A detection method and system applied to a magnetic suspension flywheel energy storage device

Abstract

The application discloses a kind of detection method and system applied to magnetic suspension flywheel energy storage equipment, it is related to flywheel detection technical field, the method includes: obtaining the data of flywheel;The data of flywheel includes the geometric parameter of flywheel, material attribute, operating parameter, the magnetic field intensity of flywheel, the local transient stress on the surface of flywheel and ultrasonic guided wave in flywheel;Flywheel digital twin model is constructed based on the data of flywheel, and the remaining useful life of flywheel is predicted by the flywheel digital twin model;The data of flywheel and the output of flywheel digital twin model are input to pre-trained federal learning model to carry out health score;Based on the predicted remaining useful life of flywheel and the health score, grading response is carried out.Through the processing scheme of the present disclosure, early failure of flywheel can be accurately perceived, equipment life is prolonged, energy storage interruption rate is reduced, and safety and operating efficiency are balanced.

Description

Technical Field

[0001] This invention relates to the field of flywheel testing technology, and in particular to a testing method and system for magnetic levitation flywheel energy storage devices. Background Technology

[0002] With its advantages of high power density and long cycle life, magnetic levitation flywheel energy storage equipment has become a core energy storage device in fields such as aerospace satellite attitude control and power grid frequency regulation and peak shaving. Its health status directly determines the system's operational safety and energy storage efficiency, which places stringent requirements on the accuracy, foresight, and reliability of testing technology.

[0003] Existing methods for detecting magnetic levitation flywheels have several key drawbacks:

[0004] First, the sensing accuracy is insufficient. Traditional strain gauges and Hall sensors can only achieve millimeter-level stress detection and millitalas-level magnetic field measurement, and cannot capture the precursor signals of micron-level stress concentration, micro-magnetic distortion, crack initiation, and early failure of magnetic bearings.

[0005] Second, the internal defect detection blind zone: vibration sensors can only monitor surface vibrations and are unable to identify deep defects such as microcracks and delamination inside composite rotors.

[0006] Third, the life prediction capability is weak, and there is a lack of fault evolution simulation combined with physical models. It can only provide passive alarms and cannot quantify the remaining life in advance.

[0007] Fourth, the diagnostic model has poor generalization ability, single device fault samples are scarce, training models are difficult to cover faults in multiple scenarios, and cross-device data sharing poses a risk of privacy leakage.

[0008] Fifth, the lack of a closed-loop control system and the disconnect between detection results and the control system make it impossible to proactively intervene based on the health status, leaving only passive shutdowns, which severely reduces equipment utilization and energy storage continuity.

[0009] The aforementioned shortcomings make it difficult for existing testing methods to meet the requirements of high-reliability scenarios, and are prone to problems such as sudden equipment failure, shortened lifespan, and energy storage interruption. There is an urgent need for a breakthrough in a multi-dimensional, high-precision, and full-process health management testing technology. Summary of the Invention

[0010] In view of this, the present disclosure provides a detection method for magnetic levitation flywheel energy storage devices, which at least partially solves the problems existing in the prior art.

[0011] In a first aspect, embodiments of this disclosure provide a detection method for magnetic levitation flywheel energy storage devices, the method comprising the following steps:

[0012] Acquire flywheel data; the flywheel data includes the flywheel's geometric parameters, material properties, operating parameters, magnetic field strength, local instantaneous stress on the flywheel surface, and ultrasonic guided waves inside the flywheel;

[0013] A digital twin model of the flywheel is constructed based on the flywheel data, and the remaining service life of the flywheel is predicted using the digital twin model.

[0014] The flywheel data and the output of the flywheel digital twin model are input into a pre-trained federated learning model for health scoring;

[0015] A graded response is performed based on the predicted remaining flywheel lifespan and the health score.

[0016] According to a specific implementation of an embodiment of this disclosure, local instantaneous stress on the surface of a flywheel is obtained based on an NV color center quantum sensor array. and local instantaneous magnetic field strength ;

[0017] based on and Generate a two-dimensional stress distribution cloud map on the rotor surface. and two-dimensional magnetic field distribution cloud map ,in, The coordinates of the rotor surface are rectangular coordinates. Sampling time;

[0018] Ultrasonic guided waves inside the flywheel are obtained by using acoustic metasurface waveguides arranged radially along the rotor.

[0019] According to a specific implementation of an embodiment of this disclosure, the method further includes:

[0020] The sound wave transmission efficiency is calculated based on the following formula. :

[0021] ;

[0022] in, The energy received by sound waves; The emitted energy of sound waves;

[0023] The centroid of the sound wave spectrum is calculated based on the following formula. :

[0024] ;

[0025] in, For the first Frequency values ​​at each frequency point; For serial number; For frequency The corresponding sound wave amplitude; This represents the number of frequency points.

[0026] According to a specific implementation of this disclosure, constructing a flywheel digital twin model based on the flywheel data includes the following steps:

[0027] Based on the flywheel data, define the flywheel digital twin model state vector:

[0028] ;

[0029] in, The state vector of the flywheel digital twin model; for The angular velocity of the magnetic levitation flywheel rotor at any given moment; for The electromagnetic force applied by the magnetic bearing at any given moment; The rotor surface temperature; for The spatial location of the flywheel predicted by the time-of-flight digital twin model Stress distribution at the location; for The spatial location of the flywheel predicted by the time-of-flight digital twin model The distribution of magnetic induction intensity at that location;

[0030] Calculate the health consistency index to quantify the differences between virtual and real models:

[0031]

[0032] in, For the health consistency index, ; This is the actual flywheel state vector, representing the actual flywheel state vector obtained directly from the sensor measurement and processing. The state vector of the flywheel digital twin model;

[0033] when At that time, recalibrate the model until ;

[0034] The method further includes: generating a two-dimensional stress distribution cloud map of the rotor surface at preset time intervals. Two-dimensional magnetic field distribution cloud map Sound wave transmission efficiency Harmony of the sound wave spectrum centroid Input a flywheel digital twin model and correct the model parameters using the Kalman filter algorithm.

[0035] According to a specific implementation of this disclosure, predicting the remaining service life of the flywheel using the flywheel digital twin model includes:

[0036] The crack propagation rate was simulated in the flywheel digital twin model based on the Paris formula:

[0037] ;

[0038] in This represents the crack propagation rate. The length of the crack; The material constant for crack propagation; The stress intensity factor amplitude; The material index for crack propagation; This represents the number of loop iterations.

[0039] The time taken for the crack length to reach the design critical value in the flywheel digital twin model is used to predict the remaining life of the flywheel.

[0040] Output digital twin to predict remaining lifetime.

[0041] According to a specific implementation of this disclosure, the step of inputting the flywheel data and the output of the flywheel digital twin model into a pre-trained federated learning model for health scoring includes:

[0042] The local federated learning model is trained based on historical data; the historical data includes local stress on the flywheel surface, local magnetic field strength, sound wave transmission efficiency, sound wave spectrum centroid, health consistency index, and time series predicting the remaining life of the flywheel.

[0043] The weight gradient parameters of all local federated learning models are sent to the cloud for global model aggregation to obtain the global gradient.

[0044] Update the global model parameters based on the global gradient;

[0045] The updated global model parameters are distributed to each client, and the client continues to train based on the new parameters, iterating until the global loss converges, and a new federated learning model is obtained.

[0046] The flywheel data and the output of the flywheel digital twin model are input into the new federated learning model to output a health score. .

[0047] According to a specific implementation of this disclosure, a graded response is performed based on the predicted remaining flywheel lifespan and the health score, including:

[0048] A comprehensive early warning index is calculated based on health score, predicted remaining life expectancy, and health consistency index. :

[0049] ;

[0050] in, Rate your health; For health consistency index; To predict remaining lifespan; Design the flywheel for maximum remaining lifespan. ;

[0051] when At that time, the warning level was Level 1;

[0052] when At that time, the warning level was Level 2;

[0053] when At that time, the warning level was Level 3;

[0054] The graded response includes:

[0055] When the warning level is Level 1, and When this happens, active vibration suppression is triggered, adjusting the PID parameters of the magnetic levitation controller to reduce the rotor vibration amplitude;

[0056] When the warning level is Level 2, and the following conditions are met: , , Under any condition, reduce the speed setpoint to decrease rotor centrifugal stress and magnetic bearing load;

[0057] When the warning level is Level 3, and the following conditions are met: , , Under any condition, an emergency shutdown is triggered, and the mechanical auxiliary bearing lock is activated to prevent rotor instability and collision, thus protecting the magnetic levitation system and the main body of the energy storage equipment.

[0058] Secondly, embodiments of this disclosure provide a detection system for magnetic levitation flywheel energy storage devices, the system comprising:

[0059] The data acquisition module is configured to acquire data from the flywheel; the flywheel data includes the flywheel's geometric parameters, material properties, operating parameters, magnetic field strength, local instantaneous stress on the flywheel surface, and ultrasonic guided waves inside the flywheel.

[0060] A digital twin module is configured to construct a digital twin model of the flywheel based on the flywheel's data, and to predict the flywheel's remaining service life using the digital twin model.

[0061] The scoring module is configured to input the flywheel data and the output of the flywheel digital twin model into a pre-trained federated learning model for health scoring;

[0062] The response module is configured to provide a graded response based on the predicted remaining flywheel lifespan and the health score.

[0063] The detection method for magnetic levitation flywheel energy storage devices disclosed in this embodiment integrates NV color center quantum sensing and acoustic metasurface waveguide technology to capture both macroscopic stress / magnetic field distortion on the rotor surface and detect internal microscopic cracks and delamination. This improves early fault detection lead time by 3-5 times compared to traditional methods, avoiding missed detection of minor anomalies. Digital twins simulate fault evolution through virtual-real synchronization and fault physics models, reducing remaining life prediction errors. A federated learning framework enables collaborative optimization of data from multiple devices, addressing the problem of insufficient fault samples from single devices and improving fault identification accuracy. Hierarchical self-healing control achieves proactive suppression of mild risks, delayed load reduction for moderate risks, and safe shutdown for severe risks, extending equipment life and reducing energy storage outage rates. Attached Figure Description

[0064] Figure 1 This is a schematic flowchart of a testing method for magnetic levitation flywheel energy storage devices provided in an embodiment of the present disclosure;

[0065] Figure 2 A flowchart illustrating a detection method for magnetic levitation flywheel energy storage devices provided in this disclosure.

[0066] Figure 3 This is a schematic diagram of a detection system for a magnetic levitation flywheel energy storage device, provided as an embodiment of the present disclosure. Detailed Implementation

[0067] The embodiments of this disclosure will now be described in detail with reference to the accompanying drawings.

[0068] The following specific examples illustrate the implementation of this disclosure. Those skilled in the art can easily understand other advantages and effects of this disclosure from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of this disclosure, and not all of them. This disclosure can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this disclosure. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0069] It should be noted that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this disclosure, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. Furthermore, this device and / or this method can be implemented using other structures and / or functionalities besides one or more of the aspects set forth herein.

[0070] Furthermore, specific details are provided in the following description to facilitate a thorough understanding of the examples. However, those skilled in the art will understand that the described aspects can be practiced without these specific details.

[0071] Figure 1 A schematic diagram of the detection method flow for magnetic levitation flywheel energy storage devices provided in this embodiment of the disclosure;

[0072] Figure 2 To and Figure 1 The flowchart of the corresponding testing method applied to magnetic levitation flywheel energy storage devices.

[0073] like Figure 1 As shown, in step S110, the flywheel data is acquired; the flywheel data includes the flywheel's geometric parameters, material properties, operating parameters, magnetic field strength, local instantaneous stress on the flywheel surface, and ultrasonic guided waves inside the flywheel.

[0074] The flywheel's geometric parameters mainly include rotor diameter and thickness, its material properties mainly include elastic modulus and Poisson's ratio, and its operating parameters mainly include rotational speed and electromagnetic force. There are no restrictions on these parameters. The acquired data can also include all the data required to build a digital twin model of the flywheel, such as temperature, angular velocity, and electromagnetic force.

[0075] In this embodiment of the invention, the local instantaneous stress on the flywheel surface is obtained based on the NV color center quantum sensor array. and local instantaneous magnetic field strength ;based on and Generate a two-dimensional stress distribution cloud map on the rotor surface. and two-dimensional magnetic field distribution cloud map ,in, The coordinates of the rotor surface are rectangular coordinates. The sampling time is used to obtain ultrasonic guided waves inside the flywheel based on acoustic metasurface waveguides arranged radially along the rotor.

[0076] In this embodiment of the invention, the method further includes:

[0077] The sound wave transmission efficiency is calculated based on the following formula. :

[0078] ;

[0079] in, The energy received by sound waves; The emitted energy of sound waves;

[0080] The centroid of the sound wave spectrum is calculated based on the following formula. :

[0081] ;

[0082] in, For the first Frequency values ​​at each frequency point; For serial number; For frequency The corresponding sound wave amplitude; This represents the number of frequency points.

[0083] More specifically, the magnetic field strength of the flywheel and the local instantaneous stress on the flywheel surface are obtained based on the following methods.

[0084] Traditional sensors cannot analyze micron-level stress concentrations or minute magnetic field anomalies, which are early signs of crack initiation and magnetic bearing failure.

[0085] This invention involves embedding an array of NV color centers (NCCs) quantum sensors at key points on the surface and inside of a flywheel rotor. Each sensor measures local instantaneous stress and magnetic field strength. The quantum sensors provide nanometer-level resolution, enabling nanometer-level health monitoring and thus detecting minute distortions in material micro-strain and electromagnetic fields before macroscopic failures occur.

[0086] Furthermore, quantum sensors based on nitrogen-vacancy (NV) color centers are implanted or plated in key parts of the flywheel rotor (such as near the magnetic bearing, in the composite laminate region, and in the stress concentration region under high-speed rotation).

[0087] Each NV color center quantum sensor simultaneously measures two physical quantities: local instantaneous stress. and local instantaneous magnetic field strength .

[0088] For all quantum sensors and Preprocessing, including noise reduction and spatial interpolation, is performed to generate a two-dimensional stress distribution cloud map of the rotor surface. and two-dimensional magnetic field distribution cloud map ,in, The coordinates of the rotor surface are rectangular coordinates. Sampling time.

[0089] In this embodiment of the invention, ultrasonic guided waves inside the flywheel are obtained based on the following method:

[0090] Acoustic metasurfaces are sensitive to micro-defects within materials, enabling the detection of deep structural changes inaccessible to traditional vibration sensors, thus complementing quantum stress data. This invention utilizes the properties of acoustic metamaterials to transform the flywheel rotor structure itself into a giant distributed sensor. By monitoring minute changes in the propagation characteristics of elastic waves (phonons) within the material, it detects microscopic damage inaccessible to traditional methods, such as interlaminar debonding in composite materials, microcrack initiation, and material fatigue.

[0091] More specifically, an acoustic metasurface waveguide (a phononic crystal structure composed of periodically arranged tungsten steel cylinders) is embedded in the prefabricated interlayer channel inside the flywheel composite rotor. The waveguide is arranged radially along the rotor and covers the core stress area of ​​the rotor.

[0092] A narrowband ultrasonic guided wave (center frequency 2.5kHz) is emitted by a piezoelectric transducer at one end of the waveguide, and the transmitted wave is received by a piezoelectric transducer at the other end, simultaneously collecting the emitted energy. With receiving energy ;

[0093] Calculate sound wave transmission efficiency The formula reflects the integrity of the material's internal structure:

[0094]

[0095] When defects (such as cracks or debonding) appear in the waveguide path, sound waves will be scattered and reflected, resulting in a reduction in the energy reaching the receiver. It will decrease significantly ( ).

[0096] Calculate the centroid of the sound wave spectrum :

[0097]

[0098] in, For the first Frequency values ​​at each frequency point; For serial number; For frequency The corresponding sound wave amplitude; The frequency point is 1024 in this invention.

[0099] Next, proceed to step S120.

[0100] In step S120, a digital twin model of the flywheel is constructed based on the flywheel data, and the remaining service life of the flywheel is predicted using the digital twin model.

[0101] The digital flywheel digital twin model assesses risks in advance through fault simulation, avoiding the dangers of physical testing, and quantifies the differences between virtual and real models through the HDT consistency index, capturing anomalies not covered by the model.

[0102] In this embodiment of the invention, constructing a digital twin model of the flywheel based on the flywheel data includes the following steps:

[0103] Based on the flywheel data, define the flywheel digital twin model state vector:

[0104] ;

[0105] in, The state vector of the flywheel digital twin model; for The angular velocity of the magnetic levitation flywheel rotor at any given moment; for The electromagnetic force applied by the magnetic bearing at any given moment; The rotor surface temperature; for The spatial location of the flywheel predicted by the time-of-flight digital twin model Stress distribution at the location; for The spatial location of the flywheel predicted by the time-of-flight digital twin model The distribution of magnetic induction intensity at that location;

[0106] Calculate the health consistency index to quantify the differences between virtual and real models:

[0107]

[0108] in, For the health consistency index, ; This is the actual flywheel state vector; The state vector of the flywheel digital twin model;

[0109] when At that time, recalibrate the model until ;

[0110] The method further includes: generating a two-dimensional stress distribution cloud map of the rotor surface at preset time intervals. Two-dimensional magnetic field distribution cloud map Sound wave transmission efficiency Harmony of the sound wave spectrum centroid Input a flywheel digital twin model and correct the model parameters using the Kalman filter algorithm.

[0111] More specifically, a high-precision digital twin model containing multi-physics coupling is constructed to fully map all the key characteristics of the physical flywheel.

[0112] 1. Digital Twin Model Construction

[0113] Based on the flywheel's geometric parameters (rotor diameter, thickness), material properties (elastic modulus, Poisson's ratio), and operating parameters (speed, electromagnetic force), a digital twin model of the flywheel is constructed, and the model's state vector is defined:

[0114]

[0115] in, State vector of flywheel digital twin model ; for The angular velocity of the magnetic levitation flywheel rotor at any given moment; for The electromagnetic force applied by the magnetic bearing at any given time is calculated by measuring the real-time current of the magnetic bearing coil through a current sensor, and then combining this with a pre-calibrated current-electromagnetic force model of the magnetic bearing. for The time flywheel in space The temperature distribution at the location is obtained through distributed temperature sensors; for The spatial location of the flywheel predicted by the time-of-flight digital twin model Stress distribution at the location; for The spatial location of the flywheel predicted by the time-of-flight digital twin model The distribution of magnetic induction intensity at that location; This is a transpose.

[0116] Will As boundary conditions, the flywheel is simulated and calculated using the finite element analysis (FEA) method. Time and space Stress distribution at the location.

[0117] Predicted flywheel position in space Stress distribution at Specifically, it includes the following steps:

[0118] First, based on the flywheel design drawings, a three-dimensional geometric model of the rotor is established in finite element software; and the rotor surface is meshed, wherein the stress concentration area (edge, spoke root) adopts a dense mesh (mesh size ≤ 1mm), and the non-critical area adopts a gradient mesh to balance accuracy and computing power; the present invention does not limit the distribution of the mesh and can be adjusted according to the actual situation.

[0119] The three-dimensional geometric model of the rotor is assigned values ​​based on the mechanical properties of the rotor material, including the elastic modulus. Poisson's ratio ,density and coefficient of thermal expansion .

[0120] Secondly, , and As input to the rotor's three-dimensional geometric model, the driving stress is calculated:

[0121] measured angular velocity Converted to centrifugal stress :

[0122]

[0123] in, Density; Radial distance; ω is the angular velocity.

[0124] Centrifugal stress is applied to the rotor edge by distributing loads.

[0125] The measured electromagnetic force of the magnetic bearing This is converted into contact pressure, which is applied to the contact area of ​​the magnetic bearing through concentrated force.

[0126] Based on measured temperature distribution Calculate thermal strain , The coefficient of thermal expansion of the material. The reference temperature is the temperature at which the material has no thermal strain. Thermal strain is introduced into the stress field through temperature field coupling.

[0127] Next, after calculating the transient stress field using an implicit solver and obtaining the stress tensor (including normal stress and shear stress in the x, y, and z directions) for each grid node, an interpolation algorithm is used to extend the discrete node stress into a continuous spatial distribution. Stress cloud map is generated.

[0128] Finally, the stress measured by the NV color center quantum sensor was... Compared with model predictions For comparison, when the error exceeds 5%, adjust the material parameters (such as elastic modulus ±5%) or mesh density, and recalculate until the error is less than 5% and the health consistency index is met. .

[0129] Predicting the flywheel's spatial position Magnetic induction intensity distribution at [location] Specifically, it includes the following steps:

[0130] First, a three-dimensional geometric model of the magnetic bearing (coil, magnetic pole) and rotor (iron core / permanent magnet) is established in electromagnetic simulation software; the three-dimensional geometric model is meshed, and a finer mesh (mesh size ≤ 0.1 mm) is used for areas with large magnetic field gradients such as air gap and magnetic pole tip; this invention does not limit the distribution of the mesh and can be adjusted according to the actual situation.

[0131] The three-dimensional geometric model is assigned values ​​based on the magnetic properties of the material, including coil conductivity, relative magnetic permeability of magnetic poles, magnetic pole saturation magnetic induction intensity and / or permanent magnet remanence and permanent magnet coercivity.

[0132] Secondly, the measured physical quantities are converted into excitation sources for the electromagnetic model:

[0133] Based on measured electromagnetic force Reverse coil current .

[0134]

[0135] in, ; This refers to the number of coil turns. The area of ​​the magnetic poles; For air gap;

[0136] Right now,

[0137] The coil current is applied to the coil in the form of current density.

[0138] If the rotor contains permanent magnets, set its magnetization direction (e.g., radial magnetization).

[0139] A balloon boundary (with the normal component of the magnetic field strength being 0) is set at the outer boundary of the model to simulate an infinitely far magnetic field environment.

[0140] Secondly, a static magnetic field solver is used to solve Maxwell's equations. , ), to obtain the magnetic induction intensity of each node. ;in, For curl operator; The magnetic field strength; Current density; Permeability is determined by the electromagnetic properties of the material or medium.

[0141] The magnetic field strength of discrete nodes is extended into a continuous spatial distribution using an interpolation algorithm. This generates a magnetic field cloud map.

[0142] Finally, the magnetic field measured by the NV color center quantum sensor was... Compared with model predictions In comparison, if the error is greater than 0.1 mT, adjust the permeability or coil current and recalculate until the error is less than 0.1 mT. .

[0143] 2. Simultaneous updates of virtual and real data

[0144] Set synchronization time interval (Preferred interval: 10 seconds) Two-dimensional stress distribution cloud map of rotor surface Two-dimensional magnetic field distribution cloud map Sound wave transmission efficiency Harmony of the sound wave spectrum centroid Input the flywheel digital twin model, and correct the model parameters (such as material damping coefficient and electromagnetic force coefficient) using the Kalman filter algorithm, so that... With the actual flywheel state vector Consistent.

[0145]

[0146] in, for The angular velocity of the magnetic levitation flywheel rotor at any given moment; for The electromagnetic force applied by the magnetic bearing at any given moment; for The time flywheel in space Temperature distribution at that location.

[0147] Calculate the health consistency index to quantify the differences between virtual and real models:

[0148]

[0149] in, For the health consistency index, The closer the value is to 1, the more consistent the model is with the actual state. This is the actual flywheel state vector; This is the state vector of the flywheel digital twin model.

[0150] 3. Fault Injection and Lifetime Prediction

[0151] In the synchronized flywheel digital twin model, a crack propagation fault is injected:

[0152] Simulation of crack propagation rate based on the Paris formula:

[0153] ;

[0154] in, This represents the crack propagation rate. The length of the crack; This represents the number of loop iterations. The material constant for crack propagation is determined by fatigue testing of the rotor material; preferably, it is used in this invention. ; The stress intensity factor amplitude is positively correlated with stress level and crack size. The material index for crack propagation indicates the material's sensitivity to crack propagation, such as the material index of a steel rotor. .

[0155] Based on the acoustic wave transmission efficiency obtained in step S110 When the crack length decreases significantly, the crack length is obtained, and the time it takes for the crack length to reach the design critical value is calculated by simulating the crack propagation rate. This time is the predicted remaining lifetime. The design critical value can be obtained by consulting the product manual; the sound wave transmission efficiency can be obtained through experimental calibration. The empirical formula relating crack length to crack length is based on The length of the crack can be obtained by observing the decrease in the crack value.

[0156] This invention combines real-time sensing data with physical simulation using a digital twin model of the flywheel. It simulates the failure evolution process in the absence of physical faults, predicts remaining lifespan in advance, and allows for advance planning of downtime for the maintenance of magnetic levitation flywheels. The quantifiable model error avoids misjudgments caused by model inaccuracy, improves the transparency and interpretability of the prediction process, and solves the black box problem of traditional data-driven models.

[0157] Next, proceed to step S130.

[0158] In step S130, the data of the flywheel and the output of the flywheel digital twin model are input into the pre-trained federated learning model for health scoring.

[0159] In this embodiment of the invention, the step of inputting the flywheel data and the output of the flywheel digital twin model into a pre-trained federated learning model for health scoring includes: training a local federated learning model based on historical data; the historical data includes local stress on the flywheel surface, local magnetic field strength, acoustic wave transmission efficiency, acoustic wave spectrum centroid, health consistency index, and a time series predicting the remaining lifespan of the flywheel; sending the weight gradient parameters of all local federated learning models to the cloud for global model aggregation to obtain a global gradient; updating the global model parameters based on the global gradient; distributing the updated global model parameters to each client, which continues training based on the new parameters, iterating until the global loss converges to obtain a new federated learning model; and inputting the flywheel data and the output of the flywheel digital twin model into the new federated learning model to output a health score. .

[0160] In this embodiment of the invention, predicting the remaining service life of the flywheel using the flywheel digital twin model includes:

[0161] The crack propagation rate was simulated in the flywheel digital twin model based on the Paris formula:

[0162] ;

[0163] in, This represents the crack propagation rate. The length of the crack; The material constant for crack propagation; The stress intensity factor amplitude; The material index for crack propagation; This represents the number of loop iterations.

[0164] The time taken for the crack length to reach the design critical value in the flywheel digital twin model is used to predict the remaining life of the flywheel.

[0165] Output digital twin to predict remaining lifetime.

[0166] More specifically, a federated learning framework is built, including a cloud server and clients; each magnetic levitation flywheel device is a client, and the cloud server is responsible for receiving the model gradients uploaded by the clients and aggregating them to generate a global model.

[0167] 1. Local Federated Learning Model Training:

[0168] Each client trains its local federated learning model using its own historical data, and the output is a health score. ; Calculate local loss during training With model weight gradient .

[0169] Specifically, the following steps are included:

[0170] Step 1: Data Acquisition

[0171] Collect historical operating data of client devices, including: input features such as local stress on the flywheel surface, local magnetic field strength, acoustic wave transmission efficiency, acoustic wave spectrum centroid, health consistency index, time series of predicted flywheel remaining life, and equipment operation and maintenance records; equipment operation and maintenance records include: fault occurrence time, fault type (such as crack, delamination) and other tagged data.

[0172] Historical data preprocessing includes data cleaning and data normalization.

[0173] Time series segmentation: The sliding window method is used to segment long time series data into multiple subsequences. The window length is set to 30 (i.e., containing 30 consecutive sampling points), the step size is set to 10, and training samples are generated in the format of (feature sequence, health label). The health label indicates whether the data of the corresponding window length has equipment operation and maintenance records and the severity of the fault.

[0174] Step 2: Feature Extraction

[0175] Spatial distribution pattern extraction is performed based on a convolutional neural network (CNN). The CNN network consists of two convolutional layers (3×3 kernel size, ReLU activation function) used for extraction. , Spatial distribution pattern

[0176] Time extraction is performed using a Long Short-Term Memory (LSTM) network. The LSTM subnetwork contains one LSTM layer (64 hidden units) for learning. , , The dynamic pattern of time.

[0177] The outputs of the CNN and LSTM are concatenated to obtain the feature vector. A health score is output through a single fully connected layer (with the sigmoid activation function). .

[0178] Step 3: Training Process

[0179] The preprocessed training samples were divided into a training set and a validation set in an 8:2 ratio.

[0180] Loss function: Mean squared error (MSE) is used as the local loss. ;

[0181] The model weight gradient is calculated using the backpropagation algorithm. ;

[0182] Optimizer: Select the Adam optimizer, set the learning rate to 0.001, and the weight decay to [value missing]. ;

[0183] Training iteration: The training batch size is 32 samples, the number of iteration rounds is 50, and the model performance is evaluated on the validation set after every 5 rounds of training. If the performance declines, training is stopped (the performance is judged to have declined when the validation set performance index shows a deteriorating trend for more than 2 consecutive rounds or the mean squared error increases by more than 5% in a single round).

[0184] Step 4: Model Evaluation and Gradient Extraction

[0185] After training, evaluate the model's health score prediction accuracy on the validation set. If the accuracy requirement is met (validation set MSE < 0.05), then extract the model's weight gradients. The model is uploaded to the cloud server to participate in global model aggregation; if the accuracy is not satisfactory, the model architecture or hyperparameters need to be adjusted and the model retrained until the accuracy requirements are met.

[0186] Step 5: Health Score Acquisition

[0187] The true health label of the sample was determined by combining equipment maintenance records with expert annotations. :

[0188] For example: If the equipment is in a fault-free and normal operating condition, it should be labeled... If the equipment has minor defects (such as micron-sized cracks or micro-delamination), please indicate the defect. If the equipment has moderate defects (such as cracks extending to 0.3mm, or delamination area increasing), indicate this. If the equipment has severe defects (such as critical cracks or magnetic bearing failure), indicate this. 0.

[0189] Local federated learning models learn input features With Real Health Labels The mapping relationship, after training, is based on the input feature vector. Model output This is the predicted health score.

[0190] Calculate local loss using the following formula :

[0191]

[0192] Where N is the number of samples in the local training set; The predicted health score for the j-th sample; Let be the true health label of the j-th sample.

[0193] The smaller the value, the more accurately the local federated learning model learns the mapping relationship between health status and input features. During training, the model continuously adjusts its parameters through the backpropagation algorithm to minimize... This improves the predictive accuracy of health scores.

[0194] Model weight gradient It is a loss function For the set of model weight parameters The partial derivatives reflect the direction and magnitude of parameter updates.

[0195] First, for a single sample Calculate the loss on the output Partial derivatives:

[0196]

[0197] By backpropagating along the model, the partial derivatives of the loss with respect to the parameters of the fully connected layer, LSTM layer, and CNN layer are calculated sequentially to obtain the weight gradient of the entire model. (Includes parameter gradients for all layers).

[0198] 2. Global model aggregation:

[0199] Step 1: Calculate the client-side aggregate weight.

[0200] The cloud assigns aggregation weights based on the amount of local data on each client, ensuring that clients with more data contribute more to the global model.

[0201] Let the first The local data volume of each client is The total amount of data from all clients is , If the total number of clients is , then the th The aggregate weight of each client is: ,satisfy This ensures that the weights are normalized.

[0202] Step 2: Gradient-weighted aggregation.

[0203] The cloud performs a weighted average of the gradients from all clients to generate a global gradient:

[0204]

[0205] Step 3: Update global model parameters.

[0206] Global model parameters are updated using global gradients in the cloud. :

[0207]

[0208] in, The global learning rate (adjusted from 0.001 to 0.01 depending on convergence) is used to achieve iterative optimization of the global model.

[0209] Step 4: Global model distribution and iteration.

[0210] The cloud will update the global model parameters The new global model is distributed to all clients, who then continue local training based on it, repeating the process of local training → gradient uploading → global aggregation until the global model converges (e.g., global loss). (Tend to stabilize).

[0211] When the relative rate of change of global loss in two consecutive rounds is less than or equal to a preset threshold When the global model converges, it is determined that:

[0212]

[0213] in, This represents the number of iterations. For the first The global loss after the next iteration; For the first The global loss after the next iteration; It is 0.01.

[0214] Failure modes of a single flywheel are extremely rare, and relying on historical data from a single device to train a fault diagnosis model will result in poor model generalization ability. The federated learning approach of this invention optimizes the model using multi-scenario data from multiple clients, improving fault identification generalization while avoiding the leakage of raw data.

[0215] Next, proceed to step S140.

[0216] At step S140, a graded response is performed based on the predicted remaining flywheel lifespan and the health score.

[0217] In this embodiment of the invention, a graded response is performed based on the predicted remaining service life of the flywheel and the health score, including:

[0218] A comprehensive early warning index is calculated based on health score, predicted remaining life expectancy, and health consistency index. :

[0219] ;

[0220] in, Rate your health; For health consistency index; To predict remaining lifespan; Design the flywheel for maximum remaining lifespan. ;

[0221] when At that time, the warning level was Level 1;

[0222] when At that time, the warning level was Level 2;

[0223] when At that time, the warning level was Level 3;

[0224] The graded response includes:

[0225] When the warning level is Level 1, and When this happens, active vibration suppression is triggered, adjusting the PID parameters of the magnetic levitation controller to reduce the rotor vibration amplitude;

[0226] When the warning level is Level 2, and the following conditions are met: , , Under any condition, reduce the speed setpoint to decrease rotor centrifugal stress and magnetic bearing load;

[0227] When the warning level is Level 3, and the following conditions are met: , , Under any condition, an emergency shutdown is triggered, and the mechanical auxiliary bearing lock is activated to prevent rotor instability and collision, thus protecting the magnetic levitation system and the main body of the energy storage equipment.

[0228] More specifically, firstly, an early warning level is established based on a health score. Predicting remaining lifespan The comprehensive early warning index is calculated based on the consistency index with health. :

[0229]

[0230] in, The maximum remaining lifespan of the flywheel is designed (i.e., the reference service life of the flywheel minus the service life already used). The higher the value, the lower the risk.

[0231] when At that time, the warning level was Level 1 (mild risk).

[0232] when At that time, the warning level was Level 2 (moderate risk).

[0233] when At that time, the warning level was Level 3 (severe risk).

[0234] Secondly, implement graded self-healing control:

[0235] When the warning level is Level 1, and When this happens, active vibration suppression is triggered, adjusting the PID parameters of the magnetic levitation controller to reduce the rotor vibration amplitude;

[0236] When the warning level is Level 2, and the following conditions are met: , , Under any condition, reduce the speed setpoint to decrease rotor centrifugal stress and magnetic bearing load;

[0237] When the warning level is Level 3, and the following conditions are met: , , Under any condition, an emergency shutdown is triggered, and the mechanical auxiliary bearing lock is activated to prevent rotor instability and collision, thus protecting the magnetic levitation system and the main body of the energy storage equipment.

[0238] Traditional detection methods can only issue alarms after a fault is detected, and cannot proactively intervene based on the health status. For example, if the rotor experiences slight vibration, the system can only wait for the vibration to intensify before shutting down, leading to the escalation of the fault. The detection method proposed in this invention for magnetic levitation flywheel energy storage devices links the detection-diagnosis-prediction results with the control system in a closed loop. Through hierarchical control, it achieves proactive suppression of minor risks, load reduction and delay of moderate risks, and safe shutdown of severe risks. This upgrades from passive operation and maintenance to proactive self-healing, extending equipment lifespan, reducing energy storage outage rates, and optimizing operational efficiency while ensuring safety.

[0239] Figure 3 The present invention illustrates a detection system 300 for magnetic levitation flywheel energy storage devices, including a data acquisition module 310, a digital twin module 320, a scoring module 330, and a response module 340.

[0240] The data acquisition module 310 is used to acquire data of the flywheel; the flywheel data includes the flywheel's geometric parameters, material properties, operating parameters, magnetic field strength, local instantaneous stress on the flywheel surface, and ultrasonic guided waves inside the flywheel;

[0241] The digital twin module 320 is used to construct a digital twin model of the flywheel based on the flywheel data, and to predict the remaining service life of the flywheel through the digital twin model of the flywheel;

[0242] The scoring module 330 is used to input the data of the flywheel and the output of the flywheel digital twin model into a pre-trained federated learning model to perform health scoring;

[0243] The response module 340 is used to perform a graded response based on the predicted remaining flywheel lifespan and the health score.

[0244] The above description is merely a specific embodiment of this disclosure, but the scope of protection of this disclosure 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 disclosure should be included within the scope of protection of this disclosure.

Claims

1. A testing method for magnetic levitation flywheel energy storage devices, characterized in that, The method includes the following steps: Acquire flywheel data; the flywheel data includes the flywheel's geometric parameters, material properties, operating parameters, magnetic field strength, local instantaneous stress on the flywheel surface, and ultrasonic guided waves inside the flywheel; A digital twin model of the flywheel is constructed based on the flywheel data, and the remaining service life of the flywheel is predicted using the digital twin model. The flywheel data and the output of the flywheel digital twin model are input into a pre-trained federated learning model for health scoring; A graded response is performed based on the predicted remaining flywheel lifespan and the health score; The step of inputting the flywheel data and the output of the flywheel digital twin model into a pre-trained federated learning model for health scoring includes: The local federated learning model is trained based on historical data; the historical data includes local stress on the flywheel surface, local magnetic field strength, sound wave transmission efficiency, sound wave spectrum centroid, health consistency index, and time series predicting the remaining life of the flywheel. The weight gradient parameters of all local federated learning models are sent to the cloud for global model aggregation to obtain the global gradient. Update the global model parameters based on the global gradient; The updated global model parameters are distributed to each client, and the client continues to train based on the new parameters, iterating until the global loss converges, and a new federated learning model is obtained. The flywheel data and the output of the flywheel digital twin model are input into the new federated learning model to output a health score. ; The construction of a digital twin model of the flywheel based on the flywheel data includes the following steps: Based on the flywheel data, define the flywheel digital twin model state vector: ; in, The state vector of the flywheel digital twin model; for The angular velocity of the magnetic levitation flywheel rotor at any given moment; for The electromagnetic force applied by the magnetic bearing at any given moment; The rotor surface temperature; for The spatial location of the flywheel predicted by the time-of-flight digital twin model Stress distribution at the location; for The spatial location of the flywheel predicted by the time-of-flight digital twin model The distribution of magnetic induction intensity at that location; Calculate the health consistency index to quantify the differences between virtual and real models: ； in, For the health consistency index, ; This is the actual flywheel state vector; The state vector of the flywheel digital twin model; when At that time, recalibrate the model until ; The method further includes: generating a two-dimensional stress distribution cloud map of the rotor surface at preset time intervals. Two-dimensional magnetic field distribution cloud map Sound wave transmission efficiency Harmony of the sound wave spectrum centroid Input a flywheel digital twin model and correct the model parameters using the Kalman filter algorithm; Based on the predicted remaining flywheel lifespan and the health score, a graded response is performed, including: A comprehensive early warning index is calculated based on health score, predicted life expectancy, and health consistency index. : ; in, Rate your health; For health consistency index; To predict remaining lifespan; Design the flywheel for maximum remaining lifespan. ; when At that time, the warning level was Level 1; when At that time, the warning level was Level 2; when At that time, the warning level was Level 3; The graded response includes: When the warning level is Level 1, and When this happens, active vibration suppression is triggered, adjusting the PID parameters of the magnetic levitation controller to reduce the rotor vibration amplitude; When the warning level is Level 2, and the following conditions are met: , , Under any condition, reduce the speed setpoint to decrease rotor centrifugal stress and magnetic bearing load; When the warning level is Level 3, and the following conditions are met: , , Under any condition, an emergency shutdown is triggered, and the mechanical auxiliary bearing lock is activated to prevent rotor instability and collision, thus protecting the magnetic levitation system and the main body of the energy storage equipment.

2. The detection method for magnetic levitation flywheel energy storage devices according to claim 1, characterized in that, Local instantaneous stress on the flywheel surface is obtained using an NV color center quantum sensor array. and local instantaneous magnetic field strength ; based on and Generate a two-dimensional stress distribution cloud map on the rotor surface. and two-dimensional magnetic field distribution cloud map ,in, The coordinates of the rotor surface are rectangular coordinates. Sampling time; Ultrasonic guided waves inside the flywheel are obtained by using acoustic metasurface waveguides arranged radially along the rotor.

3. The detection method for magnetic levitation flywheel energy storage devices according to claim 2, characterized in that, The method further includes: The sound wave transmission efficiency is calculated based on the following formula. : ; in, The energy received by sound waves; The emitted energy of sound waves; The centroid of the sound wave spectrum is calculated based on the following formula. : ; in, For the first Frequency values ​​at each frequency point; For serial number; For frequency The corresponding sound wave amplitude; This represents the number of frequency points.

4. The detection method for magnetic levitation flywheel energy storage devices according to claim 1, characterized in that, The prediction of the remaining service life of the flywheel using the flywheel digital twin model includes: The crack propagation rate was simulated in the flywheel digital twin model based on the Paris formula: ; in, This represents the crack propagation rate. The length of the crack; The material constant for crack propagation; The stress intensity factor amplitude; The material index for crack propagation; This represents the number of loop iterations. The time taken for the crack length to reach the design critical value in the flywheel digital twin model is used to predict the remaining life of the flywheel. Output digital twin to predict remaining lifetime.

5. A detection system for magnetic levitation flywheel energy storage devices, characterized in that, The system includes: The data acquisition module is configured to acquire data from the flywheel; the flywheel data includes the flywheel's geometric parameters, material properties, operating parameters, magnetic field strength, local instantaneous stress on the flywheel surface, and ultrasonic guided waves inside the flywheel. A digital twin module is configured to construct a digital twin model of the flywheel based on the flywheel's data, and to predict the flywheel's remaining service life using the digital twin model. The construction of a digital twin model of the flywheel based on the flywheel data includes the following steps: Based on the flywheel data, define the flywheel digital twin model state vector: ; in, The state vector of the flywheel digital twin model; for The angular velocity of the magnetic levitation flywheel rotor at any given moment; for The electromagnetic force applied by the magnetic bearing at any given moment; The rotor surface temperature; for The spatial location of the flywheel predicted by the time-of-flight digital twin model Stress distribution at the location; for The spatial location of the flywheel predicted by the time-of-flight digital twin model The distribution of magnetic induction intensity at that location; Calculate the health consistency index to quantify the differences between virtual and real models: ； in, For the health consistency index, ; This is the actual flywheel state vector; The state vector of the flywheel digital twin model; when At that time, recalibrate the model until ; The system further includes: generating a two-dimensional stress distribution cloud map of the rotor surface at preset time intervals. Two-dimensional magnetic field distribution cloud map Sound wave transmission efficiency Harmony of the sound wave spectrum centroid Input a flywheel digital twin model and correct the model parameters using the Kalman filter algorithm; The scoring module is configured to input the flywheel data and the output of the flywheel digital twin model into a pre-trained federated learning model for health scoring; The response module is configured to provide a graded response based on the predicted remaining flywheel lifespan and the health score; Based on the predicted remaining flywheel lifespan and the health score, a graded response is performed, including: A comprehensive early warning index is calculated based on health score, predicted life expectancy, and health consistency index. : ; in, Rate your health; For health consistency index; To predict remaining lifespan; Design the flywheel for maximum remaining lifespan. ; when At that time, the warning level was Level 1; when At that time, the warning level was Level 2; when At that time, the warning level was Level 3; The graded response includes: When the warning level is Level 1, and When this happens, active vibration suppression is triggered, adjusting the PID parameters of the magnetic levitation controller to reduce the rotor vibration amplitude; When the warning level is Level 2, and the following conditions are met: , , Under any condition, reduce the speed setpoint to decrease rotor centrifugal stress and magnetic bearing load; When the warning level is Level 3, and the following conditions are met: , , Under any condition, an emergency shutdown is triggered, and the mechanical auxiliary bearing lock is activated to prevent rotor instability and collision, thus protecting the magnetic levitation system and the main body of the energy storage equipment.

Images