An ai algorithm-based industrial big data efficient retrieval service method

Abstract

The application discloses an industrial big data efficient retrieval service method based on an AI algorithm, and relates to the technical field of data retrieval.The application first realizes millisecond-level accurate alignment of time series, text and image multi-modal data through dynamic time warping and statistical interpolation cleaning;secondly, a layered feature extraction model constructed by RoBERTa, TCN and Swin Transformer is adopted, combined with cross-modal attention fusion and GAN feature enhancement, so that the semantic and trend features of industrial data are deeply mined;in the index and retrieval stage, a high-dimensional index structure constructed by a graph neural network GAT effectively alleviates the dimension disaster, and a mixed search strategy combining HNSW and LSH greatly improves the retrieval efficiency.Industrial knowledge graph is introduced for entity alignment and logical reasoning, which makes up for the semantic gap of traditional vector retrieval, and realizes the correlation expansion based on cause and effect and hierarchical relationship.Finally, the reordering model based on LambdaMART fuses multi-dimensional scoring and context perception mechanism, and dynamically optimizes the result ranking.

Description

Technical Field

[0001] This invention relates to the field of data retrieval technology, specifically to a method for efficient industrial big data retrieval services based on AI algorithms. Background Technology

[0002] With the rapid development of Industry 4.0 and intelligent manufacturing, massive amounts of multi-source heterogeneous data are generated in industrial production processes. This data includes not only structured time-series data collected by sensors (such as temperature, pressure, and rotational speed), but also a large amount of unstructured data, such as equipment operation logs, maintenance records, and image data captured by surveillance cameras. Statistics show that over 80% of industrial big data is unstructured or semi-structured, and it is characterized by high generation frequency, complex modes, and strong correlations.

[0003] In scenarios such as industrial equipment fault diagnosis, process optimization, and predictive maintenance, efficient data retrieval is a core component. However, existing industrial data retrieval technologies still have the following significant limitations: Insufficient multi-source heterogeneous data processing capabilities: Traditional retrieval systems are typically designed for single-modal data (such as processing only time-series databases or only documents), making it difficult to process time-series, text, and image data simultaneously within a unified framework. In the data preprocessing stage, simple mean filling or static threshold denoising cannot cope with the complex noise interference and data gaps in industrial settings, and lacks a precise timestamp alignment mechanism for multimodal data, resulting in poor data correlation and an inability to recreate the real production scenario.

[0004] Superficial and fragmented feature representation: Existing feature extraction methods mostly employ manual rules or shallow machine learning models (such as TF-IDF, SIFT, etc.), making it difficult to capture the deep semantics of industrial data. For text logs, there is a lack of contextual understanding of industrial terminology; for time-series data, it is difficult to simultaneously capture short-term fluctuations and long-term trends; for image data, traditional CNNs struggle to establish global contextual relationships. More importantly, existing technologies lack cross-modal feature fusion mechanisms, resulting in semantic fragmentation of different modalities in the feature space, failing to form complementary information.

[0005] The contradiction between efficiency and accuracy in high-dimensional index retrieval: As the dimensions of deep features increase (often reaching hundreds or even thousands of dimensions), traditional inverted indexes and relational database indexes face the "curse of dimensionality," resulting in a sharp decline in retrieval efficiency. Although Approximate Nearest Neighbor (ANN) algorithms (such as LSH and HNSW) alleviate the computational pressure to some extent, in the high-dimensional vector space of industry, simply relying on vector similarity can easily ignore the topological structure and semantic relationships of the data, leading to "semantic drift" in retrieval results. That is, the returned results are mathematically close but logically unrelated in industry.

[0006] Lack of semantic understanding and knowledge reasoning: Users typically use natural language for searches (e.g., "find downtime events on production line A caused by bearing overheating in the past week"). Existing keyword matching technologies cannot accurately identify entities (e.g., "production line A", "bearing overheating") and their complex semantic relationships (time, causality, space). Furthermore, industrial data contains rich domain knowledge (e.g., equipment hierarchy, fault propagation chains). Existing retrieval systems lack support from industrial knowledge graphs, making it impossible to perform correlation expansion and logical reasoning on search results, leading to missed detections or the inability to provide in-depth root cause analysis of faults.

[0007] The ranking of search results lacks context awareness: existing ranking models are mostly based on simple statistical indicators (such as TF-IDF scores or vector cosine similarity) without comprehensively considering data quality, temporal context, user historical behavior, and real-time device status. In terms of multi-dimensional score fusion, there is a lack of adaptive re-ranking mechanisms, resulting in low relevance and usability of search results. Users need to manually sift through massive amounts of results for effective information, severely impacting the response speed of industrial operations and maintenance. Summary of the Invention

[0008] To address the aforementioned technical problems, this paper provides a method for efficient industrial big data retrieval services based on AI algorithms. This technical solution resolves the issues raised in the background section.

[0009] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for efficient industrial big data retrieval services based on AI algorithms, comprising: Acquire multi-source heterogeneous raw datasets from the industrial field, the datasets including structured time-series data, unstructured text logs, and image data; The original dataset is cleaned and aligned to generate a standard industrial data sequence. Specifically, this includes filling missing values ​​using interpolation, removing outliers based on the 3σ principle, and mapping multimodal data to a unified timestamp dimension. A multimodal deep feature extraction model is constructed to perform hierarchical feature encoding on the cleaned data, generating text semantic feature vectors, time-series trend feature vectors, and image visual feature vectors respectively. A high-dimensional vector index library is constructed based on feature vectors. A graph neural network structure is used to segment and connect the vector space to establish a mapping relationship between index nodes and data records. It receives natural language search queries input by users, performs semantic parsing and entity recognition on the query statements, and transforms them into target search feature vectors; Approximate nearest neighbor search is performed in a high-dimensional vector index library, and candidate datasets are selected by combining vector similarity calculation with inverted index matching. Based on industrial knowledge graphs, the candidate datasets are expanded and logically reasoned to calculate the semantic matching score between the candidate data and the query intent. A re-ranking model is used to perform multi-dimensional scoring fusion on candidate datasets, and a context-aware mechanism is combined to dynamically adjust the ranking of results and output the final retrieval results.

[0010] Preferably, the cleaning and alignment of the original dataset specifically includes: For structured time series data, a local weighted linear regression interpolation method based on a sliding window is used to fill missing values. The window length is dynamically adjusted according to the data sampling frequency, and the weight coefficients are determined by the autocorrelation function of the time series data. The interpolation result is calculated using the following formula: ; in, For missing moments The interpolation results, The radius of the sliding window is determined by the sampling frequency. Decide, , This is the time window coefficient, with a range of values. , For the time within the window The observed values, The weighting coefficients are derived from the autocorrelation function. The calculation yielded: ; For time series data in lag The autocorrelation coefficient at a given location is calculated using the following formula: ; in The time series mean Standard deviation For data length; based on A dynamic threshold model is constructed based on the principle of using the Isolation Forest algorithm to perform secondary verification of outliers, removing data points that exceed the threshold and have an Isolation Forest anomaly score greater than 0.8; the Isolation Forest anomaly score is calculated using the following formula: ; in, For data points Abnormal scores, for Average path length in an isolated tree for n Average path length of samples ( , For harmonic series, , It is the Euler-Macheroni constant; For unstructured text logs and image data, an event-triggered timestamp alignment strategy is adopted to extract time entities from the text and capture time from the EXIF ​​information of the image. The dynamic time warping (DTW) algorithm is used to calculate the curved path of the time series and map the multimodal data to a unified millisecond-level timestamp dimension. The distance metric of DTW adopts an improved Euclidean distance and introduces a time decay factor to reduce historical data alignment errors. The DTW distance calculation formula is: ; in, and There are two time series. For curved paths (satisfying) , For the improved Euclidean distance, ,in The time decay factor has a range of values. , They are respectively and timestamp, For feature dimensions.

[0011] Preferably, after mapping the multimodal data to a unified timestamp dimension, the process further includes feature enhancement processing, specifically including: Cross-modal attention fusion is performed on text semantic feature vectors, temporal trend feature vectors, and image visual feature vectors to construct a Transformer-based cross-modal interaction model. Text features are used as queries, temporal features and image features are used as keys and values, cross-modal attention weights are calculated, and fused feature vectors are generated. The formula for calculating cross-modal attention weights is: ; in, For text feature vectors, and For time series / image feature vectors, For feature dimension, Number of time-series / image feature vectors; fused feature vectors ; A feature enhancement module based on Generative Adversarial Network (GAN) is adopted, with the fused feature vector as the input to the generator. The discriminator uses a multilayer perceptron (MLP) to distinguish between real features and generated features. The discriminativeness of the feature vector is enhanced through adversarial training. The generator adopts a residual connection structure, and the loss function of the discriminator includes Wasserstein distance and gradient penalty terms. The generator formula is:

[0012] in, To fuse feature vectors, It is a two-layer fully connected network; The discriminator loss function is: ; in, For the true feature distribution, Input distribution to the generator, for The distribution, This is the gradient penalty coefficient; The generator loss function is .

[0013] Preferably, the construction of the multimodal deep feature extraction model specifically includes: For unstructured text logs, a pre-trained language model based on RoBERTa-wwm is used for semantic encoding. An industrial-specific vocabulary is introduced to enhance the embedding layer. Long-distance dependencies are captured through a bidirectional long short-term memory network (BiLSTM) to output text semantic feature vectors. The hidden layer states of BiLSTM are weighted and fused using an attention mechanism, and the attention weights are dynamically generated by the density of technical terms in the text. The formula for calculating attention weights is: ; in, For a moment t Attention weights , For BiLSTM at time t The hidden layer state, and Given a trainable parameter matrix and a bias vector, v For attention vectors, For text sequence length; terminology density ,in This represents the number of industry-specific technical terms in the text. Total word count of the text; For structured time-series data, a hybrid encoding model based on Temporal Convolutional Network (TCN) and self-attention mechanism is constructed. TCN uses dilated convolution to capture multi-scale time-series trends, and the self-attention layer calculates the correlation strength at different time steps to generate time-series trend feature vectors. The dilation coefficient of the dilated convolution increases exponentially, covering time spans from minutes to hours. The formula for dilated convolution is: ; in, For output features, To input timing data, For convolution kernel weights, K The kernel size is [size]. The expansion coefficient is used; the formula for calculating the self-attention association strength is: ; in, For time step i and j The strength of the association, , and The feature vector output by the self-attention layer. For feature dimensions; For image data, a hierarchical visual feature extraction model based on the Swing Transformer is adopted. This model divides the image into patches using a sliding window and calculates local self-attention. Global feature interaction is achieved through cross-window shifting operations, outputting an image visual feature vector. The number of Transformer layers is adaptively adjusted according to the image resolution; for resolutions greater than [resolution value missing], the model is used. Increased to Layer; the formula for calculating local self-attention is: ; in, These are query, key, and value matrices, respectively. For feature dimension, This is a mask matrix, and the values ​​represent the positions outside the non-overlapping window of the mask. Image resolution ,when At that time, the number of Transformer layers ,otherwise .

[0014] Preferably, the construction of a high-dimensional vector index library based on feature vectors specifically includes: The cleaned multimodal feature vectors are mapped to node features of a graph neural network. The node attributes include the L2 norm of the feature vector, the cosine similarity threshold, and the data source type identifier. Node feature vectors Generate using the following formula: ; in, The original feature vector, It is the L2 norm. The cosine similarity threshold is calculated using the following formula: , The mean cosine similarity between the feature vector and its neighboring vectors. Standard deviation Identifier for data source type; A dynamic index structure based on the graph attention network GAT is constructed. The semantic association strength between nodes is calculated through a multi-head attention mechanism to generate the edge weights between nodes. The attention coefficient adopts the LeakyReLU activation function, and the negative slope parameter is adaptively adjusted according to the sparsity of the feature vector. The formula for calculating the attention coefficient is: ; in, For nodes i and j Attention coefficient For trainable weight matrix, For attention vectors, For nodes i The set of neighboring nodes, This indicates vector concatenation; Negative slope parameter Due to the sparsity of eigenvectors ,in The number of non-zero elements. Determined by the feature dimension, ; A hierarchical graph clustering algorithm is used to partition the vector space, dividing the high-dimensional space into multiple subspace clusters. Each cluster corresponds to an index node. Nodes within a cluster are connected by a minimum spanning tree algorithm, and cross-cluster associations are established between clusters through bridge nodes. A bidirectional mapping relationship between the index node and the original data record is established, and the mapping relationship is compressed and stored using a Bloom filter. The formula for calculating the edge weights of a minimum spanning tree is: ; in, For nodes i and j edge weights, and For node feature vectors; Number of hash functions in a Bloom filter ,inm For filter bit width, n The number of elements stored, and the false positive rate. By adjusting m make .

[0015] Preferably, the semantic parsing and entity recognition of the query statement specifically includes: A joint model based on BERT-CRF is used for entity recognition. The BERT layer is pre-trained using industrial corpus, and the CRF layer introduces constraint rules to limit entity boundaries, thereby identifying entities in the query such as equipment name, process parameters, fault type and time range. The scoring function for the CRF layer is: ; in, A sequence of entity labels. For the query word sequence, The label transition score is learned by the CRF layer. For the BERT layer at time t Output Labels The score; The constraints include "time range entities must be followed by a time unit" and "device name entities must not exceed 10 characters in length"; A semantic parsing module based on dependency parsing is constructed to extract the subject-verb-object structure and modification relations of the query statement and generate a semantic role labeling sequence. The dependency relations are represented by tuples (head, relation, tail), and the relation includes three types of industry-specific relations: "temporal relation", "causal relation" and "spatial relation". The probability calculation formula for semantic role labeling is: ; in, r It is a dependency relationship. h and t These are the feature vectors of the head node and the tail node. Represents element-wise product. For a set of relations, For relationship r The weight vector; The identified entities and semantic role annotation sequences are mapped to the target retrieval feature space. The entity vector is generated by the TransE-based knowledge graph embedding method. The entity vector is then fused with the global semantic vector of the query statement through a gating mechanism to generate the target retrieval feature vector. The gating weight is determined by the centrality index of the entity in the industrial knowledge graph. The formula for generating entity vectors in TransE is: ; in, For the head entity vector, For relation vectors, For the tail entity vector, minimize Training yielded, among which A set of knowledge graph triples; The gating fusion formula is: ; in, To retrieve feature vectors for the target, This is the global semantic vector of the query statement. For the first i TransE vectors of each entity, For the number of entities, For gated functions ( , c As an indicator of entity centrality, , For nodes v The degree, N This represents the total number of nodes in the knowledge graph. d This represents the number of times the entity appears in the query. These are trainable parameters.

[0016] Preferably, performing an approximate nearest neighbor search in the high-dimensional vector index specifically includes: Candidate datasets are pre-filtered based on an inverted index structure. The key-value pairs of the inverted index are (entity attribute, list of data record IDs). The entity attributes include equipment ID, process stage, time interval, and fault code. Boolean queries are used to match the entity attributes entered by the user to filter out the initial candidate set. The formula for calculating the Boolean query match degree is:

[0017] in, q To allow users to query a set of entity attributes, d For data recording; In the initial candidate set, a vector index graph is constructed using the hierarchical navigation small-world HNSW algorithm. A greedy search is then performed to traverse the graph and retrieve the feature vectors with the highest similarity to the target. K The similarity calculation uses an improved cosine similarity, which introduces a variance weighting term for the feature vector to balance the scale differences of features of different modalities. The improved cosine similarity formula is as follows: ; in, To retrieve feature vectors for the target, Let be the candidate data feature vectors, and Σ be the covariance matrix of the feature vectors, which is obtained from the training set statistics. , The characteristic mean; Combining vector similarity with inverted index matching results, a hash bucket filtering strategy based on Locality Sensitive Hash (LSH) is adopted. Vectors are mapped to hash buckets, and precise similarity calculation is performed only within buckets with hash collisions to filter out candidate datasets. The hash function is generated by random projection of p-stable distribution. The hash function formula is: ; in, a Let b be a random vector sampled from a p-stable distribution (Cauchy distribution when p=1). Offset of uniform sampling within the area. Where is the bucket width. The hash collision condition is... .

[0018] Preferably, after selecting the candidate dataset, a candidate set pruning step is further included, specifically including: The spatiotemporal density distribution of the candidate dataset is calculated. The density clustering algorithm based on DBSCAN is used to divide the candidate data into high-density clusters and low-density clusters. The neighborhood radius ε is determined based on the average distance of the feature vectors, and the minimum number of points MinPts is set according to the typical data density of industrial scenarios. Neighborhood radius ,in , For candidate feature vectors, For the number of candidates, Scaling factor; minimum number of points ; Representative screening of candidate data within high-density clusters is performed, and cluster center data points are selected using a strategy based on maximum marginal nearest neighbor (MMN). The centrality index is determined by the local density of the data point and its distance to higher density points. Local density Distance to higher density points Centrality indicators ,choose The largest top 20% of data points; Noise filtering is performed on candidate data within low-density clusters. An anomaly detection model based on isolated forests is used to remove data points whose anomaly scores exceed a threshold. The threshold is dynamically calibrated by the historical noise rate of the industrial scenario. Anomaly scores are also calculated using the Isolation Forest anomaly score calculation formula, with a threshold. ,in The historical noise score is the average. Let be the standard deviation, when the historical noise rate hour, Increase by 10%, otherwise keep the original value.

[0019] Preferably, the step of performing association expansion and logical reasoning on the candidate dataset based on the industrial knowledge graph specifically includes: An industrial knowledge graph is constructed, which includes equipment entities, process entities, fault entities, and relationship types. The relationship types include four categories: "belongs to", "triggers", "associates", and "time series". The entity embedding method based on graph convolutional network (GCN) is used to generate the vector representation of the knowledge graph. The propagation formula for the GCN layer is: ; in, For the first l The node feature matrix of the layer, ,in A It is an adjacency matrix. It is the identity matrix. for The degree matrix, For the first l The trainable weight matrix of the layer, It is the ReLU activation function; Knowledge graph alignment is performed on entities in the candidate dataset. A dual matching strategy based on string similarity and vector similarity is adopted to link candidate entities to corresponding nodes in the knowledge graph. The string similarity is calculated using Jaro-Winkler distance, and the vector similarity is calculated using the scoring function of the TransE model. The Jaro-Winkler distance formula is: ; in, and For candidate entity and knowledge graph entity names, m To match the number of characters, t Half the number of character transpositions. l The length of the common prefix. p Prefix scaling factor; The TransE scoring function is: Double-match score in and These represent the minimum and maximum values ​​of the scoring function; Based on the description logic DL inference engine, logical reasoning is performed on the linked knowledge graph. The inference rule includes "If device A causes fault B, and fault B is associated with process C, then device A and process C are indirectly associated", generating an extended set of candidate data associations. The description logic inference formula is: ; in, Indicates conjunction. Quantifiers indicating existence; The semantic matching score between candidate data and query intent is calculated using a graph neural network model based on GraphSAGE. The input is the feature vector of the candidate data and the subgraph structure of the knowledge graph, and the output is the semantic matching score. The subgraph structure includes the 2-hop neighbor nodes and relation edges of the candidate entity. The GraphSAGE aggregation formula is:

[0020] in, For nodes v In the k The feature vector of the layer, It is the mean aggregation function. , For trainable weight matrices, semantic matching scores ,in For the number of GraphSAGE layers, and These are the parameters for the classification layer.

[0021] Preferably, the step of using a re-ranking model to perform multi-dimensional score fusion on the candidate dataset specifically includes: A multi-dimensional scoring index system is constructed, including vector similarity score, knowledge graph association strength score, temporal context matching score, and data quality score. The data quality score is calculated by weighting data integrity, collection frequency stability, and historical verification pass rate. The formula for data quality score is: ; in, For the number of valid data points, Total number of data points The standard deviation of the sampling frequency. This is the average sampling frequency. This represents the number of times the historical verification has passed. This represents the total number of checks. For the weights, where ; A LambdaMART-based learning ranking model was adopted, with multi-dimensional scoring indicators as features and manually labeled search relevance scores as labels. The re-ranking model was trained, and the model loss function was optimized using NDCG. The formula for calculating NDCG is: ; in, IDCG@ K The DCG value under ideal sorting. For the first i Relevance score of each result; The LambdaMART loss function is: ; in, For exchanging documents i and j The change in NDCG after that, and The model predicts the score. For scaling parameters; A context-aware mechanism is introduced to obtain the user's historical query sequence, current session context and device operating status parameters in real time. The context temporal dependency is modeled through a gated recurrent unit (GRU) to generate a dynamic weight vector and perform weighted fusion of multi-dimensional scoring indicators. The hidden layer state of the GRU is adaptively updated according to the entity density of the current query. The GRU update formula is: ; ; ; ; in, To update the door, To reset the door, In the candidate hidden state, Currently in a hidden state. Entity density is a contextual feature of the current query. ,in To determine the number of entities in the query. For query length, when At that time, the hidden layer dimension of GRU ,otherwise Dynamic weight vector ; Sort the candidate datasets in descending order based on the fused scores, and output the top-ranked datasets. N The results are used as the final search results. NThe system load is dynamically adjusted based on the number of returns specified by the user. N The calculation formula is: ; in, The number of returns specified by the user. The system's currently available computing resources. Resources required for processing a single result.

[0022] Compared with existing technologies, this invention provides a method for efficient industrial big data retrieval services based on AI algorithms, which has the following beneficial effects: This invention first achieves millisecond-level precise alignment of multimodal data (time series, text, and images) through dynamic time warping and statistical interpolation cleaning, solving the problems of data heterogeneity and missing data. Second, it employs a hierarchical feature extraction model built with RoBERTa, TCN, and SwinTransformer, combined with cross-modal attention fusion and GAN feature enhancement, to deeply mine the semantic and trend features of industrial data. In the indexing and retrieval stages, a high-dimensional index structure built using graph neural networks (GAT) effectively alleviates the curse of dimensionality, and a hybrid search strategy combining HNSW and LSH significantly improves retrieval efficiency. An industrial knowledge graph is introduced for entity alignment and logical reasoning, bridging the semantic gap in traditional vector retrieval and enabling association expansion based on causal and hierarchical relationships. Finally, a re-ranking model based on LambdaMART integrates multidimensional scoring and context-aware mechanisms to dynamically optimize result ranking. This method achieves accurate, real-time, and interpretable retrieval of industrial big data, significantly reducing the information screening costs for maintenance personnel. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the method flow framework of the present invention. Detailed Implementation

[0024] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0025] Example 1 Please refer to Figure 1 As shown, a method for efficient industrial big data retrieval services based on AI algorithms includes: S101. Obtain multi-source heterogeneous raw datasets from the industrial field. The datasets include structured time-series data, unstructured text logs, and image data. S102. Clean and align the original dataset to generate a standard industrial data sequence. Specifically, this includes filling missing values ​​using interpolation, removing outliers based on the 3σ principle, and mapping multimodal data to a unified timestamp dimension. S103. Construct a multimodal deep feature extraction model, perform hierarchical feature encoding on the cleaned data, and generate text semantic feature vectors, time-series trend feature vectors, and image visual feature vectors respectively. S104. Construct a high-dimensional vector index library based on feature vectors, and use a graph neural network structure to segment and connect the vector space to establish a mapping relationship between index nodes and data records. S105. Receive natural language search queries input by the user, perform semantic parsing and entity recognition on the query statements, and convert them into target search feature vectors; S106. Perform an approximate nearest neighbor search in the high-dimensional vector index library, and combine vector similarity calculation with inverted index matching to filter out candidate datasets; S107. Based on the industrial knowledge graph, perform association expansion and logical reasoning on the candidate dataset, and calculate the semantic matching score between the candidate data and the query intent. S108. Use a re-ranking model to perform multi-dimensional scoring fusion on the candidate dataset, combine it with a context-aware mechanism to dynamically adjust the result ranking, and output the final retrieval results.

[0026] Those skilled in the art will understand that the efficient industrial big data retrieval method proposed in this invention significantly improves the processing depth and retrieval accuracy of multi-source heterogeneous data. First, through dynamic time warping and statistical interpolation cleaning, millisecond-level precise alignment of time-series, text, and image multimodal data is achieved, solving the problems of data heterogeneity and missing data. Second, a hierarchical feature extraction model constructed using RoBERTa, TCN, and Swin Transformer, combined with cross-modal attention fusion and GAN feature enhancement, deeply mines the semantic and trend features of industrial data. In the indexing and retrieval stages, a high-dimensional index structure constructed using graph neural networks (GAT) effectively alleviates the curse of dimensionality, and a hybrid search strategy combining HNSW and LSH significantly improves retrieval efficiency. The introduction of industrial knowledge graphs for entity alignment and logical reasoning bridges the semantic gap in traditional vector retrieval, achieving association expansion based on causal and hierarchical relationships. Finally, a re-ranking model based on LambdaMART integrates multi-dimensional scoring and context-aware mechanisms to dynamically optimize result ranking. This method achieves accurate, real-time, and interpretable retrieval of industrial big data, significantly reducing the information screening costs for maintenance personnel.

[0027] The original dataset undergoes cleaning and alignment processes, specifically including: For structured time series data, a local weighted linear regression interpolation method based on a sliding window is used to fill missing values. The window length is dynamically adjusted according to the data sampling frequency, and the weight coefficients are determined by the autocorrelation function of the time series data. The interpolation result is calculated using the following formula: ; in, For missing moments The interpolation results, The radius of the sliding window is determined by the sampling frequency. Decide, , This is the time window coefficient, with a range of values. , For the time within the window The observed values, The weighting coefficients are derived from the autocorrelation function. The calculation yielded: ; For time series data in lag The autocorrelation coefficient at a given location is calculated using the following formula: ; in The time series mean Standard deviation For data length; based on A dynamic threshold model is constructed based on the principle of using the Isolation Forest algorithm to perform secondary verification of outliers, removing data points that exceed the threshold and have an Isolation Forest anomaly score greater than 0.8; the Isolation Forest anomaly score is calculated using the following formula: ; in, For data points Abnormal scores, for Average path length in an isolated tree for n Average path length of samples ( , For harmonic series, , It is the Euler-Macheroni constant; For unstructured text logs and image data, an event-triggered timestamp alignment strategy is adopted to extract time entities from the text and capture time from the EXIF ​​information of the image. The dynamic time warping (DTW) algorithm is used to calculate the curved path of the time series and map the multimodal data to a unified millisecond-level timestamp dimension. The distance metric of DTW adopts an improved Euclidean distance and introduces a time decay factor to reduce historical data alignment errors. The DTW distance calculation formula is: ; in, and There are two time series. For curved paths (satisfying) , For the improved Euclidean distance, ,in The time decay factor has a range of values. , They are respectively and timestamp, For feature dimensions.

[0028] After mapping multimodal data to a unified timestamp dimension, feature enhancement processing is also included, specifically: Cross-modal attention fusion is performed on text semantic feature vectors, temporal trend feature vectors, and image visual feature vectors to construct a Transformer-based cross-modal interaction model. Text features are used as queries, temporal features and image features are used as keys and values, cross-modal attention weights are calculated, and fused feature vectors are generated. The formula for calculating cross-modal attention weights is: ; in, For text feature vectors, and For time series / image feature vectors, For feature dimension, Number of time-series / image feature vectors; fused feature vectors ; A feature enhancement module based on Generative Adversarial Network (GAN) is adopted, with the fused feature vector as the input to the generator. The discriminator uses a multilayer perceptron (MLP) to distinguish between real features and generated features. The discriminativeness of the feature vector is enhanced through adversarial training. The generator adopts a residual connection structure, and the loss function of the discriminator includes Wasserstein distance and gradient penalty terms. The generator formula is:

[0029] in, To fuse feature vectors, It is a two-layer fully connected network; The discriminator loss function is: ; in, For the true feature distribution, Input distribution to the generator, for The distribution, This is the gradient penalty coefficient; The generator loss function is .

[0030] Constructing a multimodal deep feature extraction model specifically includes: For unstructured text logs, a pre-trained language model based on RoBERTa-wwm is used for semantic encoding. An industrial-specific vocabulary is introduced to enhance the embedding layer. Long-distance dependencies are captured through a bidirectional long short-term memory network (BiLSTM) to output text semantic feature vectors. The hidden layer states of BiLSTM are weighted and fused using an attention mechanism, and the attention weights are dynamically generated by the density of technical terms in the text. The formula for calculating attention weights is: ; in, For a moment t Attention weights , For BiLSTM at time t The hidden layer state, and Given a trainable parameter matrix and a bias vector, v For attention vectors, For text sequence length; terminology density ,in This represents the number of industry-specific technical terms in the text. Total word count of the text; For structured time-series data, a hybrid encoding model based on Temporal Convolutional Network (TCN) and self-attention mechanism is constructed. TCN uses dilated convolution to capture multi-scale time-series trends, and the self-attention layer calculates the correlation strength at different time steps to generate time-series trend feature vectors. The dilation coefficient of the dilated convolution increases exponentially, covering time spans from minutes to hours. The formula for dilated convolution is: ; in, For output features, To input timing data, For convolution kernel weights, K The kernel size is [size]. The expansion coefficient is used; the formula for calculating the self-attention association strength is: ; in, For time step i and j The strength of the association, , and The feature vector output by the self-attention layer. For feature dimensions; For image data, a hierarchical visual feature extraction model based on the Swing Transformer is adopted. This model divides the image into patches using a sliding window and calculates local self-attention. Global feature interaction is achieved through cross-window shifting operations, outputting an image visual feature vector. The number of Transformer layers is adaptively adjusted according to the image resolution; for resolutions greater than [resolution value missing], the model is used. Increased to Layer; the formula for calculating local self-attention is: ; in, These are query, key, and value matrices, respectively. For feature dimension, This is a mask matrix, and the values ​​represent the positions outside the non-overlapping window of the mask. Image resolution ,when At that time, the number of Transformer layers ,otherwise .

[0031] A high-dimensional vector index library is built based on feature vectors, specifically including: The cleaned multimodal feature vectors are mapped to node features of a graph neural network. The node attributes include the L2 norm of the feature vector, the cosine similarity threshold, and the data source type identifier. Node feature vectors Generate using the following formula: ; in, The original feature vector, It is the L2 norm. The cosine similarity threshold is calculated using the following formula: , The mean cosine similarity between the feature vector and its neighboring vectors. Standard deviation Identifier for data source type; A dynamic index structure based on the graph attention network GAT is constructed. The semantic association strength between nodes is calculated through a multi-head attention mechanism to generate the edge weights between nodes. The attention coefficient adopts the LeakyReLU activation function, and the negative slope parameter is adaptively adjusted according to the sparsity of the feature vector. The formula for calculating the attention coefficient is: ; in, For nodes i and j Attention coefficient For trainable weight matrix, For attention vectors, For nodes i The set of neighboring nodes, This indicates vector concatenation; Negative slope parameter Due to the sparsity of eigenvectors ,in The number of non-zero elements. Determined by the feature dimension, ; A hierarchical graph clustering algorithm is used to partition the vector space, dividing the high-dimensional space into multiple subspace clusters. Each cluster corresponds to an index node. Nodes within a cluster are connected by a minimum spanning tree algorithm, and cross-cluster associations are established between clusters through bridge nodes. A bidirectional mapping relationship between the index node and the original data record is established, and the mapping relationship is compressed and stored using a Bloom filter. The formula for calculating the edge weights of a minimum spanning tree is: ; in, For nodes i and j edge weights, and For node feature vectors; Number of hash functions in a Bloom filter ,in m For filter bit width, n The number of elements stored, and the false positive rate. By adjusting m make .

[0032] Semantic parsing and entity recognition of query statements, specifically including: A joint model based on BERT-CRF is used for entity recognition. The BERT layer is pre-trained using industrial corpus, and the CRF layer introduces constraint rules to limit entity boundaries, thereby identifying entities in the query such as equipment name, process parameters, fault type and time range. The scoring function for the CRF layer is: ; in, A sequence of entity labels. For the query word sequence, The label transition score is learned by the CRF layer. For the BERT layer at time t Output Labels The score; The constraints include "time range entities must be followed by a time unit" and "device name entities must not exceed 10 characters in length"; A semantic parsing module based on dependency parsing is constructed to extract the subject-verb-object structure and modification relations of the query statement and generate a semantic role labeling sequence. The dependency relations are represented by tuples (head, relation, tail), and the relation includes three types of industry-specific relations: "temporal relation", "causal relation" and "spatial relation". The probability calculation formula for semantic role labeling is: ; in, r It is a dependency relationship. h and t These are the feature vectors of the head node and the tail node. Represents element-wise product. For a set of relations, For relationship r The weight vector; The identified entities and semantic role annotation sequences are mapped to the target retrieval feature space. The entity vector is generated by the TransE-based knowledge graph embedding method. The entity vector is then fused with the global semantic vector of the query statement through a gating mechanism to generate the target retrieval feature vector. The gating weight is determined by the centrality index of the entity in the industrial knowledge graph. The formula for generating entity vectors in TransE is: ; in, For the head entity vector, For relation vectors, For the tail entity vector, minimize Training yielded, among which A set of knowledge graph triples; The gating fusion formula is: ; in, To retrieve feature vectors for the target, This is the global semantic vector of the query statement. For the first i TransE vectors of each entity, For the number of entities, For gated functions ( , c As an indicator of entity centrality, , For nodes v The degree, N This represents the total number of nodes in the knowledge graph. d This represents the number of times the entity appears in the query. These are trainable parameters.

[0033] Performing approximate nearest neighbor search in a high-dimensional vector index library specifically includes: Candidate datasets are pre-filtered based on an inverted index structure. The key-value pairs of the inverted index are (entity attribute, list of data record IDs). The entity attributes include equipment ID, process stage, time interval, and fault code. Boolean queries are used to match the entity attributes entered by the user to filter out the initial candidate set. The formula for calculating the Boolean query match degree is:

[0034] in, q To allow users to query a set of entity attributes, d For data recording; In the initial candidate set, a vector index graph is constructed using the hierarchical navigation small-world HNSW algorithm. A greedy search is then performed to traverse the graph and retrieve the feature vectors with the highest similarity to the target. K The similarity calculation uses an improved cosine similarity, which introduces a variance weighting term for the feature vector to balance the scale differences of features of different modalities. The improved cosine similarity formula is as follows: ; in, To retrieve feature vectors for the target, Let be the candidate data feature vectors, and Σ be the covariance matrix of the feature vectors, which is obtained from the training set statistics. , The characteristic mean; Combining vector similarity with inverted index matching results, a hash bucket filtering strategy based on Locality Sensitive Hash (LSH) is adopted. Vectors are mapped to hash buckets, and precise similarity calculation is performed only within buckets with hash collisions to filter out candidate datasets. The hash function is generated by random projection of p-stable distribution. The hash function formula is: ; in, a Let b be a random vector sampled from a p-stable distribution (Cauchy distribution when p=1). Offset of uniform sampling within the area. Where is the bucket width. The hash collision condition is... .

[0035] After selecting candidate datasets, a candidate set pruning step is also included, which specifically includes: The spatiotemporal density distribution of the candidate dataset is calculated. The density clustering algorithm based on DBSCAN is used to divide the candidate data into high-density clusters and low-density clusters. The neighborhood radius ε is determined based on the average distance of the feature vectors, and the minimum number of points MinPts is set according to the typical data density of industrial scenarios. Neighborhood radius ,in , For candidate feature vectors, For the number of candidates, Scaling factor; minimum number of points ; Representative screening of candidate data within high-density clusters is performed, and cluster center data points are selected using a strategy based on maximum marginal nearest neighbor (MMN). The centrality index is determined by the local density of the data point and its distance to higher density points. Local density Distance to higher density points Centrality indicators ,choose The largest top 20% of data points; Noise filtering is performed on candidate data within low-density clusters. An anomaly detection model based on isolated forests is used to remove data points whose anomaly scores exceed a threshold. The threshold is dynamically calibrated by the historical noise rate of the industrial scenario. Anomaly scores are also calculated using the Isolation Forest anomaly score calculation formula, with a threshold. ,in The historical noise score is the average. Let be the standard deviation, when the historical noise rate hour, Increase by 10%, otherwise keep the original value.

[0036] Based on industrial knowledge graphs, candidate datasets are expanded and logically reasoned, specifically including: An industrial knowledge graph is constructed, which includes equipment entities, process entities, fault entities, and relationship types. The relationship types include four categories: "belongs to", "triggers", "associates", and "time series". The entity embedding method based on graph convolutional network (GCN) is used to generate the vector representation of the knowledge graph. The propagation formula for the GCN layer is: ; in, For the first l The node feature matrix of the layer, ,in A It is an adjacency matrix. It is the identity matrix. for The degree matrix, For the first l The trainable weight matrix of the layer, It is the ReLU activation function; Knowledge graph alignment is performed on entities in the candidate dataset. A dual matching strategy based on string similarity and vector similarity is adopted to link candidate entities to corresponding nodes in the knowledge graph. The string similarity is calculated using Jaro-Winkler distance, and the vector similarity is calculated using the scoring function of the TransE model. The Jaro-Winkler distance formula is: ; in, and For candidate entity and knowledge graph entity names, m To match the number of characters, t Half the number of character transpositions. l The length of the common prefix. p Prefix scaling factor; The TransE scoring function is: Double-match score in and These represent the minimum and maximum values ​​of the scoring function; Based on the description logic DL inference engine, logical reasoning is performed on the linked knowledge graph. The inference rule includes "If device A causes fault B, and fault B is associated with process C, then device A and process C are indirectly associated", generating an extended set of candidate data associations. The description logic inference formula is: ; in, Indicates conjunction. Quantifiers indicating existence; The semantic matching score between candidate data and query intent is calculated using a graph neural network model based on GraphSAGE. The input is the feature vector of the candidate data and the subgraph structure of the knowledge graph, and the output is the semantic matching score. The subgraph structure includes the 2-hop neighbor nodes and relation edges of the candidate entity. The GraphSAGE aggregation formula is:

[0037] in, For nodes v In the k The feature vector of the layer, It is the mean aggregation function. , For trainable weight matrices, semantic matching scores ,in For the number of GraphSAGE layers, and These are the parameters for the classification layer.

[0038] The candidate dataset is fused using a re-ranking model with multi-dimensional scores, specifically including: A multi-dimensional scoring index system is constructed, including vector similarity score, knowledge graph association strength score, temporal context matching score, and data quality score. The data quality score is calculated by weighting data integrity, collection frequency stability, and historical verification pass rate. The formula for data quality score is: ; in, For the number of valid data points, Total number of data points The standard deviation of the sampling frequency. This is the average sampling frequency. This represents the number of times the historical verification has passed. This represents the total number of checks. For the weights, where ; A LambdaMART-based learning ranking model was adopted, with multi-dimensional scoring indicators as features and manually labeled search relevance scores as labels. The re-ranking model was trained, and the model loss function was optimized using NDCG. The formula for calculating NDCG is: ; in, IDCG@ K The DCG value under ideal sorting. For the first i Relevance score of each result; The LambdaMART loss function is: ; in, For exchanging documents i and j The change in NDCG after that, and The model predicts the score. For scaling parameters; A context-aware mechanism is introduced to obtain the user's historical query sequence, current session context and device operating status parameters in real time. The context temporal dependency is modeled through a gated recurrent unit (GRU) to generate a dynamic weight vector and perform weighted fusion of multi-dimensional scoring indicators. The hidden layer state of the GRU is adaptively updated according to the entity density of the current query. The GRU update formula is: ; ; ; ; in, To update the door, To reset the door, In the candidate hidden state, Currently in a hidden state. Entity density is a contextual feature of the current query. ,in To determine the number of entities in the query. For query length, when At that time, the hidden layer dimension of GRU ,otherwise Dynamic weight vector ; Sort the candidate datasets in descending order based on the fused scores, and output the top-ranked datasets. N The results are used as the final search results. N The system load is dynamically adjusted based on the number of returns specified by the user. N The calculation formula is: ; in, The number of returns specified by the user. The system's currently available computing resources. Resources required for processing a single result.

[0039] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A method for efficient industrial big data retrieval services based on AI algorithms, characterized in that, include: Acquire multi-source heterogeneous raw datasets from the industrial field, the datasets including structured time-series data, unstructured text logs, and image data; The original dataset is cleaned and aligned to generate a standard industrial data sequence. Specifically, this includes filling missing values ​​using interpolation, removing outliers based on the 3σ principle, and mapping multimodal data to a unified timestamp dimension. A multimodal deep feature extraction model is constructed to perform hierarchical feature encoding on the cleaned data, generating text semantic feature vectors, time-series trend feature vectors, and image visual feature vectors respectively. A high-dimensional vector index library is constructed based on feature vectors. A graph neural network structure is used to segment and connect the vector space to establish a mapping relationship between index nodes and data records. It receives natural language search queries input by users, performs semantic parsing and entity recognition on the query statements, and transforms them into target search feature vectors; Approximate nearest neighbor search is performed in a high-dimensional vector index library, and candidate datasets are selected by combining vector similarity calculation with inverted index matching. Based on industrial knowledge graphs, the candidate datasets are expanded and logically reasoned to calculate the semantic matching score between the candidate data and the query intent. A re-ranking model is used to perform multi-dimensional scoring fusion on candidate datasets, and a context-aware mechanism is combined to dynamically adjust the ranking of results and output the final retrieval results.

2. The method for efficient industrial big data retrieval service based on AI algorithm according to claim 1, characterized in that, The cleaning and alignment of the original dataset specifically includes: For structured time series data, a local weighted linear regression interpolation method based on a sliding window is used to fill missing values. The window length is dynamically adjusted according to the data sampling frequency, and the weight coefficients are determined by the autocorrelation function of the time series data. The interpolation result is calculated using the following formula: ； in, For missing moments The interpolation results, The radius of the sliding window is determined by the sampling frequency. Decide, , This is the time window coefficient, with a range of values. , For the time within the window The observed values, The weighting coefficients are derived from the autocorrelation function. The calculation yielded: ； For time series data in lag The autocorrelation coefficient at a given location is calculated using the following formula: ； in The time series mean Standard deviation For data length; based on A dynamic threshold model is constructed based on the principle of using the Isolation Forest algorithm to perform secondary verification of outliers, removing data points that exceed the threshold and have an Isolation Forest anomaly score greater than 0.8; the Isolation Forest anomaly score is calculated using the following formula: ； in, For data points Abnormal scores, for Average path length in an isolated tree for n Average path length of samples ( , For harmonic series, , It is the Euler-Macheroni constant; For unstructured text logs and image data, an event-triggered timestamp alignment strategy is adopted to extract time entities from the text and capture time from the EXIF ​​information of the image. The dynamic time warping (DTW) algorithm is used to calculate the curved path of the time series and map the multimodal data to a unified millisecond-level timestamp dimension. The distance metric of DTW adopts an improved Euclidean distance and introduces a time decay factor to reduce historical data alignment errors. The DTW distance calculation formula is: ； in, and There are two time series. For curved paths (satisfying) , For the improved Euclidean distance, ,in The time decay factor has a range of values. , They are respectively and timestamp, For feature dimensions.

3. The method for efficient industrial big data retrieval service based on AI algorithm according to claim 1, characterized in that, After mapping the multimodal data to a unified timestamp dimension, the process also includes feature enhancement processing, specifically including: Cross-modal attention fusion is performed on text semantic feature vectors, temporal trend feature vectors, and image visual feature vectors to construct a Transformer-based cross-modal interaction model. Text features are used as queries, temporal features and image features are used as keys and values, cross-modal attention weights are calculated, and fused feature vectors are generated. The formula for calculating cross-modal attention weights is: ； in, For text feature vectors, and For time series / image feature vectors, For feature dimension, Number of time-series / image feature vectors; fused feature vectors ; A feature enhancement module based on Generative Adversarial Network (GAN) is adopted, with the fused feature vector as the input to the generator. The discriminator uses a multilayer perceptron (MLP) to distinguish between real features and generated features. The discriminativeness of the feature vector is enhanced through adversarial training. The generator adopts a residual connection structure, and the loss function of the discriminator includes Wasserstein distance and gradient penalty terms. The generator formula is: ； in, To fuse feature vectors, It is a two-layer fully connected network; The discriminator loss function is: ； in, For the true feature distribution, Input distribution to the generator, for The distribution This is the gradient penalty coefficient; The generator loss function is .

4. The efficient industrial big data retrieval service method based on AI algorithm according to claim 1, characterized in that, The construction of the multimodal deep feature extraction model specifically includes: For unstructured text logs, a pre-trained language model based on RoBERTa-wwm is used for semantic encoding. An industrial-specific vocabulary is introduced to enhance the embedding layer. Long-distance dependencies are captured through a bidirectional long short-term memory network (BiLSTM) to output text semantic feature vectors. The hidden layer states of BiLSTM are weighted and fused using an attention mechanism, and the attention weights are dynamically generated by the density of technical terms in the text. The formula for calculating attention weights is: ； in, For a moment t Attention weights , For BiLSTM at time t The hidden layer state, and Given a trainable parameter matrix and a bias vector, v Here is the attention vector, where is the length of the text sequence; and the terminology density. ,in This represents the number of industry-specific technical terms in the text. Total word count of the text; For structured time-series data, a hybrid encoding model based on Temporal Convolutional Network (TCN) and self-attention mechanism is constructed. TCN uses dilated convolution to capture multi-scale time-series trends, and the self-attention layer calculates the correlation strength at different time steps to generate time-series trend feature vectors. The dilation coefficient of the dilated convolution increases exponentially, covering time spans from minutes to hours. The formula for dilated convolution is: ； in, For output features, To input timing data, For convolution kernel weights, K The kernel size is [size]. The expansion coefficient is used; the formula for calculating the self-attention association strength is: ； in, For time steps i and j The strength of the association, , and The feature vector output by the self-attention layer. For feature dimensions; For image data, a hierarchical visual feature extraction model based on the Swing Transformer is adopted. This model divides the image into patches using a sliding window and calculates local self-attention. Global feature interaction is achieved through cross-window shifting operations, outputting an image visual feature vector. The number of Transformer layers is adaptively adjusted according to the image resolution; for resolutions greater than [resolution value missing], the model is used. Increased to Layer; the formula for calculating local self-attention is: ； in, These are query, key, and value matrices, respectively. For feature dimension, This is a mask matrix, and the values ​​represent the positions outside the non-overlapping window of the mask. Image resolution ,when At that time, the number of Transformer layers ,otherwise .

5. The method for efficient industrial big data retrieval service based on AI algorithm according to claim 1, characterized in that, The construction of a high-dimensional vector index library based on feature vectors specifically includes: The cleaned multimodal feature vectors are mapped to node features of a graph neural network. The node attributes include the L2 norm of the feature vector, the cosine similarity threshold, and the data source type identifier. Node feature vectors Generate using the following formula: ； in, The original feature vector, It is the L2 norm. The cosine similarity threshold is calculated using the following formula: , The mean cosine similarity between the feature vector and its neighboring vectors. Standard deviation Identifier for data source type; A dynamic index structure based on the graph attention network GAT is constructed. The semantic association strength between nodes is calculated through a multi-head attention mechanism to generate the edge weights between nodes. The attention coefficient adopts the LeakyReLU activation function, and the negative slope parameter is adaptively adjusted according to the sparsity of the feature vector. The formula for calculating the attention coefficient is: ； in, For nodes i and j Attention coefficient For trainable weight matrix, For attention vectors, For nodes i The set of neighboring nodes, This indicates vector concatenation; Negative slope parameter Due to the sparsity of eigenvectors ,in The number of non-zero elements. Determined by the feature dimension, ; A hierarchical graph clustering algorithm is used to partition the vector space, dividing the high-dimensional space into multiple subspace clusters. Each cluster corresponds to an index node. Nodes within a cluster are connected by a minimum spanning tree algorithm, and cross-cluster associations are established between clusters through bridge nodes. A bidirectional mapping relationship between the index node and the original data record is established, and the mapping relationship is compressed and stored using a Bloom filter. The formula for calculating the edge weights of a minimum spanning tree is: ； in, For nodes i and j edge weights, and For node feature vectors; Number of hash functions in a Bloom filter ,in m For filter bit width, n The number of elements stored, and the false positive rate. By adjusting m make .

6. The efficient industrial big data retrieval service method based on AI algorithm according to claim 1, characterized in that, The semantic parsing and entity recognition of the query statement specifically includes: A joint model based on BERT-CRF is used for entity recognition. The BERT layer is pre-trained using industrial corpus, and the CRF layer introduces constraint rules to limit entity boundaries, thereby identifying entities in the query such as equipment name, process parameters, fault type and time range. The scoring function for the CRF layer is: ； in, A sequence of entity labels. For the query word sequence, The label transition score is learned by the CRF layer. For the BERT layer at time t Output Labels The score; The constraints include "time range entities must be followed by a time unit" and "device name entities must not exceed 10 characters in length"; A semantic parsing module based on dependency parsing is constructed to extract the subject-verb-object structure and modification relations of the query statement and generate a semantic role labeling sequence. The dependency relations are represented by tuples (head, relation, tail), and the relation includes three types of industry-specific relations: "temporal relation", "causal relation" and "spatial relation". The probability calculation formula for semantic role labeling is: ； in, r It is a dependency relationship. h and t These are the feature vectors of the head node and the tail node. Represents element-wise product. For a set of relations, For relationship r The weight vector; The identified entities and semantic role annotation sequences are mapped to the target retrieval feature space. The entity vector is generated by the TransE-based knowledge graph embedding method. The entity vector is then fused with the global semantic vector of the query statement through a gating mechanism to generate the target retrieval feature vector. The gating weight is determined by the centrality index of the entity in the industrial knowledge graph. The formula for generating entity vectors in TransE is: ； in, For the head entity vector, For relation vectors, For the tail entity vector, minimize Training yielded, among which A set of knowledge graph triples; The gating fusion formula is: ； in, To retrieve feature vectors for the target, This is the global semantic vector of the query statement. For the first i TransE vectors of each entity, For the number of entities, For gated functions ( , c As an indicator of entity centrality, , For nodes v The degree, N This represents the total number of nodes in the knowledge graph. d This represents the number of times the entity appears in the query. These are trainable parameters.

7. The method for efficient industrial big data retrieval service based on AI algorithm according to claim 1, characterized in that, The process of performing an approximate nearest neighbor search in a high-dimensional vector index specifically includes: Candidate datasets are pre-filtered based on an inverted index structure. The key-value pairs of the inverted index are (entity attribute, list of data record IDs). The entity attributes include equipment ID, process stage, time interval, and fault code. Boolean queries are used to match the entity attributes entered by the user to filter out the initial candidate set. The formula for calculating the Boolean query match degree is: ； in, q To allow users to query a set of entity attributes, d For data recording; In the initial candidate set, a vector index graph is constructed using the hierarchical navigation small-world HNSW algorithm. A greedy search is then performed to traverse the graph and retrieve the feature vectors with the highest similarity to the target. K The similarity calculation uses an improved cosine similarity, which introduces a variance weighting term for the feature vector to balance the scale differences of features of different modalities. The improved cosine similarity formula is as follows: ； in, To retrieve feature vectors for the target, Let be the candidate data feature vectors, and Σ be the covariance matrix of the feature vectors, which is obtained from the training set statistics. , The characteristic mean; Combining vector similarity with inverted index matching results, a hash bucket filtering strategy based on Locality Sensitive Hash (LSH) is adopted. Vectors are mapped to hash buckets, and precise similarity calculation is performed only within buckets with hash collisions to filter out candidate datasets. The hash function is generated by random projection of p-stable distribution. The hash function formula is: ； in, a Let b be a random vector sampled from a p-stable distribution (Cauchy distribution when p=1). Offset of uniform sampling within the area. Where is the bucket width. The hash collision condition is... .

8. The method for efficient industrial big data retrieval service based on AI algorithm according to claim 1, characterized in that, After selecting the candidate dataset, a candidate set pruning step is also included, specifically: The spatiotemporal density distribution of the candidate dataset is calculated. The density clustering algorithm based on DBSCAN is used to divide the candidate data into high-density clusters and low-density clusters. The neighborhood radius ε is determined based on the average distance of the feature vectors, and the minimum number of points MinPts is set according to the typical data density of industrial scenarios. Neighborhood radius ,in , For candidate feature vectors, For the number of candidates, Scaling factor; minimum number of points ; Representative screening of candidate data within high-density clusters is performed, and cluster center data points are selected using a strategy based on maximum marginal nearest neighbor (MMN). The centrality index is determined by the local density of the data point and its distance to higher density points. Local density Distance to higher density points Centrality indicator ,choose The largest top 20% of data points; Noise filtering is performed on candidate data within low-density clusters. An anomaly detection model based on isolated forests is used to remove data points whose anomaly scores exceed a threshold. The threshold is dynamically calibrated by the historical noise rate of the industrial scenario. Anomaly scores are also calculated using the Isolation Forest anomaly score calculation formula, with a threshold. ,in The historical noise score is the average. The standard deviation is the historical noise rate. hour, Increase by 10%, otherwise keep the original value.

9. The method for efficient industrial big data retrieval service based on AI algorithm according to claim 1, characterized in that, The process of expanding associations and performing logical reasoning on candidate datasets based on industrial knowledge graphs specifically includes: An industrial knowledge graph is constructed, which includes equipment entities, process entities, fault entities, and relationship types. The relationship types include four categories: "belongs to", "triggers", "associates", and "time series". The entity embedding method based on graph convolutional network (GCN) is used to generate the vector representation of the knowledge graph. The propagation formula for the GCN layer is: ； in, For the first l The node feature matrix of the layer, ,in A It is an adjacency matrix. It is the identity matrix. for The degree matrix, For the first l The trainable weight matrix of the layer, It is the ReLU activation function; Knowledge graph alignment is performed on entities in the candidate dataset. A dual matching strategy based on string similarity and vector similarity is adopted to link candidate entities to corresponding nodes in the knowledge graph. The string similarity is calculated using Jaro-Winkler distance, and the vector similarity is calculated using the scoring function of the TransE model. The Jaro-Winkler distance formula is: ； in, and For candidate entity and knowledge graph entity names, m To match the number of characters, t Half the number of character transpositions. l The length of the common prefix. p Prefix scaling factor; The TransE scoring function is: Double-match score in and These represent the minimum and maximum values ​​of the scoring function; Based on the description logic DL inference engine, logical reasoning is performed on the linked knowledge graph. The inference rule includes "If device A causes fault B, and fault B is associated with process C, then device A and process C are indirectly associated," generating an extended set of candidate data associations. The description logic inference formula is: ； in, Indicates conjunction. Quantifiers indicating existence; The semantic matching score between candidate data and query intent is calculated using a graph neural network model based on GraphSAGE. The input is the feature vector of the candidate data and the subgraph structure of the knowledge graph, and the output is the semantic matching score. The subgraph structure includes the 2-hop neighbor nodes and relation edges of the candidate entity. The GraphSAGE aggregation formula is: ； in, For nodes v In the k The feature vector of the layer, It is the mean aggregation function. , where is a trainable weight matrix, and is the semantic matching score. ,in For the number of GraphSAGE layers, and These are the parameters for the classification layer.

10. The method for efficient industrial big data retrieval service based on AI algorithm according to claim 1, characterized in that, The method of using a re-ranking model to perform multi-dimensional scoring fusion on candidate datasets specifically includes: A multi-dimensional scoring index system is constructed, including vector similarity score, knowledge graph association strength score, temporal context matching score, and data quality score. The data quality score is calculated by weighting data integrity, collection frequency stability, and historical verification pass rate. The formula for data quality score is: ； in, For the number of valid data points, Total number of data points The standard deviation of the sampling frequency. This is the average sampling frequency. This represents the number of times the historical verification has passed. This represents the total number of checks. For the weights, where ; A LambdaMART-based learning ranking model was adopted, with multi-dimensional scoring indicators as features and manually labeled search relevance scores as labels. The re-ranking model was trained, and the model loss function was optimized using NDCG. The formula for calculating NDCG is: ； in, IDCG@ K The DCG value under ideal sorting. For the first i The relevance score of each result; The LambdaMART loss function is: ； in, For exchanging documents i and j The change in NDCG after that, and The model predicts the score. For scaling parameters; A context-aware mechanism is introduced to obtain the user's historical query sequence, current session context and device operating status parameters in real time. The context temporal dependency is modeled through a gated recurrent unit (GRU) to generate a dynamic weight vector and perform weighted fusion of multi-dimensional scoring indicators. The hidden layer state of the GRU is adaptively updated according to the entity density of the current query. The GRU update formula is: ； ； ； ； in, To update the door, To reset the door, In the candidate hidden state, Currently in a hidden state. Entity density is a contextual feature of the current query. ,in To determine the number of entities in the query. For query length, when At that time, the hidden layer dimension of GRU ,otherwise Dynamic weight vector ; Sort the candidate datasets in descending order based on the fused scores, and output the top-ranked datasets. N The results are used as the final search results. N The system load is dynamically adjusted based on the number of returns specified by the user. N The calculation formula is: ； in, The number of returns specified by the user. The system's currently available computing resources. Resources required for processing a single result.

Images