An Industrial Big Data Analysis and Mining System and Method
Patent Information
- Application Number
- CN202611338586.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-31
- Publication Date
- 2026-11-17
AI Technical Summary
In existing technologies, there is physical storage isolation between time-series sensor data and unstructured data in industrial sites, which leads to problems such as cross-modal data query delays and retrieval results that do not match the actual operating conditions of the equipment. Furthermore, it lacks adaptive capabilities and cannot meet real-time requirements.
By constructing a graph-time-series inverted hybrid index tree, semantic vector nodes and time waveform feature nodes are organized in the same graph structure. Affine transformation matrices are used for dynamic correction. Combined with streaming prefetching and physical mechanism verification, joint retrieval and adaptive verification of cross-modal data are achieved.
It enables cross-modal data to be retrieved in a single query, reduces network transmission latency, ensures that the retrieval results are consistent with the current operating conditions of the equipment, and meets the real-time analysis needs of industrial sites.
Smart Images

Figure CN123115242A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of big data analysis and mining technology, specifically to an industrial big data analysis and mining system and method. Background Technology
[0002] With the deep integration of industrial internet and artificial intelligence technologies, industrial big data analysis and mining have become important technical means for scenarios such as predictive maintenance and intelligent quality inspection. Data generated in industrial sites has multimodal characteristics. High-frequency sensor data is continuously generated in the form of time series, while equipment fault images, maintenance records, expert diagnostic texts, etc. exist in unstructured form. To support intelligent diagnosis and decision-making, the industry usually introduces large language models combined with retrieval enhancement generation technology, which provides contextual support for model generation by retrieving similar cases in historical knowledge bases.
[0003] The closest existing technical solutions to this invention typically employ a multi-database parallel, post-processing stitching architecture: high-frequency sensor time-series data is stored in a dedicated time-series database and indexed by timestamps; unstructured data is encoded and stored in an independent vector database and indexed by semantic similarity; during joint queries, the two types of query results need to be manually or systematically aligned and stitched together at the application layer using timestamp fields. This process involves cross-database network transmission and redundant calculations, leading to increased query latency. Meanwhile, in existing technologies, graph index structures and statistical metric corrections are usually separated: once a graph index is built, it remains static; dynamic adjustments to the metric method are implemented only at the application layer and do not generate feedback with the index structure, forming an inherent technical path dependency of "static index and decoupled metric". This application breaks through the inherent path dependence by coupling the affine transformation matrix, which is updated with the baseline, with the graph index traversal process in real time. It also feeds back the physical mechanism verification results to the candidate set to form a closed-loop elimination mechanism, enabling the index structure itself to have working condition adaptive capabilities. This achieves the technical effect of simultaneously improving retrieval accuracy and real-time performance. This organic coupling method is not a simple superposition of existing graph indexing technology and statistical whitening methods. There is an urgent need for a solution that can uniformly organize multimodal data at the underlying data structure level and has working condition adaptive capabilities. Summary of the Invention
[0004] The technical problem to be solved by this invention is: how to break the physical storage isolation between time-series sensor data and unstructured data in industrial sites at the underlying data structure level, and achieve a single query joint hit of cross-modal data. At the same time, how to enable the retrieval system to perceive and compensate for the concept drift caused by aging of industrial equipment and environmental changes, so as to ensure that the retrieved historical knowledge is consistent with the current actual physical condition of the equipment, and to complete cross-modal joint retrieval and adaptive verification mining under the premise of meeting real-time requirements.
[0005] To address the aforementioned technical problems, this invention provides an industrial big data analysis and mining method, comprising: receiving time-series data streams and unstructured data collected by sensors in an industrial field, performing semantic encoding on the unstructured data to obtain a semantic feature vector, and performing time-series encoding on the time-series data stream after slicing it by a sliding window to obtain a time waveform feature vector;
[0006] The semantic feature vector and the corresponding timestamp interval and data category label are encapsulated into a semantic vector node, and the time waveform feature vector and the corresponding start and end timestamps and device number are encapsulated into a time waveform feature node. The semantic vector node and the time waveform feature node are organized in the same graph structure to obtain a graph time-series inverted hybrid index tree. Semantic adjacency edges are established between two nodes whose semantic similarity exceeds a preset similarity threshold, and time adjacency edges are established between two nodes whose corresponding timestamp intervals are less than a preset time window.
[0007] The time waveform feature vector during the normal operation period of the equipment is continuously collected, and the baseline mean vector and baseline covariance matrix are updated in a sliding manner. An affine transformation matrix is then generated based on the baseline covariance matrix.
[0008] When a retrieval request is received, the baseline mean vector is subtracted from the current query feature vector and the candidate feature vector in the graph time-series inverted hybrid index tree, respectively. Then, the affine transformation matrix is multiplied on the left to obtain the corrected query representation and the corrected candidate representation. The distance between the two is calculated and the candidate nodes are sorted accordingly to obtain a candidate diagnostic experience set.
[0009] Based on the above scheme, the present invention further provides a streaming prefetching and bidirectional verification mechanism: continuously monitor the distribution of sensor data, and when the deviation of the data distribution in the current time window from the recent baseline distribution exceeds a preset deviation threshold, trigger an asynchronous prefetching request to the graph time-series inverted hybrid index tree, and execute the prefetching process in parallel with the subsequent feature decoding calculation process;
[0010] The system calls a pre-set device physical mechanism model to verify the physical rationality of each candidate diagnostic experience in the candidate diagnostic experience set. Candidate diagnostic experiences that violate physical laws are eliminated to obtain verified candidate diagnostic experiences. These verified experiences are then summarized with the corresponding time-aligned sensor waveforms to generate a structured industrial big data mining report.
[0011] The present invention also provides an industrial big data analysis and mining system corresponding to the above method, including a multimodal coding layer, a combined index storage layer, an adaptive retrieval recalibration layer, and a streaming prefetching and verification layer, each layer sequentially performing the aforementioned coding, index construction, dynamic metric correction and streaming verification steps.
[0012] Compared with the prior art, the beneficial effects of the present invention by adopting the above technical solution are as follows:
[0013] Firstly, by organizing semantic vector nodes and time waveform feature nodes in the same graph time-series inverted hybrid index tree, and connecting the two types of nodes with semantic adjacency edges and time adjacency edges, a single graph traversal can simultaneously hit semantically related historical case nodes and time-aligned sensor waveform nodes, eliminating network transmission and result set matching delays caused by cross-database splicing.
[0014] Secondly, by performing sliding statistics on the equipment operating baseline and generating an affine transformation matrix that changes over time, the query vector and candidate vector are dynamically measured rather than measured by a fixed distance. This allows the search results to automatically deduct baseline offsets caused by natural aging of the equipment or changes in the environment, ensuring that the retrieved historical knowledge matches the current actual operating conditions of the equipment. This reduces the possibility of industrial illusion when the large language model generates diagnostic suggestions based on the search results.
[0015] Third, the index prefetching triggered by anomaly detection is executed in parallel with subsequent calculations, and the candidate diagnostic experience is verified in reverse by combining the physical mechanism model. This ensures the timeliness requirements of real-time analysis and early warning in industrial sites while guaranteeing the confidence of the retrieval results. Attached Figure Description
[0016] Figure 1 This is a schematic diagram of the overall system architecture of the present invention;
[0017] Figure 2 This is a flowchart illustrating the construction process of the graph-time-inverted hybrid index tree of the present invention.
[0018] Figure 3 This is a schematic diagram of the flow prefetching and bidirectional verification mining pipeline of the present invention. Detailed Implementation
[0019] The following is in conjunction with the appendix Figure 1-3 The specific embodiments of the present invention will be further described below. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention.
[0020] Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.
[0021] I. System Overall Architecture
[0022] The industrial big data analysis and mining system provided in this invention generally includes a data access layer, a multimodal coding layer, a joint index storage layer, an adaptive retrieval recalibration layer, a streaming prefetching and verification layer, and a large model interaction layer.
[0023] The data access layer is used to access time-series data streams and unstructured data from high-frequency sensors in industrial sites. The unstructured data includes at least one of defect images, maintenance manual text, and expert diagnostic records. The multimodal coding layer includes a time-series coding module and a semantic coding module, which convert the time-series data streams and unstructured data into feature vectors, respectively. The joint index storage layer carries a graph time-series inverted hybrid index tree. The adaptive retrieval recalibration layer includes a baseline statistics module and a dynamic metric transformation module. The streaming prefetching and verification layer includes an asynchronous prefetch scheduler and a physical mechanism reverse verification module. The large model interaction layer receives high-confidence reports output by the mining system and generates the final diagnostic conclusion.
[0024] The layers are connected sequentially according to the data flow: the output of the data access layer is sent to the multimodal coding layer, the encoding result of the multimodal coding layer is written to the joint index storage layer, the retrieval request is processed by the adaptive retrieval recalibration layer, triggering the asynchronous pipeline of the streaming prefetch and verification layer, and the final output of the streaming prefetch and verification layer is sent to the large model interaction layer.
[0025] II. Construction Method of Graph Time-Sequence Inverted Hybrid Index Tree
[0026] The graph-based time-series inverted hybrid index tree contains two types of nodes: semantic vector nodes and time waveform feature nodes. Semantic vector nodes are generated from unstructured data through a semantic encoding module and are labeled with content category and source timestamp range. Time waveform feature nodes are generated from sensor time-series segments through a time-series encoding module and are labeled with acquisition device number and start and end timestamps. Both types of nodes carry time range labels to record the coverage range of the corresponding data on the time axis. Nodes are connected by dynamic graph edges, which are divided into two types: semantic adjacency edges and temporal adjacency edges. Semantic adjacency edges connect nodes with high semantic similarity, and the edge weight is determined by vector similarity. Temporal adjacency edges connect nodes with overlapping or adjacent timestamp ranges, and the edge weight is determined by temporal adjacency decay weight. The hierarchical graph structure is modified based on the idea of approximate nearest neighbor graph layering. The upper layer is a coarse-grained navigation layer, and the lower layer is a fine-grained retrieval layer. During retrieval, the search range is gradually narrowed from top to bottom.
[0027] The specific method for constructing a graph-sequence inverted hybrid index tree is as follows.
[0028] The system first receives high-frequency sensor time-series data streams and unstructured data from the industrial site, and performs preprocessing operations such as denoising and missing value imputation on the time-series data, and performs format standardization processing on the unstructured data.
[0029] Next, the system inputs the preprocessed unstructured data into a pre-trained semantic embedding model to obtain a fixed-dimensional semantic feature vector. This semantic feature vector, along with the original timestamp interval of the data to which it belongs and the data category label, is then encapsulated into a semantic vector node and inserted into the index tree to be built. Simultaneously, the system slices the sensor time-series data stream into a fixed-length sliding window. Each slice is encoded by a time-series coding model trained based on sliding window contrastive learning to obtain a time waveform feature vector. This time waveform feature vector, along with the start and end timestamps corresponding to the slice and the device number, is then encapsulated into a time waveform feature node and inserted into the index tree. The generation processes of semantic vector nodes and time waveform feature nodes can be executed in parallel.
[0030] Then, the system calculates the semantic similarity and temporal adjacency of any two nodes in the index tree: when the semantic similarity between the two nodes exceeds a preset similarity threshold, a semantic adjacency edge is established, and the edge weight is taken from the similarity value; when the timestamp interval between the two nodes is less than a preset time window, a temporal adjacency edge is established, and the edge weight is calculated according to the temporal adjacency decay weight formula. The calculation method of the temporal adjacency decay weight is as follows:
[0031] ;
[0032] in, Represents a node With nodes The weight of time-adjacent edges between them. , Representing nodes respectively ,node For the timestamp of the corresponding data, when the node corresponds to a time interval, the midpoint of that interval is taken. This is the time decay coefficient, used to control the rate at which the weight decays as the time interval increases. The value can be preset according to the time scale of fault propagation in a specific industrial scenario.
[0033] Finally, the system performs layered pruning on the fully connected candidate edge set constructed above using the approximate nearest neighbor graph layering algorithm, retaining only the edges with the highest degree of each node in each layer, forming a multi-layer graph structure from coarse-grained to fine-grained, and persistently storing this multi-layer graph structure together with the time range label index, thus completing the construction of the graph time-series inverted hybrid index tree.
[0034] Through the aforementioned joint index data structure, a single query request can simultaneously hit semantically similar historical fault case nodes via semantic adjacency edges and sensor waveform nodes aligned with the time of the case via temporal adjacency edges during graph traversal. This allows cross-modal joint retrieval to be completed within a single underlying data structure, avoiding the network overhead and splicing delay of cross-database queries.
[0035] III. An Adaptive Retrieval Recalibration Method Based on Dynamic Concept Drift
[0036] The adaptive retrieval recalibration method includes a baseline statistics module, a dynamic anchor point update module, a transformation matrix generation module, a query vector correction module, and a distance calculation and sorting module. The baseline statistics module continuously receives the feature vectors of the current normal operating period of the device and maintains a sliding window. This module is marked with a feedback loop, indicating that it continuously updates as new data arrives, forming a closed-loop adaptive process.
[0037] Specifically, within the normal operating time window of the device, the system continuously collects time waveform feature vectors and updates the baseline mean vector and baseline covariance matrix in an exponential sliding manner, as follows:
[0038] ;
[0039] ;
[0040] in, Indicates the first The baseline mean vector updated at time step. Indicates the first The baseline covariance matrix updated at time step 1. Indicates the first The waveform feature vector of the device's normal operating time, collected at any time. This represents the forgetting coefficient for sliding updates, with values between 0 and 1. A larger value indicates a higher weight of the historical baseline and a slower response to mutations. (Superscript) This represents the transpose of a vector.
[0041] Based on the aforementioned baseline covariance matrix, the system generates an affine transformation matrix for whitening through matrix inverse square root operations:
[0042] ;
[0043] in, Indicates the first The affine transformation matrix generated at time t, Represents the baseline covariance matrix The inverse square root, this operation is performed by... The eigenvalues are reconstructed by taking the inverse square root of the eigenvalues after eigenvalue decomposition.
[0044] When a retrieval request arrives, the system will use the current query feature vector. Candidate feature vectors in the historical knowledge base After subtracting the corresponding baseline mean vectors, multiply by the transformation matrix on the left. The corrected representation is obtained, and the dynamic distance between the corrected query vector and the candidate vector is calculated using the following formula:
[0045] ;
[0046] in, This represents the query vector under the current baseline correction. With candidate vectors The dynamic distance between them For the affine transformation matrix generated above, This is the current query feature vector. These are candidate feature vectors from the historical knowledge base. This represents the squared value of the Euclidean norm of a vector.
[0047] The system calculates the dynamic distances mentioned above for all candidate nodes in the historical knowledge base, sorts them in ascending order of distance, and selects the top-ranked nodes as the candidate diagnostic experience set for this retrieval. Due to the transformation matrix... The baseline changes in real time as it slides and updates, so the mechanism can automatically deduct the baseline offset caused by natural aging of equipment or changes in the environment, so that the search results always reflect the true similarity relationship under the current physical conditions, without blocking the real-time streaming search process.
[0048] IV. Flow Prefetching and Two-Way Verification Excavation Production Line
[0049] The streaming prefetching and bidirectional verification mining pipeline includes an anomaly detection triggering module, an asynchronous prefetch scheduler, a multimodal parallel retrieval module, a physical mechanism reverse verification module, and a high-confidence report generation module. The anomaly detection triggering module and the asynchronous prefetch scheduler are marked with overlapping execution intervals, indicating that they are executed in parallel on the time axis. The physical mechanism reverse verification module and the multimodal parallel retrieval module are marked with bidirectional arrows, indicating that the verification results can be fed back to the retrieval module to eliminate unreasonable candidates.
[0050] Specifically, the system continuously monitors the distribution of the streaming sensor data. When the deviation of the data distribution within the current time window from the recent baseline distribution exceeds a preset deviation threshold, the system determines that the trigger condition has been met. Once the trigger condition is met, the system immediately initiates an asynchronous prefetch request to the graph time-series inverted hybrid index tree of the composite index storage layer. The relevant index fragments are loaded into memory in advance through memory mapping and prefetching mechanisms. This prefetching process is executed in parallel and overlapped with the subsequent feature decoding calculation process to mask the disk input / output time consumption.
[0051] Subsequently, the system simultaneously retrieves semantically relevant nodes and time-aligned nodes in the joint index tree according to the aforementioned graph traversal method, and performs distance calculation and sorting on the candidate results according to the aforementioned dynamic metric recalibration mechanism to obtain a preliminary candidate diagnostic experience set.
[0052] Furthermore, the system calls a pre-set equipment physical mechanism model. This model is a parameter verification rule base pre-built based on the equipment's known mechanical vibration transmission equation, thermodynamic balance equation, or industry experience formula. It takes at least one of the equipment's measured speed, load, and temperature as input and outputs the judgment results of whether the fault causes and parameter ranges implied by the candidate diagnostic experiences meet the preset physical constraints. The system verifies the physical rationality of each of the above candidate diagnostic experiences. If a candidate experience violates the physical laws of the current operating condition, it is removed from the candidate set.
[0053] Finally, the system will generate a structured industrial big data mining report by summarizing the verified candidate diagnostic experiences and their corresponding time-aligned sensor waveforms, and use this report as context input to the upper-level big language model to generate diagnostic conclusions or maintenance suggestions.
[0054] Through the asynchronous parallel design of the above modules, the present invention can overlap the time consumption of each link of anomaly detection, index prefetching, multimodal retrieval and physical verification, thereby meeting the timeliness requirements of real-time analysis and early warning of industrial production lines while ensuring retrieval accuracy.
[0055] V. Specific Application Examples
[0056] Taking a wind turbine bearing vibration monitoring scenario as an example, the implementation process of this invention is explained. The system continuously receives the time-series data stream of the vibration sensor of the wind turbine bearing, as well as unstructured data such as historical maintenance records and fault images. The multimodal coding layer encodes the above data respectively, and the generated semantic vector nodes are labeled with the timestamp interval and defect category of the maintenance record. The generated time waveform feature nodes are labeled with the start and end time of vibration signal acquisition and equipment number. The joint index storage layer organizes the above nodes into a graph time-series inverted hybrid index tree. Semantic adjacency edges are established between semantically similar historical maintenance record nodes, and time adjacency edges are established between vibration signal nodes with adjacent timestamps and maintenance record nodes.
[0057] During the long-term operation of the wind turbine, the bearing wear caused the vibration baseline to gradually rise. The adaptive retrieval and recalibration layer continuously collected the vibration feature vectors during the normal operation of the bearing, updated the baseline mean vector and baseline covariance matrix in an exponential sliding manner, and updated the affine transformation matrix accordingly. When a significant deviation of the vibration data distribution from the recent baseline was detected at a certain moment, the streaming prefetching and verification layer immediately initiated asynchronous prefetching to the index tree and called the updated affine transformation matrix in the multimodal parallel retrieval module to perform dynamic distance calculation on the current query vector and the candidate vectors in the historical knowledge base. This allowed the retrieval results to automatically deduct the baseline rise caused by bearing wear, avoiding misjudging historical normal cases before wear as abnormal matching cases. The physical mechanism reverse verification module further verified the rationality of the recalled candidate fault experiences based on the physical mechanism model of the wind turbine bearing, eliminating candidate experiences that did not conform to the physical laws of the current measured operating conditions such as speed and load. Finally, a structured high-confidence diagnostic report was generated and sent to the large language model for generating maintenance suggestions.
[0058] VI. Other Implementation Methods
[0059] In some alternative implementations, when encoding time-series data slices, in addition to using a time-series coding model trained based on sliding window contrastive learning, feature extraction of sensor data slices based on autoencoders or time-series Transformers can also be used to obtain equivalent time waveform feature vectors. The remaining joint index construction, dynamic recalibration, and pipeline verification steps remain unchanged, and the same technical effects of cross-modal joint retrieval and concept drift compensation can be achieved.
[0060] In other alternative implementations, in addition to using an affine transformation matrix based on the inverse square root of covariance, a lightweight neural network can be used to learn a nonlinear metric mapping function online. The mapping parameters are dynamically generated using the statistical features of the device baseline as input. The distance is then calculated after nonlinear correction of the query vector and the candidate vector. This alternative approach can also achieve adaptive compensation for concept drift.
[0061] Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of the present invention. Therefore, any modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the protection scope defined by the claims of the present invention.
Claims
1. An industrial big data analysis and mining method, characterized in that, include: The system receives time-series data streams and unstructured data collected by sensors in the industrial field, performs semantic encoding on the unstructured data to obtain semantic feature vectors, and performs time-series encoding on the time-series data stream after slicing it by a sliding window to obtain time waveform feature vectors. The semantic feature vector and the corresponding timestamp interval and data category label are encapsulated into a semantic vector node, and the time waveform feature vector and the corresponding start and end timestamps and device number are encapsulated into a time waveform feature node. The semantic vector node and the time waveform feature node are organized in the same graph structure to obtain a graph time-series inverted hybrid index tree. Semantic adjacency edges are established between two nodes whose semantic similarity exceeds a preset similarity threshold, and time adjacency edges are established between two nodes whose corresponding timestamp intervals are less than a preset time window. The time waveform feature vector during the normal operation period of the equipment is continuously collected, and the baseline mean vector and baseline covariance matrix are updated in a sliding manner. An affine transformation matrix is then generated based on the baseline covariance matrix. When a retrieval request is received, the baseline mean vector is subtracted from the current query feature vector and the candidate feature vector in the graph time-series inverted hybrid index tree, respectively. Then, the result is multiplied by the affine transformation matrix to obtain the corrected query representation and the corrected candidate representation. The distance between the corrected query representation and the corrected candidate representation is calculated. The candidate nodes corresponding to the candidate feature vectors are sorted according to the distance to obtain a candidate diagnostic experience set.
2. The method of claim 1, wherein, The weights of the temporal adjacent edges are determined according to the following formula: ; wherein, denotes a node between nodes , , denotes a node , a node corresponding data, the midpoint of the time interval when the node corresponds to a time interval, is a time decay coefficient for controlling the rate of decay of the weight with increasing time interval.
3. The method of claim 1, wherein, Organizing the semantic vector nodes and the temporal waveform feature nodes in the same graph structure further includes: pruning the candidate edge set composed of the semantic adjacency edges and the temporal adjacency edges according to the approximate nearest neighbor graph layering algorithm, retaining the edge with the highest degree of each node in each layer, resulting in a multi-layer graph structure composed of a coarse-grained navigation layer and a fine-grained retrieval layer; when a retrieval request is received, the search range is narrowed layer by layer from the coarse-grained navigation layer to the fine-grained retrieval layer to determine the candidate feature vectors.
4. The method according to claim 1, characterized in that: The baseline mean vector and baseline covariance matrix are updated using a sliding method, including updates according to the following formula: ; ; in, Indicates the first The baseline mean vector updated at time step. Indicates the first The baseline covariance matrix updated at time step 1. Indicates the first The waveform feature vector of the device's normal operating time, collected at any time. This represents the forgetting coefficient for sliding updates, with values between 0 and 1, indicated by the superscript. Indicates vector transpose; Generating an affine transformation matrix based on the baseline covariance matrix includes: performing eigenvalue decomposition on the baseline covariance matrix, reconstructing the eigenvalues by taking the inverse square root, obtaining the inverse square root matrix of the baseline covariance matrix, and determining the inverse square root matrix as the affine transformation matrix.
5. The method of claim 1, wherein, Also includes: The distribution of sensor data is continuously monitored. When the deviation of the data distribution in the current time window from the recent baseline distribution exceeds a preset deviation threshold, an asynchronous prefetch request is triggered to the graph time-series inverted hybrid index tree to load the relevant index fragments into memory. The asynchronous prefetch request is executed in parallel with the subsequent feature decoding calculation process. The preset equipment physical mechanism model is invoked to verify the physical rationality of each candidate diagnostic experience in the candidate diagnostic experience set. It is determined whether the fault causes and parameter ranges implied by the candidate diagnostic experience are consistent with the physical laws of the current measured working conditions. Candidate diagnostic experiences that violate the physical laws are removed from the candidate diagnostic experience set to obtain the verified candidate diagnostic experiences. The verified candidate diagnostic experiences and corresponding time-aligned sensor waveforms are summarized to generate a structured industrial big data mining report.
6. The method of claim 1, wherein, The time-series data stream is sliced into sliding window segments and then time-series encoded, including: using a time-series encoding model trained based on sliding window contrastive learning to encode each sliding window slice to obtain a fixed-dimensional time waveform feature vector.
7. An industrial big data analysis and mining system, characterized in that, include: A multimodal coding layer is used to receive time-series data streams and unstructured data collected by sensors in the industrial field, perform semantic encoding on the unstructured data to obtain semantic feature vectors, and perform time-series encoding on the time-series data stream after slicing it by a sliding window to obtain time waveform feature vectors. A composite index storage layer is used to encapsulate the semantic feature vector with the corresponding timestamp interval and data category label into a semantic vector node, and to encapsulate the time waveform feature vector with the corresponding start and end timestamps and device number into a time waveform feature node. The semantic vector node and the time waveform feature node are organized in the same graph structure to obtain a graph time-series inverted hybrid index tree. Semantic adjacency edges are established between two nodes whose semantic similarity exceeds a preset similarity threshold, and time adjacency edges are established between two nodes whose corresponding timestamp intervals are less than a preset time window. An adaptive retrieval recalibration layer is used to continuously collect time waveform feature vectors during normal device operation, update the baseline mean vector and baseline covariance matrix in a sliding manner, and generate an affine transformation matrix based on the baseline covariance matrix. Upon receiving a retrieval request, it is also used to subtract the baseline mean vector from the current query feature vector and the candidate feature vectors in the graph time-series inverted hybrid index tree, respectively, and then multiply them by the affine transformation matrix to obtain a corrected query representation and a corrected candidate representation. The distance between the corrected query representation and the corrected candidate representation is calculated, and the candidate nodes corresponding to the candidate feature vectors are sorted according to the distance to obtain a candidate diagnostic experience set.
8. The system of claim 7, wherein, It also includes a streaming prefetch and verification layer, which is used for: The distribution of sensor data is continuously monitored. When the deviation of the data distribution in the current time window from the recent baseline distribution exceeds a preset deviation threshold, an asynchronous prefetch request is triggered to the graph time-series inverted hybrid index tree to load the relevant index fragments into memory. The asynchronous prefetch request is executed in parallel with the subsequent feature decoding calculation process. The preset device physical mechanism model is invoked to verify the physical rationality of each candidate diagnostic experience in the candidate diagnostic experience set. Candidate diagnostic experiences that violate physical laws are removed from the candidate diagnostic experience set to obtain verified candidate diagnostic experiences. The verified candidate diagnostic experiences and corresponding time-aligned sensor waveforms are summarized to generate a structured industrial big data mining report.
9. The system of claim 7, wherein, The joint index storage layer is also used to prune the candidate edge set composed of the semantic adjacency edge and the temporal adjacency edge according to the approximate nearest neighbor graph layering algorithm, retaining the edge with the highest degree of each node in each layer, to obtain a multi-layer graph structure composed of a coarse-grained navigation layer and a fine-grained retrieval layer.
10. The system of claim 7, wherein, The weights of the temporal adjacent edges are determined according to the following formula: ; in, Represents a node With nodes The weight of time-adjacent edges between them. , Representing nodes respectively ,node The timestamp of the corresponding data is used, and when a node corresponds to a time interval, the midpoint of that time interval is taken. This is the time decay coefficient, used to control the rate at which the weight decays as the time interval increases.
