CN-120066023-B - Trans-GCN-based mobile robot dead reckoning method

CN120066023BCN 120066023 BCN120066023 BCN 120066023BCN-120066023-B

Abstract

The invention discloses a mobile robot dead reckoning method based on a Trans-GCN. The method comprises the following steps of S1, collecting wheel speed sensor and inertial sensor data of a wheel type mobile robot, intercepting time sequence data through a sliding window and an adaptive sliding step length, representing the sensor time sequence data into graph structure data, S2, constructing a Laplace position code for each node in a graph, linearly pooling and then converging the Laplace position code into a new embedded representation of the node, S3, embedding graph characteristic signals and coding information into a transducer, inputting output characteristics of an integral model into a full-connection layer to obtain an output vector, S4, constructing a training true value, training the model according to sample data, stopping training when the model is smaller than a threshold value, and storing the model, S5, inputting data in the condition of satellite signal missing into the trained model to obtain a two-dimensional position increment of the mobile robot in a predicted sampling time period, and realizing dead reckoning of a basic body sensing sensor signal under a GNSS signal.

Inventors

LU YONGLE
LUO YI
MA JUNJIE
HE JINGDAN
LU YANFEI
FAN JUXIANG
ZHANG YIMING
WANG YUE

Assignees

重庆邮电大学

Dates

Publication Date: 20260512
Application Date: 20250219

Claims (6)

1. The mobile robot dead reckoning method based on the Trans-GCN is characterized by comprising the following steps of: S1, collecting data of a wheel speed sensor and an inertial sensor of a wheel type mobile robot, intercepting time sequence data through a sliding window and an adaptive sliding step length, constructing sensor data with a single dimension as graph nodes in each window data, and constructing a bidirectional edge connection structure for part of nodes to form completed graph structure data; s2, constructing a graph rolling network GCN to extract the dependency relationship of data in the graph in a non-Euclidean space, wherein the dependency relationship comprises three steps of feature transmission, aggregation and updating to obtain node features and edge features; S3, decomposing an adjacency matrix of the constructed graph, constructing a position coding matrix of the nodes, integrating the position coding matrix with the linear layer and the L2 pooling layer, and then performing linear transformation on the position coding matrix and the characteristics extracted in the S2, and inputting the position coding matrix into a transducer encoder structure; S4, calculating the correlation between the target node and the neighbor node in each layer to obtain attention weight, connecting attention characteristics with original characteristic residual errors, integrating the attention characteristics with the original characteristic residual errors through a feedforward neural network FFN, and finally inputting the integrated attention characteristics into a fully connected network to obtain a predicted value of the network; S5, training the model according to the training sample data, stopping training when the verification loss is smaller than a threshold value, storing the model, and inputting the data of the wheel speed sensor and the inertial sensor into the model to obtain a dead reckoning result without GNSS signals; In the step S4, the input data respectively passes through four full connection layers to obtain four matrices of Query (Q), key (K), value (V), edge (E), and the feature data is projected to the Attention space, where the Attention head (h) and the K layer (h) have the following formula: wherein i is the number of the target node, j is the neighbor node, d K represents the dimension of the Key vector, and the Softmax function calculates each row of the matrix according to the following formula: Wherein y a represents the value of a certain row, a and column of the Attention-score matrix, y b represents the value of a certain row, b and column of the matrix, and w represents the number of columns of the matrix; When the Attention-score is calculated, residual connection is introduced, meanwhile Layernorm normalizes input data, and the characteristics of each node pass through a two-layer feedforward neural network and are used for further nonlinear mapping characteristic representation, wherein the specific formula is expressed as follows: Wherein the method comprises the steps of Is a layer of Layernorm which is formed by the steps of, Is the self-attention of the multi-head splice, W is the weight matrix, relu is the activation function, Represents the hidden state after LayerNorm normalization, The hidden state after the feedforward neural network and ReLu activation functions is added with nonlinear transformation, so that the model can learn more complex characteristics; And S4, integrating the attention characteristic with the original characteristic residual error, and finally inputting the integrated attention characteristic with the original characteristic residual error into a fully connected network to obtain a predicted value of the network, wherein the method specifically comprises the following steps of: After passing through the feedforward neural network of two layers, the self-adaptive learning rate adjusting factor is designed, namely the ratio of the current gradient norm to the weight norm, and the local learning rate scaling factor is calculated as follows: Where eta represents the global learning rate, Is that the current kth layer is at the current the gradient over the batch data is such that, λ represents the weight attenuation coefficient, and finally the node characteristic output through the Trans-GCN is as follows: And then performing inverse normalization after passing through the full connection layer to obtain the prediction output of the model.
2. The method for dead reckoning a mobile robot based on Trans-GCN according to claim 1, wherein the step S1 specifically includes: Recording three-axis angular velocity and three-axis acceleration time sequence data by using an inertial sensor of a mobile robot body, respectively recording left and right wheel two-dimensional data by using a wheel speed sensor, setting a sliding window with a certain length and a sliding step length, performing dimension reduction processing on the data, taking the mean value and standard deviation of single dimension window data, aligning a true value label with a timestamp, modeling the trained single dimension data without label data into nodes, setting edge connection between the nodes, generating a record adjacent matrix, and constructing complete graph structure training data.
3. The method for dead reckoning of mobile robot based on Trans-GCN according to claim 1, wherein the step S3 is characterized in that the constructed position coding matrix comprises three dimensions of feature data, after the feature matrix is obtained by the Laplacian matrix of the graph, the normalized feature vector and the normalized feature vector corresponding to the first m minimum feature values are selected, the normalized feature vector and the normalized feature vector are constructed into a three-dimensional initial position coding matrix together with the original feature vector, then new embedded representation is obtained through linear layer aggregation, and then final position coding together propagation learning is obtained through simplification of an L2 pooling layer.
4. The Trans-GCN-based mobile robot dead reckoning method of claim 1, wherein in step S5, the training truth value is derived from latitude and longitude data output by an RTK differential positioning system, the latitude and longitude data is converted as follows, the latitude and longitude coordinates in a geographic coordinate system are converted into a local coordinate system, r e represents the equatorial radius, r p represents the earth polar radius, lat 0 is the radian latitude of a reference point, and the effective radius at the reference point is expressed as: r ns 、r ew is used to describe the radius of curvature of the earth in different directions, affecting the position information calculation: r ns denotes the radius of curvature in the north-south direction, i.e. meridian direction, assisting in displacement calculation in the latitude direction; r ew denotes a radius of curvature in the east-west direction, i.e., in the weft direction, and assists in displacement calculation in the longitudinal direction; Therefore, the longitude and latitude coordinate change can be converted into displacement change in a plane coordinate system in a local coordinate system, lat rad and lon rad are used for representing radian values of the longitude and latitude of a target, N and E are used for representing north displacement and east displacement, and the calculation is as follows: The time stamp after the dimension reduction of the input data is aligned with the longitude and latitude time stamp, and the model can be trained and the characteristics extracted.
5. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, the processor implementing the Trans-GCN based mobile robot dead reckoning method of any of claims 1 to 4 when the program is executed.
6. A non-transitory computer readable storage medium having stored thereon a computer program, which when executed by a processor, implements the Trans-GCN based mobile robot dead reckoning method according to any of claims 1 to 4.

Description

Trans-GCN-based mobile robot dead reckoning method Technical Field The invention belongs to the field of dead reckoning of mobile robots, and particularly relates to a dead reckoning method based on a graph convolution neural network and a transducer structure. Background Under the background of the transition from modern science and technology automation to intelligent, the robot technology is widely applied to various fields as a result model of multi-disciplinary fusion such as mechanical design, sensing technology, electronic information, automatic control theory, artificial intelligence and the like. The roles of the mobile robot in various production and daily life scenes are increasingly highlighted, and the efficiency and the safety of tasks such as power inspection, security search and rescue, material distribution and the like are remarkably improved. However, the core premise of the mobile robot being able to successfully cope with various challenging tasks is its autonomous positioning capability, and autonomous positioning is also a premise of decision making and path planning. The wheel type mobile robot is widely studied as a most widely used mobile robot by means of flexible movement mode, high stability and excellent bearing capacity, and a dead reckoning system based on a wheel speed sensor and an inertial sensor of the wheel type mobile robot is widely studied. The pulse number of the wheel in each unit time is measured by the wheel encoder to calculate the wheel speed and the driving distance, but due to factors such as uneven ground friction, unstable differential speed calculation heading and the like, various accumulated errors can exist in the pure odometer positioning. IMUs (inertial measurement units) typically include two types of sensors, accelerometers and gyroscopes, for measuring and outputting acceleration and angular velocity information of a carrier, and the IMU measurement is affected by multiple noise that is gradually amplified in integration, resulting in a gradual accumulation of attitude angle and displacement errors and an exponential increase. The traditional filtering method depends on accurate system model and noise statistical characteristics, and can not be adapted to the change of a model along with dynamic environment, so that estimation errors are increased, with the continuous progress of big data and artificial intelligence technology, a data driving method based on machine learning is widely applied to solving navigation positioning related problems such as sensor calibration, positioning error divergence suppression and multi-sensor fusion, and the like, mobile robot dead reckoning based on multi-sensor fusion can be regarded as a prediction problem between multiple time sequences and sequences, and is not only limited to time angles, but also the inherent relation and interaction between sensor data can influence the prediction precision. The invention provides a position prediction model Trans-GCN integrating GCN and a Transformer, which utilizes GCN to learn a complex topological structure to capture the space dependence among data, and provides a node embedding method for fusing node position characteristics by using Laplace vectors, wherein graph characteristic signals and coding information are embedded into the Transformer to enhance the global position sensing capability of nodes in a graph, so that the expression capability of the graph structure is further improved, the model has global sensing capability during characteristic extraction, and time sequence data is modeled from multiple scales. Disclosure of Invention The present invention is directed to solving the above problems of the prior art. A mobile robot dead reckoning method based on Trans-GCN is provided. The technical scheme of the invention is as follows: a mobile robot dead reckoning method based on Trans-GCN, comprising the steps of: S1, collecting data of a wheel speed sensor and an inertial sensor of a wheel type mobile robot, intercepting time sequence data through a sliding window and an adaptive sliding step length, constructing sensor data with a single dimension into graph nodes in each window data, and constructing a bidirectional edge connection structure for part of nodes to form complete graph structure data; s2, constructing a graph rolling network GCN to extract the dependency relationship of data in the graph in a non-Euclidean space, wherein the dependency relationship comprises three steps of feature transmission, aggregation and updating to obtain node features and edge features; and S3, decomposing the adjacency matrix of the constructed graph, constructing a position coding matrix of the nodes, integrating the position coding matrix with the linear layer and the L2 pooling layer, and then performing linear transformation on the position coding matrix and the characteristics extracted in the S2, and inputting the position coding matrix into a transducer encoder st