Method for dynamically establishing knowledge graph based on multi-source heterogeneous data fusion

Abstract

The invention discloses a method for dynamically establishing a knowledge graph based on multi-source heterogeneous data fusion, and belongs to the field of intelligent robot systems. Firstly, multi-source heterogeneous data fusion is carried out, a distributed Kalman filter network is constructed, filters are designed for different sensors, and data space-time alignment is realized; carrying out conflict detection and resolution by adopting an improved D-S evidence theory, and dynamically adjusting a fusion weight through a confidence evaluation model; secondly, constructing a dynamic knowledge graph, and establishing a three-layer graph structure comprising an environment entity, a space-time relationship and probabilistic reasoning; designing an event-triggered incremental updating algorithm, starting local reconstruction when the environmental change degree exceeds a threshold value, and introducing an LSTM memory network to enhance the time sequence reasoning capability; and finally, constructing a reinforcement learning model based on the dynamic knowledge graph, designing a multi-target reward function, and finally training until the model is converged. According to the method, the positioning error is reduced through EKF / UKF combined filtering, the multi-sensor data space alignment error is reduced through a conflict resolution mechanism, and the execution efficiency and the coping capacity of the omni-directional mobile robot in a complex dynamic environment can be improved.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent robot systems, and relates to a method for fusing multi-source heterogeneous data and dynamically building a knowledge graph, specifically to a robot control system that fuses multi-modal sensor data, dynamically constructs a knowledge graph, and is trained based on reinforcement learning. Background Art

[0002] With the rapid development of artificial intelligence technology and robot technology, intelligent robot systems are increasingly widely used in various complex dynamic environments, especially in scenarios of multi-robot dynamic operations, such as warehousing logistics, intelligent distribution, and inspection. However, existing multi-robot systems still face many technical challenges in dynamic and complex environments.

[0003] Current robot systems usually rely on traditional sensor technologies and static control algorithms, resulting in insufficient adaptability in changing environments. Although multi-modal sensor data (such as lidar, cameras, temperature and humidity sensors, etc.) can provide rich environmental information, existing systems often cannot effectively fuse this data from different sources, resulting in the information island effect, which affects the system's accurate perception and understanding of the environment. This problem is particularly prominent during task execution.

[0004] Another major problem is that existing robot control technologies usually rely on fixed scheduling algorithms and preset rules, and cannot flexibly respond in dynamic environments. For example, when a robot swarm faces obstacles, formation adjustment, or task changes, the decision-making systems of existing technologies lack sufficient intelligence and adaptability to adjust the behavior of robots in real time. This leads to low task execution efficiency, and in some cases, conflicts or resource waste may occur between robots.

[0005] In addition, the existing swarm intelligence decision-making methods are also relatively single, lacking a real-time update and feedback mechanism based on a dynamic knowledge graph. Since multi-robot systems need to process a large amount of complex data and changing environments, existing technologies often cannot dynamically optimize robot behavior through effective knowledge graphs and intelligent algorithms, resulting in insufficient swarm intelligence.

[0006] Existing technologies have not fully solved the problems of multi-robot systems operating in dynamic and complex environments, especially the deficiencies in information fusion, decision optimization, and swarm intelligence feedback, which limit the performance of robot systems in practical applications. The present invention proposes a robot control system that fuses multi-modal sensor data, dynamically constructs a knowledge graph, and is based on reinforcement learning technology, aiming to solve the defects existing in the prior art and improve the intelligence level of multi-robots. Summary of the Invention

[0007] In view of the above-mentioned technical problems, the present invention provides a method for multi-source heterogeneous data fusion and dynamic knowledge graph construction, aiming to improve the execution efficiency and response ability of an omnidirectional mobile robot in a complex dynamic environment by fusing multi-source heterogeneous data and constructing a knowledge graph with learning ability.

[0008] To achieve the above object, the technical solution adopted by the present invention is as follows:

[0009] A method for multi-source heterogeneous data fusion and dynamic knowledge graph construction. First, perform multi-source heterogeneous data fusion, construct a distributed Kalman filter network, design dedicated filters for sensors such as lidar, vision, and IMU respectively to achieve data spatio-temporal alignment; at the same time, use the improved D-S evidence theory for conflict detection and resolution, and dynamically adjust the fusion weight through a confidence evaluation model. Second, construct a dynamic knowledge graph, establish a three-layer graph structure including environmental entities, spatio-temporal relationships, and probabilistic reasoning; design an event-triggered incremental update algorithm to start local reconstruction when the environmental change degree exceeds a threshold, and introduce an LSTM memory network to enhance the temporal reasoning ability. Finally, construct a reinforcement learning model based on the dynamic knowledge graph, design a multi-objective reward function, and finally train until the model converges. The specific steps are as follows:

[0010] Step S1: Multi-source heterogeneous data fusion;

[0011] Step S1.1: First, deploy the sensors of the omnidirectional mobile robot, install the lidar, vision sensor, and IMU on the omnidirectional mobile robot. Among them, the lidar samples at 10 Hz to obtain lidar data; the vision sensor collects at 30 Hz to obtain vision data; the IMU samples at 200 Hz to obtain IMU data.

[0012] Step S1.2: Then, perform time synchronization of multi-source heterogeneous data. The specific operation is to unify the spatio-temporal benchmarks of different sensors (lidar, vision sensor, IMU). For lidar data, use the extended Kalman filter (EKF), linearize the nonlinear motion during omnidirectional wheel steering through the Jacobian matrix, and then obtain the three-dimensional point cloud data of the omnidirectional mobile robot. For IMU data, use the Butterworth low-pass filter method for processing to obtain the acceleration and angular velocity of the omnidirectional mobile robot. Therefore, define the extended Kalman filter (EKF) network as:

[0013]

[0014]

[0015] Among them, is the state estimation at time k (including position x, y, and velocity); J fis the Jacobian matrix of the state transition function, used for linearizing the nonlinear motion; Q k = diag(0.1, 0.1, 0.05) represents the noise covariance; w k represents the Gaussian process noise; f(.) represents the nonlinear state transition function; u k represents the control input; represents the covariance matrix of the state estimate at time k; P k-1 represents the covariance matrix of the state estimate at time k-1.

[0016] Step S1.3: For visual data, use the unscented Kalman filter (UKF), and utilize Sigma-point sampling to maintain the nonlinear characteristics of rotational motion, thereby obtaining the pose estimate of six degrees of freedom.

[0017] Define the unscented Kalman filter (UKF):

[0018]

[0019] where, X i represents the Sigma-point set, used to approximate the nonlinear distribution; λ represents the scaling hyperparameter, usually set to -6, and its role is to adjust the Sigma-point distribution range; n = 6 represents the state dimension (x, y, θ, v_x, v_y, w); represents the state estimate at time k-1; i represents the i-th Sigma point.

[0020] The unscented Kalman filter UKF transmits the nonlinear characteristics through Sigma-point sampling, avoiding the calculation of the Jacobian matrix. Visual data can be accurately estimated for the rotational angular velocity of the omnidirectional mobile robot through Sigma-point propagation.

[0021] Step S1.4: The perception ability of the robot in a dynamic environment depends on the accurate fusion of sensor data. The above text constructs different Kalman filter networks (EKF and UKF) to solve the spatio-temporal alignment problem of heterogeneous sensors such as lidar, vision, and IMU. However, in a multi-sensor fusion system, visual data and IMU data will conflict with each other, and the traditional solutions mainly involve processing through preset rules. For example, the two sensors are assigned fixed proportion weights (such as 70% for visual data and 30% for IMU data), or the data with the latest timestamp is used to overwrite the old data. Although these methods are simple to implement, they have obvious limitations - fixed weights cannot adapt to changes in the dynamic environment. To address these problems, this paper proposes an improved conflict resolution mechanism: First, evaluate the uncertainty of sensor data (such as the error propagation based on IMU state prediction and the noise level of the vision sensor), and dynamically adjust the contribution weights of the two in the fusion. When the IMU prediction error increases, automatically increase the influence of visual data, and vice versa, strengthen the role of IMU data when the vision sensor is disturbed.

[0022] To evaluate the uncertainty of sensor data, a confidence evaluation model is defined:

[0023]

[0024] where C i represents the confidence of the sensor at time i; Acc i is the historical accuracy of the sensor, which can be obtained through offline calibration; Rel i is the device reliability coefficient, which is internally based on failure rate statistics; t delay is the data delay time; t th = 100 ms is the maximum allowable delay. The confidence is affected by three factors: historical accuracy, device health, and data delay rate.

[0025] From formula (4), the D-S evidence synthesis formula can be derived:

[0026]

[0027] where Δ BC is the conflict correction term, used to adjust the credibility assignment when the intersection of event B and event C is an empty set. In this paper, the default form is adopted, that is, the conflict part is evenly distributed, and the value is set to 0.1; K represents the conflict factor; ε = 0.001, representing the stability parameter, used to prevent division by zero error; represents the credibility of event A after the fusion of the first evidence source (IMU data) and the second evidence source (visual data); m1(B) represents the credibility of the first evidence source (IMU data); m2(C) represents the credibility of the second evidence source (visual data); A represents the intersection of event B and event C, that is, the event where the IMU data and the visual data do not conflict; B represents the event where the IMU data accuracy is accurate; C represents the event where the visual data accuracy is accurate. Through formula (5), the data of lidar, vision, and IMU sensors can be fused to obtain the fused data.

[0028] Step S2: Incremental construction of the dynamic knowledge graph;

[0029] The fused multi-source data needs to be transformed into inferable environmental knowledge. In this step, a three-layer graph structure including entity relationships, probabilistic reasoning, and historical memory is constructed, and updates are realized through an event trigger mechanism. Specifically as follows:

[0030] Step S2.1: Construct entity relationships.

[0031] Define that the basic layer of the dynamic knowledge graph represents entity relationships, specifically as:

[0032] G b = <V, E>, V = {vi |v i = (position, type, confidence)}, E = {e ij |e ij = (distance, interaction type)} (6) where, v i represents a node, and its node attributes include coordinates, entity type, confidence (0 - 1); e ij represents an edge, and its attributes include Euclidean distance (representing weight), relationship type; G b represents a weighted graph; V represents the vertex set, that is, all nodes in the graph; E represents the edge set, that is, the connection relationship between nodes.

[0033] Step S2.2: Through formula (6), the inference layer can be designed to achieve probabilistic inference, specifically as follows:

[0034]

[0035] where, σ = 0.2, representing the distance standard deviation, characterizing the uncertainty of spatial relationships; Z is the normalization factor to ensure that the sum of probabilities is 1; P(v i |Pa(v i )) represents the probability of the current node v i when its parent node Pa(v i )(prior information) is known; Pa(v i ) represents the probability of the parent node being established; Pa(.) represents the function of prior knowledge under conditional probability; vj represents node j; N(.) represents the Gaussian distribution; pos i represents the position of node i; pos j represents the position of node j.

[0036] Step S2.3: According to formula (6) and formula (7), the historical memory layer can be designed:

[0037]

[0038] where, h t is the hidden state at time t and is also the encoded historical trajectory; env t represents the environmental characteristics (such as light intensity, pedestrian flow density); h t-1 represents the hidden state at time t - 1; pos t represents the position information at time t; vel t represents the velocity information at time t; LSTM represents the long short - term memory network, which is used to selectively retain or forget historical information.

[0039] Step S2.4: So far, the three-layer design of the dynamic knowledge graph has been completed, namely the basic layer, the reasoning layer, and the memory layer of the dynamic knowledge graph. Next, the event-triggered update is designed.

[0040] Define the environmental change degree as:

[0041]

[0042] where M represents the number of monitored entities; Δε is the environmental change degree; represents the pose after the i-th moment; represents the pose before the i-th moment.

[0043] Adopt an event-triggered design update mechanism: Through the above historical memory layer, the historical entity states (including position, environment, and relationship, etc.) can be obtained, and the total displacement of all entities in the dynamic knowledge graph is calculated. When more than 30% of the entity states in the dynamic knowledge graph change, and the total displacement of the entities in the state accounts for more than the threshold of the total displacement, start local graph reconstruction, and only update the affected entity states instead of global reconstruction, reducing the computational overhead. The threshold refers to thirty percent.

[0044] When the total entity displacement exceeds the threshold, local reconstruction will be triggered. Neighbors within the query radius in the graph will be queried, new relationships will be re-added, old edges will be deleted and new edges will be added, and finally the LSTM memory will be updated.

[0045] Step S3: Build a reinforcement learning model based on the dynamic knowledge graph;

[0046] The dynamic knowledge graph provides environmental context for decision-making. In this step, a reinforcement learning model based on the dynamic knowledge graph is designed to embed the dynamic knowledge graph information into the decision-making process. When designing the reinforcement learning model framework, the self-information of the omnidirectional mobile robot, the information of the dynamic knowledge graph, and the task requirements need to be considered. Therefore, these three parameters need to be designed as the state space of the agent in the reinforcement learning model, and then the agent in the model can be trained. The state space design of the agent in the model is as follows:

[0047] s t =[vec(G t ),robot_state,task_req] (10)

[0048] where vec(G t ) is the graph embedding vector of the graph; robot_state is the state of the omnidirectional mobile robot, and task_req is parameters such as the priority of the task; s t represents the state space of the agent in the model.

[0049] Then define the multi-objective reward function:

[0050]

[0051] Among them, R t represents the reward value at time t; t curr represents the currently consumed time; t plan represents the planned time of the task; min obs d obs represents the distance to the nearest obstacle; E used represents the cumulative energy consumption; E total represents the total energy consumption.

[0052] In formula (11), the first term is the task timeliness reward; the second term is the safety reward; the third term is the energy consumption penalty.

[0053] From formula (10) and formula (11), based on the PPO (policy objective function) algorithm, a loss function can be designed and gradient descent training can be carried out. The specific formulas for these two parts are:

[0054]

[0055] Among them, γ = 0.99 is the discount factor, which balances the current and future rewards; Q' is the output of the target network, lagging behind the update of the online network; θc represents the gradient descent parameter; L represents the loss function; E[.] represents the expected value of all samples; Q(s,a) represents the state function of the current network; y represents the target value; r represents the reward value; s' represents the network state at the next moment; π(s') represents the best state selected by the network at the next moment.

[0056] Step S4: Perform feedback optimization;

[0057] During the training process of the above reinforcement learning model, in order to dynamically control the learning rate to accelerate the training process. The output of the target network can be fed back to the edge set V in the dynamic knowledge graph constructed in step two, and the feedback result in the dynamic knowledge graph can be designed as:

[0058]

[0059] Among them, η is the learning rate, set to 0.01 to control the weight update step size; Consistency(e) represents the consistency measure of the dynamic knowledge graph relationship; represents the new weight of the edge in the dynamic knowledge graph; represents the original weight of the edge in the dynamic knowledge graph; represents the gradient descent formula; G represents the graph structure of the dynamic knowledge graph; e represents the edge in the dynamic knowledge graph.

[0060] To construct the consistency metric, specifically, the multi-source heterogeneous data integrated in Step 1 is required to obtain the distance of the obstacle; secondly, the entity state information in the construction of the dynamic knowledge graph in Step 2 is also required, specifically the distance information of the entity. The consistency metric is designed as:

[0061]

[0062] where p pred is the distance of the obstacle in the dynamic knowledge graph; p obs is the entity.

[0063] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0064] (1) Through the EKF / UKF combined filtering, the present invention reduces the positioning error compared with the traditional single filtering method; through the dynamic confidence evaluation model, the robot can still maintain stability when there is a sensor conflict; in addition, through the conflict resolution mechanism, the spatial alignment error of multi-sensor data is reduced.

[0065] (2) In terms of response speed, compared with the traditional method, the event-triggered update mechanism of the present invention improves the dynamic response speed, and the LSTM memory layer can also trace back to the historical environment state, and the prediction of the obstacle movement trajectory can also be improved through the probability inference layer.

[0066] (3) The present invention designs a multi-objective reward function according to the deficiencies of the prior art and the objectives, taking into account conflict avoidance and energy consumption penalty, which can improve the training efficiency: through the hierarchical design of the multi-objective reward function (basic task completion reward + path smoothness reward + safety distance maintenance reward), the model can obtain effective gradient signals in the early stage, thereby improving the training efficiency. In addition, by adding feedback optimization, the learning rate can be dynamically controlled, and it is easier to find the optimal state and output. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 is a schematic diagram of path tracking based on multi-source heterogeneous data fusion in the present invention;

[0068] Figure 2 is a comparison diagram of the fusion data method and the use of the traditional method;

[0069] Figure 3 is a flowchart of the method for constructing a dynamic knowledge graph in the present invention;

[0070] Figure 4 is a comprehensive state diagram of the dynamic knowledge graph;

[0071] Figure 5 is a training progress curve diagram of the dynamic knowledge graph. DETAILED DESCRIPTION OF THE INVENTION

[0072] The present invention provides a method for multi-source heterogeneous data fusion and building a dynamic knowledge graph, aiming to improve the execution efficiency and response ability of a robot in a complex dynamic environment by fusing multi-source heterogeneous data and constructing a dynamic knowledge graph with learning ability. Its technical solution can be realized by constructing multi-sensor fusion trajectory simulation, dynamic knowledge graph construction and visualization, and reinforcement learning model construction, and each link needs to be completed in combination with algorithms and manual operations. The specific implementation process is as follows:

[0073] Step S1: Multi-source heterogeneous data fusion;

[0074] Step S1.1: First, deploy the sensors of the omnidirectional mobile robot, and install the lidar, vision sensor and IMU on the omnidirectional mobile robot. Among them, the lidar samples at 10Hz to obtain lidar data; the vision sensor collects at 30Hz to obtain vision data; the IMU samples at 200Hz to obtain IMU data.

[0075] Step S1.2: Then, perform time synchronization of multi-source heterogeneous data. The specific operation is to unify the spatio-temporal benchmarks of different sensors. For lidar data, the extended Kalman filter (EKF) is used. The nonlinear motion during omnidirectional wheel steering is linearized through the Jacobian matrix, and then the three-dimensional point cloud data of the omnidirectional mobile robot is obtained. For IMU data, the Butterworth low-pass filter method is used for processing to obtain the acceleration and angular velocity of the omnidirectional mobile robot. Therefore, define the extended Kalman filter (EKF) network as:

[0076]

[0077]

[0078] Among them, is the state estimation at time k (including position x, y and velocity); J f is the Jacobian matrix of the state transition function, which is used to linearize the nonlinear motion; Q k = diag(0.1, 0.1, 0.05) represents the noise covariance; w k represents the Gaussian process noise; f(.) represents the nonlinear state transition function; u k represents the control input; represents the covariance matrix of the state estimation at time k; P k-1 represents the covariance matrix of the state estimation at time k-1.

[0079] Step S1.3: For vision data, the unscented Kalman filter (UKF) is used to utilize the Sigma point sampling to maintain the nonlinear characteristics of rotational motion, and then the pose estimation of 6 degrees of freedom is obtained.

[0080] Define the Unscented Kalman Filter (UKF):

[0081]

[0082] where X i represents the set of Sigma points, which is used to approximate the non - linear distribution; λ represents the scaling hyperparameter, usually set to - 6, and its role is to adjust the distribution range of Sigma points; n = 6 represents the state dimension (x, y, θ, v_x, v_y, w); represents the state estimate at time k - 1; i represents the i - th Sigma point.

[0083] The Unscented Kalman Filter UKF transfers the non - linear characteristics through Sigma - point sampling, avoiding the calculation of the Jacobian matrix. Visual data can be propagated through Sigma points, and the angular velocity of the omnidirectional mobile robot can be accurately estimated.

[0084] Step S1.4: The perception ability of the robot in a dynamic environment depends on the accurate fusion of sensor data. Above, by constructing different Kalman filter networks (EKF and UKF), the spatio - temporal alignment problem of heterogeneous sensors such as lidar, vision, and IMU is solved. However, in a multi - sensor fusion system, visual data and IMU data will conflict with each other. The traditional solutions mainly deal with it through preset rules. For example, the two sensors are assigned fixed - ratio weights (such as 70% for visual data and 30% for IMU data), or the data with the latest timestamp is used to overwrite the old data. Although these methods are simple to implement, they have obvious limitations - fixed weights cannot adapt to the changes in the dynamic environment. To address these problems, this paper proposes an improved conflict resolution mechanism: First, by evaluating the uncertainty of sensor data (such as the error propagation based on IMU state prediction and the noise level of visual sensors), the contribution weights of the two in the fusion are dynamically adjusted. When the IMU prediction error increases, the influence of visual data is automatically increased, and vice versa, when the visual sensor is disturbed, the role of IMU data is strengthened.

[0085] To evaluate the uncertainty of sensor data, a confidence evaluation model is defined:

[0086]

[0087] where C i represents the confidence of sensor at time i; Acc i is the historical accuracy of the sensor, which can be obtained through offline calibration; Rel i is the device reliability coefficient, which is based on the failure rate statistics internally; t delay is the data delay time; t th= 100 ms is the maximum allowable delay. The confidence level is affected by three factors: historical accuracy, device health, and data latency rate.

[0088] From formula (4), the D-S evidence synthesis formula can be derived:

[0089]

[0090] where, Δ BC is the conflict correction term, which is used to adjust the credibility assignment when the intersection of event B and event C is an empty set. In this embodiment, the default form is adopted, that is, the conflict part is evenly distributed, and the value is set to 0.1; K represents the conflict factor; ε = 0.001, which represents the stability parameter and is used to prevent division-by-zero errors; represents the credibility of event A after the fusion of the first evidence source (IMU data) and the second evidence source (visual data); m1(B) represents the credibility of the first evidence source (IMU data); m2(C) represents the credibility of the second evidence source (visual data); A represents the intersection of event B and event C, that is, the event where the IMU data and the visual data do not conflict; B represents the event where the IMU data accuracy is accurate; C represents the event where the visual data accuracy is accurate. Through formula (5), the data of lidar, vision, and IMU sensors can be fused to obtain the fused data.

[0091] This embodiment adopts a multi-sensor data fusion architecture and realizes time synchronization through a timestamp alignment mechanism. With the extended Kalman filter as the core, a six-dimensional state vector including position, velocity, heading angle, and sensor bias is constructed, and the following key parameters are set: the IMU acceleration noise variance is set to 0.01, and the vision sensor noise variance is set to 0.01. In the dynamic performance verification stage, through a preset three-stage motion trajectory test (uniform linear motion, 90-degree turn, acceleration and deceleration motion), the fusion result is as Figure 1 shown. The fusion trajectory (blue) closely fits the real trajectory (green), indicating that the multi-source heterogeneous data integrated positioning effect is excellent. As Figure 2 shown, comparing the effects of each sensor and multi-sensor fusion, the black trajectory is the true value trajectory, the blue trajectory is the fusion trajectory, the pink dotted line is the single IMU trajectory, and the green dotted line is the single vision sensor. By zooming in, it can be seen that the fusion trajectory (blue solid line) has an excellent tracking effect compared to the single sensor trajectory, indicating the feasibility of this method.

[0092] Step S2: Dynamically construct an incremental knowledge graph;

[0093] The fused multi-source data needs to be transformed into inferable environmental knowledge. In this step, a three-layer graph structure including entity relationships, probabilistic reasoning, and historical memory is constructed, and updates are realized through an event trigger mechanism. As Figure 3This is a flow chart of a dynamic knowledge graph, and the specific steps are as follows:

[0094] Step S2.1: Construct entity relationships.

[0095] Define the entity relationships in the basic layer of the dynamic knowledge graph, specifically as follows:

[0096] G b = <V, E>, V = {v i |v i = (position, type, confidence)}, E = {e ij |e ij = (distance, interaction type)} (6) where v i represents a node, and its node attributes include coordinates, entity type, and confidence (0 - 1); e ij represents an edge, and its attributes include Euclidean distance (representing weight) and relationship type; G b represents a weighted graph; V represents the vertex set, that is, all nodes in the graph; E represents the edge set, that is, the connection relationships between nodes.

[0097] Step S2.2: Through formula (6), an inference layer can be designed to achieve probabilistic inference, specifically as follows:

[0098]

[0099] where σ = 0.2, representing the distance standard deviation, characterizing the uncertainty of spatial relationships; Z is a normalization factor to ensure that the sum of probabilities is 1; P(v i |Pa(v i )) represents the probability of the current node v i when its parent node Pa(v i )(prior information) is known; Pa(v i ) represents the probability of the parent node being established; Pa(.) represents the function of prior knowledge under conditional probability; vj represents node j; N(.) represents the Gaussian distribution; pos i represents the position of node i; pos j represents the position of node j.

[0100] Step S2.3: According to formula (6) and formula (7), a historical memory layer can be designed:

[0101] h t = LSTM(h t-1 , [pos t , vel t , env t ) (8)

[0102] where h tis the hidden state at time t and also encodes the historical trajectory; env t represents environmental features (such as light intensity, pedestrian flow density); h t-1 represents the hidden state at time t-1; pos t represents the position information at time t; vel t represents the speed information at time t; LSTM represents the long short-term memory network, which is used to selectively retain or forget historical information.

[0103] Step S2.4: So far, the three-layer design of the dynamic knowledge graph has been completed, namely the basic layer, the reasoning layer, and the memory layer of the dynamic knowledge graph. Next, the event-triggered update is designed.

[0104] Define the environmental change degree as:

[0105]

[0106] where M represents the number of entities being monitored; Δε is the environmental change degree; represents the pose after the i-th moment; represents the pose before the i-th moment.

[0107] Adopt an event-triggered design update mechanism: Through the above historical memory layer, the historical entity states (including position, environment, and relationships, etc.) can be obtained, and the total displacement of all entities in the dynamic knowledge graph is calculated. When more than 30% of the entity states in the dynamic knowledge graph change, and the total displacement of the entities in the state accounts for more than the threshold of the total displacement, start local graph reconstruction, and only update the affected entity states instead of global reconstruction, reducing the computational overhead. The threshold refers to thirty percent.

[0108] When the total entity displacement exceeds the threshold, local reconstruction will be triggered. Neighbors within the query radius in the graph will be queried, new relationships will be re-added, old edges will be deleted and new edges will be added, and finally the LSTM memory will be updated.

[0109] Step S3: Build a reinforcement learning model based on the dynamic knowledge graph;

[0110] The dynamic knowledge graph provides environmental context for decision-making. In this step, a reinforcement learning model based on the dynamic knowledge graph is designed to embed the dynamic knowledge graph information into the decision-making process. When designing the reinforcement learning model framework, the self-information of the omnidirectional mobile robot, the information of the dynamic knowledge graph, and the task requirements need to be considered. Therefore, these three parameters need to be designed as the state space of the agent in the reinforcement learning model, and then the agent in the model can be trained. The state space design of the agent in the model is as follows:

[0111] s t =[vec(G t), robot_state, task_req] (10)

[0112] Among them, vec(G t ) is the graph embedding vector of the graph spectrum; robot_state is the state of the omnidirectional mobile robot, and task_req is parameters such as the priority of the task; s t represents the state space of the agent in the model.

[0113] Next, define the multi-objective reward function:

[0114]

[0115] Among them, R t represents the reward value at time t; t curr represents the currently consumed time; t plan represents the planned time of the task; min obs d obs represents the distance to the nearest obstacle; E used represents the cumulative energy consumption; E total represents the total energy consumption.

[0116] In formula (11), the first term is the task timeliness reward; the second term is the safety reward; the third term is the energy consumption penalty.

[0117] From formula (10) and formula (11), based on the PPO (policy objective function) algorithm, the loss function can be designed and gradient descent training can be carried out. The specific formulas for these two parts are:

[0118]

[0119] Among them, γ = 0.99 is the discount factor, which balances the current and future rewards; Q' is the output of the target network, lagging behind the update of the online network; θc represents the gradient descent parameter; L represents the loss function; E[.] represents the expected value of all samples; Q(s,a) represents the state function of the current network; y represents the target value; r represents the reward value; s' represents the network state at the next moment; π(s') represents the best state selected by the network at the next moment.

[0120] Step S4: Perform feedback optimization;

[0121] During the training process of the above reinforcement learning model, in order to dynamically control the learning rate to accelerate the training process. The output of the target network can be fed back to the edge set V of the dynamic knowledge graph constructed in step two, and the feedback result in the dynamic knowledge graph can be designed as:

[0122]

[0123] Among them, η is the learning rate, set to 0.01 to control the weight update step; Consistency(e) represents the consistency measure of the dynamic knowledge graph relationship; represents the new weight of the edge of the dynamic knowledge graph; represents the original weight of the edge of the dynamic knowledge graph; represents the gradient descent formula; G represents the graph structure of the dynamic knowledge graph; e represents the edge in the dynamic knowledge graph.

[0124] To construct the consistency measure, specifically, the multi-source heterogeneous data integrated in Step 1 is required to obtain the distance of the obstacle; secondly, the entity state information in the construction of the dynamic knowledge graph in Step 2 is also required, specifically the distance information of the entity. The consistency measure is designed as:

[0125]

[0126] where p pred is the distance of the obstacle in the dynamic knowledge graph; p obs is the entity.

[0127] Such as Figure 4 , the dynamic knowledge graph is monitored by a visualization method, which is used to represent the state, goal, path planning of the agent in the environment, as well as the influence of dynamic and static obstacles. When the agent moves, it is detected in real time whether the current position is an obstacle. If it is an obstacle and the current number of obstacle nodes does not reach the upper limit, a new obstacle type node is added to the graph, and an avoidance relationship edge is established with the agent node. At the same time, a moving trajectory queue with a capacity of 20 steps is maintained to record the historical path of the last 20 steps. This queue adopts the first-in-first-out strategy to ensure that it will not grow infinitely and occupy memory.

[0128] In the figure, the black square represents the position of the agent in the reinforcement learning model, which serves as the starting point of path planning and connects multiple goals, including goals, path points, and obstacles; the black star represents the goal, indicating the final task of the agent. The circles and crosses are both obstacles, and the agent needs to bypass them; finally, the triangles are path points, which record the path of the agent.

[0129] Such as Figure 5 , it can be seen from the reinforcement learning training results that the final reward value stabilizes at a positive value. By viewing the training process, it can also be found that the path history of the agent is [(1,0),(1,1),(1,2),(1,3),(2,3)], and it finally reaches the target point [2 3], indicating that the reinforcement learning model based on the dynamic knowledge graph can demonstrate the path planning and obstacle avoidance strategy of the agent.

[0130] The above-described embodiments merely represent the implementation manners of the present invention, but should not be construed as limiting the scope of the present invention. It should be noted that for those skilled in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all fall within the protection scope of the present invention.

Claims

1. A method for multi-source heterogeneous data fusion and dynamically building a knowledge graph, characterized in that The method first performs multi-source heterogeneous data fusion to construct a distributed Kalman filtering network, designs filters for different sensors respectively to achieve data spatio-temporal alignment; at the same time, an improved D-S evidence theory is used for conflict detection and resolution, and the fusion weights are dynamically adjusted through a confidence evaluation model; Secondly, a dynamic knowledge graph is constructed, and a three-layer graph structure including environmental entities, spatio-temporal relationships, and probabilistic reasoning is established; an event-triggered incremental update algorithm is designed, and local reconstruction is started when the environmental change degree exceeds the threshold, and an LSTM memory network is introduced to enhance the temporal reasoning ability; finally, a reinforcement learning model based on the dynamic knowledge graph is constructed, a multi-objective reward function is designed, and finally the model is trained until convergence.

2. The method for fusing multi-source heterogeneous data and dynamically building a knowledge graph according to claim 1, characterized in that The method specifically includes the following steps: Step S1: Multi-source heterogeneous data fusion; Step S1.1: Install three sensors, namely lidar, vision sensor and IMU, on the omnidirectional mobile robot to obtain lidar data, vision data and IMU data respectively; Step S1.2: Perform time synchronization of multi-source heterogeneous data to unify the spatio-temporal benchmarks of different sensors; Step S1.3: For vision data, use the unscented Kalman filter UKF, and utilize Sigma point sampling to maintain the nonlinear characteristics of rotational motion to obtain a six-degree-of-freedom pose estimate; The unscented Kalman filter UKF transfers the nonlinear characteristics through Sigma point sampling, and the vision data propagates through Sigma points, which can accurately estimate the rotational angular velocity of the omnidirectional mobile robot; Step S1.4: Propose an improved conflict resolution mechanism: first, by evaluating the uncertainty of sensor data, dynamically adjust the contribution weights of the two in the fusion. When the IMU prediction error increases, automatically increase the influence of vision data, and vice versa, strengthen the role of IMU data when the vision sensor is interfered; To evaluate the uncertainty of sensor data, a confidence evaluation model is defined: Among them, C i represents the confidence level of the sensor at time i; Acc i is the historical accuracy of the sensor, which can be obtained through offline calibration; Rel i is the device reliability coefficient, which is based on the failure rate statistics internally; t delay is the data delay time; t th = 100ms is the maximum allowable delay; the confidence level is affected by three factors: historical accuracy, device health, and data delay rate; Step S2: Incremental construction of the dynamic knowledge graph; The fused multi-source data needs to be transformed into inferable environmental knowledge, a three-layer graph structure including entity relationships, probabilistic reasoning and historical memory, and is updated through an event-triggered mechanism; specifically as follows: Step S2.1: Construct entity relationships; Define the basic layer of the dynamic knowledge graph to represent entity relationships, specifically: G b = <V, E>, V = {v i | v i = (position, type, confidence)}, E = {e ij | e ij = (distance, interaction type)} (6) where v i represents a node, and its node attributes include coordinates, entity type, and confidence; e ij represents an edge, and its attributes include the Euclidean distance representing the weight and the relationship type; G b represents a weighted graph; V represents the vertex set, that is, all nodes in the graph; E represents the edge set, that is, the connection relationship between nodes; Step S2.2: Design the inference layer through formula (6) to achieve probabilistic reasoning, specifically as follows: Among them, σ represents the distance standard deviation, characterizing the uncertainty of spatial relationship; Z is a normalization factor to ensure that the sum of probabilities is 1; P(v i |Pa(v i )) represents the probability of the current node v i when its parent node Pa(v i ) is known; Pa(v i ) represents the probability that the parent node holds; Pa(.) represents a function of prior knowledge under conditional probability; vj represents node j; N(.) represents the Gaussian distribution; pos i represents the position of node i; pos j represents the position of node j; Step S2.3: Design the historical memory layer according to formula (6) and formula (7): h t = LSTM(h t-1 , [pos t , vel t , env t ) (8) Among them, h t is the hidden state at time t and is also the encoded historical trajectory; env t represents the environmental characteristics; h t-1 represents the hidden state at time t-1; pos t represents the position information at time t; vel t represents the velocity information at time t; LSTM represents the long short-term memory network; Step S2.4: Complete the three-layer design of the dynamic knowledge graph, namely the basic layer, inference layer and memory layer of the dynamic knowledge graph, and then design an event-triggered update; Define the environmental change degree and adopt an event-triggered design update mechanism: Obtain the historical entity states through the historical memory layer and calculate the total displacement of all entities in the dynamic knowledge graph; When more than 30% of the entity states in the dynamic knowledge graph change, and the total displacement of the entities in the state exceeds the threshold, initiate local graph reconstruction, only update the affected entity states instead of global reconstruction, reducing the computational overhead; When the total entity displacement exceeds the threshold, trigger local reconstruction, query the neighbors within the query radius in the graph, re-add new relationships, delete old edges and add new edges, and finally update the LSTM memory; Step S3: Build a reinforcement learning model based on the dynamic knowledge graph; The dynamic knowledge graph provides environmental context for decision-making. Design a reinforcement learning model based on the dynamic knowledge graph and embed the dynamic knowledge graph information into the decision-making process; When designing the reinforcement learning model framework, consider the self-information of the omnidirectional mobile robot, the information of the dynamic knowledge graph, and the task requirements, and design these three parameters as the state space of the agent in the reinforcement learning model, and then train the agent in the model; Step S4: Perform feedback optimization; Feed the output of the target network back to the edge set V in the dynamic knowledge graph built in Step 2. The feedback result in the dynamic knowledge graph is designed as: Among them, η is the learning rate, set to 0.01 to control the weight update step; Consistency(e) represents the consistency measure of the dynamic knowledge graph relationship; represents the new weight of the dynamic knowledge graph edge; represents the original weight of the dynamic knowledge graph edge; represents the gradient descent formula; G represents the graph structure of the dynamic knowledge graph; e represents the edge in the dynamic knowledge graph; Construct a consistency metric. Obtain the distance of the obstacle through the multi-source heterogeneous data fused in Step 1; And obtain the consistency metric based on the entity state information in the construction of the dynamic knowledge graph in Step 2.

3. The method for fusing multi-source heterogeneous data and dynamically building a knowledge graph according to claim 2, wherein, In the step S1.1, the lidar samples at 10Hz to obtain lidar data; The visual sensor collects at 30Hz to obtain visual data; The IMU samples at 200Hz to obtain IMU data.

4. A method for multi-source heterogeneous data fusion and dynamically building a knowledge graph according to claim 2, characterized in that, The step S1.2 is specifically as follows: For the lidar data, use the Extended Kalman Filter (EKF), linearize the nonlinear motion during omnidirectional wheel steering through the Jacobian matrix, and then obtain the three-dimensional point cloud data of the omnidirectional mobile robot; For the IMU data, use the Butterworth low-pass filter method for processing to obtain the acceleration and angular velocity of the omnidirectional mobile robot; Therefore, define the Extended Kalman Filter (EKF) network as: Among them, is the state estimate at time k (including position x, y, and velocity); J f is the Jacobian matrix of the state transition function, used to linearize the non-linear motion; Q k = diag(0.1, 0.1, 0.05) represents the noise covariance; w k represents the Gaussian process noise; f(.) represents the non-linear state transition function; u k represents the control input; represents the covariance matrix of the state estimate at time k; P k-1 represents the covariance matrix of the state estimate at time k-1.

5. A method for fusing multi-source heterogeneous data and dynamically constructing a knowledge graph according to claim 2, characterized in that, In the step S1.3, define the Unscented Kalman Filter UKF as follows: Among them, X i represents the Sigma point set, which is used to approximate the non-linear distribution; λ represents the scaling hyperparameter, usually set to -6, and its function is to adjust the Sigma point distribution range; n = 6 represents the state dimension (x, y, θ, v_x, v_y, w); represents the state estimate at time k-1; i represents the i-th Sigma point.

6. A method for fusing multi-source heterogeneous data and dynamically building a knowledge graph according to claim 2, characterized in that, In the step S1.4, further propose the D-S evidence synthesis formula from formula (4): Among them, Δ BC is a conflict correction term used to adjust the credibility assignment when the intersection of event B and event C is an empty set. It adopts the default form, that is, the conflict part is evenly distributed, and the value is set to 0.1; K represents the conflict factor; ε = 0.001, which represents the stability parameter used to prevent division-by-zero errors; represents the credibility of event A after the fusion of the first evidence source and the second evidence source. The first evidence source is IMU data, and the second evidence source is visual data; m1(B) represents the credibility of the first evidence source; m2(C) represents the credibility of the second evidence source; A represents the intersection of event B and event C, that is, the event where the IMU data and the visual data do not conflict; B represents the event where the IMU data accuracy is accurate; C represents the event where the visual data accuracy is accurate; Through formula (5), the data of lidar, vision, and IMU sensors can be fused to obtain the fused data.

7. A method for fusing multi-source heterogeneous data and dynamically building a knowledge graph according to claim 2, characterized in that In the step 2: In the step S2.1, the confidence level is 0 to 1; In the step S2.2, the distance standard deviation σ = 0.2; In the step S2.4, the threshold refers to thirty percent.

8. A method for multi-source heterogeneous data fusion and dynamically building a knowledge graph according to claim 2, characterized in that In the step S2.4, the formula for defining the environmental change degree is: Among them, M represents the number of entities to be monitored; Δε is the degree of environmental change; represents the pose after the i-th moment; represents the pose before the i-th moment.

9. A method for multi-source heterogeneous data fusion and dynamically building a knowledge graph according to claim 2, characterized in that, The step S3 is specifically as follows: First, the state space of the agent in the model is designed as follows: s t = [vec(G t ), robot_state, task_req] (10) Among them, vec(G t ) is the graph embedding vector of the graph spectrum; robot_state is the state of the omnidirectional mobile robot, and task_req are parameters such as the priority of the task; s t represents the state space of the agent in the model; Then define the multi-objective reward function: Among them, R t represents the reward value at time t; t curr represents the currently consumed time; t plan represents the planned time of the task; min obs d obs represents the distance to the nearest obstacle; E used represents the cumulative energy consumption; E total represents the total energy consumption; In formula (11), the first term is the task timeliness reward; The second term is the safety reward; The third term is the energy consumption penalty; From formula (10) and formula (11), based on the PPO (policy objective function) algorithm, design the loss function and perform gradient descent training. The specific formulas for these two parts are: Among them, γ = 0.99 is the discount factor, which balances the current and future rewards; Q' is the output of the target network and lags behind the update of the online network; θc represents the gradient descent parameter; L represents the loss function; E[.] represents the expected value of all samples; Q(s,a) represents the state function of the current network; y represents the target value; r represents the reward value; s' represents the network state at the next moment; π(s') represents the best state selected by the network at the next moment.

10. A method for multi-source heterogeneous data fusion and dynamically building a knowledge graph according to claim 2, characterized in that The consistency metric in the step 4 is designed as: Among them, p pred is the distance of the obstacle in the dynamic knowledge graph; p obs is an entity.

Images