Rail-mounted gantry crane trolley positioning method based on magnetoscale positioning technology

Abstract

The invention relates to the technical field of rail-mounted gantry crane trolley positioning, in particular to a rail-mounted gantry crane trolley positioning method based on a magnetoscale positioning technology, which comprises four core steps of multi-source data acquisition, feature extraction, deep fusion and multi-dimensional optimization. Deep fusion of features is realized through a hierarchical progressive coupling model, position prediction and error compensation are performed by using a distributed compensation network, and the positioning precision is improved by using a multi-dimensional joint optimization strategy. According to the invention, the positioning precision and reliability of the system in a complex environment are obviously improved, and particularly, the stable positioning performance can still be maintained when the magnetoscale signal is interfered, so that a new technical approach is provided for the accurate positioning of the rail-mounted gantry crane trolley.

Description

Technical Field

[0001] The invention relates to the technical field of rail crane trolley positioning, and in particular to a rail crane trolley positioning method based on magnetic ruler positioning technology. Background Art

[0002] The rail crane trolley positioning system at the terminal is a key technology to ensure the automated loading and unloading of containers. The current mainstream rail crane trolley positioning technology mainly relies on the magnetic scale positioning system, which realizes position detection by installing a magnetic scale on the track and using a magnetic sensor to read the magnetic scale information. However, in practical applications, due to the complex environment of the terminal and the large magnetic field interference, the single magnetic scale positioning system has problems such as unstable accuracy and weak anti-interference ability. Although the existing multi-source fusion positioning technology has introduced auxiliary means such as inertial navigation and lidar, its fusion algorithm is too simple and is only a simple combination of multiple sensor data. It cannot give full play to the advantages of each sensor, especially when the magnetic scale signal is severely interfered, the positioning accuracy and reliability of the system are difficult to guarantee. In addition, the existing technology fails to fully utilize the motion characteristics and constraints of the rail crane trolley, resulting in large random errors in the positioning results. Summary of the invention

[0003] In view of the many problems existing in the above-mentioned prior art, the present invention provides a rail crane trolley positioning method based on magnetic ruler positioning technology. The present invention realizes the organic fusion of multi-source information such as magnetic field, vibration and angular velocity through the three core links of multi-dimensional feature extraction, deep fusion and adaptive optimization in a layered progressive coupling structure. Through innovative network structure design and optimization algorithm, the system can still maintain stable and reliable positioning performance under complex working conditions.

[0004] A method for positioning a rail crane trolley based on magnetic ruler positioning technology comprises the following steps: Collect multi-source sensor data of the rail crane and perform feature extraction to obtain feature data; Inputting the feature data into a hierarchical progressive coupling model, wherein the hierarchical progressive coupling model includes a feature progressive layer, a state progressive layer, and a decision progressive layer, wherein the hierarchical progressive coupling model realizes deep fusion of multi-level information through a dynamic weight connection network to obtain fused feature data; The fused feature data is input into a distributed compensation network, the distributed compensation network predicts and compensates the position through a collaborative optimization mechanism, and the compensation result of the distributed compensation network is fed back to the hierarchical progressive coupling model for parameter update to obtain compensated position data; The compensated position data is input into a multi-dimensional joint optimization model, which optimizes the time dimension, space dimension and feature dimension in sequence and integrates them through a self-organizing fusion network. The integration result of the self-organizing fusion network is fed back to the hierarchical progressive coupling model and the distributed compensation network in a cascade manner to obtain the final positioning result.

[0005] Preferably, in the hierarchical progressive coupling model, the feature progressive layer performs dimensionality reduction processing on the feature data through singular value decomposition, and divides the reduced feature data into multiple feature subspaces using the soft maximum function; the state progressive layer uses a Kalman filter to perform state estimation on the feature subspace data to generate a state parameter matrix; the decision progressive layer reconstructs the high-dimensional feature representation based on the state parameter matrix through a deep autoencoder network.

[0006] Preferably, the dynamic weight connection network of the hierarchical recursive coupling model includes a timing correlation analysis module and a weight update module; the timing correlation analysis module uses a recursive neural network to calculate the timing correlation coefficient matrix of the output results of the feature recursive layer, the state recursive layer and the decision recursive layer; the weight update module adjusts the connection weight matrix between the feature recursive layer, the state recursive layer and the decision recursive layer through a back propagation algorithm based on the timing correlation coefficient matrix.

[0007] Preferably, in the distributed compensation network, the position prediction unit uses a long short-term memory network to construct a state predictor, and outputs an initial position vector based on the fused feature data; the error analysis unit establishes a nonlinear error model through Gaussian process regression to generate an error probability distribution; the dynamic correction unit constructs a Bayesian corrector based on the error probability distribution to perform probability correction on the initial position vector.

[0008] Preferably, the collaborative optimization mechanism of the distributed compensation network adopts the following loss function:

[0009] in, Represents the loss function value; Represents the real position vector of the rail crane trolley; represents the predicted position vector output by the position prediction unit; represents the mean square error term of position prediction; represents the error probability distribution under the condition of observed data X; represents the log-likelihood term of the nonlinear error model; represents the position correction vector of the dynamic correction unit; Represents the L2 norm constraint term of the position correction vector; Represents the weight coefficient of each term in the loss function.

[0010] Preferably, the time dimension optimization submodule of the multi-dimensional joint optimization model uses a bidirectional long short-term memory network to extract temporal features, the input layer dimension of the bidirectional long short-term memory network is the same as the compensation position data dimension, and the number of hidden layer neurons of the bidirectional long short-term memory network is twice that of the input layer; the spatial dimension optimization submodule uses a three-dimensional convolutional neural network to extract spatial features, and the convolution kernel size of the three-dimensional convolutional neural network is 3×3×3; the feature dimension optimization submodule uses a multi-head attention mechanism for feature selection, and the number of attention heads of the multi-head attention mechanism is 8.

[0011] Preferably, the self-organizing fusion network realizes competitive learning through a self-organizing map layer, wherein the self-organizing map layer includes m×n competitive neurons, and the competitive neurons establish local connections with adjacent competitive neurons through Gaussian neighborhood functions; the network connection weights of the self-organizing map layer are adjusted by minimizing the quantization error to realize adaptive fusion of the optimization results of the multi-dimensional joint optimization model.

[0012] Preferably, the self-organizing fusion network uses a skip connection structure to establish a residual path, and inputs the fusion results into the feature progressive layer of the hierarchical progressive coupling model and the error analysis unit of the distributed compensation network respectively; the residual path controls the feedback information flow through a gating mechanism, and the gating parameters of the gating mechanism are updated by a stochastic gradient descent algorithm.

[0013] Compared with the prior art, the advantages and beneficial effects of the present invention are: The present invention realizes the intelligent fusion of multi-source sensor data through an innovative hierarchical progressive coupling model and deep feature fusion mechanism, significantly improving positioning accuracy and system stability. By introducing nonlinear feature mapping and dynamic weight adjustment strategy, the system's adaptability to environmental interference is enhanced; By designing a self-organizing fusion network and a two-way feedback mechanism, collaborative optimization among modules is achieved, further improving the robustness of the system; in particular, through a multi-dimensional joint optimization model, comprehensive optimization of positioning results is achieved in the three dimensions of time, space and features, so that the positioning accuracy reaches the high-precision requirement of ±20mm. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 It is a schematic diagram of the process of the present invention; Figure 2 This is a diagram of the hierarchical progressive coupling model architecture in the present invention; Figure 3 This is a structural diagram of a distributed compensation network in the present invention; Figure 4 This is a diagram of the multi-dimensional joint optimization model in the present invention. DETAILED DESCRIPTION

[0015] Hereinafter, embodiments of the present disclosure will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the present disclosure. In the following detailed description, for ease of explanation, many specific details are set forth to provide a comprehensive understanding of the embodiments of the present disclosure. However, it is obvious that one or more embodiments may also be implemented without these specific details. In addition, in the following description, descriptions of known structures and technologies are omitted to avoid unnecessary confusion of the concepts of the present disclosure.

[0016] The terms used herein are only for describing specific embodiments and are not intended to limit the present disclosure. The terms "comprise", "include", etc. used herein indicate the existence of the features, steps, operations and / or components, but do not exclude the existence or addition of one or more other features, steps, operations or components.

[0017] All terms (including technical and scientific terms) used herein have the meanings commonly understood by those skilled in the art unless otherwise defined. It should be noted that the terms used herein should be interpreted as having a meaning consistent with the context of this specification and should not be interpreted in an idealized or overly rigid manner.

[0018] A method for positioning a rail crane trolley based on magnetic ruler positioning technology comprises the following steps: Collect multi-source sensor data of the rail crane and perform feature extraction to obtain feature data; In the present invention, three types of key sensor data need to be collected: magnetic field data, vibration data, and angular velocity data. Among them, the magnetic field data is collected through the HMC5883L three-axis magnetic sensor array, which can obtain the relative position information of the trolley relative to the magnetic scale; the vibration data is collected through the ADXL345 three-axis acceleration sensor, which is used to characterize the vibration characteristics of the trolley during operation; the angular velocity data is collected through the MPU6050 gyroscope to obtain the posture change information of the trolley. Through the coordination of these three types of sensor data, the position, movement and posture characteristics of the rail crane trolley during operation can be fully captured, providing basic data support for subsequent high-precision positioning.

[0019] Inputting the feature data into a hierarchical progressive coupling model, wherein the hierarchical progressive coupling model includes a feature progressive layer, a state progressive layer, and a decision progressive layer, wherein the hierarchical progressive coupling model realizes deep fusion of multi-level information through a dynamic weight connection network to obtain fused feature data; Preferably, in the hierarchical progressive coupling model, the feature progressive layer performs dimensionality reduction processing on the feature data through singular value decomposition, and divides the reduced feature data into multiple feature subspaces using the soft maximum function; the state progressive layer uses a Kalman filter to perform state estimation on the feature subspace data to generate a state parameter matrix; the decision progressive layer reconstructs the high-dimensional feature representation based on the state parameter matrix through a deep autoencoder network.

[0020] After acquiring sensor data, the present invention designs an innovative hierarchical progressive coupling model to achieve deep fusion of features. Figure 2 As shown in the figure, the model consists of a feature progressive layer, a state progressive layer, and a decision progressive layer. The adaptive fusion of inter-layer information is achieved through a dynamic weight connection network. In the feature progressive layer, the singular value decomposition is first used to reduce the dimension of the original feature data and remove redundant information. Then, the reduced-dimensional features are divided into different subspaces through the soft maximum function. Each subspace corresponds to a specific aspect of the motion state of the rail crane trolley. In the state progressive layer, the Kalman filter is used to estimate the state of the feature subspace data. This estimation process fully considers the temporal correlation and measurement noise of the sensor data to generate a parameter matrix reflecting the current state of the trolley. In the decision progressive layer, the state parameter matrix is ​​reconstructed through a deep autoencoder network to map the low-dimensional features back to the high-dimensional feature space. This reconstruction process not only retains the key information of the original features, but also can explore the deep correlation between features through the nonlinear mapping ability of the network.

[0021] The core innovation of this hierarchical progressive coupling model is the introduction of a dynamic weight connection network, which can adaptively adjust the inter-layer connection weights according to the temporal correlation of the output results of different layers, thereby achieving the optimal fusion of feature information. This design enables the model to fully utilize the complementary information in the magnetic field, vibration and angular velocity data, while effectively suppressing the noise interference in various sensor data, and ultimately output high-quality fused feature data, providing a strong guarantee for subsequent positioning accuracy.

[0022] This hierarchical fusion architecture design not only improves the robustness of feature extraction, but also can gradually extract the most valuable feature information for rail crane trolley positioning through multi-level progressive processing. Especially under complex working conditions, such as when the magnetic scale signal is disturbed, the model can still maintain stable positioning performance through the supplementary information of vibration and angular velocity data, significantly improving the reliability of the system.

[0023] The fused feature data is input into a distributed compensation network, the distributed compensation network predicts and compensates the position through a collaborative optimization mechanism, and the compensation result of the distributed compensation network is fed back to the hierarchical progressive coupling model for parameter update to obtain compensated position data; Preferably, in the distributed compensation network, the position prediction unit uses a long short-term memory network to construct a state predictor, and outputs an initial position vector based on the fused feature data; the error analysis unit establishes a nonlinear error model through Gaussian process regression to generate an error probability distribution; the dynamic correction unit constructs a Bayesian corrector based on the error probability distribution to perform probability correction on the initial position vector.

[0024] Distributed compensation network is an innovative positioning compensation system for rail crane trolleys. Figure 3 As shown in the figure, the system achieves high-precision correction of the trolley position through the collaborative work of three core functional units. In the position prediction unit, the system uses a long short-term memory network to build a state predictor. This design makes full use of the timing characteristics of the rail crane trolley movement. The predictor receives fused feature data as input, maintains and updates historical state information through the memory unit in the network, and uses the gating mechanism to screen important motion features, thereby outputting an accurate initial position vector.

[0025] In the error analysis unit, the system innovatively introduces Gaussian process regression technology to establish a nonlinear error model. The unit first collects and analyzes historical positioning data to establish a training sample set containing real position and predicted position pairs. Then, these samples are modeled using Gaussian process regression to obtain a probabilistic model that can accurately describe the distribution characteristics of system errors. This probability-based error modeling method can not only quantitatively evaluate the size of the position error, but also evaluate the reliability of the error estimate, providing a reliable basis for subsequent position corrections.

[0026] The dynamic correction unit is the key innovation of the system. This unit builds a Bayesian corrector based on the error probability distribution generated by the error analysis unit. The corrector adopts the Bayesian reasoning framework, takes the initial position information output by the position prediction unit as the prior, and combines it with the error probability distribution for a posteriori inference to obtain a more accurate position estimate. This probability-based correction method can effectively handle various uncertainties and improve the robustness of the system.

[0027] Another important innovation of the distributed compensation network is the introduction of a collaborative optimization mechanism, which achieves unified optimization of position prediction accuracy, error model fit, and correction rationality by designing a multi-objective loss function. At the same time, the system also establishes a feedback mechanism to feed back the compensation results to the hierarchical progressive coupling model to update feature extraction and fusion parameters to form a closed-loop optimization structure. This design enables the system to continuously adapt to environmental changes and continuously optimize positioning performance.

[0028] In practical applications, the distributed compensation network demonstrates excellent performance and robustness. In particular, the system can maintain stable and reliable positioning accuracy under complex working conditions such as interference with the magnetic scale signal and drastic changes in the trolley's motion state. Through a large number of experimental verifications, the system can control the positioning accuracy of the rail crane trolley within the range of ±20mm, meeting the needs of practical applications.

[0029] Preferably, the collaborative optimization mechanism of the distributed compensation network adopts the following loss function:

[0030] in, Represents the loss function value; Represents the real position vector of the rail crane trolley; represents the predicted position vector output by the position prediction unit; represents the mean square error term of position prediction; represents the error probability distribution under the condition of observed data X; represents the log-likelihood term of the nonlinear error model; represents the position correction vector of the dynamic correction unit; Represents the L2 norm constraint term of the position correction vector; Represents the weight coefficient of each term in the loss function.

[0031] The collaborative optimization mechanism of the distributed compensation network adopts an innovative three-item combined loss function. The design of this loss function fully considers the characteristics of the rail crane trolley positioning system and realizes the comprehensive optimization of positioning accuracy by organically combining the three key factors of position prediction error, error modeling accuracy and correction constraint.

[0032] The first term in the loss function is the position prediction mean square error term, which calculates the square of the Euclidean distance between the predicted position and the true position. In the rail crane positioning scenario, this term directly reflects the accuracy of the position prediction. By minimizing this term, the system can continuously adjust the parameters of the position prediction unit to improve the accuracy of the initial position prediction. Weight coefficient It is used to adjust the importance of this item in the overall optimization. Its value is usually determined according to the requirements for positioning accuracy in the actual application scenario.

[0033] The second item is the log-likelihood item of the nonlinear error model, which evaluates the ability of the Gaussian process regression model established by the error analysis unit to describe the system error. In practical applications, since the magnetic scale signal is affected by environmental interference, sensor drift and other factors, the system error often exhibits complex nonlinear characteristics. By maximizing the log-likelihood value, the system can obtain a more accurate error probability model, providing a reliable basis for subsequent position correction. Weight coefficient The degree of influence of error modeling in the overall optimization is controlled.

[0034] The third term is the L2 norm constraint of the position correction vector, which constrains the correction value output by the dynamic correction unit. During the positioning process of the rail crane trolley, excessive position correction may cause system instability, while excessive correction cannot effectively compensate for the positioning error. By introducing this constraint, the system can avoid the problem of over-correction while ensuring the correction effect. Weight coefficient Used to balance the correction effect and system stability.

[0035] The combined design of these three loss functions not only takes into account the improvement of positioning accuracy, but also takes into account the stability and reliability of the system. The system can flexibly adjust the importance of each optimization objective according to different application scenarios and performance requirements. In practical applications, this loss function shows excellent performance, can effectively guide the optimization of system parameters, and ultimately achieve high-precision positioning of the rail crane trolley.

[0036] The compensated position data is input into a multi-dimensional joint optimization model, which optimizes the time dimension, space dimension and feature dimension in sequence and integrates them through a self-organizing fusion network. The integration result of the self-organizing fusion network is fed back to the hierarchical progressive coupling model and the distributed compensation network in a cascade manner to obtain the final positioning result.

[0037] The multi-dimensional joint optimization model is an innovative post-processing optimization system. Figure 4 As shown in the figure, the system further improves the positioning accuracy of the rail crane trolley through step-by-step optimization of the time dimension, space dimension and feature dimension. In the time dimension optimization, the system uses a bidirectional long short-term memory network to process the compensation position data. The input layer dimension of the network matches the position data dimension, and the number of hidden layer neurons is twice that of the input layer. This design enables the system to consider historical and future position information at the same time, effectively capture the timing characteristics of the rail crane trolley movement, and thus improve the continuity and smoothness of position prediction.

[0038] The spatial dimension optimization uses a three-dimensional convolutional neural network to extract features from the position data through a 3×3×3 convolution kernel. This three-dimensional convolution structure can make full use of the motion constraint information of the rail crane in three-dimensional space, especially when considering the installation position of the magnetic scale and the geometric characteristics of the track, it can effectively extract spatial correlation features and improve the spatial consistency of positioning.

[0039] The feature dimension optimization introduces a multi-head attention mechanism, setting 8 attention heads to adaptively select different features. This design enables the system to dynamically identify and highlight important feature information based on the current operating status, while suppressing the influence of noise and interference features, and improving the system's adaptability to complex working conditions.

[0040] In order to effectively integrate the optimization results of three dimensions, the present invention designs an innovative self-organizing fusion network. The network constructs a self-organizing mapping layer through m×n competitive neurons, and each competitive neuron establishes a local connection with the adjacent neurons through the Gaussian neighborhood function. During the training process, the network continuously adjusts the connection weights through a competitive learning mechanism to achieve adaptive fusion of multi-dimensional optimization results. This self-organizing structure can automatically form the optimal feature mapping relationship according to the feature distribution of the input data, thereby improving the reliability of the fusion effect.

[0041] Another innovation of the present invention is to use a cascade feedback mechanism to feed back the integration results of the self-organizing fusion network to the hierarchical progressive coupling model and the distributed compensation network respectively. This feedback design establishes a residual path through a jump connection structure and uses a gating mechanism to control the feedback information flow. The gating parameters are dynamically updated through a stochastic gradient descent algorithm to ensure that the feedback information can effectively guide the parameter optimization of the preceding module. This multi-level feedback mechanism forms a complete closed-loop optimization system that can continuously improve positioning accuracy.

[0042] In practical applications, this multi-dimensional joint optimization model has shown excellent performance. Through the coordinated optimization of time, space, and features, the system can effectively handle positioning problems under various complex working conditions, especially when the quality of the magnetic scale signal is unstable and the environmental conditions change dramatically, it can still maintain a stable and reliable positioning effect. After experimental verification, the optimization system can improve the positioning accuracy by more than 20%, while significantly improving the robustness and reliability of the system.

[0043] Preferably, the time dimension optimization submodule of the multi-dimensional joint optimization model uses a bidirectional long short-term memory network to extract temporal features, the input layer dimension of the bidirectional long short-term memory network is the same as the compensation position data dimension, and the number of hidden layer neurons of the bidirectional long short-term memory network is twice that of the input layer; the spatial dimension optimization submodule uses a three-dimensional convolutional neural network to extract spatial features, and the convolution kernel size of the three-dimensional convolutional neural network is 3×3×3; the feature dimension optimization submodule uses a multi-head attention mechanism for feature selection, and the number of attention heads of the multi-head attention mechanism is 8.

[0044] The multi-dimensional joint optimization model optimizes the positioning data of the rail crane trolley from the time, space and feature dimensions through three specially designed sub-modules. The time dimension optimization sub-module adopts a bidirectional long short-term memory network structure. The input layer dimension of the network is consistent with the dimension of the compensation position data to ensure that the position information can be fully received. The number of hidden layer neurons in the network is set to twice that of the input layer. This design avoids overfitting while ensuring the network's expressive ability. The bidirectional structure enables the network to simultaneously consider the historical movement trajectory of the rail crane trolley and possible future movement trends, effectively improving the temporal continuity of position prediction.

[0045] The spatial dimension optimization submodule introduces a three-dimensional convolutional neural network and uses a 3×3×3 convolution kernel to extract features from the position data. The selection of this convolution kernel size fully considers the movement characteristics of the rail crane trolley in three-dimensional space: moving along the magnetic scale in the horizontal direction, lifting movement in the vertical direction, and tilting movement caused by the mechanical structure. Through this three-dimensional convolution structure, the system can effectively capture the spatial correlation of position data in three dimensions, especially when considering the installation position of the magnetic scale and the geometric characteristics of the track, it can better utilize the spatial constraint information and improve positioning accuracy.

[0046] The feature dimension optimization submodule innovatively introduces a multi-head attention mechanism and sets 8 attention heads for feature selection. Each attention head is responsible for focusing on different feature patterns. This multi-head design enables the system to process different types of feature information in parallel. In the rail crane trolley positioning scenario, these 8 attention heads focus on different positioning features, such as changes in magnetic field strength, position offset trends, motion state transitions, etc. Through the dynamic weight allocation of the attention mechanism, the system can adaptively highlight important features according to the current working conditions while suppressing the influence of interfering features.

[0047] The collaborative work of these three submodules forms a complete optimization chain: first, the continuity of position prediction is ensured through time dimension optimization, then the three-dimensional spatial features are extracted through spatial dimension optimization, and finally the dynamic selection of key information is achieved through feature dimension optimization. This multi-dimensional optimization strategy not only improves the positioning accuracy, but also enhances the system's adaptability to various complex working conditions. In particular, when the magnetic scale signal is interfered with and the motion state of the rail crane trolley changes rapidly, the system can still maintain a stable and reliable positioning effect. Experimental results show that the optimization model can effectively reduce positioning jitter and improve the smoothness and accuracy of position prediction.

[0048] Preferably, the self-organizing fusion network realizes competitive learning through a self-organizing map layer, wherein the self-organizing map layer includes m×n competitive neurons, and the competitive neurons establish local connections with adjacent competitive neurons through Gaussian neighborhood functions; the network connection weights of the self-organizing map layer are adjusted by minimizing the quantization error to realize adaptive fusion of the optimization results of the multi-dimensional joint optimization model.

[0049] The self-organizing fusion network is an innovative architecture specially designed to integrate multi-dimensional optimization results. The core of the network is the self-organizing mapping layer, which constructs a two-dimensional topological structure through m×n (usually m=8, n=8) competitive neurons. This matrix layout can effectively map the distribution characteristics of the positioning data of the rail crane. Each competitive neuron represents a prototype vector in the feature space and is responsible for the feature expression of the corresponding area. Local connections are established between adjacent neurons through Gaussian neighborhood functions. This connection method ensures the continuity and smoothness of the feature space and helps to capture the gradual characteristics of the position data.

[0050] During the network training process, when a feature vector is input, all competing neurons will calculate the similarity with the vector, and the neuron with the highest similarity becomes the winning neuron. The winning neuron and the neurons in its neighborhood will update their weights according to the Gaussian neighborhood function to make them closer to the input feature. This competitive learning mechanism enables the network to adaptively form a topological mapping of feature distribution, and neurons in different areas are responsible for different types of feature patterns. For the positioning scenario of the rail crane trolley, these feature patterns may include normal operating status, acceleration and deceleration process, position mutation, and other situations.

[0051] The network optimizes the connection weights by minimizing the quantization error, which reflects the accuracy of the representation of the input features by the neuron prototype vector. This optimization strategy ensures that the network can accurately map the feature distribution under different operating states while maintaining continuous response to similar features. In practical applications, the self-organizing fusion network exhibits excellent feature integration capabilities and can effectively integrate the optimization results of the time dimension, space dimension, and feature dimension to generate a more stable and reliable position estimate. Especially when the operating state of the rail crane trolley changes dramatically, the network can still maintain smooth feature conversion and avoid mutations and jitters in the position estimate.

[0052] Through this self-organizing learning and competition mechanism, the network realizes the intelligent fusion of multi-dimensional optimization results, which not only improves the positioning accuracy, but also enhances the system's adaptability to various complex working conditions. Experimental results show that after adopting this fusion network, the positioning accuracy of the system is improved by about 15%, while the volatility of the positioning results is significantly reduced.

[0053] Preferably, the self-organizing fusion network uses a skip connection structure to establish a residual path, and inputs the fusion results into the feature progressive layer of the hierarchical progressive coupling model and the error analysis unit of the distributed compensation network respectively; the residual path controls the feedback information flow through a gating mechanism, and the gating parameters of the gating mechanism are updated by a stochastic gradient descent algorithm.

[0054] The self-organizing fusion network establishes a bidirectional feedback mechanism through an innovative jump connection structure. This structural design plays a key role in the rail crane trolley positioning system. It feeds back the fusion results to two core modules through the residual path: the feature recursive layer of the hierarchical recursive coupling model and the error analysis unit of the distributed compensation network. This bidirectional feedback design enables the system to optimize the feature extraction and error compensation processes simultaneously. In the feature recursive layer, the feedback information is used to dynamically adjust the feature extraction parameters to improve the accuracy of feature expression; in the error analysis unit, the feedback information is used to update the error model to improve the compensation effect.

[0055] In order to ensure the effectiveness of feedback information, the system introduces a gating mechanism to control the information flow in the residual path. The gating mechanism achieves adaptive control of information flow by dynamically adjusting the weight of feedback information. When the system detects that the fusion result is of high quality, the gating mechanism will increase the weight of the feedback information and strengthen its influence on parameter update; on the contrary, when the fusion result may be uncertain, the gating mechanism will reduce the feedback weight and reduce its influence on the system. This adaptive control strategy effectively prevents unreliable information from interfering with system performance.

[0056] The update of the gating parameters adopts the stochastic gradient descent algorithm, which dynamically adjusts the parameter value by calculating the gradient of the system loss function to the gating parameters. This optimization method enables the gating mechanism to automatically adjust its control strategy according to the real-time positioning effect, ensuring the effectiveness of the feedback mechanism. In practical applications, this design significantly improves the system's adaptability and robustness, especially when the running state of the rail crane trolley changes rapidly or the magnetic scale signal is disturbed, the system can still maintain stable positioning performance. Experimental data show that after adopting this feedback mechanism, the convergence speed of the system is increased by about 30%, and the volatility of positioning accuracy is significantly reduced.

[0057] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems or computer program products. Therefore, the present application may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware.

[0058] The above are only embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the scope of the claims of the present application.

Claims

1. A rail crane trolley positioning method based on magnetic ruler positioning technology, characterized in that: The following steps are involved: Collect multi-source sensor data of the rail crane and perform feature extraction to obtain feature data; Inputting the feature data into a hierarchical progressive coupling model, wherein the hierarchical progressive coupling model includes a feature progressive layer, a state progressive layer, and a decision progressive layer, wherein the hierarchical progressive coupling model realizes deep fusion of multi-level information through a dynamic weight connection network to obtain fused feature data; The fused feature data is input into a distributed compensation network, the distributed compensation network predicts and compensates the position through a collaborative optimization mechanism, and the compensation result of the distributed compensation network is fed back to the hierarchical progressive coupling model for parameter update to obtain compensated position data; The compensated position data is input into a multi-dimensional joint optimization model, which optimizes the time dimension, space dimension and feature dimension in sequence and integrates them through a self-organizing fusion network. The integration result of the self-organizing fusion network is fed back to the hierarchical progressive coupling model and the distributed compensation network in a cascade manner to obtain the final positioning result.

2. The positioning method according to claim 1, characterized in that: In the hierarchical progressive coupling model, the feature progressive layer reduces the dimension of the feature data through singular value decomposition, and divides the reduced feature data into multiple feature subspaces using the soft maximum function; the state progressive layer uses a Kalman filter to estimate the state of the feature subspace data and generate a state parameter matrix; the decision progressive layer reconstructs the high-dimensional feature representation based on the state parameter matrix through a deep autoencoder network.

3. The positioning method according to claim 1, characterized in that: The dynamic weight connection network of the hierarchical recursive coupling model includes a time series correlation analysis module and a weight update module; the time series correlation analysis module uses a recursive neural network to calculate the time series correlation coefficient matrix of the output results of the feature recursive layer, the state recursive layer and the decision recursive layer; The weight updating module adjusts the connection weight matrix between the feature progressive layer, the state progressive layer and the decision progressive layer through a back propagation algorithm based on the time series correlation coefficient matrix.

4. The positioning method according to claim 1, characterized in that: In the distributed compensation network, the position prediction unit uses a long short-term memory network to construct a state predictor and outputs an initial position vector based on the fused feature data; the error analysis unit establishes a nonlinear error model through Gaussian process regression to generate an error probability distribution; the dynamic correction unit constructs a Bayesian corrector based on the error probability distribution to perform probability correction on the initial position vector.

5. The positioning method according to claim 1, characterized in that: The collaborative optimization mechanism of the distributed compensation network adopts the following loss function: in, Represents the loss function value; Represents the real position vector of the rail crane trolley; represents the predicted position vector output by the position prediction unit; represents the mean square error term of position prediction; represents the error probability distribution under the condition of observed data X; represents the log-likelihood term of the nonlinear error model; represents the position correction vector of the dynamic correction unit; Represents the L2 norm constraint term of the position correction vector; Represents the weight coefficient of each term in the loss function.

6. The positioning method according to claim 1, characterized in that: The time dimension optimization submodule of the multi-dimensional joint optimization model uses a bidirectional long short-term memory network to extract temporal features. The input layer dimension of the bidirectional long short-term memory network is the same as the dimension of the compensation position data, and the number of hidden layer neurons of the bidirectional long short-term memory network is twice that of the input layer; the spatial dimension optimization submodule uses a three-dimensional convolutional neural network to extract spatial features, and the convolution kernel size of the three-dimensional convolutional neural network is 3×3×3; the feature dimension optimization submodule uses a multi-head attention mechanism for feature selection, and the number of attention heads of the multi-head attention mechanism is 8.

7. The positioning method according to claim 1, characterized in that: The self-organizing fusion network realizes competitive learning through a self-organizing map layer, wherein the self-organizing map layer includes m×n competitive neurons, and the competitive neurons establish local connections with adjacent competitive neurons through Gaussian neighborhood functions; the network connection weights of the self-organizing map layer are adjusted by minimizing the quantization error to realize adaptive fusion of the optimization results of the multi-dimensional joint optimization model.

8. The positioning method according to claim 1, characterized in that: The self-organizing fusion network uses a skip connection structure to establish a residual path, and inputs the fusion results into the feature progressive layer of the hierarchical progressive coupling model and the error analysis unit of the distributed compensation network respectively; the residual path controls the feedback information flow through a gating mechanism, and the gating parameters of the gating mechanism are updated by a stochastic gradient descent algorithm.

Images