A method for diagnosing oil and gas facility faults based on UAV aerial photography

Abstract

This invention discloses a method for diagnosing oil and gas facility faults based on UAV aerial photography, comprising the following steps: Step (1): Collecting multi-source monitoring data using a UAV equipped with a visible light camera, a thermal imaging camera, and a gas sensor; Step (2): Spatiotemporally registering the multi-source monitoring data to generate a fused feature map; Step (3): Inputting the fused feature map into a fault diagnosis model to output the fault type and original probability; Step (4): Calculating the dynamic fault confidence level based on environmental parameters and image quality factors; Step (5): Generating graded early warning signals and maintenance decision suggestions based on the dynamic fault confidence level. This invention can integrate multi-source data, accurately register and fuse it, dynamically correct fault probabilities, achieve efficient diagnosis and graded early warning, and improve the comprehensiveness, accuracy, and response efficiency of oil and gas facility fault diagnosis.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis, and more specifically to a method for diagnosing faults in oil and gas facilities based on drone aerial photography. Background Technology

[0002] Oil and gas facilities: Oil and gas facilities encompass all kinds of key equipment and facilities in the process of oil and gas extraction, transportation, storage and processing, such as oil and gas pipelines, storage tanks, gathering and transportation stations, valves, welds, etc. Their safe and stable operation is directly related to the efficiency and safety of oil and gas production. If corrosion, perforation, leakage, abnormal temperature or other failures occur, it may lead to safety accidents or economic losses.

[0003] Application of drone aerial photography in oil and gas facility fault diagnosis: Due to their flexibility, efficiency and accessibility to high-risk areas, drones are increasingly being used for fault diagnosis of oil and gas facilities. They can be equipped with visible light cameras, thermal imaging cameras and gas sensors to collect monitoring data and provide data support for fault diagnosis.

[0004] Existing technologies suffer from limitations in providing single protection types, relying on only a single data source or processing method. This makes it difficult to comprehensively and accurately identify faults, resulting in poor protection effectiveness and adversely affecting the safe operation of oil and gas facilities. Therefore, to address this issue, this paper proposes a fault diagnosis method for oil and gas facilities based on UAV aerial photography. Summary of the Invention

[0005] The present invention solves the above-mentioned technical problems through the following technical solution, and the present invention includes the following steps:

[0006] Step (1): Collect multi-source monitoring data using a drone equipped with a visible light camera, a thermal imaging camera, and a gas sensor;

[0007] Step (2): Perform spatiotemporal registration on the multi-source monitoring data to generate a fused feature map;

[0008] Step (3): Input the fused feature map into the fault diagnosis model and output the fault type and original probability;

[0009] Step (4): Calculate dynamic fault confidence based on environmental parameters and image quality factors;

[0010] Step (5): Generate graded early warning signals and maintenance decision suggestions based on dynamic fault confidence.

[0011] Furthermore, the spatiotemporal registration in step (2) includes:

[0012] Establish a global coordinate system based on UAV pose data and equipment 3D point cloud;

[0013] Align visible light images with thermal imaging images by feature point matching;

[0014] A two-way timing compensation module is used to synchronize gas sensor data with data from other sensors.

[0015] Furthermore, the generation of the fused feature map includes:

[0016] Extract visible light texture features (Fvis), thermal imaging temperature gradient features (Fthermal), and gas concentration temporal features (Fgas);

[0017] The weighting coefficients for each modality are calculated using a gated attention unit:

[0018] ;

[0019] Perform weighted feature fusion:

[0020] ;

[0021] Where || represents the feature concatenation operation. This represents element-wise multiplication. for Corresponding weights for Corresponding weights for Corresponding weights.

[0022] Furthermore, the method also includes a step of adaptive adjustment of fusion weights:

[0023] When the image quality factor QI is lower than the set threshold, the Wvis weight coefficient is reduced;

[0024] When the environmental disturbance factor QE is lower than the set threshold, the gas characteristic-dominated mode is activated:

[0025] ;

[0026] Where η3>η1+η2, and η1,η2,η3 are preset weight allocation parameters.

[0027] Furthermore, the dynamic fault confidence calculation in step (4) includes:

[0028] Calculate the image quality factor (QI):

[0029] ;

[0030] in Ii represents the image gradient magnitude, and μ and σ are the mean and standard deviation of the gradient magnitude, respectively;

[0031] Calculate the environmental disturbance factor QE:

[0032] ;

[0033] in This is the actual temperature. This represents the actual humidity. and The preset reference temperature and humidity are denoted by k, which is the environmental interference sensitivity adjustment parameter.

[0034] Correcting the probability of failure:

[0035] ;

[0036] Where α and β are adjustable compensation coefficients.

[0037] Furthermore, the generation of the graded early warning signal in step (5) includes:

[0038] Set the dynamic confidence threshold function:

[0039] ;

[0040] Base th The baseline threshold is γ, and the sensitivity adjustment coefficient is γ.

[0041] When Padj > Threshold dynamic At that time, a warning signal corresponding to the fault level is generated.

[0042] Furthermore, the fault diagnosis model in step (3) is a multi-task deep learning network, which is trained by optimizing the loss function to achieve multi-task processing of the fused feature map, and performs the following simultaneously:

[0043] Fault region localization based on the fused feature map;

[0044] Fault type classification;

[0045] Regression analysis of damage severity;

[0046] The loss function of a network is defined as:

[0047] ;

[0048] Wherein, λ1, λ2, and λ3 are the weighting coefficients of each loss term;

[0049] This represents the loss function for the fault location task, used to optimize the network's accuracy in pinpointing the specific location of the fault. This represents the loss function for the fault type classification task, used to optimize the network's accuracy in classifying fault types (such as corrosion, leakage, etc.). The loss function represents the task of regression analysis of damage degree, which is used to optimize the network's regression accuracy on the degree of fault damage (such as corrosion depth, leakage rate, etc.).

[0050] Furthermore, the maintenance decision recommendation generation in step (5) includes:

[0051] Establish a knowledge base mapping fault types to maintenance measures;

[0052] When the dynamic fault confidence level exceeds the warning threshold, a maintenance plan is automatically matched and its implementation priority is marked.

[0053] Compared with existing technologies, this invention has the following advantages: This method for diagnosing oil and gas facility faults based on UAV aerial photography integrates multi-source data acquisition and fusion, combining visible light, thermal imaging, and gas sensing information to provide more comprehensive monitoring data and improve the comprehensiveness of fault diagnosis. Spatiotemporal registration ensures that data collected at different times and of different types maintains consistency in space and time, providing a reliable foundation for subsequent feature fusion and diagnosis. The fusion feature map generation dynamically allocates weights through a gating attention mechanism and can adaptively adjust according to image quality and environmental interference, optimizing the feature fusion effect, highlighting key features, and improving diagnostic accuracy. A multi-task deep learning network simultaneously achieves fault area localization, type classification, and damage degree regression, improving the comprehensiveness and efficiency of fault diagnosis. Dynamic fault confidence calculation combines image quality and environmental interference factors to correct the fault probability, reducing the impact of external factors on the diagnostic results and enhancing the reliability of fault assessment. The graded early warning uses dynamic thresholds, which can flexibly adjust the early warning standards according to the actual situation to achieve accurate graded early warning; maintenance decision suggestions can automatically match solutions and mark priorities, assisting in the rapid formulation of maintenance strategies and improving fault response efficiency. Attached Figure Description

[0054] Figure 1 This is the overall flowchart of the present invention. Detailed Implementation

[0055] The embodiments of the present invention are described in detail below. These embodiments are implemented based on the technical solution of the present invention, and provide detailed implementation methods and specific operation processes. However, the scope of protection of the present invention is not limited to the following embodiments.

[0056] like Figure 1 As shown, this embodiment provides a technical solution: a method for diagnosing oil and gas facility faults based on UAV aerial photography, comprising the following steps:

[0057] Step (1): Collect multi-source monitoring data using a drone equipped with a visible light camera, a thermal imaging camera, and a gas sensor;

[0058] Step (2): Perform spatiotemporal registration on the multi-source monitoring data to generate a fused feature map;

[0059] Step (3): Input the fused feature map into the fault diagnosis model and output the fault type and original probability;

[0060] Step (4): Calculate dynamic fault confidence based on environmental parameters and image quality factors;

[0061] Step (5): Generate graded early warning signals and maintenance decision suggestions based on dynamic fault confidence.

[0062] The spatiotemporal registration in step (2) includes:

[0063] Establish a global coordinate system based on UAV pose data and equipment 3D point cloud;

[0064] Align visible light images with thermal imaging images by feature point matching;

[0065] A two-way timing compensation module is used to synchronize gas sensor data with data from other sensors.

[0066] Furthermore, the generation of the fused feature map includes:

[0067] Extract visible light texture features (Fvis), thermal imaging temperature gradient features (Fthermal), and gas concentration temporal features (Fgas);

[0068] The weighting coefficients for each modality are calculated using a gated attention unit:

[0069] ;

[0070] Perform weighted feature fusion:

[0071] ;

[0072] Where || represents the feature concatenation operation. This represents element-wise multiplication. for Corresponding weights for Corresponding weights for Corresponding weights;

[0073] By establishing a global coordinate system, aligning different types of images, and synchronizing sensor data, we can ensure that multi-source monitoring data are accurately correlated in spatial and temporal dimensions, eliminate interference caused by data misalignment or asynchrony, and provide a consistent and reliable foundation for subsequent feature fusion and fault diagnosis, thereby improving the accuracy of diagnosis.

[0074] For example, when monitoring the gas pipeline of an oil and gas gathering and transportation station, the visible light camera, thermal imaging camera and gas sensor carried by the UAV collect data respectively.

[0075] Without establishing a global coordinate system based on UAV pose data and equipment 3D point cloud, pipeline segment data captured at different times may lack a unified spatial reference, making it impossible to stitch together a complete pipeline spatial distribution. Maintenance personnel will find it difficult to locate the specific location of the fault in the entire gathering and transportation station (for example, they may not be able to distinguish whether the fault is in pipeline area A or area B). However, after establishing a global coordinate system, all data is anchored under a unified spatial framework, which can clearly mark the precise coordinates of the fault point in the 3D model of the gathering and transportation station, providing clear guidance for maintenance.

[0076] Suppose a hidden crack exists at a weld in a pipeline. In a visible light image, this location might only show a slight color difference due to lighting conditions. In a thermal image, the crack would cause gas leakage and friction, resulting in a localized temperature increase. If the two images are not aligned using feature point matching, the color difference area in the visible light image and the temperature anomaly area in the thermal image might be misaligned by 5 centimeters, leading to misjudgments of two independent, normal phenomena. However, after feature point matching, the two can be accurately linked as the same location. Combining the color difference and temperature anomaly, a clear diagnosis can be made that a leak is caused by a crack at the weld.

[0077] When the drone flies to the pipeline valve, the image sensor first captures a slight deformation of the valve seal (visible light image). 1.5 seconds later, the gas sensor detects a sudden increase in the surrounding methane concentration. If the data is not synchronized via a two-way time-series compensation module, the increase in methane concentration may be incorrectly mapped to images of other equipment taken 1.5 seconds later, ignoring the correlation with the valve deformation. However, after synchronization, the data is calibrated to the same time dimension, clearly establishing the temporal correlation between the valve seal deformation and the increase in methane concentration, thus accurately determining that the gas leak is caused by valve seal failure.

[0078] The method also includes a fusion weight adaptive adjustment step:

[0079] When the image quality factor QI is lower than the set threshold, the Wvis weight coefficient is reduced;

[0080] When the environmental disturbance factor QE is lower than the set threshold, the gas characteristic-dominated mode is activated:

[0081] ;

[0082] Where η3>η1+η2, η1,η2,η3 are preset weight allocation parameters;

[0083] The advantage of the fusion feature map generation method is that by extracting multimodal features (visible light texture, thermal imaging temperature gradient, gas concentration time series) and using gated attention units to dynamically calculate the weights of each modality for weighted fusion, it can effectively integrate the key information of different types of data, highlight the features that are more important for fault diagnosis, suppress irrelevant or noisy information, and provide more accurate and effective fusion features for subsequent fault diagnosis models, thereby improving the accuracy of fault diagnosis.

[0084] For example, during the monitoring of an oil and gas storage tank, a slow leak was found caused by a tiny perforation at the bottom of the tank.

[0085] The visible light camera captured fine rust marks (texture features Fvis) on the bottom of the tank, but the features were not obvious due to the small area of ​​the rust marks and the reflection of light.

[0086] The thermal imaging camera detected a temperature gradient at the perforation point where the gas leaked and rubbed against the air, resulting in a slightly higher local temperature than the surrounding area (temperature gradient feature).

[0087] The gas sensor recorded a slow upward trend in the concentration of combustible gas around the leak point over time (time-series characteristic Fgas).

[0088] Without a fusion approach, simply splicing or treating the three features equally may lead to difficulties in identifying weak fault signals due to interference from reflected light in visible light features and improper weight allocation. However, when calculating weights using a gated attention unit, it is found that the temporal features of gas concentration and the temperature gradient features of thermal imaging are more strongly correlated with leakage faults, thus assigning them higher weights (Wgas and Wthermal are larger), while visible light texture features are given lower weights due to interference (Wvis is smaller). The Ffused model generated after weighted fusion highlights the concentration increase trend and abnormal temperature gradient, enabling the fault diagnosis model to more accurately identify minor perforation leaks at the bottom of the storage tank, avoiding missed detections due to interference from a single feature or weight imbalance.

[0089] The dynamic fault confidence calculation in step (4) includes:

[0090] Calculate the image quality factor (QI):

[0091] ;

[0092] in Ii represents the image gradient magnitude, and μ and σ are the mean and standard deviation of the gradient magnitude, respectively;

[0093] Calculate the environmental disturbance factor QE:

[0094] ;

[0095] in This is the actual temperature. This represents the actual humidity. and The preset reference temperature and humidity are used, and k is an environmental disturbance sensitivity adjustment parameter, which is used to control the influence of the sum of temperature and humidity deviations on QE: the larger the k value, the faster QE decays under the same temperature and humidity deviation (i.e., the environmental disturbance is considered more significant); the smaller the k value, the more gradual the influence of temperature and humidity deviations on QE (i.e., the lower the sensitivity to environmental disturbances).

[0096] Correcting the probability of failure:

[0097] ;

[0098] Where α and β are adjustable compensation coefficients;

[0099] By calculating the image quality factor (QI) and the environmental interference factor (QE), the original fault probability output by the fault diagnosis model is corrected to obtain the dynamic fault confidence. This can effectively reduce the interference of image quality (such as blur and noise) and environmental factors (such as temperature and humidity) on the diagnostic results, making the fault probability assessment more in line with the actual scenario, improving the reliability and accuracy of fault diagnosis, and avoiding misjudgment or missed judgment caused by external conditions.

[0100] For example, when monitoring the fault of an oil pipeline in an oil field, a drone captured a visible light image of a section of the pipeline in hazy weather (the image was blurry and the gradient information was unclear). At the same time, the actual humidity in the area was much higher than the reference humidity (significant environmental interference).

[0101] Without dynamic confidence calculation, the fault diagnosis model may be based on blurry images and gas sensor data affected by high humidity, outputting an 80% raw probability (Praw) of leakage in the pipeline. However, this result does not take into account the effects of poor image quality and environmental interference, and may lead to misjudgment (for example, the suspected leakage traces in the blurry image are actually light and shadow errors caused by smog).

[0102] First, the image quality factor QI is calculated (QI value is low due to image blur), then the environmental interference factor QE is calculated (QE value is low due to humidity deviating from the reference value). The original probability is then corrected using QI and QE to obtain the adjusted fault probability Padj. Assuming Padj drops to 50% after correction, below the dynamic threshold, the system will not generate a false leak warning, thus avoiding erroneous decisions caused by image quality and environmental interference.

[0103] Conversely, in a sunny and dry environment, the image is clear (high QI) and there is little environmental interference (high QE). The original probability Praw is 70%, and after correction, Padj may rise to 75%. If it exceeds the dynamic threshold, an early warning will be accurately generated to ensure that real faults are not missed.

[0104] Step (5) of generating the graded early warning signal includes:

[0105] Set the dynamic confidence threshold function:

[0106] ;

[0107] Base th The baseline threshold is γ, and the sensitivity adjustment coefficient is γ.

[0108] When Padj > Threshold dynamic At that time, a warning signal corresponding to the fault level is generated;

[0109] By setting a dynamic confidence threshold function, the warning threshold can be dynamically changed according to the adjusted fault probability (Padj), rather than using a fixed threshold. This allows for a more accurate match with the actual situation of different fault probabilities, avoiding over-warning (false alarms of low-probability faults) or delayed warnings (missed alarms of high-probability faults) caused by fixed thresholds. This enables more realistic tiered warnings and improves the targeting and reliability of warnings.

[0110] For example, when monitoring the separator equipment in an oil and gas processing plant, two scenarios may occur:

[0111] The equipment has a slight loose seal, and the adjusted failure probability Padj is 65%. If a fixed threshold (such as 70%) is used, it may not issue a warning because 65% is below the threshold, causing the minor fault to be overlooked. However, in the dynamic threshold function of this case, the base threshold Baseth is set to 50%, and the sensitivity coefficient γ is 20, at which point the dynamic threshold is 56.8%. Since 65% > 56.8%, the system generates a low-level warning, prompting that a routine inspection of the seal should be arranged to avoid missed detection.

[0112] A serious valve rupture occurred in the equipment, with Padj reaching 90%. The dynamic threshold function calculated it to be 64.2%. Because 90% > 64.2%, the system generated a high-level warning, prompting immediate shutdown and maintenance to prevent large-scale leakage. While a fixed threshold of 50% would also trigger a warning, the dynamic threshold automatically increases the threshold based on high-probability faults, more accurately reflecting the severity of the fault, avoiding confusion with warning levels for low-probability faults, and improving the targeted nature of emergency response.

[0113] The fault diagnosis model in step (3) is a multi-task deep learning network. This network is trained by optimizing the loss function to achieve multi-task processing of the fused feature map, which is performed simultaneously:

[0114] Fault region localization based on the fused feature map;

[0115] Fault type classification;

[0116] Regression analysis of damage severity;

[0117] The loss function of a network is defined as:

[0118] ;

[0119] Wherein, λ1, λ2, and λ3 are the weighting coefficients of each loss term;

[0120] This represents the loss function for the fault location task, used to optimize the network's accuracy in pinpointing the specific location of the fault. This represents the loss function for the fault type classification task, used to optimize the network's accuracy in classifying fault types (such as corrosion, leakage, etc.). The loss function represents the task of regression analysis of damage degree, which is used to optimize the network's regression accuracy on the degree of fault damage (such as corrosion depth, leakage rate, etc.).

[0121] Employing a multi-task deep learning network to simultaneously perform fault area localization, fault type classification, and damage degree regression analysis can improve computational efficiency by sharing the feature extraction process. Furthermore, the tasks can assist each other (e.g., fault area features assist in type judgment, and type information feeds back into damage degree assessment), thereby improving the accuracy and comprehensiveness of the overall diagnosis. At the same time, it outputs more comprehensive fault information (location, type, and damage degree), providing a more detailed basis for subsequent maintenance decisions.

[0122] For example, during fault monitoring of an oil and gas pipeline, a section of the pipeline developed a perforation leak due to long-term corrosion.

[0123] If a single-task model is used, the localization model, classification model, and damage regression model need to be trained separately: the localization model may only output that there is an abnormal area in the middle section of the pipeline, but cannot determine the fault type of the area; the classification model may determine that it is a leakage fault, but cannot locate the specific location; the damage regression model may conclude that the damage is serious, but lacks type and location information, which means that maintenance personnel need to check multiple times to clarify the specific problem.

[0124] Multi-task networks process fused feature maps simultaneously:

[0125] The fault area was located 30 to 32 meters in the middle section of the pipeline (location task).

[0126] The fault type was classified as a perforation leak caused by corrosion (classification task).

[0127] Regression analysis showed that the damage was severe, with a perforation diameter of about 5 mm and the wall thickness of the surrounding 10 cm area reduced to 40% of the design value (regression task).

[0128] The three tasks share the feature extraction process, which not only reduces the time cost of model training and running, but also makes the output results more coherent and accurate because of the complementary information between tasks (such as the texture and temperature features of the localized area help to confirm the corrosion type, and the corrosion type helps to more accurately assess the damage degree of wall thickness reduction). Maintenance personnel can directly prepare repair materials and tools of the corresponding specifications based on this information, and accurately go to the fault point for treatment, which greatly improves maintenance efficiency.

[0129] For example, the bottom of an oil and gas storage tank developed localized corrosion and perforation due to prolonged dampness.

[0130] When multi-task deep learning networks process fused feature maps, they include visible light textures, thermal imaging temperature gradients, and temporal features of gas concentration.

[0131] Fault location: By analyzing the overlapping area of ​​texture anomalies (rust) and temperature anomalies (frictional heat generation from leaking gas) in the fusion features, the fault point was accurately located as a 1.2m × 0.8m area on the south side of the bottom of the storage tank.

[0132] Fault type classification: Based on the loose corrosion characteristics of the visible light texture in this area, the local high temperature points (friction of leaking gas) in the thermal imaging, and the time sequence of methane concentration increase captured by the gas sensor, it is classified as corrosion perforation leakage caused by a humid environment.

[0133] Damage extent regression: Based on the texture ambiguity of the location area (reflecting corrosion thickness) and temperature gradient intensity (reflecting leakage), the regression results show that the corrosion depth reaches 55% of the tank wall thickness, the perforation diameter is about 3 mm, and the leakage rate is 0.02 m³ / h.

[0134] The three tasks are passed through the loss function. Collaborative optimization: The precise location provides spatial constraints for classification (eliminating interference from other areas), while the classified corrosion type helps the regression model focus on corrosion-specific damage indicators such as wall thickness reduction (rather than other damage such as impact). The final output of location, type, and damage severity information is coherent and accurate. Maintenance personnel can directly use this information to carry anti-corrosion and welding repair materials to the located area and develop targeted repair plans based on the degree of damage, avoiding the inefficiencies caused by multiple analyses required by single-task models and fragmented information.

[0135] Step (5) of generating maintenance decision recommendations includes:

[0136] Establish a knowledge base mapping fault types to maintenance measures;

[0137] When the dynamic fault confidence level exceeds the warning threshold, a maintenance plan is automatically matched and its implementation priority is marked.

[0138] By calculating the image quality factor (QI) and the environmental interference factor (QE), the original fault probability output by the fault diagnosis model is corrected to obtain a dynamic fault confidence level. This process effectively quantifies and compensates for the interference of image quality (such as blurring, noise, etc.) and environmental factors (such as temperature, humidity deviating from reference values, etc.) on the diagnostic results, making the corrected fault probability more consistent with the actual scenario. This reduces misjudgments (such as misjudging normal areas as faults due to poor image quality) or missed judgments (such as environmental interference masking the true fault signal) caused by external conditions, thereby improving the reliability and accuracy of fault assessment and providing a more accurate basis for subsequent graded early warning and maintenance decisions.

[0139] If the detected fault type in the oil and gas pipeline is minor corrosion leakage at the weld, the corresponding maintenance measure in the knowledge base is local grinding and rust removal, with a priority marked as medium. If the fault is a severe leak due to perforation of the tank wall, emergency emptying of the tank and replacement of the damaged section will be automatically matched, with a priority marked as high. Maintenance personnel can directly address high-risk faults based on the marked priorities and solutions, avoiding decision-making delays or incorrect measures.

[0140] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0141] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0142] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for diagnosing faults in oil and gas facilities based on UAV aerial photography, characterized in that, Includes the following steps: Step (1): Collect multi-source monitoring data using a drone equipped with a visible light camera, a thermal imaging camera, and a gas sensor; Step (2): Perform spatiotemporal registration on the multi-source monitoring data to generate a fused feature map; Step (3): Input the fused feature map into the fault diagnosis model and output the fault type and original probability; Step (4): Calculate dynamic fault confidence based on environmental parameters and image quality factors; The dynamic fault confidence calculation in step (4) includes: Calculate the image quality factor (QI): ； in Ii represents the image gradient magnitude, and μ and σ are the mean and standard deviation of the gradient magnitude, respectively; Calculate the environmental disturbance factor QE: ； in This is the actual temperature. This represents the actual humidity. and The preset reference temperature and humidity are used, and k is an environmental disturbance sensitivity adjustment parameter, which is used to control the influence of the sum of temperature and humidity deviations on QE: the larger the value of k, the faster QE decays under the same temperature and humidity deviation; the smaller the value of k, the more gradual the influence of temperature and humidity deviations on QE. Correcting the probability of failure: ； Where α and β are adjustable compensation coefficients. The original probability; Step (5): Generate graded early warning signals and maintenance decision suggestions based on dynamic fault confidence.

2. The method for diagnosing oil and gas facility faults based on UAV aerial photography according to claim 1, characterized in that: The spatiotemporal registration in step (2) includes: Establish a global coordinate system based on UAV pose data and equipment 3D point cloud; Align visible light images with thermal imaging images by feature point matching; A two-way timing compensation module is used to synchronize gas sensor data with data from other sensors.

3. The method for diagnosing oil and gas facility faults based on UAV aerial photography according to claim 2, characterized in that: The generation of the fused feature map includes: Extract visible light texture features (Fvis), thermal imaging temperature gradient features (Fthermal), and gas concentration temporal features (Fgas); The weighting coefficients for each modality are calculated using a gated attention unit: ; Perform weighted feature fusion: ； Where || represents the feature concatenation operation. This represents element-wise multiplication. for Corresponding weights for Corresponding weights for Corresponding weights.

4. The method for diagnosing oil and gas facility faults based on UAV aerial photography according to claim 3, characterized in that: It also includes a fusion weight adaptive adjustment step: When the image quality factor QI is lower than the set threshold, the Wvis weight coefficient is reduced; When the environmental disturbance factor QE is lower than the set threshold, the gas characteristic-dominated mode is activated: ； Where η3>η1+η2, and η1,η2,η3 are preset weight allocation parameters.

5. The method for diagnosing oil and gas facility faults based on UAV aerial photography according to claim 1, characterized in that: Step (5) of generating the graded early warning signal includes: Set the dynamic confidence threshold function: ； Base th The baseline threshold is γ, and the sensitivity adjustment coefficient is γ. When Padj > Threshold dynamic At that time, a warning signal corresponding to the fault level is generated.

6. The method for diagnosing oil and gas facility faults based on UAV aerial photography according to claim 1, characterized in that: The fault diagnosis model in step (3) is a multi-task deep learning network. This network is trained by optimizing the loss function to achieve multi-task processing of the fused feature map, which is performed simultaneously: Fault region localization based on the fused feature map; Fault type classification; Regression analysis of damage severity; The loss function of a network is defined as: ； Wherein, λ1, λ2, and λ3 are the weighting coefficients of each loss term; This represents the loss function for the fault location task. The loss function represents the fault type classification task. The loss function represents the regression analysis task for the degree of damage.

7. The method for diagnosing oil and gas facility faults based on UAV aerial photography according to claim 1, characterized in that: Step (5) of generating maintenance decision recommendations includes: Establish a knowledge base mapping fault types to maintenance measures; When the dynamic fault confidence level exceeds the warning threshold, a maintenance plan is automatically matched and its implementation priority is marked.

Images