Anomaly cause tracing and semantic retrieval integrated hybrid index method and system

Abstract

This invention belongs to, but is not limited to, the field of shale gas production technology, and particularly relates to a hybrid indexing method and system integrating anomaly causal tracing and semantic retrieval, comprising: S1: multimodal data acquisition and characterization extraction; S2: hybrid index construction; S3: retrieval and anomaly causal tracing execution. It can efficiently integrate time-series, text, and structured data from shale gas production scenarios, achieving cross-modal collaborative retrieval. Technical support: Through a full-link design of multimodal data acquisition → unified characterization extraction (structured attributes + vector features) → unified index structure of attribute layer + vector layer, SCADA time-series data, maintenance and repair text work orders, and well log structured data, which were originally stored independently, are associated through unique identifiers, allowing for simultaneous retrieval of multiple data types in a single search, avoiding the problem of one-dimensional data bias.

Description

Technical Field

[0001] This invention belongs to, but is not limited to, the field of shale gas production technology, and particularly relates to a hybrid indexing method and system that integrates anomaly causal tracing and semantic retrieval. Background Technology

[0002] Shale gas, as an important unconventional natural gas resource, has been developed and utilized on a large scale in recent years. With increasing production years, shale gas wells in most areas of China are gradually entering the aging stage, resulting in fluctuating and unstable production capacity. According to statistics for the entire year of 2024, the Southwest Oil and Gas Field had an average of 10 wells experiencing liquid accumulation per day, with an average well recovery cycle of 2.5 days and an average daily production of 15,000 cubic meters. This resulted in a daily production loss of up to 150,000 cubic meters, impacting annual production by approximately 55 million cubic meters. Improving the response efficiency to abnormal conditions such as liquid accumulation and enhancing the intelligence level of production operation and maintenance have become urgent problems to be solved in current shale gas production and operation.

[0003] In existing technologies, the identification and handling of anomalies such as wellbore fluid accumulation, compressor shutdown, and pipeline blockage during shale gas production still mainly rely on manual experience and static threshold alarm mechanisms set in SCADA (Supervisory Control and Data Acquisition) or POC (Production Operation Control) systems. This type of method has the following significant drawbacks: 1. Fixed rules and poor generalization ability: The SCADA / POC system alarms based on manually set rules, which cannot adapt to the complex working conditions brought about by the evolution of wellbore status over time and environmental changes, and is prone to missed alarms or false alarms.

[0004] 2. Reliance on manual analysis and delayed response: Current anomaly assessment still relies on the experience and manual analysis of on-site engineers, which is inefficient and makes it difficult to achieve accurate and rapid dynamic management.

[0005] 3. Low data utilization: Although downhole sensors and surface monitoring equipment have achieved the acquisition of a large amount of time-series data, the existing system has failed to fully explore the dynamic evolution patterns hidden within them.

[0006] 4. Lack of model capabilities: There is a lack of intelligent models with self-learning and automatic feature extraction capabilities, making it difficult to establish a detection system with high robustness and adaptability in multi-well and multi-dimensional data environments.

[0007] Based on the above analysis, the urgent technical problems that need to be solved in the existing technology are: multimodal data silos, low retrieval efficiency and insufficient accuracy, unexplainable causes of anomalies, poor adaptability to general technical scenarios, and lack of cross-well collaborative analysis in the current shale gas production and operation and maintenance. The ultimate goal is to achieve collaborative data retrieval, accelerated operation and maintenance response, traceability of anomaly causes, and block-level operation and maintenance decision support. Summary of the Invention

[0008] To address the problems existing in the prior art, this invention provides a hybrid indexing method and system that integrates anomaly causal tracing and semantic retrieval.

[0009] This invention is implemented as follows: a hybrid indexing method integrating anomaly causal tracing and semantic retrieval, comprising: S1: Multimodal data acquisition and differentiated event serialization representation.

[0010] This step aims to transform isolated data into a unified timeline event stream that includes business logic.

[0011] Multi-source data standardization: Access SCADA time-series data, alarm records, maintenance work orders, and well logs.

[0012] Time-series data: Collects real-time / historical data such as oil pressure and sleeve pressure from the SCADA system.

[0013] Text data: Summarizes unstructured text such as alarm records, maintenance work orders, and operation logs.

[0014] Structured data: Import structured tabular data such as well logs, process parameter tables, and equipment attributes.

[0015] Definitions and representations of three types of events: Sensor operating condition events: For time-series data such as oil pressure and casing pressure, feature engineering is used to capture sudden trend changes (such as sudden pressure drop or liquid level fluctuation). When these changes cannot be explained by previous operations, they are characterized as "sensor abnormal change events" as proxy labels for unknown downhole conditions.

[0016] External / cross-well interference events: Extract external events not controlled by this well, such as pressure surge (adjacent well measures), compressor shutdown, desander / pipeline blockage, and changes in external transmission pressure, from the logs and alarms.

[0017] Process system adjustment events: Identify proactive operation events such as process changeover, foaming dosage adjustment, and plunger system adjustment from ledgers and operation records.

[0018] Event chain integration: Merge the above three types of events according to timestamps to generate a production evolution sequence.

[0019] S2: Hybrid index construction and sequence correlation modeling.

[0020] Build an underlying architecture that supports complex attribution logic.

[0021] Hybrid index architecture: Build an attribute filtering engine (for hash and time filtering) and a vector semantic index (for event pattern similarity retrieval).

[0022] Sequence causal modeling: The sequence model Transformer is used to learn the temporal correlation and logistic probability between the "operation / external event sequence" and the "abnormal operating condition event".

[0023] Attribution Logic Library: Preset attribution rules to distinguish between "measures / interference" causes that can be directly mapped and "sensor / downhole" causes that require manual intervention.

[0024] S3: Hierarchical retrieval and abnormal causal attribution execution.

[0025] Automatic anomaly classification and attribution are achieved through a three-step process: Composite query recall: Upon receiving a query request, the system uses structured metrics and vector indexes to retrieve candidate anomaly fragments and preceding event chains.

[0026] Differentiated attribution decision: 1) Direct attribution path: If the model identifies an anomaly that is highly correlated with the sequence of events of “process adjustment” or “external interference” (correlation score exceeds the preset threshold), the measure or event is directly output as the root cause (e.g., caused by adjustment of the plunger system).

[0027] 2) Sensor attribution path: If the anomaly is only reflected in the evolution of sensor characteristics and there is no obvious related operation in the preceding sequence, considering the unmonitorable nature of the shale gas downhole environment, the system will automatically attribute the cause to "abnormal changes in sensor values" and mark it as an unknown downhole condition.

[0028] Traceability results generation: It automatically generates natural language explanations; for direct attribution, it provides operational conclusions; for sensor attribution, it prompts manual analysis in conjunction with other geological information.

[0029] S1: Multimodal data acquisition and characterization extraction.

[0030] S2: Hybrid index construction.

[0031] S3: Retrieval and anomaly causal tracing execution.

[0032] Furthermore, S1 specifically includes: first, completing the standardization processing of multi-source data from shale gas production scenarios to provide a high-quality data foundation for subsequent index construction.

[0033] Multimodal data acquisition: Access to three types of core data, covering the entire production process. The core data includes: Time-series data: Collects real-time / historical data such as oil pressure and sleeve pressure from the SCADA system.

[0034] Text data: Summarizes unstructured text such as alarm records, maintenance work orders, and operation logs.

[0035] Structured data: Import structured tabular data such as well logs, process parameter tables, and equipment attributes.

[0036] Data representation extraction: For different modal data, generate feature forms that adapt to the index; different modal data include: structured data: extract attribute fields such as well number, process type, and timestamp to form structured attribute features.

[0037] Time series data: Time series feature vectors are generated through feature engineering.

[0038] Text data: It is transformed into semantic vectors using an NLP model, preserving the core meaning of the text.

[0039] Furthermore, S2 specifically includes: building a hybrid index with a triple engine and a unified structure to achieve unified retrieval support for multi-dimensional data. Specific steps include: Step 1: Build a triple engine: Build an index based on the attribute fields of structured data to support exact matching queries; store time-series feature vectors and text semantic vectors in a vector database to build a vector index to support similarity retrieval; preset core filtering conditions for shale gas scenarios such as time periods, abnormal tags, and similar wells to form a filtering rule base and quickly narrow down the search scope.

[0040] Step 2: Design a unified index structure: Adopt a two-layer architecture of attribute layer + vector layer. The upper layer connects to the attribute filtering engine and structured index to achieve accurate filtering; the lower layer connects to the vector semantic index to support similarity calculation. The two layers of data are linked by a unique identifier to ensure cross-layer collaborative access during retrieval.

[0041] Furthermore, S3 specifically includes: a three-step process of retrieval, rearrangement, and tracing to meet complex operational needs while ensuring the interpretability of abnormal results. The specific steps include: Step 1: Cross-modal composite query recall: Receive user query requests, first filter by time period + device + abnormal tags through the attribute filtering engine, then match the target well with the structured index, and finally call the vector semantic index to expand the scope and obtain candidate search results.

[0042] Step 2: Two-stage reordering optimization: Filter candidate results according to preset rules and remove irrelevant data; call the index data + text data fusion scoring model, combine well anomaly frequency, work order urgency, vector similarity and other dimensions to calculate the comprehensive score, and sort the results in descending order of score to improve accuracy.

[0043] Step 3: Anomaly Causal Tracing: Map the anomaly nodes output by the anomaly detection model to the shale gas domain ontology knowledge graph; trace the upstream and downstream related entities of the anomaly nodes through graph traversal algorithms; automatically convert the tracing links into natural language interpretations, while also supporting users to initiate follow-up questions through semantic interfaces.

[0044] Another objective of this invention is to provide a hybrid indexing system that integrates anomaly causality tracing and semantic retrieval, thereby realizing the hybrid indexing method that integrates anomaly causality tracing and semantic retrieval, comprising: The multimodal data processing module is used for data acquisition, cleaning, and characterization extraction, and outputs structured attributes and vector features.

[0045] The hybrid index management module is used to build and update structured indicator indexes and vector semantic indexes, and maintain the attribute filtering rule base.

[0046] The retrieval service module is used to process user query requests, execute the recall-reordering process, and return the sorted retrieval results.

[0047] The causal tracing module is used to manage the ontology knowledge graph and perform anomaly mapping, association tracing, and natural language interpretation generation.

[0048] Another object of the present invention is to provide a computer device including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the hybrid indexing method integrating anomaly causal tracing and semantic retrieval.

[0049] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the hybrid indexing method that integrates anomaly causal tracing and semantic retrieval.

[0050] Another objective of this invention is to provide an information data processing terminal, which includes the aforementioned hybrid indexing system that integrates anomaly causal tracing and semantic retrieval.

[0051] Based on the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solution to be protected by this invention are as follows: 1. Strong multimodal data collaboration capabilities, completely breaking down data silos.

[0052] Core Advantages: It can efficiently integrate time-series, text, and structured data from shale gas production scenarios, enabling cross-modal collaborative retrieval. Technical Support: Through a full-chain design of multi-modal data acquisition → unified representation extraction (structured attributes + vector features) → unified index structure of attribute layer + vector layer, it allows previously independently stored SCADA time-series data, maintenance and repair text work orders, and well log structured data to be linked through unique identifiers. This enables simultaneous retrieval of multiple data types in a single search, avoiding the limitations of single-dimensional data.

[0053] 2. High search efficiency and accuracy, significantly reducing maintenance time.

[0054] Core advantages: Balancing search speed and result quality, it can quickly narrow down the search scope while accurately locating core information. Technical support.

[0055] Efficiency Guarantee: The structured index and attribute filtering engine first complete the initial screening, quickly eliminating irrelevant data and avoiding full-scale retrieval.

[0056] Accuracy Guarantee: Vector semantic index captures data similarity features, coupled with a two-stage recall-reordering mechanism (first expanding the candidate range and then optimizing the ranking through a fusion scoring model), the accuracy of core results is significantly improved compared with traditional solutions, and the retrieval time is reduced from hours to minutes.

[0057] 3. The causal explanation of anomalies is strong, solving the pain point of black box detection.

[0058] Core Advantages: Not only can it identify anomalies, but it can also automatically trace the causes of anomalies and generate easy-to-understand conclusions, eliminating the need for manual investigation. Technical Support: Using a shale gas ontology knowledge graph as the core carrier, after mapping anomaly nodes (such as pressure anomalies) to the graph, it traces upstream and downstream related entities (such as anomaly → equipment → work order) through graph traversal algorithms and automatically converts them into natural language explanations (such as pressure anomalies caused by upstream equipment maintenance failure), making previously inexplicable black-box detection results visible and traceable.

[0059] 4. High adaptability to shale gas scenarios, requiring no secondary modifications.

[0060] Core advantages: Specifically designed for shale gas production and operation, directly matching the industry's specific needs and avoiding the incompatibility of general technologies. Technical support.

[0061] The system presets core shale gas search dimensions such as time period, process, anomaly tags, and similar wells, eliminating the need for users to define complex conditions.

[0062] It focuses on the correlation logic of entities in fields such as wells, skids, processes, and abnormal events, and highlights the integration of abnormal causal tracing and semantic retrieval. Compared with general retrieval technologies, it saves the cost of secondary adaptation for industry scenarios.

[0063] 5. It has outstanding cross-well collaborative analysis capabilities, supporting block-level decision-making.

[0064] Core Advantages: It overcomes the limitations of single-well analysis, quickly pinpointing common causes of anomalies in multiple wells and supporting block-level operation and maintenance strategy adjustments. Technical Support: It supports parallel retrieval of multi-well data, enabling simultaneous access to multimodal data from multiple wells. During the causal tracing phase, it uses knowledge graphs to link multiple wells with common factors (e.g., multiple well anomalies → same process → same supplier materials), quickly identifying block-level common problems (e.g., process and material mismatch), providing data support for overall operation and maintenance optimization.

[0065] This invention proposes an integrated multimodal hybrid retrieval and causal tracing technology solution for complex industrial scenarios such as shale gas, which has significant value in engineering applications, industrial transformation, and technological paradigms. From an economic and commercial perspective, this invention transforms the original process of anomaly retrieval and cause analysis, which relied on repeated manual queries across multiple systems and experience-based judgments, into a minute-level automated retrieval and causal analysis process. This significantly reduces fault location and decision-making response time, substantially lowering system maintenance and expert labor costs, saving thousands of hours of maintenance manpower annually, and effectively reducing production losses caused by the inability to locate anomalies in a timely manner, with an estimated reduction of over 30 million cubic meters of production loss, resulting in significant economic benefits. Furthermore, this invention adopts a modular architecture design, enabling seamless integration with existing production management platforms, data platforms, and monitoring systems. It boasts low deployment costs, strong adaptability, and excellent scalability and industrialization potential.

[0066] From a technological originality perspective, existing industrial retrieval technologies are generally limited to single-modal retrieval or simple vector matching, lacking a unified hybrid indexing mechanism for time-series data, text data, structured parameters, and event data in shale gas scenarios. They also lack the ability to deeply integrate retrieval results with causal graphs to achieve automatic causal tracing. This invention, for the first time, realizes an integrated collaborative architecture of a multimodal hybrid indexing engine, a causal reasoning engine, and a knowledge graph engine, opening up the entire process chain of "retrieval—association—interpretation—tracing," filling a long-standing systemic technological gap in the field of industrial anomaly tracing and retrieval.

[0067] From a problem-solving perspective, fragmented multimodal data, data silos across systems, and the inexplicable causes of anomalies have long been core challenges in the field of industrial intelligence, proving difficult to effectively address. While the industry generally aspires to achieve unified retrieval and automatic traceability, bottlenecks such as complex index structures, difficulties in modal fusion, and high costs of causal modeling have kept related solutions largely confined to theoretical or localized applications. This invention, by introducing a triple-engine collaborative mechanism and using knowledge graphs as semantic and causal links, achieves for the first time a practically implementable unified retrieval and automatic traceability closed loop in engineering, substantially breaking through the aforementioned long-standing constraints.

[0068] Meanwhile, this invention systematically overcomes several long-standing technical biases in the field of industrial information technology, such as the belief that industrial retrieval requires only a single modality, that causal tracing is difficult to automate, and that vector indexes are unsuitable for structured filtering. Practical results of this invention demonstrate that hybrid indexing and causal reasoning are not only feasible but also possess higher efficiency, accuracy, and interpretability in complex industrial scenarios, thus achieving a significant paradigm shift and cognitive update in the technical approach. Attached Figure Description

[0069] Figure 1 This is a flowchart of a hybrid indexing method that integrates anomaly causal tracing and semantic retrieval, provided in an embodiment of the present invention.

[0070] Figure 2 This is a schematic diagram of multimodal data acquisition and characterization extraction provided in an embodiment of the present invention.

[0071] Figure 3 This is a structural diagram of a hybrid indexing system that integrates anomaly causal tracing and semantic retrieval, provided in an embodiment of the present invention.

[0072] Figure 4 This is a comparison chart of anomaly retrieval efficiency provided in an embodiment of the present invention.

[0073] Figure 5 This is a diagram showing the impact of annual gas production loss provided in an embodiment of the present invention.

[0074] Figure 6 This is a comparison chart of the accuracy of abnormal root cause attribution provided in the embodiments of the present invention.

[0075] Figure 7 This is a diagram showing the reduction in operation and maintenance costs (working hours) provided in an embodiment of the present invention.

[0076] Figure 8 This is an anomaly attribution category distribution diagram provided in an embodiment of the present invention.

[0077] Figure 9 This is a system robustness graph under different query complexities provided in the embodiments of the present invention.

[0078] Figure 10 This is a time evolution effect diagram of the production loss control provided in an embodiment of the present invention. Detailed Implementation

[0079] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0080] In shale gas production scenarios, signal data mainly originates from SCADA systems and field sensor networks, and is essentially a high-frequency, continuously sampled multivariate time series. To eliminate the noise and heterogeneity effects caused by the complex acquisition environment and frequent switching of operating conditions, this method first performs hierarchical preprocessing and feature reconstruction on the raw signal data to ensure the reliability of subsequent hybrid index construction and causal tracing.

[0081] The acquired raw signal data undergoes time alignment and integrity verification. To address issues such as inconsistent sampling periods, time drift, or missing data across different sensors, a unified time axis is used to resample multi-source signals, and short-term missing intervals are repaired using linear interpolation or forward padding. Signal segments with consecutive missing durations exceeding a preset threshold are marked as low-confidence intervals and are subject to weight reduction or removal during subsequent feature calculations, thereby preventing abnormal interpolation from interfering with model learning.

[0082] Noise suppression and outlier screening were performed on the signal data. Based on the characteristics of shale gas production conditions, physically reasonable ranges were set for key parameters such as pressure, flow rate, and temperature. A combination of amplitude limiting filtering and median filtering was used to remove abrupt changes that significantly exceeded the equipment's operating boundaries. Simultaneously, an anomaly detection strategy based on local statistical characteristics was introduced. The signal distribution within the sliding time window was analyzed to identify isolated anomalies caused by sensor jitter and transient interference, and these were marked as candidate anomalies rather than being directly input into subsequent models.

[0083] After basic cleaning, the signal data undergoes feature engineering to map it from a "continuous signal" to an "event-level representation." Specifically, the signal sequence is segmented within a preset time window, and multidimensional statistical features are calculated within each window, including mean, standard deviation, extreme value range, rate of change, and first- and second-order difference features, to characterize the signal's level, volatility, and trend changes. For parameters with clear technological implications, structured features such as threshold crossing count and duration are further extracted to enhance the ability to express changes in operating conditions.

[0084] The above features are combined and mapped into a time-series feature vector of a unified dimension. Each vector is then appended with a corresponding timestamp, hash number, and operating context information, thus forming a standardized "sensor operating event." This event retains the dynamic characteristics of the original signal while possessing discretized and indexable structural features, making it suitable for fusion modeling with text events and structured events within a unified event space.

[0085] The generated sensor operating condition events are normalized and semantically enhanced. By standardizing or normalizing features of different dimensions, parameter scale differences are eliminated. At the same time, a comparative coding method based on historical operating condition distribution is introduced, so that the feature vector can reflect the degree of deviation of the current operating condition from the historical benchmark, providing a more sensitive and stable signal representation basis for subsequent anomaly correlation modeling and causal attribution.

[0086] like Figure 1 As shown, the implementation of the method is divided into three major stages: data preprocessing, hybrid index construction, and retrieval and traceability execution. Each stage is closely connected to form a closed loop.

[0087] S1: Multimodal data acquisition and differentiated event serialization representation.

[0088] Collect time-series, text, and structured data from the shale gas production process and convert them into a unified event sequence with timestamps.

[0089] The event sequence is categorized by source as follows: sensor operating condition events, external interference events, and process regulation adjustment events.

[0090] S2: Hybrid index construction and sequence correlation modeling.

[0091] Construct a hybrid architecture consisting of attribute filtering indexes and vector semantic indexes.

[0092] The logical correlation between the external disturbance event sequence, the process regulation adjustment event sequence and the abnormal operating condition node is learned and modeled using sequence modeling.

[0093] S3: Hierarchical retrieval and abnormal causal attribution execution.

[0094] Receive query requests for abnormal events, recall relevant historical event sequences based on the hybrid index, and perform differential attribution.

[0095] If the correlation score between abnormal operating conditions and process system adjustments or external interference events exceeds a preset threshold, the root cause will be directly attributed to the corresponding process measures or external interference events.

[0096] If a logical connection with a preceding event cannot be established, the anomaly is attributed to an abnormal change in sensor values, and semantic hints are output to assist further manual analysis.

[0097] like Figure 2 As shown, the standardization of multi-source data from shale gas production scenarios is completed first, providing a high-quality data foundation for subsequent index construction.

[0098] Multimodal data acquisition: Access to three types of core data, covering the entire production process.

[0099] Time-series data: Collects real-time / historical data such as oil pressure and sleeve pressure from the SCADA system.

[0100] Text data: Summarizes unstructured text such as alarm records, maintenance work orders, and operation logs.

[0101] Structured data: Import structured tabular data such as well logs, process parameter tables, and equipment attributes.

[0102] Data representation extraction: Generate feature forms that adapt to different modalities of data.

[0103] Structured data: Extract attribute fields such as well number, process type, and timestamp to form structured attribute features.

[0104] Time series data: Time series feature vectors (Embedding) are generated through feature engineering (such as sliding window and statistical calculation).

[0105] Text data: It is transformed into semantic vectors (Embedding) using NLP models (such as BERT) to retain the core meaning of the text.

[0106] The hybrid index construction method provided in this embodiment of the invention: The core of integrated technology is to build a hybrid index with a triple engine and a unified structure to achieve unified retrieval support for multi-dimensional data.

[0107] Step 1: Build a triple engine.

[0108] Structured index index: The index is built based on the attribute fields of structured data (such as hash number, process type) and supports exact matching queries.

[0109] Vector Semantic Index: Store time-series feature vectors and text semantic vectors in a vector database to build a vector index, supporting similarity retrieval.

[0110] Attribute filtering engine: It sets up a filtering rule base by pre-setting core filtering conditions for shale gas scenarios such as time period, abnormal tags, and similar wells, which can quickly narrow down the search scope.

[0111] Step 2: Design a unified index structure.

[0112] It adopts a two-layer architecture of attribute layer + vector layer. The upper layer connects to the attribute filtering engine and structured index for accurate filtering; the lower layer connects to the vector semantic index to support similarity calculation. The data of the two layers are linked by a unique identifier (such as well ID + timestamp) to ensure cross-layer collaborative access during retrieval.

[0113] The retrieval and anomaly causal tracing execution method provided in this embodiment of the invention meets complex operation and maintenance needs through a three-step process of retrieval-reordering-tracing, while also ensuring the interpretability of abnormal results.

[0114] Step 1: Cross-pattern composite query recall.

[0115] Upon receiving a user query request (such as the liquid accumulation warnings for all wells under a certain skid over the past three days), the system first filters the time period (past three days) + equipment (a certain skid) + anomaly tag (liquid accumulation warning) using the attribute filtering engine. Then, it links the structured index to match the target wells and finally calls the vector semantic index to expand the scope (such as similar warnings for similar wells) to obtain candidate search results.

[0116] Step 2: Two-stage rearrangement optimization.

[0117] Similarity threshold filtering: Filter candidate results and remove irrelevant data based on preset rules (such as vector similarity ≥ 0.8).

[0118] Fusion scoring and reordering: The scoring model integrates indicator data and text data, calculates a comprehensive score by combining dimensions such as well anomaly frequency, work order urgency, and vector similarity, and sorts the results in descending order of score to improve accuracy.

[0119] Step 3: Tracing the cause of the anomaly.

[0120] Anomaly Mapping: Mapping the anomaly nodes (such as pressure anomalies) output by the anomaly detection model to the shale gas ontology knowledge graph.

[0121] Association tracing: Using graph traversal algorithms (such as depth-first search), trace the upstream and downstream related entities of the abnormal node (such as abnormal pressure → shutdown of upstream device A → maintenance work order for device A).

[0122] Explanation generation: Automatically converts the traceability link into natural language explanation (e.g., the pressure abnormality was caused by the shutdown of upstream device A, which has unfinished maintenance work orders), and also supports users to initiate follow-up questions through semantic interfaces (e.g., historical maintenance records of device A shutdown).

[0123] like Figure 3 As shown, the hybrid indexing system integrating anomaly causal tracing and semantic retrieval includes: The multimodal data processing module is used for data acquisition, cleaning, and characterization extraction, and outputs structured attributes and vector features.

[0124] The hybrid index management module is used to build and update structured indicator indexes and vector semantic indexes, and maintain the attribute filtering rule base.

[0125] The retrieval service module is used to process user query requests, execute the recall-reordering process, and return the sorted retrieval results.

[0126] The causal tracing module is used to manage the ontology knowledge graph and perform anomaly mapping, association tracing, and natural language interpretation generation.

[0127] Shale gas ontology knowledge graph: Defines core entities such as wells, skids, processes, sensors, and alarms, as well as their relationships (such as well-skid connections and alarm causal chains), providing a structured knowledge carrier for causal tracing.

[0128] Hybrid index architecture: Through a unified design of attribute layer and vector layer, it achieves synergy between precise matching and similarity retrieval, supporting integrated needs.

[0129] Collaboration with the data layer: Leveraging the ACID properties and data lineage capabilities of the data lake warehouse, it ensures that retrieved data can be traced back to the original SCADA data and model version, guaranteeing the accuracy and auditability of the traceability results.

[0130] Collaborate with the application layer: Provide a retrieval interface for the liquid accumulation early warning app and the anomaly detection app, such as quickly querying wells with anomaly patterns similar to well A and containing 'pump failure' work orders, to assist in operation and maintenance decisions.

[0131] The hybrid indexing system integrating anomaly causal tracing and semantic retrieval described in this invention uses a multimodal data processing module as the data entry point, a hybrid index management module as the core support, and a retrieval service module and a causal tracing module as external service and decision output units during its overall operation, forming a closed-loop processing mechanism covering data acquisition, index construction, retrieval analysis, and anomaly interpretation.

[0132] In the initial stage of system operation, the multimodal data processing module continuously receives multi-source heterogeneous data generated during shale gas production, including time-series signal data from the monitoring system, structured data from the business system, and text data from the operation and maintenance management system. This module performs unified time-base alignment, data cleaning, and normalization on the data from different sources, and generates structured attribute features and vectorized semantic features for different data formats. The structured attributes are used to describe explicit information such as wells, processes, equipment, and time, while the vector features are used to characterize operating condition change patterns and textual semantic connotations.

[0133] Subsequently, the hybrid index management module constructs a unified hybrid index architecture based on the structured attributes and vector features. This architecture maintains configurable filtering rules at the attribute layer to limit the search scope, and maintains a semantic index structure for similarity measurement at the vector layer. During system operation, the hybrid index management module continuously receives new data and performs index updates, while ensuring the correspondence between the index and the underlying data versions, so that any search result can be traced back to its corresponding data source and generation process.

[0134] When a query request is received from an upper-layer application or terminal, the retrieval service module parses the query content, breaking it down into attribute constraints and semantic query conditions, and executes a hierarchical retrieval process within the hybrid index architecture. First, based on the attribute constraints, candidate events or event sequences that meet the conditions are filtered in the attribute layer. Then, in the vector layer, semantic similarity calculation and sorting are performed on the candidate results, outputting the set of historical anomaly patterns or operating condition events most relevant to the query request.

[0135] After the retrieval is completed, the causal tracing module takes over the retrieval results and performs causal analysis on the abnormal events using the shale gas domain ontology knowledge graph. Based on predefined entity relationships and causal chain rules, this module maps abnormal operating conditions to corresponding process, equipment, or external interference entities, and performs step-by-step tracing along the association paths in the knowledge graph to identify potential triggering factors for the anomalies. For anomalies with clear causal relationships, the system generates corresponding causal explanations; for anomalies lacking valid evidence of association, they are marked as sensor anomalies or unknown anomalies, and semantic prompts for manual assessment are output.

[0136] During the overall operation of the system, each module maintains a collaborative relationship with the data layer and application layer to ensure that the retrieval and tracing results are logically verifiable, data-based traceable, and business-interpretable, thereby achieving integrated support capabilities for anomaly causal tracing and semantic retrieval.

[0137] like Figures 4 to 10 As shown, this invention achieves efficient identification, accurate attribution, and continuous optimization control of anomalies in complex production systems by constructing a collaborative working mechanism of "rapid anomaly retrieval—joint analysis of multi-source information—serialized root cause reasoning—dynamic feedback optimization." Its core lies in breaking away from the traditional passive analysis model that relies on manual experience and single rule thresholds, and integrating anomaly retrieval, root cause attribution, and operational decision-making into a unified automatic inference framework, enabling each functional module to form a closed-loop collaborative operation rather than operating in isolation.

[0138] During the anomaly retrieval phase, this invention employs a hybrid indexing and semantic association mechanism to uniformly model real-time monitoring data, historical operational characteristics, and textualized maintenance records, enabling anomaly features to be located and matched within seconds. Figure 4 As shown, the anomaly retrieval efficiency has been reduced from the significant latency of traditional manual methods to near real-time levels, securing a critical response window for subsequent analysis. This rapid retrieval result does not directly output conclusions, but rather serves as a high-confidence candidate input for root cause analysis.

[0139] In the root cause attribution stage, this invention introduces a sequence model based on temporal correlation and causal constraints to jointly reason about the multivariate evolution paths before and after the anomaly. This not only determines "where the anomaly occurred," but also characterizes "why it occurred and how it evolved." For example... Figure 6and Figure 8 As shown, this mechanism can effectively distinguish between various anomalies such as sensor failure, external disturbances, and process adjustments, significantly reducing misjudgment and attribution bias. Its accuracy is significantly better than methods that rely solely on human experience or static rules.

[0140] At the level of operation and decision-making collaboration, this invention directly maps root cause results to actionable operation and maintenance strategies and control recommendations. Through automatic tracing and strategy filtering, it reduces ineffective investigations and repetitive operations, significantly decreasing operation and maintenance hours and costs. Figure 7 As shown. Simultaneously, this result is continuously fed back into the model and index structure to dynamically adjust weights and associations, thereby improving the system's robustness under different query complexities and cross-system scenarios, such as... Figure 9 As shown.

[0141] In terms of overall effectiveness, this collaborative mechanism transformed anomaly handling from "post-event remediation" to "in-process suppression and early intervention," resulting in a continuous downward trend in production losses after system deployment. Figure 10 As shown, this invention demonstrates its stable control capability in the time dimension. In summary, this invention is not simply a superposition of anomaly detection and analysis algorithms, but rather forms an anomaly diagnosis and decision-making system with adaptive and self-evolving capabilities through deep coupling and closed-loop collaboration of multiple mechanisms, demonstrating significant technological advancement and inventiveness.

[0142] Example 1: Rapid source tracing for early warning of fluid accumulation in a single well.

[0143] 1. Scenario background: Well A in a shale gas field experienced three liquid accumulation warning alarms between October 12 and 14, 2024. However, the operation and maintenance personnel were unable to quickly locate the cause of the alarms. The traditional method requires manually reviewing the SCADA data, maintenance work orders and process switching records of the past three days, which takes more than four hours.

[0144] 2. Core user needs: Quickly obtain all data (time series, text, structured) related to the liquid accumulation warning of Well A over the past 3 days; identify the direct cause of this liquid accumulation warning to avoid blind maintenance.

[0145] 3. Technical Implementation Process Step 1: Initiate a cross-pattern compound query.

[0146] Maintenance personnel input a request through a semantic interface: data and reasons related to the fluid accumulation warning for Well A from October 12th to October 14th, 2024. The system automatically breaks down the query conditions: attribute filtering: well number = Well A, time period = October 12th to October 14th, 2024, anomaly tag = fluid accumulation warning; semantic search: related semantic keywords such as fluid accumulation warning and pressure anomaly.

[0147] Step 2: Hybrid index recall and rearrangement.

[0148] The three-engine collaborative retrieval: the attribute filtering engine first filters out the structured data (such as process parameters) of the target time period of well A; the structured index index matches the alarm records corresponding to the liquid accumulation warning label; and the vector semantic index recalls time-series data in similar scenarios (such as the oil pressure change vector during historical liquid accumulation), obtaining a total of 23 candidate results.

[0149] Two-stage reordering: First, filter out 7 low-association data based on similarity threshold (≥0.85), and then rank the downstream separation skid maintenance work order on October 12th based on the record of sudden drop in casing pressure of well A on October 13th using a fusion scoring model (combining alarm frequency + work order urgency + vector similarity).

[0150] Step 3: Tracing the cause of the anomaly.

[0151] The system maps the abnormal node of sudden drop in casing pressure to the shale gas ontology knowledge graph; it traces the correlation through graph traversal algorithm: sudden drop in casing pressure → abnormal pressure transmission of downstream separation skid → separation skid did not return to normal operating condition after maintenance on October 12; it automatically generates natural language explanation: the liquid accumulation warning of well A was caused by the failure to restore the operating condition of the downstream separation skid after maintenance. There is a maintenance work order record for the skid on October 12, and the casing pressure transmission was delayed after maintenance.

[0152] Example 2: Multi-well anomaly correlation analysis.

[0153] 1. Scenario Background: After a shale gas block adopted a new staged fracturing process in October 2024, three wells, B, C, and D, experienced continuous abnormal pressure fluctuations. The operation and maintenance team needs to confirm whether the abnormality is related to the new process and whether there are common causes across the wells.

[0154] 2. Core user requirements: Retrieve relevant data for all wells that experienced pressure fluctuations after adopting the new staged fracturing technology; analyze whether the anomalies are directly related to the new technology and whether there are common causes across wells.

[0155] 3. Technical Implementation Process Step 1: Initiate a cross-well composite query.

[0156] User input semantic request: All wells that experienced abnormal pressure fluctuations after the adoption of the new staged fracturing technology in October 2024, and the associated causes. System disassembly conditions.

[0157] Attribute filtering: Process type = staged fracturing, Time period = 2024-10, Anomaly label = pressure fluctuation.

[0158] Cross-well search: Covers all wells within the block and supports parallel querying of data from multiple wells.

[0159] Step 2: Hybrid index multimodal data fusion retrieval.

[0160] Multimodal data collaborative recall: The structured data layer screened out 12 wells using the new technology; the time series data layer recalled the pressure fluctuation time series vectors of these 12 wells through vector indexing; the text data layer matched maintenance work orders related to segmented fracturing pressure fluctuations through semantic indexing, and finally locked 45 candidate data for three wells B, C and D.

[0161] Fusion Rearrangement: The fusion scoring model places the fracturing fluid supplier work orders used by the three wells in the core position, based on the new process parameter settings, and excludes 18 independent interference data from individual wells (such as the sensor temporary fault record of well B).

[0162] Step 3: Tracing the cause of cross-well anomalies.

[0163] The pressure fluctuation anomalies of the three wells were uniformly mapped onto a knowledge graph, focusing on the correlation link between new technology, fracturing fluid, and pressure.

[0164] Commonalities were identified: all three wells used fracturing fluid from supplier X → the fracturing fluid viscosity did not match the new process parameters → uneven pressure transmission caused fluctuations.

[0165] Conclusion: The abnormal pressure fluctuations in wells B, C, and D are directly related to the new staged fracturing process. The common cause is that the viscosity of the fracturing fluid used does not match the process parameters. This fracturing fluid was provided by supplier X.

[0166] This invention addresses the need for retrieval and attribution of abnormal events during shale gas production, constructing a multi-dimensional fusion architecture combining attribute filtering, semantic vectors, and sequence modeling. For structured attributes in the event sequence, fields such as well number, stratigraphic number, fracturing section number, construction team number, production system type, equipment model, timestamp, pressure range, and gas production range are preferred. The attribute filtering index structure is preferably constructed using inverted indexes, B+ tree indexes, or hash indexes. In an exemplary embodiment, the index fields can cover string, integer, and range data, and hierarchical time indexes with daily, hourly, or minute granularity can be constructed for the time field. For semantic feature vectors, pre-trained language models (such as encoding models based on the Transformer architecture) can map event text descriptions, work records, or alarm logs into 128-dimensional, 256-dimensional, or 768-dimensional semantic vectors, and a vector semantic index based on approximate nearest neighbor algorithms (such as HNSW, IVF-PQ, or LSH) can be constructed. Hybrid index architecture preferably adopts a two-stage strategy of "filtering first and recalling later". That is, first use the attribute filtering index to limit the well area, time window or system type, and then use the vector index to recall similar event sequences to improve retrieval efficiency and accuracy.

[0167] For sequence correlation modeling, time-encoded sequence models, such as LSTM, Transformer, or attention models with time decay factors, are preferred for joint modeling of external disturbance events, process regulation adjustment events, and abnormal operating condition events. For example, the event occurrence time can be location-encoded, and a weighted correlation score combining semantic distance and time interval can be calculated. The time proximity score can be implemented using an exponential decay function, a linear decay function, or a piecewise function, with decay parameters selected at different levels such as 0.01, 0.05, or 0.1. When the correlation score exceeds a preset threshold (e.g., 0.6, 0.7, or 0.8), the anomaly is attributed to external disturbance or process regulation adjustment; otherwise, it is preferably attributed to sensor anomaly, and corresponding semantic prompts are output, such as "potential fracturing backflow disturbance" or "suspected sensor drift."

[0168] It should be noted that embodiments of the present invention can be implemented in hardware, software, or a combination of both. The hardware portion can be implemented using dedicated logic; the software portion can be stored in memory and executed by a suitable instruction execution system, such as a microprocessor or dedicated-design hardware. Those skilled in the art will understand that the above-described devices and methods can be implemented using computer-executable instructions and / or included in processor control code, for example, such code provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuitry such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field-programmable gate arrays, programmable logic devices, etc., or by software executed by various types of processors, or by a combination of the above-described hardware circuitry and software, such as firmware.

[0169] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention, and within the spirit and principles of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for constructing and attributing multi-source event sequences for anomaly analysis in shale gas production, characterized in that, Includes the following steps: Collect time-series data, text data, and structured data generated during shale gas production; The time-series data, text data, and structured data are aligned to a unified time reference. The aligned multi-source data is transformed into a sequence of events with timestamps; The event sequence is divided into sensor operating condition events, external interference events, and process regime adjustment events according to the data source. Based on the temporal relationship and semantic correlation between abnormal operating condition events and preceding event sequences, root cause attribution is performed on abnormal operating condition events.

2. The method according to claim 1, characterized in that, The time-series data includes oil pressure data and casing pressure data collected by the monitoring system. The time-series data undergoes time resampling and missing value repair processing before being converted into an event sequence.

3. The method according to claim 1, characterized in that, The text data includes alarm records, maintenance work orders, and operation logs. The text data is transformed into a vector form to represent semantic information through a natural language processing model.

4. The method according to claim 1, characterized in that, The structured data includes well log information, process parameter tables, and equipment attribute information. The structured data is constructed by extracting well number, process type, and time fields to form structured attribute features.

5. A hybrid index construction and sequence correlation modeling method for shale gas production anomaly analysis, characterized in that, Includes the following steps: Build an attribute filtering index for structured attributes in the event sequence; Construct a vector semantic index based on the semantic feature vectors in the event sequence; A hybrid index architecture is constructed based on the attribute filtering index and the vector semantic index; Sequence models are used to learn the correlation between external disturbance event sequences, process regulation adjustment event sequences, and abnormal operating condition events.

6. The method according to claim 5, characterized in that, The sequence model encodes the temporal order of events and calculates a relevance score by combining the semantic distance between events and the size of the time interval.

7. The method according to claim 5, characterized in that, When performing a retrieval, the hybrid index first limits the retrieval scope based on the attribute filtering index, and then completes the recall of similar event sequences based on the vector semantic index.

8. A hierarchical retrieval and causal attribution method for anomaly analysis in shale gas production, characterized in that, Includes the following steps: Receive query requests for abnormal operating conditions; Historical event sequences related to the abnormal operating condition events are retrieved using a hybrid index; Calculate the correlation score between abnormal operating condition events and external disturbance events or process regime adjustment events; When the correlation score exceeds the preset threshold, the abnormal operating condition event will be attributed to the corresponding external interference event or process system adjustment event. When the preset threshold is not exceeded, the abnormal operating condition event is attributed to the abnormal sensor operating condition.

9. The method according to claim 8, characterized in that, The relevance score is calculated by weighting the event semantic similarity score and the time proximity score, and the time proximity score decreases as the event time interval increases.

10. The method according to claim 8, characterized in that, When an abnormal operating condition event is attributed to an abnormal sensor condition, semantic prompts related to the abnormality are output to assist manual further analysis.

Images