CN-122017854-A - Sonar point cloud outlier detection method of self-adaptive multi-scale attention mechanism

CN122017854ACN 122017854 ACN122017854 ACN 122017854ACN-122017854-A

Abstract

The invention discloses a sonar point cloud outlier detection method of a self-adaptive multi-scale attention mechanism, which comprises the steps of collecting original sonar data, preprocessing the data to generate standardized sonar point cloud data, constructing acoustic propagation field tensors, calculating propagation direction characteristics to form multi-scale coherent domains, enabling the propagation direction characteristics to act on each coherent domain, executing cross-scale thrust flow response calculation to form propagation evolution data, respectively constructing three branches by using propagation consistency as input to obtain three types of depth prediction results, constructing three types of consistency deviations according to the three types of depth prediction results and original depths to generate outlier scores, setting outlier judgment thresholds, outputting outlier detection results and eliminating outliers. According to the invention, high-precision and self-adaptive detection of complex abnormal points in the sonar point cloud is realized by constructing the acoustic propagation field tensor and combining multi-scale attention and three-branch consistency reasoning.

Inventors

CHEN JUNYU
CEN WENJIE
LIU CHEN
ZHENG CHENMING
YUAN HAO
Xia Xianhuang
SHI TIANYU

Assignees

中交四航局第三工程有限公司
中交第四航务工程局有限公司

Dates

Publication Date: 20260512
Application Date: 20260413

Claims (8)

1. The sonar point cloud outlier detection method of the self-adaptive multi-scale attention mechanism is characterized by comprising the following steps of: the method comprises the steps of collecting original sonar data output by sonar sounding equipment, preprocessing the original sonar data, and generating standardized sonar point cloud data; Constructing an acoustic propagation field tensor based on standardized sonar point cloud data, performing gradient operation to obtain acoustic wave propagation direction characteristics, and forming a fine-scale coherent domain, a mesoscale coherent domain and a coarse-scale coherent domain according to the acoustic propagation field tensor; Respectively acting the sound wave propagation direction characteristics on a fine-scale coherent domain, a mesoscale coherent domain and a coarse-scale coherent domain, executing cross-scale thrust flow response calculation, constructing cross-scale propagation evolution data, executing cross-scale attention reasoning of acoustic and terrain coupling on the cross-scale propagation evolution data, and generating cross-scale propagation consistency representation; Taking cross-scale propagation consistency as input, respectively constructing a geometric reconstruction branch, an acoustic physical branch and a propagation field consistency branch, predicting geometric reasonable depth, acoustic physical reasonable depth and propagation field consistency depth, and obtaining three types of depth prediction results; respectively constructing geometric consistency deviation, acoustic consistency deviation and propagation field consistency deviation according to three types of depth prediction results and original depth in standardized sonar point cloud data, and generating abnormal scores of corresponding points; and setting an abnormal point judgment threshold, determining points with abnormal scores larger than the judgment threshold as abnormal points of the sonar point cloud, removing the abnormal points from the standardized sonar point cloud data, and outputting an abnormal point detection result.
2. The method for detecting abnormal values of sonar point cloud of adaptive multi-scale attention mechanism of claim 1, wherein said raw sonar data comprises point location 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. A sonar point cloud outlier detection method of an adaptive multi-scale attention mechanism as recited in claim 1, wherein said preprocessing of raw sonar data comprises performing sound speed correction, performing attitude compensation, performing time synchronization processing on the raw sonar data.
4. The method for detecting abnormal values of sonar point clouds by an adaptive multi-scale attention mechanism according to claim 1, wherein forming a fine-scale coherence domain, a mesoscale coherence domain and a coarse-scale coherence domain according to an acoustic propagation field tensor comprises: Establishing an acoustic propagation field tensor index system on the Ping sequence dimension, the propagation path discrete dimension and the beam angle discrete dimension of the standardized sonar point cloud data, calculating acoustic propagation intensity and propagation direction information for each index position, and generating acoustic propagation direction characteristics according to the propagation direction information; Selecting a coherent seed subset from an index system, wherein the coherent seed meets the requirements that the signal-to-noise ratio is not lower than a preset lower limit, the propagation direction is stable in a local window and the direction difference between the coherent seed and the adjacent index position is not more than a preset angle threshold, and the coherent seed defines a fine-scale candidate according to the local window size, a middle-scale candidate according to the middle window size and a coarse-scale candidate according to the large window size; Starting with coherent seeds, three-axis consistent region growth is performed on three discrete axes: The propagation direction is kept continuous in the Ping direction and the propagation strength is free from mutation; The monotonic advancing of the path is kept in the direction of the propagation path, and the energy is not interrupted; maintaining angular adjacency and directional variance limited in beam angular direction; Respectively generating a fine-scale coherent domain, a mesoscale coherent domain and a preliminary region of a coarse-scale coherent domain; And performing cross-scale conflict resolution and closure correction on the three types of preliminary areas, namely performing attribution judgment according to rules of long-distance continuity priority, medium-scale gradient transition priority and fine-scale edge retention priority when adjacent scales have overlapping or boundary conflict, removing low-area isolated sheets, correcting fracture boundaries, filling holes in an allowable range, and obtaining final areas of a fine-scale coherent domain, a medium-scale coherent domain and a coarse-scale coherent domain.
5. The method for detecting outliers of sonar point clouds for an adaptive multi-scale attention mechanism of claim 1, wherein said generating a cross-scale propagated consistency representation comprises: Performing intra-domain propulsion in a fine-scale coherent domain along the Ping direction, the propagation path direction and the beam angle direction according to the characteristics of the propagation direction of the sound wave, performing direction consistency accumulation on a point set meeting the requirements of continuity in the propagation direction, no abrupt change in propagation strength and limited angle adjacency, generating a fine-scale thrust flow response record, and performing propulsion and accumulation in a middle-scale coherent domain and a coarse-scale coherent domain according to the same rule respectively, generating a middle-scale thrust flow response record and a coarse-scale thrust flow response record; Performing time-path-angle alignment on the fine-scale thrust flow response record, the middle-scale thrust flow response record and the coarse-scale thrust flow response record according to the Ping sequence, the propagation path sequence and the beam angle sequence, removing incomplete fragments, and performing interpolation on vacant positions within an allowable range to form cross-scale propagation evolution data; Using transscale propagation evolution data as input, generating a transscale selection graph according to the sound velocity profile change rate, the echo intensity change rate, the local terrain gradient continuity and the direction stability mark, respectively distributing attention control factors to a fine-scale coherent domain, a mesoscale coherent domain and a coarse-scale coherent domain, starting a conflict suppression gate to shield coherent domain fragments with long-distance discontinuity, and outputting a transscale attention result of acoustic and terrain coupling; And performing weighted integration and position alignment among thrust stream response records of the fine-scale coherent domain, the mesoscale coherent domain and the coarse-scale coherent domain according to the cross-scale attention result to generate a cross-scale propagation consistency representation.
6. The method for detecting abnormal values of sonar point cloud of adaptive multi-scale attention mechanism of claim 1, wherein said obtaining three types of depth prediction results comprises: Receiving a trans-scale propagation consistency representation and an acoustic propagation field tensor, establishing a three-dimensional index consistent with a Ping sequence, a propagation path sequence and a beam angle sequence number, and positioning a point set to be processed; Constructing a geometric reconstruction branch, namely performing intra-domain assembly in the fine-scale coherent domain, the mesoscale coherent domain and the coarse-scale coherent domain according to the sequence of edge priority, continuous priority and step reservation, generating isodepth line fragments for trans-scale propagation consistency representation, monotonous splicing along Ping and propagation paths and stitching and closing adjacent fragments, and outputting a geometric reasonable depth sequence; Constructing an acoustic physical branch, namely performing path propulsion of a layered sound velocity paragraph, angle gating of an incidence angle constraint, energy gating of an echo attenuation constraint and candidate rejection of multipath inhibition based on a trans-scale propagation consistency representation, a sound velocity profile index, a beam angle index and a propagation path index, and outputting an acoustic physical reasonable depth sequence; Constructing a propagation field consistency branch, namely taking an acoustic propagation field tensor as input, performing in-field track search and coherent domain boundary alignment and long-distance connectivity check according to track following rules of a thrust flow direction field, performing scene intersection positioning on a trans-scale propagation consistency representation, and generating a propagation field consistency depth sequence; And performing alignment on the three types of output depth sequences according to the three-dimensional index, and completing segment-level consistent alignment according to the combination relation of the edge maintenance mark, the propagation continuity mark and the field direction alignment mark to form three types of depth prediction results.
7. The method for detecting abnormal values of sonar point cloud of adaptive multi-scale attention mechanism of claim 1, wherein generating abnormal scores of corresponding points comprises: establishing three-dimensional indexes corresponding to the original depth one by one for the geometric depth prediction result, the acoustic physical depth prediction result and the propagation field consistent depth prediction result, and keeping the positions of each sounding point on the Ping sequence, the propagation path sequence and the beam angle sequence consistent; forming voxel windows in three directions by taking each sounding point as a center, comparing the geometric depth prediction result in the windows with the original depth point by point, marking zero deviation if the difference is in the gradient smoothing threshold range and the edges of adjacent points keep continuous, otherwise, mapping the difference value and the edge mutation level into stepped geometric deviation grading values in a combined way; The acoustic consistency deviation is constructed by comparing the acoustic physical depth prediction result with the original depth in the same voxel window, and simultaneously searching the sound velocity profile change rate, the incident angle change amplitude and the echo intensity attenuation amplitude at the center of the voxel window, if the three acoustic indexes are all in a stable interval and the difference value does not exceed an acoustic tolerance threshold value, marking the three acoustic indexes as zero deviation, otherwise, overlapping the punishment coefficients of the acoustic indexes according to the difference value, and mapping the three acoustic indexes as multilevel acoustic deviation values; Tracking a propagation field track in an index dimension along the starting point to the tail end of the propagation direction of the sound wave, comparing the difference value of a propagation field consistent depth prediction result with the original depth, judging a deviation grade according to a propagation direction alignment mark and a communication integrity check result, recording zero deviation when the track is communicated and the direction alignment and the difference value are in a permissible range, otherwise, jointly mapping the difference value and the connectivity fracture penalty into a graded propagation field deviation value; and respectively rectifying the geometric deviation grading value, the acoustic deviation value and the propagation field deviation value into a unified dimension percentage metric interval, and carrying out weighted summarization on the three types of rectifying deviations to generate an abnormal score.
8. The method for detecting abnormal values of sonar point cloud of adaptive multi-scale attention mechanism of claim 1, wherein said outputting abnormal point detection results comprises: Receiving the anomaly scores, and enabling the anomaly scores corresponding to each point to establish three-dimensional index mapping according to the Ping number index, the propagation path index and the beam angle index; Performing local stability detection on the abnormal scores along the Ping numbering direction, screening out isolated score peak fragments caused by single-point jump, performing continuity detection on the propagation path direction to identify abnormal gaps among the score fragments, and performing angular domain smoothing on the beam angle direction to eliminate beam level random fluctuation, so as to form a score set corrected by three-dimensional consistency; Based on the three-dimensional consistency corrected scoring set, generating a self-adaptive threshold set according to the regional noise density, the coherent domain category and the local communication state, and respectively determining independent outlier scoring thresholds for the fine-scale coherent domain, the mesoscale coherent domain and the coarse-scale coherent domain; Each score in the abnormal score sequence is compared with a corresponding abnormal point evaluation threshold point by point, points with scores higher than the abnormal point evaluation threshold value are marked as abnormal points, points with scores not higher than the abnormal point evaluation threshold value are marked as normal points, and an abnormal point marking set in the three-dimensional space is generated; And outputting a sonar point cloud outlier detection result according to the outlier mark set, and recording outlier positions corresponding to the Ping number, the propagation path index and the beam angle index.

Description

Sonar point cloud outlier detection method of self-adaptive multi-scale attention mechanism Technical Field The invention relates to the technical field of sonar sounding data processing, in particular to a sonar point cloud outlier detection method of a self-adaptive multi-scale attention mechanism. Background In applications such as marine mapping, marine topography exploration, underwater engineering monitoring, etc., sonar sounding devices are widely used to acquire three-dimensional point cloud data of an underwater environment. Existing sonar point cloud processing methods typically rely on geometric rules, filtering thresholds, or outlier rejection strategies based on statistical characteristics to identify outliers caused by water body disturbances, acoustic wave scattering, multipath reflections, and equipment attitude changes. However, most of the methods are based on shallow geometric features or local consistency analysis, and lack modeling capability on acoustic characteristics such as refraction effect, multipath superposition, angular domain artifact and the like generated in the acoustic wave propagation process, so that it is difficult to accurately distinguish a real submarine structure from complex acoustic noise. In addition, the traditional method has insufficient adaptability in different scale terrain scenes, and when a submarine structure presents a slope fold, a step or a mixed terrain, the situation that the real terrain points are deleted by mistake or weak abnormal points are omitted often occurs. With the development of deep learning technology, partial researches try to apply a graph structure, a convolution network or a global feature extraction method to sonar point cloud outlier detection, but such methods still have significant limitations. Firstly, the existing model is generally constructed based on a visual framework, and cannot embody the physical rule of sound wave propagation, so that the robustness of the model is insufficient when facing the gradient change of sound velocity or the bending of a sound wave propagation path. Secondly, most models only use the feature expression of a single scale or a fixed scale, lack of a cross-scale association mechanism, and are difficult to consider both fine-scale point cloud noise and large-scale submarine structure change. Thirdly, most of anomaly decisions of the current method depend on a single prediction result or a single feature space, and the capability of comprehensively evaluating point cloud rationality from three aspects of geometry, acoustic physics and propagation fields is lacking. The prior art is difficult to realize the comprehensive identification of the abnormal value of the sonar point cloud with high precision, self-adaption and physical consistency, and particularly has insufficient performance under complex scenes such as multipath reflection, acoustic refraction, stripe artifact, weak echo abnormality and the like. Therefore, how to provide a sonar point cloud outlier detection method with an adaptive multi-scale attention mechanism is a problem that needs to be solved by those skilled in the art. Disclosure of Invention The invention aims to provide a sonar point cloud outlier detection method of a self-adaptive multi-scale attention mechanism, which utilizes technologies such as acoustic propagation field tensor modeling, multi-scale coherent domain construction, cross-scale thrust stream meaning force reasoning, geometric-acoustic-propagation field three-branch consistency judgment and the like to automatically identify and reject outliers formed by multipath reflection, acoustic refraction, stripe artifacts and weak echo interference in sonar point cloud, and realizes the deep fusion of acoustic physical constraint and data driving reasoning. The method has the advantages of strong adaptability to complex acoustic noise, high anomaly identification precision, excellent cross-scene robustness and the like, effectively improves the quality and reliability of sonar sounding data, and provides a more stable technical basis for ocean mapping and underwater target perception. According to the embodiment of the invention, the sonar point cloud outlier detection method of the self-adaptive multi-scale attention mechanism comprises the following steps: the method comprises the steps of collecting original sonar data output by sonar sounding equipment, preprocessing the original sonar data, and generating standardized sonar point cloud data; Constructing an acoustic propagation field tensor based on standardized sonar point cloud data, performing gradient operation to obtain acoustic wave propagation direction characteristics, and forming a fine-scale coherent domain, a mesoscale coherent domain and a coarse-scale coherent domain according to the acoustic propagation field tensor; Respectively acting the sound wave propagation direction characteristics on a fine-scale coherent domain, a mesoscale coherent doma