LLM-based air compression station anomaly detection method and system

Abstract

The invention discloses an LLM-based air compression station anomaly detection method and system, and relates to the technical field of air compression station anomaly management. According to the air compression station anomaly detection method based on the LLM, online operation data and offline robot inspection data of an air compression station are obtained, automatic inspection and anomaly recognition are achieved through the LLM by combining data verification with an operation and maintenance knowledge base, then anomaly types are judged based on recognition results, graded feedback is conducted, and the anomaly types are pushed to an automatic control module. And meanwhile, optimizing a robot vision algorithm by using real-time abnormal data, and finally, dividing routing inspection priorities according to offline routing inspection data of an adaptive lattice, setting different routing inspection intervals and generating a report containing a priority thermodynamic diagram. According to the process, multi-source data are fused, abnormal intelligent judgment and dynamic inspection are achieved, the intelligent and cooperative efficiency of operation and maintenance management of the unattended air compression station is improved, and then the problem that the intelligent degree of the air compression station is low from single equipment monitoring to regional cooperative management is effectively solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of air compression station anomaly management, and particularly relates to an air compression station anomaly detection method and system based on an LLM. BACKGROUND

[0002] As a cutting-edge technology in the field of artificial intelligence, a large language model (LLM) has strong language understanding, generation and reasoning capabilities, can process massive amounts of data and mine potential information, and has shown great application potential in many industries. An air compression station, as a key place for providing compressed air, has closely coordinated structural components that are connected in order through pipes to jointly ensure the stable supply of compressed air. Among them, the air compressor is the core power source of the entire system, which inhales external air and compresses it to produce high-temperature and high-pressure compressed air. The compressed air discharged from the air compressor is directly delivered to the air tank through the pipe, which can store a certain amount of compressed air and stabilize the system pressure, reducing the frequent start and stop of the compressor and prolonging the service life of the equipment.

[0003] In industrial production, the air compression station provides power support for numerous devices, and its stable operation is crucial. However, current air compression station operation and maintenance management faces many challenges. In the traditional mode, air compression station operation and maintenance management relies on long-term manual guarding. As the number of managed stations increases, the operation and maintenance cost rises significantly. At the same time, manual inspection is inefficient and difficult to monitor the equipment status in real time, leading to problems being discovered and solved in a timely manner, affecting production progress and quality.

[0004] To improve this situation, existing technologies use online systems to collect air compression station data to promote intelligent management. The online system is arranged with numerous sensors at each key component of the air compression station, which can collect data such as air compressor operating parameters, air tank pressure and temperature, dryer water removal effect, and filter pressure difference before and after. The collected data is transmitted to the monitoring center server through wired or wireless methods; the server uses a pre-set algorithm and model to preliminarily analyze and process the data, and sends out an alarm as soon as data anomalies are found; operation and maintenance personnel can quickly locate the problem components based on the alarm information, prepare maintenance tools and spare parts in advance to handle them, which to some extent reduces the workload of manual inspection and improves the operation and maintenance response speed.

[0005] For example, Chinese invention patent CN116720983A discloses a method and system for detecting power supply equipment anomalies based on big data analysis. The method includes: collecting relevant data from various power supply devices and preprocessing the collected data; building an anomaly detection model; obtaining the status assessment value of the corresponding power supply device based on the collected data; setting a first threshold, a second threshold, and a third threshold, where the first threshold < the second threshold < the third threshold; comparing each threshold with the status assessment value of the corresponding power supply device to obtain the comparison result; building a prediction model using big data analysis technology to obtain a prediction function curve; extracting the result at the time node where the predicted status assessment value > the first threshold; and selecting the corresponding execution strategy based on the results of the threshold unit and the prediction unit.

[0006] While existing online systems can achieve remote monitoring by collecting data through sensors and reduce some basic manual inspection work, their limited level of intelligence makes them unable to cope with the management challenges brought about by the increase in the number of stations. Specifically, this is reflected in the following aspects: The online system relies solely on sensor data and lacks a cross-validation mechanism with offline equipment status (such as instrument readings and equipment appearance). This makes it prone to misjudgments in the event of sensor failure or data drift, further hindering the online system's ability to distinguish between equipment malfunctions and sensor layer failures, resulting in low alarm accuracy. As the number of management stations increases, the online system cannot integrate data from multiple stations for global optimization (such as regional load scheduling and pressure balancing), making it difficult to adapt to large-scale management needs. This leads to a low level of intelligence in the transition from single-equipment monitoring to regional collaborative management of air compressor stations. Summary of the Invention

[0007] To address the technical problems in the prior art, this invention provides an LLM-based method and system for detecting anomalies in air compressor stations. The technical solution is as follows: On the one hand, an LLM-based anomaly detection method for air compressor stations is provided. This method includes: Step 1, acquiring online and offline data during the operation of the air compressor station and performing cross-validation of the data. Simultaneously, an operation and maintenance knowledge base is used to organize the daily operation and maintenance work of the air compressor station and the equipment technical manual. Combined with LLM, automatic inspection of the air compressor station is achieved to automatically identify anomalies. The online and offline data include online operational data reflecting the real-time operating status of the air compressor station and offline robot inspection data. The offline robot is used to acquire water meter data not covered by the online system and equipment appearance information. Step 2, based on the results of automatic anomaly identification using LLM, anomaly problem judgment and hierarchical feedback are performed. Simultaneously, data is acquired to drive the offline robot's vision. The algorithm dynamically optimizes real-time anomaly data. Anomaly identification is used to distinguish the types of anomalies in the operation of the air compressor station. Anomaly types include air compressor station failures and perception layer failures of online and offline robots. Hierarchical feedback is used to classify the degree of anomaly according to the anomaly type and push the result of the degree of anomaly classification to the corresponding automatic control module of the air compressor station. Step 3: Obtain the offline inspection data of the corresponding offline robot after the adaptability is qualified, divide the inspection priority to dynamically set the inspection time interval of different inspection areas, and generate an inspection report based on the inspection priority division result to visualize the distribution of each priority area in each inspection area of ​​the air compressor station. The offline inspection data is used to reflect the identification status of pressure instruments in each inspection area of ​​the air compressor station.

[0008] On the other hand, an LLM-based air compressor station anomaly detection system is provided. This system applies an LLM-based air compressor station anomaly detection method and includes: an air compressor station anomaly identification module, an air compressor station anomaly handling module, and an air compressor station inspection priority division module. The air compressor station anomaly identification module acquires online and offline data during the air compressor station's operation and performs data cross-verification. It also utilizes an operation and maintenance knowledge base (combined with LLM) to automatically identify anomalies by analyzing daily maintenance data and equipment technical manuals. The air compressor station anomaly handling module determines and grades anomalies based on the LLM's automatic anomaly identification results and acquires real-time anomaly data to drive dynamic optimization of the offline robot's visual recognition algorithm. The air compressor station inspection priority division module acquires offline inspection data from the corresponding offline robot after successful adaptation, divides inspection priorities to dynamically set inspection time intervals for different inspection areas, and generates inspection reports based on the priority division results to visualize the distribution of each priority area across the air compressor station's inspection areas.

[0009] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: 1. This invention collects online operational data and offline robot inspection data from air compressor stations, verifies them, and integrates them with an operations and maintenance knowledge base and LLM (Local Management Model) to achieve automatic inspection and anomaly identification. This compensates for the incompleteness of online data and allows for remote on-site inspection via robots during troubleshooting. Based on the LLM's identification results, it determines the type of anomaly, such as faults in the air compressor station itself or the sensing layer, categorizes them, and sends them to the automatic control module. Simultaneously, real-time anomaly data is used to optimize the robot's vision algorithm, achieving AI-driven automatic inspection without the need for manual on-site inspection and troubleshooting. Finally, based on qualified offline data, inspection priorities are assigned, inspection intervals are set, and reports with heatmaps are generated, promptly dispatching problems and solutions to maintenance personnel for rapid resolution. This method integrates online and offline data, automatically handling inspections and anomalies through LLM, eliminating the need for manual operation, making operations and maintenance faster, achieving unmanned station operation, reducing costs, and improving efficiency.

[0010] 2. By verifying the time difference between online pressure data and offline robot-captured images, and using the open interval of historical synchronization deviation as the allowable range, the system ensures that the data are in the same time dimension, avoiding verification errors caused by time asynchrony. This provides a precise data foundation for single-station equipment monitoring and solves the problem of inconsistent data benchmarks in single-equipment monitoring. By comparing the online and offline pressure differences, anomalies are categorized and marked: no anomalies ensure single-station stability; first-level anomalies trigger robot re-inspection and LLM verification to reduce single-station misjudgments; second-level anomalies trigger manual review and emergency response plans, achieving a seamless transition from single-station anomalies to regional collaborative handling. This mechanism breaks the limitations of single-equipment monitoring, reducing manual intervention through data consistency assurance and tiered collaborative handling. It upgrades air compressor stations from decentralized equipment monitoring to intelligent regional collaborative management, effectively solving the problem of low intelligence caused by inaccurate data and insufficient collaboration in traditional models.

[0011] 3. By comparing the average pressure difference between adjacent stations with historical data, regional load scheduling is automatically triggered to balance pressure in case of anomalies, and the baseline data of the knowledge base is updated when there are no anomalies. This mechanism breaks the limitations of independent monitoring of a single station and avoids the inefficiency of relying on manual scheduling for regional load imbalance in the traditional model. Through secondary identification by the robot and judgment based on the difference reduction index, the knowledge base is updated if it is qualified and an early warning is issued if it is not qualified, which reduces the misjudgment of pressure data of a single station, provides reliable basic data support for regional collaboration, avoids regional scheduling errors caused by data deviation of a single station, and forms a data closed loop between single station monitoring and regional collaboration, promoting the upgrade of air compressor station management from decentralized equipment monitoring to intelligent full-domain collaboration.

[0012] 4. Through fault-data feature mapping between LLM and the operation and maintenance knowledge base, three types of anomalies are accurately identified: On-device anomalies are determined by combining online temperature, flow rate, and offline fan status from multiple dimensions; online sensing layer anomalies are verified through pressure fluctuation differences and pipe appearance; and offline robot sensing layer anomalies rely on temperature differences before and after angle adjustments. This multi-source data linkage identification logic avoids misjudgments caused by incomplete data in single-device monitoring, providing a precise foundation for anomaly localization in regional collaboration. Anomaly severity-based feedback enables collaborative anomaly handling: On-device anomalies optimize robot parameters to improve single-station identification accuracy; online sensing layer anomalies generate work orders through regional operation and maintenance units, promoting cross-regional sensor maintenance; and offline sensing layer anomalies activate emergency plans and trigger alarms, ensuring stable regional load. This closed-loop mechanism breaks the isolation of single-station monitoring, upgrading anomaly handling from decentralized device-level responses to intelligent regional collaborative management, significantly improving the intelligence level of overall operation and maintenance.

[0013] 5. Adjustment based on real-time abnormal data: If the flow rate deviation exceeds the limit, match the supplementary lighting parameters to reduce interference; if the load rate deviation exceeds the limit, verify the physical status of the indicator. After optimization, verify the adaptability. If it is effective, complete the optimization; otherwise, issue an early warning and intervene to continuously improve the robot's recognition accuracy and provide reliable support for data verification.

[0014] Subsequently, priorities are set from one to three levels and no priority based on the average deviation value from largest to smallest, with corresponding inspection intervals ranging from 2 hours to 24 hours. After the classification, supplementary inspections are triggered to correct misjudgments, and data is synchronized to the knowledge base to ensure that the priority classification matches the dynamic on-site situation, thereby achieving accurate allocation of inspection resources and improving the efficiency of operation and maintenance response. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 A flowchart of an LLM-based anomaly detection method for air compressor stations provided in an embodiment of the present invention; Figure 2 A flowchart for data mutual verification provided in embodiments of the present invention; Figure 3 A flowchart for data cross-validation and anomaly classification provided in embodiments of the present invention; Figure 4 A schematic diagram of an LLM-based anomaly detection system for an air compressor station is provided in an embodiment of the present invention. Figure 5The diagram shows the architecture interface of the LLM-based air compressor station anomaly detection system provided in this embodiment of the invention. Detailed Implementation

[0017] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0018] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0019] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0020] This invention provides an LLM-based method for detecting anomalies in air compressor stations, such as... Figure 1 The flowchart shown is a method for detecting anomalies in an air compressor station based on LLM. The processing flow of this method may include the following steps: Step 1: Acquire online and offline data during the operation of the air compressor station and perform cross-validation. Simultaneously, utilize an operation and maintenance knowledge base (based on equipment technical manuals and daily maintenance procedures) to analyze the station's operational status. Combined with LLM (Local Management Model), implement automated inspections to automatically identify anomalies. Online and offline data include real-time operational data collected online and inspection data from offline robots. Offline robots acquire data from water meters not covered by the online system and information about equipment appearance. Cross-validation includes time synchronization and anomaly detection within a single air compressor station, as well as multi-station collaborative management (when the number of air compressor stations increases). Time synchronization is used to verify online pressure data. Based on the consistency of the time of pressure gauge data collection and offline robot photography, the air pressure anomaly determination means determining the type of air pressure problem at the air compressor station by using pressure data collected online and offline. Multi-station collaborative management is used to determine the cross-station collaborative status by comparing the average pressure difference between adjacent air compressor stations with the historical average pressure difference. The online collected operating data includes data from equipment such as air compressors, refrigerated dryers, desiccant dryers, and cooling towers during operation, as well as flow data, pressure data, and environmental temperature and humidity data on pipelines. The offline robot inspection data includes data from various instruments along the designated route and environmental data of the air compressor station during the inspection process (videos of the inspection process taken by the camera). The data from various instruments includes readings from the air compressor panel, pressure gauges, thermometers, water meters, etc.

[0021] Step two involves determining and classifying anomalies based on the results of LLM's automatic anomaly identification. Simultaneously, real-time anomaly data is acquired to drive dynamic optimization of the offline robot's visual recognition algorithm. Anomaly determination and classification feedback include data cross-validation and anomaly severity classification, as well as anomaly severity classification feedback. Anomaly determination is used to distinguish the types of anomalies occurring during the operation of the air compressor station. Anomaly types include air compressor station malfunctions and perception layer malfunctions of online and offline robots. Anomaly severity classification feedback is used to classify the severity of anomalies based on their types and pushes the results of the anomaly severity classification to the corresponding air compressor station's automatic control module. The automatic control module includes, but is not limited to, an automatic calibration unit, a regional operation and maintenance unit, and an emergency control unit.

[0022] Step 3: Obtain the offline inspection data of the corresponding offline robot after the adaptability is qualified, divide the inspection priority to dynamically set the inspection time interval of different inspection areas, and generate an inspection report based on the inspection priority division results to visualize the distribution of each priority area in each inspection area of ​​the air compressor station. The offline inspection data is used to reflect the identification status of pressure instruments in each inspection area of ​​the air compressor station, and the inspection report shows the distribution of each priority area in each inspection area of ​​the air compressor station through a priority heat map.

[0023] The inspection report is visualized using a priority heatmap as follows: The air compressor station is divided into multiple inspection areas. Based on the instrument identification status data (such as identification success rate and error rate) collected by the offline robot in each area, the necessity of inspection for each area is quantitatively assessed and priorities are assigned (first priority, second priority, and third priority). The priority heatmap uses different color gradients to intuitively present the priority distribution: the first priority area is marked in red, representing areas with frequent instrument identification anomalies that require special attention; the second priority area is marked in yellow, corresponding to areas with occasional anomalies; and the third priority area is displayed in green, representing stable areas. The specific location of each area and key instruments are also marked to facilitate maintenance personnel in quickly locating key inspection targets.

[0024] It's crucial to understand that cross-validating online and offline data is primarily about overcoming the limitations of single data dimensions and ensuring the accuracy of anomaly detection. On one hand, offline robots conduct inspections every 4-8 hours and only take a single photo, lacking data fluctuations and unable to capture dynamic anomalies in real time. While online data is real-time, it may be distorted due to sensor malfunctions; for example, a stable air compressor power output with large fluctuations in air production can detect flow meter anomalies. On the other hand, when LLM detects a rise in unit consumption from 0.186 kWh / m³ to 0.2 kWh / m³, it needs to be verified in conjunction with online air production, power, and offline equipment appearance data to rule out sensor malfunctions or misjudgments and confirm whether it's an operational anomaly. Furthermore, after offline robots discover external issues such as oil leaks, it's necessary to link this to online data such as increased oil temperature and abnormal exhaust pressure to rule out misjudgments, confirm the actual air compressor malfunction, and avoid false alarms or missed detections based on single data points, providing a reliable data foundation for accurate anomaly identification by LLM.

[0025] like Figure 2 The flowchart shown illustrates the data cross-verification process. Its specific design logic is as follows: This design revolves around air compressor station management, using data interaction verification as a foundation for multi-dimensional collaborative control. First, through collaborative management of each air compressor station and time synchronization determination within the station, the average pressure difference and synchronization time deviation are obtained. Based on a comparison between the maximum average pressure difference and the historical average, it determines whether cross-station collaboration is abnormal; by comparing the synchronization time deviation with the allowable range, it decides whether to calibrate the time. After a synchronization anomaly is detected, air compressor anomaly determination is performed, combining exhaust pressure difference and pipeline pressure difference to identify first- or second-level air compressor anomalies or the absence of anomalies. Finally, LLM (Limited Linear Modulation) is used to achieve automatic inspection and anomaly identification of air compressor stations, ensuring stable system operation.

[0026] Further understanding is needed regarding the time synchronization determination. The specific steps are as follows: First, obtain the synchronization time deviation between online and offline data collection. This deviation represents the time difference between the online pressure data collection time and the offline robot's pressure gauge image capture time. This is used to determine whether the two types of data are in the same monitoring dimension to avoid verification errors caused by time asynchrony. If the obtained synchronization time deviation exceeds the set allowable range, a preset personnel will be prompted to perform time synchronization calibration to ensure that subsequent air pressure anomaly determinations are based on pressure data in the same time dimension. Otherwise, an air pressure anomaly determination is performed. Pressure data includes exhaust pressure and pipeline pressure. Exhaust pressure is the pressure value at the corresponding pressure monitoring sensor installation node after the air compressor discharges compressed gas. Pipeline pressure is the pressure value at the corresponding pressure monitoring sensor installation node in the delivery pipeline. The set allowable range for synchronization time deviation is represented by the open interval corresponding to the maximum and minimum values ​​of historical synchronization time deviations during the mutual verification process of historical data in the operation and maintenance knowledge base. That is, it does not include cases equal to the maximum and minimum values ​​of historical synchronization time deviations.

[0027] The specific steps for determining air pressure anomalies are as follows: The online-collected air compressor discharge pressure and pipeline pressure are compared with the pressure gauge readings on the air compressor panel captured by the corresponding offline robot to obtain the discharge pressure difference and pipeline pressure difference. Simultaneously, the corresponding set values ​​are acquired. These set values ​​include the set discharge pressure difference (usually set to 0.5 bar) and the set pipeline pressure difference (usually set to 0.05 MPa). Both the discharge pressure difference and pipeline pressure difference reflect the degree of difference between the pressure data collected online and the pressure gauge readings collected by the offline robot in the pressure monitoring results. The pressure gauge readings include the air compressor discharge pressure reading and the pipeline pressure reading. The discharge pressure difference represents the difference between the air compressor discharge pressure collected by the pressure sensor and the air compressor discharge pressure reading collected by the offline robot. The absolute value of the pipeline pressure difference represents the absolute value of the difference between the pipeline pressure collected by the pressure sensor and the pipeline pressure reading collected by the offline robot. If neither the obtained exhaust pressure difference nor the pipeline pressure difference is greater than the corresponding set value, it is marked as no air pressure anomaly. If either the obtained exhaust pressure difference or the pipeline pressure difference is greater than the corresponding set value, it is marked as a Level 1 air pressure anomaly, and the offline robot is re-checked and LLM is initially verified. If both the obtained exhaust pressure difference and the pipeline pressure difference are greater than the corresponding set value, it is marked as a Level 2 air pressure anomaly, and manual review is triggered to formulate a corresponding emergency handling plan, such as closing the valve of the abnormal pipeline section to prevent a sudden drop in pressure, or temporarily sharing the load with surrounding stations. Level 1 and Level 2 air pressure anomalies reflect the increasing degree of abnormality of air pressure problems in the air compressor station, respectively.

[0028] In this embodiment, both the set exhaust pressure difference and the set pipeline pressure difference are based on the equipment safety operation manual, historical fault data statistics, and industry pressure deviation standards. In practical applications, these can be fine-tuned according to the air compressor station load fluctuation range, equipment aging level, regional air pressure requirements, etc., to adapt to the precise anomaly detection needs in different scenarios.

[0029] In this example, time synchronization and air compressor anomaly detection provide core support for unmanned air compressor stations, significantly improving operational efficiency. Specifically, time synchronization automatically verifies the time difference between online and offline data, using the open interval of historical deviations in the operational knowledge base as a standard. When the deviation exceeds the range, automatic calibration is prompted, ensuring data consistency without manual intervention. This adapts to the autonomous data verification needs of unmanned scenarios, avoiding the lag and errors of traditional manual synchronization. Air compressor anomaly detection achieves tiered processing through difference comparison: no anomalies are automatically confirmed; first-level anomalies are re-examined by the robot and verified by LLM; and second-level anomalies directly trigger manual review and cross-station scheduling.

[0030] Compared to existing technologies that rely on manual inspection and anomaly detection, and manual recording for time synchronization, this method achieves accurate data alignment under unattended operation through automated time verification and hierarchical anomaly handling; it reduces the false judgment rate of first-level anomalies by leveraging robot re-inspection and LLM verification; and it enables cross-station scheduling in conjunction with second-level anomalies, significantly surpassing the operation and maintenance efficiency of traditional manual methods, and providing an intelligent solution for large-scale air compressor station cluster management.

[0031] Furthermore, the offline robot re-inspection and LLM preliminary verification process is as follows: The air compressor station area corresponding to the first-level air compressor anomaly is obtained, and the offline robot's visual recognition algorithm is triggered to re-obtain pressure gauge readings. Simultaneously, a preliminary verification report is generated. This report visualizes the reduction rate of the difference between the re-obtained exhaust pressure difference and the pipeline pressure difference. This reduction rate represents the result of harmonic averaging of the reduction rates of the re-obtained pipeline pressure difference and the exhaust pressure difference. The reduction rate includes both the exhaust pressure difference reduction rate and the pipeline pressure difference reduction rate. The exhaust pressure difference reduction rate represents the difference between the re-obtained exhaust pressure difference and the exhaust pressure difference obtained during the air compressor anomaly determination process. The pipeline pressure difference reduction rate represents the difference between the re-obtained pipeline pressure difference and the pipeline pressure difference obtained during the air compressor anomaly determination process.

[0032] If the obtained difference reduction index is greater than the set difference reduction index (usually set to 10%, which can be fine-tuned by the preset personnel according to the equipment accuracy and environmental interference), it is determined that the offline robot re-inspection and LLM preliminary verification are qualified, and the results of the offline robot re-inspection and LLM preliminary verification are input into the operation and maintenance knowledge base to update the benchmark data of the corresponding air compressor station pressure verification; if the obtained difference reduction index is not greater than the set difference reduction index, it is determined that the offline robot re-inspection and LLM preliminary verification are unqualified, and an air compressor problem warning is issued, and the warning information is pushed to the operation and maintenance terminal.

[0033] In this embodiment, the verification effect is quantified by the difference reduction index through automatic robot re-inspection and LLM analysis. The knowledge base is updated based on the qualified results to form a data closed loop, continuously optimizing the judgment benchmark. Timely warnings are issued when non-qualified results occur to ensure that no anomalies are missed. The operation and maintenance terminal is usually deployed on the mobile terminal (such as mobile APP or tablet terminal) of the corresponding operation and maintenance personnel in the air compressor station regional management center. It supports remote reception of warning information and viewing of preliminary verification reports. The details of the anomalies can be grasped without going to the site, which greatly improves the efficiency of operation and maintenance response and adapts to the remote management needs of unattended mode.

[0034] Furthermore, the multi-station collaborative management process is as follows: When comparing various air compressor stations, if the maximum value of the average pressure difference between adjacent air compressor stations during the data mutual verification period is greater than the historical average pressure difference, it is marked as an anomaly in cross-station collaboration and triggers the regional load rescheduling mechanism to balance the operating pressure between the various air compressor stations. The average pressure difference represents the average value corresponding to the exhaust pressure difference and pipeline pressure difference between adjacent stations. If the maximum value of the average pressure difference between adjacent air compressor stations during the data mutual verification period is not greater than the historical average pressure difference, it is marked as no anomaly in cross-station collaboration, and the result of multi-station collaborative management is input into the operation and maintenance knowledge base to update the benchmark data for collaborative operation between air compressor stations, providing a more accurate reference for subsequent cross-station pressure balance scheduling. The historical average pressure difference is represented by the summation and averaging of the historical average pressure differences between adjacent air compressor stations during the historical station collaborative management process in the updated operation and maintenance knowledge base.

[0035] In this embodiment, by comparing the average pressure difference between adjacent stations with historical data, the cross-station collaborative status is accurately determined. In case of anomalies, regional load rescheduling is triggered to quickly balance the pressure between stations, avoiding equipment overload or insufficient power supply caused by pressure imbalance. In case of normal operation, the knowledge base baseline data is updated to continuously optimize collaborative operation parameters. This dynamic control and data closed-loop mechanism breaks the limitations of independent management of a single station, improves the overall stability and energy efficiency of multi-station operation, and provides reliable support for the intelligent collaborative management of large-scale air compressor station clusters.

[0036] like Figure 3 The flowchart shown illustrates the data cross-validation and anomaly severity grading process. Its specific design logic is as follows: Based on an operations and maintenance knowledge base, the process first updates the knowledge base and inputs fault-data feature mapping relationships into a Large Language Model (LLM). Using the LLM, anomalies are identified and categorized into anomalies in the air compressor station itself, the online sensing layer, and the offline robot sensing layer. Different anomalies correspond to different processing paths: anomalies in the air compressor station itself are pushed to the automatic calibration unit to optimize robot vision parameters; online anomalies are pushed to the regional operations and maintenance unit, prompting for sensor verification or replacement; offline anomalies are pushed to the emergency control unit, sending an emergency alarm to the operations and maintenance manager to achieve precise and efficient operations and maintenance response.

[0037] Further understanding is needed regarding data cross-validation and anomaly severity grading. The specific process is as follows: The fault-data feature mapping relationship in the updated maintenance knowledge base is input into the LLM for anomaly identification, outputting the anomaly identification results. If the anomaly identification results show an anomaly in the air compressor station itself, specifically manifested as: the average value of the air compressor exhaust temperature collected online during the anomaly identification period is greater than the corresponding set value (usually set to 80℃), and simultaneously, the cooling water pump flow rate is detected to drop below the rated value (usually set to 30m³ / h), and the number of times the cooling fan blades stop as captured by the offline robot is greater than the corresponding set value (usually set to 3 times); if the anomaly identification results show an anomaly in the online sensing layer, specifically... The specific manifestations are as follows: the fluctuation range of the pipeline pressure collected online during the abnormal problem identification period is greater than the corresponding set value (usually set to 20%), and the fluctuation range of the corresponding pressure gauge reading captured by the offline robot is not greater than the corresponding set value (usually set to 5%), while there are no traces of leakage in the pipeline appearance video; if the abnormal problem identification result shows that the offline robot perception layer is abnormal, the specific manifestations are as follows: the average difference between the temperature gauge reading captured by the offline robot and the corresponding temperature data collected online during the abnormal problem identification period is greater than the corresponding set value (usually set to 3℃), and the difference between the temperature gauge reading reread after adjusting the shooting angle of the offline robot and the online data is not greater than the corresponding set value (usually set to 1℃).

[0038] Specifically: Excessive online exhaust temperature and insufficient cooling water pump flow, combined with frequent cooling fan stoppages captured by the offline robot, form a "data loop." All three point to a cooling system failure, indicating a fault in the equipment itself, not a monitoring issue, thus classifying it as an anomaly in the air compressor station itself. Online pipeline pressure fluctuations far exceed thresholds, but the pressure gauge fluctuations captured by the offline robot are small, and there are no pipeline leaks, indicating distorted online data. The problem lies with online sensing devices such as pipeline pressure sensors (e.g., sensor malfunction, transmission interference), thus classifying it as an anomaly in the online sensing layer. The initial temperature readings captured by the robot deviate significantly from the online readings, but the data matches after adjusting the angle, indicating a problem with the robot's shooting angle, focus, or other sensing aspects, rather than the temperature itself or the online system itself, thus classifying it as an anomaly in the offline robot sensing layer.

[0039] In this embodiment, the settings mentioned are all based on the safety thresholds in the equipment technical manual and industry operation and maintenance standards. In practical applications, they can be fine-tuned according to the load fluctuation range of the air compressor station, the degree of equipment aging, and changes in ambient temperature and humidity, so as to achieve accurate adaptation for anomaly detection.

[0040] This example utilizes LLM (Limited Linear Modeling) combined with fault-data feature mapping to cross-validate anomaly types from multiple dimensions, avoiding misjudgments based on single data points. For anomalies on the device itself, the dynamic equipment images provided by the offline robot complement the online parameters. For anomalies at the online perception layer, the instrument readings and pipe appearance captured by the robot can corroborate whether the sensors have failed. For anomalies at its own perception layer, the robot performs self-verification through angle adjustments. This model of offline robot intervention combined with LLM logical reasoning breaks through the traditional reliance on a single data source, ensuring that anomalies can be quickly located and responded to when unattended, significantly improving the intelligence and automation level of anomaly detection in air compressor stations, and thus effectively enhancing the reliability and stability of autonomous operation and maintenance of air compressor stations.

[0041] It is important to understand that air compressor station malfunctions refer to problems with the station building and equipment such as air compressors and refrigerated dryers, such as cooling system failures (fan jamming causing exhaust temperature to exceed 80°C), refrigerant leaks in refrigerated dryers, and adsorbent failures in desiccant dryers. However, actual detection is easily affected by interference: offline robots perform inspections every 4-8 hours, making it difficult to respond promptly to sudden anomalies such as instantaneous oil leaks. Environmental dust and equipment vibration can also cause sensor data deviations. Furthermore, existing technologies have not solved the problems of sensor response delays to sudden temperature and pressure changes and signal transmission lags, further exacerbating detection delays. Existing technologies mostly rely on single parameter thresholds to determine anomalies, such as triggering an alarm when online monitoring detects unit consumption exceeding the threshold or fluctuations in gas production. This application addresses the aforementioned problem of untimely sensor monitoring by using real-time cross-calibration of online and offline data and LLM dynamic analysis to avoid detection delays. It also focuses on anomaly level classification, graded feedback, and responses to external interference, compensating for the shortcomings of narrow coverage and slow response of simple offline inspections.

[0042] Furthermore, the anomaly severity is graded and feedback is provided, specifically as follows: Anomalies in the displayed air compressor station are pushed to the automatic calibration unit in the automatic control module to trigger dynamic optimization of the offline robot's visual parameters (e.g., adjusting the brightness of the supplementary lighting, correcting the shooting angle, and updating anti-blur algorithm parameters) to reduce similar recognition biases; anomalies in the displayed online perception layer are pushed to the regional maintenance unit in the automatic control module to automatically generate a maintenance work order (including the faulty sensor model, installation location, and historical fault records) and push it to the mobile terminal of the regional maintenance personnel, prompting them to perform sensor verification or replacement; anomalies in the displayed offline robot perception layer are pushed to the emergency control unit in the automatic control module to immediately activate the emergency plan (e.g., automatically switching to the backup air compressor and closing the valves in the abnormal pipeline section), and simultaneously send an emergency alarm to the maintenance manager (including faulty equipment location and temporary handling suggestions from the maintenance knowledge base) to prevent the fault from escalating.

[0043] In this embodiment, when the anomaly severity is graded and feedback is provided, the corresponding unit of the automatic control module is accurately pushed according to different anomaly types. In addition to the automatic calibration unit, regional operation and maintenance unit, and emergency control unit, the automatic control module also includes a strategy optimization unit, which is used to dynamically optimize the anomaly grading standard, inspection priority rules and control strategy parameters (such as pressure threshold, load scheduling coefficient, etc.) based on historical anomaly data and handling results, so that the system decision logic continuously adapts to the changes in the operating status of the air compressor station.

[0044] This hierarchical feedback mechanism enables precise triage and handling of anomalies: the automatic calibration unit optimizes robot recognition parameters to reduce deviations; the regional operation and maintenance unit generates work orders to accelerate sensor repair; the emergency control unit activates contingency plans to curb the escalation of faults; and the strategy optimization unit drives system iteration. This multi-stage collaboration improves recognition accuracy and speeds up fault handling, ensuring the stable operation of the air compressor station from multiple dimensions and significantly enhancing operation and maintenance efficiency as well as the timeliness and effectiveness of anomaly handling.

[0045] Furthermore, the dynamic optimization of the offline robot visual recognition algorithm involves the following process: Real-time anomaly data is obtained based on the feedback results of anomaly severity levels. This real-time anomaly data includes instantaneous airflow deviation and station load rate deviation. The instantaneous airflow deviation reflects the difference between the pipeline flow meter readings recognized by the offline robot visual recognition and the instantaneous airflow data collected online, indicating the accuracy of the robot's recognition of the current air delivery volume. For example, if the instantaneous flow rate collected online is 25 m³ / min, while the robot's recognition value is 22 m³ / min, then the instantaneous airflow deviation is 3 m³ / min. Excessive deviation can lead to misjudgment of the real-time gas supply of a single station, affecting the dynamic balance scheduling of regional load. The station load rate deviation reflects the degree of difference between the reading of the air compressor load rate indicator identified by the offline robot and the actual load rate calculated based on online operating data. It indicates the accuracy of the robot's perception of the overall operating load of the station. For example, if the actual load rate is 70% (calculated based on online power and rated power), while the robot's identification value is 55%, then the station load rate deviation is 15%. This deviation will directly affect the load allocation decision among multiple stations, and may result in high-load stations not receiving timely load support from surrounding stations.

[0046] If the obtained instantaneous airflow deviation is not greater than the set instantaneous airflow deviation, it indicates that the visual recognition accuracy of the offline robot for the pipeline flow meter meets the requirements. Otherwise, the difference between the obtained instantaneous airflow deviation and the set instantaneous airflow deviation will be matched with the mapping relationship between the degree of flow difference and the supplementary light intensity stored in the updated operation and maintenance knowledge base. Based on the matching result (i.e., the supplementary light intensity adjustment target), the preset personnel will be prompted to adjust the spectral parameters of the current offline robot supplementary light (such as increasing the proportion of 550nm yellow-green light) to reduce the interference of reflection or shadow on the flow meter surface. If the obtained station load rate deviation is greater than the set station load rate deviation, the physical status of the indicator will be verified to prompt the preset personnel to check whether there are physical abnormalities such as dust coverage or scale wear on the surface of the load rate indicator. Otherwise, it indicates that the robot's recognition status of the load rate indicator is stable in the current environment.

[0047] After dynamic optimization of the offline robot visual recognition algorithm, adaptability verification is performed. Specifically, if the reduction in instantaneous airflow deviation and station load rate deviation are both greater than the corresponding set reduction values, it indicates that dynamic optimization is effective in improving the adaptability of the visual recognition algorithm, and the dynamic optimization of the offline robot visual recognition algorithm is completed. At the same time, inspection priority is assigned. Otherwise, a visual recognition algorithm adaptability warning is issued, and preset personnel are prompted to intervene (such as manually updating the algorithm model corresponding to the offline robot visual recognition algorithm and checking the corresponding hardware configuration). The reduction in instantaneous airflow deviation and station load rate deviation represent the differences between the instantaneous airflow deviation and station load rate deviation obtained before dynamic optimization and the instantaneous airflow deviation and station load rate deviation obtained after dynamic optimization, respectively. The set reduction values ​​include the set reduction values ​​for instantaneous airflow deviation and station load rate deviation, which are represented by the average of the historical reduction values ​​for instantaneous airflow deviation and station load rate deviation in the historical visual recognition process in the updated operation and maintenance knowledge base.

[0048] In this embodiment, the core of adjusting the spectral parameters of the offline robot's supplementary lighting is to specifically reduce interference. When the instantaneous airflow deviation exceeds the limit, the supplementary lighting adjustment target is determined based on the mapping relationship between the degree of flow difference and the supplementary lighting intensity in the operation and maintenance knowledge base: if it is due to reflective interference, the proportion of 550nm yellow-green light can be increased, as its wavelength has strong penetrating power and can reduce specular reflection on the flow meter surface; if there is a shadow, the overall proportion of 450-650nm visible light can be increased to enhance the uniformity of ambient light.

[0049] It's important to understand that the specific values ​​(550nm, 450-650nm) are used to provide precise spectral targeting for supplementary lighting adjustments: the 550nm yellow-green light can specifically match the reflective characteristics of the flowmeter's metal or glass surface, reducing specular reflection through a specific wavelength; the 450-650nm visible light covers the wavelength range sensitive to the human eye, balancing the brightness of shadow areas and ensuring the accuracy of the robot's vision algorithm in resolving the scale. In practical applications, fine-tuning can be made based on the flowmeter material (such as plastic, stainless steel), surface finish, and ambient light (such as direct sunlight or dim corners). For example, when the metal surface has strong reflectivity, the proportion of yellow-green light can be increased by 5%-10%, and in environments with a lot of dust, it can be broadened to 400-700nm to enhance penetration, ensuring that the spectral parameters dynamically adapt to the actual interference scenario.

[0050] The indirect impact of instantaneous airflow deviation and station load rate deviation is mainly achieved through the basic conditions of visual recognition: both rely on the visual acquisition capabilities of offline robots. If there are common interferences on site (such as uneven lighting or dust coverage), the recognition accuracy of the flow meter and load rate indicator may be reduced simultaneously, causing the two deviations to increase synchronously. In addition, systematic deviations in the robot's vision algorithm (such as insufficient generalization ability of the recognition model) may also affect the interpretation of the readings of the two types of instruments, causing the flow deviation and load rate deviation to fluctuate in a linked manner.

[0051] This example demonstrates how targeted optimizations (such as adjusting supplemental lighting parameters and verifying instrument status) continuously improve the robot's accuracy in recognizing key instruments, reducing misjudgments of air supply and load allocation errors caused by data deviations. Adaptability verification ensures the effectiveness of optimization, reduces the frequency of manual intervention, provides reliable data support for unmanned air compressor stations, guarantees the accuracy of stable operation of single stations and collaborative scheduling of multiple stations, and significantly improves the level of intelligent operation and maintenance and the system's anti-interference capability.

[0052] Furthermore, the inspection priority is dynamically set by defining the inspection time interval for different inspection areas. The specific process is as follows: if the obtained average deviation value is greater than the historical average deviation maximum value in the updated maintenance knowledge base, it is determined to be the first priority, and the corresponding inspection time interval is set as the first inspection time interval, usually set to 2 hours / time. The average deviation value is the offline inspection data of the offline robot during the inspection process, which is used to reflect the degree of difference between the real-time reading of the pressure instrument in each inspection area of ​​the air compressor station and the corresponding online collected data. The historical average deviation maximum value is represented by the sum of the historical deviation maximum values ​​in each historical inspection priority division process in the updated maintenance knowledge base.

[0053] If the obtained average deviation value is greater than the historical average deviation in the updated maintenance knowledge base but not greater than the maximum historical average deviation value, it is determined to be of the second priority, and the corresponding inspection interval is set as the second inspection interval, usually set to 8 hours / time. The historical average deviation is represented by the sum of all historical deviations in the historical inspection priority division process in the updated maintenance knowledge base. If the obtained average deviation value is greater than the minimum historical average deviation in the updated maintenance knowledge base but not greater than the historical average deviation, it is determined to be of the third priority, and the corresponding inspection interval is set as the third inspection interval, usually set to 16 hours / time. The minimum historical average deviation is represented by the sum of all minimum historical deviations in the historical inspection priority division process in the updated maintenance knowledge base. If the obtained average deviation value is less than the minimum historical average deviation in the updated maintenance knowledge base, it is determined to be of no priority, and the offline robot is prompted to maintain the current inspection interval, such as 24 hours / time.

[0054] After the inspection priority is assigned, a supplementary inspection is triggered to correct the priority misjudgment caused by momentary interference. The average deviation value re-acquired during the supplementary inspection is synchronized to the operation and maintenance knowledge base to avoid judgment errors caused by the lag of historical data. Through this real-time update mechanism, the operation and maintenance knowledge base can continuously adapt to the actual operating status of each inspection area, so that the priority assignment logic always keeps a dynamic match with the on-site situation.

[0055] The first, second, and third priorities are divided in descending order of average deviation value, with the corresponding inspection intervals gradually increasing, reflecting the decreasing severity of data deviation at the air compressor station and the decreasing urgency of inspection needs.

[0056] This example sets inspection intervals based on average deviation values, allowing the robot to autonomously focus on high-deviation areas (2 hours / time) and reduce inspections in low-deviation areas (maximum 24 hours / time), achieving optimal resource allocation without human intervention. It supplements inspections to correct misjudgments and updates the knowledge base, ensuring that priority allocation matches dynamic on-site conditions and preventing missed inspections due to insufficient manual inspections. The entire process requires no on-site personnel; through autonomous robot scheduling and data closed-loop optimization, it ensures the stable operation of the air compressor station in unattended mode.

[0057] This invention provides an LLM-based anomaly detection system for air compressor stations, such as... Figure 4The diagram illustrates the structure of an LLM-based air compressor station anomaly detection system. The system's processing flow includes: an air compressor station anomaly identification module, an air compressor station anomaly handling module, and an air compressor station inspection priority allocation module. The anomaly identification module acquires online and offline data during air compressor station operation and performs cross-validation. It also utilizes an operation and maintenance knowledge base (compiled from daily maintenance data and equipment technical manuals) and LLM to enable automatic inspection and anomaly identification. The anomaly handling module determines and grades anomalies based on the LLM's automatic anomaly identification results and acquires real-time anomaly data to drive dynamic optimization of the offline robot's visual recognition algorithm. The inspection priority allocation module acquires offline inspection data from the corresponding offline robot after successful adaptation, allocates inspection priorities to dynamically set inspection time intervals for different inspection areas, and generates inspection reports based on the priority allocation results to visualize the distribution of each priority area across the air compressor station's inspection areas.

[0058] like Figure 5 The diagram shown illustrates the architecture of an LLM-based air compressor station anomaly detection system, presenting a collaborative model of "online system - verification - offline robot," with each component's function clearly defined. Online System: Through IoT data acquisition, it aggregates diverse online data from air compressor stations, covering equipment operation data such as air compressors and refrigerated dryers, as well as pipeline flow, pressure, and ambient temperature and humidity, providing comprehensive data support for monitoring. Utilizing equipment monitoring (equipment dashboards, real-time monitoring, etc.) and intelligent algorithms (real-time KPI calculation, anomaly identification, etc.), combined with an operations and maintenance knowledge base (organizing operations and maintenance status and equipment manuals), it achieves automated inspection and anomaly identification via LLM, compensating for online data gaps and replacing manual preliminary troubleshooting. The interactive interface is presented as a webpage and mini-program for convenient remote viewing and operation.

[0059] The verification process acts as a bridge between the online system and the offline robot, cross-verifying online operational data and offline inspection data to ensure consistency and accuracy of data across time dimensions, stress monitoring, and other aspects. This lays a solid data foundation for anomaly detection and avoids misjudgments due to data deviations.

[0060] Offline robot inspections: Relying on technologies such as SLAM (Simultaneous Localization and Mapping) navigation and three-camera reconstruction, the robot inspects along designated routes, collecting data on water meters and equipment appearance (such as instrument readings and environmental videos) not covered by online monitoring, supplementing on-site data. It can also obtain richer equipment status information through infrared temperature measurement and noise detection. Its inspection data collaborates with online data to help comprehensively understand the status of the air compressor station; simultaneously, real-time anomaly data can drive dynamic optimization of its visual recognition algorithm, and inspection results are used for priority allocation, making operation and maintenance more intelligent and efficient.

[0061] The online system and offline robots transmit data via a network. The online system collects data using IoT and then transmits the data to the processing module via the network. Data collected by the offline robots during inspections is transmitted back to the online system in real time via wireless networks (such as Wi-Fi and 5G) to complete data interaction and verification. The system shown in the interface diagram achieves unattended intelligent operation and maintenance of air compressor stations through the integration of online and offline data and collaboration between LLM and robots, improving management efficiency and the speed of anomaly handling.

[0062] The above embodiments can be implemented, in whole or in part, by software, hardware (such as circuits), firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the flow or function according to the embodiments of the present invention is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. A computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. Available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media. Semiconductor media can be solid-state drives.

[0063] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0064] In various embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0065] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0066] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An LLM-based air compression station anomaly detection method, characterized in that, The method comprises: Step one, obtaining online and offline data in the running process of the air compression station and performing mutual data verification, and simultaneously through an operation and maintenance knowledge base for combing operation and maintenance conditions in the daily operation and maintenance process of the air compression station and equipment technical manuals, and combining LLM to realize automatic inspection of the air compression station to automatically identify abnormalities, the online and offline data comprising operation data collected online for reflecting the running state of the air compression station in real time and inspection data of an offline robot, the offline robot being used to obtain water meter data and equipment appearance information not covered online; Step two, based on the result of automatic identification of abnormalities by LLM, performing abnormal problem determination and hierarchical feedback, and simultaneously obtaining real-time abnormal data for driving dynamic optimization of a visual recognition algorithm of the offline robot, the abnormal problem determination being used to distinguish abnormal types in the running process of the air compression station, the abnormal types comprising air compression station body faults and sensing layer faults of the online and offline robots, the hierarchical feedback being used to grade the degree of abnormalities according to the abnormal types, and pushing the result of the degree of abnormality grading to an automatic control module of the corresponding air compression station; Step three, obtaining offline inspection data of the offline robot after adaptation qualification, performing inspection priority division to dynamically set the inspection time interval of different inspection areas, and simultaneously generating an inspection report based on the result of the inspection priority division to visualize the distribution position of each priority area in each inspection area of the air compression station, the offline inspection data being used to reflect the identification state of the pressure instrument of each inspection area of the air compression station.

2. The LLM-based air compression station anomaly detection method of claim 1, wherein, The mutual data verification comprises time synchronization determination, air compression abnormality determination in a single air compression station house, and multi-station house collaborative management in each air compression station house; The time synchronization determination is used to verify the time consistency of online pressure data collection and offline robot shooting of the pressure gauge, the air compression abnormality determination indicates that the air compression problem type of the air compression station is determined through online and offline collected pressure data, and the multi-station house collaborative management is used to determine the cross-station collaborative state by comparing the average pressure difference value of adjacent air compression station houses with the historical average pressure difference value; The time synchronization determination comprises the following specific steps: Obtaining a synchronization time deviation collected online and offline, the synchronization time deviation being used to determine whether the two types of data are in the same monitoring dimension to avoid verification errors caused by different time synchronization; If the obtained synchronization time deviation exceeds the set synchronization time deviation allowable range, a preset personnel is prompted to perform time synchronization calibration to ensure that subsequent air compression abnormality determination is based on pressure data in the same time dimension, otherwise, air compression abnormality determination is performed, the pressure data comprising exhaust pressure and pipeline pressure, the exhaust pressure being a pressure value at a pressure monitoring sensor installation node corresponding to the pressure of the compressed gas discharged by the air compressor, and the pipeline pressure being a pressure value at a pressure monitoring sensor installation node corresponding to the pressure of the compressed gas in the conveying pipeline.

3. The LLM-based air compression station anomaly detection method of claim 2, wherein, The air compression abnormality determination comprises the following specific steps: The online collected air compressor exhaust pressure and pipeline pressure are compared with the corresponding offline robot photographed air compressor panel pressure instrument reading values respectively to obtain the exhaust pressure difference value and the pipeline pressure difference value, and the corresponding set values are synchronously obtained, the set values include the set exhaust pressure difference value and the set pipeline pressure difference value, the exhaust pressure difference value and the pipeline pressure difference value are both used to reflect the difference degree of the online collected pressure data and the offline robot collected pressure instrument reading in the pressure monitoring result, the pressure instrument reading values include the air compressor exhaust pressure reading value and the pipeline pressure reading value; If the obtained exhaust pressure difference value and pipeline pressure difference value are both not greater than the corresponding set value, it is marked as no air pressure abnormality; If one of the obtained exhaust pressure difference value and pipeline pressure difference value is greater than the corresponding set value, it is marked as a first-level air pressure abnormality, and offline robot re-inspection and LLM preliminary verification are performed; If the obtained exhaust pressure difference value and pipeline pressure difference value are both greater than the corresponding set value, it is marked as a second-level air pressure abnormality and triggers manual review to develop a corresponding emergency treatment plan; The first-level air pressure abnormality and the second-level air pressure abnormality reflect the increasing degree of the abnormality degree of the air pressure problem in the air pressure station in turn.

4. The LLM-based air compression station anomaly detection method of claim 2, wherein, The multi-station house cooperative management has the following specific process: When comparing each air pressure station house, if the maximum value of the average pressure difference value between adjacent air pressure station houses in the data mutual verification period is greater than the historical average pressure difference value, it is marked as a cross-station cooperative abnormality and triggers a regional load rescheduling mechanism to balance the operating pressure between each air pressure station house; If the maximum value of the average pressure difference value between adjacent air pressure station houses in the data mutual verification period is not greater than the historical average pressure difference value, it is marked as a cross-station cooperative no abnormality and the results of the multi-station house cooperative management are input to the operation and maintenance knowledge base to update the baseline data of the cooperative operation between the air pressure stations.

5. The LLM-based air compression station anomaly detection method of claim 4, wherein, The offline robot re-inspection and LLM preliminary verification have the following specific process: The air pressure station house area corresponding to the first-level air pressure abnormality is obtained, and the visual recognition algorithm of the offline robot is triggered to obtain the pressure instrument reading value again, and a preliminary verification report is generated, the preliminary verification report is used to visualize the difference value reduction amplitude index of the re-obtained exhaust pressure difference value and pipeline pressure difference value; If the obtained difference value reduction amplitude index is greater than the set difference value reduction amplitude index, it is determined that the offline robot re-inspection and LLM preliminary verification are qualified, and the results after the offline robot re-inspection and LLM preliminary verification are input to the operation and maintenance knowledge base to update the baseline data of the pressure verification of the corresponding air pressure station; If the obtained difference value reduction amplitude index is not greater than the set difference value reduction amplitude index, it is determined that the offline robot re-inspection and LLM preliminary verification are unqualified, and an air pressure problem warning is performed, and the warning information is pushed to the operation and maintenance end.

6. The LLM-based air compression station anomaly detection method of claim 1, wherein, The abnormal problem judgment and grading feedback include data cross-validation and abnormal degree grading, and abnormal degree grading feedback; The data cross-validation and abnormal degree grading are used to combine the updated operation and maintenance knowledge base, identify abnormal problems through LLM, and divide abnormal types; The data cross-validation and abnormal degree grading have the following specific process: The fault-data feature mapping relationship in the updated operation and maintenance knowledge base is input into the LLM for abnormal problem identification to output an abnormal problem identification result; The abnormal problem identification result includes an ontology abnormality in the air compression station, an online sensing layer abnormality, and an offline robot sensing layer abnormality; The ontology abnormality in the air compression station indicates that the average value of the air compressor exhaust temperature collected online in the abnormal problem identification period is greater than the corresponding set value, and the cooling water pump flow is detected to be reduced to below the rated value, and the number of times that the cooling fan blade stops photographed by the offline robot is greater than the corresponding set value; The online sensing layer abnormality indicates that the fluctuation amplitude of the pipeline pressure collected online in the abnormal problem identification period is greater than the corresponding set value, and the fluctuation amplitude of the corresponding pressure instrument reading photographed by the offline robot is not greater than the corresponding set value, and there is no leakage trace in the pipeline appearance video; The offline robot sensing layer abnormality indicates that the average value of the difference between the temperature table reading photographed by the offline robot and the corresponding temperature data collected online in the abnormal problem identification period is greater than the corresponding set value, and the difference between the temperature table reading re-read after adjusting the shooting angle of the offline robot and the online data is not greater than the corresponding set value.

7. The LLM-based air compression station anomaly detection method of claim 6, wherein, The abnormality degree grading feedback, and the specific process is: The displayed ontology abnormality in the air compression station is pushed to the automatic calibration unit in the automatic control module to trigger dynamic optimization of the offline robot vision parameters, thereby reducing the same type of identification deviation; The displayed online sensing layer abnormality is pushed to the regional operation and maintenance unit in the automatic control module to automatically generate a maintenance work order and push it to the regional operation and maintenance personnel mobile terminal to prompt sensor verification or replacement; The displayed offline robot sensing layer abnormality is pushed to the emergency control unit in the automatic control module to immediately activate the emergency plan and send an emergency alarm to the operation and maintenance person in charge.

8. The LLM-based air compression station anomaly detection method of claim 7, wherein, The dynamic optimization of the offline robot vision recognition algorithm, and the specific process is: Real-time abnormal data is obtained based on the results of the abnormality degree grading feedback, the real-time abnormal data includes a transient air flow deviation and a station house load rate deviation, the transient air flow deviation is used to reflect the difference degree between the pipeline flow meter reading of the offline robot vision recognition and the transient air flow data collected online, and the station house load rate deviation is used to reflect the difference degree between the air compressor load rate indicator reading recognized by the offline robot and the actual load rate calculated based on the online operation data; If the obtained transient air flow deviation is not greater than the set transient air flow deviation, it indicates that the vision recognition accuracy of the offline robot to the pipeline flow meter meets the requirements, otherwise, the difference between the obtained transient air flow deviation and the set transient air flow deviation is matched with the mapping relationship between the flow difference degree and the light compensation intensity stored in the updated operation and maintenance knowledge base, and the matching degree result is used to prompt the preset personnel to adjust the spectrum parameters of the current offline robot light compensation lamp to reduce the interference of the flow meter surface reflection or shadow; If the obtained station house load rate deviation is greater than the set station house load rate deviation, the indicator physical state verification is performed, otherwise, it indicates that the recognition state of the robot to the load rate indicator is stable under the current environment. After the dynamic optimization of the offline robot visual recognition algorithm, adaptability verification is performed, specifically: If both the reduction amplitude of the obtained instantaneous air flow deviation and the reduction amplitude of the station house load rate deviation are greater than the corresponding set reduction amplitude, it indicates that the dynamic optimization is effective for improving the adaptability of the visual recognition algorithm, and the dynamic optimization of the offline robot visual recognition algorithm is completed, and the inspection priority division is performed. Otherwise, the visual recognition algorithm adaptability warning is performed and the preset personnel is prompted for intervention.

9. The LLM-based air compression station anomaly detection method of claim 1, wherein, The inspection priority division dynamically sets the inspection time interval of different inspection areas, and the specific process is: If the obtained average deviation value is greater than the maximum value of the historical average deviation in the updated operation and maintenance knowledge base, it is determined as the first priority, and the corresponding inspection time interval is set as the first inspection time interval. The average deviation value is used to reflect the difference between the real-time readings of the pressure instruments in each inspection area of the air compression station and the corresponding online collected data; If the obtained average deviation value is greater than the historical average deviation in the updated operation and maintenance knowledge base and not greater than the maximum value of the historical average deviation, it is determined as the second priority, and the corresponding inspection time interval is set as the second inspection time interval; If the obtained average deviation value is greater than the minimum value of the historical average deviation in the updated operation and maintenance knowledge base and not greater than the historical average deviation, it is determined as the third priority, and the corresponding inspection time interval is set as the third inspection time interval; If the obtained average deviation value is less than the minimum value of the historical average deviation in the updated operation and maintenance knowledge base, it is determined as no priority, and the offline robot is prompted to maintain the current inspection time interval; After the inspection priority division, a supplementary inspection is triggered to correct the priority misjudgment caused by instantaneous interference, and the average deviation value reacquired in the supplementary inspection process is synchronized to the operation and maintenance knowledge base to avoid the judgment deviation caused by the lag of historical data; The first priority, the second priority and the third priority are sequentially divided according to the average deviation value from large to small, and the corresponding inspection interval is gradually extended, reflecting the severity of the air compression station data deviation and the decreasing urgency of the inspection demand.

10. An LLM-based air compression station anomaly detection system applying the LLM-based air compression station anomaly detection method according to any one of claims 1-9, characterized in that, It includes: An air compression station abnormality identification module, an air compression station abnormality problem processing module, and an air compression station inspection priority division module; The air compression station abnormality identification module is used to acquire online and offline data during the operation of the air compression station and perform mutual data verification, and through the operation and maintenance knowledge base for analyzing the operation and maintenance of the air compression station during the daily operation and maintenance process and the operation and maintenance of the equipment technical specification, and combining LLM to realize automatic inspection of the air compression station to automatically identify abnormalities; The air compression station abnormality problem processing module is used to determine and feedback the abnormality according to the result of the LLM automatic identification of the abnormality, and to acquire real-time abnormal data for driving the dynamic optimization of the offline robot visual recognition algorithm; The air compression station inspection priority division module is used to acquire offline inspection data of the corresponding offline robot after the adaptability is qualified, to divide the inspection priority to dynamically set the inspection time interval of different inspection areas, and to generate an inspection report based on the result of the inspection priority division to visualize the distribution position of each priority area in each inspection area of the air compression station.

Images