A sonar point cloud outlier detection method of adaptive multi-scale attention mechanism

Abstract

The application discloses a sonar point cloud outlier detection method based on an adaptive multi-scale attention mechanism, which comprises the following steps: collecting original sonar data, preprocessing the original sonar data, and generating standardized sonar point cloud data; constructing an acoustic propagation field tensor, calculating a propagation direction feature, and forming a multi-scale coherence domain; the propagation direction feature acts on each coherence domain, cross-scale thrust flow response calculation is performed, and propagation evolution data is formed; taking propagation consistency as input, three branches are constructed respectively, and three types of depth prediction results are obtained; three types of consistency deviation are constructed according to the three types of depth prediction results and the original depth, and an abnormal score is generated; an abnormal point judgment threshold is set, and an abnormal point detection result is output, and the abnormal point is removed. Through the construction of the acoustic propagation field tensor and the combination of the multi-scale attention and the three-branch consistency reasoning, high-precision and adaptive detection of complex abnormal points in the sonar point cloud is realized.

Description

Technical Field

[0001] This invention relates to the field of sonar depth sounding data processing technology, and in particular to a sonar point cloud outlier detection method based on an adaptive multi-scale attention mechanism. Background Technology

[0002] Sonar bathymetry equipment is widely used to acquire three-dimensional point cloud data of the underwater environment in applications such as marine mapping, seabed topography surveying, and underwater engineering monitoring. Existing sonar point cloud processing methods typically rely on geometric rules, filtering thresholds, or outlier removal strategies based on statistical characteristics to identify anomalies caused by water disturbance, sound wave scattering, multipath reflection, and changes in equipment attitude. However, most of these methods are based on shallow geometric features or local consistency analysis and lack the ability to model acoustic characteristics such as refraction effects, multipath superposition, and angular artifacts generated during sound wave propagation, making it difficult to accurately distinguish between real seabed structures and complex acoustic noise. Furthermore, traditional methods are not adaptable to terrain scenarios at different scales. When the seabed structure presents slopes, steps, or mixed terrain, it often results in the accidental deletion of real terrain points or the omission of weak anomalies.

[0003] With the development of deep learning technology, some studies have attempted to apply graph structures, convolutional networks, or global feature extraction methods to sonar point cloud anomaly detection. However, these methods still have significant limitations. First, existing models are generally built based on visual frameworks, failing to reflect the physical laws of sound wave propagation, resulting in insufficient robustness when faced with changes in sound velocity gradients or curved sound wave propagation paths. Second, most models only utilize single-scale or fixed-scale feature representations, lacking cross-scale correlation mechanisms, making it difficult to balance fine-scale point cloud noise with large-scale seabed structural changes. Third, current methods mostly rely on a single prediction result or a single feature space for anomaly detection, lacking the ability to comprehensively evaluate the rationality of point clouds from geometric, acoustic physics, and propagation field perspectives.

[0004] Existing technologies struggle to achieve high-precision, adaptive, and physically consistent comprehensive identification of sonar point cloud anomalies, especially in complex scenarios such as multipath reflection, acoustic refraction, strip artifacts, and weak echo anomalies.

[0005] Therefore, how to provide an adaptive multi-scale attention mechanism for sonar point cloud outlier detection is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose an adaptive multi-scale attention mechanism for sonar point cloud outlier detection. This invention utilizes techniques such as acoustic propagation field tensor modeling, multi-scale coherence domain construction, cross-scale thrust flow attention inference, and geometric-acoustic-propagation field three-branch consistency determination to automatically identify and remove outliers in sonar point clouds caused by multipath reflection, sound wave refraction, stripe artifacts, and weak echo interference. This achieves a deep integration of acoustic physical constraints and data-driven inference. This invention possesses advantages such as strong adaptability to complex acoustic noise, high anomaly identification accuracy, and excellent cross-scene robustness, effectively improving the quality and reliability of sonar bathymetry data and providing a more stable technical foundation for marine mapping and underwater target perception.

[0007] A method for detecting outliers in sonar point clouds using an adaptive multi-scale attention mechanism according to an embodiment of the present invention includes:

[0008] Collect raw sonar data output from the sonar depth sounding equipment, preprocess the raw sonar data, and generate standardized sonar point cloud data;

[0009] An acoustic propagation field tensor is constructed based on standardized sonar point cloud data. Gradient calculations are performed to obtain the characteristics of the sound wave propagation direction. Based on the acoustic propagation field tensor, fine-scale coherence domains, mesoscale coherence domains, and coarse-scale coherence domains are formed.

[0010] The sound wave propagation direction features are applied to the fine-scale coherence domain, the mesoscale coherence domain, and the coarse-scale coherence domain respectively. Cross-scale thrust flow response calculation is performed to construct cross-scale propagation evolution data. Acoustic and terrain coupled cross-scale attention inference is performed on the cross-scale propagation evolution data to generate a cross-scale propagation consistency representation.

[0011] Using cross-scale propagation consistency representation as input, geometric reconstruction branch, acoustic physics branch and propagation field consistency branch are constructed respectively to predict geometrically reasonable depth, acoustically reasonable depth and propagation field consistent depth, and obtain three types of depth prediction results;

[0012] Based on the three types of depth prediction results and the original depth in the standardized sonar point cloud data, geometric consistency deviation, acoustic consistency deviation and propagation field consistency deviation are constructed respectively, and anomaly scores are generated for the corresponding points.

[0013] Set an anomaly detection threshold, identify points with anomaly scores greater than the threshold as anomalies in the sonar point cloud, remove them from the standardized sonar point cloud data, and output the anomaly detection results.

[0014] Optionally, the raw sonar data includes point coordinate data, Ping number data, beam angle data, echo intensity data, signal-to-noise ratio data, sound velocity profile data, and attitude information data.

[0015] Optionally, the preprocessing of the raw sonar data includes performing sound velocity correction, attitude compensation, and time synchronization processing on the raw sonar data.

[0016] Optionally, the step of forming fine-scale coherence domains, mesoscale coherence domains, and coarse-scale coherence domains based on the acoustic propagation field tensor includes:

[0017] An acoustic propagation field tensor index system is established on the Ping sequence dimension, propagation path discrete dimension, and beam angle discrete dimension of standardized sonar point cloud data. Acoustic propagation intensity and propagation direction information are calculated for each index position, and sound wave propagation direction features are generated from the propagation direction information.

[0018] A coherent seed set is selected from the index system. The coherent seeds meet the following requirements: the signal-to-noise ratio is not lower than the preset lower limit, the propagation direction is stable within the local window and the difference between the direction of the propagation seed and the direction of the adjacent index position does not exceed the preset angle threshold. Fine-scale candidates are defined according to the local window size, medium-scale candidates are defined according to the medium window size, and coarse-scale candidates are defined according to the large window size.

[0019] Starting with a coherent seed, perform triaxial consistent region growth on each of the three discrete axes:

[0020] The propagation direction remains continuous in the Ping direction, and the propagation intensity does not change abruptly.

[0021] The propagation path maintains monotonous progression without interruption of energy.

[0022] Maintaining angular adjacency and limiting directional differences in the beam angle direction;

[0023] Preliminary regions for fine-scale coherence domains, mesoscale coherence domains, and coarse-scale coherence domains are generated respectively.

[0024] Cross-scale conflict resolution and closure correction are performed on three types of preliminary regions: when there is overlap or boundary conflict between adjacent scales, the region is assigned according to the rules of prioritizing long-distance continuity, prioritizing medium-scale slope transition and prioritizing fine-scale edge preservation. Low-area isolated patches are removed, fracture boundaries are corrected, and voids within the allowable range are filled to obtain the final regions of fine-scale coherence domain, medium-scale coherence domain and coarse-scale coherence domain.

[0025] Optionally, generating a cross-scale propagation consistency representation includes:

[0026] In the fine-scale coherence domain, the propagation direction, the propagation path direction, and the beam angle direction are all propagated in the domain according to the characteristics of the sound wave propagation direction. The point set that satisfies the continuous propagation direction, the no sudden change in propagation intensity, and the angular neighbor acceptance limit is accumulated in a consistent direction to generate a fine-scale thrust flow response record. In the mesoscale coherence domain and the coarse-scale coherence domain, the same rules are followed to perform propagation and accumulation to generate mesoscale thrust flow response records and coarse-scale thrust flow response records, respectively.

[0027] The fine-scale thrust flow response records, meso-scale thrust flow response records, and coarse-scale thrust flow response records are aligned in time-path-angle order according to Ping order, propagation path order, and beam angle number. Incomplete segments are removed and gaps within the allowable range are interpolated to form cross-scale propagation evolution data.

[0028] Using cross-scale propagation evolution data as input, a cross-scale selection map is generated based on the rate of change of sound velocity profile, the rate of change of echo intensity, and the continuity and directional stability of local terrain slope. Attention control factors are assigned to the fine-scale coherence domain, the meso-scale coherence domain, and the coarse-scale coherence domain, respectively. A conflict suppression gate is enabled to shield coherence domain segments with long-distance discontinuities. The cross-scale attention results of acoustic and terrain coupling are output.

[0029] Based on the cross-scale attention results, weighted integration and position alignment are performed among the thrust flow response records in the fine-scale coherence domain, the mesoscale coherence domain, and the coarse-scale coherence domain to generate a cross-scale propagation consistency representation.

[0030] Optionally, the three types of depth prediction results obtained include:

[0031] Receive the cross-scale propagation consistency representation and the acoustic propagation field tensor, establish a three-dimensional index consistent with the Ping order, propagation path order, and beam angle number, and locate the set of points to be processed;

[0032] Construct a geometric reconstruction branch: In the fine-scale coherence domain, mesoscale coherence domain, and coarse-scale coherence domain, perform intra-domain assembly in the order of edge priority, continuity priority, and step preservation. Generate isobath segments, monotonically splice along Ping and the propagation path, and stitch adjacent segments together to close the consistent representation of cross-scale propagation, and output a geometrically reasonable depth sequence.

[0033] Construct an acoustic physics branch: Based on cross-scale propagation consistency representation and sound velocity profile index, beam angle index, and propagation path index, perform path advancement of layered sound velocity segments, angular gating of incident angle constraints, energy gating of echo attenuation constraints, and candidate elimination of multipath suppression, and output an acoustically reasonable depth sequence.

[0034] Constructing a propagation field consistency branch: Taking the acoustic propagation field tensor as input, the in-field trajectory search, coherence domain boundary alignment and long-distance connectivity verification are performed according to the trajectory following rules of the thrust flow direction field. The cross-scale propagation consistency representation is then localized by scene intersection and a propagation field consistency depth sequence is generated.

[0035] Alignment is performed on the three types of output depth sequences according to the three-dimensional index. Segment-level consistent alignment is completed based on the combination relationship of edge preservation markers, propagation continuity markers, and field orientation alignment markers to form three types of depth prediction results.

[0036] Optionally, generating the anomaly score for the corresponding point includes:

[0037] A three-dimensional index corresponding to the original depth is established for the geometric depth prediction results, acoustic physical depth prediction results and propagation field consistent depth prediction results, and the position of each sounding point is kept consistent on the Ping sequence, propagation path sequence and beam angle sequence.

[0038] Constructing geometric consistency deviation: Voxel windows are formed in three directions with each sounding point as the center. The geometric depth prediction results within the window are compared with the original depth point by point. If the difference is within the slope smoothing threshold and the edges of adjacent points remain continuously marked, it is recorded as zero deviation. Otherwise, the difference is jointly mapped to a stepped geometric deviation grade value according to the magnitude of the difference and the edge abruptness level.

[0039] Constructing acoustic consistency deviation: Within the same voxel window, the acoustic physical depth prediction result is compared with the original depth by difference. At the same time, the rate of change of sound velocity profile, the magnitude of change of incident angle, and the magnitude of echo intensity attenuation at the center of the voxel window are retrieved. If the three acoustic indicators are all in the stable range and the difference does not exceed the acoustic tolerance threshold, it is recorded as zero deviation. Otherwise, the penalty coefficient of the acoustic indicators is added according to the difference to map to a multi-level acoustic deviation value.

[0040] Constructing propagation field consistency deviation: Track the propagation field trajectory along the index dimension from the starting point to the end of the sound wave propagation direction, compare the difference between the propagation field consistency depth prediction result and the original depth, and determine the deviation level based on the propagation direction alignment mark and connectivity integrity check result. When the trajectory is connected and the direction is aligned, and the difference is within the permissible range, it is recorded as zero deviation; otherwise, the difference and connectivity breakage penalty are jointly mapped to the graded propagation field deviation value.

[0041] The geometric deviation, acoustic deviation, and propagation field deviation values ​​are respectively normalized to a percentage range with unified dimensions, and the three types of normalized deviations are weighted and summarized to generate an anomaly score.

[0042] Optionally, the output anomaly detection results include:

[0043] Receive the anomaly score and establish a three-dimensional index mapping for the anomaly score corresponding to each point according to the Ping number index, propagation path index and beam angle index;

[0044] Local stability detection is performed on abnormal scores along the Ping number direction to filter out isolated score peak segments caused by single-point jumps. Continuity detection is performed on the propagation path direction to identify abnormal gaps between score segments. Angular domain smoothing is performed on the beam angle direction to eliminate beam-level random fluctuations, forming a score set corrected for three-dimensional consistency.

[0045] Based on the scoring set after three-dimensional consistency correction, an adaptive threshold set is generated according to regional noise density, coherence domain category and local connectivity state, and independent outlier evaluation thresholds are determined for fine-scale coherence domain, mesoscale coherence domain and coarse-scale coherence domain respectively.

[0046] Each score in the abnormal score sequence is compared with the corresponding abnormal point evaluation threshold point by point. Points with scores higher than the abnormal point evaluation threshold are marked as abnormal points, and points with scores not higher than the abnormal point evaluation threshold are marked as normal points, thus generating an abnormal point label set in three-dimensional space.

[0047] Output the anomaly detection results of the sonar point cloud based on the anomaly marker set, and record the anomaly locations corresponding to the Ping number, propagation path index, and beam angle index.

[0048] The beneficial effects of this invention are:

[0049] This invention introduces an acoustic propagation field tensor into sonar point cloud processing to model the propagation path, direction, and energy evolution of sound waves in water. This allows outlier detection to no longer rely solely on geometric spatial relationships but simultaneously consider acoustic physics factors such as sound wave refraction, multipath propagation, and echo attenuation, thereby significantly improving the ability to identify complex acoustic noise. Compared to traditional methods that rely on fixed neighborhoods or geometric statistical features, this invention exhibits higher physical consistency and stability when handling scenarios such as multipath false echoes, depth shifts caused by refraction, and angular stripe artifacts.

[0050] This invention constructs fine-scale, mesoscale, and coarse-scale coherence domains, and then performs cross-scale thrust flow response calculation and acoustic-terrain coupled cross-scale attention inference on this basis. This enables the system to adapt to different terrain scales and acoustic environment changes, automatically selecting the most reasonable scale information for deep inference. This mechanism effectively improves the generalization ability of anomaly detection under different seabed topographic conditions, avoiding the phenomenon of erroneous deletion of real terrain points or missed detection of weak anomalies in sloping terrain and complex structural areas, while also possessing cross-scenario robustness.

[0051] This invention constructs a three-branch consistency reasoning system consisting of a geometric reconstruction branch, an acoustic physics branch, and a propagation field consistency branch. This system achieves a comprehensive determination of the rationality of sonar point cloud depth from three independent dimensions: geometric features, acoustic physics features, and propagation field structure. It overcomes the limitations of existing technologies that rely on a single reasoning result, improves the accuracy of anomaly identification, effectively reduces the false detection rate, and makes the output anomaly points more reliable. Overall, this invention achieves high-precision, robust, and adaptive detection of sonar point cloud anomalies, significantly improving the quality of marine mapping data and the reliability of subsequent applications. Attached Figure Description

[0052] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0053] Figure 1 This is a flowchart of a sonar point cloud outlier detection method based on an adaptive multi-scale attention mechanism proposed in this invention.

[0054] Figure 2 This is a schematic diagram of the three-branch deep inference structure of the sonar point cloud outlier detection method with an adaptive multi-scale attention mechanism proposed in this invention. Detailed Implementation

[0055] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0056] refer to Figure 1 and Figure 2 An adaptive multi-scale attention mechanism for outlier detection in sonar point clouds includes:

[0057] Collect raw sonar data output from the sonar depth sounding equipment, preprocess the raw sonar data, and generate standardized sonar point cloud data;

[0058] An acoustic propagation field tensor is constructed based on standardized sonar point cloud data. Gradient calculations are performed to obtain the characteristics of the sound wave propagation direction. Based on the acoustic propagation field tensor, fine-scale coherence domains, mesoscale coherence domains, and coarse-scale coherence domains are formed.

[0059] The sound wave propagation direction features are applied to the fine-scale coherence domain, the mesoscale coherence domain, and the coarse-scale coherence domain respectively. Cross-scale thrust flow response calculation is performed to construct cross-scale propagation evolution data. Acoustic and terrain coupled cross-scale attention inference is performed on the cross-scale propagation evolution data to generate a cross-scale propagation consistency representation.

[0060] Using cross-scale propagation consistency representation as input, geometric reconstruction branch, acoustic physics branch and propagation field consistency branch are constructed respectively to predict geometrically reasonable depth, acoustically reasonable depth and propagation field consistent depth, and obtain three types of depth prediction results;

[0061] Based on the three types of depth prediction results and the original depth in the standardized sonar point cloud data, geometric consistency deviation, acoustic consistency deviation and propagation field consistency deviation are constructed respectively, and anomaly scores are generated for the corresponding points.

[0062] Set an anomaly detection threshold, identify points with anomaly scores greater than the threshold as anomalies in the sonar point cloud, remove them from the standardized sonar point cloud data, and output the anomaly detection results.

[0063] In this embodiment, the raw sonar data includes point coordinate data, Ping number data, beam angle data, echo intensity data, signal-to-noise ratio data, sound velocity profile data, and attitude information data.

[0064] In this embodiment, the preprocessing of the raw sonar data includes performing sound velocity correction, attitude compensation, and time synchronization processing on the raw sonar data.

[0065] In this embodiment, the step of forming fine-scale coherence domains, mesoscale coherence domains, and coarse-scale coherence domains based on the acoustic propagation field tensor includes:

[0066] An acoustic propagation field tensor indexing system is established on the Ping sequence dimension, propagation path discrete dimension, and beam angle discrete dimension of standardized sonar point cloud data. For each index position, acoustic propagation intensity and propagation direction information are calculated, and sound wave propagation direction features are generated from the propagation direction information. Specifically, the calculation of acoustic propagation intensity and propagation direction information involves:

[0067] At the moment of each Ping's transmission, the sound velocity profile value and beam direction angle of the corresponding depth layer are read, and the starting direction vector of the first propagation unit is determined based on the initial direction of the ray and the local sound velocity gradient.

[0068] The propagation direction vector and propagation distance are updated sequentially for each step along the propagation path at a fixed step size. The local propagation intensity and direction after each step are obtained by combining the water absorption coefficient, interface reflection loss and geometric diffusion factor cumulative attenuation.

[0069] When the propagation reaches the boundary of the coherence domain or the echo energy is below the threshold, the recursion stops and the final propagation intensity value and normalized direction vector at the index position are recorded as the acoustic propagation intensity and sound wave propagation direction characteristics of the corresponding cell of the acoustic propagation field tensor.

[0070] A coherent seed set is selected from the indexing system. The coherent seeds satisfy the following conditions: signal-to-noise ratio not lower than a preset lower limit; propagation direction stable within a local window; and direction difference from adjacent index positions not exceeding a preset angle threshold. Fine-scale candidates are defined according to the local window size, medium-scale candidates according to the medium window size, and coarse-scale candidates according to the large window size.

[0071] The preset lower limit is 15dB, the preset angle threshold is 6°, the local window size is 3×3×3, the medium window size is 7×7×7, and the large window size is 11×11×11.

[0072] Starting with a coherent seed, perform triaxial consistent region growth on each of the three discrete axes:

[0073] The propagation direction remains continuous in the Ping direction, and the propagation intensity does not change abruptly.

[0074] The propagation path maintains monotonous progression without interruption of energy.

[0075] Maintaining angular adjacency and limiting directional differences in the beam angle direction;

[0076] Preliminary regions for fine-scale coherence domains, mesoscale coherence domains, and coarse-scale coherence domains are generated respectively.

[0077] Cross-scale conflict resolution and closure correction are performed on three types of preliminary regions: when there is overlap or boundary conflict between adjacent scales, the region is assigned according to the rules of prioritizing long-distance continuity, prioritizing medium-scale slope transition and prioritizing fine-scale edge preservation. Low-area isolated patches are removed, fracture boundaries are corrected, and voids within the allowable range are filled to obtain the final regions of fine-scale coherence domain, medium-scale coherence domain and coarse-scale coherence domain.

[0078] In this embodiment, generating a cross-scale propagation consistency representation includes:

[0079] In the fine-scale coherence domain, propagation is performed along the Ping direction, propagation path direction, and beam angle direction according to the characteristics of sound wave propagation direction. Point sets satisfying continuous propagation direction, no abrupt changes in propagation intensity, and angular adjacency acceptance limits are accumulated with consistent direction to generate fine-scale thrust flow response records. Propagation and accumulation are performed according to the same rules in the mesoscale and coarse-scale coherence domains to generate mesoscale and coarse-scale thrust flow response records, respectively. Specifically, the intra-domain propagation is performed as follows:

[0080] Starting with the coherent seed as the index, search frame by frame along the Ping direction. If the angle between the propagation directions of consecutive frames is less than the set angle difference threshold and the change in propagation intensity is within the set intensity difference range, then the frame is included in the current path and continues to move forward.

[0081] On the verified Ping sequence, the propagation path index is expanded step by step from near to far. If the directional consistency and intensity smoothness of adjacent path units meet the requirements at the same time, the unit is added in the path direction; otherwise, the path advancement is terminated at the current position.

[0082] Under the same Ping and the same path index, adjacent beam elements are checked sequentially along the positive and negative directions of the beam angle. If the directions are consistent and the energy difference does not exceed the set limit, they are merged into the current accumulation set. Otherwise, they are considered as boundary stop angular expansion, completing the intra-domain propulsion and recording the thrust flow response.

[0083] The fine-scale thrust flow response records, meso-scale thrust flow response records, and coarse-scale thrust flow response records are aligned in time-path-angle order according to Ping order, propagation path order, and beam angle number. Incomplete segments are removed and gaps within the allowable range are interpolated to form cross-scale propagation evolution data.

[0084] Using cross-scale propagation evolution data as input, a cross-scale selection map is generated based on the rate of change of sound velocity profile, the rate of change of echo intensity, and markers of local topographic slope continuity and directional stability. Attention control factors are assigned to the fine-scale coherence domain, the mesoscale coherence domain, and the coarse-scale coherence domain, respectively. A conflict suppression gate is enabled to shield coherence domain segments with long-distance discontinuities. The output is the cross-scale attention result of acoustic and topographic coupling, where:

[0085] Generate a cross-scale selection map, specifically as follows:

[0086] The cross-scale propagation evolution data is unfolded into a continuous frame sequence in the Ping direction. For each frame, three indices are calculated: the rate of change of sound velocity profile, the rate of change of echo intensity, and the continuity of local terrain slope. Valid frames are then identified by marking them according to directional stability.

[0087] Sliding statistics are performed on the effective frames according to the fine-scale, medium-scale and coarse-scale windows respectively. The dominant scale type is determined based on the rate of change of sound velocity profile and slope continuity. The three scales at the same position are prioritized in combination with the rate of change of echo intensity to form a frame-level scale optimization matrix.

[0088] The frame-level scale optimization matrix is ​​fused by three-dimensional interpolation along the propagation path and beam angle directions. The selected, candidate, or shielded status is marked for each scale according to the optimization results, and a cross-scale selection map is generated.

[0089] The attention control factor determines the contribution ratio of the three types of coherent domain continuum feature fusion based on the cross-scale selection map: high weight is given to the scale marked as selected, medium weight is given to the candidate scale, and the weight of the scale marked as masked by cross-scale conflict or lack of continuity is set to zero, so as to avoid noise or incoherent fragments interfering with deep inference and to achieve an adaptive weight allocation of physical and terrain consistency among fine-scale, medium-scale and coarse-scale information.

[0090] Enabling a collision suppression gate to shield coherent domain segments with long-distance discontinuities is specifically as follows:

[0091] When continuously traversing the same coherent domain, monitor the index spacing between adjacent frames; when the frame spacing exceeds the set cross-frame threshold and the propagation direction stability mark fails at the same time, accumulate the interval to the discontinuity length counter; if the accumulated discontinuity length exceeds the preset threshold, trigger the collision suppression gate, mark the current coherent domain segment as shielded; stop further accumulation in the propagation direction, and synchronously write the shielding mark into the cross-scale selection map.

[0092] Based on the cross-scale attention results, weighted integration and position alignment are performed among the thrust flow response records in the fine-scale coherence domain, the mesoscale coherence domain, and the coarse-scale coherence domain to generate a cross-scale propagation consistency representation.

[0093] In this embodiment, obtaining three types of depth prediction results includes:

[0094] Receive the cross-scale propagation consistency representation and the acoustic propagation field tensor, establish a three-dimensional index consistent with the Ping order, propagation path order, and beam angle number, and locate the set of points to be processed;

[0095] Constructing a geometric reconstruction branch: Within the fine-scale, mesoscale, and coarse-scale coherence domains, perform intra-domain assembly in the order of edge priority, continuity priority, and step preservation. This involves generating isobath segments, monotonically splicing along the Ping and propagation path, and stitching adjacent segments together to achieve cross-scale propagation consistency representation. The output is a geometrically reasonable depth sequence, where:

[0096] The process of generating contour segments for cross-scale propagation consistency representation is as follows:

[0097] The point with the highest propagation consistency in the same Ping is set as the seed based on the depth value. Adjacent points within the depth tolerance range and with valid consistency markings are searched in the positive and negative directions of the propagation path, and synchronously expanded according to the beam angle direction.

[0098] When the search spans different Pings, if the depth difference and orientation stability of adjacent Pings simultaneously meet the threshold conditions, the corresponding point in the new frame is merged into the current segment; during the segment growth process, slope changes are continuously detected, and when the local slope change exceeds the set limit, the extension is stopped and the segment boundary is recorded.

[0099] Finally, all segments are sorted by depth and connectivity is verified. Continuous and closed sets of isobath points are saved as isobath segments of the geometric reconstruction branch.

[0100] Constructing an acoustic physics branch: Based on cross-scale propagation consistency representation and sound velocity profile index, beam angle index, and propagation path index, it performs path advancement for layered sound velocity segments, angular gating for incident angle constraints, energy gating for echo attenuation constraints, and candidate elimination for multipath suppression, outputting an acoustically reasonable depth sequence. Specifically, the execution of path advancement for layered sound velocity segments, angular gating for incident angle constraints, energy gating for echo attenuation constraints, and candidate elimination for multipath suppression involves:

[0101] After segmenting according to the sound velocity profile, starting from the sound velocity segment corresponding to the current Ping and beam angle, it gradually advances along the propagation path. At each segment boundary, the refraction direction is recalculated and the process continues until the path energy or connectivity determination is completed.

[0102] During the propulsion process, the incident angle is detected in real time. If the incident angle exceeds the preset angular threshold, the propagation is interrupted by angular gating. Simultaneously, the cumulative attenuation of the echo intensity is monitored. When the local energy is lower than the preset energy threshold, energy gating is triggered to eliminate candidate points with insufficient energy.

[0103] The propagation time and spatial connectivity of the retained candidate paths and direct paths are compared. Paths with insufficient connectivity are marked as multipath echoes and eliminated. Finally, the continuous path points that meet the physical constraints are written into the acoustically physical reasonable depth sequence.

[0104] Constructing a propagation field consistency branch: Taking the acoustic propagation field tensor as input, performing intra-field trajectory search, coherence domain boundary alignment, and long-distance connectivity verification according to the trajectory following rules of the thrust flow direction field, performing scene intersection localization on the cross-scale propagation consistency representation, and generating a propagation field consistency depth sequence. Specifically, the intra-field trajectory search, coherence domain boundary alignment, and long-distance connectivity verification according to the trajectory following rules of the thrust flow direction field are as follows:

[0105] Guided by the thrust flow direction vector in the acoustic propagation field tensor, we start from the seed point of the cross-scale propagation consistency representation and follow the direction field continuously, recording the path nodes step by step, and checking the direction consistency and energy smoothness at each step until we reach the direction field attenuation boundary or the path interruption position.

[0106] When the path crosses the boundaries of coherence domains at fine, medium and coarse scales, the trajectory nodes are reverted to the nearest valid edge of the coherence domain by referring to the boundary index, correcting the spatial offset that may be caused by scale switching.

[0107] After completing the alignment of the coherent domain boundaries, a connectivity check is performed along the entire path. If any segment experiences a sudden change in direction, energy break, or node loss exceeding the set distance, the segment is terminated and backtracked to the previous stable node. The set of depth values ​​for the complete trajectory that passes the connectivity check is recorded and written into the consistent depth sequence of the propagation field.

[0108] Alignment is performed on the three types of output depth sequences according to the three-dimensional index. Segment-level consistent alignment is completed based on the combination relationship of edge preservation markers, propagation continuity markers, and field orientation alignment markers to form three types of depth prediction results.

[0109] In this embodiment, generating the anomaly score for the corresponding point includes:

[0110] A three-dimensional index corresponding to the original depth is established for the geometric depth prediction results, acoustic physical depth prediction results and propagation field consistent depth prediction results, and the position of each sounding point is kept consistent on the Ping sequence, propagation path sequence and beam angle sequence.

[0111] Constructing geometric consistency deviation: Voxel windows are formed in three directions with each sounding point as the center. The geometric depth prediction results within the window are compared with the original depth point by point. If the difference is within the slope smoothing threshold and the edges of adjacent points remain continuously marked, it is recorded as zero deviation. Otherwise, the difference is jointly mapped to a stepped geometric deviation grade value according to the magnitude of the difference and the edge abruptness level.

[0112] Constructing acoustic consistency deviation: Within the same voxel window, the acoustic physical depth prediction result is compared with the original depth by difference. At the same time, the rate of change of sound velocity profile, the magnitude of change of incident angle, and the magnitude of echo intensity attenuation at the center of the voxel window are retrieved. If the three acoustic indicators are all in the stable range and the difference does not exceed the acoustic tolerance threshold, it is recorded as zero deviation. Otherwise, the penalty coefficient of the acoustic indicators is added according to the difference to map to a multi-level acoustic deviation value.

[0113] Constructing propagation field consistency deviation: The propagation field trajectory is tracked along the index dimension from the starting point to the end point of the sound wave propagation direction. The difference between the predicted propagation field consistency depth and the original depth is compared. The deviation level is determined based on the propagation direction alignment mark and connectivity integrity check results. When the trajectory is connected and the direction is aligned, and the difference is within the permissible range, it is recorded as zero deviation; otherwise, the difference is jointly mapped to a graded propagation field deviation value using the combined mapping of the difference and connectivity breakage penalty. Specifically, the combined mapping of the difference and connectivity breakage penalty to a graded propagation field deviation value is as follows:

[0114] The absolute difference between the consistent depth and the original depth of the propagation field is calculated at each trajectory node. The absolute difference is divided into low difference interval, medium difference interval, and high difference interval, and the corresponding basic deviation levels are set as level one, level two, and level three, respectively.

[0115] The trajectory connectivity is broken by counting the number of breaks. If the distance between missing consecutive nodes is less than the first threshold, the penalty coefficient is set to 0; if the distance between missing nodes is between the first threshold and the second threshold, the penalty coefficient is set to 1; if the distance between missing nodes exceeds the second threshold, the penalty coefficient is set to 2.

[0116] The weights corresponding to the basic deviation level and the penalty coefficient are summed. If the total value is 0, it is still recorded as zero deviation. If the total value is 1, it is recorded as level 1 deviation. If the total value is 2, it is recorded as level 2 deviation. If the total value is greater than 2, it is recorded as level 3 deviation. The obtained classification results are written into the propagation field consistency deviation sequence.

[0117] The geometric deviation, acoustic deviation, and propagation field deviation values ​​are respectively normalized to a percentage range with unified dimensions, and the three types of normalized deviations are weighted and summarized to generate an anomaly score.

[0118] In this embodiment, the output of the anomaly detection result includes:

[0119] Receive the anomaly score and establish a three-dimensional index mapping for the anomaly score corresponding to each point according to the Ping number index, propagation path index and beam angle index;

[0120] Local stability detection is performed on abnormal scores along the Ping number direction to filter out isolated score peak segments caused by single-point jumps. Continuity detection is performed on the propagation path direction to identify abnormal gaps between score segments. Angular domain smoothing is performed on the beam angle direction to eliminate beam-level random fluctuations, forming a score set corrected for three-dimensional consistency.

[0121] Based on the scoring set after three-dimensional consistency correction, an adaptive threshold set is generated according to regional noise density, coherence domain category, and local connectivity state. Independent outlier evaluation thresholds are determined for fine-scale, mesoscale, and coarse-scale coherence domains, respectively. Specifically, the generation of the adaptive threshold set according to regional noise density, coherence domain category, and local connectivity state involves:

[0122] First, the scoring set is divided into regions according to spatial grids. The mean, variance, and high score ratio of each grid are calculated. The noise density level of the region is obtained based on the high score ratio. Then, the corresponding initial threshold tables are set for the three coherence domains: fine, medium, and coarse.

[0123] The initial threshold is adjusted by weighting according to the regional noise density level: the threshold is appropriately raised in areas with high noise density and appropriately lowered in areas with low noise density, thus obtaining the first round of adaptive thresholds for each grid.

[0124] Check the local connectivity state within each grid. If the point set is completely connected, keep the first round threshold unchanged. If there is a break or jump, increase the threshold by one level based on the original threshold. Finally, generate an adaptive anomaly point evaluation threshold set for three types of coherence domains: fine-scale, medium-scale, and coarse-scale, and assign the corresponding threshold to each point.

[0125] Three thresholds are used in the fine-scale coherence domain: 0.35 in the low-noise region, 0.45 in the medium-noise region, and 0.55 in the high-noise region.

[0126] The corresponding thresholds for the mesoscale coherence domain are 0.40 for low noise, 0.50 for medium noise, and 0.60 for high noise.

[0127] The threshold settings for the coarse-scale coherence region are slightly higher: 0.45 for the low-noise region, 0.55 for the medium-noise region, and 0.65 for the high-noise region.

[0128] The thresholds are set in layers based on the differences in noise sensitivity and terrain detail requirements of each coherence domain at different scales: fine-scale coherence domains focus on local geometric details and are easily disturbed by isolated noise, so a relatively low threshold range of 0.35-0.55 is used to retain more real micro-topographic information; medium-scale coherence domains take into account both local and overall terrain trends and need to moderately suppress noise while avoiding the deletion of edge points, so the threshold is raised by 0.05 overall; coarse-scale coherence domains mainly capture large-scale continuous structures across Ping and across beams and have the strongest resistance to random noise, but are easily inflated by stripe artifacts, so a threshold range of up to 0.45-0.65 must be used to avoid misjudgment.

[0129] Each score in the abnormal score sequence is compared with the corresponding abnormal point evaluation threshold point by point. Points with scores higher than the abnormal point evaluation threshold are marked as abnormal points, and points with scores not higher than the abnormal point evaluation threshold are marked as normal points, thus generating an abnormal point label set in three-dimensional space.

[0130] Output the anomaly detection results of the sonar point cloud based on the anomaly marker set, and record the anomaly locations corresponding to the Ping number, propagation path index, and beam angle index.

[0131] Example 1:

[0132] To verify the feasibility of this invention in practice, it was applied to a shallow-sea survey data acquisition task in a certain sea area. The water depth in this area varies between 9 and 38 meters, and it exhibits local seabed scour channels, turbidity zones caused by tidal currents, and attitude disturbances caused by wind and waves. To complete a survey of a photovoltaic offshore pile foundation, a survey vessel conducted multibeam echo sounding in this area for three consecutive days in June 2025, accumulating approximately 1.65 million point cloud echo points. According to the on-site sound velocity meter records, the sound velocity profile in this sea area exhibits a significant layered structure at depths of 6–12 meters, causing small-angle refraction of sound waves during propagation and easily leading to local depth shifts. Furthermore, due to the superposition of tides and waves, multipath false echoes appeared in some Ping sequences, forming continuous strip-shaped anomalies. Traditional filtering methods using geometric neighborhood statistics and slope constraints in this area are prone to mistakenly deleting genuine shallow trench topography, resulting in high costs for manual secondary screening.

[0133] During data processing, the raw observation files exported from the survey vessel are preprocessed, including sound velocity correction, attitude compensation, and time synchronization. Taking one section of the survey line as an example, this line is approximately 540 meters long and contains 4,980 Pings, with an average of about 230 measurement points per Ping. After preprocessing, the system establishes an acoustic propagation field tensor index system based on the corrected point cloud. Due to the existence of tiers in the sound velocity profile, the curvature of the propagation path of the same beam varies at different incident angles. The method of this invention can automatically form three types of coherence domains in the tensor according to the evolution of the propagation path: fine-scale, mesoscale, and coarse-scale. The fine-scale coherence domain corresponds to the local energy direction stability region, the mesoscale coherence domain reflects the point set with a consistent slope, and the coarse-scale coherence domain covers the sound wave propagation trend across Pings, thus providing a structural basis for cross-scale thrust flow response calculation.

[0134] The system then applies the acoustic propagation direction characteristics to three types of coherence domains, performing cross-scale thrust flow response calculations within these domains. Since the tidal current velocity in this sea area was approximately 1.4 meters per second at the time of data acquisition, some beam echo energy exhibited slight disturbances. The processing mechanism of this invention can identify segments with directional deviations, energy fluctuations, and angular instability within the coherence domains, automatically eliminating disturbance segments that are ineffective for propagation behavior, and constructing continuous cross-scale propagation evolution data. The system performs acoustic and topographic coupled cross-scale attention inference on the evolution data. Because this survey line includes both gentle muddy terrain and small-scale scour gullies, attention inference can assign different weights to different scales based on regional acoustic stability and topographic trends, improving cross-scene adaptability.

[0135] In the three-branch depth inference stage, the geometric reconstruction branch successfully preserved the terrain edge structure in the shallow trench area; the acoustic physics branch generated a depth that better conformed to the laws of physical propagation in areas with strong local refraction effects; and the propagation field consistency branch continuously tracked depth changes along the propagation direction, making the multipath false echoes and the real terrain significantly different in the continuity of the propagation field, facilitating differentiation. Finally, the system constructs three types of consistency biases based on the three types of depth predictions and the original depth, and generates anomaly scores by combining regional noise density.

[0136] Table 1. Comparison of sonar point cloud anomaly detection results in a certain sea area

[0137]

[0138] As can be seen from the data in Table 1, this invention has a significant advantage in the accuracy of anomaly identification. Under the same survey line point cloud scale, the number of manually labeled anomalies is 14,200. The traditional neighborhood statistical method and the single-scale depth model correctly remove 11,157 and 11,948 anomalies, respectively, while this invention removes 13,745, achieving a correct removal rate of 96.8%. This demonstrates that this invention effectively improves the identification capability of complex anomalies such as multipath reflections, refraction migrations, and stripe artifacts through acoustic propagation field tensors, multi-scale coherence domains, and cross-scale thrust flow attention mechanisms, significantly outperforming existing technologies.

[0139] In terms of weak anomaly identification and real terrain preservation, this invention also demonstrates higher stability. The weak anomaly identification rate is improved from 63.5% in the traditional method to 91.3%, and the false deletion rate of real terrain is reduced from 3.7% in the traditional method to 0.9% in this invention. This improvement stems from the three-branch consistency inference mechanism proposed in this invention, which includes a geometric reconstruction branch, an acoustic physics branch, and a propagation field consistency branch. This mechanism allows each point to be comprehensively verified from three dimensions: geometry, acoustic physics, and propagation field, effectively reducing the missed detection of weak anomalies and the false deletion of real terrain.

[0140] In terms of processing efficiency and adaptability to complex noise, the method of this invention maintains good engineering usability. Traditional methods achieve a processing efficiency of approximately 21,000 points per second, and 34,000 points per second for a single-scale model, while this invention reaches 52,000 points per second. In the three subjective capability indicators of multipath artifact recognition, refraction migration recognition, and stripe artifact processing, this invention demonstrates strong performance. This indicates that this invention not only achieves high-precision anomaly detection but also maintains consistent processing stability and efficiency under real sea conditions, which is of great significance for improving the quality of marine mapping results.

[0141] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for detecting outliers in sonar point clouds using an adaptive multi-scale attention mechanism, characterized in that, include: Collect raw sonar data output from the sonar depth sounding equipment, preprocess the raw sonar data, and generate standardized sonar point cloud data; An acoustic propagation field tensor is constructed based on standardized sonar point cloud data. Gradient calculations are performed to obtain the characteristics of the sound wave propagation direction. Based on the acoustic propagation field tensor, fine-scale coherence domains, mesoscale coherence domains, and coarse-scale coherence domains are formed. The sound wave propagation direction features are applied to the fine-scale coherence domain, the mesoscale coherence domain, and the coarse-scale coherence domain respectively. Cross-scale thrust flow response calculation is performed to construct cross-scale propagation evolution data. Acoustic and terrain coupled cross-scale attention inference is performed on the cross-scale propagation evolution data to generate a cross-scale propagation consistency representation. Using cross-scale propagation consistency representation as input, geometric reconstruction branch, acoustic physics branch and propagation field consistency branch are constructed respectively to predict geometrically reasonable depth, acoustically reasonable depth and propagation field consistent depth, and obtain three types of depth prediction results; Based on the three types of depth prediction results and the original depth in the standardized sonar point cloud data, geometric consistency deviation, acoustic consistency deviation and propagation field consistency deviation are constructed respectively, and anomaly scores are generated for the corresponding points. Set an anomaly detection threshold, identify points with anomaly scores greater than the threshold as anomalies in the sonar point cloud, remove them from the standardized sonar point cloud data, and output the anomaly detection results.

2. The method for detecting outliers in sonar point clouds using an adaptive multi-scale attention mechanism according to claim 1, characterized in that, The raw sonar data includes point coordinate data, Ping number data, beam angle data, echo intensity data, signal-to-noise ratio data, sound velocity profile data, and attitude information data.

3. The method for detecting outliers in sonar point clouds using an adaptive multi-scale attention mechanism according to claim 1, characterized in that, The preprocessing of the raw sonar data includes performing sound velocity correction, attitude compensation, and time synchronization processing on the raw sonar data.

4. The method for detecting outliers in sonar point clouds using an adaptive multi-scale attention mechanism according to claim 1, characterized in that, The process of forming fine-scale, mesoscale, and coarse-scale coherence domains based on the acoustic propagation field tensor includes: An acoustic propagation field tensor index system is established on the Ping sequence dimension, propagation path discrete dimension, and beam angle discrete dimension of standardized sonar point cloud data. Acoustic propagation intensity and propagation direction information are calculated for each index position, and sound wave propagation direction features are generated from the propagation direction information. A coherent seed set is selected from the index system. The coherent seeds meet the following requirements: the signal-to-noise ratio is not lower than the preset lower limit, the propagation direction is stable within the local window and the difference between the direction of the propagation seed and the direction of the adjacent index position does not exceed the preset angle threshold. Fine-scale candidates are defined according to the local window size, medium-scale candidates are defined according to the medium window size, and coarse-scale candidates are defined according to the large window size. Starting with a coherent seed, perform triaxial consistent region growth on each of the three discrete axes: The propagation direction remains continuous in the Ping direction, and the propagation intensity does not change abruptly. The propagation path maintains monotonous progression without interruption of energy. Maintaining angular adjacency and limiting directional differences in the beam angle direction; Preliminary regions for fine-scale coherence domains, mesoscale coherence domains, and coarse-scale coherence domains are generated respectively. Cross-scale conflict resolution and closure correction are performed on three types of preliminary regions: when there is overlap or boundary conflict between adjacent scales, the region is assigned according to the rules of prioritizing long-distance continuity, prioritizing medium-scale slope transition and prioritizing fine-scale edge preservation. Low-area isolated patches are removed, fracture boundaries are corrected, and voids within the allowable range are filled to obtain the final regions of fine-scale coherence domain, medium-scale coherence domain and coarse-scale coherence domain.

5. The method for detecting outliers in sonar point clouds using an adaptive multi-scale attention mechanism according to claim 1, characterized in that, The generation of cross-scale propagation consistency representations includes: In the fine-scale coherence domain, the propagation direction, the propagation path direction, and the beam angle direction are all propagated in the domain according to the characteristics of the sound wave propagation direction. The point set that satisfies the continuous propagation direction, the no sudden change in propagation intensity, and the angular neighbor acceptance limit is accumulated in a consistent direction to generate a fine-scale thrust flow response record. In the mesoscale coherence domain and the coarse-scale coherence domain, the same rules are followed to perform propagation and accumulation to generate mesoscale thrust flow response records and coarse-scale thrust flow response records, respectively. The fine-scale thrust flow response records, meso-scale thrust flow response records, and coarse-scale thrust flow response records are aligned in time-path-angle order according to Ping order, propagation path order, and beam angle number. Incomplete segments are removed and gaps within the allowable range are interpolated to form cross-scale propagation evolution data. Using cross-scale propagation evolution data as input, a cross-scale selection map is generated based on the rate of change of sound velocity profile, the rate of change of echo intensity, and the continuity and directional stability of local terrain slope. Attention control factors are assigned to the fine-scale coherence domain, the meso-scale coherence domain, and the coarse-scale coherence domain, respectively. A conflict suppression gate is enabled to shield coherence domain segments with long-distance discontinuities. The cross-scale attention results of acoustic and terrain coupling are output. Based on the cross-scale attention results, weighted integration and position alignment are performed among the thrust flow response records in the fine-scale coherence domain, the mesoscale coherence domain, and the coarse-scale coherence domain to generate a cross-scale propagation consistency representation.

6. The method for detecting outliers in sonar point clouds using an adaptive multi-scale attention mechanism according to claim 1, characterized in that, The obtained three types of depth prediction results include: Receive the cross-scale propagation consistency representation and the acoustic propagation field tensor, establish a three-dimensional index consistent with the Ping order, propagation path order, and beam angle number, and locate the set of points to be processed; Construct a geometric reconstruction branch: In the fine-scale coherence domain, mesoscale coherence domain, and coarse-scale coherence domain, perform intra-domain assembly in the order of edge priority, continuity priority, and step preservation. Generate isobath segments, monotonically splice along Ping and the propagation path, and stitch adjacent segments together to close the consistent representation of cross-scale propagation, and output a geometrically reasonable depth sequence. Construct an acoustic physics branch: Based on cross-scale propagation consistency representation and sound velocity profile index, beam angle index, and propagation path index, perform path advancement of layered sound velocity segments, angular gating of incident angle constraints, energy gating of echo attenuation constraints, and candidate elimination of multipath suppression, and output an acoustically reasonable depth sequence. Constructing a propagation field consistency branch: Taking the acoustic propagation field tensor as input, the in-field trajectory search, coherence domain boundary alignment and long-distance connectivity verification are performed according to the trajectory following rules of the thrust flow direction field. The cross-scale propagation consistency representation is then localized by scene intersection and a propagation field consistency depth sequence is generated. Alignment is performed on the three types of output depth sequences according to the three-dimensional index. Segment-level consistent alignment is completed based on the combination relationship of edge preservation markers, propagation continuity markers, and field orientation alignment markers to form three types of depth prediction results.

7. The method for detecting outliers in sonar point clouds using an adaptive multi-scale attention mechanism according to claim 1, characterized in that, The generation of anomaly scores for corresponding points includes: A three-dimensional index corresponding to the original depth is established for the geometric depth prediction results, acoustic physical depth prediction results and propagation field consistent depth prediction results, and the position of each sounding point is kept consistent on the Ping sequence, propagation path sequence and beam angle sequence. Constructing geometric consistency deviation: Voxel windows are formed in three directions with each sounding point as the center. The geometric depth prediction results within the window are compared with the original depth point by point. If the difference is within the slope smoothing threshold and the edges of adjacent points remain continuously marked, it is recorded as zero deviation. Otherwise, the difference is jointly mapped to a stepped geometric deviation grade value according to the magnitude of the difference and the edge abruptness level. Constructing acoustic consistency deviation: Within the same voxel window, the acoustic physical depth prediction result is compared with the original depth by difference. At the same time, the rate of change of sound velocity profile, the magnitude of change of incident angle, and the magnitude of echo intensity attenuation at the center of the voxel window are retrieved. If the three acoustic indicators are all in the stable range and the difference does not exceed the acoustic tolerance threshold, it is recorded as zero deviation. Otherwise, the penalty coefficient of the acoustic indicators is added according to the difference to map to a multi-level acoustic deviation value. Constructing propagation field consistency deviation: Track the propagation field trajectory along the index dimension from the starting point to the end of the sound wave propagation direction, compare the difference between the propagation field consistency depth prediction result and the original depth, and determine the deviation level based on the propagation direction alignment mark and connectivity integrity check result. When the trajectory is connected and the direction is aligned, and the difference is within the permissible range, it is recorded as zero deviation; otherwise, the difference and connectivity breakage penalty are jointly mapped to the graded propagation field deviation value. The geometric deviation, acoustic deviation, and propagation field deviation values ​​are respectively normalized to a percentage range with unified dimensions, and the three types of normalized deviations are weighted and summarized to generate an anomaly score.

8. The method for detecting outliers in sonar point clouds using an adaptive multi-scale attention mechanism according to claim 1, characterized in that, The output anomaly detection results include: Receive the anomaly score and establish a three-dimensional index mapping for the anomaly score corresponding to each point according to the Ping number index, propagation path index and beam angle index; Local stability detection is performed on abnormal scores along the Ping number direction to filter out isolated score peak segments caused by single-point jumps. Continuity detection is performed on the propagation path direction to identify abnormal gaps between score segments. Angular domain smoothing is performed on the beam angle direction to eliminate beam-level random fluctuations, forming a score set corrected for three-dimensional consistency. Based on the scoring set after three-dimensional consistency correction, an adaptive threshold set is generated according to regional noise density, coherence domain category and local connectivity state, and independent outlier evaluation thresholds are determined for fine-scale coherence domain, mesoscale coherence domain and coarse-scale coherence domain respectively. Each score in the abnormal score sequence is compared with the corresponding abnormal point evaluation threshold point by point. Points with scores higher than the abnormal point evaluation threshold are marked as abnormal points, and points with scores not higher than the abnormal point evaluation threshold are marked as normal points, thus generating an abnormal point label set in three-dimensional space. Output the anomaly detection results of the sonar point cloud based on the anomaly marker set, and record the anomaly locations corresponding to the Ping number, propagation path index, and beam angle index.

Images