An AUV-oriented multi-element information collaborative decision-making method based on intention recognition

Abstract

The application provides a multi-element information collaborative decision method based on intention recognition for an autonomous underwater robot in a weak communication and partial observation environment, so as to improve the adaptability and intelligent decision-making ability of an AUV group in a complex dynamic environment. Specifically, the application provides an intention recognition method based on decision consensus. Firstly, a decision consensus among AUVs is constructed through fuzzy reasoning, and key features such as a packet loss rate, a distance, and a number of partners are mapped to a behavior space, so that a preliminary intention inference is efficiently realized with a minimum amount of data exchange. Subsequently, a residual convolutional neural network is used to further strengthen feature extraction, and precise intention prediction is realized. Finally, by completing global state information and optimizing decision input, a dynamic perception-intention reasoning-collaborative decision closed loop is constructed, and more efficient task planning and cooperation efficiency of the AUV group are realized.

Description

Technical Field

[0001] This invention belongs to the field of collaborative decision-making for multiple autonomous underwater vehicles (AUVs), specifically a multi-source collaborative decision-making method based on intent recognition for AUVs. Background Technology

[0002] AUV collaborative operations face complex underwater communication environments characterized by low bandwidth, high latency, and high packet loss rates. Traditional trajectory prediction methods rely on extrapolating historical motion data, which struggles to handle non-inertial motion and sudden behavioral changes, and only provide location information, failing to support high-level decision-making needs. Intent prediction, by fusing target behavior patterns, environmental constraints, and task information, can predict the target's high-level action intentions in advance, improving adaptability to dynamic environments. However, current research on intent recognition theory in the field of underwater multi-agent collaboration is still in its early stages, suffering from problems such as a lack of clear intent mapping relationships, poor data quality, and severe accumulation of navigation errors, resulting in insufficient intent prediction accuracy and limiting practical applications. Therefore, there is an urgent need to research efficient intent recognition methods adapted to weak communication and partial observation conditions to enhance the autonomy and intelligent decision-making capabilities of multi-AUV collaborative tasks. Summary of the Invention

[0003] This invention proposes a multi-dimensional collaborative decision-making method based on intent recognition for Autonomous Underwater Vehicles (AUVs) in environments with weak communication and partial observation, aiming to improve the adaptability and intelligent decision-making capabilities of AUV swarms in complex and dynamic environments. Addressing the limitations of traditional trajectory prediction methods in scenarios involving non-inertial motion and sudden behavioral changes, this invention utilizes high-level intent prediction that combines AUV behavior patterns within the swarm, environmental constraints, and task information to identify potential actions by AUVs in advance, thereby enhancing the system's ability to cope with uncertainty.

[0004] The technical solution adopted by the present invention to achieve the above objectives is as follows:

[0005] A collaborative decision-making method based on intent recognition using multi-source information for AUVs includes the following steps:

[0006] 1) Determine the behavioral intentions and characteristics affecting changes in intentions during AUV collaborative tasks;

[0007] 2) Construct an intent reasoning model based on residual convolutional neural networks, and use this model to predict the behavioral intent of other AUVs besides the local AUV;

[0008] 3) Calculate the future state of the AUV population based on the prediction results, and combine it with the state of the local AUVs to construct the input feature set of the behavior model;

[0009] 4) Construct a behavior model based on fuzzy reasoning, and use the model to perform fuzzy reasoning to obtain the behavior intent of the local AUV.

[0010] The behavioral intent includes: communication intent Intention to reserve space and the intent of detection These represent optimized communication, rapid approach to the target, and optimized detection, respectively.

[0011] The features affecting the change of intent include: relative positioning point direction RDP, relative PNCT direction RDPN, average detection error ADE, number of partner AUVs NoC, data reception success rate PRR, distance to positioning point DPP, distance to parent topology node DPN, detection error DE, and variance of detection error VDE.

[0012] Step 2) includes the following steps:

[0013] 2.1) Obtain AUV's DPP, DPN, RDP, RDPN, NoC, and ADE data;

[0014] 2.2) Data features are extracted through an initial convolutional layer with 16 filters, and then processed sequentially by batch normalization, ReLU activation function, and max pooling with a stride of 2.

[0015] 2.3) The processed features are passed through four residual blocks in sequence, with the number of filters in the residual blocks gradually increasing from 16 to 256, and each residual block uses skip connections;

[0016] 2.4) After processing with residual blocks, the spatial dimension of the features is normalized to a fixed size of 1 by an adaptive average pooling layer, and the normalized data is flattened.

[0017] 2.5) Pass the flattened feature data to the fully connected layer and output the category prediction results.

[0018] The future states of the AUV population include: DPP, DPN, and NoC, and the states of the local AUVs include: PRR, DE, and VDE.

[0019] Step 4) includes the following steps:

[0020] 4.1) Use Gaussian membership function and trigonometric membership function to fuzzify the input and output respectively;

[0021] 4.2) Construct a fuzzy rule table;

[0022] 4.3) Use the Mamdani type fuzzy implication operator and combine it with the maximum-minimum fuzzy synthesis operation for inference calculation, and obtain the output result through the weighted average defuzzification method.

[0023] Step 4.1) includes the following steps:

[0024] 4.1.1) Define fuzzy domains for each input feature and set thresholds within the domains. Represent the fuzzy values ​​of different ranges within the fuzzy domains as large L, medium M, and small S.

[0025] 4.2.2) The input features are calculated using a Gaussian membership function, and the fuzzy value corresponding to the feature is determined based on the calculation result. The Gaussian membership function is:

[0026] ;

[0027] in, Represents the input value. Represents the center of the Gaussian function. It is the standard deviation;

[0028] 4.2.3) Setting the output The output is selected using a triangular membership function, which is:

[0029] ;

[0030] in, Represents the output value. It is the left endpoint, i.e., the minimum value, corresponding to a membership degree of 0; It is the center point, i.e., the maximum value, with a membership degree of 1; It is the right endpoint, i.e., the maximum value, and the corresponding membership degree is 0.

[0031] The fuzzy rule table is specifically as follows:

[0032] i) If NoC is L, then output ;

[0033] ii) If NoC is M, PRR is M, DPP is M, and DPN is M, then the output is... ;

[0034] iii) If NoC is S, PRR is S, and DPN is M, then the output is... ;

[0035] iv) If NoC is S, PRR is L, and DPN is S, then output ;

[0036] v) If DPP is L, then output ;

[0037] vi) If DPP is M, VDE is S, and DE is L, then output ;

[0038] vii) If DPP is M, VDE is L and DE is L, then the output is... ;

[0039] viii) If DPP is M, VDE is L, and DE is S, then the output is... ;

[0040] ix) If DPP is M, VDE is S, and DE is S, then the output is... ;

[0041] x) If DPP is S, then output .

[0042] Step 4.3) includes the following steps:

[0043] 4.3.1) Reasoning is performed on a single fuzzy rule using a Mamdani-type fuzzy implication operator, and the activation intensity of the rule antecedent is applied to the membership function of the rule consequent through minimum operation;

[0044] 4.3.2) The "maximum-minimum" fuzzy synthesis method is used to synthesize the reasoning results of multiple rules, and the membership value with the highest activation degree is selected in the same output domain to obtain the fuzzy set;

[0045] 4.3.3) The weighted average defuzzification method is used to defuzzify the fuzzy set to obtain the final AUV behavior intention reasoning result.

[0046] The present invention has the following beneficial effects and advantages:

[0047] 1. This invention proposes an intent recognition method based on decision consensus. By effectively integrating key features such as packet loss rate, distance, and number of partners through fuzzy reasoning, it can efficiently infer intent under extremely limited data exchange conditions, significantly improving the adaptability of AUV systems to weak underwater communication conditions.

[0048] 2. This invention utilizes residual convolutional neural networks to enhance feature extraction capabilities, overcoming the impact of underwater navigation errors and data gaps, achieving more accurate high-level intent prediction, and effectively improving the system's decision-making reliability in certain observation environments.

[0049] 3. This invention, through a closed-loop structure of dynamic perception-intent reasoning-collaborative decision-making, feeds back the intent prediction results to the decision input, completes and optimizes the global state information, and significantly improves the autonomy of AUV group task planning and overall collaborative efficiency. Attached Figure Description

[0050] Figure 1 This is a schematic diagram of the overall structure of the AUV intent recognition based on decision consensus proposed in this invention;

[0051] Figure 2 This describes the information flow of the present invention in a practical system.

[0052] Figure 3 A schematic diagram of the membership function for the input feature NoC and the intended output;

[0053] Figure 4 This is a schematic diagram of the neural network architecture used in this invention;

[0054] Figure 5 These are schematic diagrams of confusion matrices obtained using different methods;

[0055] Figure 6 A schematic diagram showing the changes in loss values ​​during the training process of different methods;

[0056] Figure 7 The diagram illustrates the initial distribution under two typical communication topology scenarios.

[0057] Where (a) represents scenario A: good communication topology conditions, and (b) represents scenario B: extreme communication topology conditions;

[0058] Figure 8 This is a schematic diagram illustrating the trend of DAP_total over time in scenario A.

[0059] Figure 9 This is a schematic diagram showing the trajectories of the two methods in completing the collaborative search task in scenario A;

[0060] Wherein (a) is a collaborative decision-making method without intent recognition, and (b) is a collaborative decision-making method proposed in this invention (combined with intent recognition IR);

[0061] Figure 10 This is a schematic diagram illustrating the trend of DAP_total over time in scenario B.

[0062] Figure 11 This is a schematic diagram showing the trajectories of the two methods in completing the collaborative search task in scenario B;

[0063] Wherein (a) is a collaborative decision-making method without intent recognition, and (b) is a collaborative decision-making method (combined with intent recognition IR) proposed in this invention. Detailed Implementation

[0064] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0065] This invention proposes an intent recognition method based on decision consensus. First, it constructs a decision consensus among AUVs through fuzzy reasoning, mapping key features such as packet loss rate, distance, and number of companions to the behavior space, achieving preliminary intent inference with minimal data exchange. Then, a residual convolutional neural network is used to further enhance feature extraction, achieving accurate intent prediction. Finally, by supplementing global state information and optimizing decision input, a dynamic perception-intent reasoning-collaborative decision-making closed loop is constructed, achieving more efficient task planning and collaboration among AUV groups.

[0066] Figure 1 This paper presents the overall structure of the multi-AUV intent recognition method proposed in this invention. The decision-making process of the AUV swarm is unified based on the BDI framework. To establish the mapping relationship between behavioral states and intents, the concept of "landmarks" is introduced to standardize the intent types throughout the swarm's decision-making process, including optimized communication, optimized detection, and rapid target approach. The method employs a behavior modeling-intent reasoning architecture. In the behavior modeling stage, fuzzy reasoning is used to build the AUV's behavior model, mapping observed state information (such as position, velocity, heading, mission status, etc.) to the action space to generate the AUV's behavioral intent. In the intent reasoning stage, ResCNN is used to process the data from the behavior modeling stage to predict the AUV's target intent. This two-layer architecture simplifies data transmission and communication costs while ensuring computational efficiency and improving the accuracy of intent reasoning.

[0067] Taking collaborative search tasks as an example, Figure 2 The information flow of this invention in a practical system is demonstrated. The flow begins with the collection of key state data from other AUVs, including parameters such as DPP, DPN, RDP, RDPN, NoC, and ADE. This data is first processed by a pre-trained ResCNN to predict the intentions of other AUVs. Subsequently, the system combines the predicted intentions with data received via underwater acoustic communication to calculate the future states of other AUVs. A fuzzy inference model uses DPP, PRR, DPN, NoC, DE, and VDE as inputs to determine the intentions of the local AUVs, ensuring consensus on intentions within the entire AUV group. Finally, the identified intentions are used in the path planning process, enabling AUVs to adaptively adjust their decisions and effectively participate in cooperative search tasks.

[0068] A collaborative decision-making method based on intent recognition using multi-source information for AUVs mainly includes the following steps:

[0069] 1) Determine the behavioral intent in AUV collaborative operation tasks;

[0070] 2) Identify the characteristics that influence changes in intent during AUV collaborative work tasks;

[0071] 3) Construction of behavioral models based on fuzzy reasoning;

[0072] 4) Intent reasoning based on residual convolutional neural networks.

[0073] Step 1) includes the following steps:

[0074] Intent refers to the specific actions taken by an AUV to complete a collaborative task, based on its understanding of global state information. AUVs need to dynamically adjust their decisions at fixed time intervals according to the current global situation and generate immediate intents. Taking a collaborative search task as an example, an AUV swarm cannot complete a target search through a single decision. Throughout the mission, not only is continuous communication necessary, but detection accuracy also needs to be constantly improved, requiring AUVs to make multiple dynamic adjustments to complete the task. Therefore, this invention categorizes intents in collaborative search tasks into three types: optimized communication, optimized detection, and rapid target approach.

[0075] Step 2) includes the following steps:

[0076] The intent of an AUV is manifested through its actions and states, and the mapping from behavior and state to intent is determined by a cooperative search decision-making process. Therefore, it is essential to first identify the factors influencing this decision-making process. This invention identifies several key factors, including distance to the occupant site, success rate of data reception, distance to the parent node in the communication topology (PNCT), number of buddies, distance to the target, current detection error, and variance of the detection error. If these factors are obtained in a timely manner, the same decision-making method can be used to predict the intent of other AUVs, facilitating adjustments to cooperative strategies. However, due to the inherent characteristics of underwater acoustic communication, such as significant delays and packet loss, this information becomes unreliable at the receiving end and fails to reflect the true state of the group. Therefore, intent recognition is crucial for predicting the intent of other AUVs.

[0077] Intent recognition requires the fusion and analysis of received information, making the selection of appropriate feature inputs for the model crucial. The factors mentioned above are key to intent recognition, but some information, such as packet loss rate and detection error, varies drastically over time and directly impacts intent. These factors can adversely affect AUV intent recognition under high latency and packet loss conditions. Therefore, the following information was ultimately selected as model inputs: distance to the occupant site, distance to the PNCT, relative orientation to the occupant site, relative orientation to the PNCT, number of partners, and average detection error. Through comprehensive analysis of these factors, the key factors influencing AUV intent changes can be more accurately captured, thereby improving the accuracy of intent recognition and collaborative efficiency.

[0078] Step 3) includes the following steps:

[0079] This invention employs fuzzy reasoning to construct an AUV behavior model, aiming to unify the decision-making methods of an AUV group. The core idea is to ensure that identical or similar information consistently leads to consistent decision results, thereby forming a decision consensus among AUVs. In specific implementation, Gaussian membership functions and triangular membership functions are used to fuzzify the input and output, respectively. Through multiple rounds of testing, a complete fuzzy rule table is constructed. During the inference calculation process, a Mamdani-type fuzzy implication operator is used, combined with max-min fuzzy synthesis operations for inference calculation. Finally, the output result is obtained through a weighted average defuzzification method.

[0080] The fuzzification process involves defining several fuzzy subsets for each input variable within the fuzzy domain and assigning membership functions to these subsets. Membership functions describe the degree of mapping for each fuzzy subset within the domain and are typically represented graphically. This invention defines the fuzzy domain of NoC as [0, n-1], where n represents the total number of AVUs. The fuzzy domain of PRR is [0, 1], DPP is [-3000, 0], DPN is [-1000, 0], the detection error is [0, 10], and the detection error variance is [0, 3]. The linguistic values ​​for all these domains are represented as "Large (L)", "Medium (M)", and "Small (S)".

[0081] Because the smooth curve characteristic of the Gaussian membership function is well-suited for handling both continuous and fuzzy data, it is chosen here. The mathematical expression for the Gaussian membership function is:

[0082] (1)

[0083] in, Represents the input value. Represents the center of the Gaussian function. It is the standard deviation (control curve width).

[0084] For the output, this invention designs three types of output behaviors. ,in Represents the intent of communication. This represents the intention to reserve a space. This represents the intent to probe. For the selected membership function in the output, which is a trigonometric membership function, the mathematical expression is:

[0085] (2)

[0086] in, It is the left endpoint (minimum value), corresponding to a membership degree of 0; It is the center point (maximum value), corresponding to a membership degree of 1; It is the right endpoint (maximum value), corresponding to a membership degree of 0.

[0087] The membership functions of the input feature NoC and the intended output are as follows: Figure 3 As shown.

[0088] Fuzzy rules are used to define the relationship between input conditions and output variables. By listing all possible preconditions and their corresponding consequences, a complete set of fuzzy rules can be constructed to guide the intelligent decision-making of AUVs under different environmental conditions, as shown below:

[0089] i) IF NoC is L, THEN

[0090] ii) IF NoC is M, PRR is M, DPP is M and DPN is M, THEN

[0091] iii) IF NoC is S, PRR is S and DPN is M, THEN

[0092] iv) IF NoC is S, PRR is L and DPN is S, THEN

[0093] v) IF DPP is L, THEN

[0094] vi) IF DPP is M, VDE is S and DE is L, THEN

[0095] vii) IF DPP is M, VDE is L and DE is L, THEN

[0096] viii) IF DPP is M, VDE is L and DE is S, THEN

[0097] ix) IF DPP is M, VDE is S and DE is S, THEN

[0098] x) IF DPP is S, THEN

[0099] This invention employs a Mamdani-type fuzzy implication operator and a "maximum-minimum" fuzzy synthesis operation to calculate the inference result of each fuzzy rule. Subsequently, the outputs of all fuzzy rules are aggregated using the maximum value method to obtain the final inference result. This method effectively handles fuzzy inference problems with multiple inputs and outputs, ensuring intelligent decision optimization for AUVs in complex environments.

[0100] This invention employs a weighted average defuzzification method to defuzzify the results of fuzzy inference. The calculation formula is as follows:

[0101] (3)

[0102] in, This represents the final output value after defuzzification; This represents the membership degree of the rule. This represents the output value corresponding to the i-th fuzzy rule; The number of fuzzy rules that meet the conditions.

[0103] Step 4) includes the following steps:

[0104] This invention classifies other AUVs' intentions using feature information. However, due to the high latency and high packet loss rate of underwater acoustic communication, some key information that fluctuates significantly over time, such as PRR, DE, and VDE, can negatively impact classification accuracy. Therefore, this invention uses substitute variables with smaller time variations but still reflecting necessary features to replace this information, thereby improving model stability. The final classification model input features include: DPP, DPN, relative direction to the positioning point (RDP), relative direction to the PNCT (RDPN), NoC, and the average detection error (ADE). The classification model output still corresponds to three intention types: optimized communication, optimized detection, and rapid approach to the target.

[0105] The network architecture used in this invention is as follows: Figure 4 As shown, the model consists of an initial convolutional layer (Conv) followed by multiple residual blocks, each containing two convolutional layers. The output of the final residual block passes through a fully connected layer and a Softmax activation function to achieve multi-class prediction. To improve the model's stability and convergence, the output of each convolutional layer undergoes batch normalization (BN), followed by a ReLU activation function, and a Dropout mechanism is introduced to prevent overfitting.

[0106] This network is optimized for one-dimensional data processing. It first extracts features through an initial convolutional layer with 16 filters, followed by batch normalization, ReLU activation, and max pooling with a stride of 2, reducing data dimensionality while extracting low-level features. Subsequently, the network introduces four residual blocks, where the number of filters gradually increases from 16 to 256 to enhance the model's ability to learn complex patterns. Each residual block employs skip connections to effectively prevent the vanishing gradient problem and improve deep feature learning capabilities.

[0107] After passing through the residual blocks, the network uses an adaptive average pooling layer to normalize the spatial dimension to a fixed size of 1, regardless of the input length. Subsequently, after a flattening operation, the data is passed to a fully connected layer (containing 256 units), which ultimately outputs the class prediction result.

[0108] Figure 5 The confusion matrices obtained using different methods are shown, reflecting the accuracy of classification, the number of misclassifications, and the overall performance of the model.

[0109] Figure 6 The changes in the loss curves during the training process of different methods are shown to analyze the training process of different methods.

[0110] To verify the performance of the proposed intention recognition method based on decision consensus in collaborative search tasks, a comparative experiment was designed to compare the method with a collaborative decision-making method that does not employ intention recognition.

[0111] Figure 7 The initial distributions under two communication topology conditions are shown.

[0112] In the experiments, all methods were conducted under the same initial conditions, and the performance of the proposed methods was comprehensively analyzed using evaluation metrics such as search time and task success rate. At each sampling time, the total distance from all AUVs to their respective occupancy points was calculated to quantify the progress of the cooperative search task. The calculation method is as follows:

[0113]

[0114] in, Represents the Euclidean norm. Represents the number of AUVs, Represents AUV i Location, Represents the corresponding AUV i The coordinates of the occupant point.

[0115] Figure 8 Scene A was shown The trend over time shows that the method proposed in this invention (combined with intent recognition IR) achieves faster speeds compared to methods that do not use IR. Descent speed. This indicates that intent recognition can improve the efficiency of collaborative search, enabling AUVs to reach the target location more quickly.

[0116] Figure 9 The simulation results for the two methods in scenario A demonstrate the trajectories achieved in the cooperative search task. The comparison results in scenario A clearly validate the advantages of the proposed CBIR method in cooperative search tasks. Figure 9 (a) It can be observed that the AUV's search path contains a significant amount of redundant movement. This is mainly because traditional methods rely on global communication and strictly adhere to communication constraints for path planning, making it difficult to optimize task execution through intelligent decision-making. Figure 9 (b) demonstrates a more efficient path distribution, where AUVs can predict and infer the intentions of other AUVs, thereby achieving more reasonable collaboration and avoiding unnecessary path overlap. This result shows that the CBIR method effectively reduces redundant paths and improves the efficiency of path planning through intention prediction and inference mechanisms.

[0117] Figure 10 Scene B was shown Despite increasingly stringent communication topology environments, the proposed method still maintains better search efficiency over time.

[0118] Under extreme communication topology conditions, the traditional method and the intent prediction-based CBIR method show significant differences in performance in simulations. Figure 11 (a) illustrates the search path of the traditional method under sparse communication conditions. Due to limited communication, AUVs cannot exchange information in real time, leading to collaboration failure. When some AUVs lose communication, they cannot continue to participate effectively in the task, ultimately causing the team to separate and the search task to fail. The 1.58km distance indicated by the arrow exceeds the communication radius (1km), exacerbating the risk of task failure.

[0119] In comparison, Figure 11 (b) illustrates the search path using the intent prediction-based CBIR method. Under the same extreme communication conditions, even with prolonged communication breakdowns between AUVs, the CBIR method can still achieve global decision-making based on local information through intent prediction and reasoning mechanisms. Each AUV can infer the intent of the entire team and dynamically adjust its own path, effectively avoiding task failures caused by communication interruptions in traditional methods. Ultimately, this method successfully completed the collaborative search task.

Claims

1. A multi-source collaborative decision-making method based on intent recognition for AUVs, characterized in that, Includes the following steps: 1) Determine the behavioral intentions and characteristics affecting changes in intentions during AUV collaborative tasks; 2) Construct an intent reasoning model based on residual convolutional neural networks, and use this model to predict the behavioral intent of other AUVs besides the local AUV; 3) Calculate the future state of the AUV population based on the prediction results, and combine it with the state of the local AUVs to construct the input feature set of the behavior model; 4) Construct a behavior model based on fuzzy reasoning, and use the model to perform fuzzy reasoning to obtain the behavior intent of the local AUV.

2. The multi-source collaborative decision-making method based on intent recognition for AUVs according to claim 1, characterized in that, The behavioral intent includes: communication intent Intention to reserve space and the intent of detection These represent optimized communication, rapid approach to the target, and optimized detection, respectively.

3. The multi-source collaborative decision-making method based on intent recognition for AUVs according to claim 1, characterized in that, The features affecting the change of intent include: relative positioning point direction RDP, relative PNCT direction RDPN, average detection error ADE, number of partner AUVs NoC, data reception success rate PRR, distance to positioning point DPP, distance to parent topology node DPN, detection error DE, and variance of detection error VDE.

4. The multi-source collaborative decision-making method based on intent recognition for AUVs according to claim 1, characterized in that, Step 2) includes the following steps: 2.1) Obtain AUV's DPP, DPN, RDP, RDPN, NoC, and ADE data; 2.2) Data features are extracted through an initial convolutional layer with 16 filters, and then processed sequentially by batch normalization, ReLU activation function, and max pooling with a stride of 2. 2.3) The processed features are passed through four residual blocks in sequence, with the number of filters in the residual blocks gradually increasing from 16 to 256, and each residual block uses skip connections; 2.4) After processing with residual blocks, the spatial dimension of the features is normalized to a fixed size of 1 by an adaptive average pooling layer, and the normalized data is flattened. 2.5) Pass the flattened feature data to the fully connected layer and output the category prediction results.

5. A multi-source collaborative decision-making method based on intent recognition for AUVs according to claim 1, characterized in that, The future states of the AUV population include: DPP, DPN, and NoC, and the states of the local AUVs include: PRR, DE, and VDE.

6. A multi-source collaborative decision-making method based on intent recognition for AUVs according to claim 1, characterized in that, Step 4) includes the following steps: 4.1) Use Gaussian membership function and trigonometric membership function to fuzzify the input and output respectively; 4.2) Construct a fuzzy rule table; 4.3) Use the Mamdani type fuzzy implication operator and combine it with the maximum-minimum fuzzy synthesis operation for inference calculation, and obtain the output result through the weighted average defuzzification method.

7. A multi-source collaborative decision-making method based on intent recognition for AUVs according to claim 6, characterized in that, Step 4.1) Includes the following steps: 4.1.1) Define fuzzy domains for each input feature and set thresholds within the domains. Represent the fuzzy values ​​of different ranges within the fuzzy domains as large L, medium M, and small S. 4.2.2) The input features are calculated using a Gaussian membership function, and the fuzzy value corresponding to the feature is determined based on the calculation result. The Gaussian membership function is: ； in, Represents the input value. Represents the center of the Gaussian function. It is the standard deviation; 4.2.3) Setting the output The output is selected using a triangular membership function, which is: ； in, Represents the output value. It is the left endpoint, i.e., the minimum value, corresponding to a membership degree of 0; It is the center point, i.e., the maximum value, with a membership degree of 1; It is the right endpoint, i.e., the maximum value, and the corresponding membership degree is 0.

8. A multi-source collaborative decision-making method based on intent recognition for AUVs according to claim 6, characterized in that, The fuzzy rule table is specifically as follows: i) If NoC is L, then output ; ii) If NoC is M, PRR is M, DPP is M, and DPN is M, then the output is... ; iii) If NoC is S, PRR is S, and DPN is M, then the output is... ; iv) If NoC is S, PRR is L, and DPN is S, then output ; v) If DPP is L, then output ; vi) If DPP is M, VDE is S, and DE is L, then output ; vii) If DPP is M, VDE is L and DE is L, then the output is... ; viii) If DPP is M, VDE is L, and DE is S, then the output is... ; ix) If DPP is M, VDE is S, and DE is S, then the output is... ; x) If DPP is S, then output .

9. A multi-source collaborative decision-making method based on intent recognition for AUVs according to claim 6, characterized in that, Step 4.3) includes the following steps: 4.3.1) Reasoning is performed on a single fuzzy rule using a Mamdani-type fuzzy implication operator, and the activation intensity of the rule antecedent is applied to the membership function of the rule consequent through minimum operation; 4.3.2) The "maximum-minimum" fuzzy synthesis method is used to synthesize the reasoning results of multiple rules, and the membership value with the highest activation degree is selected in the same output domain to obtain the fuzzy set; 4.3.3) The weighted average defuzzification method is used to defuzzify the fuzzy set to obtain the final AUV behavior intention reasoning result.

Images