A data island prediction method, medium and system based on asset big data

Abstract

The application provides a data island prediction method, medium and system based on asset big data, belongs to the technical field of the data island prediction method, medium and system based on asset big data, and through a data integration engine, the application connects various heterogeneous data islands and constructs a multi-island heterogeneous knowledge graph, utilizes a multi-island optimal feature compression algorithm based on an information bottleneck principle to independently compress features in each island, and then performs product space fusion, triggers hierarchical resource isolation scheduling and incremental flow calculation to guarantee prediction real-time performance, and through a fault risk score adjustment function, dynamically adjusts the number of graph attention heads to output the final fault prediction result, thereby solving the technical problem that the equipment fault features of cross-heterogeneous data islands cannot be jointly compressed and cross-island physical coupling state representation without sharing original data.

Description

Technical Field

[0001] This invention belongs to the technical field of data silo prediction methods, media and systems based on asset big data. Specifically, it relates to a data silo prediction method, media and system based on asset big data. Background Technology

[0002] In the field of industrial asset management, equipment health status prediction relies on the comprehensive analysis of sensor time-series data distributed across various business subsystems. Traditional methods typically aggregate data from various business systems into a unified data warehouse, then use support vector machines, random forests, or shallow neural networks to extract features and construct fault classification models, or leverage cross-table lookups in relational databases to achieve multi-source data fusion. However, with the expansion of the Industrial Internet of Things (IIoT), data silos have formed between various business systems. Equipment sensor data is scattered across heterogeneous platforms such as production execution systems, energy management systems, and equipment management systems, which differ fundamentally in data patterns, sampling frequencies, and semantic definitions. Current industrial predictive maintenance scenarios often employ federated learning frameworks to train shared models without moving the original data, or use graph neural networks to model equipment topology. However, neither of these methods simultaneously addresses the collaborative problem of compressing high-dimensional redundant features across heterogeneous silos and modeling physical coupling relationships across silos. This leads to the curse of dimensionality when processing joint features from multiple silos, and the inability to capture the true physical causal links between silos while ensuring data remains local, ultimately severely limiting the accuracy and real-time performance of fault prediction. In other words, existing technologies have a technical problem that the characteristics of device faults across heterogeneous data silos cannot be jointly compressed and physically coupled across silos without sharing the original data. Summary of the Invention

[0003] In view of this, the present invention provides a method, medium and system for predicting data silos in asset big data, which can solve the technical problem in the prior art that the fault characteristics of equipment across heterogeneous data silos cannot be jointly compressed and physically coupled across silos without sharing the original data.

[0004] The present invention is implemented as follows: The first aspect of the present invention provides a method for predicting data silos in asset big data, comprising the following steps:

[0005] The data integration engine connects various heterogeneous data silos, automatically extracts metadata and time-series data from each silo and establishes hierarchical relationships between data source nodes, data table nodes and physical subsystem nodes through the metadata registry center, forming a multi-island heterogeneous knowledge graph.

[0006] For the original features of each node in the multi-island heterogeneous knowledge graph, the multi-island optimal feature compression algorithm based on the information bottleneck principle is adopted. After compression by each island's independent encoder, the effective feature dimension is compressed to 2% to 5% of the original feature dimension through product space fusion, thus obtaining the compressed feature set of each island.

[0007] The compressed feature sets of each island are input into the cross-island state representation model. After processing through three layers—local aggregation layer within an island, physical coupling aggregation layer across islands, and global readout layer—the device health status representation vector and confidence level are output. When the confidence level is lower than the confidence level threshold, the graph structure of the multi-island heterogeneous knowledge graph is updated and attention propagation is re-executed. The process is repeated for a maximum of 5 rounds until the device health status representation vector converges.

[0008] The equipment fault risk score is calculated based on the cosine similarity between the equipment health status representation vector and the preset fault category standard vector set. When the equipment fault risk score exceeds the fault warning threshold, the hierarchical resource isolation scheduler is triggered to allocate a dedicated computing quota to the current prediction task and activate the incremental streaming computing mode to replace the full batch processing, keeping the prediction latency within 500ms.

[0009] When a new type of equipment or a new type of fault is detected, and the number of labeled samples in the corresponding island is lower than the small sample trigger threshold, the cross-island state representation model executes no more than 5 inner loop gradient steps, updates the task-specific adaptation layer parameters, and outputs the updated equipment fault risk score.

[0010] Based on three parameters—the average equipment failure risk score, the normalized value of the convergence round of the state representation iteration, and the utilization rate of the compressed feature dimension—a comprehensive adjustment value is calculated using a failure risk score adjustment function. The number of attention heads in the cross-island state representation model is dynamically adjusted according to the interval to which the comprehensive adjustment value belongs. The final failure prediction result is then output and pushed to the business system via the data service bus.

[0011] Specifically, the multi-island heterogeneous knowledge graph refers to organizing device sensor nodes, data table nodes, and physical subsystem nodes belonging to different business systems into a unified heterogeneous graph structure according to hierarchical relationships. The node features are composed of statistical quantities of the corresponding device sensor time series data, and the edge types include three categories: belonging to, referencing, and converted from.

[0012] Specifically, the multi-island optimal feature compression algorithm based on the information bottleneck principle refers to formalizing the multi-island joint feature extraction problem into an information bottleneck optimization problem. Under the constraint of maximizing the mutual information between compressed features and fault labels, it minimizes the mutual information between compressed features and original features, using Lagrange multipliers. To balance compression ratio and prediction correlation, the mutual information estimation is transformed into a differentiable variational lower bound using the variational information bottleneck method, and the optimal parameters of each isolated encoder are solved by gradient descent.

[0013] Specifically, the product space fusion refers to concatenating the compressed feature vectors output by each isolated encoder along the feature dimension to form a joint compressed feature representation. The compressed feature components of each isolated encoder remain separable in position after concatenation, which is used for subsequent separation and quantification of the information contribution of each isolated encoder.

[0014] The cross-island state representation model integrates two parts: a cross-island heterogeneous graph attention network and a cross-island federated meta-learning fast adaptation network. The cross-island heterogeneous graph attention network consists of three layers: a local aggregation layer within an island, a cross-island physical coupling aggregation layer, and a global readout layer. The device health status representation vector output by the global readout layer is directly used as the input sequence frame of the outer long short-term memory network in the cross-island federated meta-learning fast adaptation network.

[0015] Specifically, the cross-island physical coupling aggregation layer refers to encoding the cross-island physical coupling relationship as graph sparsification constraint edges based on the heat transfer equation and electromagnetic equation prior constraints encoded in the physical mechanism knowledge graph. The graph sparsification algorithm is used to automatically identify feature subsets that are highly related to the state of the target device, and graph attention aggregation is performed on the sparsified coupling graph to capture cross-island cascade physical effects.

[0016] The cross-island federated meta-learning fast adaptation network adopts a meta-learner-driven two-layer long short-term memory network structure. The outer long short-term memory network encodes cross-task general temporal representations, while the inner long short-term memory network is a task-specific adaptation layer that stores fast weight parameters. During federated training, each island locally calculates the meta-gradient of the task-specific adaptation layer, and the federated server aggregates the meta-gradients of each island and updates the global meta-initialization parameters.

[0017] Specifically, the calculation of the equipment failure risk score involves calculating the cosine similarity between the equipment health status representation vector and each standard vector in the preset failure category standard vector set, taking the maximum cosine similarity value and mapping it to the interval between 0 and 1 through a linear transformation to obtain the equipment failure risk score.

[0018] Specifically, the hierarchical resource isolation scheduler divides computing resources into multiple isolated quota pools based on the type of prediction task. Fault warning queries prioritize the use of the dedicated quota pool for fault warnings, while other queries use the remaining quota pool. Resource isolation avoids interference from concurrent queries on the real-time performance of fault warnings.

[0019] Specifically, the incremental streaming computing mode refers to triggering the calculation and updating of equipment failure risk scores one by one or in small batches according to the order of data arrival, rather than waiting for all data to arrive before batch processing, thereby controlling the end-to-end prediction latency to within 500ms.

[0020] Specifically, the small sample trigger threshold refers to the number of samples at which the F1 value no longer significantly increases with the increase of the number of samples after rapid adaptation, by gradually increasing a small number of real labeled samples in the interval [5, 50] with a step size of 5 in the synthetic sample expansion experiment.

[0021] Specifically, the fault risk score adjustment function refers to the function based on the average equipment fault risk score. Normalized value of the convergence iteration of state representation and the utilization rate of compressed feature dimensions The weighted sum of the three parameters yields the comprehensive adjustment value. The formula is The weighting coefficient , , satisfy .

[0022] Specifically, the parameter for dynamically adjusting the number of attention points in the graph based on the interval to which the comprehensive adjustment value belongs refers to when... Time-based attention head count Adjusted to 4; when hour Adjusted to 8; when hour Adjusted to 12; when hour Adjusted to 16.

[0023] Specifically, the establishment of the training dataset for the cross-island state representation model refers to collecting the time-series records of sensors of each island device from the historical database of each business system, dividing them into sample units according to the device number and time window, using a conditional generative adversarial network to synthesize scarce fault category samples under physical constraints, ensuring that the proportion of labeled samples for each fault category is not less than 1%, dividing the training dataset according to the boundaries of each island, with each island independently holding its own local training dataset, and dividing the training set, validation set, and test set according to 80%, 10%, and 10%, respectively.

[0024] The training of the cross-island state representation model is carried out in three stages: the first stage uses unlabeled historical data from each island to train the cross-island heterogeneous graph attention network using a contrastive self-supervised pre-training method; the second stage trains a two-layer long short-term memory network using a federated meta-learning method, with the federated aggregation rounds determined by the early stopping criterion of global validation loss; the third stage performs small-sample fine-tuning on labeled sample units, fixes the slow weights of the cross-island heterogeneous graph attention network, and only updates the fast weights of the task-specific adaptation layer, with the fine-tuning steps not exceeding 5 steps.

[0025] A second aspect of the present invention provides a computer-readable storage medium storing program instructions that, when executed in a computer, are used to perform the aforementioned method for predicting data silos in asset big data.

[0026] A third aspect of the present invention provides a data silo prediction system based on asset big data, comprising the aforementioned computer-readable storage medium, wherein the system is a computer, the computer-readable storage medium is disposed within the system, and the system is provided with a microprocessor for executing program instructions stored in the computer-readable storage medium.

[0027] This invention unifies the hierarchical association structure of metadata and sensor time-series features across multiple isolated data islands by constructing a heterogeneous knowledge graph. It utilizes a multi-island optimal feature compression algorithm based on the information bottleneck principle to independently compress features locally on each island, then fuses them through a product space to form a joint compressed feature. This is further processed through a three-layer structure of local aggregation within each island, cross-island physical coupling aggregation, and global readout to extract the device health status representation vector. This solves the problem that effective joint feature compression cannot be achieved when the original data of each island remains local. Furthermore, this invention embeds prior constraints from the physical mechanism knowledge graph into a graph sparsification algorithm, ensuring that the cross-island physical coupling aggregation layer propagates information only along coupling paths with physical causal significance. This avoids interference from spurious statistical correlations on state representation, solving the problem that cross-island physical coupling relationships cannot be effectively modeled without data sharing. In summary, this invention solves the technical problem mentioned in the background art that device fault features across heterogeneous data islands cannot achieve joint compression and cross-island physical coupling state representation without sharing the original data. Attached Figure Description

[0028] Figure 1 This is a flowchart of the method of the present invention.

[0029] Figure 2 A graph showing the relationship between the utilization rate of each island's compressed feature dimension and the distribution of equipment failure risk scores.

[0030] Figure 3 This is a weight distribution diagram of sparse coupling path nodes in the cross-island physical coupling aggregation layer. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0032] like Figure 1 The diagram shown is a flowchart of a data silo prediction method based on asset big data provided by the first aspect of the present invention. This method includes the following steps:

[0033] S01. Connect various heterogeneous data islands through the data integration engine, automatically extract metadata and device sensor time series data from each island, and establish hierarchical relationships between data source nodes, data table nodes and physical subsystem nodes through the metadata registration center to form a multi-island heterogeneous knowledge graph.

[0034] S02. For the original features of each node in the multi-island heterogeneous knowledge graph, the multi-island optimal feature compression algorithm based on the information bottleneck principle is adopted. After compression by each island's independent encoder, the effective feature dimension is compressed to 2% to 5% of the original feature dimension through product space fusion, and the compressed feature set of each island is obtained.

[0035] S03. Input the compressed feature sets of each island into the cross-island state representation model. After processing through three layers—local aggregation layer within the island, physical coupling aggregation layer across the island, and global readout layer—output the device health status representation vector and confidence level. When the confidence level is lower than the confidence level threshold, update the graph structure of the multi-island heterogeneous knowledge graph and re-execute attention propagation. Iterate for a maximum of 5 rounds until the device health status representation vector converges.

[0036] S04. Calculate the equipment fault risk score based on the cosine similarity between the equipment health status representation vector and the preset fault category standard vector set. When the equipment fault risk score exceeds the fault warning threshold, trigger the hierarchical resource isolation scheduler to allocate a dedicated computing quota to the current prediction task and activate the incremental streaming computing mode to replace the full batch processing, keeping the prediction latency within 500ms.

[0037] S05. When a new type of equipment or a new type of fault is detected, and the number of labeled samples in the corresponding island is lower than the small sample trigger threshold, the cross-island state representation model executes no more than 5 inner loop gradient steps, updates the task-specific adaptation layer parameters, and outputs the updated equipment fault risk score.

[0038] S06. Based on three parameters—the average equipment failure risk score, the normalized value of the convergence round of the state representation iteration, and the utilization rate of the compressed feature dimension—a comprehensive adjustment value is calculated using the failure risk score adjustment function. The number of attention heads in the cross-island state representation model is dynamically adjusted according to the interval to which the comprehensive adjustment value belongs. The final failure prediction result is then output and pushed to the business system via the data service bus.

[0039] The multi-island heterogeneous knowledge graph refers to organizing device sensor nodes, data table nodes, and physical subsystem nodes belonging to different business systems into a unified heterogeneous graph structure according to hierarchical relationships. The node features are composed of statistical quantities of the corresponding device sensor time series data. The edge types include three categories: belonging, referencing, and transformed from. The multi-island heterogeneous knowledge graph serves as a unified input carrier for the multi-island optimal feature compression algorithm based on the information bottleneck principle and the cross-island state representation model.

[0040] The principle of the multi-island optimal feature compression algorithm based on the information bottleneck principle is as follows: The multi-island joint feature extraction problem is formalized into an information bottleneck optimization problem. Under the constraint of maximizing the mutual information between compressed features and fault labels, the mutual information between compressed features and original features is minimized, using Lagrange multipliers... Balancing compression ratio and prediction correlation; utilizing the variational information bottleneck method to transform mutual information estimation into a differentiable variational lower bound, and solving for the optimal parameters of each isolated encoder through gradient descent; compressing different isolated data using each isolated encoder separately, and then fusing through product space to ensure that the information contribution of each isolated data point is separable and quantifiable; adaptively determining the compressed feature dimension as the minimum sufficient statistical dimension required for the prediction task; the objective function of the variational information bottleneck is expressed as follows: ;in For compression features, For fault labels, Original features To compress the mutual information between features and fault labels, To compress the mutual information between the original features and the original features, To normalize the reference mutual information, The balance coefficient between compression ratio and predictive correlation The values ​​are determined by taking values ​​in the interval [0.1] using a step size of 0.1 on a historical multi-island dataset. A grid search experiment was conducted to verify the result corresponding to the highest F1 score for fault classification. The value is determined; the multi-island optimal feature compression algorithm based on the information bottleneck principle enables the subsequent cross-island state representation model to process only the minimum sufficient feature subset carrying the maximum fault prediction information, remove redundant and noisy features between islands, effectively suppress the curse of dimensionality in the high-dimensional sparse feature space, and at the same time, the independent encoding of each island ensures the traceability of the information contribution of each island, providing high-quality input for the cross-island state representation model.

[0041] The specific structure of the cross-island state representation model is as follows: The cross-island state representation model integrates two parts: a cross-island heterogeneous graph attention network and a cross-island federated meta-learning fast adaptation network. The cross-island heterogeneous graph attention network consists of three layers: the first layer is a local aggregation layer within each island, which aggregates local features for similar device nodes within each island using a multi-head graph attention mechanism. Each attention head independently learns the contribution weights of its neighboring nodes, and the multi-head outputs are concatenated and linearly transformed to obtain the local feature representation within the island; the second layer is a cross-island physical coupling aggregation layer, which encodes the cross-island physical coupling relationship as graph sparsification constraint edges based on the heat transfer equation and electromagnetic equation prior constraints encoded in the physical mechanism knowledge graph, and uses a graph sparsification algorithm to automatically identify and target devices. The first layer performs graph attention aggregation on a sparsely coupled graph, capturing cross-island cascaded physical effects, based on a highly correlated subset of features. The second layer is a global readout layer, which uses attention-weighted summation to read the features of all graph nodes into a device health status representation vector and outputs confidence scores simultaneously. When the confidence score is lower than the confidence threshold, the graph structure of the multi-island heterogeneous knowledge graph is dynamically updated and attention propagation is re-executed, with a maximum of 5 iterations. The number of graph attention heads in the local aggregation layer within an island, the cross-island physical coupling aggregation layer, and the global readout layer is adjusted by the fault risk score. The function's output is dynamically adjusted; the cross-island federated meta-learning fast adaptation network adopts a meta-learner-driven two-layer long short-term memory network structure: the outer long short-term memory network encodes cross-task general temporal representations, with the hidden layer dimension consistent with the dimension of the device health status representation vector output by the global readout layer; the inner long short-term memory network is a task-specific adaptation layer, storing fast weight parameters, and performing task-specific fast adaptation after receiving the output of the outer long short-term memory network; during federated training, each island locally calculates the meta-gradient of the task-specific adaptation layer, and the federated server aggregates the meta-gradients of each island and updates the global meta-initialization parameters, i.e., slow weights; when new devices or new fault types appear, the inner long short-term memory network performs fast adaptation using a small number of labeled samples, with no more than 5 steps of inner loop gradient steps; the number of steps of the inner loop gradient steps is dynamically determined by the early stopping criterion of validation loss; the connection between the cross-island heterogeneous graph attention network and the cross-island federated meta-learning fast adaptation network is as follows: the device health status representation vector output by the global readout layer is directly used as the input sequence frame of the outer long short-term memory network, realizing end-to-end transfer from graph structure features to temporal representations.

[0042] The steps for establishing the training dataset for the cross-island state representation model specifically include: collecting time-series records of sensors from each island device from the historical database of each business system, and dividing them into sample units according to device number and time window; labeling sample units with fault records with corresponding fault category labels, and retaining sample units without fault labels as unsupervised pre-training samples; using conditional generative adversarial networks to synthesize scarce fault category samples under physical constraints, with the synthesis ratio determined by the ratio of various fault samples to normal samples, until the proportion of labeled samples for each fault category is not less than 1%; dividing the training dataset according to the boundaries of each island, with each island independently holding its own local training dataset and not sharing original data across islands; and allocating 80% of the local training dataset from each island for training, 10% for validation, and 10% for testing.

[0043] The training steps of the cross-island state representation model specifically include: First, using unlabeled historical data from each island, a three-layer structure of the cross-island heterogeneous graph attention network is trained using a contrastive self-supervised pre-training method to learn a general representation of device state. The number of pre-training rounds is determined by the convergence of the representation quality index on the validation set. Second, a two-layer long short-term memory network is trained using a federated meta-learning method. Each island performs an internal loop gradient step to update fast weights, and the federated server aggregates the meta-gradients of each island and updates the slow weights. The number of federated aggregation rounds is determined by the early stopping criterion of the global validation loss. Third, a few-sample fine-tuning is performed on labeled sample units with fault category labels. The slow weights of the cross-island heterogeneous graph attention network are fixed, and only the fast weights of the task-specific adaptation layer are updated. The number of fine-tuning steps does not exceed 5. The confidence threshold is determined by traversing the candidate value range on the validation set and experimenting with the highest convergence rate of the device health state representation vector iteration, with no more than 5 iteration rounds.

[0044] The cross-island state representation model integrates a cross-island heterogeneous graph attention network and a cross-island federated meta-learning fast adaptation network. This allows the model to capture local device features within an island and cross-island cascaded physical coupling features without the sensor time-series data of each island device leaving the local environment. Meta-learning initialization enables the model to quickly adapt to new fault types. The attention weights of the cross-island heterogeneous graph attention network directly reflect the contribution of each neighboring node to the state of the target device. Combined with physical mechanism knowledge graph constraints, the model learns physically meaningful real causal links rather than spurious statistical correlations, thereby improving the interpretability of the final fault prediction results and the generalization stability under distribution changes.

[0045] The equipment failure risk score is calculated as follows: the cosine similarity between the equipment health status representation vector and each standard vector in the preset failure category standard vector set is calculated, and the maximum cosine similarity value is linearly transformed to map to the interval between 0 and 1 to obtain the equipment failure risk score; the failure warning threshold is obtained by traversing the interval on the test set with a step size of 0.01. The candidate values ​​within the range are determined by the value corresponding to the highest harmonic mean of precision and recall.

[0046] The confidence level is calculated as follows: the maximum class probability value after softmax normalization of the global readout layer output is taken as the confidence level for the current round; the method for determining the confidence level threshold is described in the relevant description of the training steps of the cross-island state representation model.

[0047] The small sample trigger threshold is the upper limit of the number of labeled samples for the corresponding new type of isolated fault. This threshold is determined by using a step size of 5 in the synthetic sample expansion experiment within the specified interval. Gradually increase a small number of real labeled samples and observe the change in the F1 value of the validation set after rapid adaptation. Take the number of samples when the F1 value no longer increases significantly with the increase of the number of samples as the small sample trigger threshold.

[0048] The input parameters and calculation method of the fault risk score adjustment function are as follows: based on the average fault risk score of the equipment within the current time window. Normalized value of the convergence iteration of state representation and the utilization rate of compressed feature dimensions The weighted sum of the three parameters yields the comprehensive adjustment value. The formula is expressed as follows: ;in The normalization benchmark for equipment failure risk scores is set to 1. The maximum number of iterations is 5. For the original feature dimension, , , The weighting coefficients are satisfied. The three weighting coefficients are adjusted on the validation set with a step size of 0.1. , , A grid search experiment was conducted to verify the determination of the weight coefficient combination that yields the highest F1 score for fault classification; when Attention count Adjusted to 4; when hour, Adjusted to 8; when hour, Adjusted to 12; when hour, The number of attention heads was adjusted to 16; the candidate values ​​of 4, 8, 12, and 16 were determined through ablation experiments under different load and sample distribution conditions, with the optimal head number combination corresponding to the comprehensive score of prediction accuracy and computational delay; the comprehensive adjustment value of the output result of the fault risk score adjustment function. The number of graph attention heads used to dynamically adjust the cross-island state representation model in step S06. .

[0049] Among them, the average score of the equipment failure risk The equipment failure risk score is derived from the outputs of steps S04 and S05, and is obtained by taking the arithmetic mean of all output equipment failure risk scores within the current time window; the state representation iteration convergence rounds are derived from the actual iteration rounds of attention propagation in step S03; the compressed feature dimension utilization rate... The effective feature dimension derived from the multi-island optimal feature compression algorithm based on the information bottleneck principle in step S02. Compared with the original feature dimension The ratio.

[0050] The hierarchical resource isolation scheduler refers to dividing computing resources into multiple isolated quota pools based on the type of prediction task. Fault warning queries take priority in using the dedicated quota pool for fault warnings, while other queries use the remaining quota pool. Resource isolation avoids interference from concurrent queries on the real-time performance of fault warnings.

[0051] The incremental streaming computing mode refers to triggering the calculation and updating of equipment failure risk scores one by one or in small batches according to the order of data arrival, rather than waiting for all data to arrive before batch processing, thereby controlling the end-to-end prediction delay to within 500ms.

[0052] The product space fusion refers to concatenating the compressed feature vectors output by each isolated encoder along the feature dimension to form a joint compressed feature representation. The compressed feature components of each isolated island remain separable after concatenation, which facilitates subsequent separation and quantification of the information contribution of each isolated island. The joint compressed feature representation is the compressed feature set of each isolated island.

[0053] The conditional generative adversarial network refers to introducing fault category labels and physical constraints as conditional inputs to the generator and discriminator in the generative adversarial network framework, so that the generator can synthesize fault category sensor time-series samples that meet the physical law constraints, which is used to expand the number of labeled sample units for scarce fault categories.

[0054] The aforementioned contrastive self-supervised pre-training refers to constructing positive sample pairs by applying different random augmentation transformations to the time-series samples of the same device's sensors under the condition of no fault category label, and using the time-series samples of sensors from different devices or different time periods as negative samples. The cross-island heterogeneous graph attention network is trained so that the device health status representation vectors of the positive sample pairs are close to each other and the device health status representation vectors of the negative sample pairs are far apart, thereby learning a general device state representation related to fault prediction.

[0055] The physical mechanism knowledge graph refers to encoding prior physical knowledge in fields such as equipment heat transfer equations and electromagnetic equations into a graph structure. Nodes are physical quantities or physical subsystems, and directed edges represent the direction and intensity of causal influences. This serves as the prior constraint input for the graph sparsification algorithm, ensuring that the cross-island physical coupling aggregation layer propagates information only along physically meaningful coupling paths.

[0056] The graph sparsification algorithm refers to pruning the full set of edges in a multi-island heterogeneous knowledge graph based on the prior physical constraints in the physical mechanism knowledge graph, retaining the subset of edges that are highly related to the state of the target device, and compressing the number of associated nodes corresponding to the effective feature dimensions to less than 5% of the original number of nodes. The compressed subset of edges constitutes a sparsified coupling graph, which serves as the graph structure input for the cross-island physical coupling aggregation layer.

[0057] It's important to explain that the task-specific adaptation layer is a core component of the inner long short-term memory network in the cross-island federated meta-learning fast adaptation network. Essentially, it's a set of network layers storing "fast weights" parameters. Unlike the outer long short-term memory network, which is responsible for learning the general temporal representation across tasks, the task-specific adaptation layer is specifically designed to quickly adjust its parameters using a very small number of locally labeled samples (below the small sample trigger threshold) through gradient updates in no more than 5 steps when facing new devices or new types of faults. This allows the model to rapidly adapt to the specificities of the current task. During the federated training phase, each island independently calculates the meta-gradient of this layer locally and uploads it to the federated server for aggregation, thereby continuously optimizing its initialization starting point (i.e., the initial values ​​of the slow weights relative to the fast weights). This allows it to converge to good performance with only a few steps when encountering new tasks. Throughout the adaptation process, the slow weights of the outer network remain fixed, ensuring that the general representation capability across islands is not compromised. Simultaneously, the task-specific adaptation layer undertakes all the "personalized fine-tuning" responsibilities, achieving decoupling and synergy between generalization and fast adaptation.

[0058] The specific implementation of step S01 is as follows: The data integration engine connects to the data interfaces of various business subsystems in an adapter mode, including heterogeneous platforms such as the production execution system, energy management system, and equipment management system. It automatically crawls the table structure metadata, field semantic annotations, and time-series sampling records of each isolated system, with a sampling frequency typically between 1Hz and 10Hz. The metadata registry center registers the device sensor nodes, data table nodes, and physical subsystem nodes within each isolated system as unified node identifiers according to a three-level hierarchical relationship based on the device number and system affiliation. The edge types between nodes are labeled according to three types of directed edges: belonging, referencing, and derived. The original feature vector of each node is constructed by concatenating the statistical quantities such as the mean, variance, maximum, minimum, and autocorrelation coefficient of the corresponding sensor time-series data within a fixed time window, with a reference time window length of 60s to 300s. After the above processing, the node and edge information of each isolated system is organized into a unified heterogeneous graph structure, i.e., a multi-isolated heterogeneous knowledge graph, serving as the unified input carrier for subsequent algorithms.

[0059] The specific implementation of step S02 is as follows: The multi-island optimal feature compression algorithm based on the information bottleneck principle formalizes the multi-island joint feature extraction problem into an information bottleneck optimization problem, with the goal of maximizing the compression of features. With fault labels Mutual Information Minimize the compression feature under constraints With original features Mutual Information Using Lagrange multipliers A balance is struck between compression ratio and prediction relevance. The variational information bottleneck method replaces mutual information estimation with a differentiable variational lower bound, allowing the objective function to be solved via standard backpropagation gradient descent. The values ​​are selected within the range [0.1, 2.0] using a grid search with a step size of 0.1. The value corresponding to the highest F1 score in the validation set for fault classification is selected, with a reference value of 0.5 to 1.0. Each island is configured with an independent encoder. The encoder input is the original feature vector of the corresponding island node, and the output is a compressed feature vector. The effective feature dimension is adaptively compressed to 2% to 5% of the original feature dimension. The output compressed feature vectors of each island's independent encoder are sequentially concatenated along the feature dimension to form a joint compressed feature representation in the product space. The compressed feature components of each island still maintain positional separability after concatenation, which facilitates independent tracing of the information contribution of each island. The joint compressed feature representation is the compressed feature set of each island, which serves as the input to the cross-island state representation model.

[0060] The specific implementation of step S03 is as follows: The cross-island state representation model consists of two end-to-end connections: a cross-island heterogeneous graph attention network and a cross-island federated meta-learning fast adaptation network. The local aggregation layer within each island of the cross-island heterogeneous graph attention network uses a multi-head graph attention mechanism to aggregate local compressed features for similar device nodes within each island. Each attention head independently learns the contribution weights of neighboring nodes. The multi-head outputs are concatenated and linearly transformed to obtain the local feature representation within the island. The initial number of attention heads is 8. The cross-island physical coupling aggregation layer, based on the prior constraints of the heat transfer equation and electromagnetic equation encoded in the physical mechanism knowledge graph, prunes the full edge set of the multi-island heterogeneous knowledge graph using a graph sparsification algorithm, retaining a subset of edges highly related to the target device state. This compresses the number of effectively associated nodes to less than 5% of the original number of nodes. Graph attention aggregation is then performed on the sparsely coupled graph to capture cross-island cascaded physical effects. The global readout layer uses an attention-weighted summation method to read the features of all nodes in the graph into a device health status representation vector. The maximum class probability value after softmax normalization is taken as the confidence output, with a confidence threshold of 0.7 to 0.85. When the confidence falls below the threshold, the graph structure of the multi-island heterogeneous knowledge graph is dynamically updated based on the current attention weight distribution, and attention propagation is re-executed, iterating a maximum of 5 rounds until the Euclidean distance between two adjacent rounds of the device health status representation vector is lower than the convergence criterion threshold (reference value is...). The device health status representation vector is then used as the input sequence frame of the outer long short-term memory network, and further processed by the task-specific adaptation layer of the inner long short-term memory network to output the final temporally enhanced status representation.

[0061] The specific implementation of step S04 is as follows: The cosine similarity between the device health status representation vector and each standard vector in the preset fault category standard vector set is calculated. The maximum cosine similarity value is then linearly transformed and mapped to the interval between 0 and 1 to obtain the device fault risk score. The fault warning threshold is determined by iterating through candidate values ​​in the interval [0.50, 0.95] on the test set with a step size of 0.01, and the value corresponding to the highest harmonic mean of precision and recall is selected, with a reference value of 0.75 to 0.85. When the device fault risk score exceeds the fault warning threshold, the hierarchical resource isolation scheduler divides the computing resources into a fault warning dedicated quota pool and a general quota pool according to the prediction task category. Fault warning type queries preferentially occupy the fault warning dedicated quota pool, while other concurrent queries use the remaining general quota pool. Resource isolation avoids concurrent load interfering with the real-time performance of the warning. At the same time, the incremental streaming computing mode is activated, which triggers calculations and updates the equipment failure risk score in real time according to the order of data arrival, keeping the end-to-end prediction delay within 500ms, instead of waiting for all data to arrive before batch processing.

[0062] The specific implementation of step S05 is as follows: The system continuously monitors the device type registry and fault category annotation library of each island. When a new type of device or fault type is detected and the number of annotated samples in the corresponding island is lower than the small sample trigger threshold, the rapid adaptation process is triggered. The small sample trigger threshold is determined by gradually increasing the number of real annotated samples in the interval [5, 50] with a step size of 5 in the synthetic sample expansion experiment and observing the change in the F1 value of the validation set. The number of samples corresponding to when the F1 value no longer increases significantly is taken, with a reference value of 20 to 30. The inner long short-term memory network of the cross-island state representation model uses a small number of local annotated samples to perform no more than 5 inner loop gradient steps. The number of inner loop steps is dynamically determined by the early stopping criterion of validation loss, that is, when the validation loss no longer decreases, it terminates early and only updates the fast weight parameters of the task-specific adaptation layer. The slow weights of the cross-island heterogeneous graph attention network remain fixed, thereby achieving rapid adaptation to new fault types while retaining the cross-island general representation capability, and outputting the updated device fault risk score.

[0063] The specific implementation of step S06 is as follows: the fault risk score adjustment function uses the average equipment fault risk score within the current time window. Normalized value of the convergence iteration of state representation and the utilization rate of compressed feature dimensions The weighted sum of the three parameters yields the comprehensive adjustment value. The formula is ,in Take 1, Take 5 as the weighting coefficient. , , Determined through a grid search of the validation set, the reference value is... , , .in accordance with The interval to which it belongs, when Time-based attention head count Adjusted to 4, when hour Adjusted to 8, when hour Adjusted to 12, when hour The number of graph attention heads was adjusted to 16, with candidate values ​​of 4, 8, 12, and 16 determined by the comprehensive scoring of ablation experiments under different load and sample distribution conditions. The number of graph attention heads in the local aggregation layer within each island, the cross-island physical coupling aggregation layer, and the global readout layer were all updated synchronously based on the above adjustment results. Finally, the fault prediction results were pushed to each business system via the data service bus to complete the complete prediction loop.

[0064] It's important to note that the first key technical approach is based on the principle of information bottlenecks, which involves independent encoding of isolated islands and product space fusion. Traditional methods often directly concatenate high-dimensional original features into the model when fusing multi-source data, leading to the curse of dimensionality and interference between redundant and noisy features in isolated islands. The information bottleneck principle, from an information theory perspective, ensures that the compressed features are the minimum sufficient statistics required for fault prediction. Independent encoding of isolated islands ensures that the compression process does not depend on sharing the original data, and product space fusion preserves the separability of information contributions from each island. Thus, high-quality compression of joint features across isolated islands is achieved without the data leaving its local location. The second key technical approach is cross-island heterogeneous graph attention aggregation constrained by a physical mechanism knowledge graph. Traditional graph neural networks are prone to introducing spurious statistical correlation paths when aggregating information on dense graphs. However, embedding prior physical knowledge such as heat transfer equations and electromagnetic equations into the graph sparsification algorithm ensures that information propagates only along physical causal links. The graph attention weights thus directly correspond to the real physical coupling strength, significantly improving the physical rationality of cross-island state representations and the generalization stability under distribution variations. The third key technical approach is the decoupling of slow and fast weights in federated meta-learning. Traditional federated learning only aggregates model parameters, making it difficult to quickly adapt to new fault types. In a two-layer long short-term memory network, the outer slow weights continuously learn general temporal representations across isolated systems under federated aggregation, while the inner fast weights quickly adapt on a small number of locally labeled samples. This decoupling allows the model to simultaneously possess cross-isolated generalization ability and rapid adaptation ability within a single isolated system. The synergistic effect of these three technical approaches is as follows: information bottleneck compression provides low-noise, high-quality input for graph attention aggregation; physical constraint graph sparsification provides physically meaningful state representations for federated meta-learning; and federated meta-learning ensures continuous self-optimization of the entire process under data-non-sharing constraints. Together, these three elements constitute a complete closed-loop technical system from feature compression and cross-isolated modeling to rapid adaptation.

[0065] It should be noted that this invention also addresses the following technical problem: In industrial equipment fault prediction scenarios, the scarcity of fault category samples leads to severe imbalance in model training. Traditional oversampling methods struggle to guarantee the physical plausibility of synthesized samples on multi-island heterogeneous sensor data, resulting in a significant decrease in the model's prediction accuracy for scarce fault categories. This invention introduces a conditional generative adversarial network (GAN), embedding fault category labels and physical constraints simultaneously into the conditional inputs of the generator and discriminator. This allows the generator to synthesize fault category sensor time-series samples that satisfy physical constraints. The synthesis ratio is dynamically determined by the ratio of each fault category sample to normal samples, until the proportion of labeled samples for each fault category is no less than 1%. This mechanism ensures the consistency between the synthesized samples and the actual physical process, effectively solving the technical problem of physical plausibility of scarce fault category samples in multi-island scenarios. This invention also addresses the technical problem of difficulty in guaranteeing real-time fault prediction under high-concurrency loads. The hierarchical resource isolation scheduler divides computing resources into mutually isolated quota pools, cutting off the path for concurrent queries to preempt computing resources for fault early warning from the system scheduling level; the incremental streaming computing mode eliminates the fixed delay introduced by waiting for data accumulation in the full batch processing mode by triggering updates one by one or in small batches. The two work together to ensure that the end-to-end prediction latency of fault early warning can still be stably controlled within 500ms in high-concurrency scenarios, solving the technical problem of resource competition between real-time early warning and system throughput.

[0066] A second aspect of the present invention provides a computer-readable storage medium storing program instructions that, when executed in a computer, are used to perform the aforementioned method for predicting data silos in asset big data.

[0067] A third aspect of the present invention provides a data silo prediction system based on asset big data, comprising the aforementioned computer-readable storage medium, wherein the system is any one of a computer, a server, or a microcontroller, the computer-readable storage medium is disposed within the system, and the system is provided with a microprocessor that executes the program instructions stored in the computer-readable storage medium.

[0068] Specifically, the principle of this invention is as follows: The fundamental reason why this invention can solve the above-mentioned core technical problems is that: the information bottleneck principle theoretically ensures that each isolated encoder minimizes the amount of redundant information transmission while maximizing the retention of fault prediction-related information; the product space fusion maintains the separability of the compressed feature components of each isolated island in the feature dimension, so that the cross-island joint features still have sufficient statistical power without the need to aggregate the original data; the three-layer heterogeneous graph attention structure of the cross-island state representation model constrains the edge set sparsity with the physical mechanism knowledge graph, limiting the information aggregation path of the graph neural network to the physical causal link, so that the attention weights directly correspond to the real physical coupling strength rather than statistical pseudo-correlation; the federated meta-learning mechanism of the two-layer long short-term memory network enables the slow weights to continuously optimize the cross-task general temporal representation under the aggregation of meta-gradients of each isolated island, and the fast weights to quickly adapt to new fault types on a small number of locally labeled samples, thereby achieving a synergistic improvement in cross-island generalization ability and single-island rapid adaptation ability under the constraint of data localization, and ensuring the physical rationality of state representation and the accuracy of prediction from a mechanism perspective.

[0069] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0070] The specific implementation of step S01 is as follows: The data integration engine connects to the data interfaces of various heterogeneous business subsystems in an adapter mode, automatically extracting the table structure metadata, field semantic annotations, and device sensor time-series sampling records of each isolated system. The metadata registration center registers the device sensor nodes, data table nodes, and physical subsystem nodes within each isolated system as unified node identifiers according to a three-level hierarchical relationship based on the device number and system affiliation. Edge types are labeled according to three types of directed edges: belonging, referencing, and derived. The node feature vector is constructed by concatenating the statistics of the corresponding sensor time-series data within a fixed time window. Let the first... The original feature vectors of the nodes are ,in For the original feature dimension, From the mean ,variance Maximum value Minimum value and autocorrelation coefficient ( , The number of time delays (empirically 5) is used to construct the structure, specifically represented as follows:

[0071] ;

[0072] In the formula, The normalized baseline value for the mean is the maximum absolute value of the historical mean values ​​of each sensor. The normalized baseline value for the variance is the maximum historical variance of each sensor. The normalized reference value for the maximum value is the maximum absolute value of the historical maximum values ​​of each sensor. The normalized reference value for the minimum value is the maximum absolute value of the historical minimum values ​​of each sensor; superscript Indicates transpose , , , Each term is normalized using the maximum absolute value of its corresponding historical statistic as the benchmark, making each term a dimensionless scalar; autocorrelation coefficient Its value range is Since it is a dimensionless scalar, it requires no additional normalization and its dimensions are consistent with those of the other quantities. The entire vector is dimensionless. Autocorrelation coefficient. The calculation is expressed as follows:

[0073] ;

[0074] In the formula, For nodes The corresponding sensor in the first Observations at each sampling time The total number of sampling points within the time window For the first One time delay step; the numerator and denominator have the same physical dimensions, after division It is a dimensionless scalar. After the above processing, the information of each isolated node and edge is organized into a unified heterogeneous graph structure, namely a multi-isolated heterogeneous knowledge graph, which serves as a unified input carrier for subsequent algorithms.

[0075] The specific implementation of step S02 is as follows: The multi-island optimal feature compression algorithm based on the information bottleneck principle formalizes the multi-island joint feature extraction problem into an information bottleneck optimization problem. The variational information bottleneck objective function is expressed as follows:

[0076] ;

[0077] In the formula, The objective function value Let the normalization baseline value of the objective function be 1, so that... Dimensionless scalar For compression features Fault Label Original features To compress the mutual information between features and fault labels Mutual information between compressed features and original features To normalize the baseline mutual information, take the training set... The estimated value makes and Both are dimensionless scalars, and the dimensions on both sides of the equal sign are consistent. The coefficient representing the balance between compression ratio and prediction correlation is a dimensionless scalar. It is determined by performing a grid search within the interval [0.1, 2.0] on a historical multi-island dataset with a step size of 0.1, using the value corresponding to the highest F1 score for fault classification on the validation set. A reference value is taken from 0.5 to 1.0. The variational information bottleneck method is used to integrate mutual information... The estimate is replaced by a variational upper bound. The estimate is replaced with a variational lower bound, allowing the objective function to be solved via standard backpropagation gradient descent. Each isolated encoder will generate the original features Mapped to compressed features ,in For the first The first isolated island The normalized original feature vector of each node. For the first The original feature dimensions of an isolated island For the first The first isolated island The compressed feature vector of each node, For the first The compressed feature dimension of each isolated island satisfies The encoder output follows a conditional Gaussian distribution, as described below:

[0078] ;

[0079] In the formula, For the first The variational posterior distribution parameterized by an isolated encoder For the first Network parameters of an isolated encoder The mean vector output by the encoder is generated by the encoder network. Forward computation of input This is the variance vector output by the encoder, where each component is a positive real number. This represents a diagonal matrix formed by using vector elements as diagonal elements. The mean is The covariance matrix is multivariate Gaussian distribution The sampling employs a reparameterization technique, described as follows:

[0080] ;

[0081] In the formula, The standard deviation is the element-wise square root of the variance vector. To and Standard Gaussian noise vector of the same dimension for 3D identity matrix element-wise product , and All are dimensionless vectors, with uniform dimensions on both sides of the equals sign. The compressed features from each isolated feature are fused and concatenated through a product space to form a joint compressed feature vector, as described below:

[0082] ;

[0083] In the formula, Jointly compressed feature vectors Total number of isolated islands The total dimension of the joint compressed features is the sum of the dimensions of each isolated compressed feature; the isolated compressed feature components remain positionally separable after splicing. It is a dimensionless vector.

[0084] The specific implementation of step S03 is as follows: the local aggregation layer within the island affects the first... Nodes within an isolated island Employing a multi-head graph attention mechanism, the first Attention coefficient of each attention head The statement is as follows:

[0085] ;

[0086] In the formula, For attention head index, , To increase the number of attention heads, the fault risk score adjustment function in step S06 is dynamically adjusted. For the first First The linear transformation matrix of the isolated island. The attention head hidden layer dimension has an empirical value of 64. For the first First Learnable attention vectors of isolated islands For nodes In the Set of neighboring nodes within an isolated island This represents vector concatenation. This is an inner product operation, the result of which is a dimensionless scalar; LeakyReLU is a linear rectified activation function with a negative slope, the empirical value of which is 0.2. For dimensionless normalized weights, satisfying For the neighbor node traversal index, and They have the same meaning and are used for normalizing the denominator during summation to distinguish it from the numerator. Node On the deserted island Feature representation after local aggregation ( The dimension representing local features within the isolated island, empirical value and (Same) is obtained by concatenating multiple outputs and then performing a linear transformation, as described below:

[0087] ;

[0088] In the formula, For the first The output linear transformation matrix of the island Indicates to The outputs of each attention head are concatenated sequentially along the vector dimension, resulting in the following concatenation: dimensionless vector For dimensionless vectors, the dimensions on both sides of the equals sign are unified. The cross-island physical coupling aggregation layer, based on prior constraints from the physical mechanism knowledge graph, prunes the entire edge set using a graph sparsification algorithm, retaining only those with coupling strength exceeding the sparsification threshold. Cross-island edges (with an empirical value of 0.1) form a sparse coupling graph, compressing the number of effectively associated nodes to less than 5% of the original number of nodes. Graph attention aggregation, consistent with the local aggregation layer structure within the islands, is then performed on the sparse coupling graph to obtain the cross-island coupling feature representation. ,in The node feature dimension after cross-island coupling aggregation is empirically set to 512. The global readout layer uses an attention-weighted summation method to read the node features of the entire image into a device health status representation vector. The normalized representation of the attention weights for all nodes in the graph is as follows:

[0089] ;

[0090] In the formula, Learnable query vectors for the global readout layer For the set of nodes in the entire graph For the index of the nodes in the full graph, and Both are node indices, used for normalizing the denominator during summation to distinguish them from the numerator. This is an inner product operation, and the result is a dimensionless scalar. This is a scaling factor used to prevent the gradient from vanishing due to excessively large inner product values. For dimensionless normalized weights, satisfying Equipment health status representation vector The calculation is expressed as follows:

[0091] ;

[0092] In the formula, and All A dimensionless eigenvector, with both sides of the equation having the same dimensions. Confidence level. The calculation is expressed as follows:

[0093] ;

[0094] In the formula, The weight matrix of the global readout layer classification layer This represents the total number of fault categories (including normal categories). To maximize the class index output by softmax, This is a category traversal index used for normalizing the sum of the denominator to distinguish it from the numerator. express The The element, i.e., the th element Logit value of the class It is a dimensionless confidence scalar. ( The confidence threshold is determined by iterating through the candidate value range on the validation set to find the value with the highest convergence rate and no more than 5 iterations (reference value is 0.7 to 0.85). When this value is reached, the graph structure of the multi-island heterogeneous knowledge graph is updated and attention propagation is re-executed, for a maximum of 5 iterations, until two adjacent iterations... The Euclidean distance is lower than the convergence criterion threshold (empirical value is ). Equipment health status representation vector Subsequently, the sequence frames, which are then used as input to the outer long short-term memory network, are further processed by the task-specific adaptation layer of the inner long short-term memory network to output an enhanced temporal state representation.

[0095] The specific implementation of step S04 is: to convert the device health status representation vector Compared with the preset fault category standard vector set ( For the first Class Fault Standard Vector , This represents the total number of fault categories. (Calculated from the mean of the state representation vectors corresponding to the training samples of each fault category) Cosine similarity is calculated separately, and equipment fault risk score is determined. The calculation is expressed as follows:

[0096] ;

[0097] In the formula, The dot product of vectors results in a dimensionless scalar (because...). and (All are dimensionless vectors) and They are vectors and The Euclidean norms of these numbers are all dimensionless positive real numbers; cosine similarity The range of values ​​is To define the lower bound of the linear transformation, all samples from the historical validation set are taken. minimum value The upper bound for the linear transformation is determined by taking all samples from the historical validation set. The maximum value; if Let the denominator be 1 to avoid division by zero. Mapped to The interval is a dimensionless scalar, and the dimensions on both sides of the equals sign are the same. Fault warning threshold. The value was determined by traversing the interval [0.50, 0.95] on the test set with a step size of 0.01, using the value corresponding to the highest harmonic mean of precision and recall, with a reference value of 0.75 to 0.85. At that time, the hierarchical resource isolation scheduler is triggered and the incremental streaming computing mode is activated, and the end-to-end prediction latency is controlled within 500ms.

[0098] The specific implementation of step S05 is as follows: When a new type of equipment or a new type of fault is detected and the number of corresponding isolated labeled samples is lower than the small sample trigger threshold (reference value is 20 to 30), the inner long short-term memory network uses a small number of local labeled samples to execute no more than 5 inner loop gradient steps. The gradient update within the loop is expressed as follows:

[0099] ;

[0100] In the formula, For the first The task-specific adaptation layer's fast weight parameter vector updated within the inner loop. The initial fast weights for the inner loop are the initial values ​​of the slow weights obtained from federated meta-learning training. , The actual number of inner loop steps is dynamically determined by the early stopping criterion based on validation set loss. The learning rate for the inner loop is a dimensionless positive real number, with an empirical value ranging from 0.01 to 0.05. for right The gradient vector, and Same dimension To support the cross-entropy loss on a set, it is expressed as follows:

[0101] ;

[0102] In the formula, To support the number of samples in the set For the first The first sample class-specific hot tags For the inner long short-term memory network, use the current fast weights The output of the first The first sample Class prediction probability Dimensionless scalar and Dimensions are consistent, and the dimensions on both sides of the equal sign are unified; after rapid adaptation, the inner long short-term memory network... The updated device failure risk score is output for the parameters; the slow weights of the cross-island heterogeneous graph attention network remain fixed throughout the inner loop.

[0103] The specific implementation method of step S06 is: comprehensive adjustment value The calculation is expressed as follows:

[0104] ;

[0105] In the formula, The comprehensive adjustment value is a dimensionless scalar. This is the arithmetic mean of the equipment failure risk scores within the current time window, calculated from all equipment failure risk scores output in steps S04 and S05, and is a dimensionless scalar. The normalization benchmark for equipment failure risk scores is set to 1, making Dimensionless scalar The actual convergence round of attention propagation in step S03 is a positive integer. The maximum number of iterations is set to 5, making... Dimensionless scalar For joint compression of total feature dimension For the original feature dimension, make Dimensionless scalar , , These are weighting coefficients, all of which are dimensionless positive real numbers, satisfying... The combination with the highest F1 score is determined by a grid search of the validation set (step size 0.1). A reference value is [value missing]. , , All three terms are dimensionless scalars, and after weighted summation... The whole is dimensionless, and the dimensions on both sides of the equal sign are unified; based on The interval to which it belongs, when Time-based attention head count Adjusted to 4, when hour Adjusted to 8, when hour Adjusted to 12, when hour After the number of attention heads in the three layers of graph—the local aggregation layer within the island, the physical coupling aggregation layer across islands, and the global readout layer—is adjusted to 16, the final fault prediction result is pushed to the business system via the data service bus.

[0106] To better understand and implement this invention, the following is a specific application scenario of the invention, Example 2: To verify the effect of the invention, the technicians built a test environment and selected the turbine unit of a large thermal power generator unit as the test object. The unit involves three independent business subsystems: thermal control system, electrical monitoring system and lubrication oil system, which constitute three typical heterogeneous data islands. The three islands are deployed on physically isolated servers, and the original sensor data is not allowed to be transmitted across islands.

[0107] Technicians first execute step S01, connecting the data interfaces of the three isolated islands via the data integration engine in adapter mode. This automatically extracts metadata and sensor time-series sampling records from each island. The thermal control island contains 128 sensors (temperature, pressure, flow, etc.); the electrical monitoring island contains 64 sensors (voltage, current, power factor, etc.); and the lubrication oil island contains 32 sensors (oil temperature, oil pressure, flow, etc.). The sampling frequency is uniformly set to 5Hz. The metadata registry organizes the 224 sensor nodes, corresponding data table nodes, and three physical subsystem nodes into a multi-island heterogeneous knowledge graph based on device number and attribution. Edge types are labeled as belonging to, referencing, and converted from. Node features are constructed by concatenating the mean, variance, maximum, minimum, and autocorrelation coefficients within a 120-second time window. The original feature dimensions... The value is 512.

[0108] Then, step S02 is executed, where the multi-island optimal feature compression algorithm based on the information bottleneck principle optimizes each island's independent encoder separately. The grid search determined the value to be 0.7. The compressed feature dimensions of the three isolated islands were 12, 6, and 4, respectively. The combined effective feature dimension... The utilization rate of compressed feature dimensions is 22. The compression ratio is approximately 4.3%, meeting the design requirement of 2% to 5%. The compressed feature vectors of each isolated island are concatenated in a product space to form a 22-dimensional joint compressed feature, and the positions of each isolated island component in the concatenated vector are separable.

[0109] Technicians input the joint compressed features into the cross-island state representation model according to step S03. The physical mechanism knowledge graph encodes the heat transfer equation and rotor electromagnetic equation of the turbine unit. The graph sparsification algorithm prunes the entire edge set based on prior constraints, compressing the number of effective associated nodes from 224 to 11, with a sparsity ratio of approximately 4.9%. Initial graph attention head number Taking 8, the global readout layer outputs a 512-dimensional device health status representation vector. The first-round confidence score is 0.91, which is higher than the confidence threshold of 0.80. No iterative graph structure update is required, and the state table collects convergence rounds. The value is 1. The device health status representation vector is then processed by the outer long short-term memory network and the inner long short-term memory network to output the enhanced temporal status representation.

[0110] In step S04, the cosine similarity between the state representation vector and the preset set of 7 types of fault standard vectors is calculated. The maximum cosine similarity corresponds to the lubricating oil system oil film eddy fault category. The equipment fault risk score after mapping is 0.83, and the fault warning threshold is determined to be 0.78 after traversing the test set. The fault risk score exceeds the warning threshold, triggering the hierarchical resource isolation scheduler to allocate dedicated quota pool resources for fault warning to this prediction task and activating the incremental streaming computing mode. The actual measured end-to-end prediction latency is 347ms, which meets the design requirement of less than 500ms. The statistics of the fault risk scores of each island in this round of prediction are shown in Table 1:

[0111] Table 1. Statistical Table of Failure Risk Scores for Each Isolated Equipment

[0112]

[0113] Average equipment failure risk score within the current time window The score is 0.67. Due to a new type of oil film instability mode that was recently developed after the maintenance of the lubricating oil island, the corresponding number of labeled samples is 18, which is lower than the small sample trigger threshold of 25. Step S05 is executed. The inner long short-term memory network uses 18 labeled samples to perform 3 inner loop gradient steps to verify early stop of loss. After the fast weight update, the equipment failure risk score is corrected to 0.87, which is further higher than the warning threshold, confirming the triggering of the warning.

[0114] Execute step S06, comprehensive adjustment value by , , calculate, , , , The number of heads in the image The value is adjusted to 8, while maintaining the current configuration. The final fault prediction result is pushed to the device management business system via the data service bus. The core parameters of the entire process are summarized in Table 2:

[0115] Table 2 Summary of Core Operating Parameters

[0116]

[0117] like Figure 2 As shown, in this embodiment, the utilization rate of each island compression feature dimension and the corresponding fault risk score distribution exhibit a characteristic pattern where the lubricating oil island contributes the highest risk score under low compression ratios, confirming the effectiveness of information bottleneck compression in retaining the maximum fault-related information on small-dimensional islands. Figure 3 As shown, the graph sparsification results of the cross-island physical coupling aggregation layer demonstrate the sparse coupling paths of 11 highly relevant nodes selected from 224 candidate nodes. The coupling edge weight between the oil temperature node and the rotation speed node is the highest, which is consistent with the physical causal relationship corresponding to the heat transfer equation, verifying the rationality of the physical mechanism knowledge graph constraint.

[0118] Compared to traditional methods that aggregate raw data from isolated islands and directly input it into a shallow classification model, this invention represents a fundamental advancement in principle. Traditional methods suffer from a linear increase in feature dimensionality with the number of islands, failing to fundamentally alleviate the curse of dimensionality. Information bottleneck compression, however, guarantees from an information theory perspective that the compressed features are minimum sufficient statistics, theoretically suppressing the curse of dimensionality. Traditional graph neural networks, in their information aggregation on dense graphs, cannot distinguish between physical causal paths and statistical pseudo-correlation paths. In contrast, the graph sparsity constrained by the physical mechanism knowledge graph allows attention weights to directly reflect the true physical coupling strength. The model learns physically meaningful causal links rather than spurious statistical correlations, thus ensuring generalization stability in scenarios with varying distributions. Traditional federated learning only aggregates model parameters and cannot quickly adapt to new fault types. The decoupling mechanism of slow and fast weights in a two-layer long short-term memory network, under the constraint of data non-sharing, simultaneously achieves continuous optimization of cross-island general representations and rapid adaptation to new fault types, filling the gap in meta-learning capabilities of traditional federated learning.

[0119] It should be noted that the variables involved in this invention are explained in detail in Tables 3 and 4.

[0120] Table 3. Variable Explanation Table (Part 1)

[0121]

[0122] Table 4. Variable Explanation Table (Part Two)

[0123]

[0124] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting data silos in asset big data, characterized in that, Includes the following steps: The data integration engine connects various heterogeneous data silos, automatically extracts metadata and time-series data from each silo and establishes hierarchical relationships between data source nodes, data table nodes and physical subsystem nodes through the metadata registry center, forming a multi-island heterogeneous knowledge graph. For the original features of each node in the multi-island heterogeneous knowledge graph, the multi-island optimal feature compression algorithm based on the information bottleneck principle is adopted. After compression by each island's independent encoder, the effective feature dimension is compressed to 2% to 5% of the original feature dimension through product space fusion, thus obtaining the compressed feature set of each island. The compressed feature sets of each island are input into the cross-island state representation model. After processing through three layers—local aggregation layer within an island, physical coupling aggregation layer across islands, and global readout layer—the device health status representation vector and confidence level are output. When the confidence level is lower than the confidence level threshold, the graph structure of the multi-island heterogeneous knowledge graph is updated and attention propagation is re-executed. The process is repeated for a maximum of 5 rounds until the device health status representation vector converges. The equipment failure risk score is calculated based on the cosine similarity between the equipment health status representation vector and the preset failure category standard vector set. When the equipment failure risk score exceeds the failure warning threshold, the hierarchical resource isolation scheduler is triggered to allocate a dedicated computing quota to the current prediction task and activate the incremental streaming computing mode to replace the full batch processing, keeping the prediction latency within 500ms. When a new type of equipment or a new type of fault is detected, and the number of labeled samples in the corresponding island is lower than the small sample trigger threshold, the cross-island state representation model executes no more than 5 inner loop gradient steps, updates the task-specific adaptation layer parameters, and outputs the updated equipment fault risk score. Based on three parameters—the average equipment failure risk score, the normalized value of the convergence round of the state representation iteration, and the utilization rate of the compressed feature dimension—a comprehensive adjustment value is calculated using a failure risk score adjustment function. The number of attention heads in the cross-island state representation model is dynamically adjusted according to the interval to which the comprehensive adjustment value belongs. The final failure prediction result is then output and pushed to the business system via the data service bus.

2. The data silo prediction method based on asset big data according to claim 1, characterized in that, The aforementioned multi-island heterogeneous knowledge graph specifically refers to organizing device sensor nodes, data table nodes, and physical subsystem nodes belonging to different business systems into a unified heterogeneous graph structure according to hierarchical relationships. The node features are composed of statistical quantities of the corresponding device sensor time-series data, and the edge types include three categories: belonging to, referencing, and converted from.

3. The data silo prediction method based on asset big data according to claim 2, characterized in that, The multi-island optimal feature compression algorithm based on the information bottleneck principle specifically refers to formalizing the multi-island joint feature extraction problem into an information bottleneck optimization problem. Under the constraint of maximizing the mutual information between compressed features and fault labels, it minimizes the mutual information between compressed features and original features, using Lagrange multipliers. To balance compression ratio and prediction correlation, the mutual information estimation is transformed into a differentiable variational lower bound using the variational information bottleneck method, and the optimal parameters of each isolated encoder are solved by gradient descent.

4. The data silo prediction method based on asset big data according to claim 3, characterized in that, The product space fusion specifically refers to concatenating the compressed feature vectors output by each isolated encoder along the feature dimension to form a joint compressed feature representation. The compressed feature components of each isolated island remain separable after concatenation, which is used for subsequent separation and quantification of the information contribution of each isolated island.

5. The data silo prediction method based on asset big data according to claim 4, characterized in that, The cross-island state representation model integrates two parts: a cross-island heterogeneous graph attention network and a cross-island federated meta-learning fast adaptation network. The cross-island heterogeneous graph attention network consists of three layers: a local aggregation layer within an island, a cross-island physical coupling aggregation layer, and a global readout layer. The device health status representation vector output by the global readout layer is directly used as the input sequence frame of the outer long short-term memory network in the cross-island federated meta-learning fast adaptation network.

6. The data silo prediction method based on asset big data according to claim 5, characterized in that, The cross-island physical coupling aggregation layer specifically refers to encoding the cross-island physical coupling relationship as graph sparsification constraint edges based on the heat transfer equation and electromagnetic equation prior constraints encoded in the physical mechanism knowledge graph. The graph sparsification algorithm is used to automatically identify feature subsets that are highly related to the state of the target device, and graph attention aggregation is performed on the sparsified coupling graph to capture cross-island cascade physical effects.

7. The data silo prediction method based on asset big data according to claim 6, characterized in that, The cross-island federated meta-learning fast adaptation network adopts a two-layer long short-term memory network structure driven by a meta-learner. The outer long short-term memory network encodes cross-task general temporal representations, while the inner long short-term memory network is a task-specific adaptation layer that stores fast weight parameters. During federated training, each island locally calculates the meta-gradient of the task-specific adaptation layer, and the federated server aggregates the meta-gradients of each island and updates the global meta-initialization parameters.

8. The data silo prediction method based on asset big data according to claim 7, characterized in that, The calculation of the equipment failure risk score specifically refers to calculating the cosine similarity between the equipment health status representation vector and each standard vector in the preset failure category standard vector set, taking the maximum cosine similarity value and mapping it to the interval between 0 and 1 through a linear transformation to obtain the equipment failure risk score.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores program instructions, which, when executed in a computer, are used to perform the data silo prediction method based on asset big data as described in any one of claims 1-8.

10. A data silo prediction system based on asset big data, characterized in that, The system comprises the computer-readable storage medium of claim 9, wherein the system is a computer, the computer-readable storage medium is disposed within the system, and the system is provided with a microprocessor that executes program instructions stored in the computer-readable storage medium.

Images