Flood discharge tunnel automatic inspection and erosion defect intelligent detection method

Abstract

The invention relates to the technical field of flood discharge tunnel detection, and discloses a flood discharge tunnel automatic inspection and erosion defect intelligent detection method. The method comprises the following steps: acquiring structural state data in the flood discharge tunnel and performing standardization processing; dividing standard classification data according to preset hydrological working condition intervals, and extracting multi-modal sensing features of each interval to generate a working condition feature set; monitoring the switching frequency of the hydrological working condition interval, and activating a dynamic analysis strategy to identify an abnormal working condition interval when the switching frequency exceeds a preset threshold value; performing feature decoupling on the working condition feature set of the non-abnormal working condition interval according to the matching degree of the historical erosion stage feature and the current multi-mode sensing feature, and separating out an erosion feature subset; on the basis of coupling relevance of mechanical parameters and hydraulic parameters of the tunnel structure, parameter collaborative correction is carried out on the erosion feature subset; and finally, reversely tracing a potential defect source region along a defect evolution path by utilizing the corrected erosion feature subset, determining a repair priority and outputting a maintenance instruction.

Description

Technical Field

[0001] The present invention relates to the technical field of flood discharge tunnel detection, and in particular to a method for automatic inspection and intelligent detection of erosion defects in flood discharge tunnels. Background Art

[0002] As a critical facility in a water conservancy project, flood discharge tunnels are subject to complex hydrological conditions such as high-speed water flow, fluctuating water pressure, and varying sediment concentrations. Their structural integrity is directly linked to the operational safety of the water conservancy hub. Over time, tunnels are prone to corrosion defects such as concrete spalling, steel corrosion, and crack expansion. If not discovered and addressed promptly, these defects can lead to serious consequences such as structural instability and a reduction in flood discharge capacity. Inspections of flood discharge tunnels often rely on manual labor, requiring personnel to enter the tunnels and collect data using visual observation and handheld measurement devices. This method is not only inefficient but also limited by environmental factors such as the narrow tunnel space, dim lighting, and high water noise. It also presents blind spots and safety hazards. Furthermore, manual inspections struggle to achieve continuous monitoring, limiting their ability to detect and track corrosion defects early and dynamically. With the advancement of sensing technology, some projects have begun using sensor arrays to monitor tunnel structural conditions. However, existing methods often analyze data under a single hydrological condition, ignoring the impact of variations in hydrological conditions (such as flow, water level, and velocity) on structural condition monitoring data. Tunnels experience varying loads and structural responses under different hydrological conditions. Lumping data from these different conditions can lead to inaccurate feature extraction, which in turn affects the reliability of defect detection. Existing defect detection methods often fail to effectively separate corrosion signatures from other interfering factors (such as temperature changes and structural vibrations) when processing multimodal sensor data. This leads to feature redundancy and increases the difficulty of defect identification. Furthermore, the ability to trace the evolution path of corrosion defects and potential defect sources is insufficient, making it difficult to formulate targeted maintenance strategies, thus hindering the scientific and timely nature of tunnel maintenance. Summary of the Invention

[0003] The object of the present invention is to provide a method for automatic inspection and intelligent detection of corrosion defects in flood discharge tunnels to solve the problems raised in the above background technology.

[0004] To achieve the above objectives, the present invention provides a method for automatic inspection and intelligent detection of corrosion defects in flood discharge tunnels, the method comprising: Collect structural status data inside the flood discharge tunnel and perform standardization processing; Classify the structural state data according to the preset hydrological operating condition interval division standard, extract the multimodal sensing features within each hydrological operating condition interval, and generate a set of operating condition features; Monitor the switching frequency of hydrological operating intervals. When the switching frequency exceeds a preset frequency threshold, activate a dynamic analysis strategy to identify abnormal operating intervals. Based on the matching degree between the historical corrosion stage characteristics and the current multimodal sensing characteristics, a feature decoupling operation is performed on the operating condition feature set of the non-abnormal operating condition interval to separate the corrosion feature subset; Based on the coupling correlation between tunnel structural mechanical parameters and hydraulic parameters, the parameters of the erosion feature subset are collaboratively corrected. The corrected corrosion feature subset is used to reversely trace the potential defect source area along the defect evolution path, determine the repair priority and output maintenance instructions.

[0005] Preferably, the collecting of structural status data inside the flood discharge tunnel and performing standardization processing includes: Synchronously acquire laser scanning data, acoustic detection data, and seepage pressure monitoring data, and perform time synchronization alignment operations on all data sources; Point cloud noise reduction is performed on the time-aligned laser scanning data, spectrum filtering is performed on the acoustic detection data, and an outlier removal algorithm is used on the seepage pressure monitoring data; The processed laser scanning data, acoustic detection data and seepage pressure monitoring data were resampled at a uniform spatial resolution to construct a standardized three-dimensional state data set.

[0006] Preferably, the structural state data is classified according to a preset hydrological operating condition interval division standard, multimodal sensing features within each hydrological operating condition interval are extracted, and a set of operating condition features is generated, including: The standardized three-dimensional state data set is divided into multiple hydrological operating ranges according to the water level change amplitude, including low water level operating range, medium water level operating range and high water level operating range; Extract the surface deformation characteristics of laser scanning data, the medium attenuation characteristics of acoustic detection data, and the water pressure gradient characteristics of seepage pressure monitoring data within each hydrological operating range respectively; The surface deformation characteristics, medium attenuation characteristics and water pressure gradient characteristics within the same hydrological operating condition interval are integrated according to the spatial grid to form a multidimensional operating condition feature set.

[0007] Preferably, the surface deformation characteristics of the laser scanning data include the curvature distribution matrix and the point cloud density gradient, the medium attenuation characteristics of the acoustic wave detection data include the acoustic wave propagation delay spectrum and the energy attenuation coefficient, and the water pressure gradient characteristics of the seepage pressure monitoring data include the pressure conduction rate and the pulsation frequency.

[0008] Preferably, the switching frequency of the monitoring hydrological operating condition interval is activated, when the switching frequency exceeds a preset frequency threshold, to identify the abnormal operating condition interval, including: Based on the sliding time window, the number of switching times of the hydrological operating interval is counted, the switching frequency per unit time is calculated, and it is determined whether the switching frequency exceeds the preset frequency threshold; If the switching frequency exceeds the preset frequency threshold, the multimodal sensing features of adjacent hydrological operating conditions are matched using a waveform alignment algorithm to extract cross-operating condition stability indicators. When the cross-operating condition stability index is lower than the stability threshold, the hydrological operating condition interval corresponding to the current time window is marked as an abnormal operating condition interval, and the re-extraction process of the multimodal sensing features of the interval is triggered.

[0009] Preferably, the extraction of the cross-operating-condition stability index includes: performing cross-correlation calculation on the acoustic wave propagation delay spectra of adjacent hydrological operating condition intervals, performing trend similarity calculation on the pressure conduction rate, and weightedly fusing the calculation results to generate the cross-operating-condition stability index.

[0010] Preferably, the feature decoupling operation is performed on the operating condition feature set of the non-abnormal operating condition interval based on the degree of matching between the historical erosion stage features and the current multimodal sensing features to separate the erosion feature subset, including: For the current multimodal sensing features in the working condition feature set of the non-abnormal working condition interval, calculate the similarity matrix between them and the historical erosion stage features; When the maximum value of the similarity matrix exceeds the preset similarity threshold, the frequency domain bandpass decomposition mode is used to extract the erosion feature subset from the working condition feature set; When the maximum value of the similarity matrix is ​​lower than the preset similarity threshold, based on the correlation between the acoustic wave energy attenuation coefficient and the structural resonance frequency, the frequency band where the characteristic energy is concentrated is selected for directional decomposition to generate a erosion feature subset.

[0011] Preferably, the collaborative parameter correction of the erosion feature subset based on the coupling correlation between the tunnel structure mechanical parameters and the hydraulic parameters includes: According to the nonlinear relationship between the elastic modulus of concrete and the surface curvature, deformation compensation correction is performed on the point cloud density gradient in the erosion feature subset. Based on the correlation between hydraulic load distribution and seepage pressure pulsation, dynamic pressure correction is performed on the water pressure gradient characteristics in the erosion feature subset. The corrected point cloud density gradient is fused with the water pressure gradient feature to generate a erosion feature subset with coordinated parameter correction.

[0012] Preferably, the method of using the corrected corrosion feature subset to reversely trace the potential defect source area along the defect evolution path, determine the repair priority and output the maintenance instruction includes: Load a pre-built defect evolution path map, which contains the defect transmission relationship and influence weight between structural units; Identify abnormal characteristic parameters from the erosion feature subset after parameter collaborative correction and match the corresponding nodes in the defect evolution path map; Traverse in reverse along the conduction direction of the defect evolution path to locate the potential defect source unit with abnormal characteristic parameters; Calculate the repair priority weight based on the number of abnormal characteristic parameters and the depth of the conduction path of the potential defect source unit; Based on the repair priority weight and defect development rate, a maintenance instruction containing the target unit coordinates and maintenance time is generated.

[0013] Preferably, the maintenance instructions are converted into a three-dimensional coordinate instruction set through geographic information coding and automatically output to the tunnel maintenance execution terminal.

[0014] Compared with the prior art, the present invention has the following beneficial effects: This method for automatic inspection and intelligent detection of corrosion defects in flood discharge tunnels eliminates the reliance on manual inspections through automated data collection and standardized processing. It can continuously and stably obtain structural status data in complex tunnel environments, reduce the interference of environmental factors on data collection, expand the monitoring range, eliminate detection blind spots, and improve the comprehensiveness and objectivity of data collection. Structural state data is classified according to pre-set hydrological operating condition interval criteria, and multimodal sensing features within each interval are extracted to generate a set of operating condition features, fully considering the impact of different hydrological conditions on the structural state. By correlating data with specific operating conditions, the extracted features are more closely aligned with the actual structural response, avoiding feature distortion caused by mixing data from different operating conditions and providing a more reliable foundation for subsequent defect detection. By monitoring the switching frequency of hydrological operating ranges and activating a dynamic analysis strategy to identify abnormal operating ranges when the switching frequency exceeds a preset threshold, the system can promptly capture abnormal operating condition changes. These abnormal operating conditions are often accompanied by sudden changes in structural stress. Through targeted analysis, potential structural risks can be warned in advance, eliminating interference for subsequent erosion feature extraction and ensuring the accuracy of structural status analysis under normal operating conditions. Based on the degree of match between historical corrosion stage characteristics and current multimodal sensing features, a feature decoupling operation is performed on the operating condition feature set of the non-abnormal operating range to separate the corrosion feature subset, effectively separating the corrosion features from other irrelevant features. This process reduces the interference of irrelevant information on defect detection, highlights the key characteristics of corrosion defects, makes subsequent defect analysis more targeted, and improves the accuracy of corrosion defect identification. Based on the coupled correlation between tunnel structural mechanical parameters and hydraulic parameters, a coordinated parameter correction was implemented for the corrosion feature subset, fully integrating the mechanical properties of the structure and the influence of hydraulic forces. This correction further optimized the corrosion feature subset to better reflect the true state of corrosion defects, reduce feature deviations caused by parameter mismatches, and enhance the scientific nature of corrosion defect detection. Using the modified corrosion feature subset to trace back potential defect source areas along the defect evolution path, we can deeply explore the root causes and development trends of corrosion defects. Based on this, we determine repair priorities and output maintenance instructions, enabling targeted maintenance work. We develop reasonable maintenance plans based on defect severity and potential risks, optimizing the allocation of maintenance resources and improving the pertinence and effectiveness of tunnel maintenance. This helps extend the service life of flood discharge tunnels and ensure the safe and stable operation of water conservancy hubs. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 This is a working principle diagram of the method for automatic inspection and intelligent detection of erosion defects of flood discharge tunnels according to the present invention; Figure 2 Flowchart for the collection and standardization of structural status data; Figure 3 Flowchart for hydrological condition interval classification and multimodal feature extraction; Figure 4 This is a flow chart of the dynamic analysis strategy for abnormal operating conditions. DETAILED DESCRIPTION

[0016] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0017] See also Figure 1 The present invention provides a method for automatic inspection and intelligent detection of corrosion defects in flood discharge tunnels, the method comprising: Laser scanning data, acoustic wave detection data, and seepage pressure monitoring data are collected from within the flood discharge tunnel. After time synchronization alignment, point cloud noise reduction, spectral filtering, and outlier removal, the data is resampled to a uniform spatial resolution to construct a standardized three-dimensional state dataset. Hydrological operating conditions are divided into intervals (low, medium, and high) based on preset water level fluctuations. Surface deformation characteristics of the laser data, dielectric attenuation characteristics of the acoustic data, and water pressure gradient characteristics of the seepage pressure data within each interval are extracted and integrated into a multidimensional operating condition feature set based on a spatial grid. The frequency of hydrological operating condition switching is monitored in real time. If the number of switching times per unit time exceeds a preset threshold, a dynamic analysis strategy is activated: a waveform alignment algorithm is used to match the multimodal features of adjacent operating conditions. A cross-condition stability index is calculated. If the interval falls below the threshold, it is marked as an abnormal operating condition and features are re-extracted. For the non-abnormal operating condition feature set, a similarity matrix is ​​calculated between the current multimodal features and the features of historical erosion stages. Based on the similarity threshold, either frequency domain bandpass decomposition or energy-concentrated frequency band directional decomposition is selected to isolate a subset of erosion features. Based on the nonlinear relationship between the concrete elastic modulus and surface curvature, the point cloud density gradient is corrected. The hydraulic pressure gradient characteristics are corrected by combining the correlation between hydraulic load and seepage pressure pulsation, generating a subset of corroded features with coordinated parameter correction. A pre-built defect evolution path map (including the conduction relationships and weights between structural units) is loaded. Anomalous parameters are identified from the corrected corroded features and matched to the map nodes. The conduction path is then traversed in reverse to locate potential defect source units. The repair priority weight is calculated based on the number of anomalous parameters and the path depth. Finally, instructions with 3D coordinates and maintenance time are generated and output to the execution terminal.

[0018] Example 1: See Figure 2 During the collection and standardization of data on the internal structural status of the flood discharge tunnel, three parallel monitoring systems were deployed to operate in coordination. The laser scanning system uses a rotating phase ranging laser array, which moves at a constant speed of 1 meter per second along the axis of the tunnel, emitting a 905-nanometer pulsed laser beam to scan the lining surface. No fewer than 2,000 original coordinate points are generated per square meter, containing three-dimensional position information and reflection intensity values. The acoustic wave detection system installs a piezoelectric broadband sensor array at the key sections of the tunnel. The sensors are arranged in an equilateral triangle topology with a spacing of 0.5 meters. A linear frequency-modulated pulse signal with a center frequency of 50kHz is excited in real time, and the receiving end collects the time domain waveform after the sound wave penetrates the lining layer. The seepage pressure monitoring system obtains pore water pressure data through a pre-buried vibrating wire sensor network. The sensors are embedded in the lining structure with a vertical spacing of 2 meters, and the sampling frequency is set to 100Hz.

[0019] All monitoring data is synchronized using a unified timing protocol. The laser scanner's built-in GNSS receiver outputs a PPS pulse signal as a reference clock source. The acoustic sensor array's FPGA controller calibrates its local clock using the IEEE 1588 protocol. The piezometer's RS485 bus uses time-division multiple access to adjust timestamps. All raw data packets are time-stamped with microsecond-level accuracy, ensuring time deviation within ±2 milliseconds.

[0020] Multi-stage noise reduction is performed on the raw laser scanning data: First, a statistical outlier filter is applied to calculate the average distance of the point cloud with a neighborhood radius of 0.8 meters, removing discrete noise points exceeding 2.5 times the standard deviation. A surface reconstruction algorithm based on moving least squares is then used to fit a local surface at a 1.2 mm scale to suppress fluctuation noise caused by surface water film scattering. Acoustic data processing utilizes a digital signal chain: the raw waveform is converted to the frequency domain using a 16-bit ADC and a 1024-point fast Fourier transform with a Blackman-Harris window function. The spectrum is processed using a cascaded filter bank, with a low-pass filter cutoff frequency set to 500 Hz to eliminate background water noise and a high-pass filter cutoff frequency set to 15 kHz to suppress electromagnetic interference. The final reconstructed time-domain signal retains only the characteristic frequency band from 0.5 kHz to 15 kHz. Dynamic threshold cleaning is implemented in the preprocessing of seepage pressure data: a sliding window with a length of 60 seconds is established, and the upper quartile Q3 and lower quartile Q1 of the pressure value within the window are calculated; when a data point exceeds the range of [Q1-1.5×(Q3-Q1),Q3+1.5×(Q3-Q1)], it is automatically replaced by the mean of the previous and next two points; the window parameters are updated in real time to cope with transient water hammer effects.

[0021] The spatial normalization process utilizes a unified grid space: the tunnel axis is defined as the Z-axis coordinate reference, and the radial section is divided into 360 sectors with a 1-degree angular resolution. The laser point cloud is resampled to a 0.5 cm × 0.5 cm planar grid, with four characteristic quantities calculated for each grid cell: mean reflection intensity, point cloud density, standard deviation of the normal vector angle, and median curvature. Acoustic data are spatially mapped: within a 1-meter axial segment, the acoustic delay spectrum eigenvalues ​​of the sensor nodes are mapped to a 100 × 100 grid node using the Kriging interpolation algorithm, generating a medium acoustic property field. Key parameters include the main frequency offset, the logarithm of the energy attenuation, and the waveform distortion coefficient. Pressure monitoring data are expanded from discrete measurement point pressure values ​​into a two-dimensional pressure gradient field with a 1-cm resolution using radial basis function interpolation. The derived vectors include the pressure conduction direction angle, gradient modulus, and pulsation spectrum entropy. The final standardized three-dimensional state dataset adopts a hierarchical data structure: the base layer stores geospatial coordinates (XYZ geodetic coordinates in the WGS84 coordinate system), the attribute layer contains geometric feature arrays (100×100×N grids), acoustic feature tensors (3×100×100 matrices), and hydraulic feature matrices (2×100×100). The data storage format uses HDF5 binary encapsulation.

[0022] The key parameter configurations implemented in this process include: a laser scanning frequency of 200Hz, an acoustic sampling rate of 1MHz, and a seepage pressure monitoring time constant of 0.1 seconds. All processing algorithms are executed in real time at the edge computing node, and the processing delay of single-section data does not exceed 50 milliseconds. The standardized result set ensures consistency through a verification mechanism: after each 1-meter axis segment data is generated, the spatial alignment error between the geometric layer and the acoustic layer is automatically detected. When the grid node position deviation exceeds 2 mm, the interpolation correction process is triggered until the 0.5 mm positioning accuracy requirement is met. The processed 3D status data set generates a version tag, which contains a metadata combination of the equipment serial number, time interval, and working condition label.

[0023] Example 2: See Figure 3 The division of hydrological operating conditions and the extraction of multimodal sensing features are based on a standardized three-dimensional state data set. Water level monitoring data serves as the primary basis for operating condition classification. A dual-redundant measurement system of float-type water level gauges and radar water level gauges is used to obtain the absolute height of the water surface in the flood discharge tunnel in real time. The ratio of the water level height to the tunnel diameter is the core parameter for operating condition classification, with three fixed intervals set: low water level conditions correspond to a ratio of 0% to 30%, medium water level conditions correspond to 30% to 70%, and high water level conditions correspond to 70% to 100%. Water level data is updated at a rate of 1 second, and a sliding average filter is used to eliminate instantaneous fluctuations caused by wave disturbances.

[0024] After the standardized three-dimensional state data set is input into the working condition classification module, the data is first time-segmented according to the water level label. The data subset within each working condition interval is processed independently to ensure the consistency of the working conditions of feature extraction. Under low water level conditions, the laser scanning data focuses on analyzing the micro-deformation characteristics of the lining surface. By calculating the local variation of the point cloud normal vector, a curvature distribution matrix is ​​generated to reflect the concave and convex change trend of the structural surface. At the same time, the point cloud density differences of adjacent grid cells are statistically analyzed to construct a density gradient field for identifying possible material loss or cavitation areas. Under low water level conditions, the acoustic wave detection data mainly focuses on the attenuation characteristics of the medium, extracts the time delay spectrum curve on the acoustic wave propagation path, analyzes the offset of the main frequency peak, and calculates the logarithm of the energy attenuation coefficient to quantify the change in acoustic impedance inside the concrete. The seepage pressure monitoring data focuses on analyzing the static water pressure distribution, calculating the pressure difference change rate between grid cells, deriving the pressure conduction rate, and extracting the instantaneous frequency characteristics of the pressure pulsation through Hilbert transform.

[0025] The feature extraction process for the mid-water level condition incorporates hydraulic load factors in addition to the low-water level. In addition to analyzing surface deformation, laser scanning data also calculates the rate of change in point cloud density along the direction of water flow to identify possible signs of hydraulic erosion. Acoustic data incorporates multipath propagation analysis, using time-of-day positioning technology from a sensor array to identify potential cracks or voids within the lining. Seepage pressure monitoring data under the mid-water level condition focuses on dynamic pressure fluctuations, extracting the spectral characteristics of pressure pulsation and analyzing their correlation with water flow rate.

[0026] Feature extraction for high-water-level conditions further enhances analysis of fluid-structure interaction effects. Laser scanning data utilizes a high-frequency sampling mode to capture the transient vibration response of the lining structure under high-pressure water flow. The acoustic detection system switches to a broadband excitation mode, stimulating acoustic signals at higher frequencies and enhancing the ability to detect deep-seated defects. Seepage pressure monitoring data records peak pressure events at high water levels, statistically analyzing the spatial distribution of extreme pressure values ​​and performing time-domain alignment analysis with structural vibration data.

[0027] After completing the extraction of single working condition features, all feature parameters are integrated according to a unified spatial grid. Geometric deformation features include curvature distribution matrix, point cloud density gradient, and surface roughness index, which are stored in the structural feature layer. Acoustic features include the main peak position of the acoustic wave delay spectrum, energy attenuation coefficient, and waveform distortion, which are stored in the medium feature layer. Hydraulic features include pressure conduction rate, pulsation frequency, and pressure gradient vector, which are stored in the hydraulic feature layer. The three groups of feature layers are linked through spatial coordinate indexes to form a multi-dimensional working condition feature set with 1 cm grid as the basic unit. Each grid unit contains a 14-dimensional feature vector, covering quantitative indicators in three dimensions: geometry, acoustics, and hydraulics.

[0028] The feature set is stored in a hierarchical database structure. The raw data layer retains time series waveforms, the feature extraction layer stores statistical parameters, and the high-level abstraction layer records operating condition labels and spatial topological relationships. A data versioning mechanism ensures the traceability of the feature set. Each update records metadata such as water level intervals, timestamps, and sensor calibration parameters. The feature extraction algorithm uses adaptive parameter configuration, automatically adjusting the analysis window size and feature sensitivity threshold based on the operating condition type, ensuring detection consistency under different water level conditions.

[0029] Key technical features include a dual-redundancy check mechanism for water level data to ensure reliable operating condition classification; a parallel multimodal feature extraction process for efficient data utilization; and standardized spatial grid mapping to ensure precise alignment of data from different sensors. The feature set generation cycle is synchronized with data acquisition, and a single complete analysis takes less than 200 milliseconds, meeting the timeliness requirements of real-time monitoring. The completed multidimensional operating condition feature set is output to the dynamic analysis module as the input data source for anomaly detection and defect identification.

[0030] Example 3: See Figure 4 , the dynamic analysis of abnormal working conditions is achieved by real-time monitoring of the switching behavior of hydrological working condition intervals. The system deploys an independent status monitoring module to continuously receive the working condition label data stream from the water level classification module. The monitoring window adopts a sliding time mechanism, the window length is fixed at 60 minutes, and the sliding step is set to 1 minute. Every time a water level interval change event occurs in the window, that is, when switching between low water level, medium water level, and high water level, the counter automatically accumulates. When the cumulative number of switching in the window exceeds the preset threshold, the dynamic analysis process is triggered. The threshold is dynamically adjusted according to the historical data of tunnel operation, and the initial value is set to 5 times / hour.

[0031] After the dynamic analysis is started, the system intercepts the transition period data of the two adjacent operating conditions within the abnormal window, and the time range covers 10 minutes before the switch to 10 minutes after the switch. The acoustic detection data is first pre-processed, and the delay spectrum is zero-meaned to eliminate the DC offset, and then the matching degree of the two operating condition data is calculated by normalizing the cross-correlation coefficient. The pressure conduction rate sequence uses the dynamic time warping algorithm to align the time axis, compensate for the phase delay caused by water inertia, and then calculate the cosine angle value of the trend similarity. The cross-operating condition stability index is generated by weighted fusion of the analysis results of acoustic wave and pressure data: ; in, represents the cross-operating stability index, is the normalized cross-correlation coefficient of the acoustic wave delay spectrum, The cosine similarity represents the trend of pressure transmission rate. When the value is lower than 0.75, it is determined that the operating condition interval corresponding to the current time window is abnormal.

[0032] The marking of abnormal working conditions adopts a three-level classification mechanism: the first-level abnormality corresponds to , the secondary anomaly corresponds to , the third level abnormality corresponds to Different levels trigger differentiated response strategies: Level 1 anomalies initiate the data review process, Level 2 anomalies dispatch local re-inspection equipment, and Level 3 anomalies trigger a full-section emergency scan. During the review process, the system automatically records the spatial distribution pattern of anomaly features, including the coordinates of areas with sudden changes in acoustic delay and pressure conduction blockage points.

[0033] When re-extracting features, the system prioritizes the time point with the smallest hydraulic difference from the original acquisition time. Laser scanning uses anti-interference mode, increasing the transmission power by 20% to improve the signal-to-noise ratio. The acoustic sensor switches to broadband pulse coding mode to enhance the ability to identify multiple reflections. Seepage pressure monitoring uses high-frequency sampling mode, increasing the sampling rate to 500Hz to capture transient fluctuations. Newly acquired data undergoes the same normalization process as the original data, replacing any abnormal records in the original feature set.

[0034] The dynamic analysis module features a built-in self-learning mechanism that automatically updates the threshold parameter library after each exception handling process. New records include 12 environmental parameters, including the current hydraulic load spectrum, ambient temperature compensation factors, and equipment operating status codes. The threshold adjustment algorithm uses a gradual adaptation strategy, with each update not exceeding 15% of the historical average to avoid sudden sensitivity changes caused by a single abnormal event.

[0035] Abnormal condition analysis results are output as a structured report covering three core dimensions: anomaly level, spatiotemporal location, and feature deviation. The temporal dimension is accurate to the second, while the spatial dimension uses a combination of tunnel mileage and zoning codes for location. Feature deviation quantitatively displays the percentage difference between acoustic wave and pressure data and the original operating baseline, assisting manual review and judgment. A unique tracking ID is generated for each abnormal event, linking the original data snapshot with the processing log, supporting full lifecycle traceability.

[0036] The technical features of this implementation are reflected in three aspects: a sliding window mechanism balances real-time detection with statistical reliability; multimodal data fusion analysis improves the robustness of anomaly identification; and a hierarchical response strategy optimizes resource allocation. The average dynamic analysis process takes less than 90 seconds, of which 60 seconds are spent on data verification and collection, and 30 seconds are spent on feature recalculation and update. The system continuously monitors computing load during operation. When concurrent anomalies exceed processing capacity, it automatically activates a degraded mode, performing only key parameter verification to ensure basic security.

[0037] Historical records of stability indicators form a baseline database of operating status, with trend analysis performed weekly. The baseline database utilizes a rolling storage strategy, retaining the last 12 months of data. Expired records are archived after feature compression. Long-term data analysis identifies macro-regularities such as seasonal water level fluctuation patterns and equipment performance degradation curves, which are used to optimize the initial settings of threshold parameters. The anomaly analysis module integrates with the maintenance management system in real time. Level 3 anomalies directly trigger the generation of maintenance work orders, which are then pushed to the relevant responsible terminals.

[0038] Example 4: The process of decoupling corrosion features and coordinating parameter correction is carried out for a multi-dimensional feature set in the non-abnormal operating condition range. The system first accesses the historical corrosion feature database, which stores typical feature templates classified by material degradation stage, including 12 types of benchmark patterns, such as curvature mutation characteristics in the microcrack stage, acoustic wave attenuation patterns in the cavitation stage, and pressure gradient changes caused by chemical corrosion. Each grid cell in the current operating condition feature set is matched with the historical template, and the improved Euclidean distance is used to measure similarity: ; in, Indicates the grid cells and the The weighted distance of historical templates, For the current unit dimensional eigenvalues, Is the template Dimensional benchmark value, is the feature weight coefficient (geometric feature weight 0.4, acoustic feature weight 0.3, hydraulic feature weight 0.3). When the similarity corresponding to the value exceeds 85%, the system determines that there is a clear erosion pattern.

[0039] The feature decoupling operation uses different strategies depending on the matching results. For high similarity, a finite impulse response filter bank is used to perform frequency domain decomposition. The filter passband strictly corresponds to the typical frequency range of historical corrosion: microcrack features are concentrated in the 3-8 kHz range, cavitation features are located in the 12-18 kHz range, and chemical corrosion exhibits a broadband response of 0.5-2 kHz. The transition band width of each filter is set to 20% of the passband, and the stopband attenuation is no less than 60 dB. Through this process, a 3-5 dimensional core corrosion feature subset is extracted from the original 14-dimensional feature vector.

[0040] When the similarity falls below the threshold, the system activates the resonance frequency band analysis mode. Based on the continuous wavelet transform of the acoustic energy attenuation spectrum, frequency bands with prominent structural resonance characteristics are identified. The sliding integral of the energy proportion of each frequency band is calculated. When the energy integral proportion of a frequency band (such as 1200±50Hz) exceeds 65%, wavelet packet decomposition is performed on that frequency band. The number of decomposition layers is adaptively determined based on the frequency band width, typically 4-6 layers, to generate a subset of corrosion features with frequency positioning accuracy. This process also monitors the spatial consistency of the frequency band energy to eliminate local anomalies caused by sensor failure.

[0041] The parameter collaborative correction stage addresses systematic errors in the erosion feature subset. Concrete structural mechanics correction uses an elastic modulus compensation algorithm. Based on the nonlinear characteristics of the curvature measurement and the material constitutive relationship, a compensation function for the point cloud density gradient is established. This function takes into account eight influencing factors, such as concrete age and ambient humidity, and dynamically adjusts the compensation coefficient. For hydraulic parameters, a simplified fluid-solid coupling model is established based on the theory of viscous fluid mechanics. The seepage pressure pulsation amplitude is input as a boundary condition, and the true water pressure gradient is inverted and calculated. During the correction process, the physical constraint relationship between geometric deformation and hydraulic characteristics is maintained to avoid feature distortion caused by excessive correction.

[0042] The corrected erosion feature subset is then fused to generate the final output. Geometric corrections, acoustic features, and hydraulic corrections are spatially aligned to form a new 8-dimensional feature tensor. This tensor is stored in a hierarchical structure: the base layer retains the original measurements, the correction layer stores the adjustments, and the fusion layer records the final feature vector. A quality flag is assigned to each grid cell, reflecting the confidence level of the correction process. The value ranges from 0 to 1 and is determined by data integrity, the applicability of the correction model, and the degree of environmental interference.

[0043] The system continuously optimizes characteristic decoupling parameters during operation. Each time a corrosion pattern is successfully identified, its characteristic parameters are automatically extracted and updated in the template library. New templates must pass spatiotemporal consistency verification: repeated occurrence in three or more different sections, with a duration exceeding 24 hours. The parameter correction model is recalibrated monthly using data derived from measured material parameters from the tunnel sampling test section.

[0044] The technical features of this implementation are as follows: a dynamic update mechanism for historical templates enables the system to adapt to new material degradation patterns; a combined frequency domain decomposition and resonance analysis strategy allows for the identification of both common corrosion damage and specific defects; and parameter correction incorporates physical constraints to avoid distortion caused by purely mathematical processing. The entire process averages 15 milliseconds per grid cell, and the massively parallel computing architecture supports processing of over 100,000 spatial cells per second.

[0045] The corrected feature subset is output with complete processing metadata, including the decoupling strategy code, the corrected parameter version, and the quality mark calculation log. This data is shared with the subsequent defect tracing module to assess the reliability of the location results. The system establishes a feedback loop for correction results. If subsequent maintenance inspections reveal misjudgments or missed detections, the feature processing records of the relevant cases will be marked, triggering an expert review process and adjusting the algorithm parameters.

[0046] During long-term operation, the feature decoupling module forms a knowledge base of typical pattern recognition, organized and stored by dimensions such as material type, age, and environmental category. This knowledge base supports analogical reasoning. When historical data is lacking for new facilities, feature templates from similar environments can be used for transfer learning. An audit trail is maintained for all corrective actions, meeting the traceability requirements of project quality management. The processing history of individual feature vectors can be fully reproduced.

[0047] Example 5: The defect tracing and maintenance decision-making process is based on a subset of corroded features after parameter collaborative correction. The system loads a pre-built defect evolution path map, which is stored in a directed weighted graph structure. Nodes represent the smallest structural unit (1 meter × 1 meter block), and edges represent the defect transmission path. The edge weights are calculated based on the hydraulic scour direction and stress transfer path. The weights of downstream units are usually set to a specific proportion of the upstream units to reflect the attenuation characteristics of defect propagation. The map data is derived from a combination of historical maintenance records, material experimental data, and numerical simulation results. Topology optimization is performed monthly based on newly added inspection results.

[0048] When identifying abnormal characteristic parameters from the corrected erosion feature subset, the system sets dual judgment conditions. In terms of geometric characteristics, the point cloud density gradient exceeds the preset threshold and lasts for more than three detection cycles; in terms of acoustic characteristics, the increase in the acoustic wave attenuation coefficient compared with the baseline value reaches the warning level; in terms of hydraulic characteristics, the decrease in the pressure conduction rate and the abnormal pulsation frequency appear at the same time. If any dimension meets the abnormal condition, it will be marked as a suspicious node. If all three dimensions are abnormal at the same time, it will be directly judged as a defective unit. The positions of all abnormal nodes in the map are quickly located using a spatial hash algorithm. The matching process takes into account the surface characteristics of the tunnel structure and converts the three-dimensional coordinates into a two-dimensional mapping unfolded along the axis.

[0049] The reverse tracing process utilizes an improved depth-first search algorithm. Starting from the abnormal node, upstream nodes are recursively accessed along the incoming edge, with the tracing depth limited to five levels. The weight attenuation factor for each level of the conduction path is dynamically adjusted, taking into account the current hydraulic load and the material's remaining life coefficient. During the tracing process, a path credibility index is calculated in real time, combining the integrity of the conduction path, the recurrence rate of historical defects, and the matching degree of the current abnormal characteristics. When multiple potential defect sources are accessed, the system establishes a temporary decision tree to evaluate each candidate source's ability to explain the downstream anomaly, prioritizing the conduction paths that cover the most abnormal characteristics.

[0050] A multi-factor weighted model is used to calculate repair priority. The base weight is derived from the number of abnormal parameters, with each abnormal dimension assigned a different score: geometric anomalies are assigned 1.0 points, acoustic anomalies are assigned 1.2 points, and hydraulic anomalies are assigned 0.8 points. The depth of the conductive path is converted using an exponential decay function, with the primary path receiving full weight and the weight decreasing with each deeper level. The material degradation rate is used as a dynamic adjustment factor, linearly interpolated based on the trend change rate of the three most recent test results. The final priority score is normalized and output on a percentage basis, divided into four response levels: emergency (80-100 points), high (60-79 points), medium (40-59 points), and low (0-39 points).

[0051] The maintenance instruction generation module converts engineering semantics into a machine-executable format. The three-dimensional coordinates of the target unit are converted using a tunnel-specific coordinate system. The vertical coordinates are accurately defined to the centimeter using mileage stakes, and the circumferential coordinates are divided into 36 10-degree sections. Maintenance time windows are automatically set based on priority. Urgent tasks generate immediate response instructions. High-priority tasks are scheduled for resolution within six hours, medium-priority tasks are included in the 24-hour work order schedule, and low-priority tasks are queued for weekly maintenance. The instruction data package is encapsulated in the standard JSON-LD format and contains 12 types of structured information, including spatial positioning data, defect type codes, and recommended process parameters.

[0052] A closed-loop feedback mechanism is established for command transmission and execution. After receiving the command, the maintenance terminal first reviews on-site conditions and verifies defect characteristics using portable testing equipment. Confirmation results are transmitted back to the central system in real time. If the deviation between on-site conditions and the command exceeds the allowable range, the command correction process is triggered. During execution, key operating parameters, including material usage, process time, and treatment results, are recorded. This data is fed back into the defect map as new knowledge. After maintenance is completed, the system automatically schedules verification testing to ensure that defect elimination meets standard requirements.

[0053] The technical features of this implementation are reflected in three aspects: graph-driven traceability logic maintains the physical plausibility of defect location; a multi-level weighting model optimizes the allocation of limited resources; and a closed-loop instruction flow ensures coordinated decision-making and execution. The entire processing flow, from feature input to instruction output, takes an average of three minutes, which can be reduced to 45 seconds in emergency mode. The system continuously monitors resource usage during operation. When the number of concurrent defects exceeds processing capacity, a load balancing strategy is automatically activated to prioritize analysis resources for high-risk defects.

[0054] Over its long-term operation, the defect traceability system has accumulated a case knowledge base. Each processed defect case records the complete lifecycle data, including the time of first occurrence, development process, treatment measures, and subsequent tracking results. The knowledge base supports multi-dimensional search capabilities and can perform statistical analysis based on criteria such as defect type, location distribution, and material batch. Regularly generated trend reports reveal the temporal and spatial patterns of defect occurrence, providing data support for adjusting tunnel health management strategies. All maintenance instructions and their execution results are retained for at least the design service life of the structure, meeting the full lifecycle quality management requirements of major infrastructure.

[0055] The system maintains a library of exception handling plans, pre-setting standard response plans for frequently occurring defect types. When a known defect pattern is detected, it automatically associates the treatment process parameters, required material specifications, and safety protection requirements in the plan. The plan library is updated quarterly to incorporate the latest research findings and industry best practices. For new defects, the system activates an expert collaboration mode, distributing abnormality feature data to multiple responsible terminals. After collecting treatment suggestions, temporary solutions are developed and converted to standard plans after proven effectiveness in practice.

[0056] The maintenance command system is deeply integrated with subsystems such as inventory management, personnel scheduling, and safety monitoring. High-priority commands automatically trigger the spare parts allocation process, ensuring timely material delivery. Maintenance tasks requiring high-altitude work also require simultaneous verification of personnel qualifications. Commands involving confined space operations are forcibly linked to ventilation monitoring data. This integrated approach significantly shortens the response cycle from problem discovery to intervention, reducing the average decision-making time in traditional manual processes from hours to minutes.

[0057] Augmented reality technology is used to visualize defect location results. Maintenance personnel use smart terminals to view a 3D defect distribution map within the tunnel. Abnormal areas are color-coded with priority levels, and dynamic arrows indicate the conduction paths. Clicking on a specific defect point reveals detailed characteristic parameters, historical change curves, and recommended treatment plans. The visualization system supports collaborative annotation by multiple users, allowing new findings discovered on-site to be annotated directly on the 3D model. This information is then synchronized with the technical support team in real time.

[0058] The system incorporates a multi-level review mechanism to ensure decision reliability. The first-level review, automatically performed by the algorithm, checks the logical consistency of feature extraction and graph matching. The second-level review, performed by on-duty engineers, focuses on examining the basis for identifying high-risk defects. The third-level review involves interdisciplinary consultations for disputed cases. All review records are stored alongside the original decision data, forming a complete chain of quality accountability. Regular spot checks are conducted on the decision-making process of processed cases to assess the consistency between system judgments and human experience, and to continuously optimize algorithm parameters.

[0059] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0060] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for automatic inspection and intelligent detection of corrosion defects in flood discharge tunnels, characterized in that: The following steps are involved: Collect structural status data inside the flood discharge tunnel and perform standardization processing; Classify the structural state data according to the preset hydrological operating condition interval division standard, extract the multimodal sensing features within each hydrological operating condition interval, and generate a set of operating condition features; Monitor the switching frequency of hydrological operating intervals. When the switching frequency exceeds a preset frequency threshold, activate a dynamic analysis strategy to identify abnormal operating intervals. Based on the matching degree between the historical corrosion stage characteristics and the current multimodal sensing characteristics, a feature decoupling operation is performed on the operating condition feature set of the non-abnormal operating condition interval to separate the corrosion feature subset; Based on the coupling correlation between tunnel structural mechanical parameters and hydraulic parameters, the parameters of the erosion feature subset are collaboratively corrected. The corrected corrosion feature subset is used to reversely trace the potential defect source area along the defect evolution path, determine the repair priority and output maintenance instructions.

2. The method for automatic inspection and intelligent detection of erosion defects in flood discharge tunnels according to claim 1 is characterized in that: The collecting of structural status data inside the flood discharge tunnel and performing standardization processing include: Synchronously acquire laser scanning data, acoustic detection data, and seepage pressure monitoring data, and perform time synchronization alignment operations on all data sources; Point cloud noise reduction is performed on the time-aligned laser scanning data, spectrum filtering is performed on the acoustic detection data, and an outlier removal algorithm is used on the seepage pressure monitoring data; The processed laser scanning data, acoustic detection data and seepage pressure monitoring data were resampled at a uniform spatial resolution to construct a standardized three-dimensional state data set.

3. The method for automatic inspection and intelligent detection of erosion defects of flood discharge tunnels according to claim 1 is characterized in that: The structural state data is classified according to the preset hydrological operating condition interval division standard, and the multimodal sensing features within each hydrological operating condition interval are extracted to generate an operating condition feature set, including: The standardized three-dimensional state data set is divided into multiple hydrological operating ranges according to the water level change amplitude, including low water level operating range, medium water level operating range and high water level operating range; Extract the surface deformation characteristics of laser scanning data, the medium attenuation characteristics of acoustic detection data, and the water pressure gradient characteristics of seepage pressure monitoring data within each hydrological operating range respectively; The surface deformation characteristics, medium attenuation characteristics and water pressure gradient characteristics within the same hydrological operating condition interval are integrated according to the spatial grid to form a multidimensional operating condition feature set.

4. The method for automatic inspection and intelligent detection of corrosion defects of flood discharge tunnels according to claim 3 is characterized in that: The surface deformation characteristics of the laser scanning data include the curvature distribution matrix and the point cloud density gradient, the medium attenuation characteristics of the acoustic wave detection data include the acoustic wave propagation delay spectrum and the energy attenuation coefficient, and the water pressure gradient characteristics of the seepage pressure monitoring data include the pressure conduction rate and the pulsation frequency.

5. The method for automatic inspection and intelligent detection of corrosion defects of flood discharge tunnels according to claim 1 is characterized in that: The switching frequency of the monitoring hydrological operating condition interval is activated, and when the switching frequency exceeds a preset frequency threshold, a dynamic analysis strategy is activated to identify the abnormal operating condition interval, including: Based on the sliding time window, the number of switching times of the hydrological operating interval is counted, the switching frequency per unit time is calculated, and it is determined whether the switching frequency exceeds the preset frequency threshold; If the switching frequency exceeds the preset frequency threshold, the multimodal sensing features of adjacent hydrological operating conditions are matched using a waveform alignment algorithm to extract cross-operating condition stability indicators. When the cross-operating condition stability index is lower than the stability threshold, the hydrological operating condition interval corresponding to the current time window is marked as an abnormal operating condition interval, and the re-extraction process of the multimodal sensing features of the interval is triggered.

6. The method for automatic inspection and intelligent detection of corrosion defects of flood discharge tunnels according to claim 5 is characterized in that: The extraction of the cross-operating-condition stability index includes: performing a cross-correlation operation on the acoustic wave propagation delay spectra of adjacent hydrological operating-condition intervals, performing a trend similarity calculation on the pressure conduction rate, and weightedly fusing the calculation results to generate the cross-operating-condition stability index.

7. The method for automatic inspection and intelligent detection of corrosion defects of flood discharge tunnels according to claim 1 is characterized in that: According to the matching degree between the historical corrosion stage characteristics and the current multimodal sensing characteristics, the feature decoupling operation is performed on the operating condition feature set of the non-abnormal operating condition interval to separate the corrosion feature subset, including: For the current multimodal sensing features in the working condition feature set of the non-abnormal working condition interval, calculate the similarity matrix between them and the historical erosion stage features; When the maximum value of the similarity matrix exceeds the preset similarity threshold, the frequency domain bandpass decomposition mode is used to extract the erosion feature subset from the working condition feature set; When the maximum value of the similarity matrix is ​​lower than the preset similarity threshold, based on the correlation between the acoustic wave energy attenuation coefficient and the structural resonance frequency, the frequency band where the characteristic energy is concentrated is selected for directional decomposition to generate a erosion feature subset.

8. The method for automatic inspection and intelligent detection of corrosion defects of flood discharge tunnels according to claim 1 is characterized in that: The method of implementing parameter collaborative correction on the erosion feature subset based on the coupling correlation between the tunnel structure mechanical parameters and the hydraulic parameters includes: According to the nonlinear relationship between the elastic modulus of concrete and the surface curvature, deformation compensation correction is performed on the point cloud density gradient in the erosion feature subset. Based on the correlation between hydraulic load distribution and seepage pressure pulsation, dynamic pressure correction is performed on the water pressure gradient characteristics in the erosion feature subset; The corrected point cloud density gradient is fused with the water pressure gradient feature to generate a erosion feature subset with coordinated parameter correction.

9. The method for automatic inspection and intelligent detection of corrosion defects of flood discharge tunnels according to claim 1 is characterized in that: The method of using the modified corrosion feature subset to reversely trace the potential defect source area along the defect evolution path, determine the repair priority and output maintenance instructions includes: Load a pre-built defect evolution path map, which contains the defect transmission relationship and influence weight between structural units; Identify abnormal characteristic parameters from the erosion feature subset after parameter collaborative correction and match the corresponding nodes in the defect evolution path map; Traverse in reverse along the conduction direction of the defect evolution path to locate the potential defect source unit with abnormal characteristic parameters; Calculate the repair priority weight based on the number of abnormal characteristic parameters and the depth of the conduction path of the potential defect source unit; Based on the repair priority weight and defect development rate, a maintenance instruction containing the target unit coordinates and maintenance time is generated.

10. The method for automatic inspection and intelligent detection of corrosion defects in flood discharge tunnels according to claim 9, characterized in that: The maintenance instructions are converted into three-dimensional coordinate instruction sets through geographic information coding and automatically output to the tunnel maintenance execution terminal.

Images