A tunnel crack detection method based on multi-sensor fusion

Abstract

This invention discloses a tunnel crack detection method based on multi-sensor fusion, specifically relating to the field of tunnel defect detection technology. It addresses the problems of incomplete information acquisition, insufficient anti-interference capabilities, and susceptibility to missed, false, and unstable crack detection in complex tunnel environments caused by existing single-sensor detection methods. By collecting ultrasonic data, infrared thermal imaging data, and lidar point cloud data from the target tunnel section, and establishing a unified time reference, unified spatial coordinate reference, and multi-source correspondence for the multi-source data, a fusion judgment is performed based on the crack characterization feature set. This overcomes the problems of incomplete information acquisition, insufficient anti-interference capabilities, and susceptibility to missed and false detections caused by single-sensor methods in complex tunnel environments, achieving stable identification and accurate judgment of tunnel cracks.

Description

Technical Field

[0001] This invention relates to the field of tunnel defect detection technology, specifically a tunnel crack detection method based on multi-sensor fusion. Background Technology

[0002] During long-term operation or construction, tunnels are susceptible to cracks on the lining surface and near-surface due to factors such as surrounding rock deformation, lining shrinkage, temperature and humidity changes, load fluctuations, and water erosion. Cracks not only weaken the overall load-bearing capacity of the lining but can also further lead to leakage, spalling, and decreased durability. Therefore, timely detection and assessment of tunnel cracks are crucial. Currently, tunnel crack detection typically employs methods such as manual inspection, visible light image recognition, single infrared thermal imaging, single ultrasonic testing, or single laser scanning. Manual inspection relies heavily on experience, resulting in limited efficiency and consistency. Visible light or infrared methods primarily target surface characterization and are easily affected by light, water stains, dirt, and surface texture. Ultrasonic testing can reflect some internal response information but is sensitive to the placement of measuring points, coupling conditions, and on-site working conditions. Laser scanning is better suited for acquiring surface contours and spatial morphology but struggles to provide effective characterization of the internal state of cracks and local thermal anomalies.

[0003] In real-world tunnel scenarios, the lining surface often suffers from dampness and reflection, dust accumulation, localized obstruction, surface undulations, and disturbances from the movement of inspection vehicles. Single detection methods offer limited information, often only reflecting a portion of crack characteristics. It's difficult to simultaneously assess the surface morphology, thermal anomalies, and near-surface response of cracks during a single inspection. This is especially true in sections with micro-cracks, early-stage cracks, or areas with strong background interference, leading to missed or false detections. Even when suspected cracks are identified, the stability of determining their location, extension, and severity is insufficient, impacting the usability of the detection results in subsequent verification, damage registration, and maintenance decisions.

[0004] Therefore, how can existing technologies effectively correspond and comprehensively judge relevant information from different detection methods in complex tunnel detection environments to improve the integrity, stability, and reliability of tunnel crack detection results? Summary of the Invention

[0005] The objective of this invention is achieved through the following technical solutions.

[0006] To address the shortcomings of existing technologies, this invention provides a tunnel crack detection method based on multi-sensor fusion to solve the problems mentioned in the background section.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a tunnel crack detection method based on multi-sensor fusion, comprising:

[0008] S1. Collect ultrasonic data, infrared thermal image data, and lidar point cloud data of the target tunnel section to establish a multi-source raw dataset for the corresponding detection section.

[0009] S2. Perform denoising, correction, and coordinate transformation processing on the multi-source raw dataset to establish a unified time reference and a unified spatial coordinate reference;

[0010] S3. Based on a unified time reference and a unified spatial coordinate reference, perform registration on ultrasonic data, infrared thermal image data, and lidar point cloud data to establish multi-source correspondence in the same segment;

[0011] S4. Extract ultrasonic depth response features, infrared thermal anomaly features, and surface morphology features based on multi-source correspondence to generate a crack characterization feature set;

[0012] S5. Based on the crack characterization feature set and the quality of each source data, assign fusion weights, perform fusion judgment, and generate crack characterization results.

[0013] S6. Based on the crack characterization results, determine the crack location, direction, length, width, depth, and grade, and generate a crack detection record.

[0014] Furthermore, S1 includes:

[0015] Before data collection, the ultrasonic probe, infrared thermal imager, and lidar were synchronized with a unified clock, their installation positions were verified, and their zero points were calibrated.

[0016] During the acquisition process, echo sequences, thermal images, and point cloud frames are grouped side-by-side using an observation window.

[0017] After the section data acquisition is completed, a multi-source raw dataset is generated, which includes a measurement point table, thermal image frame set, point cloud frame set, equipment status record, calibration record, and operating environment record.

[0018] Furthermore, S2 includes:

[0019] Retrieve ultrasonic data, infrared thermal imaging data, and lidar point cloud data;

[0020] Noise reduction and calibration were performed based on the equipment calibration results.

[0021] Complete coordinate transformation based on installation location, orientation angle, and reference point coordinates;

[0022] The recording time of various data is converted into the section operation time, and the measurement point position, thermal image position, and point cloud position are converted into the same target tunnel section coordinates.

[0023] Furthermore, S3 includes:

[0024] Based on a unified time reference, ultrasonic records, infrared thermal images, and lidar point cloud records located within the same observation window were selected.

[0025] Based on a unified spatial coordinate reference, the position of the measurement point is mapped to the thermal imaging area and the position of the point cloud surface to form candidate corresponding units;

[0026] The correspondence between multiple sources in the same segment is confirmed based on the chronological order, spatial proximity, and consistency of changes within a continuous window.

[0027] Furthermore, S4 includes:

[0028] Based on the converted timestamps, ultrasonic records, infrared thermal images, and lidar point cloud records located in the same observation window are grouped into candidate groups.

[0029] Based on the segment coordinates, the measurement point location, the center location of the thermal imaging area, and the local center location of the point cloud, candidate groups with spatial proximity are selected.

[0030] The correspondence between multiple sources was confirmed based on the consistency of the positional changes of the observation windows before and after.

[0031] Furthermore, S5 includes:

[0032] The quality of ultrasonic data is determined based on the coupling state, the quality of infrared thermal image data is determined based on the clarity of the thermal image, the quality of lidar point cloud data is determined based on the point cloud coverage and pose continuity, and the corresponding fusion weights are determined based on the quality of each source data.

[0033] Furthermore, based on the fusion weights, the ultrasonic depth response characteristics, infrared thermal anomaly characteristics, and surface morphology characteristics at the same measurement point location are used to check for crack candidate conditions.

[0034] The locations of measuring points that meet the crack candidate conditions are determined as crack candidate locations;

[0035] Crack segments are determined based on the spatial continuity of adjacent candidate crack locations and the unidirectional extension relationship within the front and rear observation windows.

[0036] Furthermore, S6 includes:

[0037] The location and length of the crack are determined based on the segment coordinates of the first and last measuring points of the crack segment. The direction of the crack is determined based on the direction of the center line connecting the candidate locations. The width of the crack is determined based on the surface morphology and infrared thermal anomaly characteristics. The depth of the crack is determined based on the ultrasonic depth response characteristics. The crack grade is determined based on the length, width, depth and overall confirmation grade.

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] 1. By collecting ultrasonic data, infrared thermal image data, and lidar point cloud data of the target tunnel section, and performing denoising, correction, and coordinate transformation on the multi-source raw datasets, a unified time reference and a unified spatial coordinate reference are established. Then, based on the unified time reference and the unified spatial coordinate reference, a multi-source correspondence relationship for the same section is established. Based on the multi-source correspondence relationship, ultrasonic depth response features, infrared thermal anomaly features, and surface morphology features are extracted. Furthermore, the fusion judgment is performed by combining the fusion weights assigned to each source data quality. This overcomes the problems of incomplete information acquisition, insufficient anti-interference ability, and easy omission and false detection of existing single sensing methods in the complex environment of tunnels, and achieves stable identification, accurate judgment, and improved consistency of detection results for tunnel cracks.

[0040] 2. After completing the fusion judgment, the location, direction, length, width, depth and grade of the crack are further determined based on the crack characterization results. A crack detection record containing crack segment number, crack location, direction, length, width, depth and grade is generated. This improves the crack detection results from single crack discovery to structured characterization results for engineering applications. The crack detection results have the technical effects of being able to locate, quantify, classify and trace, which facilitates the continuous use of subsequent inspection and verification, disease record updates and maintenance decisions. Attached Figure Description

[0041] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0042] Figure 1 This is a schematic flowchart of a tunnel crack detection method based on multi-sensor fusion according to the present invention.

[0043] Figure 2 This is a schematic diagram of a multi-sensor tunnel crack detection system.

[0044] Figure 3 A schematic diagram for establishing a unified time reference and a unified spatial coordinate reference;

[0045] Figure 4 Establish a schematic diagram for the source-to-source correspondence;

[0046] Figure 5 This is a schematic diagram of crack characterization feature extraction and fusion determination. Detailed Implementation

[0047] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0048] Example: Combined with Appendix Figure 1-3 This embodiment provides a tunnel crack detection method based on multi-sensor fusion, including:

[0049] S1. Collect ultrasonic data, infrared thermal image data, and lidar point cloud data of the target tunnel section to establish a multi-source raw dataset for the corresponding detection section. The specific implementation is as follows:

[0050] This step is deployed using tunnel inspection vehicles, track inspection platforms, or manually pushed inspection trolleys. It is applicable to operating tunnels, construction tunnels, and maintenance and re-inspection sections where the secondary lining surface is accessible. The target tunnel section is a continuously defined lining section to be inspected, and on-site, it can be set to a testing length of 20 to 100 meters as per the work order issued before each operation. Ultrasonic data is the echo sequence collected after the probe couples and emits at the lining surface, reflecting the reflection from the internal interface of the lining. During the residence period at each measuring point, the amplitude, voltage (volts), arrival time (microseconds), and sampling frequency (units) are continuously collected and recorded. The measurement speed is terasecond per second and the measurement point coordinates are recorded. The typical measurement point spacing can be set from 10 mm to 50 mm. The infrared thermal image data is an image of the temperature distribution of the lining surface in the same section. It is continuously acquired at a fixed frame rate when the inspection vehicle passes by at a constant speed, and the temperature value in degrees Celsius, pixel position, imaging time and lens attitude are recorded. The temperature calibration error is controlled within ±2 degrees Celsius. The lidar point cloud data is a set of spatial points on the lining surface in the same section. It is formed frame by frame during the vehicle's movement and the three-dimensional coordinates in millimeters, reflection intensity, scanning time and equipment attitude are recorded. The ranging error is controlled within ±5 mm.

[0051] Before data acquisition begins on-site, a unified clock synchronization, installation location verification, and zero-point calibration are performed on the three types of equipment. The calibration results are written into the work record according to the equipment number, version number, executor, and execution time. During the acquisition process, a 100-millisecond observation window is used as the basic observation window. The echo sequence, thermal image frame, and point cloud frame formed within the same observation window are collected side by side according to the timestamp. One supplementary acquisition is allowed for missing frames. If two consecutive observation windows are not filled, it is recorded as a gap and the original mark is retained. The interpolated value is not used to replace the measured value. The original file is written to the local solid-state drive of the on-site industrial control computer using an append-only storage method. The media generates unique record keys based on project number, section number, equipment number, timestamp, and sequence number. When the same record key arrives again, only the first successfully written version is retained and written to the deduplication log to ensure consistent order and traceability of duplicates. After the section data acquisition is completed, a multi-source raw dataset for the corresponding detection section is formed. The content includes at least the measurement point table, thermal image frame set, point cloud frame set, equipment status record, calibration record, and operating environment record. These are saved in separate files in the same directory and an index list is generated for direct retrieval by section number and time range in the next stage.

[0052] Upstream and downstream communication is achieved via vehicle Ethernet or industrial serial port. The maximum latency for a single observation window can be set to 500 milliseconds. If writing to disk fails, it will retry twice. If it still fails, it will switch to a backup directory and record the reason code. If coupling failure, lens obstruction, radar echo saturation, or clock drift exceeds 5 milliseconds, the current observation window will be frozen immediately, the fault record will be written, and the measured data of other sensors will continue to be retained to avoid failure of the entire operation.

[0053] Preferably, under the conditions of lining surface temperature of 5°C to 35°C, inspection speed of 0.3 m / s, section length of 40 m, measuring point spacing of 20 mm, thermal image frame rate of 30 frames / s, and point cloud frequency of 10 Hz, 2000 ultrasonic measuring points, 1200 thermal image frames, and 400 point cloud frames can be formed in a single section, with a subsequent retrieval success rate of no less than 99.5%. Alternatively, lidar point cloud data can be replaced by structured light contour scanning data, as long as the same section, unified timing, unified pose recording, and the same aperture collection method are maintained.

[0054] S2. Perform denoising, correction, and coordinate transformation processing on the multi-source raw dataset to establish a unified time reference and a unified spatial coordinate reference. The specific implementation is as follows:

[0055] Denoising, correction, and coordinate transformation are performed on the multi-source raw datasets to establish a unified time reference and a unified spatial coordinate reference. In one embodiment, after the target tunnel section is acquired, the on-site industrial control computer immediately retrieves the ultrasonic data, infrared thermal image data, and lidar point cloud data written to the local solid medium in the previous stage, segment by segment. The ultrasonic data is taken from the measurement point table and echo sequence file to obtain the measurement point coordinates, echo amplitude, arrival time, sampling frequency, and coupling state. The infrared thermal image data is taken from the thermal image frame set to obtain the frame number, frame time, pixel temperature, lens attitude, and temperature measurement calibration record. The lidar point cloud data is taken from the point cloud frame set to obtain the point coordinates, reflection intensity, scanning time, and equipment attitude.

[0056] Noise reduction refers to removing isolated abrupt changes that are inconsistent with the continuous surface changes of the lining within adjacent observation windows after the formation of their respective signals. Ultrasonic waves remove spike interference by comparing the peak jumps and noise levels of the continuous echo envelope during the residence period of a single measurement point. Infrared thermal images remove instantaneous hot spots by comparing the temperature jumps of the same pixel between consecutive frames. LiDAR point clouds remove outliers by comparing the point spacing and local surface continuity between adjacent scan lines. Correction refers to restoring the original quantities to actual measurements under a unified aperture based on the equipment calibration results written in the work record before acquisition. Ultrasonic waves correct the arrival time according to the probe zero point and propagation time base. Infrared thermal images correct the pixel temperature according to the blackbody calibration results. LiDAR corrects the point coordinates according to the installation angle and ranging deviation. Coordinate transformation refers to converting the position quantities recorded under the coordinates of each device to the position quantities under the coordinates of the same tunnel section. The installation position, orientation angle, and reference point coordinates used for the transformation are all taken from the on-duty calibration record. After transformation, they are uniformly expressed along the tunnel extension direction, the transverse direction of the lining, and the direction from the normal to the lining surface, with the length unit being millimeters.

[0057] A unified time reference refers to converting the recording times of the three types of data to the segment operation times under the same clock. On-site, an industrial control computer can be used as the sole time source. Each record retains both the original timestamp and the converted timestamp. If the difference between the two exceeds 5 milliseconds, it is judged as clock drift and written into the fault record. A unified spatial coordinate reference refers to converting the measurement point positions, thermal image positions, and point cloud positions in the three types of data to the coordinates of the same target tunnel segment, which facilitates the establishment of multi-source correspondences for the next step according to the same segment.

[0058] During the initial processing, a 100-millisecond observation window was used as the smallest merging unit. It was allowed that one frame of thermal image or one frame of point cloud was missing in a single observation window and the missing measurement mark was retained. If two consecutive observation windows were missing, the completion was stopped and the window was marked as an unalignable area. For the same record key that arrived repeatedly, only the first successfully corrected version was retained. Subsequent duplicate records were only registered in the deduplication log to ensure consistent order and idempotent results.

[0059] Upon completion, a multi-source organized dataset, a time reference list, and a spatial reference list are generated after noise reduction and correction. The files and index list are appended and stored together in the organized subdirectory under the original segment directory, along with the rule version number, parameter version number, device version number, execution time, and execution subject, for the next step to directly call according to the segment number and observation window range.

[0060] Upstream and downstream transmission is carried out via vehicle Ethernet. The maximum delay for single-segment processing can be set to 3 seconds, and the number of concurrent segments can be set to 2. If disk writing fails, it will retry 2 times. If it still fails, it will switch to the backup directory and retain the incomplete mark.

[0061] Regarding on-site constraints, this step is applicable to sections where the inspection speed is stable, the equipment attitude can be recorded, and the lining surface is visible. It is not applicable to large-area water obstruction, continuous probe decoupling, or continuous radar lockout. In the event of the above situations, only the affected observation windows are frozen, while the remaining observation windows are retained to continue processing and recording the cause codes. The cause codes in the minimum feasible set can include time drift, temperature measurement mismatch, coordinate mismatch, write failure, and record duplication.

[0062] Preferably, under the conditions of a section length of 40 meters, an observation window of 100 milliseconds, an ultrasonic measuring point spacing of 20 millimeters, a thermal imaging frame rate of 30 frames per second, and a point cloud frequency of 10 Hz, a single section can obtain 1920 effective ultrasonic measuring points, 1186 effective thermal imaging frames, and 398 effective point cloud frames after noise correction. The time reference difference is controlled within 3 milliseconds, the spatial reference verification error is controlled within 4 millimeters, and the success rate of calling the next step reaches 99.5%. For example, a vehicle-mounted positioning combination can also replace the fixed installation reference to complete the coordinate transformation. As long as the same time source, the same spatial diameter, and the same version of the tracking method are maintained, equivalent implementation can be achieved.

[0063] S3. Based on a unified time reference and a unified spatial coordinate reference, register ultrasonic data, infrared thermal image data, and lidar point cloud data to establish multi-source correspondence relationships for the same segment. The specific implementation is as follows:

[0064] Based on a unified time reference and a unified spatial coordinate reference, registration is performed on ultrasonic data, infrared thermal imaging data, and lidar point cloud data to establish a multi-source correspondence relationship within the same section. In one embodiment, after noise reduction, correction, and coordinate transformation, the on-site industrial control computer retrieves and organizes ultrasonic data, infrared thermal imaging data, and lidar point cloud data from the subdirectories according to the section number and observation window order. Ultrasonic data includes the measurement point coordinates, arrival time, echo amplitude, and coupling state; infrared thermal imaging data includes the frame time, pixel temperature, lens attitude, and intra-frame position; and lidar point cloud data includes the point coordinates, reflection intensity, scanning time, and equipment pose. During registration, measurement points, thermal imaging frames, and point cloud frames with a time difference not exceeding 5 milliseconds are first screened within a 100-millisecond observation window according to the unified time reference. Then, the measurement point positions are mapped to the thermal imaging area and point cloud surface positions according to the unified spatial coordinate reference, forming candidate corresponding units within the same local area of ​​the lining. A candidate corresponding unit refers to a group of ultrasonic data within the same observation window whose position difference along the tunnel extension direction and lateral position difference both fall within a preset tolerance zone. Records, thermal imaging records, and point cloud records are used. The tolerance band is determined by the equipment installation accuracy, inspection speed, and lining surface undulation, and is fixed in the parameter version before operation. Then, each candidate corresponding unit is compared one by one in terms of time sequence, spatial proximity, and consistency of changes within a continuous window. The consistency of changes within a continuous window is obtained by comparing the movement of the measuring point position, the movement of the thermal imaging area center, and the movement of the local position of the point cloud in two adjacent observation windows. If the three changes are consistent in direction and the difference does not exceed the corresponding tolerance, it is confirmed as a multi-source correspondence in the same segment. If the same ultrasonic record corresponds to multiple thermal imaging records or multiple point cloud records, only the group with the smallest time difference and the shortest spatial distance is retained, and the remaining relationships are written into the deduplication log to ensure unique order and idempotent results. For observation windows that lack any source record, only the incomplete correspondence mark is registered, and the results of adjacent windows are not directly replaced. When two consecutive observation windows cannot form a complete correspondence, the registration of that local range is stopped and the previously confirmed relationship is retained to avoid mistakenly merging records from different time periods or different locations onto the same crack object.

[0065] Upon completion, a multi-source correspondence list for the same segment is generated. The list includes at least the segment number, observation window number, ultrasonic record number, thermal image record number, point cloud record number, time difference, spatial difference, confirmation status, rule version number, and execution time. This list, along with the index list, is appended to the registration subdirectory under the segment directory, allowing the next step to directly extract the crack characterization feature set based on the observation window range. The upstream and downstream are connected via vehicle-mounted Ethernet. The maximum registration delay for a single segment can be set to 2 seconds, and the number of concurrent segments can be set to 2. If the write fails, it will retry twice. If it still fails, it will be transferred to the backup directory and the reason code will be recorded. The reason code may include time mismatch, spatial mismatch, record duplication, and index write failure. This step is applicable to segments where the inspection speed fluctuation is controlled, the pose recording is continuous, and the lining surface is unobstructed.

[0066] Preferably, under the conditions of a segment length of 40 meters, an observation window of 100 milliseconds, a time tolerance of 5 milliseconds, and a spatial tolerance of 10 millimeters, 1850 sets of complete multi-source correspondences can be formed, and 70 sets of incomplete correspondence markers can be formed, with a success rate of 99.5% for the next step. For example, corresponding units can also be established first based on the point cloud surface position, and then the thermal image position and ultrasonic measurement point position can be checked in reverse. As long as the same time reference, the same spatial coordinate reference, and the same confirmation rule are maintained, equivalent registration can be achieved.

[0067] S4. Based on multi-source correspondence, extract ultrasonic depth response features, infrared thermal anomaly features, and surface morphology features to generate a crack characterization feature set. The specific implementation is as follows:

[0068] In one embodiment, the infrared thermal anomaly feature can be represented by the local temperature difference between the crack candidate region and the background region. Let the average temperature of the crack candidate region be... The average temperature of the background area is The local temperature difference It can be represented as:

[0069]

[0070] The candidate crack region is selected from the local thermal image area corresponding to the current ultrasonic measurement point location; the background region is selected from the lining surface area in a preset neighborhood outside the local thermal image area, excluding obstructed areas, reflective areas, and areas without measurement. Average temperature and All values ​​are obtained by averaging the corresponding pixel temperatures within the same thermal image frame. The pixel temperatures used are those after temperature measurement calibration and noise reduction. Local temperature differences are used to characterize the degree of thermal anomaly between the crack candidate area and the surrounding lining surface. When there are differences in heat exchange conditions near the crack, internal voids affecting the heat transfer path, or surface microcracks altering the local heat distribution, this value will shift relative to the background area. To avoid misjudgments caused by single-frame jitter, this application preferably uses values ​​obtained within a continuous observation window. They should be incorporated into subsequent assessments of the stability of thermal anomaly locations, rather than being determined solely based on the instantaneous values ​​of a single observation window.

[0071] In one embodiment, the infrared thermal anomaly feature can also be represented by the temperature gradient amplitude at the thermal anomaly boundary. Let the thermal image temperature field be... The temperature gradient magnitude at the thermal anomaly boundary It can be represented as:

[0072]

[0073] in, and These represent the horizontal and vertical positions of pixels in the current thermal image frame, respectively. and These represent the rates of change of the temperature field along two directions, which can be obtained by the temperature difference between the current pixel and its adjacent pixels. The larger the temperature gradient amplitude, the more concentrated the temperature change at that location, and the closer it is to the thermal anomaly boundary. In this application, this quantity is not used alone to determine whether a crack exists, but rather together with the local temperature difference to verify the range of the infrared thermal anomaly boundary. When the temperature gradient amplitude of the candidate region boundary remains concentrated and stable at the corresponding position within a continuous observation window, it indicates that the current thermal anomaly boundary is continuous and can be used for subsequent width verification and fusion determination.

[0074] In one embodiment, the surface topography feature can be represented by the amount of local depression. Let the local reference surface be the point cloud containing the first... The normal distance of each valid point is The total number of valid points is The amount of local depression It can be represented as:

[0075]

[0076] The local reference surface is obtained by fitting the point cloud corresponding to the continuous lining surfaces on both sides of the crack candidate region; the normal distance The vertical deviation of valid points in the point cloud from the local reference surface is taken from the denoised, corrected, and coordinate-transformed lidar point cloud data. The local concavity is used to characterize the degree of morphological deviation of the crack candidate region relative to the surrounding continuous lining surface; this value increases when openings, grooves, or continuous linear concavities appear on the surface. To avoid misjudgment due to isolated outliers, valid points participating in the calculation should be located within the local area corresponding to the current candidate group and have passed the point cloud coverage check; when there are insufficient valid points within the local area, this value is not directly calculated, but the missing marker is retained for subsequent fusion weight adjustment.

[0077] Based on a unified time and spatial coordinate reference, registration is performed on ultrasonic data, infrared thermal imaging data, and lidar point cloud data to establish a multi-source correspondence within the same segment. In one implementation, after completing the previous step of data processing, the on-site industrial control computer sequentially reads ultrasonic data, infrared thermal imaging data, and lidar point cloud data from the processing subdirectory under the segment directory, according to the segment number and observation window order. Specifically, the ultrasonic data reads the measurement point coordinates, echo amplitude, arrival time, sampling frequency, and coupling state; the infrared thermal imaging data reads the frame time, pixel temperature, lens attitude, and intra-frame position; and the lidar point cloud data reads the point coordinates, reflection intensity, scanning time, and... Equipment position; the unified time reference adopts the section operation time converted in the previous stage, and the unified spatial coordinate reference adopts the section coordinates along the tunnel extension direction, the lining transverse direction, and the lining normal direction. During registration, the converted timestamps of the three types of records are first compared within the same observation window. Records whose time differences fall within the preset tolerance zone are grouped into the same candidate group. Then, the positions of the ultrasonic measuring points in the candidate group are projected onto the infrared thermal imaging area and the lidar point cloud surface, and their extension direction position difference, transverse position difference, and normal position difference in the section coordinates are compared. Among them, the candidate group refers to a group of records that simultaneously meet the conditions of time proximity and spatial proximity within the same observation window. Time proximity is... The timestamp is obtained by comparing and converting within a 100-millisecond observation window. Spatial proximity is obtained by comparing the positions of the measuring points, the center position of the thermal imaging area, and the local center position of the point cloud within the same local lining area. The local lining area can be set as a rectangular area centered on the ultrasonic measuring point position. The size of the area is determined by the equipment installation accuracy, inspection speed, and lining undulation amplitude, and is written into the parameter version before operation. For each candidate group, it is further checked whether the positional changes in two consecutive adjacent observation windows are consistent. The positional changes are obtained by comparing the segmental coordinate displacement of the same candidate group in the preceding and following observation windows. If the directions of the ultrasonic measuring point displacement, the thermal imaging area center displacement, and the point cloud local center displacement are consistent, the positional changes are considered. If the difference does not exceed the tolerance, the multi-source correspondence of the candidate group in the same segment is confirmed. If the same ultrasound record matches multiple thermal images or multiple point cloud records at the same time, only one group is retained in the order of time difference first and then spatial difference. The remaining matching relationships are only registered in the deduplication log and do not participate in subsequent calls to ensure order stability and result idempotency. For incomplete candidate groups caused by missing frames in thermal images, missing frames in point clouds, or failure of coupling state, only the incomplete mark is retained and it is not directly replaced by the previous and next observation windows. When two consecutive observation windows cannot form a complete candidate group, the registration of the local range is stopped and a reason code is written. The reason code may include time mismatch, spatial mismatch, missing frames, duplicate records, and pose loss.

[0078] Upon completion, a multi-source correspondence list for the same segment is generated. The list must at least record the segment number, observation window number, ultrasonic record number, infrared thermal image record number, lidar point cloud record number, time difference, spatial difference, confirmation status, parameter version number, rule version number, and execution time. This list is appended to the registration subdirectory under the segment directory along with the index list and is then called by the next step in the order of segment number and observation window number. Upstream and downstream transmission is via vehicle Ethernet. The maximum registration delay for a single segment can be set to 2 seconds, and the number of concurrent segments can be set to 2. If the index writing fails, it will be retried twice. If it still fails, the backup directory will be switched and the incomplete mark will be retained. This step is applicable to segments where the inspection speed is continuously measurable, the equipment position is continuously obtainable, and the lining surface is unobstructed. When tunnel inspection images and location records are involved, they are only used locally in a closed loop and the original records are not made available to the outside.

[0079] Preferably, under the conditions of a section length of 40 meters, an observation window of 100 milliseconds, a time tolerance of 5 milliseconds, an extension direction tolerance of 10 millimeters, a lateral tolerance of 10 millimeters, and a normal tolerance of 8 millimeters, 1850 complete multi-source correspondences can be formed, along with 70 incomplete markers, and the success rate of calling the next step reaches 99.5%. For example, candidate groups can also be established first using the surface position of the lidar point cloud as a reference, and then the infrared thermal image position and ultrasonic measurement point position can be checked in reverse. As long as the same time reference, the same spatial coordinate reference, and the same deduplication order are maintained, equivalent registration can be achieved.

[0080] S5. Based on the crack characterization feature set and the quality of each source data, assign fusion weights, perform fusion judgment, and generate crack characterization results. The specific implementation is as follows:

[0081] Based on the crack characterization feature set and the fusion weights assigned to the data quality of each source, a fusion judgment is performed to generate crack characterization results. In one embodiment, after the on-site industrial control computer completes the writing of the crack characterization feature set, it reads the measurement point location, ultrasonic depth response characteristics, infrared thermal anomaly characteristics, surface morphology characteristics, and missing markers from the feature subdirectory according to the section number and observation window number. At the same time, it reads the coupling state, thermal image clarity, point cloud coverage, and pose continuity corresponding to the observation window. The data quality of each source is obtained separately within the same observation window. The coupling state is obtained by comparing the stability of continuous echoes and amplitude fluctuations in the ultrasonic recording. The thermal image clarity is obtained by comparing the continuity of the thermal image area boundary and the temperature jumps of adjacent pixels. The point cloud coverage is obtained by comparing the number of effective points and the length of the missing band within the local area. The pose continuity is obtained by comparing the change in device pose within the two observation windows. The thresholds, order, and retention rules used in the above comparisons are all written into the parameter version before the operation and are recorded along with the section operation record.

[0082] The fusion weight is determined separately for each source data quality within each observation window. Sources with higher data quality correspond to higher fusion weights, and sources with lower data quality correspond to lower fusion weights. When a source has a missing data flag, that source does not participate in the main judgment of this observation window; only the missing record is retained, and the overall confirmation level of that observation window is reduced. During fusion judgment, the ultrasonic depth response characteristics, infrared thermal anomaly characteristics, and surface morphology characteristics at the same measurement point are checked item by item to see if they meet the crack candidate conditions. Crack candidate conditions are determined by comparing whether the depth response change is continuous, the thermal anomaly location is stable, and the morphology change is continuous within two consecutive observation windows. If two of these conditions are met, the crack candidate conditions are considered to be met. If multiple measurement points meet the candidate conditions under the corresponding fusion weights, the measurement point location is recorded as a crack candidate location. Then, the continuity of adjacent measurement point locations is checked along the segment extension direction. If adjacent crack candidate locations are spatially continuous and extend in the same direction within the preceding and following observation windows, they are merged and confirmed as the same crack segment. If only a single source meets the candidate conditions, only a single-source prompt mark is registered, and it is not directly confirmed as a crack segment. If multiple overlapping crack candidate locations appear in the same observation window, a unique judgment result is retained in the order of first the overall fusion weight high and then the spatial continuity length short. The remaining results are written to the deduplication log to ensure order stability and result idempotency.

[0083] In one embodiment, the ultrasonic data quality, infrared thermal image data quality, and lidar point cloud data quality are respectively... , , The corresponding fusion weight , , It can be represented as:

[0084]

[0085] in, The results were obtained by comparing the stability and amplitude fluctuation of the continuous echoes recorded by ultrasound within the current observation window. This was obtained by comparing the continuity of the thermal imaging region boundary and the temperature jumps between adjacent pixels. The quality values ​​are obtained by comparing the number of valid points, the length of the missing band, and the continuity of the device pose within a local area. Before entering the normalization calculation, these three data quality values ​​should be converted to the same scoring caliber to ensure comparability between different sources. The normalized fusion weights sum to 1, controlling the contribution ratio of different source features in the overall judgment within the same observation window. When any source has a missing label, the corresponding quality value of that source does not participate in the current normalization summation; instead, the fusion weights are redistributed to the remaining valid sources, and the overall confirmation level is simultaneously reduced.

[0086] In one embodiment, the normalized ultrasonic depth response characteristic value, infrared thermal anomaly characteristic value, and surface morphology characteristic value are respectively... , , Then the comprehensive judgment value of the current measuring point location It can be represented as:

[0087]

[0088] in, It is obtained by comprehensively converting the delay of the stable reflection segment, the amplitude attenuation of adjacent reflection segments, and the change in echo shape. It is obtained by comprehensively converting local temperature differences, temperature gradient amplitude, and the stability of thermal anomaly locations; The result is obtained by comprehensively calculating the amount of local depression, linear continuous change, and surface undulation deviation. All the above characteristic values ​​are converted to the same comparison caliber before being included in this formula, so that they can participate in the calculation together in the same comprehensive judgment value. Comprehensive judgment value This is used to characterize the overall extent to which the current measuring point location simultaneously exhibits crack internal response, thermal anomaly manifestations, and surface morphology changes; when When the preset candidate threshold is reached and the corresponding feature remains stable within a continuous observation window, the location of the measuring point is recorded as a candidate crack location; when If the preset candidate threshold is not met, the location will not be recorded as a candidate crack location. The candidate threshold is determined by any one of the following: historical samples of similar tunnels, manually verified samples, or results from on-site test runs, and is fixed in the parameter version and recorded with the work log.

[0089] Upon completion, crack characterization results are generated, including at least the segment number, observation window number, measurement point location, crack candidate status, crack segment number, overall confirmation level, fusion weight of each source, missing identifier, rule version number, parameter version number, and execution time. These results are appended to the judgment subdirectory under the segment directory along with the index list, for subsequent stages to access in order of crack segment number and segment number. Upstream and downstream are connected via vehicle-mounted Ethernet. The maximum delay for single-segment fusion judgment can be set to 2 seconds, and the number of concurrent segments can be set to 2. If the index write fails, it will retry twice. If it still fails, it will switch to a backup directory and record the reason code. The reason code can include insufficient quality, interruption of continuity, duplicate record, and write failure.

[0090] This step applies to sections where at least two types of sensor recordings are available and the pose is continuous. It does not apply to local areas where three types of recordings are missing simultaneously or where two consecutive observation windows cannot form valid candidate conditions. When dealing with original tunnel images, poses, and section recordings, they are only used in local closed loops and the original content is not made available to the outside.

[0091] Preferably, under the conditions of a section length of 40 meters and an observation window of 100 milliseconds, when the ultrasonic depth response feature remains stable for two consecutive windows, the local temperature difference of the infrared thermal anomaly feature reaches more than 1.2 degrees Celsius, and the local depression of the surface morphology feature reaches more than 0.8 millimeters, 186 crack candidate locations can be formed, which are then merged to obtain 23 crack segments, and the success rate of the next step reaches 99.5%. For example, crack candidate locations can also be screened first according to surface morphology features, and then ultrasonic depth response features and infrared thermal anomaly features can be introduced to correct the overall confirmation level. As long as the crack characterization feature set, the quality of each source data, the fusion weight, and the verification rules of two consecutive observation windows are kept consistent, equivalent implementation can be achieved.

[0092] S6. Based on the crack characterization results, determine the crack location, direction, length, width, depth, and grade, and generate a crack detection record. The specific implementation is as follows:

[0093] Based on the crack characterization results, the crack location, direction, length, width, depth, and grade are determined, and a crack detection record is generated. In one embodiment, after completing the writing of the judgment sub-directory, the on-site industrial control computer reads the crack candidate status, crack segment number, overall confirmation grade, measuring point location, fusion weight of each source, and missing identifier in the order of segment number and crack segment number. It also checks back the ultrasonic depth response characteristics, infrared thermal anomaly characteristics, surface morphology characteristics, and multi-source correspondence list of the same segment within the corresponding observation window. The crack location is determined by the segment coordinates of the first and last measuring points of the crack segment in a unified spatial coordinate reference. The position along the tunnel extension direction, the transverse position of the lining, and the normal position are recorded in millimeters. The position value is obtained by comparing the center coordinates of adjacent candidate positions in two consecutive observation windows and combining them with the segment merging boundary. The direction is determined by the centerline formed by connecting the candidate positions in the same crack segment. The orientation is determined by comparing the changes in the centerline direction between adjacent measuring points and adjacent observation windows, and then taking the stable direction as the orientation of the segment. The length is determined by the segment coordinate distance between the first and last positions of the same crack segment. The width is obtained after joint verification of the local opening range corresponding to the surface morphology characteristics and the infrared thermal anomaly boundary range. The depth is obtained by segment mapping of the delay of the stable reflection segment in the ultrasonic depth response characteristics. The mapping relationship, segment boundaries, and verification order are all fixed to the parameter version before the operation and are left with the operation record. The grade is determined by the length, width, depth, and overall confirmation grade. Within the same crack segment, first check whether the depth exceeds the set boundary, then check whether the width and length exceed the corresponding boundaries, and finally combine the overall confirmation grade to determine the final grade. If any key quantity has a missing indicator, the grade is downgraded to a conservative grade and a verification mark is retained. A high grade judgment is not given directly.

[0094] In one embodiment, the delay of the stable reflection segment can be further used for crack depth calculation. Let the propagation speed of the ultrasonic wave in the current lining material be... The delay of the stable reflection segment relative to the launch point is Then the crack depth It can be represented as:

[0095]

[0096] Among them, the speed of transmission The acoustic propagation velocity in the lining material corresponding to the current target tunnel section can be determined and written into the parameter version by any one of the following: construction design data, calibration results of test blocks of the same material, or on-site echo calibration results; delay amount This time difference is calculated by comparing the time corresponding to the emission start point with the start time of the first stable reflection segment that meets the duration requirement within the same measurement point's dwell period. This time difference is taken from the denoised and corrected echo sequence. The "divide by 2" in the formula is used to eliminate the influence of the pulse round-trip propagation path on depth conversion, because the recorded delay includes both round-trip propagation times after the ultrasonic wave travels from the emission point to the crack reflection interface and back to the probe. To ensure consistent calculation accuracy, the propagation velocity used in depth conversion within the same section uses the same version of parameters and is not changed individually for each measurement point. When coupling fails or the stability of the continuous echo does not meet the requirements, this formula is not used to directly calculate the depth; instead, the missing identifier is used in the subsequent conservative judgment process.

[0097] In one embodiment, the crack length can be determined by the cumulative distance of discrete points along the crack centerline. Let the discrete points along the crack centerline be sequentially... Then the crack length It can be represented as:

[0098]

[0099] Each discrete point is obtained by connecting the candidate crack positions within the same crack segment in sequence according to the segment coordinates. The point coordinates are derived from the measurement point positions under a unified spatial coordinate reference. Compared with the simple straight-line distance between the first and last points, the cumulative distance segment can reflect the actual extension length of the crack segment when there are bends or deflections on the lining surface, and is more suitable for subsequent grade determination.

[0100] In one embodiment, let the projections of the first and last points of the crack centerline in the section coordinates be respectively... and Then the direction angle of the crack It can be represented as:

[0101]

[0102] in, Indicates the position coordinates along the tunnel's extension direction. This indicates the position coordinates of the lining in the transverse direction. The strike angle is used to characterize the degree of deflection of the crack segment relative to the tunnel extension direction. After comparing the changes in the centerline direction within a continuous observation window, the stable direction is taken as the strike of the crack segment.

[0103] To ensure continuity, this step uses the same time reference, spatial coordinate reference, missing identifier, rule version number, and parameter version number as the previous step. Only one official record is generated for the same crack segment. If the same segment number and crack segment number are calculated again later, the first official record is retained according to the first-come, first-served principle, and subsequent records are only written to the revision log to ensure order stability and idempotency of results. If two consecutive write attempts fail, the record is moved to the backup directory and a reason code is registered. The reason code may include segment breakage, depth missing, width conflict, record duplication, and index write failure.

[0104] Upon completion, a crack detection record is generated, which includes at least the section number, crack segment number, crack location, direction, length, width, depth, grade, overall confirmation grade, missing information, rule version number, parameter version number, and execution time. This record is appended to the record subdirectory under the section directory along with an index list for subsequent inspection and verification, defect record updates, and maintenance decision-making. The upstream and downstream connections are made via vehicle-mounted Ethernet. The maximum delay for generating a single section record can be set to 2 seconds, and the number of concurrent sections can be set to 2.

[0105] This step applies to sections that have formed crack segments and have at least locational continuity information and two types of valid features. It does not apply to local areas where the segment continuity cannot be maintained in two consecutive observation windows. When it involves original tunnel images, poses and section records, they are only saved locally in a closed loop and the original content is not made available to the outside.

[0106] Preferably, under the conditions of a section length of 40 meters and an observation window of 100 milliseconds, 23 crack detection records can be obtained, with crack lengths ranging from 120 mm to 1860 mm, widths ranging from 0.4 mm to 3.6 mm, and depths ranging from 8 mm to 62 mm. The final grade distribution is 9 low-grade, 10 medium-grade, and 4 high-grade, with a subsequent retrieval success rate of 99.5%. For example, the width can also be determined separately by the opening range of the lidar point cloud surface. As long as the same section, unified spatial coordinate reference, and the same grade judgment boundary are maintained, equivalent implementation can be achieved.

[0107] In the operational scenario shown in this embodiment: taking the secondary lining inspection operation of the left-line exit section of an operating highway tunnel as an example, after closing a single lane at night, the maintenance unit uses a tunnel inspection vehicle equipped with ultrasonic probes, infrared thermal imagers, lidar, and on-site industrial control computers to perform crack detection on the target tunnel section. The target tunnel section is a continuous lining section of 40 meters, which is pre-assigned by the work order. The inspection vehicle moves at a constant speed of 0.3 meters per second along the tunnel extension direction. Before the operation begins, on-site personnel complete the unified clock synchronization, installation position verification, and zero-point calibration of the ultrasonic probes, infrared thermal imagers, and lidar, and write the equipment number, parameter version number, rule version number, executor, and execution time into the operation record.

[0108] After the inspection begins, the ultrasonic probe continuously performs coupled emission and echo acquisition on the lining surface at a 20 mm measurement point spacing. The infrared thermal imager acquires temperature distribution images of the lining surface at 30 frames per second, and the lidar acquires a set of spatial points on the lining surface at 10 Hz. The on-site industrial control computer collects the echo sequence, thermal image frames, and point cloud frames formed in the same window in parallel with a 100 millisecond observation window to establish a multi-source raw dataset for the corresponding inspection section. When local water seepage and reflection occur in the middle of the section during the inspection, the infrared thermal image data will be missing in the current observation window. The on-site industrial control computer allows one supplementary acquisition. If the data is still not complete after the supplementary acquisition, the observation window will be marked as a gap and the original mark will be retained. At the same time, the ultrasonic data and lidar point cloud data will continue to be retained to avoid interruption of the entire operation.

[0109] After the section data acquisition is completed, the on-site industrial control computer immediately retrieves the multi-source raw dataset from the local solid-state medium. It performs peak removal and arrival time correction on the ultrasonic data, instantaneous hot spot removal and pixel temperature correction on the infrared thermal image data, and outlier removal and point coordinate correction on the lidar point cloud data. Based on the on-duty calibration record, the measurement point position, thermal image position, and point cloud position are uniformly converted to the same target tunnel section coordinates, forming a multi-source processed dataset, time reference list, and spatial reference list after noise correction. Subsequently, according to the unified time reference and unified spatial coordinate reference, ultrasonic records, thermal image records, and point cloud records with a time difference of no more than 5 milliseconds and a spatial position difference within the tolerance zone are screened out in each 100-millisecond observation window to establish a multi-source correspondence list for the same section.

[0110] For the confirmed multi-source correspondence, the on-site industrial control computer further extracts ultrasonic depth response features, infrared thermal anomaly features, and surface morphology features within the same local area of ​​the same measuring point. Specifically, within the range of 24.36 meters to 25.82 meters, ultrasonic records in multiple consecutive observation windows show a delayed stable reflection segment, infrared thermal images show local temperature differences higher than the background band, and lidar point cloud records show continuous linear depressions, thus forming a crack characterization feature set corresponding to this section. The on-site industrial control computer then assigns fusion weights based on the coupling state, thermal image clarity, point cloud coverage, and pose continuity within the same observation window, and performs fusion judgment on the three types of features at the same measuring point location. When at least two types of features simultaneously meet the crack candidate conditions in two consecutive observation windows, the measuring point location is recorded as a crack candidate location, and the continuity of adjacent candidate locations is checked along the extension direction of the section. In this embodiment, a total of 12 continuous crack candidate locations were confirmed within the range of 24.36 meters to 25.82 meters, and were finally merged into one crack segment.

[0111] For this crack segment, the on-site industrial control computer reviewed the ultrasonic depth response characteristics, infrared thermal anomaly characteristics, surface morphology characteristics, and multi-source correspondence within the corresponding observation window. It determined that the crack location was within the range of 24.36 meters to 25.82 meters in the target tunnel section, with a small angle between its direction and the tunnel extension direction. It was identified as a near-longitudinal extension crack along the lining surface, with a length of 1460 mm. The width was determined to be 1.8 mm based on the local opening range corresponding to the surface morphology characteristics and after verification with the infrared thermal anomaly boundary. The depth was determined to be 34 mm based on the mapping of the delay of the stable reflection section. Combined with the overall confirmation level, its final level was determined to be medium, and a corresponding crack detection record was generated.

[0112] The crack detection record, along with the section number, crack segment number, crack location, direction, length, width, depth, grade, missing information, rule version number, parameter version number, and execution time, is written to the record subdirectory under the section directory for subsequent inspection and verification, disease ledger updates, and maintenance decision-making. If the same section number and crack segment number are obtained through subsequent reruns or repeated writing, only the first official record will be retained, and subsequent records will be written to the revision log.

[0113] Through on-site operations in the 40-meter target tunnel section, a total of 23 crack detection records were generated, including 9 low-level, 10 medium-level, and 4 high-level records. After the records were generated, the maintenance unit manually reviewed 5 of them. The review results were consistent with the crack location and level conclusions, indicating that the implementation process can stably complete the entire closed loop from multi-source acquisition, unified processing, multi-source registration, feature extraction, fusion judgment to crack detection record generation in the operating tunnel site, and can reproduce the detection results consistent with the technical concept of this invention.

[0114] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for detecting tunnel cracks based on multi-sensor fusion, characterized in that, include: S1. Collect ultrasonic data, infrared thermal image data, and lidar point cloud data of the target tunnel section to establish a multi-source raw dataset for the corresponding detection section. S2. Perform denoising, correction, and coordinate transformation processing on the multi-source raw dataset to establish a unified time reference and a unified spatial coordinate reference; S3. Based on a unified time reference and a unified spatial coordinate reference, register ultrasonic data, infrared thermal image data, and lidar point cloud data to establish a multi-source correspondence relationship for the same segment; wherein, the registration is performed using a criterion formula. and Filter out spatiotemporally correlated observations, where: The timestamp of the sensor data to be registered. As the base timestamp, Preset time tolerance; The spatial coordinates of the data to be registered. As the reference spatial coordinate point, Preset spatial tolerance zone; S4. Extract ultrasonic depth response features, infrared thermal anomaly features, and surface morphology features based on multi-source correspondence to generate a crack characterization feature set; S5. Based on the crack characterization feature set and the quality of each source data, assign fusion weights, perform fusion judgment, and generate crack characterization results. S6. Based on the crack characterization results, determine the crack location, direction, length, width, depth, and grade, and generate a crack detection record.

2. The tunnel crack detection method based on multi-sensor fusion according to claim 1, characterized in that, S1 includes: Before data collection, the ultrasonic probe, infrared thermal imager, and lidar were synchronized with a unified clock, their installation positions were verified, and their zero points were calibrated. During the acquisition process, echo sequences, thermal images, and point cloud frames are grouped side-by-side using an observation window. After the section data acquisition is completed, a multi-source raw dataset is generated, which includes a measurement point table, thermal image frame set, point cloud frame set, equipment status record, calibration record, and operating environment record.

3. The tunnel crack detection method based on multi-sensor fusion according to claim 1, characterized in that, S2 include: Retrieve ultrasonic data, infrared thermal imaging data, and lidar point cloud data; Noise reduction and calibration were performed based on the equipment calibration results. Complete coordinate transformation based on installation location, orientation angle, and reference point coordinates; The recording time of various data is converted into the section operation time, and the measurement point position, thermal image position, and point cloud position are converted into the same target tunnel section coordinates.

4. The tunnel crack detection method based on multi-sensor fusion according to claim 1, characterized in that, S3 includes: Based on a unified time reference, ultrasonic records, infrared thermal images, and lidar point cloud records located within the same observation window were selected. Based on a unified spatial coordinate reference, the position of the measurement point is mapped to the thermal imaging area and the position of the point cloud surface to form candidate corresponding units; The correspondence between multiple sources in the same segment is confirmed based on the chronological order, spatial proximity, and consistency of changes within a continuous window.

5. The tunnel crack detection method based on multi-sensor fusion according to claim 1, characterized in that, S4 include: Based on the converted timestamps, ultrasonic records, infrared thermal images, and lidar point cloud records located in the same observation window are grouped into candidate groups. Based on the segment coordinates, the measurement point location, the center location of the thermal imaging area, and the local center location of the point cloud, candidate groups with spatial proximity are selected. The correspondence between multiple sources was confirmed based on the consistency of the positional changes of the observation windows before and after.

6. The tunnel crack detection method based on multi-sensor fusion according to claim 1, characterized in that, S5 includes: The quality of ultrasonic data is determined based on the coupling state, the quality of infrared thermal image data is determined based on the clarity of the thermal image, the quality of lidar point cloud data is determined based on the point cloud coverage and pose continuity, and the corresponding fusion weights are determined based on the quality of each source data.

7. The tunnel crack detection method based on multi-sensor fusion according to claim 6, characterized in that: The fusion weight is dynamically generated based on the quality evaluation factors of each source data, and the calculation formula is as follows: ,in: It is the fusion weight of the i-th type of sensor; It is an ultrasonic quality evaluation factor; It is an infrared quality evaluation factor; It is a point cloud quality evaluation factor.

8. A tunnel crack detection method based on multi-sensor fusion according to claim 6 or 7, characterized in that: Based on the fusion weights, the ultrasonic depth response characteristics, infrared thermal anomaly characteristics, and surface morphology characteristics at the same measuring point are used to check the crack candidate conditions. The locations of measuring points that meet the crack candidate conditions are determined as crack candidate locations; Crack segments are determined based on the spatial continuity of adjacent candidate crack locations and the unidirectional extension relationship within the front and rear observation windows.

9. A tunnel crack detection method based on multi-sensor fusion according to claim 1, characterized in that, S6 include: The location and length of the crack are determined based on the segment coordinates of the first and last measuring points of the crack segment. The direction of the crack is determined based on the direction of the center line connecting the candidate locations. The width of the crack is determined based on the surface morphology and infrared thermal anomaly characteristics. The depth of the crack is determined based on the ultrasonic depth response characteristics. The crack grade is determined based on the length, width, depth and overall confirmation grade.

Images