An intelligent detection and navigation system and method for inland waterway vessels

Abstract

The application discloses an intelligent detection and navigation system and method for inland river ships, which comprises the following steps: step one, collecting multi-modal perception data during the navigation of the inland river ship; step two, constructing an initial cognitive graph; step three, dynamically modeling each semantic state node by using an improved liquid-state neuron network, embedding a symbolic logic constraint in state transition, and generating a navigation control partial derivative vector; step four, calculating the tension value of a causal directed edge and performing a structure evolution operation to generate a self-evolution cognitive graph; step five, constructing a candidate path set; step six, calculating the graph winding degree index of each candidate path; step seven, constructing a path decision vector and performing a path backstepping reduction operation to screen a target optimal path; and step eight, converting the target optimal path into a navigation control instruction and outputting the navigation control instruction to a ship control system. The application combines cognitive graph modeling and neural network reasoning to realize intelligent navigation control of the inland river ship.

Description

Technical Field

[0001] This invention relates to the field of intelligent ship navigation technology, and in particular to an intelligent detection and navigation system and method for inland waterway vessels. Background Technology

[0002] With the rapid development of intelligent inland waterway navigation technology, the demand for efficient, safe, and autonomous navigation of vessels in complex water environments is constantly increasing. Existing inland waterway vessel navigation methods mainly rely on single sensor data for environmental perception and employ rule-driven or traditional search algorithms to complete path planning and control command generation. In practical applications, these methods generally suffer from the following problems:

[0003] Current systems have limited capabilities in fusing and modeling multimodal sensing data, making it difficult to effectively integrate heterogeneous information from multiple sources such as radar, optical images, infrared images, hydrological information, sonar data, and voice commands. This results in fragmented sensing results and a lack of unified semantic expression. The cognitive modeling stage fails to establish a dynamically evolving knowledge structure, and there is insufficient modeling of the temporal relationships and causal structures between semantic states, making it difficult to support state reasoning and map updates in complex navigation tasks. The path planning process lacks quantitative analysis of comprehensive factors such as conflict events in the navigation history, the frequency of human intervention, and path complexity. Path selection often deviates from the optimal decision, leading to navigation interruptions or frequent corrections. The generation of control commands is usually achieved through fixed mapping or static strategies, failing to dynamically adjust the navigation strategy based on control bias variables between semantic states, resulting in large path tracking errors and insufficient stability.

[0004] Therefore, how to provide an intelligent detection and navigation system and method for inland waterway vessels is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an intelligent detection and navigation system and method for inland waterway vessels. This invention integrates semantic graph modeling, improved liquid neural networks, and navigation control partial derivation mechanisms. By constructing a self-evolving cognitive graph, it achieves intelligent path selection and dynamic navigation control for vessels in complex inland waterway environments. It has the advantages of strong semantic understanding, high path selection accuracy, and good navigation execution stability. It can effectively cope with interference from multiple factors such as path conflicts, structural evolution, and control sensitivity, thereby improving the autonomous decision-making ability and safe passage efficiency of the inland waterway intelligent shipping system.

[0006] An intelligent detection and navigation method for inland waterway vessels according to an embodiment of the present invention includes the following steps:

[0007] Step 1: Collect multimodal perception data during the navigation of inland waterway vessels and construct a semantic tensor;

[0008] Step 2: Construct an initial cognitive graph based on the semantic tensor, including semantic state nodes, causal directed edges, and edge weight values;

[0009] Step 3: Use an improved liquid neural network to dynamically model each semantic state node in the initial cognitive map, embed symbolic logic constraints in the state transition, and generate navigation control bias variables for the semantic state nodes;

[0010] Step 4: Calculate the tension value of the causal directed edge based on the edge weight value and navigation control bias, and perform structural evolution operation on the causal directed edge whose tension value is greater than the preset tension threshold to generate a self-evolving cognitive graph.

[0011] Step 5: In the self-evolutionary cognitive graph, extract multiple candidate paths from the current semantic state node to the target semantic state node, and construct a candidate path set;

[0012] Step 6: Based on the out-degree value, conflict index, and historical human intervention frequency of the semantic state nodes in the candidate paths, calculate the graph entanglement index of each candidate path;

[0013] Step 7: Based on the graph entanglement index, maximum node conflict index, average tension value, and average magnitude of the navigation control bias vector of all semantic state nodes for each candidate path, construct a path decision vector, and perform a path back-pushing reduction operation on the candidate path set to filter and obtain the target optimal path.

[0014] Step 8: Convert the target optimal path into navigation control commands and output them to the ship control system for path tracking and navigation.

[0015] Optionally, the multimodal sensing data specifically includes radar data, optical images, infrared images, hydrological information, sonar data, and voice command information. By extracting features from different modal sensing data and uniformly encoding them, and then splicing and fusing them, a semantic tensor is generated.

[0016] Optionally, step two specifically involves:

[0017] Based on the semantic tensor, multimodal fusion features at different time steps are extracted as candidate state units;

[0018] The candidate state units are clustered according to the cosine similarity between the multimodal fusion features. Candidate state units with a cosine similarity greater than a set similarity threshold are merged to generate semantic state nodes. The multimodal fusion features of the merged candidate state units are averaged and weighted and concatenated to obtain the aggregated features of the semantic state nodes.

[0019] Based on the temporal sequence relationship in the semantic tensor, causal directed edges are established between adjacent semantic state nodes to form state transition channels between semantic state nodes;

[0020] For each causal directed edge, the edge weight value is calculated as follows: the cumulative number of times the semantic state node connected by the causal directed edge moves from the previous semantic state node to the next semantic state node in the historical samples is counted and multiplied by a preset first weighting factor to obtain a first calculated value; the aggregate feature Euclidean distance between the semantic state nodes connected by the causal directed edge is calculated and multiplied by a preset second weighting factor to obtain a second calculated value; the first calculated value is subtracted from the second calculated value, and the result is subjected to non-negativity constraint processing. If it is less than zero, it is set to zero, which is used as the edge weight value of the causal directed edge;

[0021] An initial cognitive graph is generated based on the semantic state nodes, causal directed edges, and edge weight values.

[0022] Optionally, step three specifically includes:

[0023] The aggregated features of each semantic state node in the initial cognitive graph are used as input, and dynamic modeling is performed through an improved liquid neural network. The improved liquid neural network includes an input encoding module, a liquid state evolution module, a symbolic logic embedding module, and a partial derivative output module.

[0024] The input encoding module is used to normalize and encode the aggregated features of semantic state nodes to generate a time step input vector sequence.

[0025] The liquid state evolution module is used to construct a state update structure with temporal memory capability. At each time step, the current input vector is weighted and combined with the state response vector generated at the previous time step, and mapped through the hyperbolic tangent function to obtain the state response vector at the current time step.

[0026] The symbolic logic embedding module is used to embed symbolic logic constraints during the state update process. It constructs a propositional symbolic logic expression based on the reachability rules between semantic state nodes, and converts the symbolic logic expression into a Boolean logic vector with the same dimension as the state response vector, where a value of 1 indicates that the logical constraint is satisfied, and a value of 0 indicates that the logical constraint is violated.

[0027] Perform a Hadamard product operation on the state response vector at each time step and the corresponding Boolean logic vector, and assign zero to the position dimension of the state response vector that violates the logical constraints, so as to realize real-time logical control of the state evolution path.

[0028] The partial derivative output module is used to generate navigation control bias parameters corresponding to each semantic state node, specifically including:

[0029] The state response vector after embedding symbolic logic constraints is dimensionally aligned with the feature vector of the preset navigation target state, and the numerical difference in each feature dimension is calculated to obtain the navigation deviation vector.

[0030] The navigation deviation vector is sequentially mapped to the rate of change of each feature dimension. Using numerical differentiation, the influence rate of the change of each feature dimension on the target state difference is calculated to obtain the navigation control deviation vector of the semantic state node.

[0031] Optionally, step four specifically involves:

[0032] The first tension calculation value is obtained by multiplying the edge weight value of the causal directed edge with the magnitude of the navigation control bias vector of the front-end semantic state node.

[0033] The second tension calculation value is obtained by multiplying the edge weight value of the causal directed edge with the magnitude of the navigation control bias vector of the backend semantic state node.

[0034] The first tension calculation value and the second tension calculation value are weighted and summed to obtain the tension value of the causal directed edge;

[0035] The tension value is compared with a preset tension threshold to identify causal directed edges where the tension value is greater than the preset tension threshold.

[0036] For each causal directed edge whose tension value is greater than a preset tension threshold, perform a structural evolution operation, including:

[0037] Starting from the current semantic state node, search backwards in chronological order for a set of candidate semantic state nodes within a preset number of hops.

[0038] In the set of candidate semantic state nodes, select candidate semantic state nodes that have historical transfer records with the current semantic state node and whose aggregate feature Euclidean distance is less than a preset distance threshold.

[0039] Establish a new causal directed edge between the current semantic state node and the selected target candidate semantic state node, and calculate the edge weight value based on the historical transition count and the Euclidean distance of the aggregation feature.

[0040] Delete causal directed edges whose original tension values ​​are greater than the preset tension threshold, and retain the newly created edge structure;

[0041] Based on the updated semantic state nodes, causal directed edges, and edge weights, a self-evolving cognitive graph is constructed.

[0042] Optionally, step five specifically includes:

[0043] In the self-evolutionary cognitive graph, the current semantic state node is set as the starting point and the target semantic state node is set as the ending point. Path search is performed using a graph traversal method with a limited maximum number of hops.

[0044] Starting from the current semantic state node, traverse forward along the causal directed edges in sequence, jumping to a semantic state node that has a causal edge connection with the current semantic state node and has a state transition record in the historical samples, until the target semantic state node is reached or the preset maximum number of jumps is reached.

[0045] The semantic state node sequences generated during the traversal process are filtered, and ordered semantic state node sequences that meet the requirements of continuous connection, no node repetition, and historical transition records at each step are selected as candidate paths, and a candidate path set is constructed.

[0046] Optionally, step six specifically includes:

[0047] Extract the out-degree value of each semantic state node in the candidate path. The out-degree value is the number of subsequent semantic state nodes connected to the semantic state node in the self-evolutionary cognitive graph.

[0048] Extract the conflict index for each semantic state node. The conflict index is the total number of conflict events that have occurred during the historical navigation process for that semantic state node. The conflict events include:

[0049] An event occurs where the ship reverts to its path after entering the semantic state node;

[0050] The event in which the ship changes its navigation instructions at the semantic state node;

[0051] The event that triggers navigation interruption by the ship at the semantic state node;

[0052] Statistical analysis of the frequency of human intervention for each semantic state node in historical navigation tasks;

[0053] The entanglement contribution value of the semantic state node is obtained by multiplying the out-degree value, conflict index, and associated historical human intervention frequency of the semantic state node by the corresponding preset weighting factor.

[0054] The entanglement contribution values ​​of all semantic state nodes in the candidate path are accumulated to obtain the graph entanglement index of the candidate path.

[0055] Optionally, step seven specifically includes:

[0056] Extract the graph entanglement index of candidate paths as a path complexity factor;

[0057] Extract the conflict index of all semantic state nodes in the candidate path, and select the maximum value as the maximum node conflict index, which serves as the path conflict risk factor.

[0058] Extract the tension values ​​of all causal directed edges in the candidate path, calculate the average tension value, and use it as the path tension load element;

[0059] The magnitude of the navigation control bias vectors of all semantic state nodes in the candidate path is statistically analyzed, and the average magnitude is calculated as the overall control sensitivity element of the path.

[0060] The path complexity elements, path conflict risk elements, path tension load elements, and path overall control sensitivity elements are arranged in sequence and combined to form a path decision vector.

[0061] Perform a path backpropagation and reduction operation on the candidate path set, the path backpropagation and reduction operation including:

[0062] Starting from the target semantic state node, trace back along the candidate path to the current semantic state node and record the corresponding path decision vector;

[0063] The multiple candidate paths obtained through reverse tracing are compared layer by layer, and candidate paths whose graph entanglement index in the path decision vector exceeds the preset entanglement threshold are eliminated.

[0064] Among the remaining candidate paths, candidate paths whose maximum node conflict index in the path decision vector exceeds a preset conflict threshold are further eliminated;

[0065] If multiple candidate paths still exist, select the candidate path with the smallest average tension value; if multiple candidate paths with the smallest average tension value exist, further select the candidate path with the largest average navigation control bias magnitude length as the target optimal path.

[0066] Optionally, step eight specifically includes:

[0067] Arrange the navigation control bias variables corresponding to each semantic state node in the target optimal path in chronological order to form a path control sequence;

[0068] The ship's turning angle is determined based on the directional change of each navigation control deflection in the path control sequence, and the ship's speed is determined based on the magnitude of the navigation control deflection.

[0069] The steering angle and the travel speed are encoded into navigation control commands and output sequentially to the ship control system to complete path tracking navigation.

[0070] An intelligent detection and navigation system for inland waterway vessels according to an embodiment of the present invention includes the following modules:

[0071] The perception data acquisition module is used to collect multimodal perception data during the navigation of inland waterway vessels and construct semantic tensors;

[0072] The cognitive graph construction module is used to construct an initial cognitive graph based on the semantic tensor, and generate semantic state nodes, causal directed edges and edge weight values;

[0073] The dynamic modeling module is used to model semantic state nodes by improving the liquid neural network, embed symbolic logic constraints in the state transition, and generate navigation control bias variables.

[0074] The graph evolution module is used to calculate the tension value of causal directed edges and perform structural evolution operations on causal directed edges whose tension values ​​are greater than a preset tension threshold to generate a self-evolving cognitive graph.

[0075] The path search module is used to extract multiple candidate paths from the current semantic state node to the target semantic state node in the self-evolving cognitive graph and construct a candidate path set.

[0076] The entanglement calculation module is used to calculate the graph entanglement index of each candidate path based on the out-degree value, conflict index, and historical human intervention frequency of the semantic state node.

[0077] The path selection module is used to construct a path decision vector based on the candidate path's graph entanglement index, maximum node conflict index, average tension value, and average magnitude of the navigation control bias vector, and to perform a path back-pushing and reduction operation to select the target optimal path.

[0078] The control command output module is used to convert the target optimal path into navigation control commands and output them to the ship control system for path tracking and navigation.

[0079] The beneficial effects of this invention are:

[0080] This invention addresses the problems of insufficient multimodal perception fusion, lack of global optimization capability in path planning, and unstable control output of inland waterway vessels in complex aquatic environments by constructing an integrated framework of self-evolving cognitive graph and improved liquid neural network. It proposes a full-process navigation method from semantic tensor construction to path control command output. In the semantic graph construction stage, a semantic tensor is generated through unified encoding and fusion of multimodal perception data, constructing an initial cognitive graph containing semantic state nodes, causal directed edges, and edge weights. Navigation control bias variables are dynamically generated by combining symbolic logic constraints. During graph evolution, a tension mechanism is introduced. By identifying edges with abnormal tension and performing structural reconstruction, the graph undergoes adaptive evolution, improving its expressive ability in terms of path complexity and state transition rationality. In the path planning stage, the graph entanglement index, maximum node conflict index, average tension value, and all semantic state nodes of each candidate path are integrated. The average magnitude of the navigation control bias vector is used to construct the path decision vector. Combined with path back-calculation and reduction operations, multi-level path selection is achieved, ensuring that path selection has low complexity, low risk, and high controllability. In the control output stage, the steering angle and speed of the ship are determined by the directional changes and magnitude of the navigation control bias vector in the path control sequence, and the output is a continuous and executable navigation control command. Ultimately, intelligent navigation capability with path evolution, interpretable decision-making, and implementable control driven by multi-source information is realized, which significantly improves the safety of autonomous navigation and mission completion efficiency of ships in inland waterway scenarios. Attached Figure Description

[0081] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0082] Figure 1 This is an overall flowchart of an intelligent detection and navigation method for inland waterway vessels proposed in this invention;

[0083] Figure 2 This is a schematic diagram of the structure of an intelligent detection and navigation system for inland waterway vessels proposed in this invention. Detailed Implementation

[0084] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0085] refer to Figure 1 A method for intelligent detection and navigation of inland waterway vessels includes the following steps:

[0086] Step 1: Collect multimodal perception data during the navigation of inland waterway vessels and construct a semantic tensor;

[0087] Step 2: Construct an initial cognitive graph based on the semantic tensor, including semantic state nodes, causal directed edges, and edge weight values;

[0088] Step 3: Use an improved liquid neural network to dynamically model each semantic state node in the initial cognitive map, embed symbolic logic constraints in the state transition, and generate navigation control bias variables for the semantic state nodes;

[0089] Step 4: Calculate the tension value of the causal directed edge based on the edge weight value and navigation control bias, and perform structural evolution operation on the causal directed edge whose tension value is greater than the preset tension threshold to generate a self-evolving cognitive graph.

[0090] Step 5: In the self-evolutionary cognitive graph, extract multiple candidate paths from the current semantic state node to the target semantic state node, and construct a candidate path set;

[0091] Step 6: Based on the out-degree value, conflict index, and historical human intervention frequency of the semantic state nodes in the candidate paths, calculate the graph entanglement index of each candidate path;

[0092] Step 7: Based on the graph entanglement index, maximum node conflict index, average tension value, and average magnitude of the navigation control bias vector of all semantic state nodes for each candidate path, construct a path decision vector, and perform a path back-pushing reduction operation on the candidate path set to filter and obtain the target optimal path.

[0093] Step 8: Convert the target optimal path into navigation control commands and output them to the ship control system for path tracking and navigation.

[0094] In this embodiment, the multimodal sensing data specifically includes radar data, optical images, infrared images, hydrological information, sonar data, and voice command information. By extracting features from different modal sensing data and uniformly encoding them, and then splicing and fusing them to generate a semantic tensor.

[0095] In this invention, the construction of the semantic tensor is achieved based on the fusion processing of multimodal perception data. Specifically, feature extraction operations are first performed on data sources from different modalities: for radar data, a target echo feature and reflection intensity feature are extracted using a target detection method based on time-frequency analysis; for optical and infrared images, spatial texture features and thermal imaging region features are extracted using a convolutional neural network; for hydrological information, including water level, water flow velocity, and flow direction, time-series features are extracted using statistical modeling; for sonar data, deep structural features are extracted using spatiotemporal filtering; and for voice command information, semantic segmentation and a BERT model are used to extract command semantic labels. After preprocessing, the feature vectors of each modality are uniformly mapped to a shared embedding space and temporally synchronized through feature alignment. Finally, a semantic tensor is constructed by splicing and fusion to provide input support for cognitive graph modeling and dynamic path planning.

[0096] In this embodiment, step two specifically includes:

[0097] Based on the semantic tensor, multimodal fusion features at different time steps are extracted as candidate state units;

[0098] The candidate state units are clustered according to the cosine similarity between the multimodal fusion features. Candidate state units with a cosine similarity greater than a set similarity threshold are merged to generate semantic state nodes. The multimodal fusion features of the merged candidate state units are averaged and weighted and concatenated to obtain the aggregated features of the semantic state nodes.

[0099] Based on the temporal sequence relationship in the semantic tensor, causal directed edges are established between adjacent semantic state nodes to form state transition channels between semantic state nodes;

[0100] For each causal directed edge, the edge weight value is calculated as follows: the cumulative number of times the semantic state node connected by the causal directed edge moves from the previous semantic state node to the next semantic state node in the historical samples is counted and multiplied by a preset first weighting factor to obtain a first calculated value; the aggregate feature Euclidean distance between the semantic state nodes connected by the causal directed edge is calculated and multiplied by a preset second weighting factor to obtain a second calculated value; the first calculated value is subtracted from the second calculated value, and the result is subjected to non-negativity constraint processing. If it is less than zero, it is set to zero, which is used as the edge weight value of the causal directed edge;

[0101] An initial cognitive graph is generated based on the semantic state nodes, causal directed edges, and edge weight values.

[0102] In this embodiment, step three specifically includes:

[0103] The aggregated features of each semantic state node in the initial cognitive graph are used as input, and dynamic modeling is performed through an improved liquid neural network. The improved liquid neural network includes an input encoding module, a liquid state evolution module, a symbolic logic embedding module, and a partial derivative output module.

[0104] The input encoding module is used to normalize and encode the aggregated features of semantic state nodes to generate a time step input vector sequence.

[0105] The liquid state evolution module is used to construct a state update structure with temporal memory capability. At each time step, the current input vector is weighted and combined with the state response vector generated at the previous time step, and mapped through the hyperbolic tangent function to obtain the state response vector at the current time step, which is used to represent the dynamic behavior characteristics of the semantic state node in the continuous time evolution process.

[0106] The symbolic logic embedding module is used to embed symbolic logic constraints during the state update process. It constructs a propositional symbolic logic expression based on the reachability rules between semantic state nodes, and converts the symbolic logic expression into a Boolean logic vector with the same dimension as the state response vector, where a value of 1 indicates that the logical constraint is satisfied, and a value of 0 indicates that the logical constraint is violated.

[0107] Perform a Hadamard product operation on the state response vector at each time step and the corresponding Boolean logic vector, and assign zero to the position dimension of the state response vector that violates the logical constraints, so as to realize real-time logical control of the state evolution path.

[0108] The partial derivative output module is used to generate navigation control bias parameters corresponding to each semantic state node, specifically including:

[0109] The state response vector after embedding symbolic logic constraints is dimensionally aligned with the feature vector of the preset navigation target state, and the numerical difference in each feature dimension is calculated to obtain the navigation deviation vector.

[0110] The navigation deviation vector is sequentially mapped to the rate of change of each feature dimension. Using numerical differentiation, the influence rate of the change of each feature dimension on the target state difference is calculated to obtain the navigation control deviation vector of the semantic state node.

[0111] This invention sets The state response vector of a semantic state node after embedding symbolic logic constraints. The feature vector of the preset navigation target state, This is the navigation deviation vector, which represents the difference across each feature dimension. Indicates the first The impact rate of input changes on navigation deviation across each feature dimension Indicates the total number of feature dimensions. Let represent the final calculated navigation control deflection, then:

[0112] ;

[0113] ;

[0114] The above calculations are performed using numerical differentiation, which represents the change in navigation deviation caused by a small perturbation in each input feature dimension, and is used to characterize the influence of semantic state nodes in the navigation control process.

[0115] In this embodiment, step four specifically refers to:

[0116] The first tension calculation value is obtained by multiplying the edge weight value of the causal directed edge with the magnitude of the navigation control bias vector of the front-end semantic state node.

[0117] The second tension calculation value is obtained by multiplying the edge weight value of the causal directed edge with the magnitude of the navigation control bias vector of the backend semantic state node.

[0118] The first tension calculation value and the second tension calculation value are weighted and summed to obtain the tension value of the causal directed edge, which is used to characterize the state transition strength and control sensitivity of the causal directed edge in the self-evolutionary cognitive graph.

[0119] The tension value is compared with a preset tension threshold to identify causal directed edges where the tension value is greater than the preset tension threshold.

[0120] For each causal directed edge whose tension value is greater than a preset tension threshold, perform a structural evolution operation, including:

[0121] Starting from the current semantic state node, search backwards in chronological order for a set of candidate semantic state nodes within a preset number of hops.

[0122] In the set of candidate semantic state nodes, select candidate semantic state nodes that have historical transfer records with the current semantic state node and whose aggregate feature Euclidean distance is less than a preset distance threshold.

[0123] Establish a new causal directed edge between the current semantic state node and the selected target candidate semantic state node, and calculate the edge weight value based on the historical transition count and the Euclidean distance of the aggregation feature.

[0124] Delete causal directed edges whose original tension values ​​are greater than the preset tension threshold, and retain the newly created edge structure;

[0125] Based on the updated semantic state nodes, causal directed edges, and edge weights, a self-evolving cognitive graph is constructed.

[0126] In this embodiment, step five specifically includes:

[0127] In the self-evolutionary cognitive graph, the current semantic state node is set as the starting point and the target semantic state node is set as the ending point. Path search is performed using a graph traversal method with a limited maximum number of hops.

[0128] Starting from the current semantic state node, traverse forward along the causal directed edges in sequence, jumping to a semantic state node that has a causal edge connection with the current semantic state node and has a state transition record in the historical samples, until the target semantic state node is reached or the preset maximum number of jumps is reached.

[0129] The semantic state node sequences generated during the traversal process are filtered, and ordered semantic state node sequences that meet the requirements of continuous connection, no node repetition, and historical transition records at each step are selected as candidate paths, and a candidate path set is constructed.

[0130] In this embodiment, step six specifically includes:

[0131] Extract the out-degree value of each semantic state node in the candidate path. The out-degree value is the number of subsequent semantic state nodes connected to the semantic state node in the self-evolutionary cognitive graph.

[0132] Extract the conflict index for each semantic state node. The conflict index is the total number of conflict events that have occurred during the historical navigation process for that semantic state node. The conflict events include:

[0133] An event occurs where the ship reverts to its path after entering the semantic state node;

[0134] The event in which the ship changes its navigation instructions at the semantic state node;

[0135] The event that triggers navigation interruption by the ship at the semantic state node;

[0136] The frequency of human intervention for each semantic state node in historical navigation tasks is statistically analyzed to characterize its navigation stability and control reliability.

[0137] The entanglement contribution value of the semantic state node is obtained by multiplying the out-degree value, conflict index, and associated historical human intervention frequency of the semantic state node by the corresponding preset weighting factor.

[0138] The entanglement contribution values ​​of all semantic state nodes in the candidate path are accumulated to obtain the graph entanglement index of the candidate path, which is used to reflect the overall structural complexity, conflict risk and navigation intervention cost of the path.

[0139] In this embodiment, step seven specifically includes:

[0140] Extract the graph entanglement index of candidate paths as a path complexity factor;

[0141] Extract the conflict index of all semantic state nodes in the candidate path, and select the maximum value as the maximum node conflict index, which serves as the path conflict risk factor.

[0142] Extract the tension values ​​of all causal directed edges in the candidate path, calculate the average tension value, and use it as the path tension load element;

[0143] The magnitude of the navigation control bias vectors of all semantic state nodes in the candidate path is statistically analyzed, and the average magnitude is calculated as the overall control sensitivity element of the path.

[0144] The path complexity elements, path conflict risk elements, path tension load elements, and path overall control sensitivity elements are arranged in sequence and combined to form a path decision vector.

[0145] Perform a path backpropagation and reduction operation on the candidate path set, the path backpropagation and reduction operation including:

[0146] Starting from the target semantic state node, trace back along the candidate path to the current semantic state node and record the corresponding path decision vector;

[0147] The multiple candidate paths obtained through reverse tracing are compared layer by layer, and candidate paths whose graph entanglement index in the path decision vector exceeds the preset entanglement threshold are eliminated.

[0148] Among the remaining candidate paths, candidate paths whose maximum node conflict index in the path decision vector exceeds a preset conflict threshold are further eliminated;

[0149] If multiple candidate paths still exist, select the candidate path with the smallest average tension value; if multiple candidate paths with the smallest average tension value exist, further select the candidate path with the largest average navigation control bias magnitude length as the target optimal path.

[0150] In the implementation of this invention, a path back-tracking elimination strategy is adopted, which involves tracing backward from the target semantic state node to the current semantic state node, to select the optimal target path from the candidate path set. Compared with the traditional forward path scoring method, this strategy has stronger target orientation and structural controllability in path selection, and can more effectively meet the dual requirements of target accuracy and path stability in actual inland waterway vessel navigation tasks.

[0151] Specifically, in the cognitive graph structure, the causal directed edges between semantic state nodes have dynamic evolution capabilities, and the path structure may be adaptively updated due to high tension relationships. If a forward traversal method is used for path scoring, all possible paths need to be scored and sorted one by one starting from the current semantic state node. This will significantly increase the consumption of computational resources when the path structure is frequently updated or the number of paths is large. Furthermore, it is easy to mistakenly select an overall unreasonable path due to the high local scores of the path's initial segment.

[0152] In contrast, the path backtracking elimination strategy takes the target semantic state node as the starting point for the reverse path, backtracks candidate paths layer by layer, and performs a phased elimination operation based on key indicators in the path decision vector (including graph entanglement index, maximum node conflict index, average tension value, and average navigation control bias vector value) at each layer. This can achieve step-by-step path compression and target alignment, thereby improving the accuracy and convergence speed of path selection.

[0153] In addition, the reverse elimination method can prioritize meeting the structural and semantic constraints of the navigation endpoint during the path selection process, avoiding the situation where a path with a high overall path score but an unstable target state or unmet constraints is incorrectly selected. This is more in line with the path selection principle of prioritizing the target and global optimization in actual navigation scenarios.

[0154] In this embodiment, step eight specifically includes:

[0155] Arrange the navigation control bias variables corresponding to each semantic state node in the target optimal path in chronological order to form a path control sequence;

[0156] The ship's turning angle is determined based on the directional change of each navigation control deflection in the path control sequence, and the ship's speed is determined based on the magnitude of the navigation control deflection.

[0157] The steering angle and the speed of travel are encoded into navigation control commands and output sequentially to the ship control system to complete path tracking navigation;

[0158] During the generation of navigation control commands, in order to achieve the mapping from semantic state sequences to specific control actions, motion parameters are transformed based on the navigation control bias variables of each semantic state node in the optimal target path. Specifically, the navigation control bias variables contain both direction information and modulus information, corresponding to the ship's steering control and propulsion speed adjustment, respectively.

[0159] First, the direction of the navigation control yaw vector is used to determine the ship's turning angle. Using the current ship's course as a reference coordinate system, the spatial direction of the current navigation control yaw vector is extracted. By calculating the angle between the navigation control yaw vector and the ship's current direction of motion, the angle deviation required for turning is obtained, which is used to adjust the ship's rudder angle and achieve course correction.

[0160] Secondly, the magnitude of the navigation control bias vector is used to determine the ship's speed. The magnitude of the bias vector represents the control sensitivity of the semantic state node in the navigation strategy; a larger value indicates a more significant impact of the current position on the navigation target. This invention normalizes the magnitude of the navigation control bias vector and weights it with the set maximum speed to generate the speed control value for the current step. In this way, the speed can be increased in high-control-sensitive sections, while the speed can be reduced or stable operation maintained in areas with weaker control, thereby achieving dynamic speed control of the intelligent navigation path.

[0161] refer to Figure 2 An intelligent detection and navigation system for inland waterway vessels includes the following modules:

[0162] The perception data acquisition module is used to collect multimodal perception data during the navigation of inland waterway vessels and construct semantic tensors;

[0163] The cognitive graph construction module is used to construct an initial cognitive graph based on the semantic tensor, and generate semantic state nodes, causal directed edges and edge weight values;

[0164] The dynamic modeling module is used to model semantic state nodes by improving the liquid neural network, embed symbolic logic constraints in the state transition, and generate navigation control bias variables.

[0165] The graph evolution module is used to calculate the tension value of causal directed edges and perform structural evolution operations on causal directed edges whose tension values ​​are greater than a preset tension threshold to generate a self-evolving cognitive graph.

[0166] The path search module is used to extract multiple candidate paths from the current semantic state node to the target semantic state node in the self-evolving cognitive graph and construct a candidate path set.

[0167] The entanglement calculation module is used to calculate the graph entanglement index of each candidate path based on the out-degree value, conflict index, and historical human intervention frequency of the semantic state node.

[0168] The path selection module is used to construct a path decision vector based on the candidate path's graph entanglement index, maximum node conflict index, average tension value, and average magnitude of the navigation control bias vector, and to perform a path back-pushing and reduction operation to select the target optimal path.

[0169] The control command output module is used to convert the target optimal path into navigation control commands and output them to the ship control system for path tracking and navigation.

[0170] To verify the feasibility of this invention in practice, it was applied to a typical inland waterway characterized by narrow waterways, dense vessel traffic, complex intersecting waterways, and interference from dock entry / exit flows and bridge areas. Traditional navigation systems often encounter problems such as path reversal, frequent adjustments to navigation commands, and control response delays in these waterways, heavily relying on crew experience for manual intervention, which presents certain safety hazards and efficiency bottlenecks.

[0171] In this invention, a medium-sized cargo ship equipped with the aforementioned intelligent inland waterway vessel detection and navigation system was selected as the test platform. Utilizing the system's sensing data acquisition module, multimodal sensing information from radar, optical images, infrared images, hydrological sensors, sonar data, and voice commands was collected in real time and uniformly encoded to generate a semantic tensor. The system then constructs an initial cognitive graph, forming semantic state nodes through clustering and similarity calculations, and extracts causal directed edges and edge weights based on historical state sequences to form an evolvable state transition structure.

[0172] During actual navigation, the system utilizes an improved liquid neural network to dynamically model each semantic state node and embeds symbolic logic constraints to generate navigation control bias vectors. In the path search phase, the system automatically identifies multiple candidate paths from the current position to the target position, constructs a path decision vector, and selects the optimal target path for navigation command output through a path back-calculation and elimination mechanism.

[0173] The system demonstrated high stability and intelligence during operation. Statistical data showed that the system completed navigation without human intervention in over 90% of the test route. In complex intersection areas and dock sections, path deviation was consistently kept within 1.5 meters, and navigation command responses were smooth, avoiding the frequent corrections and command jumps common in traditional navigation systems.

[0174] In the path control performance evaluation, the system evaluated seven candidate paths, and the results are shown in Table 1 below.

[0175] Table 1 Comparison of Path Selection and Control Indicators

[0176]

[0177] Analyzing the data in Table 1 above, the entanglement degree of path 3 is only 13.8, which is the lowest among all paths. This means that the path has the simplest entanglement structure and the clearest transfer relationship, which is conducive to the stable navigation of ships.

[0178] Regarding conflict risk, Path 3 has a maximum conflict index of 2, which is the lowest among all paths. This indicates that in historical navigation missions, the critical state nodes on this path have triggered fewer rollbacks, interruptions, or command fluctuations, demonstrating higher stability and security.

[0179] Regarding path tension load, the average tension value of path 3 is 0.68, which is the minimum among all paths. This indicates that the control tension between state transitions in the path is the smallest, and the structure tends to be more inclined towards low interference and high reliability, which helps the navigation system to perform path tracking more smoothly.

[0180] According to the rule of prioritizing the path with the smallest average tension value among the remaining paths, path 3 already has a significant advantage in this round of screening. Furthermore, according to the supplementary rule that if there are multiple candidate paths with the smallest average tension value, the path with the largest average control bias magnitude length is selected, the average navigation control bias magnitude length of path 3 is 1.60, which is greater than the average navigation control bias magnitude length of path 5 (1.55), and is also the highest among all paths. This indicates that it has the strongest ability to regulate navigation deviation and possesses the strongest response sensitivity and precise control capability.

[0181] Ultimately, in the column for whether it is the optimal path to the target, path 3 was marked as yes, and in actual verification, the deviation of this path from the actual ship track was 0.42 meters, demonstrating excellent control accuracy.

[0182] This embodiment introduces semantic tensors, initial cognitive graphs, self-evolving cognitive graphs, and improved liquid neural networks into the navigation process of inland waterways. This enables deep fusion and dynamic modeling of multimodal perception data. Symbolic logic constraints are embedded in the state transition. Combined with tension calculation and structural evolution mechanisms, it effectively solves the problems of large interference and insufficient stability in path selection in complex waters by traditional methods. It can adaptively adjust the cognitive graph structure and select the optimal path based on the path back-propagation reduction strategy, and accurately convert it into control command output. This improves the intelligence level and safety reliability of ship navigation. It has the advantages of strong real-time performance, high robustness, and good scalability.

[0183] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for intelligent detection and navigation of inland waterway vessels, characterized in that, Includes the following steps: Step 1: Collect multimodal perception data during the navigation of inland waterway vessels and construct a semantic tensor; Step 2: Construct an initial cognitive graph based on the semantic tensor, including semantic state nodes, causal directed edges, and edge weight values; Step 3: Use an improved liquid neural network to dynamically model each semantic state node in the initial cognitive map, embed symbolic logic constraints in the state transition, and generate navigation control bias variables for the semantic state nodes; Step 4: Calculate the tension value of the causal directed edge based on the edge weight value and navigation control bias, and perform structural evolution operation on the causal directed edge whose tension value is greater than the preset tension threshold to generate a self-evolving cognitive graph. Step 5: In the self-evolutionary cognitive graph, extract multiple candidate paths from the current semantic state node to the target semantic state node, and construct a candidate path set; Step 6: Based on the out-degree value, conflict index, and historical human intervention frequency of the semantic state nodes in the candidate paths, calculate the graph entanglement index of each candidate path; Step 7: Based on the graph entanglement index, maximum node conflict index, average tension value, and average magnitude of the navigation control bias vector of all semantic state nodes for each candidate path, construct a path decision vector, and perform a path back-pushing reduction operation on the candidate path set to filter and obtain the target optimal path. Step 8: Convert the target optimal path into navigation control commands and output them to the ship control system for path tracking and navigation.

2. The intelligent detection and navigation method for inland waterway vessels according to claim 1, characterized in that, The multimodal sensing data specifically includes radar data, optical images, infrared images, hydrological information, sonar data, and voice command information. By extracting features from different modal sensing data and uniformly encoding them, and then splicing and fusing them, a semantic tensor is generated.

3. The intelligent detection and navigation method for inland waterway vessels according to claim 1, characterized in that, Step two specifically involves: Based on the semantic tensor, multimodal fusion features at different time steps are extracted as candidate state units; The candidate state units are clustered according to the cosine similarity between the multimodal fusion features. Candidate state units with a cosine similarity greater than a set similarity threshold are merged to generate semantic state nodes. The multimodal fusion features of the merged candidate state units are averaged and weighted and concatenated to obtain the aggregated features of the semantic state nodes. Based on the temporal sequence relationship in the semantic tensor, causal directed edges are established between adjacent semantic state nodes to form state transition channels between semantic state nodes; For each causal directed edge, the edge weight value is calculated as follows: the cumulative number of times the semantic state node connected by the causal directed edge moves from the previous semantic state node to the next semantic state node in the historical samples is counted and multiplied by a preset first weighting factor to obtain a first calculated value; the aggregate feature Euclidean distance between the semantic state nodes connected by the causal directed edge is calculated and multiplied by a preset second weighting factor to obtain a second calculated value; the first calculated value is subtracted from the second calculated value, and the result is subjected to non-negativity constraint processing. If it is less than zero, it is set to zero, which is used as the edge weight value of the causal directed edge; An initial cognitive graph is generated based on the semantic state nodes, causal directed edges, and edge weight values.

4. The intelligent detection and navigation method for inland waterway vessels according to claim 1, characterized in that, Step three specifically involves: The aggregated features of each semantic state node in the initial cognitive graph are used as input, and dynamic modeling is performed through an improved liquid neural network. The improved liquid neural network includes an input encoding module, a liquid state evolution module, a symbolic logic embedding module, and a partial derivative output module. The input encoding module is used to normalize and encode the aggregated features of semantic state nodes to generate a time step input vector sequence. The liquid state evolution module is used to construct a state update structure with temporal memory capability. At each time step, the current input vector is weighted and combined with the state response vector generated at the previous time step, and mapped through the hyperbolic tangent function to obtain the state response vector at the current time step. The symbolic logic embedding module is used to embed symbolic logic constraints during the state update process. It constructs a propositional symbolic logic expression based on the reachability rules between semantic state nodes, and converts the symbolic logic expression into a Boolean logic vector with the same dimension as the state response vector, where a value of 1 indicates that the logical constraint is satisfied, and a value of 0 indicates that the logical constraint is violated. Perform a Hadamard product operation on the state response vector at each time step and the corresponding Boolean logic vector, and assign zero to the position dimension of the state response vector that violates the logical constraints, so as to realize real-time logical control of the state evolution path. The partial derivative output module is used to generate navigation control bias parameters corresponding to each semantic state node, specifically including: The state response vector after embedding symbolic logic constraints is dimensionally aligned with the feature vector of the preset navigation target state, and the numerical difference in each feature dimension is calculated to obtain the navigation deviation vector. The navigation deviation vector is sequentially mapped to the rate of change of each feature dimension. Using numerical differentiation, the influence rate of the change of each feature dimension on the target state difference is calculated to obtain the navigation control deviation vector of the semantic state node.

5. The intelligent detection and navigation method for inland waterway vessels according to claim 1, characterized in that, Step four specifically involves: The first tension calculation value is obtained by multiplying the edge weight value of the causal directed edge with the magnitude of the navigation control bias vector of the front-end semantic state node. The second tension calculation value is obtained by multiplying the edge weight value of the causal directed edge with the magnitude of the navigation control bias vector of the backend semantic state node. The first tension calculation value and the second tension calculation value are weighted and summed to obtain the tension value of the causal directed edge; The tension value is compared with a preset tension threshold to identify causal directed edges where the tension value is greater than the preset tension threshold. For each causal directed edge whose tension value is greater than a preset tension threshold, perform a structural evolution operation, including: Starting from the current semantic state node, search backwards in chronological order for a set of candidate semantic state nodes within a preset number of hops. In the set of candidate semantic state nodes, select candidate semantic state nodes that have historical transfer records with the current semantic state node and whose aggregate feature Euclidean distance is less than a preset distance threshold. Establish a new causal directed edge between the current semantic state node and the selected target candidate semantic state node, and calculate the edge weight value based on the historical transition count and the Euclidean distance of the aggregation feature. Delete causal directed edges whose original tension values ​​are greater than the preset tension threshold, and retain the newly created edge structure; Based on the updated semantic state nodes, causal directed edges, and edge weights, a self-evolving cognitive graph is constructed.

6. The intelligent detection and navigation method for inland waterway vessels according to claim 1, characterized in that, Step five specifically involves: In the self-evolutionary cognitive graph, the current semantic state node is set as the starting point and the target semantic state node is set as the ending point. Path search is performed using a graph traversal method with a limited maximum number of hops. Starting from the current semantic state node, traverse forward along the causal directed edges in sequence, jumping to a semantic state node that has a causal edge connection with the current semantic state node and has a state transition record in the historical samples, until the target semantic state node is reached or the preset maximum number of jumps is reached. The semantic state node sequences generated during the traversal process are filtered, and ordered semantic state node sequences that satisfy continuous connection, no node repetition, and historical transition records at each step are selected as candidate paths, and a candidate path set is constructed.

7. The intelligent detection and navigation method for inland waterway vessels according to claim 1, characterized in that, Step six specifically involves: Extract the out-degree value of each semantic state node in the candidate path. The out-degree value is the number of subsequent semantic state nodes connected to the semantic state node in the self-evolutionary cognitive graph. Extract the conflict index for each semantic state node. The conflict index is the total number of conflict events that have occurred during the historical navigation process for that semantic state node. The conflict events include: An event occurs where the ship reverts to its path after entering the semantic state node; The event in which the ship changes its navigation instructions at the semantic state node; The event that triggers navigation interruption by the ship at the semantic state node; Statistical analysis of the frequency of human intervention for each semantic state node in historical navigation tasks; The entanglement contribution value of the semantic state node is obtained by multiplying the out-degree value, conflict index, and associated historical human intervention frequency of the semantic state node by the corresponding preset weighting factor. The entanglement contribution values ​​of all semantic state nodes in the candidate path are accumulated to obtain the graph entanglement index of the candidate path.

8. The intelligent detection and navigation method for inland waterway vessels according to claim 1, characterized in that, Step seven specifically involves: Extract the graph entanglement index of candidate paths as a path complexity factor; Extract the conflict index of all semantic state nodes in the candidate path, and select the maximum value as the maximum node conflict index, which serves as the path conflict risk factor. Extract the tension values ​​of all causal directed edges in the candidate path, calculate the average tension value, and use it as the path tension load element; The magnitude of the navigation control bias vectors of all semantic state nodes in the candidate path is statistically analyzed, and the average magnitude is calculated as the overall control sensitivity element of the path. The path complexity elements, path conflict risk elements, path tension load elements, and path overall control sensitivity elements are arranged in sequence and combined to form a path decision vector. Perform a path backpropagation and reduction operation on the candidate path set, the path backpropagation and reduction operation including: Starting from the target semantic state node, trace back along the candidate path to the current semantic state node and record the corresponding path decision vector; The multiple candidate paths obtained through reverse tracing are compared layer by layer, and candidate paths whose graph entanglement index in the path decision vector exceeds the preset entanglement threshold are eliminated. Among the remaining candidate paths, candidate paths whose maximum node conflict index in the path decision vector exceeds a preset conflict threshold are further eliminated; If multiple candidate paths still exist, select the candidate path with the smallest average tension value; if multiple candidate paths with the smallest average tension value exist, further select the candidate path with the largest average navigation control bias magnitude length as the target optimal path.

9. The intelligent detection and navigation method for inland waterway vessels according to claim 1, characterized in that, Step eight specifically involves: Arrange the navigation control bias variables corresponding to each semantic state node in the target optimal path in chronological order to form a path control sequence; The ship's turning angle is determined based on the directional change of each navigation control deflection in the path control sequence, and the ship's speed is determined based on the magnitude of the navigation control deflection. The steering angle and the travel speed are encoded into navigation control commands and output sequentially to the ship control system to complete path tracking navigation.

10. An intelligent detection and navigation system for inland waterway vessels, comprising executing the intelligent detection and navigation method for inland waterway vessels as described in any one of claims 1 to 9, characterized in that, Includes the following modules: The perception data acquisition module is used to collect multimodal perception data during the navigation of inland waterway vessels and construct semantic tensors; The cognitive graph construction module is used to construct an initial cognitive graph based on the semantic tensor, and generate semantic state nodes, causal directed edges and edge weight values; The dynamic modeling module is used to model semantic state nodes by improving the liquid neural network, embed symbolic logic constraints in the state transition, and generate navigation control bias variables. The graph evolution module is used to calculate the tension value of causal directed edges and perform structural evolution operations on causal directed edges whose tension values ​​are greater than a preset tension threshold to generate a self-evolving cognitive graph. The path search module is used to extract multiple candidate paths from the current semantic state node to the target semantic state node in the self-evolving cognitive graph and construct a candidate path set. The entanglement calculation module is used to calculate the graph entanglement index of each candidate path based on the out-degree value, conflict index, and historical human intervention frequency of the semantic state node. The path selection module is used to construct a path decision vector based on the candidate path's graph entanglement index, maximum node conflict index, average tension value, and average magnitude of the navigation control bias vector, and to perform a path back-pushing and reduction operation to select the target optimal path. The control command output module is used to convert the target optimal path into navigation control commands and output them to the ship control system for path tracking and navigation.

Images