Depth perception-based method and system for identifying berth obstacles of water decontamination robot

Abstract

The invention provides a method and a system for identifying a berth obstacle of a water decontamination robot based on depth perception, and the method comprises the steps: executing the interweaving tracing of a multi-source signal track, and fusing the signals of a robot and peripheral fixed sensing equipment to generate an interlaced obstacle signal track; constructing a dynamic symbiotic feature model, mining a dynamic symbiotic relationship between the trajectory and multiple feature libraries, and generating an association rule; then implementing hierarchical reverse calibration, and adjusting a model association rule weight by using historical and actual data; reconstructing an obstacle holographic three-dimensional contour, and analyzing node attributes and features to generate contour information; and finally, outputting a self-adaptive evasion instruction, transmitting the contour information to a robot motion control unit, generating motion control parameters in combination with the parameters of the robot and the operation scene, and packaging the motion control parameters into an instruction. According to the invention, the accuracy and adaptability of berth obstacle identification of the water trash cleaning robot are improved, and the operation efficiency and safety are improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence, and more specifically, to a method and system for recognizing obstacles at berths of a waterborne cleaning robot based on depth perception. Background Technology

[0002] In aquatic wastewater cleanup operations, robots need to accurately identify obstacles in berth areas to ensure safe and efficient operation. Traditional methods for obstacle identification in aquatic wastewater cleanup robots often rely on single sensing devices, such as sensors mounted on the robot itself. However, single sensing devices have limitations, including limited sensing range and susceptibility to environmental interference, leading to insufficient accuracy and comprehensiveness in obstacle identification. For example, the robot's own sensors may fail to accurately obtain crucial information such as the depth and location of obstacles due to water surface reflections and water flow fluctuations. Furthermore, traditional methods lack effective fusion and in-depth analysis of multi-source sensing signals, making it difficult to uncover complex features hidden within the signals. This hinders their adaptability to different berth environments, different types of obstacles, and dynamically changing operational scenarios, ultimately affecting the robot's obstacle avoidance capabilities and operational efficiency. Summary of the Invention

[0003] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a method for recognizing obstacles at berths of a depth-sensing-based waterborne waste cleaning robot, the method comprising: The system performs multi-source signal trajectory interweaving and tracing, receives the original sensing signal output by the depth sensing component of the water cleaning robot, and simultaneously accesses the auxiliary sensing signal transmitted by the fixed sensing equipment around the berth. It then fuses the original sensing signal and the auxiliary sensing signal according to the time and space dimensions to generate a fused sensing signal, marks the abnormal fluctuation points in the fused sensing signal, tracks the propagation trajectory of the abnormal fluctuation points, and interweaves and associates the signal nodes of different trajectories to form an interwoven obstacle signal trajectory. A dynamic symbiotic feature model is constructed by inputting intertwined obstacle signal trajectories into a preset artificial intelligence framework, linking the berth environment feature library, obstacle history feature library, and real-time operation feature library, mining the dynamic symbiotic relationship between the trajectory and berth environment features, obstacle history features, and real-time operation features, transforming the symbiotic relationship into multiple association rules at different levels, and embedding them into the framework to generate a dynamic symbiotic feature model. Hierarchical reverse calibration is implemented by calling up historical obstacle identification data stored locally on the water cleaning robot and actual obstacle measurement data collected on site, feeding dynamic symbiotic feature models in different batches, capturing the deviations at different levels between predicted obstacle information and actual obstacle information, adjusting the model association rule weights according to the level of deviation, and completing the hierarchical reverse calibration. Reconstruct the holographic 3D contour of the obstacle, load the calibrated dynamic symbiotic feature model into the holographic contour analysis unit, input the spatial occupancy attributes and signal features of the interwoven obstacle signal trajectory analysis node, associate the attributes and features to form the basic contour framework, fill in the detailed features to generate the holographic 3D contour information of the obstacle. The system outputs adaptive avoidance commands, transmitting the holographic 3D contour information of the obstacle to the motion control unit of the waterborne cleaning robot. It analyzes the obstacle's position, size, and spatial posture in the contour information, and combines the robot's body length, width, and height dimensions, the range of motion angles of each joint, the maximum output speed and torque of the drive motor, the minimum turning radius in the working state, and the working scenario to convert them into motion control parameters and encapsulate them into adaptive avoidance commands.

[0004] In another aspect, embodiments of the present invention also provide a depth-sensing-based underwater cleaning robot berth obstacle recognition system, including a processor and a machine-readable storage medium connected to the processor. The machine-readable storage medium is used to store programs, instructions, or code, and the processor is used to execute the programs, instructions, or code in the machine-readable storage medium to implement the above-described method.

[0005] Based on the above, this invention, through multi-source signal trajectory interleaving and tracing, integrates signals from the depth perception component of the aquatic cleaning robot and fixed perception devices around the berth, comprehensively acquiring information about the berth area, effectively expanding the perception range, reducing the limitations of single perception devices, and improving the ability to capture obstacle signals. A dynamic symbiotic feature model is constructed to deeply explore the dynamic symbiotic relationship between obstacle signal trajectories and multiple feature libraries, forming association rules that enable the model to more accurately understand the characteristics of obstacles in different scenarios, enhancing its adaptability to complex environments. Layered reverse calibration is implemented, using historical and actual data to adjust the model layer by layer, effectively reducing prediction bias and improving the accuracy of model predictions. The holographic 3D contour of the obstacle is reconstructed, ultimately outputting adaptive avoidance commands, enabling the robot to flexibly and safely avoid obstacles based on its own parameters and the operating scenario, comprehensively improving the operating efficiency and safety of the aquatic cleaning robot in the berth area. Attached Figure Description

[0006] Figure 1 This is a schematic diagram of the execution flow of the obstacle recognition method for the waterborne cleaning robot based on depth perception provided in the embodiments of the present invention.

[0007] Figure 2 This is a schematic diagram of exemplary hardware and software components of the depth-sensing-based underwater cleaning robot berth obstacle recognition system provided in an embodiment of the present invention. Detailed Implementation

[0008] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 This is a flowchart illustrating a method for recognizing obstacles at berths of a depth-sensing-based underwater cleaning robot, provided in one embodiment of the present invention. The following is a detailed description of this method.

[0009] Step S110: Perform multi-source signal trajectory interweaving and tracing, receive the original sensing signal output by the depth sensing component of the water cleaning robot, simultaneously access the auxiliary sensing signal transmitted by the fixed sensing equipment around the berth, fuse the original sensing signal and the auxiliary sensing signal according to the time dimension and the spatial dimension to generate a fused sensing signal, mark the abnormal fluctuation points in the fused sensing signal, track the propagation trajectory of the abnormal fluctuation points, and interweave the signal nodes of different trajectories to form an interwoven obstacle signal trajectory.

[0010] In this embodiment, a surface debris removal robot is used as an example in a cargo berth in calm water. This berth is frequently used by bulk carriers, and potential obstacles include wooden pallets (sunken debris) dropped during transport, plastic containers discarded by crew members (floating garbage), concrete blocks left over from berth construction (reef-like objects), and fishing nets left behind by fishermen (abandoned fishing gear). The robot needs to identify these obstacles in real time during debris removal operations to avoid collisions that could damage the equipment or interrupt the debris removal task.

[0011] Step S111: Start the signal receiving module to receive the raw sensing signal output by the depth sensing component of the waterborne cleaning robot. The raw sensing signal includes the depth reflection signal, spatial position signal, signal attenuation signal and signal frequency signal of the signal point in the berth area.

[0012] The signal receiving module employs a dual-channel data acquisition card, supporting simultaneous reception of analog and digital signals. The depth sensing components include a 24-beam mechanical scanning sonar, a 3D lidar, and an underwater high-definition camera. The depth reflection signal is the echo signal received after the sonar emits sound waves; its amplitude is related to the surface material of the obstacle and the distance. The spatial position signal is calculated using the lidar's time-of-flight method, containing the three-dimensional coordinate components of the obstacle in the robot's local coordinate system. The signal attenuation signal reflects the energy loss of the sound wave during transmission and reception, and is related to the propagation path length and water absorption characteristics. The signal frequency signal is the spectral distribution of the reflected signal; different obstacle materials lead to differences in the frequency components of the reflected signal. For example, metal surfaces have relatively rich high-frequency components, while wooden surfaces have a higher proportion of low-frequency components. During reception, the module preprocesses the raw signal, including automatic gain control to avoid signal saturation and specific-frequency notch filtering to remove power frequency interference.

[0013] Step S112: Access the auxiliary sensing signals transmitted by the fixed sensing devices around the berth through a preset communication protocol. The auxiliary sensing signals include depth signals of the berth edge area, near-shore obstacle reflection signals, water flow disturbance signals, and water temperature correlation signals.

[0014] The default communication protocol is Modbus RTU based on TCP / IP, with data transmission achieved via fiber optic Ethernet, ensuring real-time transmission rates. Fixed sensing equipment around the berth is deployed at a certain distance from the berth edge on shore lighthouses and mooring bollards. This includes: a pulse depth sounder to collect depth signals in the berth edge area, preventing robots from approaching shallow waters; an FMCW radar to monitor reflected signals from near-shore fixed obstacles such as crash barriers; a Doppler current meter to collect water flow disturbance signals, including velocity vectors and flow direction angles; and a platinum resistance thermometer to collect water temperature correlation signals at different depths, compensating for the influence of sound velocity variations with water temperature. Auxiliary sensing signals are converted from analog to digital and transmitted in data packets. Each data packet contains a device ID, timestamp, signal type identifier, and signal value field.

[0015] Step S113: Start the signal fusion unit, load the time-space dual-dimensional alignment algorithm, extract the timestamp information of the original sensing signal and the auxiliary sensing signal, arrange the signal sequence in order, extract the spatial coordinates and unify the coordinate system through the coordinate transformation algorithm.

[0016] The signal fusion unit is built on an FPGA chip, and its parallel processing capability meets the real-time fusion requirements of multi-source signals. The temporal and spatial dual-dimensional alignment algorithm includes a time alignment sub-algorithm and a spatial alignment sub-algorithm. During time alignment, the UTC timestamps of the original and auxiliary sensing signals are extracted to millisecond accuracy. A linear interpolation method is used to resample signals with non-overlapping timestamps, synchronizing the two types of signals on the time axis to form a signal sequence with equal time intervals. The sequence length is determined based on the size of the working area and the sensor sampling frequency. During spatial alignment, the original sensing signal uses the robot's body coordinate system, while the auxiliary sensing signal uses the berth's geodetic coordinate system. A coordinate transformation algorithm unifies the two types of signals to the berth's geodetic coordinate system. The transformation process must consider the influence of the robot's real-time pose on the coordinates of the original sensing signal.

[0017] Step S1131: Extract the coordinate system parameters of the original sensing signal. The parameters include the position of the coordinate origin, the direction of the coordinate axes, the coordinate units, and the coordinate precision.

[0018] The coordinate system parameters of the original sensing signals are read by the robot control system. The coordinate origin is the real-time coordinate of the robot's center of mass in the geodetic coordinate system. The coordinate axis directions are determined by the roll, pitch, and yaw angles measured by the IMU. The X-axis points to the robot's current heading, and the Y and Z axes are determined using the right-hand rule. The coordinate unit is uniformly meters. The coordinate accuracy is determined by the sensor performance indicators. The ranging accuracy of the lidar is within its technical parameters, while the depth-finding accuracy of the sonar is affected by the beam opening angle and water scattering, and can reach the manufacturer's stated value under calm water conditions. These parameters are stored in the memory of the signal fusion unit in the form of a structure, and the coordinate origin position and coordinate axis directions are updated at regular time intervals.

[0019] Step S1132: Extract the coordinate system parameters of the auxiliary sensing signal and record the differences between the original sensing signal and the auxiliary sensing signal coordinate system in terms of origin, direction, unit and precision.

[0020] The coordinate system parameters of the auxiliary sensing signals are obtained from the configuration file of the fixed sensing equipment. The origin of the coordinate system is the geodetic coordinate of the port surveying reference point; the direction of the coordinate axes is the orientation of the geodetic coordinate system; the coordinate unit is also meters; the coordinate accuracy is determined by the equipment's technical specifications. By comparing the coordinate system parameters of the two types of signals, the differences are calculated: the origin difference is the difference between the real-time coordinates of the robot's center of mass and the coordinates of the surveying reference point; the direction difference is the angle between each axis of the robot's body coordinate system and each axis of the geodetic coordinate system; the units are the same; the accuracy difference is the standard deviation of different sensors in each coordinate axis direction.

[0021] Step S1133: Load the preset coordinate transformation algorithm, input the coordinate system parameters of the original sensing signal and the auxiliary sensing signal, and calculate the transformation matrix and deviation correction coefficient.

[0022] The preset coordinate transformation algorithm is a Bursa model for spatial rectangular coordinate transformation, containing seven transformation parameters. The coordinate system parameters extracted in steps S1131 and S1132 are input into the algorithm, where the origin difference is used as the initial value for the translation parameter, the direction difference as the initial value for the rotation parameter, and the scale parameter is set to one. By selecting uniformly distributed common points within the robot's working area, their coordinates in the original and auxiliary sensing coordinate systems are collected. The optimal estimate of the transformation parameters is calculated using the least squares adjustment method, resulting in the transformation matrix. Deviation correction coefficients are used to compensate for system errors. Through residual analysis of historical transformation data, a robust estimation method is used to calculate the linear correction terms for each coordinate axis direction.

[0023] Step S1134: Substitute all spatial coordinates of the auxiliary sensing signal into the transformation matrix, and calculate the transformed coordinate data by combining the deviation correction coefficient.

[0024] The spatial coordinates of the auxiliary sensing signal are stored in array form, with each coordinate point containing three components. During the transformation process, the coordinate points are first rotated and translated using a transformation matrix. Then, the transformed coordinates are corrected for deviations by performing proportional and translational corrections for each coordinate axis. The transformed coordinate data is stored in the same data structure as the original sensing signal.

[0025] Step S1135: Select uniformly distributed feature signal points and compare the consistency between the original sensing signal coordinates and the converted auxiliary sensing signal coordinates.

[0026] Feature signal points were selected using a grid-based method, dividing the work area into equally spaced grids on the XY plane, with one feature point selected at the center of each grid. Each feature point must be covered by both the original sensing signal and the auxiliary sensing signal, and the signal strength must be higher than a preset threshold. The number of selected points met statistical requirements, ensuring coverage of the entire work area, including the nearshore, central, and peripheral regions. For comparison, the coordinate difference of the same feature point in the original sensing coordinate system and the transformed auxiliary sensing coordinate system was calculated. The consistency evaluation index was the root mean square error of the coordinate difference, and the root mean square error in all three coordinate axes of all feature points was calculated.

[0027] Step S1136: Adjust the transformation matrix parameters and deviation correction coefficients to correct the transformation deviation until the coordinates of the original sensing signal and the auxiliary sensing signal are consistent with the preset standard, thus completing the coordinate system unification.

[0028] The preset standard is that the root mean square error (RMSE) in each coordinate axis direction is less than a set threshold. If the RMSE in a certain direction exceeds the threshold, the distribution pattern of the feature point coordinate differences is first analyzed. If a systematic deviation is observed, the process returns to step S1133 to adjust the translation parameters of the transformation matrix; if a linear trend is observed, the proportional coefficient in the deviation correction coefficient is adjusted. After adjustment, steps S1134 and S1135 are re-executed, with the number of iterations not exceeding the set upper limit. When the RMSE in all directions meets the preset standard, the coordinate system is considered unified, and the final transformation matrix and deviation correction coefficients are stored.

[0029] Step S114: Align the original sensing signal and the auxiliary sensing signal at the same coordinate point and the same time stamp, perform amplitude superposition and phase calibration, complete the signal superposition of all coordinate points and timestamps, and generate the fused sensing signal.

[0030] Signal alignment at the same coordinate point and timestamp is achieved through the spatiotemporal alignment results in step S113. For each spatiotemporal alignment point, corresponding data is extracted from the original sensing signal and the converted auxiliary sensing signal. Amplitude superposition uses a weighted average method, with weights determined based on the signal-to-noise ratio (SNR), assigning greater weights to signals with higher SNR. Phase calibration is performed on the sonar signal, using a cross-correlation method to calculate the phase difference between the original sensing signal and the auxiliary sensing signal. Then, phase compensation is applied to the auxiliary signal to ensure phase consistency between the two. After amplitude superposition and phase calibration are completed for all spatiotemporal alignment points, the results are organized by coordinate point and timestamp to form a fused sensing signal dataset. The data format includes fields such as coordinates, timestamp, amplitude, phase, and frequency.

[0031] Step S115: Load the signal fluctuation identification unit, retrieve the berth normal sensing signal range data built into the unit, cover all signal points of the fused sensing signal according to the preset scanning order, and compare the signal point values ​​with the normal range data.

[0032] The signal fluctuation identification unit is built on an embedded microprocessor, and the built-in normal berth sensing signal range data is obtained through historical data statistics. It collects fused sensing signals from periods of unobstructed access at the berth over the past year, statistically analyzes parameters such as signal amplitude and frequency at each coordinate point, and calculates a specific confidence interval as the normal range. The preset scanning sequence is a serpentine scan, starting from the upper left corner of the working area, scanning along the X-axis to the lower right corner, then scanning back the previous row, sequentially covering all signal points. The scanning step size matches the sensor's spatial resolution. During comparison, for each signal point, its real-time value is compared with the upper and lower limits of the normal range. If the value exceeds the upper limit or falls below the lower limit, it is marked as a potential abnormal fluctuation point. During the comparison process, the unit performs a moving average filter on the signal, with the window size determined based on the signal change rate to reduce misjudgments caused by instantaneous noise.

[0033] Step S116: Mark signal points with values ​​exceeding the normal range as abnormal fluctuation points, start the trajectory tracking unit to track the propagation trajectory of abnormal fluctuation points, and record the temporal sequence, spatial coordinates and signal strength changes of signal nodes on the path.

[0034] Abnormal fluctuation point markings are binarized, assigning a value of one to points exceeding the normal range and zero to normal points. The trajectory tracking unit employs a multi-target tracking algorithm based on extended Kalman filtering, with the state vector containing the three-dimensional position and velocity of the point. During initialization, tracking tracks are established for points marked as abnormal in two consecutive scan cycles, with the initial state covariance matrix set based on sensor measurement noise. In the prediction phase, the point position at the next moment is predicted based on the current state and the uniform motion model; in the update phase, the measured points in the new scan cycle are associated with the predicted positions. If the association is successful, the track state is updated; otherwise, it is marked as track lost. Recorded signal node information includes temporal sequence, spatial coordinates, and signal strength changes.

[0035] Step S117: Activate the trajectory interleaving unit, analyze the spatial overlapping areas and temporal associated nodes of different propagation paths, interleave the signal nodes and associated nodes in the overlapping areas, and generate interleaved obstacle signal trajectories.

[0036] The trajectory interleaving unit first analyzes the propagation paths of all abnormal fluctuation points to identify spatially overlapping areas and temporally correlated nodes between different paths. Spatially overlapping areas are determined by calculating the intersection of the minimum bounding boxes of the paths, while temporally correlated nodes are determined by comparing the timestamp differences between path nodes. Then, signal nodes and temporally correlated nodes within the overlapping areas are associated, based on spatial distance and time difference thresholds. If the spatial distance and time difference between two nodes are both less than the set thresholds, they are considered correlated. Finally, the associated nodes are arranged according to time sequence and spatial location to form an interleaved obstacle signal trajectory. The trajectory data includes the temporal index, spatial coordinates, signal strength, and associated node ID of each node.

[0037] Step S120: Construct a dynamic symbiotic feature model. Input the intertwined obstacle signal trajectory into a preset artificial intelligence framework, link the berth environment feature library, obstacle history feature library and real-time operation feature library, mine the dynamic symbiotic relationship between the trajectory and berth environment features, obstacle history features and real-time operation features, transform the symbiotic relationship into multiple association rules at different levels, and embed them into the framework to generate a dynamic symbiotic feature model.

[0038] Step S121: Input the interwoven obstacle signal trajectory into the preset artificial intelligence framework, which has a built-in feature association mining algorithm, dynamic learning module and rule embedding module.

[0039] The pre-defined artificial intelligence framework employs a deep learning framework, with a built-in graph neural network algorithm for feature association mining, capable of processing graph-structured data of interwoven obstacle signal trajectories. The dynamic learning module includes an online learning mechanism that adjusts model parameters based on new input trajectory data and feedback results. The rule embedding module transforms the mined association rules into connection weights and bias terms of the neural network, enabling either hard or soft encoding of the rules. The framework's network structure comprises an input layer, hidden layers, and an output layer. The number of nodes in the input layer matches the dimension of the trajectory features, the hidden layers employ a multilayer perceptron structure, and the output layer provides predictions of obstacle type and attributes.

[0040] Step S122: Call the berth environment feature library, which stores the flow characteristics, water level characteristics, berth structure characteristics, water temperature characteristics, and water quality characteristics of different berth scenarios and under different hydrological conditions.

[0041] The berth environmental feature database is a relational database that stores data including: the length, width, and water depth distribution of different berth scenarios; and the flow velocity, direction, water level, water temperature, water transparency, pH value, and dissolved oxygen content under different hydrological conditions. The data is indexed by time and space dimensions and supports queries by berth number, date, hydrological condition type, etc. When accessed, the database retrieves matching environmental feature data based on the current berth number and real-time hydrological monitoring data, such as the average flow velocity, water level, and water temperature for the current date.

[0042] Step S123: Call the obstacle history feature database, which stores the signal trajectory features, spatial morphology features, reflected signal features, attenuated signal features, and frequency change features of floating garbage, sunken debris, reefs, and abandoned fishing gear.

[0043] The obstacle history feature database is stored using a distributed file system. Each obstacle type corresponds to a dataset containing sample signal trajectory data, spatial morphological parameters, reflected signal amplitude, attenuation coefficient, frequency spectrum, etc. The data is stored in HDF5 format, supporting efficient read, write, and query operations. Upon retrieval, based on the preliminary classification results of interwoven obstacle signal trajectories, historical feature data of similar obstacle types is retrieved, such as the historical trajectory curvature and mean signal strength of floating debris.

[0044] Step S124: Call the real-time operation feature library, which stores the sensor signal feature thresholds corresponding to the robot's current operation time period, operation area, and operation mode.

[0045] The real-time operation feature library is an in-memory database that stores data including: light intensity and ambient noise levels during different operation periods; obstacle density and typical obstacle types in different operation areas; sensor operating parameters and signal characteristic thresholds for different operation modes. The data is updated in real time and pushed by the robot's operation control system. When accessed, parameters such as the corresponding sensing signal strength threshold and frequency change threshold are obtained based on the current operation period, operation area number, and operation mode.

[0046] Step S125: Extract the core features of the intertwined obstacle signal trajectory. The core features include trajectory curvature change, node density distribution, signal intensity gradient, trajectory intertwining frequency, and signal phase shift.

[0047] The trajectory curvature change is obtained by calculating the reciprocal of the radius of the arc formed by three adjacent nodes on the trajectory; the node density distribution is the number of signal nodes per unit volume; the signal intensity gradient is the ratio of the difference in signal intensity between adjacent nodes to their distance; the trajectory interleaving frequency is the number of times different trajectories intersect per unit time; and the signal phase shift is the difference between the phase of the reflected signal and the phase of the transmitted signal. Feature extraction is achieved using the sliding window method, with the window size determined based on the trajectory sampling rate. Feature calculations are performed on the trajectory data within each window to obtain the core feature vector.

[0048] Step S126: Extract environmental features from the berth environment feature library that match the trajectory acquisition, extract historical obstacle features from the obstacle history feature library that are similar to the core features, and extract the corresponding operation feature thresholds from the real-time operation feature library.

[0049] When extracting environmental features, data such as water flow velocity, water level, and water temperature in the same area during the same period are retrieved from the berth environmental feature database based on the timestamp and spatial coordinates of the trajectory acquisition. When extracting historical obstacle features, the K-nearest neighbor algorithm is used to calculate the distance between the current core feature vector and the feature vectors of each sample in the historical feature database, and the closest samples are selected as similar historical obstacle features. When extracting operational feature thresholds, the corresponding signal strength threshold, frequency change threshold, etc. are queried from the real-time operational feature database based on the current operational parameters.

[0050] Step S127: Activate the dynamic symbiotic mining module within the framework, divide the trajectory acquisition time into continuous time segments, divide the berth area at fixed intervals to form several spatiotemporal units, and extract the core features, environmental features, historical features, and operational features within each spatiotemporal unit.

[0051] The dynamic symbiotic mining module first divides the trajectory acquisition time into continuous time segments of equal length, with the segment length determined by the obstacle's movement speed. Then, it divides the berth area into a three-dimensional grid at fixed intervals, forming spatial regions. The time segments and spatial regions are combined to form spatiotemporal units. For each spatiotemporal unit, the core feature vector, environmental feature parameters, historical feature similarity values, and operational feature thresholds are extracted and stored in a feature matrix. Rows in the matrix correspond to spatiotemporal units, and columns correspond to feature terms.

[0052] Step S128: Mine the dependencies and changing trends of core features, environmental features, historical features, and operational features in different spatiotemporal units, form a dynamic symbiotic relationship description, and transform it into multiple association rules at different levels according to priority and association strength.

[0053] Step S1281: Construct a feature association analysis matrix, horizontally arranging all feature items of core features, environmental features, historical features and operational features, and vertically arranging the divided spatiotemporal units.

[0054] The feature association analysis matrix is ​​a two-dimensional table where rows represent spatiotemporal units and columns represent feature terms. Feature terms include core features such as trajectory curvature changes and node density distribution, environmental features such as water flow velocity and water temperature, historical features such as similarity values, and operational features such as threshold parameters. Matrix elements are the values ​​of feature terms within the corresponding spatiotemporal unit, such as the water flow velocity value and trajectory curvature change value of a certain spatiotemporal unit.

[0055] Step S1282: Extract the numerical sequence of each feature item within each spatiotemporal unit, and analyze the dependency relationship between each feature item and all other feature items using feature correlation analysis.

[0056] For each pair of feature terms in the matrix, their numerical sequences across all spatiotemporal units are extracted. The Pearson correlation coefficient is used to analyze the dependency relationship; the larger the absolute value of the correlation coefficient, the stronger the dependency. The direction of the correlation is also analyzed: a positive correlation indicates that an increase in one feature value leads to an increase in the other, while a negative correlation indicates that an increase in one feature value leads to a decrease in the other.

[0057] Step S1283: Record the existence status, association direction and synchronicity of changes of the dependency relationship to form a preliminary association record.

[0058] The existence state is categorized into strong dependency, moderate dependency, and weak dependency, based on the absolute value of the correlation coefficient; the association direction is categorized into positive correlation and negative correlation; the synchronicity of changes is determined by comparing whether the changing trends of the feature item numerical sequences are consistent. The preliminary association record contains information such as feature item pairs, existence state, association direction, and synchronicity of changes.

[0059] Step S1284: Track the changes in the dependency relationships of the same feature combination in different spatiotemporal units, and analyze the correlation between the changing trend and the spatiotemporal unit attributes.

[0060] For the feature combinations in the preliminary association records, we track the changes in parameters such as correlation coefficient and association direction in different spatiotemporal units, and analyze the relationship between these changes and the temporal and spatial attributes of the spatiotemporal units, such as whether the correlation coefficient increases with the increase of water temperature or decreases with the increase of distance from the shore.

[0061] Step S1285: Filter out the dependencies that change regularly with time and space, and supplement the triggering conditions and constraints of the relationship changes.

[0062] Regular changes refer to the monotonic or periodic changes in dependency parameters over time and space. Triggering conditions are the threshold values ​​of spatiotemporal unit attributes that cause the dependency to exhibit regular changes, such as the correlation coefficient starting to increase when water temperature exceeds a certain value; constraint conditions are the range of characteristic values ​​within which the dependency holds, such as the dependency being significant only when signal strength is within a certain interval.

[0063] Step S1286: Integrate the preliminary association records and the results of the trend analysis to form a dynamic symbiotic relationship description that includes feature combinations, dependencies, trends, and triggering conditions.

[0064] The description of dynamic symbiotic relationships is structured text, containing feature combination identifiers, dependency types, descriptions of change trends, triggering conditions, constraints, etc. Each description corresponds to a dynamic symbiotic relationship.

[0065] Step S129: Embed multiple association rules at different levels into the feature processing module and decision module of the artificial intelligence framework, test the effectiveness of the rules through the framework's built-in verification module, and generate a dynamic symbiotic feature model.

[0066] Association rules are categorized into three levels of priority: high, medium, and low. Priority is determined by the strength and importance of the dependency relationship. High-priority rules are embedded in the output layer of the decision module, directly influencing obstacle type judgment; medium-priority rules are embedded in the hidden layer, affecting feature combination weights; and low-priority rules are embedded in the feature processing module for feature selection. The validation module employs cross-validation, dividing historical data into training and validation sets. The training set is used to generate rules, while the validation set is used to test the recognition accuracy of the rules. When the accuracy reaches a set threshold, the rule is considered valid, and a dynamic co-occurrence feature model is generated.

[0067] Step S130: Perform hierarchical reverse calibration by calling up the historical obstacle identification data stored locally by the water cleaning robot and the actual measured obstacle data collected on site, feeding the dynamic symbiotic feature model in different batches, capturing the deviations between the predicted obstacle information and the actual obstacle information at different levels, adjusting the weights of the model association rules according to the level of deviation, and completing the hierarchical reverse calibration.

[0068] Step S131: Call the historical obstacle recognition data in the local storage unit of the water cleaning robot. The data covers the historical interwoven obstacle signal trajectories and actual obstacle information of different operation periods, floating garbage, sunken debris, reefs, and abandoned fishing gear.

[0069] The local storage unit is a solid-state drive array, employing RAID5 technology to ensure data reliability. Historical obstacle identification data is archived by timestamp, with each data file containing feature vectors of historical interwoven obstacle signal trajectories, as well as information such as the type, location, size, and material of the actual obstacle. When accessed, data is retrieved by time range and obstacle type through the file system interface.

[0070] Step S132: Collect actual obstacle data on site through manual annotation, obtain the on-site interwoven obstacle signal trajectory and corresponding actual obstacle information, and supplement it to the historical obstacle identification data to form a calibration data set.

[0071] Manual annotation is performed by professional operators using a robotic monitoring terminal. Based on underwater camera images and sensor data, the operators annotate information such as the type, location coordinates, and outline dimensions of the actual obstacles. The annotated data is correlated with the signal trajectories of interwoven obstacles on-site to form calibration samples, which are then added to historical data to constitute a calibration dataset.

[0072] Step S133: Divide the calibration data set into different batches according to floating garbage, sunken debris, reefs, abandoned fishing gear and operation scenarios, and feed the intertwined obstacle signal trajectory in each batch of data into the dynamic symbiotic feature model.

[0073] The calibration dataset is batched according to the combination of obstacle type and operating scenario. Each batch contains calibration samples of the same type of obstacle under the same operating scenario. The batching is achieved through a clustering algorithm to ensure that the samples in each batch are evenly distributed. During feeding, the interwoven obstacle signal trajectories are input into the dynamic symbiotic feature model in batch order, and the model outputs predicted obstacle information.

[0074] Step S134: Obtain the predicted obstacle information output by the model. The predicted obstacle information includes obstacle contour parameters, obstacle type identifier, obstacle location coordinates and obstacle signal features.

[0075] The predicted obstacle information is the model's output tensor. After decoding, the following parameters are obtained: obstacle contour parameters, including length, width, and height; obstacle type identifier, which is a numerical code, such as zero or one representing floating debris; obstacle location coordinates, which are three-dimensional geodetic coordinates; and obstacle signal characteristics, such as the predicted signal strength mean and frequency peak. The information is stored in JSON format, containing field names and corresponding values.

[0076] Step S135: Compare the predicted obstacle information with the corresponding actual obstacle information in the calibration dataset, divide the deviation levels according to core feature deviation, secondary feature deviation and edge feature deviation, and record the specific data differences of each level of deviation.

[0077] Core feature deviations include incorrect obstacle type identification, position coordinate deviation exceeding the set range, and contour parameter deviation exceeding the set proportion; secondary feature deviations include mean signal strength deviation and peak frequency deviation; edge feature deviations include phase shift deviation and polarization characteristic deviation. During comparison, the absolute and relative differences between the predicted and actual values ​​are calculated and recorded in the deviation record table.

[0078] Step S136: For different deviation levels, load the weight adjustment range calculation module, input the deviation data and the built-in deviation weight influence coefficient table, and calculate the basic adjustment range of each level association rule.

[0079] The weight adjustment calculation module calculates the adjustment magnitude based on the influence coefficient of the deviation level and the magnitude of the deviation value. In the influence coefficient table, the core feature deviation has the largest influence coefficient, followed by the secondary feature deviation, and the marginal feature deviation has the smallest. The adjustment magnitude is the product of the deviation value and the influence coefficient; positive deviations increase the rule weight, and negative deviations decrease the rule weight.

[0080] Step S137: Combine the current performance indicators of the dynamic symbiotic feature model to correct the basic adjustment range, formulate a hierarchical weight adjustment scheme, and locate the storage location of the weight parameters of each level association rule.

[0081] Step S1371: Test the current performance metrics of the dynamic symbiotic feature model, which include obstacle recognition accuracy, contour reconstruction accuracy, type judgment accuracy, and response time.

[0082] Obstacle recognition accuracy is the ratio of the number of correctly identified obstacle samples to the total number of samples; contour reconstruction accuracy is the root mean square error between the predicted contour parameters and the actual parameters; type identification accuracy is the ratio of the number of samples with correctly identified types to the total number of samples; response time is the average time for the model to process one sample. Testing is conducted by running a test dataset containing various obstacle samples.

[0083] Step S1372: Compare the current performance indicators with the preset performance standards, locate the shortcomings in the model performance, and determine the deviation level and specific content of the feature items corresponding to the shortcomings.

[0084] The preset performance standards are the set accuracy threshold, precision threshold, correctness threshold, and response time threshold. During comparison, the current performance metrics are compared with the preset standards to identify those that fail to meet the standards. The analysis then examines which deviation levels and features primarily affect these metrics; for example, a low type identification accuracy might be caused by incorrect type identification within the core feature deviation.

[0085] Step S1373: For the deviation level corresponding to the performance shortcoming, increase the weight adjustment range of the association rule at that level to enhance the improvement effect of the adjustment on the shortcoming.

[0086] Based on the severity of the performance bottleneck, the adjustment range of the corresponding deviation level is increased proportionally. For example, if the type judgment accuracy is less than five percentage points below the standard, the adjustment range of the core feature deviation level is multiplied by 1.5. For feature items related to the bottleneck, the adjustment range is increased separately; for example, additional adjustments are added to rules related to type identification errors.

[0087] Step S1374: Analyze the mutual influence of adjustment ranges at different deviation levels, correct the adjustment ranges to avoid adjustment conflicts, and ensure that the performance of each level reaches a balanced state after adjustment.

[0088] Adjustment conflicts refer to situations where adjustments at different levels have opposing effects on the same performance metric. For example, increasing the adjustment of core features might improve accuracy but reduce response time. Sensitivity analysis is used to assess the mutual impact of adjustment magnitudes, and trade-off adjustments are made to resolve conflicts, ensuring that performance metrics at all levels remain within preset standard ranges.

[0089] Step S1375: Based on the results of the rationality verification of the adjustment range, the final weight adjustment range of each hierarchical association rule is determined.

[0090] Reasonableness verification is achieved through performance testing of the adjusted model. Test indicators include the degree of reduction in bias at each level, the magnitude of improvement in performance indicators, and the degree of model overfitting. Based on the verification results, the adjustment range is fine-tuned to determine the final adjustment range value.

[0091] Step S138: Modify the weight parameter values ​​according to the adjustment plan, refeed the adjusted model to the validation dataset, repeat the adjustment and validation steps, and complete the stratified reverse calibration.

[0092] Based on the final adjustment range, modify the weight parameters of the association rules in the dynamic symbiotic feature model. This modification is achieved through the model parameter interface, supporting parameter location by hierarchy and rule ID. After modification, test the model performance using a validation dataset. If all performance metrics meet the standards, calibration is complete; otherwise, return to step S136 to recalculate the adjustment range until the model performance meets the standards.

[0093] Step S140: Reconstruct the holographic 3D contour of the obstacle, load the calibrated dynamic symbiotic feature model into the holographic contour analysis unit, input the spatial occupancy attributes and signal features of the interwoven obstacle signal trajectory analysis nodes, associate the attributes and features to form the basic contour framework, and fill in the detailed features to generate the holographic 3D contour information of the obstacle.

[0094] Step S141: Load the calibrated dynamic co-occurrence feature model into the holographic contour analysis unit, and initialize the analysis parameters, coordinate system and feature mapping rules within the unit.

[0095] The holographic contour analysis unit is a GPU-accelerated computing module. When loading the model, the model parameters are transferred from memory to GPU memory. The initialization analysis parameters include voxel resolution and contour smoothing factor. The coordinate system adopts the berth geodetic coordinate system. The feature mapping rule is a correspondence table between signal features and contour structure, such as high-intensity reflection signals corresponding to hard surfaces.

[0096] Step S142: Input the interwoven obstacle signal trajectory to the holographic contour analysis unit, and analyze the spatial occupancy attributes, signal reflection characteristics, signal attenuation characteristics and signal frequency characteristics of the signal nodes on the trajectory through the dynamic symbiotic feature model.

[0097] The spatial occupancy attributes of signal nodes include occupied volume and surface normal vector; signal reflection characteristics include reflection coefficient and polarization angle; signal attenuation characteristics include attenuation coefficient and propagation distance; and signal frequency characteristics include dominant frequency and bandwidth. Analysis is achieved through the forward propagation of the model. After trajectory data is input into the model, the attribute features of each node are output after feature extraction and classification layers.

[0098] Step S143: Activate the attribute association module, associate the spatial occupancy attributes of all signal nodes, arrange them in spatial coordinate order to form the basic outline framework of the obstacle, and mark the core area, transition area and edge area of ​​the framework.

[0099] The attribute association module employs a region growing algorithm. Starting from the seed node, it merges nodes that are spatially close and have similar occupancy attributes into regions, forming a 3D mesh framework. The core region is characterized by high node density and strong signal strength; the transition region is the transition zone between the core and edge regions; and the edge regions are characterized by sparse nodes and weak signal strength. Labeling is achieved by assigning region labels to mesh vertices.

[0100] Step S144: Extract detailed signal features of the interlaced obstacle signal trajectory. The detailed features include signal phase shift, signal polarization features, signal harmonic features, and signal pulse features.

[0101] Signal phase shift is the lag of the reflected signal phase relative to the transmitted signal; signal polarization characteristics include polarization direction angle and degree of polarization; signal harmonic characteristics are the amplitudes of frequency components that are integer multiples of the fundamental frequency in the spectrum; signal pulse characteristics include pulse width and pulse interval. Detail features are extracted from the original signal using wavelet transform and Hilbert-Huang transform.

[0102] Step S145: Establish the mapping relationship between detailed signal features and basic contour framework, and determine the framework coordinate interval and specific range of structural region corresponding to different types of detailed signal features.

[0103] Step S1451: Extract key parameters of detailed signal features, extract the offset angle and change period of signal phase offset, extract the polarization direction and intensity distribution of signal polarization features, extract the harmonic order and amplitude of signal harmonic features, and extract the pulse width and interval of signal pulse features.

[0104] The offset angle is the maximum value of the phase offset, and the change period is the period of phase offset change with time; the polarization direction is the mean value of the polarization angle, and the intensity distribution is the distribution curve of polarization degree in different directions; the harmonic order is the ratio of the harmonic frequency to the fundamental frequency, and the amplitude is the ratio of the amplitude of the harmonic component to the amplitude of the fundamental frequency; the pulse width is the duration of the pulse signal, and the interval is the time difference between adjacent pulses.

[0105] Step S1452: Analyze the coordinate system parameters and structural partitioning rules of the basic outline frame, and record the coordinate range, structural function and feature requirements of each region of the frame.

[0106] The coordinate system parameters include the origin position, coordinate axis direction, and unit; the structural partitioning rules are the thresholds for dividing the core area, transition area, and edge area; the coordinate range is the minimum and maximum value of each area in the X, Y, and Z axis directions; the structural function is the role of each area in the contour, such as the core area being the main obstacle; the feature requirements are the types of detailed features required by each area, such as the core area requiring high-resolution reflection features.

[0107] Step S1453: Call the feature location mapping rule library in the dynamic symbiotic feature model to associate the key parameters of the detailed signal features with the feature requirements of the framework structure region.

[0108] The feature location mapping rule base stores the correspondence between detailed feature parameters and structural regions, such as large harmonic amplitudes corresponding to the metal surface of the core region. A fuzzy matching algorithm is used to match the key parameters of the detailed features with the conditions in the rule base to find the corresponding structural regions.

[0109] Step S1454: Match the specific coordinate intervals of the frame and the specific ranges of the structural region corresponding to different detailed signal features through the mapping rule base.

[0110] During matching, the rule base is queried based on the detailed feature parameter values ​​to obtain the corresponding frame coordinate range, such as the X-axis coordinate range of the core region when the phase offset angle is within a certain range. The specific range of the structural region is the three-dimensional spatial range of that region within the frame, such as the X-coordinate of the core region from a certain value to a certain value, the Y-coordinate from a certain value to a certain value, and the Z-coordinate from a certain value to a certain value.

[0111] Step S1455: Generate a detailed feature frame location mapping table, and label the key parameters, corresponding frame coordinate intervals, structural regions and filling priorities of different types of detailed signal features.

[0112] The mapping table is a two-dimensional table. Rows represent detailed feature instances, and columns include feature type, key parameter name, parameter value, frame coordinate range, structural region name, and fill priority. Fill priority is determined based on the degree of influence of the feature on the contour accuracy, with key features in the core region having the highest priority.

[0113] Step S1456: Select typical obstacle samples to verify the accuracy of the mapping table and mark feature items whose mapping deviation exceeds the preset range.

[0114] Typical obstacle samples are standard samples of different types of obstacles. The overlap is calculated by comparing the coordinate range of the detailed feature mapping with the coordinate range of the actual contour of the sample. Feature items with an overlap of less than a set threshold are marked as deviation items.

[0115] Step S1457: Adjust the corresponding mapping rule parameters, re-verify the adjusted mapping table, until the mapping deviation of all feature items meets the preset standard, and form the final mapping relationship.

[0116] Adjust the rule parameters corresponding to the deviation items in the mapping rule base, such as modifying the interval threshold for the phase offset angle. Regenerate the mapping table and verify it, iterating multiple times until the overlap of all feature items reaches the set threshold, thus obtaining the final mapping relationship.

[0117] Step S146: Map the detail signal features one by one to the corresponding positions of the basic contour frame, and fill in the unfilled detail structures in the frame.

[0118] Based on the final mapping relationship, the parameter values ​​of each detailed signal feature are assigned to the grid nodes in the corresponding coordinate range of the basic contour frame, and the attribute values ​​of the nodes are filled, such as assigning the polarization direction angle to the corresponding nodes in the edge region. For the unfilled blank areas in the frame, interpolation is performed based on the feature values ​​of adjacent nodes to fill in the detailed structure.

[0119] Step S147: Activate the contour optimization module to optimize the smoothness of contour edges and the clarity of detail features, and generate complete obstacle holographic 3D contour information.

[0120] The contour optimization module employs a bilateral filtering algorithm to smooth noise while preserving contour edges; a sharpening algorithm enhances the contrast of detailed features, improving clarity. The optimized contour information is stored in a 3D point cloud format, containing attributes such as the coordinates, color (reflecting signal strength), and normal vector of each point.

[0121] Step S150: Output adaptive avoidance command, transmit the holographic 3D contour information of the obstacle to the motion control unit of the water cleaning robot, analyze the obstacle position, size and spatial posture in the contour information, combine the robot body length, width and height dimensions, the range of motion angles of each joint, the maximum output speed and torque of the drive motor, the minimum turning radius in the working state and the working scene, convert them into motion control parameters and encapsulate them into adaptive avoidance command.

[0122] For example, in step S151: the encrypted data transmission channel is activated to transmit the holographic three-dimensional contour information of the obstacle to the motion control unit of the water cleaning robot, and the data verification mechanism is activated during the transmission process.

[0123] The encrypted data transmission channel uses the AES encryption algorithm to encrypt the contour information before transmitting it via the CAN bus. The data verification mechanism uses CRC checksum, adding a checksum to the end of the data packet. The receiving end calculates the checksum and compares it with the sending end; if they match, the data is accepted; otherwise, a retransmission is requested.

[0124] Step S152: The motion control unit starts the contour analysis module to analyze the obstacle's three-dimensional coordinates, contour boundary dimensions, spatial attitude angles and surface features in the obstacle holographic three-dimensional contour information.

[0125] The contour analysis module processes the received point cloud data, calculates the three-dimensional coordinates (point cloud centroid) of the obstacle using principal component analysis, obtains the contour boundary dimensions (length, width, and height) by calculating the minimum bounding box of the point cloud, calculates the spatial attitude angles (roll angle, pitch angle, and yaw angle) by the statistical distribution of the point cloud normal vectors, and infers surface features (such as smoothness and hardness) by signal strength and polarization characteristics.

[0126] Step S153: Retrieve the robot's own size data, joint range of motion data, power output parameters and operation scenario limitation data, and store them in the parsing module's temporary cache.

[0127] The robot's own dimensions include its length, width, and height, as well as the extended dimensions of the cleaning device; joint range of motion data includes the rotation angle range of each joint of the robotic arm; power output parameters include the maximum speed and maximum torque of the drive motor; and operational scenario limitations include the minimum turning radius, maximum speed, and safe distance. This data is read from the robot's parameter configuration file and stored in the parsing module's memory cache.

[0128] Step S154: Associate the obstacle parsing data with the robot's own data and scene constraint data, and calculate the movement path, turning angle, movement range and execution sequence required for the robot to avoid obstacles.

[0129] The path planning uses the A* algorithm, starting from the robot's current position, ending at the target work point, and using obstacle contours as obstacles to plan a collision-free path. The steering angle is calculated based on the curvature of the path; the motion amplitude includes the change in the speed of the drive motor and the rotation angle of the servo motor; the execution timing is calculated based on the path length and speed to determine the start time and duration of each action.

[0130] Step S155: Convert the calculation results into motion control parameters for each actuator of the robot. The parameters include drive motor speed, steering servo angle, robotic arm posture and motion duration.

[0131] The drive motor speed is calculated based on the moving speed and reduction ratio; the steering servo angle is calculated based on the steering angle and transmission ratio; the robotic arm posture is obtained by solving inverse kinematics to obtain the angles of each joint; the duration of the action is calculated based on the amplitude of the action and the maximum speed of the actuator. The control parameters are digital quantities, and their range is within the rated range of the actuator.

[0132] Step S156: Start the parameter encapsulation module, encapsulate motion control parameters according to the robot motion control protocol, and add parameter verification code and execution priority identifier.

[0133] The action control protocol defines the format, order, and verification method of the parameters. During encapsulation, the control parameters are arranged according to the protocol format, a checksum is calculated as the verification code, and an execution priority identifier (high, medium, low) is added according to the importance of the action. The encapsulated data packet contains a header, parameter area, checksum, and priority field.

[0134] Step S157: Set the verification code verification process and priority execution logic, and embed the additional information segment of the encapsulated data.

[0135] The verification code verification process involves the receiving end calculating a checksum for the received parameter area and comparing it with the checksum. If they match, the action is executed; otherwise, the data is discarded. The priority execution logic is that higher-priority actions interrupt lower-priority actions, and actions of the same priority are executed sequentially. These processes and logic are embedded in the additional information segment of the data packet in the form of instruction codes.

[0136] Step S158: Verify the integrity and identifiability of the packaged data so that the motion control unit can correctly parse the verification code, priority identifier and motion control parameters.

[0137] Verification is achieved by simulating the receiving process. The motion control unit's parsing module analyzes the received encapsulated data, extracting the verification code, priority identifier, and control parameters. It checks whether the parameters are within the valid range and whether there are any logical conflicts. If the verification passes, the data is considered complete and identifiable; otherwise, it is re-encapsulated.

[0138] Step S159: Generate adaptive avoidance instructions that match the robot's action logic, transmit them to the robot's actuator control module, and start the instruction execution feedback mechanism.

[0139] The adaptive avoidance command is the final encapsulated data packet, transmitted to the actuator control module via the internal bus. The command execution feedback mechanism collects the execution status in real time through the actuator's position and force sensors and feeds it back to the motion control unit. If a deviation is detected, the control parameters are dynamically adjusted.

[0140] Figure 2 The illustration shows exemplary hardware and software components of a depth-sensing-based underwater cleaning robot berth obstacle recognition system 100, which can implement the ideas of this application, according to some embodiments of this application. For example, a processor 120 can be used in the depth-sensing-based underwater cleaning robot berth obstacle recognition system 100 and to perform the functions in this application.

[0141] The depth-sensing-based obstacle recognition system 100 for aquatic cleaning robots can be a general-purpose server or a special-purpose server; both can be used to implement the depth-sensing-based obstacle recognition method for aquatic cleaning robots described in this application. Although only one server is shown in this application, for convenience, the functions described in this application can be implemented in a distributed manner on multiple similar platforms to balance the processing load.

[0142] For example, the depth-sensing-based obstacle recognition system 100 for a marine cleaning robot may include a network port 110 connected to a network, one or more processors 120 for executing program instructions, a communication bus 130, and various forms of storage media 140, such as a disk, ROM, or RAM, or any combination thereof. Exemplarily, the depth-sensing-based obstacle recognition system 100 may also include program instructions stored in ROM, RAM, or other types of non-transitory storage media, or any combination thereof. The methods of this application can be implemented according to these program instructions. The depth-sensing-based obstacle recognition system 100 also includes an I / O interface 150 between the computer and other input / output devices.

[0143] For ease of explanation, only one processor is described in the depth-sensing-based underwater cleaning robot berth obstacle recognition system 100. However, it should be noted that the depth-sensing-based underwater cleaning robot berth obstacle recognition system 100 of this application may also include multiple processors. Therefore, the steps performed by one processor as described in this application may also be performed jointly by multiple processors or individually. For example, if the processor of the depth-sensing-based underwater cleaning robot berth obstacle recognition system 100 performs steps A and B, it should be understood that steps A and B may also be performed jointly by two different processors or individually by one processor. For example, the first processor performs step A, the second processor performs step B, or the first processor and the second processor jointly perform steps A and B.

[0144] Furthermore, this embodiment of the invention also provides a readable storage medium, wherein computer-executable instructions are preset in the readable storage medium, and when the processor executes the computer-executable instructions, the above-mentioned method for recognizing obstacles at berths of a depth-sensing-based waterborne cleaning robot is implemented.

[0145] It should be noted that, in order to simplify the description of the present invention and thus help to understand one or more embodiments of the invention, multiple features may sometimes be grouped into one embodiment, drawing or description thereof in the foregoing description of the embodiments of the present invention.

Claims

1. A method for recognizing obstacles at berths of a depth-sensing-based underwater debris-clearing robot, characterized in that, The method includes: The system performs multi-source signal trajectory interweaving and tracing, receives the original sensing signal output by the depth sensing component of the water cleaning robot, and simultaneously accesses the auxiliary sensing signal transmitted by the fixed sensing equipment around the berth. It then fuses the original sensing signal and the auxiliary sensing signal according to the time and space dimensions to generate a fused sensing signal, marks the abnormal fluctuation points in the fused sensing signal, tracks the propagation trajectory of the abnormal fluctuation points, and interweaves and associates the signal nodes of different trajectories to form an interwoven obstacle signal trajectory. A dynamic symbiotic feature model is constructed by inputting intertwined obstacle signal trajectories into a preset artificial intelligence framework, linking the berth environment feature library, obstacle history feature library, and real-time operation feature library, mining the dynamic symbiotic relationship between the trajectory and berth environment features, obstacle history features, and real-time operation features, transforming the symbiotic relationship into multiple association rules at different levels, and embedding them into the framework to generate a dynamic symbiotic feature model. Hierarchical reverse calibration is implemented by calling up historical obstacle identification data stored locally on the water cleaning robot and actual obstacle measurement data collected on site, feeding dynamic symbiotic feature models in different batches, capturing the deviations at different levels between predicted obstacle information and actual obstacle information, adjusting the model association rule weights according to the level of deviation, and completing the hierarchical reverse calibration. Reconstruct the holographic 3D contour of the obstacle, load the calibrated dynamic symbiotic feature model into the holographic contour analysis unit, input the spatial occupancy attributes and signal features of the interwoven obstacle signal trajectory analysis node, associate the attributes and features to form the basic contour framework, fill in the detailed features to generate the holographic 3D contour information of the obstacle. The system outputs adaptive avoidance commands, transmitting the holographic 3D contour information of the obstacle to the motion control unit of the waterborne cleaning robot. It analyzes the obstacle's position, size, and spatial posture in the contour information, and combines the robot's body length, width, and height dimensions, the range of motion angles of each joint, the maximum output speed and torque of the drive motor, the minimum turning radius in the working state, and the working scenario to convert them into motion control parameters and encapsulate them into adaptive avoidance commands.

2. The method for recognizing obstacles at berths of a depth-sensing-based underwater cleaning robot according to claim 1, characterized in that, The process of performing multi-source signal trajectory interweaving and tracing involves receiving the original sensing signal output by the depth sensing component of the waterborne cleaning robot, simultaneously accessing the auxiliary sensing signal transmitted by fixed sensing devices around the berth, fusing the original and auxiliary sensing signals according to the time and spatial dimensions to generate a fused sensing signal, marking abnormal fluctuation points in the fused sensing signal, tracking the propagation trajectory of abnormal fluctuation points, and interweaving and associating signal nodes of different trajectories to form an interwoven obstacle signal trajectory, including: The signal receiving module is activated to receive the raw sensing signals output by the depth sensing component of the waterborne cleaning robot. The raw sensing signals include the depth reflection signals, spatial position signals, signal attenuation signals, and signal frequency signals of the signal points in the berth area. The auxiliary sensing signals transmitted by fixed sensing devices around the berth are accessed through a preset communication protocol. The auxiliary sensing signals include depth signals in the berth edge area, near-shore obstacle reflection signals, water flow disturbance signals, and water temperature-related signals. The signal fusion unit is activated, a time-space dual-dimensional alignment algorithm is loaded, the timestamp information of the original sensing signal and the auxiliary sensing signal is extracted and the signal sequence is arranged in order, and the spatial coordinates are extracted and the coordinate system is unified through a coordinate transformation algorithm. Align the original sensing signal and the auxiliary sensing signal at the same coordinate point and the same timestamp, perform amplitude superposition and phase calibration, complete the signal superposition of all coordinate points and timestamps, and generate a fused sensing signal; Load the signal fluctuation recognition unit, retrieve the berth normal sensing signal range data built into the unit, cover all signal points of the fused sensing signal according to the preset scanning order, and compare the signal point values ​​with the normal range data; Signal points with values ​​exceeding the normal range are marked as abnormal fluctuation points. The trajectory tracking unit is activated to track the propagation path of the abnormal fluctuation points and record the temporal sequence, spatial coordinates, and signal strength changes of the signal nodes along the path. The trajectory interleaving unit is activated to analyze the spatial overlap areas and temporal associated nodes of different propagation paths, interleave the signal nodes and associated nodes within the overlap areas, and generate interleaved obstacle signal trajectories.

3. The method for recognizing obstacles at berths of a depth-sensing-based underwater debris-clearing robot according to claim 1, characterized in that, The construction of the dynamic symbiotic feature model involves inputting interwoven obstacle signal trajectories into a pre-defined artificial intelligence framework, linking the berth environment feature library, obstacle history feature library, and real-time operation feature library, mining the dynamic symbiotic relationships between the trajectories and berth environment features, obstacle history features, and real-time operation features, transforming the symbiotic relationships into multiple association rules at different levels, and embedding them into the framework to generate the dynamic symbiotic feature model, including: The interwoven obstacle signal trajectory is input into a preset artificial intelligence framework, which has built-in feature association mining algorithm, dynamic learning module and rule embedding module. Call the berth environment feature library, which stores water flow features, water level features, berth structure features, water temperature features, and water quality features under different berth scenarios and hydrological conditions; Access the obstacle history feature database, which stores the signal trajectory features, spatial morphology features, reflected signal features, attenuated signal features, and frequency change features of floating garbage, sunken debris, reefs, and abandoned fishing gear; Call the real-time operation feature library, which stores the sensor signal feature thresholds corresponding to the robot's current operation time, operation area, and operation mode; The core features of the intertwined obstacle signal trajectory are extracted. These core features include trajectory curvature variation, node density distribution, signal intensity gradient, trajectory intertwining frequency, and signal phase shift. Extract environmental features from the berth environment feature library that match the trajectory acquisition, extract historical obstacle features from the obstacle history feature library that are similar to the core features, and extract the corresponding operation feature thresholds from the real-time operation feature library; The dynamic symbiotic mining module within the framework is activated, the trajectory acquisition time is divided into continuous time segments, the berth area is divided at fixed intervals to form several spatial regions, and the core features, environmental features, historical features and operational features within each spatiotemporal unit are extracted. The dependencies and trends of core features, environmental features, historical features, and operational features within different spatiotemporal units are explored to form a dynamic symbiotic relationship description, which is then transformed into multiple association rules at different levels according to priority and association strength. Multiple association rules at different levels are embedded into the feature processing and decision-making modules of an artificial intelligence framework. The validity of the rules is tested through the framework's built-in verification module, generating a dynamic symbiotic feature model.

4. The method for recognizing obstacles at berths of a depth-sensing-based underwater debris-clearing robot according to claim 1, characterized in that, The hierarchical reverse calibration involves calling upon historical obstacle identification data stored locally by the waterborne debris removal robot and actual obstacle measurement data collected on-site. This data is then fed into a dynamic symbiotic feature model in different batches to capture deviations at different levels between predicted and actual obstacle information. The model association rule weights are adjusted according to the level of deviation to complete the hierarchical reverse calibration, including: The system retrieves historical obstacle recognition data from the local storage unit of the waterborne debris removal robot. The data covers historical interwoven obstacle signal trajectories and actual obstacle information for different operating periods, floating garbage, sunken debris, reefs, and abandoned fishing gear. The actual obstacle data measured on site was collected by manual annotation, and the interwoven obstacle signal trajectory and corresponding actual obstacle information were obtained to supplement the historical obstacle identification data to form a calibration data set. The calibration data sets are divided into floating debris, sunken debris, reefs, abandoned fishing gear and operation scenarios, and then the interwoven obstacle signal trajectories in each batch of data are fed into the dynamic symbiotic feature model. Obtain the predicted obstacle information output by the model. The predicted obstacle information includes obstacle contour parameters, obstacle type identifier, obstacle location coordinates and obstacle signal features. Compare the predicted obstacle information with the corresponding actual obstacle information in the calibration dataset, divide the deviation levels according to core feature deviation, secondary feature deviation and marginal feature deviation, and record the specific data differences of each level of deviation. For different deviation levels, load the weight adjustment range calculation module, input deviation data and the built-in deviation weight influence coefficient table, and calculate the basic adjustment range of each level of association rule; Based on the current performance indicators of the dynamic symbiotic feature model, the basic adjustment range is corrected, a hierarchical weight adjustment scheme is formulated, and the storage location of the weight parameters of each level association rule is located. Modify the weight parameter values ​​according to the adjustment plan, refeed the adjusted model to the validation dataset, repeat the adjustment and validation steps, and complete the stratified reverse calibration.

5. The method for recognizing obstacles at berths of a depth-sensing-based underwater cleaning robot according to claim 1, characterized in that, The reconstructed obstacle holographic 3D contour involves loading the calibrated dynamic symbiotic feature model into the holographic contour analysis unit, inputting the spatial occupancy attributes and signal features of the interwoven obstacle signal trajectory analysis nodes, associating attributes and features to form a basic contour framework, and filling in detailed features to generate obstacle holographic 3D contour information, including: The calibrated dynamic co-occurrence feature model is loaded into the holographic contour analysis unit, and the analysis parameters, coordinate system and feature mapping rules within the unit are initialized. The interwoven obstacle signal trajectory is input to the holographic contour analysis unit, and the spatial occupancy attributes, signal reflection characteristics, signal attenuation characteristics and signal frequency characteristics of the signal nodes on the trajectory are analyzed through the dynamic symbiotic feature model. Start the attribute association module to associate the spatial occupancy attributes of all signal nodes, arrange them in spatial coordinate order to form the basic outline framework of the obstacle, and mark the core area, transition area and edge area of ​​the framework. Extract detailed signal features from the interleaved obstacle signal trajectory. These features include signal phase shift, signal polarization characteristics, signal harmonic characteristics, and signal pulse characteristics. Establish the mapping relationship between detailed signal features and basic contour framework, and determine the framework coordinate interval and specific range of structural region corresponding to different types of detailed signal features; The detailed signal features are mapped one by one to the corresponding positions of the basic contour frame to fill in the unfilled detailed structures in the frame; The contour optimization module is activated to improve the smoothness of contour edges and the clarity of detailed features, generating complete holographic 3D contour information of the obstacle.

6. The method for recognizing obstacles at berths of a depth-sensing-based underwater cleaning robot according to claim 2, characterized in that, The extraction of spatial coordinates unifies the coordinate system through a coordinate transformation algorithm, including: Extract the coordinate system parameters of the original sensing signal, including the position of the coordinate origin, the direction of the coordinate axes, the coordinate units, and the coordinate precision; Extract the coordinate system parameters of the auxiliary sensing signal and record the differences between the original sensing signal and the auxiliary sensing signal coordinate system in terms of origin, direction, unit and precision; Load the preset coordinate transformation algorithm, input the coordinate system parameters of the original sensing signal and the auxiliary sensing signal, and calculate the transformation matrix and deviation correction coefficient; Substitute all the spatial coordinates of the auxiliary sensing signal into the transformation matrix, and calculate the transformed coordinate data by combining the deviation correction coefficient; Select uniformly distributed feature signal points and compare the consistency between the original sensing signal coordinates and the converted auxiliary sensing signal coordinates; Adjust the transformation matrix parameters and deviation correction coefficients to correct the transformation deviation until the coordinates of the original sensing signal and the auxiliary sensing signal are consistent with the preset standard, thus completing the coordinate system unification.

7. The method for recognizing obstacles at berths of a depth-sensing-based underwater cleaning robot according to claim 3, characterized in that, The mining of core features, environmental features, historical features, and operational features reveals their dependencies and changing trends across different spatiotemporal units, forming a dynamic symbiotic relationship description, including: Construct a feature association analysis matrix, horizontally arranging all feature items of core features, environmental features, historical features and operational features, and vertically arranging the divided spatiotemporal units; Extract the numerical sequence of each feature item within each spatiotemporal unit, and analyze the dependency relationship between each feature item and all other feature items using feature correlation analysis. Record the existence status, association direction, and synchronicity of changes of dependencies to form a preliminary association record; Track the changes in the dependencies of the same feature combination in different spatiotemporal units, and analyze the correlation between the changing trend and the spatiotemporal unit attributes; Filter out dependencies that exhibit regular changes over time and space, and supplement the triggering and constraint conditions for these changes. By integrating preliminary correlation records and trend analysis results, a dynamic symbiotic relationship description is formed, which includes feature combinations, dependencies, trends, and triggering conditions.

8. The method for recognizing obstacles at berths of a depth-sensing-based underwater cleaning robot according to claim 4, characterized in that, The adjustment range of the current performance index correction based on the dynamic symbiotic feature model includes: Test the current performance metrics of the dynamic symbiotic feature model, which include obstacle recognition accuracy, contour reconstruction accuracy, type judgment accuracy, and response time; By comparing the current performance metrics with the preset performance standards, we can identify the shortcomings in the model's performance and determine the corresponding deviation levels and specific features of the shortcomings. For the deviation level corresponding to the performance weakness, increase the weight adjustment range of the association rule at that level to enhance the improvement effect of the adjustment on the weakness; Analyze the mutual influence of adjustment ranges at different deviation levels, correct the adjustment ranges to avoid adjustment conflicts, and ensure that the performance of each level reaches a balanced state after adjustment. Based on the results of the rationality verification of the adjustment range, the final weight adjustment range of each level association rule is determined.

9. The method for recognizing obstacles at berths of a depth-sensing-based underwater cleaning robot according to claim 5, characterized in that, The process of establishing the mapping relationship between detailed signal features and the basic contour framework includes: Key parameters for extracting detailed signal features: signal phase shift (extracting offset angle and change period), signal polarization features (extracting polarization direction and intensity distribution), signal harmonic features (extracting harmonic order and amplitude), and signal pulse features (extracting pulse width and interval). Analyze the coordinate system parameters and structural partitioning rules of the basic outline frame, and record the coordinate range, structural function and feature requirements of each region of the frame; Call the feature location mapping rule library in the dynamic symbiotic feature model to associate the key parameters of detailed signal features with the feature requirements of the framework structure region; The specific coordinate range, structural region, and filling priority of the frame corresponding to different detailed signal features are matched by a mapping rule base. Generate a detailed feature frame location mapping table, and label the key parameters, corresponding frame coordinate ranges, structural regions and filling priorities of different types of detailed features; Select typical obstacle samples to verify the accuracy of the mapping table, mark feature items whose mapping deviation exceeds the preset range, and adjust the corresponding mapping rule parameters; Re-verify the adjusted mapping table until the mapping deviations of all feature items meet the preset standards, thus forming the final mapping relationship.

10. A depth-sensing-based obstacle recognition system for a waterborne cleaning robot, characterized in that, The depth-sensing-based underwater cleaning robot berth obstacle recognition system includes a processor and a memory, the memory and the processor being connected. The memory is used to store programs, instructions or code, and the processor is used to execute the programs, instructions or code in the memory to implement the depth-sensing-based underwater cleaning robot berth obstacle recognition method according to any one of claims 1-9.

Images