Intelligent positioning method for boiler tube leakage

Abstract

The application provides a boiler furnace tube leakage intelligent positioning method, relates to the technical field of boiler detection and monitoring in electric power engineering, and has the beneficial effects that: first, the adaptive dynamic noise suppression technology is used to effectively distinguish the airflow noise and the leakage sound; second, the sub-millimeter level accurate positioning of the leakage point is realized through the double-field coupling path planning and the multi-modal space-time calibration technology; and third, the positioning response time is greatly shortened through the partition response scheduling strategy.

Description

Technical Field

[0001] This invention relates to the field of boiler detection and monitoring technology in power engineering, and in particular to an intelligent method for locating boiler tube leaks. Background Technology

[0002] Boiler tubes are the core heat-receiving components of thermal power plants, and their operating status directly determines the safety and economy of the unit. Boiler tube leakage is a common fault in boiler operation. If it cannot be detected and located in time, it may lead to serious safety accidents, affect power generation efficiency and equipment life. Therefore, real-time and accurate leakage monitoring technology is needed. Current technology mainly relies on audio sensors to perform acoustic monitoring of the boiler tube area and analyze changes in sound pressure level to determine whether a leak has occurred.

[0003] However, existing technologies using audio sensors to monitor boiler tube leaks can function normally in the absence of other signal interference. But the complex operating conditions inside a boiler, such as airflow fluctuations near the tubes caused by internal airflow or maintenance work, can lead to the audio sensor misinterpreting airflow noise as leakage sound, resulting in false alarms. Frequent false alarms not only put operators under stress but also interfere with their judgment and operational decisions regarding actual leaks. They may even lead to "alarm fatigue," reducing the efficiency of response to genuine leak alarms and increasing the risk of accidents. For example, interfering alarms may be mistakenly ignored as normal signals, or operators may be distracted by frequently dealing with false alarms. Therefore, existing technologies fail to effectively distinguish between airflow noise and leakage sound and lack the ability to adaptively process noise characteristics under different boiler load conditions, resulting in a false alarm rate exceeding 25% under high loads and a missed alarm rate of 15% under low loads. Summary of the Invention

[0004] To address the shortcomings of the existing technologies, the technical problem to be solved by this invention is to provide an intelligent method for locating boiler tube leaks. This method can intelligently schedule a cluster of drones to perform multimodal data acquisition, adaptively suppress airflow noise interference according to different boiler load conditions, and achieve sub-millimeter-level precise location of the leak point.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: The present invention provides an intelligent location method for boiler tube leakage, comprising the following steps: S1. Obtain boiler operating parameters, acoustic signals, airflow parameters, and UAV node environmental parameters based on the boiler operating status; S2. Construct an airflow fluctuation model based on boiler operating parameters, acoustic signals, and airflow parameters, and extract effective leakage sound signals through improved wavelet threshold filtering. ; S3. Based on the UAV node environmental parameters and effective leakage acoustic signals Divide the leakage response into partitions, calculate the scheduling priority of each node, filter the optimal UAV node, and select the optimal UAV node to perform the positioning task; S4. Based on the optimal UAV node and boiler operating parameters, construct a dual-field coupling model of leakage field and structural field, and plan the optimal flight path; S5. Based on the optimal flight path planned in step S4, and by fusing infrared thermal imaging data, BeiDou positioning data, and IMU attitude data, the final coordinates of the leak point are output using a spatiotemporal calibration algorithm. ; S6. Based on the final coordinates of the leak point Positioning deviation and effective leakage sound signal The steps S2 and S4 are dynamically adjusted to form a closed-loop optimization.

[0006] In a preferred embodiment, the boiler operating parameters in step S1 include real-time load. Furnace tube surface temperature The acoustic signals include the timing signals of the sound pressure level in the furnace tube area. The airflow parameters include the airflow velocity around the furnace tubes. The environmental parameters of the drone nodes include the location information of at least two drone nodes. Remaining battery power and signal strength .

[0007] In the preferred embodiment, in step S2, the real-time load is used. airflow speed and sound pressure level timing signal The specific steps for constructing the airflow fluctuation model in S2 are as follows: S21. Based on real-time load For sound pressure level timing signals implement Layer wavelet decomposition yields the wavelet coefficients of each layer. ,in Number of decomposition layers ( ), For coefficient index; Decomposition layer number Based on the real-time load output by S1 Adaptive adjustment; S22, Based on real-time load and Construct an airflow wave model based on wavelet coefficients. Calculate the first Filtering threshold of layer wavelet coefficients Filter threshold The calculation formula is: ; in, For the first Layer noise standard deviation, through wavelet coefficients The median value was calculated. for The signal length; This is the boiler's rated load. The rated airflow velocity around the furnace tube; This is the load influence factor; This is the airflow influence coefficient; S23. Wavelet coefficients of each layer Perform soft threshold processing; S24. Based on the wavelet coefficients processed in step S23 Perform wavelet reconstruction and output a valid leakage acoustic signal. .

[0008] In the preferred embodiment, the number of decomposition layers in step S21 is... Based on the real-time load output by S1 The adaptive adjustment rule is: when and When the preset high load threshold is used, ;when and When the preset low load threshold is used, ;when hour, ;in .

[0009] In the preferred embodiment, in step S3, the drone node position is output through S1. Remaining battery power Signal strength and S2 output The specific steps of step S3, which involves dividing the leakage response area, are as follows: S31. Based on the valid leakage sound signal The energy distribution divides the furnace tube monitoring area into at least [number missing]. One response partition Determine the center coordinates of each partition. ; S32. Obtain each response partition Internal drone nodes Based on the location of the drone node Remaining battery power Signal strength and real-time load Calculate each response partition Internal drone node priority Priority The calculation formula is: ; in, For nodes To the partition center The three-dimensional distance; ~ Preset weighting coefficients; , , These are the drone's maximum battery power, maximum signal strength, and maximum boiler load, respectively. S33. Select the priority calculated in step S32. The drone node corresponding to the maximum value is selected as the optimal execution node, and the position of the optimal execution node is output. .

[0010] In the preferred embodiment, in step S3, the optimal node position is determined. Partition Center and furnace tube surface temperature The specific steps for constructing the dual-field coupling model of the leakage field and the structural field in S4 are as follows: S41. Obtain a valid leakage sound signal maximum value According to the maximum value Define the attraction of the leakage field According to the surface temperature of the furnace tube Define the repulsive force of the structural field ; S42. Calculate the total force on the UAV based on the dual-field coupling model of leakage field and structural field. The drone's flight speed is updated using Newton's second law; S43. Perform a rationality check on the updated flight speed. If the speed exceeds the preset maximum speed limit, perform speed truncation and output the corrected flight path point sequence Path.

[0011] In the preferred embodiment, the leakage field attraction in step S41 The calculation formula is: ; Among them, the attractive force of the leakage field Used to guide the drone to the output of step S31 near; To preset the attraction coefficient, for The maximum value, For drones to distance, For pointing Unit vector; repulsive force of structural field The calculation formula is: ; Among them, the repulsive force of the structural field Used to prevent drones from touching high-temperature furnace tubes; As a preset repulsion coefficient, The distance from the drone to the furnace tube structure. This is the unit vector that deviates from the furnace tube.

[0012] In the preferred embodiment, step S5 is as follows: S51. Based on the corrected flight path point sequence Path, acquire thermal images of the furnace tubes using an infrared camera. Extract the image coordinates corresponding to the abnormal temperature at the leak point. The real-time location of the drone is collected via a BeiDou module. Real-time attitude angles of the UAV are collected via IMU. Real-time attitude angle Including roll, pitch, and yaw; S52, Based on the data collected in step S51 and Construct the camera extrinsic matrix [R|t], where For rotation matrix, The translation vector is combined with the camera's preset intrinsic parameter matrix. Establish the spatial coordinates of the leak point. With image coordinates Mapping relationship; introduction of IMU drift compensation amount For spatial coordinates Perform calibration to obtain calibrated coordinates. ; Among them, IMU drift compensation amount Estimated using the Kalman filter algorithm; S53, Minimize using a nonlinear optimization algorithm and The reprojection error is used to output the final coordinates of the leak point. .

[0013] In the preferred embodiment, the value of S22 is dynamically adjusted in step S6. , and S41 , The parameter adjustment rule is: when the positioning deviation... and When setting a high deviation threshold, increase to ,in To preset the upward adjustment coefficient, reduce and to , ,in The preset adjustment coefficient; when and Fine-tuning when setting a low deviation threshold to ,in This is the preset fine-tuning coefficient; when At that time, keep all parameters constant.

[0014] This invention provides an intelligent method for locating leaks in boiler tubes. Through the coordination of the aforementioned structures, it offers the following advantages compared to existing methods: First, through adaptive dynamic noise suppression technology, airflow noise and leakage noise are effectively distinguished, avoiding false identification caused by airflow fluctuations and maintenance interference in the boiler, and effectively reducing false alarm rate and missed alarm rate. Secondly, through dual-field coupling path planning and multimodal spatiotemporal calibration technology, sub-millimeter level precise positioning of the leak point is achieved. The optimal flight path that avoids high-temperature dense furnace tubes can be planned. At the same time, through the multimodal spatiotemporal calibration of the infrared camera, Beidou positioning module and IMU carried by the UAV, the spatial position of the leak point can be captured, avoiding monitoring failure caused by positioning deviation or UAV collision with furnace tubes, ensuring the accuracy of leak positioning and the safety of UAV operation. Third, the partitioned response scheduling strategy significantly shortens the location response time. It can quickly select the optimal node in the drone swarm that is close to the leak area, has sufficient remaining power, and has stable signal transmission to execute the task. It does not require relying on individual drones to inspect one by one, which greatly shortens the time from discovering a suspected leak to completing accurate location. This helps operators to quickly intervene and deal with the situation to reduce the risk of working fluid loss and accident expansion. Attached Figure Description

[0015] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a flowchart illustrating the overall process structure of the method of the present invention. Figure 2 This is a schematic diagram of wavelet decomposition and thresholding for dynamic noise suppression in this invention; Figure 3 This is a schematic diagram of the force field model for the dual-field coupled path planning of the present invention; Figure 4 This is a schematic diagram of coordinate mapping for multimodal spatiotemporal calibration of the present invention; Figure 5 This is a diagram of the MM-CAFNet multimodal fusion model architecture of the present invention; Figure 6 This is a schematic diagram of the CBAM attention module structure of the present invention; Figure 7 This is a training curve diagram of the model of the present invention; Figure 8 This is a bar chart comparing the classification accuracy of different models in this invention. Detailed Implementation

[0016] To better understand the purpose, system architecture, and functional implementation of this embodiment, the embodiments and features in the embodiments of this application can be combined with each other without conflict. The exemplary embodiments disclosed in this application will be described below with reference to the accompanying drawings, including specific technical details disclosed in this embodiment to aid understanding; however, these details should be considered exemplary rather than restrictive. Therefore, those skilled in the art should understand that various improvements and adjustments can be made to the embodiments described herein without departing from the scope and core ideas of the invention. Similarly, for clarity, detailed descriptions of well-known technologies, functions, and structures (such as the conventional data transmission process of a boiler DCS system, the basic flight control principles of a UAV, etc.) are omitted in the following description.

[0017] In the field of boiler inspection and monitoring technology in power engineering, acoustic monitoring methods are widely used for boiler tube leak detection. By collecting acoustic signals from the tube area and analyzing their characteristics, real-time monitoring of the leak status can be achieved. In many application scenarios, acoustic monitoring methods can effectively identify specific frequency band sound wave signals generated by leaks, providing early warning information to operators.

[0018] As boiler operating conditions become increasingly complex, acoustic monitoring technology has gradually evolved from single-sensor detection to multi-sensor fusion detection to meet the monitoring needs of harsh environments such as high temperature and high pressure. Multimodal detection systems can fuse and analyze various information such as acoustic signals, temperature signals, and airflow parameters, typically yielding more accurate detection results than single acoustic monitoring. In the precise location of boiler tube leaks under complex operating conditions, multimodal detection systems usually need to be combined with intelligent algorithms to achieve accurate leak identification and location. However, conventional monitoring methods often struggle to effectively distinguish between leak signals and interference noise when dealing with variable operating conditions. Existing UAV monitoring solutions are mostly limited to single-modal data acquisition, lacking deep integration with boiler operating parameters (load, tube temperature) and airflow parameters. Furthermore, path planning does not fully consider the dense distribution of tubes and high-temperature characteristics, making positioning accuracy susceptible to environmental interference, and emergency response efficiency difficult to match the unit's operation and maintenance needs.

[0019] Current technologies for boiler tube leak detection primarily employ fixed-threshold acoustic signal processing methods. These methods are limited to a single sound pressure level threshold for signal processing and static noise filtering for interference mitigation. Therefore, these boiler tube leak detection methods have limitations and are ill-suited to the high standards of accuracy and reliability required under complex operating conditions.

[0020] Example 1 like Figures 1-5 As shown, in boiler tube leak location, this embodiment is used to accurately identify the leak point area (temperature anomaly area) in the infrared thermogram and classify the effective leak features in the acoustic signal after filtering out interference. .

[0021] In the preferred embodiment, the model architecture adopts a dual-branch multimodal cross-attention fusion network (MM-CAFNet), as detailed below: The model adopts an end-to-end architecture of "dual-branch feature extraction, cross-attention fusion, and joint classification and localization," such as... Figure 5 As shown, the functions of each module are as follows: (1) Input layer: multimodal data preprocessing Specifically, the 1280×960 infrared thermal image acquired by the infrared image branch input. The preprocessing is 256×256×3, that is, the temperature channel is converted to RGB: R=high temperature area, G=normal temperature area, B=low temperature area, eliminating the inference delay caused by the excessive size of the original image.

[0022] Specifically, the acoustic signal branch input is an effective leakage acoustic signal. That is, the sampling frequency is 1kHz, the duration is 3s, the preprocessed 1D time series data is 1×3000, and then converted into a 32×32 time-frequency graph (preserving the time-frequency characteristics) through short time Fourier transform (STFT) to adapt to 1D-CNN input.

[0023] (2) Feature extraction branch Specifically, the image branch uses an improved ResNet-50 (with the addition of a channel-spatial attention module). To address the feature blurring issue caused by high-temperature distortion of the furnace tubes, a CBAM attention module is inserted after the 3rd and 4th residual blocks of the ResNet-50, such as... Figure 6 As shown, channel attention focuses on the temperature anomaly area, i.e., the temperature at the leak point is 50~200℃ higher than the surrounding area, and spatial attention suppresses background noise, such as temperature interference from the furnace tube support structure.

[0024] Channel attention involves compressing the 256×256×2048 feature map into a 1×1×2048 image using global average pooling, and then outputting the channel weights through a fully connected layer. Enhance the feature weights of temperature anomaly channels.

[0025] Spatial attention involves compressing the feature map output from channel attention to 256×256×1 using Conv2d, and then activating it to obtain spatial weights. Suppress leaks in non-leakage areas, such as where there are no abnormal temperatures on the pipe wall.

[0026] Among them, the acoustic branch is: 1D-CNN+BiLSTM. In this embodiment, considering the temporal correlation of acoustic signals, i.e., leakage sound has continuous time-frequency characteristics while interference noise is random bursts, a design is made to "extract local time-frequency features using 1D-CNN and capture long-term temporal dependencies using BiLSTM": The 1D-CNN layer consists of three convolutional layers with kernel sizes of 5, 3, and 3, a stride of 1, and a channel count that increases from 64 to 128 to 256. The time-frequency peak characteristics, that is, the time-frequency peak of the leakage sound is concentrated in 500~1000Hz.

[0027] The BiLSTM layer consists of two bidirectional LSTM layers (256 hidden dimensions) that capture the temporal correlation of acoustic signals, such as the duration of leakage sound > 0.5s and interference noise < 0.2s, and outputs a 256-dimensional temporal feature vector.

[0028] (3) Cross-attention fusion layer Specifically, to address the "heterogeneous gap" in multimodal features—that is, images represent spatial features while acoustics represent temporal features—a cross-attention mechanism is employed to achieve feature interaction. Among them, image features The 256×256×2048 vector is obtained by global average pooling, and the acoustic feature F_{ac} (256-dimensional) is mapped to a 2048-dimensional vector through a fully connected layer.

[0029] Specifically, the cross-attention weights are calculated as follows: ,in As a feature dimension, enhance the cross-modal feature weights related to leakage.

[0030] Among them, fusion features Balance the contributions of dual modes (e.g., when infrared is blocked, the acoustic feature weight is automatically increased).

[0031] (4) Output layer The classification head consists of a fully connected layer and Softmax, which outputs a binary classification result of "normal / leaking" with a confidence level of Conf (0~1). A Conf ≥ 0.9 indicates a valid leak.

[0032] The positioning head: the coordinates (u', v') of the leak point output by the fully connected layer on the 256×256 image are restored to the coordinates of the original 1280×960 image using a coordinate mapping formula. : , Because it is scaled 5 times from 256 to 1280.

[0033] In the preferred scheme, the sources and scale of the dataset used to construct the dataset during model training are as follows: In this embodiment, the infrared image dataset (15,000 images) is as follows: Among them, the data acquisition equipment is: FLIRA655sc infrared camera (1280×960, temperature measurement 300~600℃). The operating conditions covered are: 300MW (3,000 sheets), 600MW (5,000 sheets), and 1000MW (7,000 sheets). Among them, the leakage parameters are: leakage size 0.1mm (2,000 sheets), 0.5mm (5,000 sheets), 1mm (4,000 sheets), 2mm (4,000 sheets). The labels include: manually marked leak point coordinates (u,v) and operating condition type.

[0034] In this embodiment, the acoustic signal dataset (20,000 records) is as follows: Among them, the data acquisition device is a PCB356A16 acoustic sensor (1kHz sampling, 3s duration). The signal types are as follows: effective leakage sound (8,000), normal airflow noise (6,000), and maintenance interference noise (6,000). The labels are categorized into three types: "Leakage / Normal / Interference," which indicate the start and end times of the leakage sound.

[0035] In this embodiment, data augmentation (improving generalization): Among them, infrared image enhancement includes random rotation (0~10°), scaling (0.8~1.2 times), Gaussian blur (σ=0.1~0.3), and temperature channel offset (simulating ±5℃ temperature measurement error).

[0036] Among them, acoustic signal enhancement includes time shift (±0.2s), additive Gaussian noise (SNR=20~30dB), and amplitude scaling (0.8~1.2 times, analog signal attenuation).

[0037] In this embodiment, the dataset is divided as follows: The training set consists of 70% (10,500 images + 14,000 acoustic signals). The validation set consists of 15% (2,250 images + 3,000 acoustic signals). The test set comprises 15% (2,250 images + 3,000 acoustic signals). The test set conditions do not overlap with the training set. For example, if the training set does not have an 850MW load sample, the test set will include it.

[0038] In the preferred scheme, all models use a unified environment to eliminate the influence of hardware differences. The specific training environment and parameters are shown in Table 1 below: Table 1

[0039] In this embodiment, key metrics during the training process are monitored in real time using TensorBoard, such as... Figure 7 As shown, the core monitoring results and analysis are presented, with the loss curve as follows: Total training set loss: gradually decreased from the initial 2.3 (when parameters were randomly initialized), decreasing to 0.5 after 10 epochs, decreasing to 0.1 after 30 epochs, and stabilizing at 0.08 after 50 epochs; Total loss of the validation set: decreased to 0.6 after 10 epochs, decreased to 0.13 after 30 epochs, and stabilized at 0.12 after 50 epochs; Specifically, the difference between the training set and the validation set loss was always <0.04, indicating no overfitting, as data augmentation and AdamW weight decay effectively suppressed overfitting.

[0040] In this embodiment, the accuracy curve is as follows: Training set classification accuracy: gradually improved from an initial 65% (random guessing level), reaching 95% after 20 epochs, 99% after 40 epochs, and stabilizing at 99.2% after 50 epochs; Validation set classification accuracy: increased to 92% after 20 epochs, 97.5% after 40 epochs, and stabilized at 98.1% after 50 epochs; Specifically, the accuracy on the validation set is close to that on the training set, indicating that the model generalizes well and has no underfitting (because the model complexity is adapted to the size of the dataset).

[0041] In this embodiment, the positioning error curve is as follows: The validation set positioning error, which is at the pixel level, corresponds to a 256×256 image: it gradually decreases from an initial 8 pixels (corresponding to the original image of 40mm), decreasing to 2 pixels (10mm) at 20 epochs, decreasing to 0.8 pixels (4mm) at 40 epochs, and stabilizing at 0.6 pixels (3mm) at 50 epochs. Specifically, the localization error continued to decrease with each training epoch, indicating that the parameter learning of the localization head was effective and that the features fused by cross-attention provided sufficient support for the localization task.

[0042] The performance comparison and verification of the preferred scheme are as follows: Based on a 15% independent test set, a quantitative comparison was performed on six key metrics between the core model MM-CAFNet and three contrasting models. The results are shown in Table 2. Figure 8As shown in Table 2 below, the comparative logic and advantages analysis are as follows: Table 2

[0043] In this embodiment, Table 2 and... Figure 8 The key comparative conclusions are as follows: In this embodiment, the advantages in classification accuracy are as follows: This model achieves 98.5% accuracy compared to Model 1 (82.3%): the addition of an acoustic branch to MM-CAFNet offsets the high-temperature distortion of infrared images under high load (accuracy of 98.2% compared to 75.1% under 1000MW load). This model achieves 98.5% better performance than Model 3 (91.2%) because cross-attention fusion is better at capturing bimodal associations than "direct splicing" (the recognition of fused features is improved by 40%), avoiding misjudgments caused by single-modal interference.

[0044] In this embodiment, the inference speed is balanced as follows: This model takes 42ms / sample, which is slower than the comparison model 2 (28ms). Although the single-modal structure is simple, it is much faster than the comparison models 1 and 3. The speed meets the requirement that the total time for the S5 steps is ≤5s, that is, 42ms for a single sample inference and only 4.2s for 100 samples. Furthermore, the speed can be further increased to 25ms / sample through model quantization (INT8).

[0045] In this embodiment, the positioning accuracy breakthrough is as follows: The localization error of this model is 0.3mm, which is far better than the comparison model 3 (0.8mm): because the localization head of MM-CAFNet is trained based on cross-attention fusion features, it has stronger localization robustness than "spoofing features" (the sensitivity to temperature gradient is improved by 2 times). Compared to models 1 and 2, which lack positioning functionality and require manual assistance, this model achieves "end-to-end positioning," eliminating the additional error of 0.8~1.5mm caused by manual judgment.

[0046] In this embodiment, the advantages in generalization and false alarm rate are as follows: The accuracy of this model fluctuated by only 0.4% (97.8%~98.2%) under loads of 300~1000MW, compared to 4.4% (75.1%~79.5%) for Model 1, because multimodal fusion reduced the dependence on a single operating condition; The false alarm rate of this model is 1.2%, which is significantly lower than that of the comparison models 1 (6.8%), 2 (5.3%), and 3 (3.5%). This is due to the combined effect of the confidence threshold (0.9) and the high discriminativeness of the cross-attention feature, which reduces the situation of "misjudging interference as leakage".

[0047] In the preferred solution, based on its effectiveness in industrial scenarios, the core model is deployed and tested in the field, and the results are verified as follows: Field verification was conducted on the boilers of 300MW~1000MW units at Hubei Energy Group Ezhou Power Generation Co., Ltd., simulating three typical leakage scenarios to verify the industrial applicability of the core model MM-CAFNet: In this embodiment, the verification scenarios covering common leakage types in actual operation are shown in Table 3 below: Table 3

[0048] In this embodiment, the verification process and method are as follows: Data acquisition: Simultaneously deploy an FLIRA655sc infrared camera (3m from the furnace tube) and a PCB356A16 acoustic sensor (1m from the furnace tube) to acquire infrared images and acoustic signals in various scenarios. Each data acquisition session lasts 5 seconds. Model inference: Input the collected data into MM-CAFNet and compare it with Model 1 (VGG16), and output the predicted coordinates of the leakage point. ; Truth Value Acquisition: After shutdown, the inner wall of the furnace tube was observed using an Olympus IPLEXNX industrial endoscope (1920×1080 resolution, 10m detection depth) to determine the coordinates of the actual leak point. The calibration was performed in conjunction with the boiler CAD drawings, achieving an accuracy of ±0.05mm. Error calculation: Calculate positioning deviation To evaluate the accuracy of the model.

[0049] In this embodiment, the verification results are shown in Table 4 below: Table 4: Field Verification Positioning Error Table (Unit: mm)

[0050] In this embodiment, the verification conclusion is as follows: The model has an average positioning error of 0.32mm in three scenarios, and for targets ≤0.5mm, it fully meets the sub-millimeter positioning requirements. Under conditions of high airflow noise (Scenario 1), smoke and dust (Scenario 2), and maintenance interference (Scenario 3), the error is consistently <0.4mm, while the comparative model 1 has an error exceeding 1mm due to its single-mode limitation. The number of false alarms is only 1 per 100 tests, with a false alarm rate of 1%, which meets the "low false alarm" operation and maintenance requirements of industrial sites (avoiding frequent downtime for maintenance) and can be directly applied.

[0051] Example 2 The intelligent location method for boiler tube leakage provided by the present invention is described below. The intelligent location method for boiler tube leakage described below can be referred to in correspondence with the pre-trained model above, and is further explained in conjunction with Example 1. like Figures 1 to 4 The structure shown. Figure 1 The intelligent location method for boiler tube leakage provided in this application includes the following steps: In this embodiment, S1, multi-source data acquisition: provides accurate and synchronous basic data for subsequent dynamic noise suppression, zone scheduling, path planning, and multimodal positioning. It performs the acquisition of boiler operating parameters, acoustic signals, airflow parameters, and environmental parameters. The acquisition methods and functions of each parameter are as follows: Boiler operating parameter acquisition: including real-time load With furnace tube surface temperature Among them, real-time load Data is collected via the boiler DCS (Distributed Control System), using 4-20mA analog signal transmission, with a measurement range of 0-120% rated load (e.g., 0-360MW for a 300MW unit), and a sampling frequency of 1Hz. This data is used in subsequent steps S2 to construct an airflow fluctuation model (load affects airflow noise intensity) and S3 to calculate drone priority (higher load areas have a higher risk of leakage); furnace tube surface temperature. Temperature is collected by armored thermocouples (such as K-type thermocouples, with a temperature range of 0~1200℃ and an accuracy of ±1℃). One thermocouple is arranged every 5m along the length of the furnace tube and embedded inside the insulation layer of the furnace tube close to the tube wall. This is used to construct the repulsive force of the structural field in the subsequent S4 process (high-temperature furnace tubes require stronger obstacle avoidance force).

[0052] Acoustic signal acquisition: Acquiring the timing signal of sound pressure level in the furnace tube area. An array of acoustic sensors (such as the PCB356A16 piezoelectric sensor, with a measurement range of 40~180dB and a frequency response of 2Hz~20kHz) is used. The sensor density is one per 2~3m², covering all furnace tube areas such as water-cooled walls and superheaters. The sampling frequency is 1~2kHz (matching the leakage sound frequency of 500~1000Hz and the airflow noise frequency of 200~1500Hz), and the sampling duration is 3~8s (3s for emergency scenarios and 5~8s for normal scenarios). These are used in subsequent S2 to extract effective leakage sound signals (distinguishing between leakage sound and airflow noise).

[0053] Airflow parameter acquisition: Acquire airflow velocity around the furnace tubes. A miniature hot-wire anemometer (such as the TSI9545 model, measuring range 0~30m / s, accuracy ±0.1m / s) is used, installed on the support structure between the furnace tubes, 0.5~1m from the tube wall (to avoid the high temperature of the tube wall affecting the sensor), with a sampling frequency of... Consistent (1~2kHz), used for subsequent S2 construction of airflow fluctuation model, that is, airflow velocity directly affects airflow noise intensity, and the filter threshold needs to be adjusted in conjunction with the load.

[0054] Environmental parameter acquisition: Collect location information from at least two UAV nodes. Remaining battery power and signal strength Location information The single-point positioning accuracy is achieved by using a BeiDou + GPS dual-mode positioning module (such as the CoreStar UM220IVN) mounted on the drone. Data is collected at 1m, and three-dimensional coordinates (x, y, z) are output in meters; remaining battery power. Data is collected via a battery level detection chip (such as MAX17048, measurement range 0~100%, accuracy ±1%) from the drone's battery management system (BMS); signal strength is also measured. The RSSI (Received Signal Strength Indicator) value detected by the drone's wireless communication module (such as a 433MHz RF module), in dBm, ranges from 120 to 30 dBm; environmental parameters are used in subsequent S3 selection of the optimal drone node (nodes with close proximity, sufficient power, and strong signal are given priority).

[0055] After the data acquisition is completed, all parameters are synchronously transmitted to the central control system via industrial Ethernet (such as the Profinet protocol), and the boiler operating parameters, acoustic signals, airflow parameters and environmental parameters are output as the input basis for steps S2 to S5.

[0056] In a preferred embodiment, the array-type acoustic sensor is installed in step S1 as follows: it is attached to the furnace tube support steel structure using a magnetic mounting base (high temperature resistant to 200℃), with the sensor axis perpendicular to the furnace tube axis (to ensure maximum reception of leakage sound), and the spacing error between adjacent sensors is... 50mm (to ensure uniform coverage); the insertion depth of the armored thermocouple is 1 / 2 of the furnace tube wall thickness (e.g., if the furnace tube wall thickness is 10mm, the insertion depth is 5mm), to avoid damaging the furnace tube if inserted too deeply or causing inaccurate temperature measurement if inserted too shallowly.

[0057] In this embodiment, as shown Figure 2 As shown, S2, Dynamic Noise Suppression: To accurately extract effective leakage sound from the acoustic signal containing airflow noise and ensure accurate leakage area division in subsequent zone scheduling, based on the real-time load output from S1... airflow speed and sound pressure level timing signal An airflow fluctuation model was constructed and improved wavelet threshold filtering was performed to finally output an effective leakage sound signal. Specifically, it includes: S21. Wavelet Decomposition: To decompose... The system filters out airflow noise at different frequency bands based on frequency stratification (e.g., low-frequency airflow noise under low load, high-frequency airflow noise under high load), according to the real-time load output from S1. right implement Layer wavelet decomposition ( , The number of decomposition layers, (As coefficient index), the wavelet coefficients of each layer are obtained. Number of decomposition layers The adaptive adjustment rule is: when and To preset a high load threshold, such as 80% of the rated load, This means that high-frequency airflow noise accounts for a high proportion under high load, and more layers of decomposition and separation of 1000~1500Hz noise are needed; when and To preset a low load threshold, such as 30% of the rated load, That is, the noise frequency distribution is uniform under medium load, and the 4-layer decomposition balances the effect and efficiency. when hour, Furthermore, under low load conditions, the noise is mainly low-frequency airflow noise of 200~500Hz, and a 3-layer decomposition is sufficient to meet the requirements; among them, the wavelet basis function is selected as the db4 wavelet, which has good time-frequency locality and is suitable for non-stationary leakage sound signals.

[0058] In a preferred embodiment, the signal length N of the wavelet decomposition in S21 satisfies (like hour, This ensures no data truncation during the inverse wavelet transform; the selection criterion for the db4 wavelet is: by comparing the reconstruction errors of the db2, db4, and db6 wavelets, the db4 wavelet has the smallest reconstruction error for the leakage sound signal. (3%), adapted to non-stationary leakage sound signal characteristics.

[0059] S22. Airflow Association Threshold Calculation: To make the filter threshold follow... and Dynamically changing and precisely matching the airflow noise intensity under different operating conditions, based on the S1 output. and Constructing an airflow fluctuation model (model core: airflow noise intensity and...) , (positive correlation), calculate the first Filtering threshold of layer wavelet coefficients The specific steps are as follows: First, based on the wavelet coefficients output by S21 Using the relationship between the median and standard deviation of Gaussian noise, through the formula " "Calculate the first Layer noise standard deviation ,in For the first The method uses the median of the absolute values ​​of the layer wavelet coefficients to robustly avoid the impact of outliers on noise estimation. Secondly, determine the basic threshold term " ( for The signal length, (This is a general wavelet threshold correction term, ensuring coverage of over 99% of the noise figure). Finally, an airflow fluctuation correction term is introduced. ",in The rated load of the boiler (e.g., 300MW). The rated airflow velocity around the furnace tube (e.g., 5 m / s). Load influence factor (range of values) Based on experimental data: for every 100MW change in load, the airflow noise intensity changes by 2%~5%. Airflow influence coefficient (range of values) Based on experimental data: for every 1 m / s change in airflow velocity, the airflow noise intensity changes by 1% to 3%); Finally, through the formula " "Calculation yielded the first Layer threshold This threshold can vary. Enlarge Increase the size to ensure that effective leakage noise is not filtered out under high load and high airflow velocity, and that no excess noise is retained under low load and low airflow velocity.

[0060] S23. Coefficient Thresholding: To retain the wavelet coefficients corresponding to the valid leakage signal and suppress the coefficients corresponding to airflow noise, the wavelet coefficients of each layer output from S21 are processed. Perform soft thresholding when When the effective leakage signal coefficient is determined, it is calculated using the formula " "Correction coefficient (to eliminate signal abrupt changes near the threshold); when..." When, it is determined as the airflow noise figure, set This achieves the initial separation of leakage signals from airflow noise.

[0061] S24. Signal Reconstruction: To restore the processed wavelet coefficients to a time-domain acoustic signal for easier subsequent analysis of the leakage area, the wavelet coefficients processed in S23 are reconstructed. Perform an inverse wavelet transform (reconstruction) operation to output an acoustic signal containing only the valid leakage information. The signal frequency is concentrated in the range of 500~1000Hz, which matches the frequency of the leakage sound. This signal will serve as the core basis for dividing the leakage area in step S3.

[0062] In this embodiment, S3, partitioned response scheduling: To avoid response lag in single-drone inspections, the optimal drone is quickly scheduled to accurately cover the leak area, based on the drone node positions output by S1. Remaining battery power Signal strength and S2 output The process involves dividing the leak response area into partitions, calculating the priority of each node, and finally selecting the optimal drone node. Specifically, this includes: S31. Leakage Area Zoning: To break down a large furnace tube monitoring area (e.g., 1000~5000m²) into smaller areas (each zone 50~100m²), reducing the drone's flight distance and response time, based on the output of S2... The energy distribution, i.e., the energy value The energy concentration area corresponds to the potential leakage area. The K-means clustering algorithm is used to divide the furnace tube monitoring area into at least two response partitions. During the clustering process, with The energy value is the clustering feature, and the number of clusters. According to energy entropy Determined, i.e., energy entropy , For partitioning Total sound energy: when ( When energy is dispersed and there are many potential leakage areas, ; when ( When energy is moderately concentrated, ; when When energy is concentrated and the leakage area is clearly defined, ; After partitioning, calculate each partition. Geometric center coordinates As the target flight point for the UAV, the partition information is output to S32.

[0063] In a preferred embodiment, the initial cluster center selection method of the K-means clustering algorithm in S31 is: random selection. indivual The initial center is selected based on the position of the M largest energy value, ensuring it is close to potential leakage areas. The clustering iteration stops when the position of the cluster center changes. 10mm or number of iterations 20 times to avoid cluster oscillations.

[0064] S32. Multi-objective priority calculation: To comprehensively evaluate the UAV node pairs in the partition. To assess adaptability, the optimal node is selected based on its proximity, sufficient battery power, strong signal strength, and load compatibility, according to the output of S1. , , and Calculate each partition Internal drone node priority The specific calculation logic is as follows: Distance term ": For nodes arrive The three-dimensional distance, the closer the distance, the faster the response, weight (Distance is a key factor affecting response time) Battery level item ": This represents the maximum battery capacity of the drone; the more battery power, the better it can complete the positioning task. (Weight) (Take 0.1 for emergency scenarios and 0.2 for routine scenarios); Signal strength item ": The maximum signal strength (ideal strength for wireless communication) is the maximum signal strength; the stronger the signal, the more stable the data transmission. (Weight) ; Load adaptation item ": The rated load is the maximum load. The higher the load, the higher the risk of leakage. Nodes need to be prioritized for scheduling, with appropriate weights. (Take 0.3 for high-load scenarios and 0.2 for low-load scenarios); Finally, through the formula " "Calculate priority, where" This ensures a balanced weighting of all factors.

[0065] S33. Optimal Node Selection: To determine the unique UAV that will perform the localization task, for each partition... Filter the priority calculated by S32 The drone node corresponding to the maximum value is selected as the optimal execution node; if multiple nodes have the same priority (difference), the node with the highest priority is selected. If the value is 0.01, then further compare the node with... signal strength ,choose Higher-order nodes (to ensure stable data transmission); after determining the optimal node, output the position of that node. (As the starting point of the flight) and positioning mission instructions (including) (Coordinates, flight speed limits, etc.) provide start and destination point information for the S4 step of flight path planning.

[0066] In this embodiment, as shown Figure 3 As shown, S4, dual-field coupled path planning: To ensure that the UAV, while flying towards the leak area, neither touches the high-temperature furnace tube (to avoid equipment damage) nor takes a detour (to ensure response time), the optimal node position output by S3 is used. Partition Center and the surface temperature of the furnace tube output by S1 A dual-field coupling model of the leakage field and structural field is constructed to plan the optimal flight path, specifically including: S41. Dual-Field Model Construction: To simultaneously achieve the two objectives of "approaching the leakage area" and "avoiding the high-temperature furnace tubes," the attractive force of the leakage field and the repulsive force of the structural field are defined, as follows: Leakage field attraction Used to guide drones to output to S31 Proximity, the magnitude of attraction maximum value Positive correlation, that is The larger the size, the more severe the leak, requiring stronger attraction for guidance, and to be guided by drones. distance The square of the negative correlation means that the closer the distance, the stronger the attraction. To avoid flying over the target point, the formula " "calculate; in, The preset attraction coefficient is set to 0.3~0.5 N·m² / dB, with 0.5 for high leakage scenarios and 0.3 for low leakage scenarios. To point from the drone's current location The unit vector ensures the direction of attraction is correct.

[0067] Structural field repulsion This is used to prevent drones from touching the high-temperature furnace tubes; the magnitude of the repulsive force is related to the output of S1. Positive correlation, that is The higher the temperature, the higher the furnace tube temperature, requiring stronger repulsion force, which is related to the distance between the drone and the furnace tube structure. The cubic negative correlation means that the closer the distance, the greater the repulsive force, thus avoiding collisions, as shown by the formula " "calculate; in, The preset repulsion coefficient is used. Furnace tube surface temperature output with S1 Dynamic adjustment: when ( (High-temperature furnace tubes) ; when ( (Medium-temperature furnace tubes) ; when (When using low-temperature furnace tubes) ; in, , This refers to the rated operating temperature of the furnace tube.

[0068] This is a unit vector pointing from the furnace tube to the current position of the drone, ensuring that the direction of the repulsive force is away from the furnace tube.

[0069] In a preferred embodiment, the distance from the drone to the furnace tube structure in S41 The calculation method is as follows: the shortest distance between the drone and the furnace tube is measured using a lidar mounted on the drone (such as the DJI Livox Mid-40, ranging range 0.1~40m, accuracy ±2cm), and this distance is used as the basis for the calculation. The input values ​​ensure accurate distance measurement.

[0070] S42. Resultant Force Calculation and Velocity Update: To determine the real-time flight speed of the UAV and ensure that speed changes conform to the laws of mechanics, i.e., to avoid flight instability caused by rapid acceleration and deceleration, the total force acting on the UAV is calculated based on the dual-field model constructed in S41. (Vector sum); Then according to Newton's second law Combined with the preset quality of the drone (e.g., 0.5kg, standard weight for a small drone) and preset control cycle (e.g., 0.1s, balancing real-time performance and stability), through the formula " "Update the drone's flight speed;" Among them, the initial velocity (The drone starts from a standstill) After each speed update, it is necessary to check whether the speed direction is pointing in the correct direction. ,deviation 5°, otherwise fine-tune. Directional coefficient.

[0071] S43. Path Correction and Output: To prevent the drone from losing control due to excessive speed (especially in densely populated pipeline sections), the flight speed updated in S42 is validated for reasonableness. If the speed... ,in The upper limit of safe flight speed for small drones is then used to truncate the execution speed: Keeping the direction of velocity unchanged, adjust the magnitude of velocity to The formula is " "; if the speed magnitude If the speed remains constant, then the speed is constant. Preset the mass of the drone. For preset control cycle; Based on the corrected velocity, the real-time position of the drone is calculated. Generate flight path point sequence ,in For the number of waypoints, ensure and distance 0.1m; Output Path to S5 to provide path constraints for the collection of multimodal positioning data in step S5. The UAV flies along the Path to ensure that the collected data covers the leaked area.

[0072] In this embodiment, as shown Figure 4 As shown, S5, multimodal spatiotemporal calibration and positioning: To eliminate the influence of high temperature inside the boiler (causing camera parameter drift), electromagnetic interference (causing BeiDou positioning deviation), and UAV attitude changes (causing image coordinate shift) on positioning, and to achieve sub-millimeter level leak point positioning, based on the path point sequence Path output by S4, infrared thermal imaging, BeiDou positioning, and IMU attitude data are fused, and the precise coordinates of the leak point are output through a spatiotemporal calibration algorithm, specifically including: S51. Multimodal Data Acquisition: To obtain the "image position and pose" 3D data of the leak point, on the path planned in S4, when the UAV flies to each path point... Simultaneously, the following data collection will be performed: Infrared thermal imaging data acquisition: An industrial-grade infrared camera (such as FLIRA655sc, specific parameters: resolution 1280×960 pixels, frame rate 30fps, temperature measurement range 20~1000℃, temperature measurement accuracy ±2% or ±2℃ (whichever is greater), lens focal length 25mm fixed focus (to avoid distortion and ensure accurate mapping between image coordinates and spatial coordinates)) is used to acquire thermal images of the furnace tubes. Because leakage of the working fluid at the leak point will cause local temperature anomalies (50-200°C higher than the normal pipe wall temperature), a threshold segmentation algorithm (such as the Otsu algorithm) is used to segment the fluid at the leak point. Extracting the image coordinates of the leak point ( Horizontal pixel coordinates (Vertical pixel coordinates), output Up to S52.

[0073] BeiDou positioning data acquisition: High-precision BeiDou modules are used (such as the Hexin Xingtong UM220IVN, specific parameters: working frequency band B1I / B2I (compatible with BeiDou-2 / 3 system), single-point positioning accuracy). 1m (non-differential), differential positioning accuracy 0.1m (with differential), data update rate 1Hz, anti-interference measures: built-in narrowband filtering, frequency hopping technology (to resist electromagnetic interference inside the boiler)), collecting the real-time position of the drone at the current path point. (A spatial rectangular coordinate system, with the center point of the boiler's zero-meter level as the origin), output Up to S52.

[0074] IMU attitude data acquisition: Employs a high-precision inertial measurement unit, such as the ADIADIS16488. Specific parameters: Measurement range: roll / pitch ±45°, yaw ±180°, zero-bias stability. 0.1° / h (to reduce attitude drift during long-duration flights), 100Hz data output rate (higher than the UAV control cycle to ensure real-time attitude data), acquiring the UAV's real-time attitude angles at the current path point. ,in For roll angle, The pitch angle, (yaw angle), output Up to S52.

[0075] After data collection is complete, the data will be transmitted via the drone's internal bus (such as the SPI protocol). , , Synchronously transmit to the data processing unit to ensure that the timestamps of the three are not misaligned. 10ms (to avoid positioning errors caused by time and space asynchrony).

[0076] In a preferred embodiment, the lens of the infrared camera in S51 is selected based on the following criteria: the furnace tube monitoring distance is 1~5m (the safe distance between the drone and the furnace tube), the 25mm focal length lens has a field of view of 24°×18°, which can cover a furnace tube area with a diameter of 0.5~2m (without frequent camera angle adjustments), and the distortion rate is low. 1% (to ensure that the image coordinates are not significantly distorted).

[0077] In a preferred embodiment, the Kalman filter process noise in S52 With observation noise The covariance matrix is ​​set as follows: (Process noise covariance, reflecting the slow change in IMU drift). (Observe the noise covariance, which reflects the error characteristics of BeiDou positioning) to ensure fast filtering convergence speed and high stability.

[0078] S52. Spatiotemporal calibration model construction: To establish the image coordinates of the leak point Spatial coordinates The precise mapping relationship eliminates attitude drift and positioning deviation, and the calibration is performed in three steps: Camera extrinsic parameter matrix construction: The extrinsic parameter matrix [R|t] is used to describe the transformation from the UAV coordinate system (with the IMU as the origin) to the world coordinate system (consistent with the BeiDou positioning coordinate system), where For the rotation matrix (attitude angles output by S51) calculate), The translation vector is output by S51. Directly used as a translation vector, since the camera and BeiDou module are fixed in their relative positions; rotation matrix The calculation formula is: ; in, , , These are the roll, pitch, and yaw angles output by the S51, ensuring that the rotation matrix accurately reflects the UAV's attitude.

[0079] Introduction to camera intrinsic parameter matrix: Intrinsic parameter matrix These are inherent parameters of the camera, determined by factory calibration, and the formula is: ; in, Pixels, calculated based on an infrared camera with a focal length of 25mm and a pixel size of 12μm: The actual calibration value is 1200. (No pixel distortion) Pixels Pixel, image center, i.e., the midpoint of a resolution of 1280×960; intrinsic parameter matrix is ​​used to project three-dimensional points in the camera coordinate system into two-dimensional image points.

[0080] Spatial coordinate mapping and drift compensation: Based on the pinhole camera model, the spatial coordinates of the leak point... With image coordinates The mapping relationship is as follows: ; in, for coordinates for The coordinates; Because IMUs can drift during long-term operation (due to zero-bias stability), IMU drift compensation is introduced. ; State equations estimated using Kalman filtering. Observation equation , For process noise, To observe noise; right Perform calibration to obtain calibrated coordinates. ,in The compensation coefficient; Based on experimental data: 80% of IMU drift is systematic drift, which can be eliminated through compensation.

[0081] S53. Accuracy Optimization and Output: To further reduce reprojection errors, i.e. and The coordinate difference projected onto the image ensures sub-millimeter positioning accuracy, and the Levenberg-Marquardt nonlinear optimization algorithm is used to minimize the reprojection error. The error formula is:

[0082] in, The number of path points (e.g., 10 path points, with 1 set of data collected for each path point). For the Euclidean norm; through iterative optimization (number of iterations) 50 iterations, error convergence threshold ),Adjustment The coordinates of the leak point are determined to ensure the error meets the preset accuracy requirements; after optimization, the final coordinates of the leak point are output. Positioning accuracy 0.5mm.

[0083] S6. Feedback Optimization: To continuously improve the accuracy and efficiency of subsequent positioning tasks, a closed-loop optimization mechanism is established based on the leak point coordinates output by S5. Positioning deviation : , The coordinates of the actual leak point will be confirmed during shutdown and maintenance. and S2 output This reflects the noise suppression effect: signal-to-noise ratio 20dB indicates excellent suppression effect; otherwise, it is considered poor. S22 should be dynamically adjusted. , and S41 , The specific adjustment rules are as follows: when ( (high deviation) and Signal-to-noise ratio <20dB (poor noise suppression): Increase to (Enhancing the attractiveness of the leakage field and improving path guidance accuracy), reducing to , to (Reduce the excessive influence of airflow fluctuations on the threshold and improve noise suppression effect); when ( (mean deviation) and Signal-to-noise ratio 20dB (Excellent noise suppression): Fine-tuning to (Slightly enhance repulsion force, optimize obstacle avoidance path, and reduce positioning deviation caused by detours); when (Low deviation) and Signal-to-noise ratio 20dB (Excellent noise suppression): Keep all parameters constant; The adjusted parameters are stored in the system parameter library and used as the initial parameters for the next positioning task, thereby achieving continuous optimization of algorithm performance.

[0084] In a preferred embodiment, the actual leak point coordinates in S6 The acquisition method is as follows: After shutdown, use an industrial endoscope (such as the Olympus IPLEX NX, resolution 1920×1080, detection depth) to obtain the data. Observe the inner wall of the furnace tube (10m) to determine the location of the leak point; combine the three-dimensional design drawings of the boiler (CAD model, accuracy ±1mm) to calculate the world coordinate system coordinates of the actual leak point.

[0085] In a preferred embodiment, a device self-test step is also included: before data acquisition in S1, self-tests are performed on the acoustic sensor, infrared camera, BeiDou module, and IMU: the acoustic sensor acquires a standard acoustic signal (such as a 1kHz, 100dB sine wave signal) and verifies the signal-to-noise ratio of the output signal. 30dB; infrared camera captures a standard blackbody (temperature 500℃) to verify temperature measurement error. 2℃; Number of satellites received by the Beidou module 8 IMUs (to ensure effective positioning); IMU idle for 10 minutes to verify zero-bias stability. 0.1° / h; if the self-test fails, an alarm signal is issued and the backup device is switched to avoid positioning errors caused by equipment failure.

[0086] In a preferred embodiment, the total response time of S3 to S5 10s (emergency scenario) 8s), specifically the time allocation is: S3 partition scheduling 2s, S4 path planning 3s, S5 data acquisition and positioning 5s; by optimizing algorithm complexity (such as the number of wavelet decomposition layers) 5 layers, number of clustering iterations 20 times), improve hardware processing speed (use FPGA chip to accelerate signal processing) to achieve time constraints.

[0087] To make the technical principles, steps, logic, and effects of this technical solution clearer, the following describes embodiments 1 and 2 in detail with reference to specific application scenarios.

[0088] Examples 3, 4, and 5 are all based on boiler tube scenarios for thermal power generating units ranging from 300MW to 1000MW, covering three typical operating conditions: high load, low load dense area, and emergency maintenance. The parameter settings of each example are deeply adapted to the scenario requirements to ensure that the solution can be implemented.

[0089] Example 3 In this embodiment, the scenario of locating a leak in the boiler tubes of a high-load boiler, i.e., P=1000MW, is applied. =550℃, pipe spacing =1m; In this embodiment, the algorithm parameters are set and the reasons are as follows: S1: Sampling frequency = 1kHz, sampling duration Sensor density = 1 per 2m², high load airflow noise fluctuation frequency (500~1000Hz), short sampling interval can capture instantaneous leakage sound; S21: Number of decomposition levels For P=1000MW, multiple layers are needed to decompose and separate high-frequency airflow noise (1000Hz) and leakage noise (500~800Hz). S22: Load Influence Coefficient = 0.002. Based on experimental data, for every 100MW deviation of P from the rated value, the noise threshold needs to be adjusted by 2% to ensure that the threshold is adapted to high load noise. S32: Weighting coefficient , High-load areas have a high risk of leakage, so nodes that are close to the load and suitable for the load should be prioritized for allocation; S41: Attraction Coefficient Repulsion coefficient ,Right now =550℃>500℃, increase the repulsion coefficient to prevent drones from getting close to the high-temperature furnace tubes; S51: Infrared camera frame rate = 30fps, temperature measurement range 20~1000℃, high load leakage point temperature difference >80℃, high frame rate can capture dynamic temperature changes.

[0090] In this embodiment, the execution steps are demonstrated as follows: Input data: , , =[118dB,119dB,115dB,120dB,116dB] (5s sampled data), UAV node , , ; , , ; S1: Collect the above parameters and output the load. Sound pressure level sequence Node data; S21: Yes Perform a 5-level wavelet decomposition to obtain the wavelet coefficients of the 3rd level. =[12,15,8,18,10]; S22: Calculation ,Right now: ; S23: All <64.3, therefore Other layer coefficients are processed according to the same rules; S24: Reconstructed (Extract leakage sound and filter out 118~120dB airflow noise); S31: Calculate energy entropy (2) H 5) Partitioning Leakage partition center ; S32: Calculation ,Right now: ; ; ; S33: Select The largest node is 1, output ; S41: , , ; , , ; S42: , speed magnitude No need to truncate; S43: Output path ; S51: Infrared Image extract Beidou IMU posture ; S52: External parameter matrix Internal Reference , mapping IMU drift , (The same applies to Y and Z); S53: Reprojection error Output ; S6: Positioning Deviation Keep the parameters unchanged.

[0091] In this embodiment, the selection criteria are as follows: The parameters in this embodiment are designed for high-load requirements of "high noise and high temperature": wavelet 5-layer decomposition effectively separates 120dB airflow noise and 115dB leakage sound, dual-field coupling model avoids the UAV from getting close to the 550℃ high-temperature furnace tube, and multi-modal calibration achieves a positioning accuracy of 0.8mm, solving the problems of 25% false alarm rate and 3mm positioning deviation in existing technologies under high load.

[0092] Example 4 In this embodiment, it is applied to a scenario of positioning in a densely populated area of ​​furnace tubes under low load, i.e., P=300MW. =300℃, pipe spacing =0.3m; In this embodiment, the algorithm parameters are set and the reasons are as follows: S1: Sampling frequency = 500Hz, sampling duration It is stable under low load noise (80~90dB) and the signal-to-noise ratio is improved by extending the sampling time; S21: Number of decomposition levels P=300MW, 3-layer decomposition can separate low-frequency leakage sound (200~500Hz). S22: Load impact coefficient = 0.001. Low load noise fluctuation is small, and the coefficient is lowered to avoid over-filtering. S32: Weighting coefficient , In densely populated areas, signals attenuate quickly (due to pipe obstruction), so prioritize nodes with high signal strength. S41: Attraction Coefficient Repulsion coefficient ,Right now =300℃<500℃, reducing the repulsion coefficient to adapt to dense pipelines (spacing 0.3m); S51: Infrared camera resolution = 1280×960, temperature measurement accuracy ±0.5℃, small leakage point in dense areas (<0.5mm), high resolution ensures capture of minute temperature anomalies.

[0093] In this embodiment, the execution steps are demonstrated as follows: Input data: , , =[85dB,86dB,83dB,88dB,84dB], drone node , , ; , , ; S1: Output , Node data; S21: 3-level wavelet decomposition, coefficients of the second level =[8,10,6,12,7]; S22: ,Right now: ; S23: , , reconstructed ; S31: Energy Entropy partition Leakage partition center ; S32: ,Right now: ; ; ; Preferred Higher node 1; S41: , ; S42: ,speed Output path ; S51: , IMU ; S52: External parameter matrix , mapping ; S53: Error Output ; S6: Deviation Keep the parameters unchanged.

[0094] In this embodiment, the selection criteria are as follows: The parameters in this embodiment are designed for the needs of "pipeline obstruction and weak signal" in dense areas: 3-layer wavelet decomposition retains 84dB of leakage sound, dual-field coupling model avoids pipelines with a spacing of 0.3m, positioning accuracy is 0.5mm, solving the problems of 18% path jamming rate and 15% false alarm rate of existing technology in dense areas.

[0095] Example 5 In this embodiment, it is applied to an emergency leak location scenario before boiler shutdown, i.e., P=600MW, T_p=400℃, and the response requirements are... 8s; In this embodiment, the algorithm parameters are set and the reasons are as follows: S1: Sampling frequency = 2kHz, sampling duration Emergency scenarios require rapid response, and short sampling times compress the time. S21: Number of decomposition levels P=600MW (rated load), 4-layer decomposition balance speed and filtering effect; S22: Load influence coefficient = 0.0015, rated load noise is stable, coefficient adapts to moderate fluctuations; S32: Weighting coefficient , In emergency scenarios, priority is given to nodes that are closer in distance, and the weight of power consumption is reduced. S41: Attraction Coefficient Repulsion coefficient Balancing speed and safety; S51: Infrared camera frame rate = 60fps, Kalman filter period = 0.05s, quickly captures dynamic leak points.

[0096] In this embodiment, the execution steps are demonstrated as follows: Input data: , , =[100dB,102dB,98dB], drone node , , ; , , ; S1~S3 complete sampling and scheduling within 3 seconds, that is: Select node 1; S4: Generates a path within 2 seconds, speed ; S5: Complete positioning within 3 seconds and output... ; Total response time: 3+2+3=8s, which meets the emergency requirements.

[0097] In this embodiment, the selection criteria are as follows: The parameters of this embodiment are designed for the "fast response" requirement: short sampling time (3s), high frame rate camera (60fps) compresses the total response time to 8s, positioning accuracy is 0.6mm, solves the problem of response time >30s in the prior art, and is suitable for emergency leak investigation before furnace shutdown.

[0098] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact through communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other.

[0099] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0100] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of this application.

Claims

1. A method for intelligently locating leaks in boiler tubes, characterized in that, Includes the following steps: S1. Obtain boiler operating parameters, acoustic signals, airflow parameters, and UAV node environmental parameters based on the boiler operating status; S2. Construct an airflow fluctuation model based on boiler operating parameters, acoustic signals, and airflow parameters, and extract effective leakage sound signals through improved wavelet threshold filtering. ; The specific steps of S2 are as follows: S21. Based on real-time load Timing signal of sound pressure level in the furnace tube area implement Layer wavelet decomposition yields the wavelet coefficients of each layer. ,in The number of decomposition levels and , For coefficient index; Decomposition layer number Based on the real-time load output by S1 Adaptive adjustment; S22, Based on real-time load and the airflow velocity around the furnace tube Construct an airflow wave model based on wavelet coefficients. Calculate the first Filtering threshold of layer wavelet coefficients Filter threshold The calculation formula is: ； in, For the first Layer noise standard deviation, through wavelet coefficients The median value was calculated. for The signal length; This is the boiler's rated load. The rated airflow velocity around the furnace tube; This is the load influence factor; This is the airflow influence coefficient; S23. Wavelet coefficients of each layer Perform soft threshold processing; S24. Based on the wavelet coefficients processed in step S23 Perform wavelet reconstruction and output a valid leakage acoustic signal. ; S3. Based on the UAV node environmental parameters and effective leakage acoustic signals Divide the leakage response into partitions, calculate the scheduling priority of each node, filter the optimal UAV node, and select the optimal UAV node to perform the positioning task; S4. Based on the optimal UAV node and boiler operating parameters, construct a dual-field coupling model of leakage field and structural field, and plan the optimal flight path; S5. Based on the optimal flight path planned in step S4, and by fusing infrared thermal imaging data, BeiDou positioning data, and IMU attitude data, the final coordinates of the leak point are output using a spatiotemporal calibration algorithm. ; S6. Based on the final coordinates of the leak point Positioning deviation and effective leakage sound signal The steps S2 and S4 are dynamically adjusted to form a closed-loop optimization.

2. The intelligent location method for boiler tube leakage according to claim 1, characterized in that, In step S1, the boiler operating parameters include real-time load. Furnace tube surface temperature The acoustic signals include the timing signals of the sound pressure level in the furnace tube area. The airflow parameters include the airflow velocity around the furnace tubes. The environmental parameters of the drone nodes include the locations of at least two drone nodes. Remaining battery power and signal strength .

3. The intelligent location method for boiler tube leakage according to claim 2, characterized in that, Decomposition of layers in step S21 Based on the real-time load output by S1 The adaptive adjustment rule is: when and When the preset high load threshold is used, ; when and When the preset low load threshold is used, ; when hour, ; in .

4. The intelligent location method for boiler tube leakage according to claim 2, characterized in that, In step S3, the drone node position is output through S1. Remaining battery power Signal strength and S2 output The specific steps of step S3, which involves dividing the leakage response area, are as follows: S31. Based on the valid leakage sound signal The energy distribution divides the furnace tube monitoring area into at least [number missing]. One response partition Determine the center of each zone ; S32. Obtain each response partition Internal drone nodes Based on the location of the drone node Remaining battery power Signal strength and real-time load Calculate each response partition Internal drone node priority Priority The calculation formula is: ； in, For nodes To the partition center The three-dimensional distance; ~ Preset weighting coefficients; , , These are the drone's maximum battery power, maximum signal strength, and maximum boiler load, respectively. S33. Select the priority calculated in step S32. The drone node corresponding to the maximum value is selected as the optimal execution node, and the position of the optimal execution node is output. .

5. The intelligent location method for boiler tube leakage according to claim 4, characterized in that, In step S4, the optimal execution node position is determined. Partition Center and furnace tube surface temperature The specific steps for constructing the dual-field coupling model of the leakage field and the structural field in S4 are as follows: S41. Obtain a valid leakage sound signal maximum value Based on the valid leakage acoustic signal maximum value Define the attraction of the leakage field According to the surface temperature of the furnace tube Define the repulsive force of the structural field ; S42. Calculate the total force acting on the UAV based on the dual-field coupling model of leakage field and structural field. The drone's flight speed is updated using Newton's second law; S43. Perform a rationality check on the updated flight speed. If the speed exceeds the preset maximum speed limit, perform speed truncation and output the corrected flight path point sequence Path.

6. The intelligent location method for boiler tube leakage according to claim 5, characterized in that, In step S41, the leakage field attraction The calculation formula is: ; Among them, the attractive force of the leakage field Used to guide the drone to the output of step S31 near; To preset the attraction coefficient, For effective leakage acoustic signals The maximum value, For drones to distance, For pointing , unit vector; Structural field repulsion The calculation formula is: ; Among them, the repulsive force of the structural field Used to prevent drones from touching high-temperature furnace tubes; As a preset repulsion coefficient, The distance from the drone to the furnace tube structure. This is the unit vector that deviates from the furnace tube.

7. The intelligent location method for boiler tube leakage according to claim 5, characterized in that, The specific steps of step S5 are as follows: S51. Based on the corrected flight path point sequence Path, acquire thermal images of the furnace tubes using an infrared camera. Extract the image coordinates corresponding to the abnormal temperature at the leak point. The real-time location of the drone is collected via a BeiDou module. ; Real-time attitude angles of the UAV are collected via IMU. Real-time attitude angle Including roll, pitch, and yaw; S52, Based on the data collected in step S51 and Construct the camera extrinsic matrix [R|t], where For rotation matrix, The translation vector is combined with the camera's preset intrinsic parameter matrix. Establish the spatial coordinates of the leak point. Image coordinates corresponding to abnormal temperature at the leak point The mapping relationship; Introducing IMU drift compensation Spatial coordinates of the leak point Perform calibration to obtain calibrated coordinates. ; Among them, IMU drift compensation amount Estimated using the Kalman filter algorithm; S53, Minimize using a nonlinear optimization algorithm and The reprojection error is used to output the final coordinates of the leak point. .

8. The intelligent location method for boiler tube leakage according to claim 6, characterized in that, In step S6, dynamically adjust S22. , and S41 , The parameter adjustment rules are as follows: When positioning deviation and When setting a high deviation threshold, increase to ,in To preset the upward adjustment coefficient, reduce and to , ,in The preset adjustment coefficient; when and Fine-tuning when setting a low deviation threshold to ,in This is the preset fine-tuning coefficient; when At that time, keep all parameters constant.

Images