Unit state information fusion system and method based on multi-source heterogeneous data lake

Abstract

The application discloses a kind of unit state information fusion system and method based on multi-source heterogeneous data lake, including extracting the metadata information of multi-source heterogeneous data from data lake, construct and continuously update semantic alignment graph, and output uniform intermediate representation layer;Receive time series data and non-time series data, extract features by time series flow feature extraction network and non-time series flow feature extraction network respectively, realize bidirectional interactive fusion using cross attention mechanism, generate fusion feature sequence;Construct self-supervised loss function, and carry out incremental fine-tuning without manual annotation;Adopt multi-index drift detection mechanism to judge whether to trigger incremental update;Build hierarchical abstract index.The application systematically solves the problem of hydropower unit data fusion: dynamic semantic alignment provides standardized input for fusion, cross-modal attention improves fault identification accuracy, incremental self-supervised learning realizes unannotated continuous optimization and reduces computational cost, and explainable query shortens fault positioning time.

Description

Technical Field

[0001] This invention relates to the field of industrial big data processing and intelligent operation and maintenance technology, and in particular to a unit status information fusion system and method based on a multi-source heterogeneous data lake. Background Technology

[0002] With the continuous improvement of the digitalization and intelligence level of hydropower equipment, modern hydropower units (including mixed-flow, axial-flow, through-flow, and pumped-storage units) have accumulated extremely rich multi-source heterogeneous condition monitoring data during operation. According to industry statistics, a single large hydropower unit (such as a 700MW mixed-flow unit) can generate hundreds of gigabytes of condition data daily, while the annual data volume of a cascade hydropower station group in a river basin exceeds petabytes. Data lake technology, due to its architectural advantages of "storing everything and binding during schema retrieval," is gradually becoming a core technology solution for big data management in the hydropower industry. However, in practical engineering applications, the fusion of hydropower unit condition information based on data lakes faces the following key challenges: First, there is the semantic gap problem of multi-source heterogeneous data: the condition monitoring data of hydropower units comes from a wide range of sources and is highly heterogeneous. Common data sources include: Turbine operating data: Time-series parameters such as head, flow rate, output, efficiency, and opening degree, obtained from the computer monitoring system; Vibration swing data: X / Y swing of the upper guide, lower guide, and water guide bearings, top cover vibration, tailwater vortex pressure pulsation, sampling frequency up to several kilohertz; Temperature data: thrust bearing bearing temperature, guide bearing bearing temperature, stator core temperature, cooling water temperature, etc. Operating status data: Unit operating status (generating / phase regulation / pumping / shutdown), active / reactive power, excitation current, etc.; Unstructured data: Inspection records (text), infrared thermal imaging images, ultrasonic flaw detection waveforms, maintenance reports (PDF / document), underwater robot inspection videos, etc.

[0003] Significant semantic inconsistencies exist in the descriptions of the same physical quantity from different data sources. For example, the field name for "water guide bearing temperature" may vary across different power plants and systems, using names such as "WBrg_Temp," "water guide bearing temperature," "guide_bearing_temp," and "T_G_BRG." Furthermore, the units (Celsius / Fahrenheit), precision, and alarm threshold definitions also differ. Traditional data integration methods rely on manually creating mapping tables. However, in a data lake environment, with unit capacity expansion, sensor upgrades, and monitoring system updates, field patterns continuously evolve, making static mapping tables unable to adapt dynamically. This results in a large amount of data becoming "dark data," which is difficult to effectively integrate and utilize.

[0004] Secondly, there is the challenge of cross-modal fusion of hydraulic, mechanical, and electrical multi-physics fields: the fault evolution of hydropower units involves the coupling effects of three physical fields: hydraulic, mechanical, and electrical. A typical example is the pressure pulsation caused by the tailrace eddy, which not only generates low-frequency pressure fluctuations in the hydraulic channel but also exacerbates the unit's vibration swing, potentially leading to power fluctuations. This means that data from a single mode cannot comprehensively characterize the unit's health status. However, the time scales and physical characteristics of data from different modes vary significantly. Timing signals (vibration swing, pressure pulsation, power, etc.): sampling frequencies range from 0.1Hz to 5000Hz, exhibiting high frequency and strong time dependence; Non-time-series data (inspection text, infrared images, maintenance records): characterized by being discrete, sparse, and semantically rich.

[0005] Existing methods often employ simple feature concatenation or weighted averaging, failing to capture the deep physical correlations between time-series signals and unstructured data. For example, there is a strong correlation between low-frequency pressure pulsations (hydraulic factors) caused by changes in head / flow rate and "abnormal noise in the tailrace area" in inspection texts, but traditional fusion methods struggle to establish such cross-modal physical-semantic alignment.

[0006] Next is the issue of concept drift and continuous model updates under complex hydraulic conditions: Hydropower units exhibit significant multi-condition and variable head operation characteristics. Taking pumped storage units as an example, their operating conditions cover multiple modes such as power generation, pumping, phase regulation, and start-up / shutdown, with head variations reaching tens of meters, leading to continuous changes in data distribution (concept drift). Specifically, this manifests as follows: Seasonal drift: The distribution of water head and flow rate differs between the wet and dry seasons, resulting in differences in the vibration characteristics of the generating units; Operating condition switching drift: When switching from power generation mode to pumping mode, the direction of hydraulic excitation force changes and the vibration characteristic distribution changes abruptly. Aging drift: As the service life increases, the bearing clearance increases and the blades wear, and the characteristic distribution baseline under normal conditions gradually shifts.

[0007] Existing fusion models mostly employ offline batch training, resulting in fixed model parameters once deployed, making them unable to adapt to changes in data distribution. Full retraining, on the other hand, faces the dual challenges of high computational resource consumption (a single retraining session can take several days) and high annotation costs (due to the scarcity of fault samples under new operating conditions). While some incremental learning methods attempt online updates, they typically rely on large amounts of labeled data and are prone to catastrophic forgetting—that is, forgetting old operating condition features after learning new hydraulic operating conditions.

[0008] Finally, there is the rigid requirement for interpretability of the fusion results in hydropower unit operation and maintenance: the hydropower industry has extremely high requirements for equipment safety and the reliability of operation and maintenance decisions. When the fusion system outputs a judgment of "abnormal vibration caused by the tailrace vortex," the operation and maintenance engineer must know: Which sensor's vibration swing data is out of range? Which inspection text recorded the abnormal sound? Which area in the infrared thermogram shows abnormal temperature? Does the current head / flow condition exacerbate the development of the vortex? Lack of interpretability will lead to the following problems: The maintenance team was unable to pinpoint the root cause of the malfunction, delaying the best time to address it. False alarms cannot be quickly confirmed, causing unnecessary downtime for inspection; It is difficult to accumulate reusable diagnostic knowledge, and experience cannot be passed on.

[0009] According to surveys and statistics in the hydropower industry, over 40% of cases saw maintenance personnel refuse to adopt intelligent diagnostic suggestions due to the inability to interpret the fusion model, severely hindering the implementation of smart hydropower construction.

[0010] To address the aforementioned issues, several technical solutions have been explored. For example, Chinese Patent CN121259986A discloses a comprehensive alarm classification processing method and platform based on multi-source heterogeneous data fusion. This method organizes and fuses multi-source data by constructing a data asset graph. However, this solution uses a static knowledge graph, which cannot adapt to pattern evolution in a data lake environment. Chinese Patent CN121351019A discloses a multimodal feature encoding and cross-modal adaptive fusion method that uses a knowledge graph to guide semantic alignment. However, its alignment process still relies on a manually constructed static graph and does not address the robustness issue when sensors are missing. The academic paper "Residual Perception Self-Attention Deep Fusion Network for IIoT Device State Recognition" (published in the journal Internet of Things, 2025) proposed a cross-modal attention fusion network, which achieved an accuracy of 95.24% on the PHM2010 dataset. However, this scheme was only validated in a laboratory environment and did not take into account the practical constraints of industrial data lakes, such as large scale, dynamic evolution, and scarce annotations.

[0011] In summary, existing technologies have significant shortcomings in the following aspects: (1) the semantic alignment method is static and cannot adapt to the evolution of data lake models; (2) multimodal fusion lacks robust design for missing sensors; (3) model updates rely on labeled data and are prone to forgetting old knowledge; (4) the fusion results lack a traceable and verifiable interpretation mechanism. Therefore, there is an urgent need for a unit status information fusion system and method that can systematically solve the above problems. Summary of the Invention

[0012] The technical problem solved by this invention is to provide a system and method for fusing unit status information based on a multi-source heterogeneous data lake.

[0013] This application provides a method for fusing unit status information based on a multi-source heterogeneous data lake, comprising the following steps: S1: Dynamic Semantic Alignment Metadata information of multi-source heterogeneous data is extracted from the data lake, a semantic alignment graph is constructed and continuously updated. The semantic alignment graph uses data lake fields as nodes and semantic equivalence relationships between fields as edges. The semantic evolution history is recorded through a version graph database, and the fields of each data source are mapped to a unified state ontology through an incremental graph neural network, outputting a unified intermediate representation layer. S2: Multimodal Joint Attention Fusion The multimodal joint attention fusion engine receives time-series and non-time-series data after normalization processing by the unified intermediate representation layer. Features are extracted by the time-series and non-time-series feature extraction networks within the engine, respectively. A cross-attention mechanism is then used to achieve bidirectional interactive fusion of time-series and non-time-series features, generating a fused feature sequence. The fused feature sequence is then pooled along the time dimension to obtain the global fused feature vector for each time window. During the training phase of the multimodal joint attention fusion engine, a masked modality training strategy is adopted to randomly mask some modality data to enhance robustness to sensor omissions. S3: Incremental Self-Supervised Learning Update The system receives the fused feature sequence generated in step S2, constructs a self-supervised loss function through a time-series comparison prediction task, and incrementally fine-tunes the multimodal joint attention fusion engine without manual annotation. At the same time, it adopts an elastic weight consolidation mechanism to prevent catastrophic forgetting and a multi-index drift detection mechanism to determine whether incremental updates are triggered. S4: Interpretable query A hierarchical summary index is constructed, comprising a first-level index for storing coarse-grained time-component-state information, a second-level index for storing the global fusion feature vector after time pooling for each time window, and a third-level index for storing attention weights and the original data path. In response to natural language queries, relevant fusion feature fragments are located through multi-level retrieval, corresponding attention weights are extracted to form an evidence chain, and interpretable query results are generated.

[0014] This application also provides a unit status information fusion system based on a multi-source heterogeneous data lake, comprising the following parts: The dynamic semantic alignment module is configured to: extract metadata information from multi-source heterogeneous data in the data lake, construct and continuously update the semantic alignment graph, and map the fields of each data source to a unified state ontology through an incremental graph neural network, and output a unified intermediate representation layer. The multimodal joint attention fusion engine, connected to the dynamic semantic alignment module, is configured to: receive time-series and non-time-series data after normalization processing by the unified intermediate representation layer; extract features through the time-series and non-time-series feature extraction networks respectively; then use a cross-attention mechanism to achieve bidirectional interactive fusion, generate a fused feature sequence, and obtain a global fused feature vector; wherein, a masked modality training strategy is adopted during the training phase of the multimodal joint attention fusion engine; The incremental self-supervised learning module is connected to the dynamic semantic alignment module and the multimodal joint attention fusion engine, respectively. It is configured to receive the fused feature sequence output by the multimodal joint attention fusion engine, and perform incremental fine-tuning of the multimodal joint attention fusion engine without manual annotation. At the same time, it adopts an elastic weight consolidation mechanism to prevent catastrophic forgetting and adopts a multi-index drift detection mechanism to determine whether to trigger incremental update. An interpretable query interface is connected to the dynamic semantic alignment module, the multimodal joint attention fusion engine, and the incremental self-supervised learning module, respectively. It is configured to: construct a hierarchical summary index, which stores time-component-state coarse-grained information, global fusion feature vectors, attention weights, and original data paths; respond to natural language queries, extract the corresponding attention weights to form an evidence chain, and generate interpretable query results for visualization.

[0015] Compared with existing technologies, the unified intermediate representation layer constructed in the dynamic semantic alignment step of this invention provides standardized input for the multimodal joint attention fusion step, enabling the fusion engine to focus on deep correlation learning across the hydraulic, mechanical, and electrical physical fields without processing differences in the original fields. This significantly reduces the number of model parameters while greatly improving fault identification accuracy. The fused feature sequence output by the multimodal joint attention fusion step also serves as training data for the incremental self-supervised learning update step, allowing the time-series comparison prediction task to achieve label-free continuous model optimization without additional feature engineering, thus reducing the computational cost of incremental updates. The multimodal joint attention fusion step also incorporates cross-attention. The attention weight matrix generated by the mechanism directly serves as evidence for interpretable query steps, giving the system a "naturally interpretable" capability. Without requiring additional interpretive models, it can trace the original sensor data, text records, or image regions upon which each judgment is based through hierarchical summary indexing. This reduces the time for maintenance personnel to understand the fusion results from an average of 15 minutes to less than 1 minute, achieving zero-illusion output in 1000 random queries. Simultaneously, the drift detection mechanism in the incremental self-supervised learning update step can perceive changes in field semantics during the dynamic semantic alignment step and trigger incremental updates to the semantic graph, forming an end-to-end closed loop of "data layer change → model layer adaptation → query layer interpretability." These synergistic effects reduce the manual cost of semantic alignment by 80%, increase the data availability of the data lake from less than 60% to over 95%, reduce the accuracy of old operating condition fusion by no more than 3% after 20 consecutive incremental updates, and shorten the fault root cause location time by more than 60%. This comprehensively constructs a new intelligent operation and maintenance paradigm of "data-driven - model adaptation - human-machine collaboration" for the complex operating environment of hydropower units. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention.

[0017] Figure 1 This is a flowchart of the unit status information fusion method based on a multi-source heterogeneous data lake according to the present invention; Figure 2 This is a flowchart of step S1 of the present invention; Figure 3 This is a flowchart of step S3 of the present invention; Figure 4 This is a structural block diagram of the unit status information fusion system based on a multi-source heterogeneous data lake according to the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0019] Example 1 like Figure 1 As shown, this invention provides a method for fusing unit status information based on a multi-source heterogeneous data lake, comprising the following steps: S1: Dynamic Semantic Alignment Metadata information from multi-source heterogeneous data is extracted from the data lake, and a semantic alignment graph is constructed and continuously updated. The semantic alignment graph uses data lake fields as nodes and semantic equivalence relationships between fields as edges. The semantic evolution history is recorded through a version graph database, and the fields of each data source are mapped to a unified state ontology through an incremental graph neural network, outputting a unified intermediate representation layer. Specifically, such as Figure 2 As shown, step S1 specifically includes the following steps: S11: Extract field names, field comments, enumeration value lists, and data types from the metadata of the data lake (Apache Iceberg or Apache Hudi can be used). Use a pre-trained language model (such as all-MiniLM-L6-v2) to concatenate the above information into a text description and encode it into a field feature vector, denoted as:

[0020] in, This represents the feature vector of the i-th field. This is the text description after concatenating the i-th field. For the encoding function of the pre-trained language model, In this embodiment, the feature vector dimension is taken as... ; express A 3D real vector space, where the index i is the field index, i = 1, 2, 3, ... ,in, This represents the total number of fields in the data lake.

[0021] S12: Construct the initial semantic alignment graph ,in For a collection of field nodes, For a set of semantically equivalent edges, Calculate any two field nodes and Cosine similarity between

[0022] in, Let be the similarity between node i and node j, with a value range of [-1, 1]. It is the dot product of two vectors; For vectors The Euclidean norm.

[0023] when When, add an edge between nodes i and j, with an edge weight of . ,in In this embodiment, a preset similarity threshold is used. Simultaneously, a set of predefined standard ontology nodes (such as "water guide bearing temperature", "upper guide sway", "water head", "flow rate", etc.) is introduced. Establish initial connections with each field node.

[0024] S13: An embedded learning process for the semantic alignment graph is performed using a relational graph convolutional network (R-GCN). Node i is at the... Layer embedding vector The updated formula is:

[0025] in, This refers to the layer index of a graph neural network. , The total number of network layers is taken as in this embodiment. ; For node i at the th Layer embedding vectors, Initialize to The vector after linear projection; R is a set of relation types (such as synonyms, parent-child, and unit conversions), and r is a relation type variable. , indicating from node To the node The specific relationship type of the edges; Let i be the set of neighboring nodes of node i under relation r, that is, there exists a node of type r. The edge from node Pointing to node All nodes A set; = The normalization function is equal to the node. In relationship The number of neighbors; For relation r in the th case Layer weight matrix; It is the weight matrix for self-connections (i.e., nodes pointing to themselves); For the activation function, this embodiment uses ReLU, defined as follows: ; For the hidden layer dimension, in this embodiment, we take... ; Neighboring nodes In the Layer embedding vector.

[0026] The training objective uses a contrastive loss function to make the embeddings of semantically connected nodes closer together and the embeddings of unconnected nodes farther apart:

[0027] in, The cosine similarity function is used. It is an exponential function. For nodes The final embedding vector, For temperature parameters, in this embodiment, we take... , Let k be the set of all neighboring nodes of node i (including various relation types); k is the index of the negative sample node. express Not a node The neighbors.

[0028] S14: When a new field During the integration process, local retraining is performed only on the new field node and its neighborhood, calculating the similarity between the new node and all existing nodes. Add edge weights The connection is fixed. Only the embeddings of new nodes and their neighboring nodes are updated, and the timestamp of each update is recorded. Store in a versioned graph database; S15: Output a unified intermediate representation layer, for each field i Mapping to standard ontologies o The confidence level is calculated by the attention mechanism:

[0029] in, The normalized exponential function transforms any real vector into a probability distribution, and its i-th... The definition of each component is ; Let be the embedding vector of the standard ontology node o; o is the standard ontology node variable. ; To sum the index variable, traverse all standard ontology nodes; For fields Mapping to standard ontologies The confidence level satisfies .

[0030] The standard ontology with the highest confidence level is selected as the mapping result:

[0031] in, A function that maximizes the set of points of the argument variable and returns a value that maximizes the set of points of the argument variable. To obtain the maximum value value; For fields The final mapping to the standard ontology.

[0032] The final output format is ,in, A timestamp is used to establish or update the mapping. Downstream systems use this timestamp to convert the raw data. Five-tuple format.

[0033] S2: Multimodal Joint Attention Fusion The multimodal joint attention fusion engine receives time-series and non-time-series data after standardized processing by a unified intermediate representation layer. Features are extracted by the time-series and non-time-series feature extraction networks within the engine, respectively. A cross-attention mechanism is then used to achieve bidirectional interactive fusion of time-series and non-time-series features, generating a fused feature sequence. The fused feature sequence is then pooled along the time dimension to obtain the global fused feature vector for each time window. During the training phase of the multimodal joint attention fusion engine, a masked modality training strategy is employed, randomly masking some modal data to enhance robustness against sensor omissions. Specifically, step S2 includes the following steps: S21: Temporal Stream Feature Extraction. Input multivariate temporal data. Where T is the number of time steps within the time window, and C is the number of sensor channels (including head, flow rate, upper guide sway, lower guide sway, water guide sway, top cover vibration, pressure pulsation, active power, etc.). The first layer uses a deformable convolutional network (DCN) to handle the variable sampling rate problem; the second layer uses a temporal convolutional network (TCN) to output a temporal feature map.

[0034] In this context, DCN(⋅) is a deformable convolutional network function that adaptively processes data with varying sampling rates by learning offsets, and TCN(⋅) is a temporal convolutional network function that uses dilated convolutions to expand the receptive field. As a time-series feature dimension, this embodiment takes... .

[0035] S22: Non-temporal stream feature extraction. For text data, the DistilBERT pre-trained language model is used to extract sentence vectors, which are then reduced to 128 dimensions through a fully connected layer, denoted as... For infrared thermal imaging images, a lightweight MobileNetV3 convolutional neural network is used to extract global features, which are also mapped to 128 dimensions, denoted as... If multiple non-time-series data exist simultaneously, extract them separately and then use self-attention pooling to obtain a single non-time-series feature vector:

[0036] in, For the feature vector of text data, Given the feature vectors of the image data, Pooling(⋅) is the self-attention pooling function that merges a variable number of input vectors into a single output vector. As a non-temporal feature dimension, this embodiment takes .

[0037] S23: Cross-attention fusion. Define the cross-attention function:

[0038] in, For query matrix; The key matrix; It is a value matrix; This is the transpose of the key matrix. The dimension of the key vector; This is a normalized exponential function, applied independently to each row of the matrix; The output dimension is .

[0039] First attention (chronological → non-chronological):

[0040]

[0041] in, The query matrix is ​​generated from time-series features; The key matrix is ​​generated from non-temporal features; This is a value matrix generated from non-temporal features; , , The weight matrix is ​​a learnable matrix; The output of the first attention.

[0042] Second attention (non-temporal → temporal):

[0043]

[0044] in, The query matrix is ​​generated from non-time-series features; The key matrix is ​​generated from temporal features; This is a value matrix generated from time-series features; , , The weight matrix is ​​a learnable matrix; is the output of the second attention; d is the attention hiding dimension, which is set to 128 in this embodiment. The two attention outputs are concatenated and then fed forward to generate a fused feature vector:

[0045] in, This is a vector concatenation function that concatenates vectors along the feature dimensions. For a repeating function, Repeated T times along the time dimension, we get The matrix; FFN is a feedforward neural network containing two linear layers and a ReLU activation function; Z To fuse the feature matrix, its first... Travelogue ; To fuse feature dimensions, this embodiment .

[0046] S24: Masked modality training. During the training phase, probabilistic... Randomly set the input features of a certain modality to zero vectors, and simultaneously input a mask vector into the model. Indicates the currently available modalities. The model learns to output stable fused features even under conditions where any modality is missing; S25: Temporal Pooling and Output. Adaptive average pooling is performed on the fused feature vector Z along the temporal dimension to obtain the global fused feature vector:

[0047] in, Let Z be the t-th row (i.e., the fused feature vector at time t); t is the time step index. .

[0048] At the same time, the original fusion feature sequence is preserved. For use in step S3.

[0049] S3: Incremental Self-Supervised Learning Update The system receives the fused feature sequence generated in step S2, constructs a self-supervised loss function through a time-series comparison prediction task, and incrementally fine-tunes the multimodal joint attention fusion engine without manual annotation. At the same time, it adopts an elastic weight consolidation mechanism to prevent catastrophic forgetting and a multi-index drift detection mechanism to determine whether incremental updates are triggered. Specifically, such as Figure 3 As shown, step S3 specifically includes the following steps: S31: Time-series comparison prediction task: Define the prediction head network (Small-scale MLP), using the fused feature vector at the current time t As input, predict the future and fuse the feature vector after a preset step size Δ. The self-supervised loss function is:

[0050] Where N is the batch size, i.e., the number of samples used in one training session; and n is the sample index within the batch. ; For the first One sample in The fused feature vector at each time step; for One sample in The fused feature vector at any given time; adaptively adjusted according to the data sampling frequency. In this embodiment, the vibration data is taken as... Operating condition data acquisition ; It is the Euclidean norm (L2 norm); This is the prediction head network function.

[0051] S32: Elastic Weight Consolidation (EWC) Mechanism: Calculates the parameters of the multimodal joint attention fusion engine under historical tasks. Diagonal elements of the Fisher information matrix:

[0052] in, This is a vector of all parameters of the multimodal joint attention fusion engine. , This represents the total number of parameters. For the multimodal joint attention fusion engine One parameter, ; Distribution of historical data; For input data variables, express Follows the historical data distribution; To output label variables; Given input and parameters Time-based multimodal joint attention fusion engine output The conditional probability; Let be the expectation operator with respect to the distribution of historical data; For the Fisher information matrix Diagonal elements, measuring parameters The importance of historical missions.

[0053] The total loss function during incremental fine-tuning is:

[0054] in, Supervised loss is used (if labeled data is available); otherwise, take... ; For self-monitoring losses; As the weight for self-supervised loss, in this embodiment, we take... ; For the first incremental update The values ​​of the parameters; The EWC regularization coefficient is taken in this embodiment. .

[0055] S33: Multi-index drift detection: Simultaneous detection of three drift indices: Maximum mean difference (MMD):

[0056] in For kernel mapping, data is mapped to the regenerating kernel Hilbert space. ; For the new batch of data set, A collection of historical data; This represents the number of samples in the new batch of data. The number of samples for historical data; The norm in the reproducing kernel Hilbert space; threshold Dynamically adjusted ; Changes in classifier confidence: threshold ; in, For the multimodal joint attention fusion engine, the prediction confidence random variable on the new batch of data; This represents the prediction confidence random variable of the multimodal joint attention fusion engine on historical data.

[0057] Feature spatial density variation: threshold ; in, For new batches of data in fusion features The kernel density estimate at that location; For historical data in feature fusion The estimated kernel density at that location.

[0058] A voting mechanism is used to trigger an incremental update when at least two metrics exceed their corresponding thresholds. S34: Checkpoint rollback mechanism: Save a snapshot of the multimodal joint attention fusion engine before each incremental update. After the update, on a small-scale validation set On a test sample (including historically representative samples), calculate the change in accuracy for older tasks:

[0059] in, This is the parameter vector of the multimodal joint attention fusion engine before incremental update; This is the incrementally updated parameter vector of the multimodal joint attention fusion engine; like (This embodiment takes) If the update fails, the user will reject the update and roll back to the previous version. .

[0060] S35: Incremental fine-tuning scheduling. Only applies to new batches of data. and playback buffer samples Training epochs, of which The number of epochs for incremental training is taken in this embodiment. ; epoch is the unit of training rounds, one epoch means traversing all training data once; Learning rate set to ,in In this embodiment, the initial learning rate is set to... ; The learning rate used for incremental training. Only update the parameters of the last two layers of the fusion engine and the prediction head, and freeze the rest of the parameters.

[0061] S4: Interpretable query A hierarchical summary index is constructed, which includes a first-level index for storing coarse-grained information on time, component, and state; a second-level index for storing the global fusion feature vector after time pooling for each time window; and a third-level index for storing attention weights and the original data path. In response to natural language queries, relevant fusion feature fragments are located through multi-level retrieval, the corresponding attention weights are extracted to form an evidence chain, and interpretable query results are generated.

[0062] Specifically, step S4 includes the following steps: S41: Construct a hierarchical summary index.

[0063] First-level index: by time window length Divide the data into seconds, and record the unit component identifier p, the comprehensive status label s (normal / attention / warning / alarm), and the semantic hash value LSH ( ) within each window. Locality-sensitive hashing (LSH) is used for storage. Second-level index: Stores the global fusion feature vector after time pooling for each time window. Using the HNSW approximate nearest neighbor index structure, the Euclidean distance between the query vector q and the index vector is calculated using the retrieval algorithm.

[0064] Third-level index: Stores the attention weight matrix generated by the cross-attention mechanism within each time window. (The attention weight matrix of temporal features to non-temporal features) and The K positions with the highest weights in the (weight matrix of non-temporal features on temporal features) and their weight values And the storage path of the original data corresponding to the location in the data lake. .

[0065] S42: A three-level cache architecture. Defines the residency time for each cache level. Hour, Hour, The query response uses a progressive display strategy.

[0066] S43: Natural Language Query Processing. The user inputs the query text. It is then converted into a query vector q using the same encoder as in step S11. In the second-level index, retrieve the k=3 most similar fusion feature windows:

[0067] Where W is the set of all time windows, For window index variables, This indicates a specific time window; For window The global fusion feature vector; These are the three windows in the search results that are most similar to the query vector.

[0068] S44: Evidence Chain Extraction. For each retrieved window w, read the top-K positions of the attention weights from the third-level index. Where K=10, the evidence list is obtained by reverse mapping. ;in: For the first in the fusion feature sequence At a certain point in time, ; is the column index in the attention weight matrix, indicating the th of the fused feature vector. One dimension, ; For position Attention weights at each location, range of values This indicates the degree to which the position contributes to the model's judgment; This is the original data lake storage path.

[0069] S45: Generate interpretable query results. For common queries, a structured template generation method is used; for complex queries, a restricted large language model generation method is used, with evidence-priority prompts constraining the generation range and setting temperature parameters. .

[0070] Example 2 This invention provides a unit status information fusion system based on a multi-source heterogeneous data lake. The data lake, built using Apache Iceberg, is deployed at a large hydropower station (referencing the configuration of the Longtan Hydropower Station), and includes the following data from a 700MW mixed-flow hydropower unit: Turbine operating data: head (H), flow rate (Q), output (P), efficiency (η), guide vane opening (Y), sampling frequency 1Hz, from computer monitoring system; Vibration swing data: upper guide swing in X / Y direction, lower guide swing in X / Y direction, water guide swing in X / Y direction, top cover vibration, tailwater pipe pressure pulsation, sampling frequency 2560Hz, from vibration monitoring system; Temperature data: thrust bearing pad temperature (8 points), upper guide pad temperature (4 points), lower guide pad temperature (4 points), water guide pad temperature (4 points), cooling water temperature, sampling frequency 0.1Hz; Unstructured data: Daily inspection text records (approximately 500 words), infrared thermal imaging images (collected every 15 minutes), and monthly maintenance reports in PDF format.

[0071] like Figure 4 As shown, the system includes the following parts: The data access module is configured to store the above data in its original format into the original area of ​​the data lake, retaining the complete timestamp and source identifier; The dynamic semantic alignment module is configured to: extract metadata information from multi-source heterogeneous data in the data lake, construct and continuously update the semantic alignment graph, and map the fields of each data source to a unified state ontology through an incremental graph neural network, and output a unified intermediate representation layer. Specifically, JanusGraph was deployed as a versioned graph database. The initial version constructed a semantic alignment graph containing approximately 500 field nodes, covering all sensor fields. The system performs incremental update detection every 24 hours. When a new field is discovered (such as a newly installed ultrasonic flow measuring device), the similarity between that field and existing nodes is automatically calculated and a semantic connection is established. The entire process requires no manual intervention. Validated on a simulation dataset, the field alignment accuracy reached 88.5%, exceeding the manual mapping scheme by approximately 12 percentage points.

[0072] The multimodal joint attention fusion engine, connected to the dynamic semantic alignment module, is configured to: receive temporal and non-temporal data after normalization processing by a unified intermediate representation layer; extract features through temporal and non-temporal flow feature extraction networks respectively; and then use a cross-attention mechanism to achieve bidirectional interactive fusion, generating a fused feature sequence and obtaining a global fused feature vector; wherein, a masked modality training strategy is adopted during the training phase of the multimodal joint attention fusion engine; Specifically, it is deployed on a GPU server (NVIDIA A10), with a time-series stream window size T=256 (corresponding to approximately 0.1 seconds of vibration data and approximately 256 seconds of operating condition data, unified to the time axis through interpolation), and outputs fused feature dimensions. Mask modality training with By randomly masking vibration data or non-time-series data, and training for 3 epochs, the fusion accuracy decreased by only 3.8% in the test simulating the failure of the water bearing temperature sensor, which is far better than the 31% decrease when training without masking (comparative data from ablation experiments).

[0073] The incremental self-supervised learning module is connected to the dynamic semantic alignment module and the multimodal joint attention fusion engine, respectively. It is configured to receive the fused feature sequence output by the multimodal joint attention fusion engine and perform incremental fine-tuning of the multimodal joint attention fusion engine without manual annotation. At the same time, it adopts an elastic weight consolidation mechanism to prevent catastrophic forgetting and a multi-index drift detection mechanism to determine whether to trigger incremental updates. Specifically, this module triggers drift detection daily at 00:00. Taking the actual operation of a pumped storage unit (referencing the configuration of the Huizhou Pumped Storage Hydropower Plant) during the transition from the wet season to the dry season (approximately September) as an example, the Maximum Mean Difference (MMD) index increased from 0.02 to 0.07, with a confidence level change of 12%, triggering an incremental update. The system completed fine-tuning within 3 epochs, taking approximately 15 minutes (compared to 6 hours for full retraining). The fusion accuracy of the model trained during the wet season for dry season data improved from 78% to 89%.

[0074] An interpretable query interface is connected to the dynamic semantic alignment module, the multimodal joint attention fusion engine, and the incremental self-supervised learning module, respectively. It is configured to: build a hierarchical summary index, which stores time-part-state coarse-grained information, global fusion feature vectors, attention weights, and original data paths; respond to natural language queries, extract the corresponding attention weights to form an evidence chain, and generate interpretable query results and visualize them. Specifically, this module is deployed on a web server, providing a RESTful API and a front-end interface. When maintenance personnel query "the cause of the abnormal water conductance swing at 10:23:05 on September 15, 2025," the system returns a coarse-grained judgment within 85ms ("increased water conductance swing caused by tailrace vortex"), and an evidence summary within 220ms (water conductance swing amplitude in the X direction is 0.38mm, contributing 41%; tailrace pipe pressure pulsation frequency is 0.8 times the rotational frequency, contributing 32%; the daily inspection record mentions "abnormal noise from the tailrace pipe," contributing 18%). Clicking the link opens the original vibration waveform file. Simulated user testing (with 35 engineers participating) showed that the adoption rate of the system output by maintenance personnel increased from that of traditional methods. Upgraded to User satisfaction rating: .

[0075] Example 3 This embodiment illustrates the above methods or modules using specific examples. 1. The specific implementation process of dynamic semantic alignment graph 1) Field Characterization. The field "W_Brg_Temp" (water guide bearing temperature, in °C) for a specific hydroelectric power station was extracted from Iceberg metadata. It has no enumerated values, is of data type FLOAT, and was encoded into a 768-dimensional feature vector using all-MiniLM-L6-v2. Another field from the same power station, "water guide bearing temperature_1#", was encoded in the same way.

[0076] 2) Graph construction. The cosine similarity between two field vectors is 0.93, exceeding the threshold. =0.7, semantic equivalent edges are established. Meanwhile, the initial similarity between the system pre-configured standard ontology node "water guide bearing temperature" and the above two fields is 0.89 and 0.91 respectively, and connections are also established.

[0077] 3): GNN embedding learning. After 2 layers of R-GCN propagation, field nodes of the same physical quantity aggregate in the embedding space, and nodes of different physical quantities are separated. After training is completed, the mapping confidence between "W_Brg_Temp" and the standard ontology "water guide bearing temperature" is 0.96.

[0078] 4) Dynamic update. After a certain unit modification, a new sensor field "W_Brg_Temp_New" is added. The system automatically calculates its similarity with existing nodes: the similarity with "W_Brg_Temp" is 0.95, the similarity with "water guide pad temperature_1#" is 0.91, and the similarity with the standard ontology is 0.90. The system automatically establishes connections, performs local retraining to update the new node embedding, and the entire process takes about 2 seconds. The versioned graph database records this update time as 14:23:05 on June 15, 2025.

[0079] 5) Output unified representation. The system outputs mapping table entries:

[0080] After the downstream system reads this table, it automatically converts the raw data into a unified format, such as (component="water guide bearing", attribute="pad temperature", value=62.5, timestamp=2025-06-15 14:30:00, confidence=0.95).

[0081] 2. Specific implementation of multimodal joint attention fusion 1) Network structure: Time series stream: Deformable convolution layer (kernel_size=3, offset_lr=0.1) → temporal convolutional network (4 layers of dilated convolution, dilation=1,2,4,8) → output F t ∈R 256×128 Non-time series stream: DistilBERT (6 layers, 768 dimensions) → fully connected layer FC (768→128); MobileNetV3 (input 224×224×3) → fully connected layer (1280→128) Cross-attention module: embedding dimension , number of attention heads num_heads=8 Feed-forward network: FC(256→128)+ReLU+Dropout(0.1) 2) Training configuration: Dataset: Six months of operational data from a hydropower station, containing 1200 labeled samples of three types of faults (tailrace vortex, bearing wear, and rotor imbalance). Optimizer: AdamW, initial learning rate

[0082] Number of training rounds: E =50 epochs Mask training probability:

[0083] Batch size:

[0084] 3) Training results: Fusion accuracy: 94.2% (compared to 79.8% for traditional feature stitching methods) The accuracy of single-modal missing data decreased by 4.1% after masked training (compared to a decrease of 31.2% without masked training). Inference time: 12ms per forward pass (GPU A10) 4) Case Demonstration: Inputting data for a specific operating condition, with a head of 80m and guide vane opening of 65%, a 0.8 harmonic component (characteristic frequency of the wake vortex band) was detected in the vibration data. Cross-attention weighting shows that the pressure pulsation channel (0.8 harmonic amplitude) in the time-series flow has a weight of 0.42 for the inspection text "wake pipe abnormal noise" in the non-time-series flow, while the non-time-series flow has a weight of 0.38 for this time-series channel. The fusion engine outputs a probability of 0.89 for "wake vortex band".

[0085] 3. Specific implementation of incremental self-supervised learning 1) Time series comparison and prediction task: Prediction step size Δ: For vibration data, Δ = 50 (corresponding to approximately 0.02 seconds); for operating condition data, Δ = 10 (corresponding to 10 seconds). Prediction head network: Linear(128→64) + ReLU + Linear(64→128) Validation: When trained on 200 hours of unlabeled data, the predicted MSE converged to 0.023.

[0086] 2) Drift detection parameters: Maximum mean difference (MMD):

[0087] MMD threshold: A dual-window strategy is adopted, with a short window (1 hour). =0.05, long window (24 hours) =0.03 Confidence change threshold:

[0088] Density change threshold:

[0089] 3) Triggering Case: Taking the pumped storage unit of Huizhou Pumped Storage Hydropower Plant as an example, when switching from power generation mode to pumping mode: MMD increased from 0.02 to 0.09 (exceeding 0.05). Changes in classifier confidence

[0090] density ratio

[0091] Two of the three indicators (MMD and density) exceeding the threshold → triggers incremental update 4) Incremental fine-tuning results: Number of training rounds: Learning rate

[0092] Time: 15 minutes vs. 6 hours of full retraining (based on GPU A10 test) Accuracy decreased under old operating conditions (power generation conditions): (Decrease of 2.1%), meeting the rollback threshold.

[0093] Accuracy improved under the new operating condition (pumping condition): from 76% to 88%. 5) Checkpoint rollback case: After an incremental update, the accuracy of the old test case on the validation set dropped from 94% to 86%, and the accuracy changed;

[0094] because The system automatically rejects the update and rolls back to the previous version, while simultaneously sending an alert to the operations engineer to check the quality of the training data.

[0095] 4. The specific implementation of the interpretable query interface. 1) Index building: First layer: LSH index, 128-bit hash, stored in HBase, table structure (rowkey = time window + component, column = status label + hash value). Second layer: FAISS HNSW index, parameter M=16 ef construction =200, storing 1 million global fusion feature vectors, index file size approximately 1.2GB The third layer stores the top-K attention weights, where K=10, each window is approximately 2KB, and the annual data volume is approximately 7GB (calculated based on a 5-minute window). 2) Three-level cache test (based on simulated query load test): L1 cache hit rate: 43%, average latency: 8ms L2 cache hit rate: 31% Average latency: 45ms L3 cache hit rate: 26%, average latency: 185ms Overall average response time: 98ms (initial response) → 189ms (complete chain of evidence response) 3) Query example: The user entered: "Why does the water guide pendulum increase at 10:23:05?" System processing flow: (1) Query encoding: The query text is converted into a query vector using the same encoder as in step S11.

[0096] (2) Second-level index retrieval: Calculate the Euclidean distance between the query vector and the index vector:

[0097] Search results: top-1 window index

[0098] (3) Third-level index reading: Read the attention weight top-of-minimum of this window. Location set The top 5 are:

[0099]

[0100]

[0101]

[0102]

[0103] (4) Reverse mapping: Water conductance oscillation in the X direction: value x = 0.38 mm Inspection report excerpt: "At 10:20, an unusual noise was detected in the tailpipe area during the inspection." (5) Generate output: Judgment basis The model determines that "increased water conductance swing" is a "wake vortex warning" state.

[0104] Main basis: 1. The X-axis amplitude of the water conductance swing is 0.38 mm, exceeding the threshold of 0.30 mm, contributing 41%. 2. Pressure pulsations in the tailrace pipe exhibit a 0.8 times frequency component (vortex band characteristic frequency), contributing 32%. 3. The inspection record (10:20) mentions "abnormal noise from the tailpipe," contributing 15%. Detailed data: Vibration waveform: / lake / vibration / 2025-09-15 / 10-23-05.csv Inspection log: / lake / text_logs / 2025-09-15.txt The following describes the complete deployment of this invention in a cascade hydropower station group in a certain river basin: 1) Hardware configuration: Data lake cluster: 5 nodes, each with 64 CPU cores, 256GB RAM, and 100TB NVMe SSD, running Apache Iceberg + Trino. Converged computing cluster: 3 nodes, each with 2×NVIDIA A10 GPUs, running PyTorch+FAISS Web server: 2 nodes, 32-core CPU, 64GB RAM, running FastAPI + Nginx 2) Software environment: Operating System: Ubuntu 22.04 LTS Data Lake Engine: Apache Iceberg 1.4.0 Graph database: JanusGraph 0.6.0 Deep learning framework: PyTorch 2.0.0 Vector Index: FAISS 1.7.4 Front-end framework: Vue.js + ECharts 3) Simulation deployment and operation indicators (based on backtesting of historical data from hydropower units): Field alignment accuracy: 88.5% → improved to 94.2% after manual verification. Fault identification accuracy: tailrace vortex 92.3%, bearing wear 89.7%, rotor imbalance 91.2%. Incremental update count: Triggered 47 times, executed successfully 42 times, and rolled back 5 times (success rate 89.4%). Query response: Approximately 200 queries per day, with an average first response time of 86ms. User satisfaction: A sample survey of 35 operations and maintenance engineers showed a satisfaction rating of 4.2 / 5.0. 4) Typical Fault Case: In November 2025, the system detected an abnormal increase in the water conduction swing of a certain unit, and the fusion engine output a "tailrace vortex warning" with a confidence level of 0.92. The maintenance engineer traced the evidence chain through an interpretable query interface, confirming that the water conduction swing amplitude in the X direction reached 0.41 mm, the tailrace pipe pressure pulsation frequency was 0.78 times the rotational speed, and the inspection record showed "abnormal noise from the tailrace pipe." A minor overhaul was then arranged. The inspection revealed three detached air inlet plates at the tailrace pipe inlet; after repair, the vibration returned to normal. The entire process, from warning to repair completion, took 5 days, 12 days shorter than the traditional method (relying on inspection and step-by-step investigation), thus avoiding the potential risk of blade breakage.

[0105] The specific numerical values ​​cited in this embodiment (including but not limited to swing amplitude, bearing temperature, pressure pulsation frequency, fault identification accuracy, etc.) are partly derived from actual operating data of hydropower units reported in publicly available literature, and partly from experimental verification results of this invention under simulation environment and laboratory conditions. The simulation experiment is based on a dataset constructed from typical operating conditions of hydropower units, and the experimental environment configuration is consistent with actual industrial deployment conditions, ensuring that the experimental data are representative and repeatable.

[0106] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of this application are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indication will also change accordingly.

[0107] Furthermore, the use of terms such as "first" and "second" in this application is for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. Additionally, the technical solutions of the various embodiments can be combined with each other, but only on the basis of being achievable by those skilled in the art. When the combination of technical solutions is contradictory or impossible to implement, such a combination of technical solutions should be considered non-existent and not within the scope of protection claimed in this application.

[0108] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0109] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for fusing unit status information based on a multi-source heterogeneous data lake, characterized in that, Includes the following steps: S1: Dynamic Semantic Alignment Metadata information of multi-source heterogeneous data is extracted from the data lake, a semantic alignment graph is constructed and continuously updated. The semantic alignment graph uses data lake fields as nodes and semantic equivalence relationships between fields as edges. The semantic evolution history is recorded through a version graph database, and the fields of each data source are mapped to a unified state ontology through an incremental graph neural network, outputting a unified intermediate representation layer. S2: Multimodal Joint Attention Fusion The multimodal joint attention fusion engine receives time-series and non-time-series data after normalization processing by the unified intermediate representation layer. Features are extracted by the time-series and non-time-series feature extraction networks within the engine, respectively. A cross-attention mechanism is then used to achieve bidirectional interactive fusion of time-series and non-time-series features, generating a fused feature sequence. The fused feature sequence is then pooled along the time dimension to obtain the global fused feature vector for each time window. During the training phase of the multimodal joint attention fusion engine, a masked modality training strategy is adopted to randomly mask some modality data to enhance robustness to sensor omissions. S3: Incremental Self-Supervised Learning Update The system receives the fused feature sequence generated in step S2, constructs a self-supervised loss function through a time-series comparison prediction task, and incrementally fine-tunes the multimodal joint attention fusion engine without manual annotation. At the same time, it adopts an elastic weight consolidation mechanism to prevent catastrophic forgetting and a multi-index drift detection mechanism to determine whether incremental updates are triggered. S4: Interpretable query A hierarchical summary index is constructed, comprising a first-level index for storing coarse-grained time-component-state information, a second-level index for storing the global fusion feature vector after time pooling for each time window, and a third-level index for storing attention weights and the original data path. In response to natural language queries, relevant fusion feature fragments are located through multi-level retrieval, corresponding attention weights are extracted to form an evidence chain, and interpretable query results are generated.

2. The unit status information fusion method based on a multi-source heterogeneous data lake as described in claim 1, characterized in that, Step S1 specifically includes the following steps: Extract field names, field comments, enumeration value lists, and data types from the metadata of the data lake, and encode the above information into field feature vectors using a pre-trained language model; Calculate the cosine similarity between field feature vectors, and establish semantically equivalent edges between corresponding field nodes when the similarity exceeds a preset threshold; A relational graph convolutional network is used to learn the embeddings of semantic alignment graphs. The training objective is to make the embeddings of semantically connected nodes close together and the embeddings of unconnected nodes far apart. When a new field is added, local retraining is performed on the new field node and its neighborhood, and the update timestamp is recorded and stored in the versioned graph database.

3. The unit status information fusion method based on a multi-source heterogeneous data lake as described in claim 2, characterized in that, The temporal flow feature extraction network in step S2 includes deformable convolutional layers and temporal convolutional network layers, used to process multivariate temporal data with varying sampling rates and durations, and output temporal feature maps; the non-temporal flow feature extraction network includes a lightweight pre-trained language model for text data and a lightweight convolutional neural network for image data, and outputs non-temporal feature vectors.

4. The unit status information fusion method based on a multi-source heterogeneous data lake as described in claim 3, characterized in that, The cross-attention mechanism in step S2 includes the following parts: First Attention: Using the temporal feature map as the query and the non-temporal feature vector as the key and value, calculate the attention weight of the temporal feature to the non-temporal feature; Second attention; Using non-time-series feature vectors as queries and time-series feature maps as keys and values, calculate the attention weight of non-time-series features on time-series features; The outputs of the first attention and the second attention are concatenated and then fed forward to generate a fused feature vector. Preferably, the mask modality training strategy in step S2 is as follows: during the model training phase, the input feature of a certain modality is randomly set to a zero vector with a preset probability, and at the same time, a mask vector indicating the currently available mask vector is input to the model, so that the model can still output stable integration features under the condition of missing modalities.

5. The unit status information fusion method based on a multi-source heterogeneous data lake as described in claim 4, characterized in that, Step S2 also includes an active learning strategy: selecting samples from the data lake whose model prediction confidence is within a preset low confidence interval, and having them labeled by domain experts to achieve model performance with a preset labeling cost at a preset proportion.

6. The unit status information fusion method based on a multi-source heterogeneous data lake as described in claim 5, characterized in that, Step S3 specifically includes the following steps: S31: Temporal Comparison Prediction Task: Using the fused feature vector at the current time as input, the prediction head network predicts the fused feature vector at the same time after a preset step size, and the prediction error is used as the self-supervised loss function. S32: Elastic weight consolidation mechanism: Calculate the Fisher information matrix of each parameter of the multimodal joint attention fusion engine under historical tasks, and apply regularization constraints to the changes of important parameters during incremental fine-tuning; S33: Multi-index drift detection: Simultaneously calculates three indices: maximum mean difference, classifier confidence change, and feature space density change. Incremental update is triggered when at least two indices exceed the corresponding threshold. S34: Checkpoint rollback mechanism: Before each incremental update, a snapshot of the multimodal joint attention fusion engine is saved. After the update, it is tested on a validation set containing historical representative samples. If the accuracy of the old task drops beyond a preset threshold, the update is rejected and rolled back to the previous version. Preferably, the maximum mean difference threshold for the multi-index drift detection is a dynamically adaptive threshold, which is automatically adjusted within the range of 0.03 to 0.08 based on the statistical results of historical drift events; the preset step size of the time series comparison prediction task is adaptively adjusted according to the data sampling frequency, and the higher the sampling frequency, the larger the preset step size.

7. The unit status information fusion method based on a multi-source heterogeneous data lake as described in claim 6, characterized in that, The hierarchical summary index in step S4 is specifically as follows: First-level index: Divided according to a preset time window, it records the unit component identifier, comprehensive status label and semantic hash value in each window, and uses local sensitive hash storage; The second-level index stores the global fusion feature vector after time pooling for each time window, and supports vector similarity retrieval using an approximate nearest neighbor index structure. The third-level index stores a preset number of positions with the highest weights in the attention weight matrix generated by the cross-attention mechanism within each time window, as well as the storage path of the original data corresponding to those positions in the data lake. The attention weight matrix includes the attention weight matrix of temporal features on non-temporal features and the attention weight matrix of non-temporal features on temporal features.

8. The unit status information fusion method based on multi-source heterogeneous data lake as described in claim 7, characterized in that, Step S4, which generates interpretable query results, specifically includes the following steps: For common queries, a structured template generation method is adopted. The structured template includes the state judgment result, several pieces of evidence with the highest contribution weight and their original data sources. For complex queries, a restricted large language model generation method is adopted. The generation range is constrained by evidence priority prompts, the output is forced to include formatted results with evidence identifiers, and the generated content is post-processed and verified to ensure that each statement is supported by corresponding evidence.

9. The unit status information fusion method based on multi-source heterogeneous data lake as described in claim 8, characterized in that, Step S4 also includes a three-level caching architecture: Level 1 cache: Stores frequently accessed queries and their results within a preset time period; Second-level cache: Stores the global fusion feature vectors and corresponding interpretation results for each time window within the most recent preset time period; Level 3 cache: Full storage of the hierarchical summary index; The query response adopts a progressive display strategy: it first returns coarse-grained status judgment results, then asynchronously returns evidence summaries, and loads the complete evidence chain as needed.

10. A unit status information fusion system based on a multi-source heterogeneous data lake, characterized in that, Includes the following parts: The dynamic semantic alignment module is configured to: extract metadata information from multi-source heterogeneous data in the data lake, construct and continuously update the semantic alignment graph, and map the fields of each data source to a unified state ontology through an incremental graph neural network, and output a unified intermediate representation layer. The multimodal joint attention fusion engine, connected to the dynamic semantic alignment module, is configured to: receive time-series and non-time-series data after normalization processing by the unified intermediate representation layer; extract features through the time-series and non-time-series feature extraction networks respectively; then use a cross-attention mechanism to achieve bidirectional interactive fusion, generate a fused feature sequence, and obtain a global fused feature vector; wherein, a masked modality training strategy is adopted during the training phase of the multimodal joint attention fusion engine; The incremental self-supervised learning module is connected to the dynamic semantic alignment module and the multimodal joint attention fusion engine, respectively. It is configured to receive the fused feature sequence output by the multimodal joint attention fusion engine, and perform incremental fine-tuning of the multimodal joint attention fusion engine without manual annotation. At the same time, it adopts an elastic weight consolidation mechanism to prevent catastrophic forgetting and adopts a multi-index drift detection mechanism to determine whether to trigger incremental update. An interpretable query interface is connected to the dynamic semantic alignment module, the multimodal joint attention fusion engine, and the incremental self-supervised learning module, respectively. It is configured to: construct a hierarchical summary index, which stores time-component-state coarse-grained information, global fusion feature vectors, attention weights, and original data paths; respond to natural language queries, extract the corresponding attention weights to form an evidence chain, and generate interpretable query results for visualization.

Images