Five-phase permanent magnet speed regulation control method and system for monorail crane

Abstract

The application discloses a five-phase permanent magnet speed regulation control method and system for a monorail crane, belongs to the technical field of monorail cranes, and collects multi-source operation state information of a five-phase motor to construct a multi-axis state topology graph based on a graph neural network; on the basis, a graph time sequence prediction model is established to predict a future disturbance trend and generate state prior information; optimal control strategy selection and priority adaptive switching are performed according to the prior information; control vector redundancy reconstruction is performed under a fault condition; and a generalized hyper-surface regulator is embedded in a sliding mode fault-tolerant controller, and the control gain is dynamically corrected in combination with a coupling strength coefficient; the method has the advantages of strong adaptability, high fault tolerance and high speed regulation precision, and is suitable for a permanent magnet motor control system under complex working conditions such as a coal mine monorail crane.

Description

Technical Field

[0001] This invention relates to the field of monorail technology, specifically to a five-phase permanent magnet speed control method and system for monorails. Background Technology

[0002] Currently, monorail systems are widely used in underground coal mine transportation, and their operating efficiency and equipment safety heavily depend on the accuracy of drive control. Traditional three-phase asynchronous motor drive control generally suffers from problems such as large torque fluctuations, slow speed regulation, and unstable start-stop, making it difficult to meet the requirements of complex underground working conditions such as heavy loads, slopes, and frequent start-stops.

[0003] Five-phase permanent magnet synchronous motors are gradually being used in motor drive systems due to their excellent torque density and multiple redundant degrees of freedom. However, most existing control methods are based on extensions of three-phase control systems and lack accurate modeling of the essential structure of five-phase multi-vector space. Especially under nonlinear conditions such as single-phase or two-phase faults or sudden changes in electromagnetic parameters, the control system is difficult to maintain stable operation.

[0004] Furthermore, current control strategies are mostly based on fixed-parameter PI control or basic vector control algorithms, which have poor adaptability to external disturbances, model uncertainties and internal nonlinear coupling. Especially in complex downhole scenarios, they cannot achieve both high response and high stability and high fault tolerance.

[0005] Therefore, there is an urgent need for a new speed regulation method that combines the multi-dimensional spatial decoupling capability of a five-phase motor with multi-model switching control, adaptive neural network prediction compensation and fault reconstruction control strategies to solve the problems of slow response, poor fault tolerance and low speed regulation accuracy in the existing technology, and realize the high-performance operation of the monorail system under complex working conditions. Summary of the Invention

[0006] The purpose of this invention is to provide a five-phase permanent magnet speed control method and system for monorail cranes, so as to solve the shortcomings in the prior art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a five-phase permanent magnet speed control method for monorail cranes, comprising:

[0008] S100: Collect stator current, voltage, speed and temperature rise signals of a five-phase permanent magnet synchronous motor, and construct a multi-axis state topology diagram based on a graph neural network;

[0009] S200. Based on the multi-axis state topology diagram, a multi-dimensional state prediction model for the motor is constructed, and a graph convolution mechanism is used to predict the torque disturbance trend in the future Δt time period, generating state prior information W.

[0010] S300. Based on the prior state information W, select the matching optimal control strategy set, including two types of nonlinear speed controllers: model predictive controller and sliding mode fault-tolerant compensation controller, and adaptively switch between them based on the priority switching algorithm during operation.

[0011] S400: When an abnormal operation of one or more phases is detected, the drive control vector is reconstructed based on the redundancy distribution to form a heterogeneous control state.

[0012] The S500 incorporates an improved generalized hypersurface regulator in the sliding mode fault-tolerant compensation controller, which adjusts the control gain in real time based on the coupling strength coefficient of the GNN output.

[0013] Preferably, the construction of the multi-axis state topology graph based on the graph neural network specifically includes:

[0014] The stator current, back electromotive force and phase angle data of each phase of the five-phase motor are collected, and the data are processed by sliding window segmentation and multidimensional normalization to form a time series state tensor.

[0015] A weighted adjacency matrix is ​​constructed based on the interphase flux coupling degree and the historical operating frequency to generate a dynamic graph structure representing the multi-axis electromagnetic structure;

[0016] We introduce a graph convolution module with residuals from graph neural networks to propagate features of the graph structure, and construct a state topology graph that integrates temporal and spatial characteristics by combining the current change trend between nodes.

[0017] By assigning high weights to key phase nodes through a graph attention mechanism, a prior state graph is output for speed regulation decisions.

[0018] Preferably, the construction of the multidimensional state prediction model of the motor and the generation of state prior information W includes:

[0019] Using the generated state topology graph as input, a graph time series prediction network that integrates spatial graph structure and time series features is established, wherein the state embedding of each node includes current derivative, electric angular frequency perturbation and historical mutation response factor;

[0020] By jointly modeling stacked graph convolutional networks and gated recurrent units, the state propagation process of graph structures within the future Δt prediction period is realized, capturing the evolution trajectory of perturbation trends between multiple axes;

[0021] Cluster analysis is performed on the torque change direction, amplitude and correlation in the predicted node features to extract the disturbance prior features and generate prior information W containing spatial disturbance hotspots and trend directions.

[0022] Preferably, the adaptive switching based on the priority switching algorithm includes:

[0023] The perturbation features in the prior state information W are mapped to the controller state set, and a dynamic weight matrix is ​​generated accordingly to evaluate the adaptability score of each type of controller in the current prediction environment.

[0024] A hierarchical priority selection mechanism is constructed, and different control strategies are ranked by a multi-factor fuzzy evaluation function. The ranking criteria include response delay, prediction offset and disturbance matching rate.

[0025] Temporary weight reduction restrictions are implemented for controllers with continuous prediction deviations exceeding limits, thereby realizing the optimal controller invocation strategy based on dynamic adjustment of the disturbance environment.

[0026] Preferably, the reconstructed drive control vector based on redundancy distribution includes:

[0027] The graph attention mechanism is used to detect abnormal current patterns in a five-phase system, identify faulty phase nodes, and perform dynamic attribution analysis on their output amplitude change rate.

[0028] Based on the spatial redundancy distribution characteristics of the five-phase structure, after removing the weight of the faulty phase, the remaining four phases are non-uniformly redistributed in amplitude, and a phase rotation transformation is introduced to maintain the electromagnetic torque balance of the system.

[0029] The reconstructed result is fused with the original controller output vector to generate a heterogeneous control state, which contains composite drive signals with unequal amplitude and unequal phase.

[0030] Preferably, the embedding of the improved generalized hypersurface adjuster in the sliding mode fault-tolerant compensation controller includes:

[0031] A generalized sliding mode hypersurface expression containing nonlinear dynamic boundary terms and current-coupled disturbance terms is constructed, and a residual driving factor is introduced to adjust the control target surface;

[0032] The node coupling strength coefficients output by the graph neural network are injected as dynamic weights into the hypersurface regulator to characterize the response coupling nonuniformity of the five-phase system between different nodes.

[0033] Nonlinear correction of sliding mode gain based on coupling weights enables adaptive adjustment of control strength for multiphase systems under sudden disturbances and fault conditions.

[0034] This invention also provides a five-phase permanent magnet speed control system for monorail cranes, comprising:

[0035] The state perception and modeling module collects stator current, voltage, speed and temperature rise signals of a five-phase permanent magnet synchronous motor and constructs a multi-axis state topology graph based on a graph neural network.

[0036] The disturbance prediction and prior analysis module constructs a multi-dimensional state prediction model for the motor based on the multi-axis state topology diagram, and uses a graph convolution mechanism to predict the torque disturbance trend in the future Δt time period, generating state prior information W.

[0037] The adaptive control strategy selection module selects the optimal set of control strategies based on the prior state information W, including two types of nonlinear speed controllers: model predictive controllers and sliding mode fault-tolerant compensation controllers, and adaptively switches between them based on a priority switching algorithm during operation.

[0038] The fault-tolerant reconfiguration control module, when detecting an abnormal operation of one or more phases, reconfigures the drive control vector based on redundant distribution to form a heterogeneous control state.

[0039] The sliding mode compensation and gain adjustment module embeds an improved generalized hypersurface regulator in the sliding mode fault-tolerant compensation controller, which adjusts the control gain in real time according to the coupling strength coefficient of the GNN output.

[0040] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0041] 1. This invention, by introducing graph neural network modeling technology and a disturbance prediction mechanism, breaks through the limitations of traditional speed control methods in modeling three-phase motors. For the first time, it constructs a state topology graph that integrates multi-axis current coordination relationships and temporal variation characteristics, enabling forward-looking prediction of future disturbance trends. This design allows the controller to obtain prior state information before selecting a control strategy, thereby significantly improving the responsiveness and decision-making accuracy of the speed control system in complex dynamic environments.

[0042] 2. This invention innovatively introduces a redundant control vector reconstruction mechanism and a sliding mode gain adjustment method based on adaptive coupling strength. This not only achieves fault-tolerant drive control in multiphase systems but also effectively avoids the stability failure problem of traditional controllers under multiphase mismatch conditions through regional coupling modeling. The overall solution possesses advantages such as high adaptability, high robustness, and strong fault tolerance, making it particularly suitable for permanent magnet speed control applications in high-dynamic, high-redundancy, and high-reliability scenarios such as underground monorail cranes. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0044] Figure 1 This is a mind map of the method of the present invention.

[0045] Figure 2This is a mind map of the system modules of the present invention. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0047] Example 1, please refer to Figure 1 As shown in this embodiment, a five-phase permanent magnet speed control method for monorail cranes includes: acquiring stator current, voltage, speed, and temperature rise signals of a five-phase permanent magnet synchronous motor, and constructing a multi-axis state topology diagram based on a graph neural network; based on the multi-axis state topology diagram, constructing a multi-dimensional state prediction model of the motor, and using a graph convolution mechanism to predict the torque disturbance trend in the future time period Δt, generating state prior information W; selecting a matching optimal control strategy set based on the state prior information W, including two types of nonlinear speed controllers: model predictive controllers and sliding mode fault-tolerant compensation controllers, and adaptively switching them based on a priority switching algorithm during operation; when an abnormal operation of one or more phases is detected, reconstructing the drive control vector based on redundant distribution to form a heterogeneous control state; embedding an improved generalized hypersurface regulator in the sliding mode fault-tolerant compensation controller, and adjusting the control gain in real time according to the coupling strength coefficient output by the GNN.

[0048] Specifically, constructing a multi-axis state topology graph based on a graph neural network includes:

[0049] The stator current, back electromotive force and phase angle data of each phase of the five-phase motor are collected, and the data are processed by sliding window segmentation and multidimensional normalization to form a time series state tensor.

[0050] Specifically, the five stator currents Ia, Ib, Ic, Id, Ie, back electromotive force Ea, Eb, ..., Ee, and the electrical angles θa, ..., θe of each phase winding are acquired in real time through the high sampling rate data acquisition module.

[0051] Using the control cycle as the unit, a fixed-length sliding time window is set (e.g., a 20ms window, sliding once every 5ms). Each type of signal acquired above is segmented in the time dimension, so that each signal forms an n×t subsequence structure, where n is the number of signal channels and t is the number of time steps.

[0052] z-score normalization (mean 0, standard deviation 1) was used to independently normalize the three physical quantities of current, back electromotive force and angle, respectively, in order to eliminate the influence of different physical quantities and enhance the model’s sensitivity to input changes and convergence.

[0053] The segmented normalized signals are concatenated into a three-dimensional tensor T∈Rc×t×f, where c is the sampling channel (current, voltage, angle) and f is the phase axis channel, realizing a structured expression of the space-time-variable dimensions, and providing a temporal-spatial parallel modeling foundation for subsequent graph neural network input.

[0054] A weighted adjacency matrix is ​​constructed based on the interphase flux coupling degree and the historical operating frequency to generate a dynamic graph structure representing the multi-axis electromagnetic structure;

[0055] Based on the structural parameters of the five-phase motor and online sampling data, the magnetic flux coupling degree Mij between any two phase windings is calculated in real time. This value is obtained through the mutual inductance formula and the relative current change rate, reflecting the strength of electromagnetic coupling between phases in space.

[0056] Dynamic Time Warping (DTW) is performed on the current sequences of each phase during the historical operating cycle, and the frequency of similar fluctuations in time sequence is statistically analyzed as a time correlation index of the coordinated changes between phases.

[0057] The two weighting factors, flux linkage coupling degree Mij and cooperative frequency Fij, are integrated using a normalized fusion formula. (where 0 < α < 1), construct the weighted adjacency matrix A ∈ R5×5 between the five phase nodes;

[0058] The adjacency matrix A is dynamically refreshed in each control cycle, and the edge connection state is determined by the weight change trend of each edge in the graph, forming a dynamic graph structure in time series, which is used to characterize the coupling characteristics of the multi-axis system as time changes.

[0059] We introduce a graph convolution module with residuals from graph neural networks to propagate features of the graph structure, and construct a state topology graph that integrates temporal and spatial characteristics by combining the current change trend between nodes.

[0060] Graph convolution model initialization: Based on the weighted adjacency matrix A and state tensor T constructed above, a graph neural network model is constructed, where each node represents a phase winding and each edge weight represents electromagnetic coupling strength and cooperative trend; initialize graph convolutional layers with residual connections to prevent gradient vanishing or overfitting caused by multi-layer propagation.

[0061] In each layer of the graph neural network, an improved graph convolution operator is employed. ,in For the features of the l-th layer nodes, σ is a trainable weight matrix, and σ is an activation function, enabling multi-level extraction of spatially coupled features.

[0062] Current mutation feature fusion: The rate of change of current at each node in the time series is statistically analyzed, the higher-order derivative (jerk) feature is extracted, the mutation relationship between nodes is identified, and the attention weight of the graph convolutional layer is adjusted by introducing a weight perturbation function to achieve a sensitive response to the current mutation relationship;

[0063] State topology graph generation: The convolutional node features are fused with the current change trend to form a graph embedding representation Z∈Rn×d that simultaneously reflects the coupling of the motor structure and the dynamic fluctuations. This constructs a state topology graph that integrates spatiotemporal information, which is used as the input judgment basis for subsequent speed control strategies.

[0064] By assigning high weights to key phase nodes through a graph attention mechanism, a prior state graph for speed regulation decision-making is output.

[0065] Based on the node feature representation of the graph convolution output, a graph attention mechanism is introduced to assign trainable attention weight coefficients αij to each node, which are used to characterize the response sensitivity of node i to neighboring node j.

[0066] By calculating the rate of change of state, frequency of historical anomalies and magnitude of sudden changes of each node in multiple time windows, “key phase nodes” with high current fluctuations or frequent participation in coordinated changes are identified as candidate sets of key points for speed regulation.

[0067] Within each control cycle, the attention coefficient αij is updated in real time based on the mutation trend of key nodes and the response intensity of their neighboring nodes, thereby increasing the semantic expression weight of key nodes in the graph structure and suppressing the information diffusion of unimportant nodes.

[0068] The full-graph embedding result after fusion node attention weighting generates a state graph Zattn, which serves as a priori input for fusing spatial coupling, temporal abrupt changes, and key location information, and is used as the basis for real-time speed adjustment judgment by the controller decision module.

[0069] The construction of the multidimensional state prediction model for the motor and the generation of prior state information W include:

[0070] Graph-based time-series prediction network construction: This invention first uses the state topology graph generated in the previous control cycle as the input data basis. This topology graph integrates the spatial coupling relationships between each phase winding and the node state evolution information. Based on this, a graph-based time-series prediction network that integrates spatial graph structure and time series characteristics is constructed. Specifically, the state embedding vector of each node includes, but is not limited to: the derivative change of the node current in the current cycle, the disturbance intensity index of the electrical angular frequency, and the response amplitude factor of abrupt events in historical cycles, thereby taking into account both the internal physical fluctuations and dynamic response behavior of the motor in state modeling.

[0071] Graph Convolution-GRU Joint Modeling: To enhance the system's ability to predict future disturbance trends, this invention further employs a stacked graph convolutional network and gated recurrent units (GRUs) for joint modeling. The graph convolutional module handles the spatial coupling propagation between nodes, while the GRU module captures the evolutionary patterns over time. This joint model effectively simulates the state propagation process of the recursive graph structure within the future Δt prediction period, mimicking the evolution trajectory of disturbance trends between phases of a five-phase motor and enhancing the system's forward-looking identification capability for nonlinear fluctuations.

[0072] Disturbance Migration Detection and Prior Map Generation: Based on the graph prediction output described above, this invention introduces a disturbance migration detection module to identify the potential torque change direction, disturbance amplitude, and correlation within the overall graph structure of each node in future cycles. Through feature clustering analysis and a disturbance focusing mechanism, spatial disturbance hotspots and trend evolution paths are extracted, ultimately forming state prior information W for subsequent speed control strategy invocation. This information not only possesses quantitative prediction capabilities for future disturbances but also provides a structured representation of disturbance intensity distribution and node risk levels.

[0073] The adaptive handover based on the priority switching algorithm includes:

[0074] Each disturbance feature in the prior state information W is mapped to the controller state set, and a dynamic weight matrix is ​​generated accordingly to evaluate the adaptability of each control strategy in the current disturbance prediction environment.

[0075] Specifically, the disturbance characteristics include parameters such as disturbance direction vector, disturbance intensity amplitude, node activity distribution, and disturbance duration prediction, which are respectively mapped to the set of indicators related to response rate, stability robustness, and regulation capability in the controller state dimension;

[0076] By using the mapping function between disturbance parameters and controller characteristic indicators, a dynamic adaptive weight matrix is ​​calculated. The matrix values ​​reflect the controller's response matching degree and adjustment capability in the current disturbance environment, providing a quantitative basis for subsequent control strategy selection.

[0077] Construct a control strategy priority ranking mechanism based on a multi-factor fuzzy evaluation function to realize dynamic scheduling and switching among multiple strategies;

[0078] Specifically, by using the dynamic weight matrix and disturbance type features, the fuzzy weight score of each type of controller is calculated under multiple evaluation dimensions such as "response delay", "disturbance fitting degree" and "historical error convergence speed". A priority sorting index is constructed using fuzzy inference rules to form a dynamic priority controller sequence.

[0079] This priority ranking mechanism can maintain the stability and sensitivity of the ranking strategy when the relationship between disturbance type and controller performance changes nonlinearly, effectively improving the intelligence and real-time performance of control switching decisions.

[0080] To prevent system oscillations caused by frequent controller switching, this invention introduces a penalty mechanism based on error feedback to temporarily reduce the weight of controllers that are not adapted to the disturbance environment.

[0081] Specifically, the output error of each control strategy under the current disturbance prediction period is monitored in real time. If prediction deviation exceeds the limit continuously, a penalty mechanism is triggered, the weight of the corresponding controller in the priority matrix is ​​reduced, a minimum cooling period is set, and it is prohibited from participating in the next round of priority competition.

[0082] This mechanism improves the accuracy and robustness of the system speed control by limiting the frequency of repeated calls to high-error strategies, ensuring that the control strategy switching path maintains global convergence and stage stability under dynamic disturbance conditions.

[0083] Reconstructing the driving control vector based on redundancy distribution specifically includes:

[0084] When an operational abnormality is detected in one or more phase windings, the system initiates a vector reconfiguration mechanism based on the redundancy distribution characteristics of the five-phase drive structure to ensure that the motor drive system can still operate stably in a non-perfectly symmetrical state.

[0085] Specifically, the graph attention mechanism is used to focus on the current fluctuation characteristics between nodes, and the abnormal phase nodes are identified by combining the output trend and disturbance direction change of each node in the prior information W of the predicted state.

[0086] For the identified faulty phase nodes, the rate of change of current amplitude and historical fluctuation characteristics are further calculated to achieve dynamic attribution of current anomalies and accurately distinguish them into transient disturbances, contact faults or permanent damage types, so as to provide a basis for subsequent control strategy selection.

[0087] Based on the fault identification results, the system activates the redundancy vector reconstruction module to perform topology adjustment on the five-dimensional driving vector output by the controller;

[0088] Specifically, the control vector components of the marked faulty phases are removed, and non-uniform amplitude redistribution is implemented in the remaining four phases according to the difference between the driving load direction and the current amplitude. At the same time, a phase rotation matrix transformation algorithm is introduced to correct the phase angle offset in order to maintain the stability of the direction of the synthesized electromagnetic torque vector and avoid output torque imbalance.

[0089] Finally, the reconstructed four-phase drive vector is fused with the reference signal of the original controller to form a heterogeneous control state vector, which has a composite structure of non-equal amplitude and non-equal phase.

[0090] This control signal is suitable for motor drive control under fault-tolerant conditions. Its structural features can automatically reconfigure the output mode according to the motor state, maintain the torque continuity and dynamic balance of the overall system when the motor is running asymmetrically, and improve the system's operational stability and reliability under fault conditions.

[0091] To enhance the robustness and adaptability of sliding mode fault-tolerant control strategies under complex disturbance conditions, this invention embeds an improved generalized hypersurface regulator in the sliding mode compensation controller to achieve adaptive dynamic adjustment of the control gain under coupled heterogeneous states in multiphase systems.

[0092] First, a generalized sliding mode hypersurface expression is constructed, incorporating nonlinear dynamic boundary terms and current-coupled perturbation terms, to... The control target surface is established in the form of Ψ(t), where Ψ(t) represents the disturbance function, γ is the adaptive disturbance gain coefficient, and S is the sliding mode hypersurface function, representing the dynamic deviation trajectory between the current system state and the target control state. The controller aims to make S→0 to ensure that the system state converges to the desired trajectory or steady state. x is the derivative of the system state variable, usually representing the instantaneous rate of change of the controlled object (such as motor current, speed, or torque); it is used to describe the dynamic response trend of the system. λx represents the linear regulation term of the state variable.

[0093] In addition, a "residual driving factor" based on the evolution trend of disturbance residuals is introduced to dynamically adjust the target range of the sliding surface, so that the system has a certain degree of soft boundary control capability and is suitable for robust adjustment under inaccurate modeling or dynamic fault disturbance environment.

[0094] The present invention further incorporates the node coupling strength coefficients output by the graph neural network as a dynamic adjustment input into the sliding mode hypersurface expression;

[0095] The coupling strength coefficient is the response correlation factor between adjacent nodes learned by the graph convolutional neural network during the state topology propagation process, which reflects the strength of the current disturbance cooperation between each phase under different time windows.

[0096] By introducing corresponding coupling weight factors into the sliding surface expression for each node, the differential adjustment of the control gain distribution for different phase axes is achieved, enabling the sliding mode controller to have regional response adaptability and improving the fault tolerance of control behavior to structural asymmetry.

[0097] Based on the above coupling weights, the present invention uses a nonlinear adjustment function to correct the gain of the sliding mode controller in real time. The gain correction function can automatically adjust its growth rate and cutoff threshold according to the disturbance density, node response delay and prediction residual.

[0098] This dynamic gain adjustment mechanism ensures that the sliding mode controller automatically focuses on the control energy distribution area under different node response strength changes, realizing high responsiveness and high stability adjustment of multiphase systems to sudden disturbances and occasional fault conditions, and improving the overall flexibility and robustness of the control system.

[0099] Example 2: To verify the feasibility and effectiveness of the present invention in five-phase permanent magnet speed control for monorail cranes, an example of a five-phase permanent magnet synchronous motor is used, including:

[0100] A five-phase permanent magnet synchronous motor of model PM-5P-1200 is used as the drive source, and a high-precision current sensor and speed encoder are installed to sample the current, voltage and speed of each phase in real time at 1kHz; a temperature sensor is used to collect the temperature rise of the motor stator or winding for status early warning; the main controller adopts an edge computing platform (such as an ARM Cortex-A53+FPGA SoC), and simultaneously deploys a graph neural network inference module and a sliding mode control algorithm.

[0101] State topology graph construction: Every 20ms, a time window containing 4 steps (5ms / step) is constructed;

[0102] The phase currents, back electromotive force, and phase angles are spliced ​​together to form a tensor with the shape of [3 channels × 4 steps × 5 phases].

[0103] Calculate the magnetic flux coupling between each phase online (the formula is based on the mutual inductance coefficient and the current derivative).

[0104] The DTW method is used to statistically analyze the phase current timing synchronization rate and then integrates the coupling degree value to generate a normalized weight matrix A.

[0105] A three-layer residual graph convolutional network is used to process the adjacency graph A and the tensor input, integrating high-order current jerk loss sensing.

[0106] The graph attention mechanism is applied to identify key nodes and output a state topology graph Z_attn for subsequent prediction and switching decisions.

[0107] Perturbation prediction and prior graph generation: Use a trained graph convolution + GRU network to predict the state for the next 50ms (Δt);

[0108] Cluster analysis is performed on the prediction results to identify high-perturbation nodes, forming a priori map W, which is then output to the control switching module.

[0109] The applicability of MPC, sliding mode controller, etc. is evaluated based on parameters such as disturbance direction, intensity, and duration in W, and the priority dynamic matrix is ​​calculated.

[0110] Fuzzy rules are used to evaluate metrics such as response time and error convergence speed, and the ranking controller executes priorities.

[0111] If the error continues for more than 3 control cycles, the controller will be given low weight and a cooling period of ≥100ms will be set to avoid oscillation.

[0112] If the attention of a node increases rapidly by more than +30% and the current amplitude changes abruptly, it is determined to be a faulty phase; the faulty phase signal is removed, the control weights are redistributed according to the current amplitude ratio of the remaining four phases, and the flux deviation is compensated by the phase rotation algorithm; the new control vector is fused with the output of the old controller to maintain the equivalent electromagnetic torque.

[0113] Establish the expression for the generalized hypersurface, and assume... , where Ψ(t) is the output coupling strength coefficient function of GNN; the control parameters λ and γ are automatically adjusted as the coupling coefficient changes; a high coupling value increases the sliding mode gain coefficient to enhance robustness and vibration suppression capability;

[0114] Experiment and effect verification:

[0115] Operating Condition 1: Idle acceleration from 0 to 1500 rpm, switching response time < 60ms;

[0116] Operating Condition 2: 100% load instantaneous drop, the sliding mode controller successfully engages, recovery time ~80ms, temperature rise <5°C;

[0117] Operating Condition 3: Simulated phase C fault triggers, the control system automatically removes the phase and switches to four-phase control, the electric torque is retained at >85%, and the operation is smooth and without vibration.

[0118] Example 3, please refer to Figure 2 As shown in this embodiment, a five-phase permanent magnet speed control system for monorail cranes includes:

[0119] The state perception and modeling module collects stator current, voltage, speed and temperature rise signals of a five-phase permanent magnet synchronous motor and constructs a multi-axis state topology graph based on a graph neural network.

[0120] The disturbance prediction and prior analysis module constructs a multi-dimensional state prediction model for the motor based on the multi-axis state topology diagram, and uses a graph convolution mechanism to predict the torque disturbance trend in the future Δt time period, generating state prior information W.

[0121] The adaptive control strategy selection module selects the optimal set of control strategies based on the prior state information W, including two types of nonlinear speed controllers: model predictive controllers and sliding mode fault-tolerant compensation controllers, and adaptively switches between them based on a priority switching algorithm during operation.

[0122] The fault-tolerant reconfiguration control module, when detecting an abnormal operation of one or more phases, reconfigures the drive control vector based on redundant distribution to form a heterogeneous control state.

[0123] The sliding mode compensation and gain adjustment module embeds an improved generalized hypersurface regulator in the sliding mode fault-tolerant compensation controller, which adjusts the control gain in real time according to the coupling strength coefficient of the GNN output.

[0124] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A five-phase permanent magnet speed control method for monorail cranes, characterized in that: include: S100: Collect stator current, voltage, speed and temperature rise signals of a five-phase permanent magnet synchronous motor, and construct a multi-axis state topology diagram based on a graph neural network; S200. Based on the multi-axis state topology diagram, a multi-dimensional state prediction model for the motor is constructed, and a graph convolution mechanism is used to predict the torque disturbance trend in the future Δt time period, generating state prior information W. S300. Based on the prior state information W, select the matching optimal control strategy set, including two types of nonlinear speed controllers: model predictive controllers and sliding mode fault-tolerant compensation controllers, and adaptively switch between them during operation based on a priority switching algorithm, including: The perturbation features in the prior state information W are mapped to the controller state set, and a dynamic weight matrix is ​​generated accordingly to evaluate the adaptability score of each type of controller in the current prediction environment. A hierarchical priority selection mechanism is constructed, and different control strategies are ranked by a multi-factor fuzzy evaluation function. The ranking criteria include response delay, prediction offset and disturbance matching rate. Temporary weight reduction restrictions are implemented for controllers with continuous prediction deviations exceeding limits, thereby realizing the optimal controller invocation strategy based on dynamic adjustment of the disturbance environment; S400: When an abnormal operation of one or more phases is detected, the drive control vector is reconstructed based on the redundancy distribution to form a heterogeneous control state. S500 incorporates an improved generalized hypersurface regulator within the sliding mode fault-tolerant compensation controller, which adjusts the control gain in real-time based on the coupling strength coefficient output by the GNN, including: A generalized sliding mode hypersurface expression containing nonlinear dynamic boundary terms and current-coupled disturbance terms is constructed, and a residual driving factor is introduced to adjust the control target surface; The node coupling strength coefficients output by the graph neural network are injected as dynamic weights into the hypersurface regulator to characterize the response coupling nonuniformity of the five-phase system between different nodes. Nonlinear correction of sliding mode gain based on coupling weights enables adaptive adjustment of control strength for multiphase systems under sudden disturbances and fault conditions.

2. The five-phase permanent magnet speed control method for a monorail crane according to claim 1, characterized in that: The construction of the multi-axis state topology graph based on graph neural network specifically includes: The stator current, back electromotive force and phase angle data of each phase of the five-phase motor are collected, and the data are processed by sliding window segmentation and multidimensional normalization to form a time series state tensor. A weighted adjacency matrix is ​​constructed based on the interphase flux coupling degree and the historical operating frequency to generate a dynamic graph structure representing the multi-axis electromagnetic structure; We introduce a graph convolution module with residuals from graph neural networks to propagate features of the graph structure, and construct a state topology graph that integrates temporal and spatial characteristics by combining the current change trend between nodes. By assigning high weights to key phase nodes through a graph attention mechanism, a prior state graph is output for speed regulation decisions.

3. The five-phase permanent magnet speed control method for a monorail crane according to claim 1, characterized in that: The construction of the multidimensional state prediction model for the motor and the generation of prior state information W include: Using the generated state topology graph as input, a graph time series prediction network that integrates spatial graph structure and time series features is established, wherein the state embedding of each node includes current derivative, electric angular frequency perturbation and historical mutation response factor; By jointly modeling stacked graph convolutional networks and gated recurrent units, the state propagation process of graph structures within the future Δt prediction period is realized, capturing the evolution trajectory of perturbation trends between multiple axes; Cluster analysis is performed on the torque change direction, amplitude and correlation in the predicted node features to extract the disturbance prior features and generate prior information W containing spatial disturbance hotspots and trend directions.

4. The five-phase permanent magnet speed control method for a monorail crane according to claim 1, characterized in that: The reconstructed driving control vector based on redundancy distribution includes: The graph attention mechanism is used to detect abnormal current patterns in a five-phase system, identify faulty phase nodes, and perform dynamic attribution analysis on their output amplitude change rate. Based on the spatial redundancy distribution characteristics of the five-phase structure, after removing the weight of the faulty phase, the remaining four phases are non-uniformly redistributed in amplitude, and a phase rotation transformation is introduced to maintain the electromagnetic torque balance of the system. The reconstructed result is fused with the original controller output vector to generate a heterogeneous control state, which contains composite drive signals with unequal amplitude and unequal phase.

5. A five-phase permanent magnet speed control system for monorail cranes, used to implement the five-phase permanent magnet speed control method for monorail cranes as described in any one of claims 1-4, characterized in that: include: The state perception and modeling module collects stator current, voltage, speed and temperature rise signals of a five-phase permanent magnet synchronous motor and constructs a multi-axis state topology graph based on a graph neural network. The disturbance prediction and prior analysis module constructs a multi-dimensional state prediction model for the motor based on the multi-axis state topology diagram, and uses a graph convolution mechanism to predict the torque disturbance trend in the future Δt time period, generating state prior information W. The adaptive control strategy selection module selects the optimal set of control strategies based on the prior state information W, including two types of nonlinear speed controllers: model predictive controllers and sliding mode fault-tolerant compensation controllers, and adaptively switches between them based on a priority switching algorithm during operation. The fault-tolerant reconfiguration control module, when detecting an abnormal operation of one or more phases, reconfigures the drive control vector based on redundant distribution to form a heterogeneous control state. The sliding mode compensation and gain adjustment module embeds an improved generalized hypersurface regulator in the sliding mode fault-tolerant compensation controller, which adjusts the control gain in real time according to the coupling strength coefficient of the GNN output.

Images