit Network (ST-GCN-GRU) To simultaneously capture the spatial dependencies and temporal evolution of node sequences in a dynamic graph, a multi-scale spatiotemporal graph convolution-gated recurrent unit network was designed. This network is an improved coupling of graph convolution and recurrent units, and its specific structure is as follows.

[0117] (1) Multi-scale temporal convolution front end Before entering the recurrent unit, multi-scale convolution is performed on the input feature time window of each node to capture temporal patterns at different granularities. For each node... Length is input fragment It passes through three one-dimensional convolutional branches in parallel: Branch A: Core size 3, expansion rate 1, output channel ; Branch B: Core size 3, expansion rate 2, output channel ; Branch C: Core size 5, expansion rate 1, output channel .

[0118] Each branch is followed by batch normalization and ReLU activation, and causal padding is used to maintain the time length. The results of the three branches are concatenated along the channel dimension to obtain a result with dimension [dimensional value missing]. The features are then further processed by a 1×1 convolution (i.e., a one-dimensional convolution with a kernel size of 1) to increase the number of channels from... Compress to Forming multi-scale time features This front-end design can simultaneously capture short-term mutations, medium-term cycles, and long-term trends, providing rich temporal representations for subsequent spatiotemporal modeling.

[0119] (2) Graph Convolution-Gated Recurrent Unit (GC-GRU) Multi-scale temporal feature sequences are used as input to GC-GRU units. GC-GRU is a spatial generalization of the standard GRU, replacing matrix multiplications within all gates with graph convolution operations, thereby aggregating neighbor information when reading the current input and the previous hidden state. In the... layer( , (Default value is 2), the calculation process is as follows: ; in For the sigmoid function, This is element-wise multiplication. and Both are Chebyshev graph convolution operators. Specifically, given the input signal... and adjacency matrix Graph convolution is defined as: ; Let be the order of the Chebyshev polynomial, and take... That is, using the information of the second-order neighborhood of each node. For the reason Calculate the scaling graph Laplacian matrix. for Chebyshev polynomials These are learnable weights. The graph convolutions in each gate are independently parameterized, with both input and output dimensions being [missing information]. .

[0120] The GC-GRU layer utilizes a dynamic graph at each time step. Transmitting information makes the hidden state It includes both the historical memory of nodes and the influence of neighbors at the current moment. Different layers are stacked: first layer input. The first output of the temporal convolution front end Step features (dimensions) The input to subsequent layers is the hidden state of the previous layer. Finally, the output of the top layer at the last time step is taken. As a spatiotemporal code for each node.

[0121] This unit fixes the graph convolution kernel to a third-order Chebyshev polynomial, achieving efficient spatial local aggregation. At the same time, the combination of multi-scale temporal front-end and gated loop constitutes a complete characterization of spatiotemporal dependence.

[0122] (3) Time series forecasting and decision evaluation Spatiotemporal coding matrix Send in two mission heads: Prediction Head: A two-layer fully connected network, the first layer will... Dimension mapping to The second layer will have a dimension and the activation function is ReLU; Dimension mapping to Dimension (prediction step size), no activation function, directly outputs the future. Predicted values ​​of indicators for each step .

[0123] Decision evaluation head: A three-layer multilayer perceptron, with the number of hidden units as follows: and The activation function is ReLU, and the output layer consists of a single node activated by sigmoid, outputting a risk score. This represents the probability that an entity will experience operational anomalies in the future. The supervision label for this score is derived from the sparse anomaly tensor in step 2. Anomaly marker: If the absolute value of the sparse component of the corresponding reference source feature of an entity exceeds a preset threshold within the target time window (e.g., ... If the sample is an anomaly, it is marked as such.

[0124] The two heads share a spatiotemporal code and are trained together, enabling the spatiotemporal code to serve both prediction accuracy and risk identification.

[0125] 4. Training Program The training data source consists of all entity fusion feature sequences obtained in step 4 and the target index sequences extracted from the clean tensor in step 2. The timeline is divided into a training set (first 70% of time steps), a validation set (middle 15%), and a test set (last 15%) to ensure no future information leakage.

[0126] Samples are generated using a sliding window method: for each entity For each feasible starting point, take a time window. As input, the corresponding target is the future. Real indicator value (Feature dimensions are extracted from a clean tensor reference source). The total loss function consists of prediction loss, decision evaluation auxiliary loss, and graph regularization term.

[0127] Prediction loss: Smoothed L1 loss is used, which is insensitive to outliers. ; in For the training sample set, exist At that time Otherwise .

[0128] Decision evaluation loss: using binary cross-entropy loss. The monitoring signal is based on the sparse anomaly tensor from step 2. The generated exception label.

[0129] Graph regularization: Encourages sparsity of the adjacency matrix in the L1 sense and smoothness between adjacent time steps. ; This ensures smooth changes in graph structure and avoids drastic topological changes caused by noise at a single time step.

[0130] Total loss: ; Set them to 0.2 and 0.05 respectively.

[0131] Training uses the Adam optimizer with an initial learning rate of The performance decays by a factor of 0.5 every 20 epochs. The batch size is 64 entity fragments. Gradient clipping is used, with a maximum gradient norm of 5.0 to prevent numerical instability in the calculation of the graph Laplacian matrix. Prediction error is monitored on the validation set (smoothed L1), and early stopping is triggered if there is no improvement for 10 consecutive epochs. The maximum number of training epochs is 100.

[0132] After training, inference is performed across all timelines to obtain the future prediction sequence for each entity at each time step. and risk score These outputs are passed directly to step 6.

[0133] Step 6 will perform anomaly tracing based on prediction bias and risk score, and correlate the sparse anomaly tensor separated in Step 2. The original features that caused the anomaly are traced back, and the offsets learned in step 4 are used. Analyze the patterns of delays from various sources. Furthermore, risk scores and prediction curves can be presented to decision-makers through a visual interface to assist them in developing timely intervention measures.

[0134] In step 6, based on the comparison results of the predicted values ​​of operational indicators with preset thresholds and the risk score, tiered early warning information is generated. Combined with the anomaly patterns recorded in the sparse anomaly tensor, recommended adjustment strategies are output, and a visual decision support interface is generated, specifically: Step 6, as the final execution stage of the overall technical solution, transforms the analysis results of the preceding steps into early warning information and strategy suggestions for business users, presented through a visual interface. Its inputs come from the prediction and evaluation output of Step 5, as well as the anomaly records and offset information accumulated in Steps 2 and 4. The outputs are actionable early warning signals, a list of adjustment strategies, and interactive chart data, directly serving operational monitoring and decision-making intervention. The specific process is as follows: 1. Step 6 receives the following data: Forecast and Risk Output (from Step 5): Each Operating Entity In the future Indicator prediction sequence at each time step and the current risk score. .

[0135] Historical anomalous components (from step 2): sparse anomalous tensors separated in step 2 Records of various data sources Outliers or structural breaks that occur in history and in the present moment.

[0136] Cross-source time offset records (from step 4): Sources learned in step 4 offset time series relative to reference source This reflects the delayed changes in each data source at the entity level.

[0137] Raw data and fused features (from steps 1 and 3): used for front-end visualization backfilling to ensure the traceability of displayed information.

[0138] 2. Early warning information generation The core basis for early warning is to compare the predicted indicators with preset thresholds and make a comprehensive judgment by combining risk scores and abnormal records.

[0139] (1) Threshold setting The thresholds are set based on business experience, while also referencing the results of the normal fluctuation range analysis in step 2. For key operational indicators, the clean tensor from step 2 is used as a basis. The historical mean and standard deviation of this indicator are used to set three threshold levels: Attention threshold The indicator deviates from its historical mean by 0.5 standard deviations. Warning threshold Deviation from 1 standard deviation; Severity threshold If the deviation is more than 2 standard deviations, or if the system is in a warning state for 3 consecutive time steps.

[0140] The threshold can be adjusted according to actual business needs, and the historical fluctuation levels of each entity obtained in step 2 can be fine-tuned to eliminate false alarms caused by individual differences.

[0141] (2) Early warning judgment rules For each entity The predicted sequence is evaluated step-by-step: like or (Both sides) Mark as attention and issue low priority prompts; like or Mark it as a warning and generate a medium-priority alarm; like If more than 3 consecutive warnings are issued, the alert will be marked as severe and a high-priority alarm will be triggered.

[0142] Risk scoring is also introduced. As a correction: if Furthermore, if the prediction fails to reach the warning threshold, the alarm level will be automatically upgraded by one level; if Furthermore, if the level only reaches the level of concern, it can be downgraded to ignore. This mechanism reduces false negatives and false positives caused by hard threshold decisions.

[0143] The early warning information is recorded in a structured manner, including entity number, trigger time, predicted value, the extent to which the threshold is exceeded, current risk score, and alarm level, for subsequent push notifications.

[0144] 3. Recommended adjustment strategy generation Based on the early warning information, the system automatically searches the knowledge base and outputs targeted adjustment suggestions. The knowledge base includes a predefined set of business rules, such as: If the warning originates from signs of supply chain disruption (step 2 reveals that the entity has an abnormally sparse component in its procurement source), and the offset shows that the delay in logistics sensor data continues to increase, it is recommended to check the delivery status of upstream suppliers and consider the strategy of activating alternative supply channels. If the warning source is a decline in market demand (predicted sales will continue to fall), and the increased deviation of the public opinion source suggests that the information is lagging, it is recommended to shorten the market monitoring cycle and launch promotional activities. If an early warning is accompanied by a high risk score and there is no clear single source of anomaly, the comprehensive risk response plan will be triggered. It is recommended that management review the operational process and strengthen the verification of multi-source information.

[0145] Strategy matching uses a rule engine: based on the entity's industry and abnormal feature tags (from... The algorithm combines the parsing of data with the time offset trend to retrieve the corresponding strategy template. The strategy output is a list, sorted by priority. Each strategy includes an action suggestion, related data source, and a brief description of the expected impact.

[0146] 4. Visual Interface Construction To support rapid understanding and decision-making, step 6 constructs various types of visualizations, with the data provided to the front-end for rendering via a back-end interface.

[0147] Warning Overview Dashboard: Using the geographical distribution of entities or an organizational tree as a background, each entity's current highest alert level is indicated by color (green / yellow / orange / red). Clicking on an entity allows you to drill down to view detailed prediction curves, historical risk scores, and anomaly markers for each source.

[0148] Predictive Comparison Curve: Plot the actual values ​​of historical indicators of the entity, the predicted values ​​from step 5, and the warning threshold line on the same time axis, and mark the future prediction range with shaded areas to intuitively present the trend and risk window.

[0149] Cross-source delay heatmap: Displays the offset of each entity from each source. The magnitude and stability of latency are represented by varying color intensity over time. Source-entity combinations with abnormally large latency are highlighted to remind users to check the data acquisition process for that source.

[0150] Strategy Recommendation Panel: Displays recommended actions for the selected entity in card format. Each card includes a brief explanation of the reason and a link to relevant evidence of the anomaly. Users can confirm the action or mark it as processed.

[0151] The visualization components work together, allowing users to click on alarm entities on the dashboard to simultaneously update drill-down charts, forming an analytical loop.

[0152] The final output of step 6 includes: Warning message queue: Pushes alerts to relevant personnel's instant messaging or email systems according to the alarm level; Strategy Recommendation Report: A list of adjustment strategies generated daily or on demand, which can be directly exported as a document; Visualization configuration file: includes chart data and layout definitions, and is rendered by the front end as an interactive interface.

[0153] The processing results after the early warning is confirmed can be fed back to the system to adjust the focus of the prediction model in step 5, or as empirical data for updating the anomaly detection threshold in step 2, so as to achieve continuous optimization.

[0154] This invention also provides an operational decision support analysis system based on multi-source heterogeneous big data fusion, applied to the aforementioned operational decision support analysis method based on multi-source heterogeneous big data fusion. This system comprises a data acquisition module, a data preprocessing module, a tensor anomaly detection and feature extraction module, a deformable cross-attention fusion module, a dynamic graph construction and prediction module, and a decision support output module connected sequentially. The output of the data acquisition module is connected to the input of the data preprocessing module; the output of the data preprocessing module is connected to the input of the tensor anomaly detection and feature extraction module; the output of the tensor anomaly detection and feature extraction module is connected to the input of the deformable cross-attention fusion module; the output of the deformable cross-attention fusion module is connected to the input of the dynamic graph construction and prediction module; and the output of the dynamic graph construction and prediction module is connected to the input of the decision support output module. Furthermore, each module is connected to a shared distributed storage system to exchange intermediate data, model parameters, and raw records required for visualization backfilling. The following is a detailed description of each module: 1. Data Acquisition Module The data acquisition module is used to acquire and time-series align multi-source heterogeneous operational data as described in step 1. This module consists of multiple data access adapters and a time alignment unit. The data access adapters can be IoT message queue clients (such as MQTT, Kafka Consumer), web crawler frameworks (such as Scrapy), enterprise relational database connectors (such as JDBC / ODBC), or file system monitoring tools, depending on the data source type. Each adapter converts raw data of different frequencies and formats into a unified time-series structure and timestamps it. The time alignment unit can use a time-series database engine that supports sliding window aggregation (such as InfluxDB's continuous query) or a custom time window resampling program, running on a Linux server or edge gateway device, to align the data from each source to a unified time grid, generating the data tensor output in step 1. and observation mask tensor .

[0155] 2. Data Preprocessing Module The data preprocessing module corresponds to the cleaning, denoising, and normalization processes in step 1. It can utilize a server-side in-memory computing environment integrated with data processing libraries such as Pandas and NumPy, or a distributed computing framework such as Apache Spark to process industrial-grade massive datasets. This module's functions include missing value interpolation, outlier detection and labeling based on interquartile ranges, source-specific normalization, and text embedding. Ultimately, it constructs a preprocessed multi-source tensor suitable for tensor decomposition and its corresponding missing mask, then passes the data to subsequent modules.

[0156] 3. Tensor Anomaly Detection and Feature Extraction Module This module corresponds to robust tensor decomposition and anomaly separation in step 2, and compressed feature extraction in step 3. Physically, this module is implemented using a computing unit consisting of one or more graphics processing units (GPUs, such as NVIDIA A100) or dedicated tensor computation chips (such as TPUs), accelerating tensor operations through the CUDA or ROCm framework. Module functions include: Online adaptive weight adjustment for low-rank extraction and sparse anomaly decomposition is performed to separate the clean tensor output from step 2. and sparse anomaly tensors ; Perform Tucker decomposition and feature mode truncation as described in step 3 on each source clean tensor to generate compressed features for each entity. and cross-source time offset prior .

[0157] This module is also responsible for storing the decomposed sparse patterns into an anomaly database for use by the decision support output module.

[0158] 4. Deformable cross-attention fusion module This module implements the functionality of step 4 and mainly consists of a neural network inference processor (such as a GPU or FPGA) configured with efficient linear interpolation units and cross-attention computation kernels. Internally, the module loads pre-trained deformable cross-attention network parameters, including an offset prediction subnetwork (implemented by two layers of one-dimensional causal convolutions with a kernel size of 3) and a multi-head cross-attention component. During runtime, this module receives compressed features. and offset prior The process sequentially performs offset fine-tuning, deformable temporal sampling based on linear interpolation, and multi-head attention fusion to output an entity fusion feature sequence. It also outputs the final offset time series of each source. For subsequent visualization use.

[0159] 5. Dynamic graph construction and prediction module This module implements the dynamic graph construction and temporal decision analysis in step 5 and is the core computational module of the system. It consists of a multi-scale temporal convolutional front-end (including three parallel one-dimensional convolutional branches with kernel sizes of 3 / dilation rate 1, 3 / dilation rate 2, and 5 / dilation rate 1, followed by 1×1 convolutional compression), stacked GC-GRU layers, and a prediction-risk assessment dual-head network. The module is deployed on a GPU server with model parallelism support and utilizes graph neural network operators (such as PyG and DGL) from deep learning frameworks (such as PyTorch and TensorFlow) for computation. The specific workflow is as follows: constructing a self-attention dynamic graph adjacency matrix online from fused features. By aggregating neighbor node information through Chebyshev graph convolution operations and performing gated loop updates, the final output is a multi-step prediction index for the future. and risk score This module is also responsible for updating network weights using historical data during the training phase and supporting real-time incremental prediction during the inference phase.

[0160] 6. Decision Support Output Module This module corresponds to step 6 and mainly consists of three parts: a rules engine server, a message push middleware, and a front-end visualization service. The rules engine can use an interpreter based on Drools or a self-developed rules script, running on the application server. Based on preset thresholds, risk scores, and sparse patterns from the tensor anomaly detection and feature extraction module, it generates alarm levels and a list of recommended strategies. The message push middleware sends alerts to relevant personnel via email, SMS, or enterprise office software interfaces. The visualization server, based on Web technologies (such as a Node.js / Java backend and a Vue.js frontend), renders prediction curves, heatmaps, and strategy panels as interactive charts and provides them to the management terminal via HTTP / WebSocket protocol, enabling users to complete the closed loop from data perception to decision-making.

[0161] The present invention also provides an embodiment that provides a detailed description of the above-described system and method, specifically including: Example: To verify the effectiveness of the technical solution proposed in this invention, an experiment was conducted using a real operational dataset from a chain retail enterprise. This dataset covers a 365-day time span, involves 10 store entities, and includes four types of data sources: internal operational data (12 fields including sales, inventory, and costs), external market data (5 fields including competitor prices and industry indices), supply chain sensor data (8 channels including temperature, humidity, and vibration), and user feedback text data (online reviews and social media, embedded into 768-dimensional features by a pre-trained model). All data sources contain varying degrees of random missing data (missing rate approximately 8%–15%) and manually injected anonymized outliers (approximately 3%). The experiment used the first 70% of days as the training set, the middle 15% as the validation set, and the last 15% as the test set, with a future prediction step size of H = 7 days.

[0162] In the experiment, the method of the present invention was compared with the following three baselines: Baseline 1: Independent time series forecasting, each entity and indicator is modeled independently using the ARIMA model, without using multi-source fusion and graph structure.

[0163] Baseline 2: Static graph fusion method, which uses a fixed fully connected graph instead of a dynamic graph construction module, and the rest is the same as the present invention.

[0164] Baseline 3: Standard robust tensor decomposition method, replacing the adaptive weights in step 2 with fixed equal weights, while keeping the rest unchanged.

[0165] The evaluation indicators used were the root mean square error (RMSE) and mean absolute percentage error (MAPE) of the predicted operating indicators for the next 7 days, as well as the AUC value for risk warning. The experimental results are shown in Table 1.

[0166] Table 1. Performance comparison of different methods on the test set As shown in Table 1, the complete method of this invention significantly outperforms all baselines in both prediction accuracy and risk identification capability. Specifically, compared to fixed equal-weight tensor decomposition, the adaptive weighting mechanism reduces RMSE by approximately 31.6%, validating the effectiveness of the innovation in step 2; compared to static graph fusion, dynamic graph construction reduces RMSE by approximately 44.6%, validating the crucial role of the dynamic graph mechanism in step 5 in capturing the evolutionary relationships between entities. The risk AUC reaches 0.92, indicating that the risk assessment head in step 5 can reliably provide early warnings of operational anomalies.

[0167] To further visualize the prediction results, Figure 2 The paper presents a comparison between the projected sales curve and the actual value for a typical store over the next 7 days. The projected value closely follows the trend of the actual value, and it can still effectively capture the actual trend during periods of large fluctuations, demonstrating the synergistic advantages of multi-source fusion and spatiotemporal modeling.

[0168] Figure 3 The estimated time delay of each data source relative to the reference source is displayed as a heatmap over time. As shown in the graph, market data sources exhibit a lag of approximately 1-3 days for most periods, with this lag widening during promotional seasons (days 100-130); sensor data delays are relatively stable. The offsets learned in step 4 effectively reveal these asynchronous patterns, providing a basis for optimizing the data acquisition process.

[0169] Embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0170] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0171] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0172] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0173] Contents not described in detail in this specification are prior art known to those skilled in the art. It is hereby indicated that the above description is intended to help those skilled in the art understand this invention, but does not limit the scope of protection of this invention. Any equivalent substitutions, modifications, improvements, or simplifications of the above descriptions that do not depart from the essential content of this invention fall within the scope of protection of this invention.

Claims

1. An operational decision support analysis method based on the fusion of multi-source heterogeneous big data, characterized in that, include: Step 1: Obtain operation-related time-series data from multiple data sources, perform time resampling, missing value marking, anomaly candidate marking and normalization on the operation-related time-series data, and construct a data tensor and corresponding observation mask tensor with unified time granularity according to the time dimension, feature dimension and operation entity dimension; Step 2: For each data tensor from a data source, at the effective observation position defined by the observation mask tensor, make the sum of the clean tensor and the sparse outlier tensor match the data tensor at the effective observation position, and separate the data tensor into a clean tensor and a sparse outlier tensor by minimizing the weighted nuclear norm and sparsity constraints. Step 3: Perform feature compression on the clean tensors of each data source, extract compressed features with uniform dimensions and retain entity information, and use the cross-correlation of multi-source time series to estimate the time offset prior of each data source relative to the selected reference source. Step 4: Using compressed features and time-shifted priors as input, multi-source feature fusion is performed through a deformable cross-attention network to output a fused feature sequence at the entity level; Step 5: Based on the fused feature sequence, dynamically construct an adjacency graph representing the strength of association between entities at each time step, and input the adjacency graph and the fused feature sequence together into a spatiotemporal prediction network consisting of multi-scale temporal coding and graph convolutional gated recurrent units, and output the predicted values ​​of operational indicators and risk scores of each entity in future time steps. Step 6: Based on the comparison results of the predicted values ​​of operational indicators with the preset thresholds and the risk score, generate graded early warning information, combine the abnormal patterns recorded in the sparse anomaly tensor to output recommended adjustment strategies, and generate a visual decision support interface.

2. The operational decision support analysis method based on multi-source heterogeneous big data fusion according to claim 1, characterized in that, Step 2, specifically includes the adaptive robust tensor decomposition: An optimization objective is constructed by combining the weighted sum of the nuclear norms of the expansion matrices of each module with the L1 norm of the sparse outlier tensor. Under the constraint of the observation mask, the alternating direction multiplier method is used to solve the problem. In each iteration, the nuclear norm weight of each module is dynamically updated based on the spectral norm of the expansion matrices of the clean tensor obtained in the previous iteration. This results in modules with larger spectral norms receiving smaller weights to retain information, and modules with smaller spectral norms receiving larger weights to suppress noise.

3. The operational decision support analysis method based on multi-source heterogeneous big data fusion according to claim 2, characterized in that, The nuclear norm weight of the nth tensor in the rth iteration Determine as follows: ； In the formula, The matrix is ​​obtained by expanding the clean tensor from the previous iteration along the nth tensor modulus. It is a constant greater than zero.

4. The operational decision support analysis method based on multi-source heterogeneous big data fusion according to claim 1, characterized in that, In step 3, the feature compression specifically includes: For each clean tensor from a data source, perform high-order singular value decomposition, truncate the factor matrix corresponding to the feature dimension, retain the column vectors that make the singular value cumulative energy ratio reach a preset ratio, and then use the truncated factor matrix to project the clean tensor along the feature dimension to a low-dimensional core tensor. The core tensor or its feature matrix after being split by entity is used as the compressed feature.

5. The operational decision support analysis method based on multi-source heterogeneous big data fusion according to claim 4, characterized in that, In step 4, the deformable cross-attention network generates continuous learnable offsets for each time step through the offset prediction sub-network, determines the sampling time position of the data source to be aligned based on the continuous learnable offsets, obtains the features at the sampling time position using linear interpolation, and then outputs the entity-level fused feature sequence using a multi-head cross-attention mechanism.

6. The operational decision support analysis method based on multi-source heterogeneous big data fusion according to claim 5, characterized in that, The offset prediction subnetwork consists of a one-dimensional causal convolutional layer. It receives the joint input of the reference source features and the source features to be aligned concatenated along the channel, outputs the initial offset estimate, and then obtains the learnable offset within the constraint range by mapping the activation function with the time offset prior as the center and the preset relaxation range as the boundary.

7. The operational decision support analysis method based on multi-source heterogeneous big data fusion according to claim 5, characterized in that, In step 5, an adjacency graph representing the strength of associations between entities is dynamically constructed, specifically as follows: Temporal context-enhanced encoding is performed on the fused features at each time step. The association score matrix between entities is calculated using a self-attention mechanism. The top k largest connections are retained based on the score value of each row and normalized to obtain a directed weighted adjacency matrix, where the value of k is logarithmically related to the total number of entities.

8. The operational decision support analysis method based on multi-source heterogeneous big data fusion according to claim 7, characterized in that, In step 5, the spatiotemporal prediction network includes: a multi-scale temporal convolutional front-end, a graph convolutional gated recurrent unit, a prediction head, and a risk assessment head. The multi-scale temporal convolutional front-end has multiple parallel one-dimensional convolutional branches with different dilation rates and kernel sizes to capture temporal patterns of different granularities. The outputs of each branch are fused and input into the stacked graph convolutional gated recurrent unit. The graph convolutional gated recurrent unit integrates neighbor information in the adjacency graph into the node hidden state update through Chebyshev graph convolution. Finally, the prediction head outputs the predicted values ​​of future multi-step operational indicators, and the risk assessment head outputs the risk score.

9. The operational decision support analysis method based on multi-source heterogeneous big data fusion according to claim 8, characterized in that, In step 6, tiered early warning information is generated, specifically as follows: Multiple threshold levels are set based on the historical statistics of the target indicators in the clean tensor. The predicted value is compared with the threshold to obtain the initial alarm level. The alarm level is then upgraded or downgraded based on the risk score. The recommended adjustment strategy is matched and output by the rule engine based on the change trend of the abnormal feature labels and offsets extracted from the sparse abnormal tensor and the preset knowledge base.

10. An operational decision support and analysis system based on the fusion of multi-source heterogeneous big data, applied to the operational decision support and analysis method based on the fusion of multi-source heterogeneous big data as described in any one of claims 1-9, characterized in that, include: The data acquisition module, data preprocessing module, tensor anomaly detection and feature extraction module, deformable cross-attention fusion module, dynamic graph construction and prediction module, and decision support output module are connected in sequence. The data acquisition module is used to obtain raw operational data from multiple data sources and record timestamps and entity identifiers; The data preprocessing module is used to perform time resampling, missing data marking, anomaly candidate marking, normalization and tensor construction on the original operational data to generate data tensors and observation mask tensors. The tensor anomaly detection and feature extraction module is used to perform robust tensor decomposition based on adaptive weight adjustment, separate clean tensors and sparse anomaly tensors, and perform feature compression based on Tucker decomposition on clean tensors to generate compressed features and cross-source time offset priors. The deformable cross-attention fusion module is used to generate a learnable offset constrained by the time offset prior through the offset prediction sub-network, perform deformable time sampling and multi-head cross-attention fusion, and output a fusion feature sequence at the entity granularity. The dynamic graph construction and prediction module is used to dynamically construct a directed adjacency graph based on the fused feature sequence, and to perform spatiotemporal prediction using a multi-scale temporal convolutional front-end and stacked graph convolutional gated recurrent units, outputting predicted values ​​of operational indicators and risk scores. The decision support output module is used to generate graded early warnings based on the comparison results of predicted values ​​and preset thresholds and risk scores, and to combine the abnormal patterns and offset trends in sparse anomaly tensors to match recommendation strategies and generate a visual decision support interface.

Images