Flywheel data real-time monitoring method and system, electronic equipment and storage medium

Abstract

The invention relates to the technical field of flywheel monitoring, and discloses a flywheel data real-time monitoring method and system, electronic equipment and a storage medium. The method comprises the following steps: synchronously acquiring multi-dimensional dynamic parameters such as flywheel temperature gradient distribution and mechanical vibration spectrum through a multi-channel signal acquisition interface, and filtering through an adaptive noise suppression layer to generate a standardized data frame; a parallel computing architecture is adopted to complete feature decoupling and establish a dynamic incidence matrix, a three-dimensional spatio-temporal evolution model is constructed based on the matrix, and the abnormal evolution trajectory of each subsystem is accurately predicted. And generating a hierarchical early warning instruction set and starting a data snapshot, and meanwhile, rendering multi-dimensional data on a visual interface to form a monitoring instrument panel. A manual correction instruction is received through a bidirectional data channel, an early warning threshold value and model parameters are dynamically adjusted, the running state of the flywheel is comprehensively and accurately monitored, and stable running of a flywheel system is guaranteed.

Description

Technical Field

[0001] This invention relates to the field of flywheel monitoring technology, specifically to a method and system for real-time monitoring of flywheel data, electronic equipment, and storage media. Background Technology

[0002] As a core component in energy storage, power transmission, and attitude control, flywheels are widely used in key fields such as aerospace, new energy power generation, and industrial automation. The stability of their operation directly affects the safe and reliable operation of the entire system; therefore, real-time and comprehensive monitoring of the flywheel's condition is of significant practical importance. With the development of flywheel technology towards higher speeds and greater precision, the coupling degree of its operating parameters has significantly increased, making traditional monitoring methods insufficient to meet monitoring needs and presenting many unresolved technical limitations.

[0003] Existing technologies for flywheel data acquisition have limitations. They struggle to simultaneously acquire multiple key parameters reflecting the flywheel's operating status, resulting in insufficient data completeness and an inability to fully reflect the flywheel's true operating condition. Regarding noise processing, conventional filtering algorithms often employ fixed-parameter filtering modes, which are ill-suited to the dynamic changes in noise characteristics during flywheel operation. This can easily lead to distortion of useful signals or incomplete noise suppression, affecting the accuracy of subsequent data processing.

[0004] In the feature extraction and anomaly prediction stages, traditional methods are inefficient and fail to meet the needs of real-time monitoring. Furthermore, their poor feature separation of collected data results in insufficient sensitivity for identifying anomalous signals. In addition, existing anomaly prediction models exhibit poor dynamic adaptability, failing to effectively reflect the correlation between parameters of various flywheel subsystems and making it difficult to accurately predict anomaly evolution trajectories, often leading to delayed or false warnings.

[0005] Regarding monitoring visualization and parameter adjustment, existing monitoring systems primarily display single data points, lacking integrated presentation of multi-dimensional data. This makes it difficult for operators to quickly and intuitively grasp the flywheel's operating status. Furthermore, most systems lack flexible parameter adjustment mechanisms; once warning thresholds and model parameters are set, they cannot be dynamically corrected based on actual operating conditions, resulting in poor adaptability. These problems severely restrict the monitoring effectiveness of flywheel monitoring systems, increase the risk of flywheel system failures, and fail to meet the high reliability requirements of modern industry for flywheel equipment. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for real-time monitoring of flywheel data, an electronic device, and a storage medium to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides a method for real-time monitoring of flywheel data, the method comprising: Multidimensional dynamic parameters of the flywheel during operation are synchronously acquired through a multi-channel signal acquisition interface. These multidimensional dynamic parameters serve as the raw signals and include temperature gradient distribution, mechanical vibration spectrum, electromagnetic field fluctuation characteristics, and spatial attitude change. The acquired raw signal stream is processed by time-domain filtering through an adaptive noise suppression layer to generate purified standardized data frames. A parallel computing architecture is used to decouple the features of standardized data frames, separate the steady-state components and transient components, and establish a dynamic correlation matrix; A three-dimensional spatiotemporal evolution model is constructed based on a dynamic correlation matrix, and the abnormal evolution trajectory of each subsystem of the flywheel is predicted through model deduction. Based on the abnormal evolution trajectory and warning threshold, a hierarchical warning instruction set is generated and synchronously transmitted to the visual interactive interface. At the same time, a data snapshot mechanism is activated to save the current system status panorama and obtain full-dimensional data at the warning trigger time. Real-time data streams, early warning signals, and historical comparison curves are rendered synchronously in the visual interactive interface to form a multi-dimensional monitoring dashboard. Based on the warning trigger and full-dimensional data, a two-way data channel is established to receive manual correction instructions, and the warning threshold and model parameters are dynamically adjusted according to the information of the multi-dimensional monitoring dashboard.

[0008] Preferably, the step of synchronously acquiring multi-dimensional dynamic parameters of the flywheel during operation via a multi-channel signal acquisition interface includes: Connect the signal output terminal of the heterogeneous sensor array to the input terminal of the multi-channel signal acquisition interface, configure the sampling clock synchronization protocol, and enable the sensors to achieve nanosecond-level time alignment. The system receives raw sampling points from each sensor through a multi-channel signal acquisition interface. A circular buffer queue is designed to store the raw sampling points of each channel. When the queue reaches a preset depth, data packet reassembly is triggered. The data packet is a collection of raw sampling points of each channel after nanosecond-level time alignment. The reassembled data packets are then stamped with timestamps and device fingerprints to form standardized data blocks with identification features.

[0009] Preferably, the step of performing time-domain filtering on the acquired raw signal stream through an adaptive noise suppression layer includes: A hybrid dictionary containing an environmental noise database and a device feature database is constructed, and interference components in the original signal are decomposed through sparse coding. The local signal-to-noise ratio is calculated using a sliding window mechanism, and frequency domain notch or wavelet threshold denoising strategy is dynamically selected based on the signal-to-noise ratio result to denoise the signal. Baseline drift compensation is performed on the denoised signal segment to eliminate system errors caused by sensor zero-point offset and generate purified standardized data frames.

[0010] Preferably, the feature decoupling of standardized data frames using a parallel computing architecture includes: Deploy a CPU-GPU collaborative computing pipeline, with the standardized data frame as the processing object. The CPU thread pool is responsible for extracting time-domain statistical features, while the GPU stream processor performs frequency-domain transformation and mode decomposition to initially separate potential steady-state components and transient components. Establishing a shared memory swapping area enables synchronization of intermediate results between heterogeneous computing units, providing data collaboration support for feature decoupling; The coupling strength between steady-state and transient components is evaluated by mutual information entropy. When the entropy value exceeds the critical threshold, the recalculation process is initiated to ensure accurate separation of steady-state and transient components.

[0011] Preferably, the construction of the three-dimensional spatiotemporal evolution model based on dynamic correlation moments includes: The flywheel shell is discretized into a finite element mesh. The physical quantity observation values ​​of each mesh node are extracted from the dynamic correlation matrix and injected into the corresponding finite element mesh nodes to provide a spatialized data foundation for the three-dimensional spatiotemporal evolution model. Using finite element meshes of physical quantity observations as model input, the spatiotemporal convolution kernel parameters are trained using particle swarm optimization algorithm, enabling the model to capture the path characteristics of anomaly propagation. Based on the spatiotemporal convolution kernel parameters, a transfer learning framework is introduced to load a pre-trained material fatigue model, which accelerates the convergence speed of local hotspot regions.

[0012] Preferably, the generation of the hierarchical early warning instruction set includes: Define a three-level state space including minor anomalies, moderate risks, and serious failures. Each state corresponds to a different color code, and preset feature judgment thresholds for the abnormal evolution trajectory in each state to construct the judgment basis for hierarchical early warning. Based on the three-level state space and feature judgment threshold, a rule engine-based decision tree is developed. The abnormal evolution trajectory output by the three-dimensional spatiotemporal evolution model is input into the decision tree. By calculating the slope change rate of the trajectory and comparing it with the preset threshold, the warning level is automatically matched and the initial graded warning instruction is generated. The generated initial graded early warning instructions are graded, and the initial graded early warning instructions are output as the final graded early warning instruction set.

[0013] Preferably, establishing a bidirectional data channel includes: A dedicated topic channel is established in the message middleware as the core transmission carrier of the two-way data channel. The dedicated topic channel is used to receive manual correction instructions, which include manually labeled abnormal samples. Using the received manual correction instructions as the input source, an incremental learning mechanism is designed to integrate manual feedback into the parameter update process of the spatiotemporal evolution model and dynamically adjust the model parameters. The received manual correction instructions are classified and statistically analyzed by type. When three consecutive correction instructions of the same type are received, a model retraining task is automatically generated to improve the model iteration closed loop of the two-way data channel.

[0014] Preferably, the present invention further includes a flywheel data real-time monitoring system for implementing the flywheel data real-time monitoring method described above, the system comprising: The multi-channel signal synchronous acquisition module is used to synchronously acquire multi-dimensional dynamic parameters of the flywheel during operation through the multi-channel signal acquisition interface. The multi-dimensional dynamic parameters include temperature gradient distribution, mechanical vibration spectrum, electromagnetic field fluctuation characteristics and spatial attitude change. The adaptive noise suppression processing module is used to pass the acquired raw signal stream through the adaptive noise suppression layer for time-domain filtering to generate purified standardized data frames. The parallel computing and feature decoupling module is used to decouple features of standardized data frames using a parallel computing architecture, separating steady-state components from transient components and establishing a dynamic correlation matrix. The three-dimensional spatiotemporal evolution model construction module is used to construct a three-dimensional spatiotemporal evolution model based on a dynamic correlation matrix, and predict the abnormal evolution trajectory of each subsystem of the flywheel through model deduction; The graded early warning and data snapshot module is used to generate graded early warning instruction sets based on the abnormal evolution trajectory, and at the same time activate the data snapshot mechanism to save a panoramic view of the current system status; The multi-dimensional visualization and interaction module is used to synchronously render real-time data streams, early warning signals, and historical comparison curves in the visualization and interaction interface, forming a multi-dimensional monitoring dashboard. The two-way manual correction feedback module is used to establish a two-way data channel to receive manual correction instructions and dynamically adjust the early warning threshold and model parameters.

[0015] Preferably, the present invention further includes an electronic device, the device comprising a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the method described above.

[0016] Preferably, the present invention further includes a computer-readable storage medium storing computer instructions that, when executed by a processor, implement the entire process of the method described above.

[0017] Compared with the prior art, the beneficial effects of the present invention are: During the data acquisition phase, multi-dimensional dynamic parameters are synchronously acquired through a multi-channel signal acquisition interface, breaking the limitations of traditional single-channel acquisition. This allows for comprehensive capture of changes in various physical quantities during flywheel operation, ensuring the integrity and synchronization of data acquisition and providing rich and reliable foundational data for subsequent analysis. The application of an adaptive noise suppression layer adaptively adjusts filtering parameters based on the dynamic changes in noise characteristics during flywheel operation. While effectively suppressing noise interference, it maximizes the preservation of the original characteristics of useful signals, avoiding signal distortion problems caused by fixed-parameter filtering and significantly improving the quality of standardized data frames.

[0018] The adoption of a parallel computing architecture significantly improves the processing efficiency of feature decoupling, enabling rapid separation of steady-state and transient components and the establishment of a dynamic correlation matrix. This solves the problem of low efficiency in traditional serial computing and meets the data processing speed requirements of real-time monitoring. The three-dimensional spatiotemporal evolution model constructed based on the dynamic correlation matrix can fully explore the dynamic relationships between various parameters and accurately predict the abnormal evolution trajectories of each subsystem of the flywheel. Compared with traditional static models, it is more adaptable to the dynamic changes in the flywheel's operating state, identifies potential abnormal risks in advance, and buys time for timely countermeasures.

[0019] The combination of tiered early warning instruction set generation and data snapshot mechanism not only provides targeted early warning information based on the severity of anomalies, facilitating operators to adopt different handling strategies according to the actual situation, but also completely preserves a panoramic view of the system status at the time of the anomaly, providing complete raw data support for subsequent fault analysis and tracing. The multi-dimensional monitoring dashboard integrates real-time data streams, early warning signals, and historical comparison curves to present the flywheel's operating status in an intuitive way, reducing the monitoring difficulty for operators, helping to quickly grasp key information, and improving the efficiency of monitoring work.

[0020] The bidirectional data channel allows operators to input manual correction commands into the system based on actual operating conditions and experience, enabling dynamic adjustments to warning thresholds and model parameters. This enhances the system's flexibility and adaptability, allowing it to meet monitoring needs under different operating conditions and avoiding the insufficient adaptability issues of fixed-parameter modes. Overall, this method, through the synergistic effect of each link, achieves refined monitoring of the flywheel's operating status, effectively improving the accuracy of anomaly identification and the timeliness of warnings, helping to reduce the probability of failures and ensuring the long-term stable operation of the flywheel system. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the core process of the flywheel data real-time monitoring method; Figure 2 Temperature monitoring trend chart; Figure 3The flowchart is for adaptive noise suppression and time-domain filtering; Figure 4 This is a time-frequency joint analysis diagram of flywheel vibration monitoring; Figure 5 A flowchart for constructing a three-dimensional spatiotemporal evolution model. Detailed Implementation

[0022] 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, and 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.

[0023] Please see Figure 1 This invention provides a method for real-time monitoring of flywheel data. The method includes: synchronously acquiring multi-dimensional dynamic parameters of the flywheel during operation through a multi-channel signal acquisition interface. These multi-dimensional dynamic parameters specifically include temperature gradient distribution, mechanical vibration spectrum, electromagnetic field fluctuation characteristics, and spatial attitude changes. The acquired raw signal stream is first processed by an adaptive noise suppression layer for time-domain filtering. This process effectively separates environmental noise from inherent equipment noise, generating purified standardized data frames. Subsequently, the system employs a parallel computing architecture to efficiently decouple the features of these standardized data frames, separating the mixed signals into steady-state and transient components, and establishing a dynamic correlation matrix based on their inherent relationship. Based on this data, a three-dimensional spatiotemporal evolution model is constructed. This model can simulate the continuous changes in the physical state of the flywheel and predict the possible abnormal evolution trajectories of each subsystem through deduction. According to the model prediction results, the system generates a hierarchical early warning instruction set and simultaneously activates a data snapshot mechanism to completely save the current system state panorama. All real-time data streams, early warning signals, and historical data comparison curves are synchronously rendered on a visual interactive interface, forming an intuitive multi-dimensional monitoring dashboard. The system also establishes a two-way data channel, which can receive manual correction instructions from operators, thereby dynamically adjusting the warning threshold and model parameters to achieve continuous optimization of the monitoring system.

[0024] The aforementioned data snapshot mechanism saves a panoramic view of the current system state. Specifically, while the hierarchical early warning instruction set is generated, the data snapshot module is triggered to call the real-time output data of the multi-channel signal synchronous acquisition module, adaptive noise suppression processing module, parallel computing and feature decoupling module, and three-dimensional spatiotemporal evolution model construction module. This includes the original sampling points of multi-dimensional dynamic parameters at the current moment, the purified standardized data frames, the separated steady-state and transient components, the dynamic correlation matrix, and the prediction results of abnormal evolution trajectories. The above data is integrated into a panoramic data package of the system state according to a preset format and stored in the historical database, realizing the complete retention of all-dimensional data at the time of early warning triggering. The real-time data stream, early warning signals, and historical data are simultaneously rendered on the visual interactive interface. The curves are compared to form a multi-dimensional monitoring dashboard. Specifically, the multi-dimensional visualization interaction module obtains real-time multi-dimensional dynamic parameter data streams from the multi-channel signal synchronous acquisition module and renders them in curve form. It obtains a graded early warning instruction set from the graded early warning and data snapshot module and renders the early warning indicators according to preset color codes (yellow for minor anomalies, orange for moderate risks, and red for serious faults). It retrieves historical data under the same type of working conditions from the historical database and generates historical comparison curves. The real-time data stream curves, early warning indicators, and historical comparison curves are displayed synchronously in different functional areas of the same visualization interaction interface, forming a multi-dimensional monitoring dashboard that covers real-time monitoring, early warning prompts, and historical comparison functions, allowing operators to intuitively grasp the flywheel's operating status.

[0025] See Figure 2 The graph uses time (seconds) as the horizontal axis and temperature (°C) as the vertical axis to visually represent the dynamic changes of three key temperature parameters—casing temperature, bearing temperature, and motor temperature—over time during flywheel operation. Different curves in the graph correspond to these three temperature parameters: the motor temperature curve is generally in a higher range, showing a gradual upward trend over time, reflecting the cumulative temperature effect caused by energy loss during continuous motor operation. The bearing temperature curve remains between the motor temperature and the casing temperature, with a relatively gentle change, reflecting the stable temperature change of the bearing under lubrication and friction. The casing temperature curve is generally at a lower level, with the smallest change over time, consistent with the physical characteristic that the casing, as an external structure, is less affected by internal heat sources.

[0026] Example 1: See Figure 3The hardware foundation of the multi-channel signal acquisition interface is a heterogeneous sensor array. This array includes a high-precision thermocouple group for measuring temperature gradient distribution, a piezoelectric accelerometer cluster for capturing mechanical vibration spectra, a Hall effect sensor array for sensing electromagnetic field fluctuations, and a microelectromechanical system (MEMS) inertial measurement unit (IMU) for tracking spatial attitude changes. The high-precision thermocouple group, piezoelectric accelerometer cluster, Hall effect sensor array, and IMU inertial measurement unit output raw sampling point data for temperature gradient distribution, mechanical vibration spectrum, electromagnetic field fluctuation characteristics, and spatial attitude changes, respectively. The core of synchronous acquisition lies in the precise implementation of the sampling clock synchronization protocol. This protocol employs a hardware clock correction mechanism based on a precision clock synchronization protocol. The local clock of each sensor node undergoes periodic offset and drift compensation through the master clock source. This compensation process is completed at the physical layer using synchronization pulse signals, ensuring that the sampling times of the temperature sensor group, vibration accelerometer group, magnetic field detection unit, and inertial measurement device maintain nanosecond-level time alignment accuracy, thereby guaranteeing a strict correspondence between different physical quantities on the time axis. The raw sampling data output from each independent signal acquisition channel in the sensor array is used as a data stream. The temporary storage and management of the data stream relies on the design of a circular buffer queue. Each independent signal acquisition channel is allocated a first-in, first-out circular buffer, and the depth of the circular buffer is set to 2048 sampling points based on the sampling rate and data packet processing cycle. When the data filling volume of the circular buffer of any channel reaches the preset depth threshold, the hardware direct memory access controller triggers a high-priority interrupt signal, which notifies the data packet reassembly process to start. The data packet reassembly process reads the sampling points within the same time window from the circular buffers of all channels, packages these sampling points from different sensors into a data packet according to a predefined binary format, embeds a microsecond-level timestamp obtained from a global precision clock in the header of the data packet, and appends device fingerprint information containing sensor serial number, calibration coefficient version number, and cyclic redundancy check code to the tail of the data packet, ultimately forming a standardized data block with complete identification and time stamp.

[0027] The adaptive noise suppression layer's processing flow begins with the construction of a hybrid dictionary, which consists of two parts: an environmental noise library and a device feature library. The environmental noise library is constructed by extracting feature vectors from background noise and electromagnetic interference signals collected over a long period, including when the flywheel is stationary and in typical operating environments. The device feature library stores the inherent patterns of the flywheel and sensors under normal operating conditions, such as characteristic frequencies and harmonic components. The noise suppression process employs a sparse coding method, representing the input original signal segment as a linear combination of atomic vectors in the hybrid dictionary. An optimization algorithm is used to solve for the sparsest representation coefficients, thereby decomposing the signal into useful and interference components. The interference components mainly correspond to atoms in the dictionary related to environmental noise and inherent device noise. Dynamic evaluation of the local signal-to-noise ratio (SNR) is achieved through a sliding window mechanism. A Hanning window with 256 sampling points slides across the signal with a 50% overlap. For each signal segment within the window, its power spectral density is quickly calculated, and the ratio of the power of the main characteristic frequency component to the power of the remaining frequency bands is used as an estimate of the local SNR. Based on the real-time calculated signal-to-noise ratio (SNR), the system dynamically selects a denoising strategy: when the SNR is low and the interference exhibits narrowband characteristics, a frequency domain notch filter is activated. This filter is based on an adaptive lattice structure and can accurately form a deep stopband at the interference frequency; when the SNR is extremely low and the noise has broadband non-stationary characteristics, a threshold denoising strategy based on discrete wavelet transform is activated, and an improved Bayesian shrinkage rule is used to process the wavelet coefficients.

[0028] The baseline drift compensation stage targets system errors introduced by sensor zero-point offset and low-frequency interference. The compensation process employs a polynomial fitting algorithm. The algorithm estimates the baseline drift trend from the denoised signal segment, uses least squares to fit a cubic polynomial onto the signal segment, and the resulting smooth curve is considered the baseline drift estimate. This estimate is then subtracted from the original signal segment to obtain a baseline-stabilized and centered signal segment. All signal segments processed in the above steps are recombined to generate purified, standardized data frames that meet the requirements of subsequent processing formats. The data frame structure includes a frame header, valid data payload, and a frame tail check sequence. The entire adaptive noise suppression layer ensures that the data input to the feature decoupling module has high fidelity and low noise levels, creating conditions for subsequent accurate analysis. The synchronization, integrity, and purity of sensor data are the cornerstones of the reliable operation of the entire monitoring system. The above implementation details the entire process from physical signal acquisition to preliminary data preprocessing. The parameters and algorithm selection for each step have been rigorously considered to match the stringent requirements of the flywheel system for monitoring data quality. The depth of the annular buffer, the length of the sliding window, and the order of the polynomial fitting can all be adjusted in the system configuration file to adapt to the specific needs of different flywheel models or different working scenarios.

[0029] Example 2: The core of the CPU-GPU collaborative computing pipeline is the task partitioning and cooperation mechanism between the CPU thread pool and the GPU stream processors. The CPU thread pool is created during system initialization and contains a set of worker threads, the number of which matches the number of physical cores in the CPU. These worker threads are specifically responsible for processing the time-domain statistical feature extraction task of standardized data frames, with each thread independently processing data from one or more channels. The extracted time-domain features include, but are not limited to, the effective value, average amplitude, root square amplitude, kurtosis index, peak factor, impulse factor, and waveform factor of the signal. Calculating these features involves performing arithmetic operations such as windowing, summation, squaring, and square root extraction on the data frames. These operations have complex logic control but relatively low computational intensity, making them suitable for execution by the CPU's general-purpose computing cores. The numerous stream processors on the GPU are organized into multiple stream multiprocessors, forming a massively parallel computing engine. Stream processors are responsible for executing computationally intensive frequency domain transformations and mode decomposition algorithms. For example, they perform Fast Fourier Transform (FFT) on data frames to obtain the power spectrum, execute Short-Time Fourier Transform (SFT) to observe the spectrum's changes over time, or use Empirical Mode Decomposition (EMD) algorithms to adaptively decompose non-stationary signals into a series of intrinsic mode functions (EMFs). The single-instruction, multi-threaded architecture of graphics processing units (GPUs) allows the same transformation operations to be performed simultaneously on thousands of data points within a data frame, significantly accelerating the extraction of frequency domain features.

[0030] The shared memory swap area is a key component for efficient data interaction between the CPU and GPU. It allocates a region in host memory that is shared and accessed by all computing units. This region is divided into multiple buffers to temporarily store time-domain feature vectors calculated by the CPU, frequency-domain feature matrices calculated by the GPU, and eigenmode function components of each order after mode decomposition. Data synchronization is achieved through a memory-mapped file mechanism. After the CPU thread writes the calculation results to a designated buffer in shared memory, it updates a corresponding status flag. The GPU kernel polls this flag before startup; when new data is detected, it directly reads the data from shared memory for processing, and vice versa. This design avoids the overhead of frequent data copying between host memory and device memory, achieving low-latency synchronization and exchange of intermediate results between heterogeneous computing units. Mutual information entropy calculation is a quantitative method for evaluating the coupling strength between steady-state and transient components. The steady-state and transient component sequences required for mutual information entropy calculation rely on the collaborative computation output of the CPU and GPU for their raw data processing. Steady-state components typically refer to the slowly changing, highly periodic parts of a signal, while transient components represent the sudden, non-periodic parts. Calculating mutual information entropy requires first estimating the joint probability distribution and individual marginal probability distributions of the steady-state and transient component sequences. The estimation process uses histogram methods or kernel density estimation to discretize continuous data values ​​into multiple intervals. Based on these probability distributions, the entropy of the steady-state component, the entropy of the transient component, and their joint entropy are calculated using the definition of Shannon entropy. Mutual information entropy is the sum of the steady-state and transient component entropies minus the joint entropy. The result is a non-negative real value; a larger mutual information entropy value indicates a stronger statistical dependence between the two components, i.e., a higher degree of coupling.

[0031] The system presets a critical entropy threshold, which is determined based on the statistical quantiles of the mutual information entropy of steady-state and transient components under a large amount of historical normal data. When the mutual information entropy calculated for the current data frame exceeds the critical threshold, the feature decoupling module triggers a recalculation process. This recalculation process does not simply repeat the previous decomposition algorithm; instead, it employs a more advanced and computationally complex variational mode decomposition algorithm. The variational mode decomposition algorithm constructs and solves a variational problem to adaptively determine the center frequency and bandwidth of each eigenmode function. This more effectively decomposes the signal corresponding to the normalized data frame generated after adaptive noise suppression into narrowband components with practical physical meaning, thus potentially separating strongly coupled mixed modes. The new decomposition result generated by the recalculation process replaces the initial decoupling result and is used to update the dynamic correlation matrix. The dynamic correlation matrix is ​​a square matrix where rows and columns correspond to different feature components, and the values ​​of the matrix elements represent the correlation coefficient or mutual information between corresponding component pairs. This matrix quantitatively characterizes the dynamic correlation relationships between different physical processes within the flywheel system.

[0032] The application of the variational mode decomposition algorithm in the recomputation process relies on a parallel computing architecture: A CPU-GPU collaborative computing pipeline serves as the execution vehicle. The CPU thread pool first preprocesses the current data frame, completing data format conversion and block partitioning, and then transmits the partitioned data to the GPU stream processor. Based on the core logic of the variational mode decomposition algorithm, the GPU stream processor first constructs a variational problem including data fidelity and regularization terms. Through an iterative optimization process, it adaptively searches for the optimal center frequency and bandwidth of each intrinsic mode function, ensuring that each narrowband component corresponds to a specific physical process in the flywheel's operation. After decomposition, the GPU transmits each intrinsic mode function component to the shared memory swap area. The CPU thread pool reads data from the shared memory swap area, verifies the validity of each component, and obtains the separated steady-state and transient components, which is the new decomposition result generated by the recomputation process. The new decomposition result will replace the steady-state and transient components obtained from the initial decoupling, and be used to recalculate the correlation coefficient or mutual information between each characteristic component, thereby updating the dynamic correlation matrix and ensuring that the dynamic correlation matrix can accurately characterize the dynamic correlation relationship between different physical processes inside the flywheel system.

[0033] When the flywheel is running at high speed, a vibration acceleration sensor mounted on the drive-end bearing housing detects an abnormal high-frequency signal mixed with strong rotational frequency vibration and electromagnetic noise. The raw voltage signal is converted into a digital sequence by a data acquisition card, forming a standardized data frame containing 2048 consecutive vibration acceleration sampling points at a sampling frequency of 51.2 kHz. This data frame is then sent to the parallel computing pipeline of the feature decoupling module. A worker thread in the CPU thread pool immediately takes over the data frame, performing windowing processing to reduce spectral leakage and calculating a series of time-domain statistical features. These features include the effective value, peak value, kurtosis index, waveform factor, and impulse factor of the vibration acceleration. The calculated feature values ​​are organized into a time-domain feature vector. The entire data frame is transmitted to the global video memory of the graphics processor via the PCIe bus. A CUDA core on the graphics processor is launched, with 256 threads configured as a thread block. Multiple such thread blocks execute the Fast Fourier Transform algorithm in parallel on the streaming multiprocessor. Each thread processes a subset of data points within a data frame, collaboratively transforming the signal from the time domain to the frequency domain to generate a complex spectral sequence. Another dedicated kernel calculates the magnitude of this spectral sequence, obtaining a power spectral density estimate that clearly shows the energy distribution of the signal across its various frequency components. A third kernel attempts to perform empirical mode decomposition (EMD), decomposing the non-stationary vibration signal into a series of intrinsic mode function (EMF) components arranged from high to low frequencies. The calculated frequency domain feature matrix and EMF components are copied from the graphics processor's video memory back to host memory and written to a buffer specifically reserved for frequency domain data in the shared memory swap area. After completing time-domain feature extraction, the CPU thread also writes the time-domain feature vector to another designated location in the shared memory swap area. A synchronization management process monitors the buffer's status flags, and once the time-domain and frequency-domain features are ready, it notifies the feature coupling analysis routine to begin work. The feature coupling analysis routine reads the time-domain features representing the overall signal energy from shared memory as a proxy for the steady-state components, and simultaneously reads the first-order EMF components obtained from EMD as a proxy for the transient components. The analysis routine uses the histogram method to estimate the joint probability distribution and individual marginal probability distributions of the two surrogate sequences. Based on these distributions, the entropy of the steady-state component sequence, the entropy of the transient component sequence, and the joint entropy of the two are calculated. The mutual information entropy is calculated using a formula, and the result is a dimensionless numerical value. The mutual information entropy is based on the entropy of the steady-state component sequence. Entropy of transient component sequences and their joint entropy The calculation formula is as follows:

[0034] Joint Entropy The formula for calculating the uncertainty of the joint distribution of steady-state and transient components is as follows:

[0035] in: It is a steady-state component sequence Data points fall into Interval and transient component sequence Data points fall into Joint probability of intervals, These are the entropy of the steady-state component sequence and the entropy of the transient component sequence, as defined above.

[0036] Assuming the calculated mutual information entropy value is significantly higher than the critical threshold obtained from historical normal data statistics, this result indicates a strong nonlinear coupling between the steady-state energy level of the bearing vibration signal and the high-frequency transient impact. Simple empirical mode decomposition (EMD) may not have effectively separated the independent fault features. The system triggers a recalculation process. The recalculation process calls a more computationally intensive but more accurate variational mode decomposition (VMD) algorithm to replace the previous EMD. The VMD algorithm iteratively optimizes to find a set of eigenmode functions with optimal center frequency and bandwidth, enabling clearer extraction of the weak impact components near the bearing outer ring fault characteristic frequency. The newly decomposed components are again used to calculate the mutual information entropy with the steady-state features. This time, the entropy value is lower than the critical threshold, indicating that the decomposition effect meets the requirements. Based on the VMD results and time-domain and frequency-domain features, the dynamic correlation matrix is ​​updated. The matrix clearly shows a strong correlation between the high-frequency impact components and the bearing housing temperature rise.

[0037] See Figure 4 In the real-time monitoring method for flywheel data, the multi-dimensional analysis of vibration signals relies on a joint time-frequency domain processing flow. Specifically, the raw vibration signal is acquired through a multi-channel synchronous acquisition interface, forming time-domain waveform data. During signal preprocessing, the Hanning window function is applied for windowing optimization, effectively suppressing spectral leakage and improving frequency resolution. The frequency domain transformation stage employs a GPU parallel computing architecture to perform a fast Fourier transform, generating a high-precision power spectral density estimate, thereby accurately extracting the frequency components of fault characteristics. The core parameters of the time-frequency domain processing, designed to achieve multi-dimensional analysis of vibration signals and accurate extraction of fault characteristics, are configured as key parameters. These key parameters include a time-domain sampling window covering 0-40ms (including two frequency cycles), a frequency band analysis range set to 0-5kHz, a fundamental frequency of 50Hz, and the bearing fault characteristic frequency locked at 2500Hz. Figure 4The paper uses a dual coordinate system to visually present the correspondence between time-domain waveform distortion and frequency-domain energy concentration, verifying the separation effect of steady-state and transient components during feature decoupling. The parameterized design of the entire process ensures sensitive detection of early faults, including the selection of window function type, spectrum calculation step size, and fault frequency threshold setting, all of which match the requirements for constructing the dynamic correlation matrix.

[0038] Example 3: See Figure 5 The construction of the three-dimensional spatiotemporal evolution model begins with the geometric discretization of the flywheel structure. Key components such as the flywheel shell, shaft, and bearings are meshed using computer-aided engineering software, generating a finite element mesh model composed of a large number of finite elements. Each finite element contains basic information such as node coordinates, element type, and material properties. The mesh model accurately represents the three-dimensional geometry and physical structure of the flywheel. During model operation, the observation values ​​of multi-dimensional dynamic parameters such as temperature, vibration acceleration, and strain acquired by the multi-channel signal acquisition system are mapped and injected into the corresponding nodes of the finite element mesh according to the actual installation positions of the sensors on the flywheel. Each node stores a state vector that changes over time, containing physical field variables such as temperature, vibration displacement, and stress components at that location, thus constructing a three-dimensional data field in digital space that evolves synchronously with the physical flywheel. The spatiotemporal convolution kernel is the core mathematical tool for the model to simulate anomaly propagation. The total number of nodes in the finite element mesh is the core data dimension foundation for the spatiotemporal convolution kernel to achieve parameter optimization and anomaly evolution trajectory prediction. The convolution kernel is a function with specific weight distributions in both spatial and temporal dimensions. Spatial convolution kernels define the extent and degree to which a node is influenced by its neighboring nodes, while temporal convolution kernels characterize the inertia or delay effect of state variables evolving over time. Particle swarm optimization (PSO) is used to train the unknown parameter set in this complex spatiotemporal convolution kernel. The optimization process initializes a swarm of "particles," each representing a candidate solution for the convolution kernel parameters. The particles fly through the parameter space, updating their velocity and position by iteratively tracking their individual and swarm historical best positions. The optimization objective is to find a set of parameters that minimizes the difference between the model's predicted state and the actual observed data, typically quantified by a loss function.

[0039] in: This represents the set of parameters that need to be optimized in the spatiotemporal convolution kernel, including weights and bias terms. This represents the total number of time steps used for training. Indicates the total number of nodes in the finite element mesh. (Subscript) Identify a specific node in the grid. Indicates in parameter set Next, node At any moment The model predicts the state vector. Represents a node At any moment The actual observed state vector. It is to assign nodes The weighting coefficients are used to emphasize the prediction accuracy of key areas (such as known weak points). The norm of a vector is used to measure the difference between predicted and observed values. Loss function. The dimensions depend on the state vector The dimensions of the vector are consistent with the square of the dimensions of the state vector, since both sides of the formula are the sum of squares of the norms.

[0040] The transfer learning framework and the spatiotemporal convolution kernel parameters work together to accurately predict the abnormal evolution trajectory of a flywheel in a 3D spatiotemporal evolution model. From the perspective of parameter initialization, the pre-trained material fatigue model loaded by the transfer learning framework has learned the general laws of crack initiation and propagation in metals and composite materials under cyclic loading. The material property-related parameters it contains can be used as initial reference values ​​for the spatiotemporal convolution kernel parameters, providing initial configuration for parameters in the spatiotemporal convolution kernel that involve material physical behavior (such as field variable coupling coefficients and spatial influence weights), avoiding the need to train these parameters from random values ​​and reducing the blindness of parameter optimization. From the perspective of optimization efficiency, the introduction of the transfer learning framework can accelerate the convergence speed of spatiotemporal convolution kernel parameters, especially for local hot spots in flywheels (such as finite element mesh nodes with high stress and high temperature gradients): Since the pre-trained model has mastered the general material fatigue behavior law, when the spatiotemporal convolution kernel adjusts the parameters using flywheel monitoring data, it does not need to relearn the basic material properties. It can directly focus on the propagation law of nodal physical quantities of specific flywheel structures (such as the rim and bearing connection area), quickly adjust the parameters to adapt to the actual working state of the flywheel, and enable the spatiotemporal convolution kernel to reach the optimal parameter configuration more quickly. This improves the prediction response speed of the three-dimensional spatiotemporal evolution model to the abnormal evolution trajectory of local hot spots, while ensuring that the optimized spatiotemporal convolution kernel parameters can more accurately characterize the spatiotemporal evolution relationship of physical quantities between finite element mesh nodes, providing reliable support for the model to infer abnormal evolution trajectories.

[0041] The introduction of the transfer learning framework aims to accelerate the convergence speed of the model's prediction of material fatigue behavior in local hotspot regions. The framework loads a pre-trained sub-model on a large-scale general-purpose materials fatigue experimental dataset. This pre-trained model has learned the general laws governing crack initiation and propagation in metals and composite materials under cyclic loading. In the flywheel monitoring scenario, the 3D spatiotemporal evolution model is not trained from scratch, but rather uses the parameters of the pre-trained model as initial values ​​and fine-tunes them using flywheel monitoring data. This transfer learning strategy enables the model to utilize knowledge gained from numerous experiments to quickly adapt to the actual operating state of the flywheel, significantly improving the prediction accuracy and response speed for abnormal evolution in local stress concentration areas. At each time step, the model receives the latest sensor data, updates the state vectors of each node, and uses optimized spatiotemporal convolution kernels to infer the system state for several future time steps, thereby generating predictions of the abnormal evolution trajectories of each subsystem of the flywheel, providing a forward-looking basis for graded early warning.

[0042] During continuous high-load operation of the flywheel, a temperature sensor mounted on the edge of the composite rim reported a slow but steadily rising temperature reading, while a nearby vibration sensor detected a subtle spectral change. A feature decoupling module separated these signals and identified a dynamic correlation between the temperature rise and specific high-frequency vibration components. This data was then fed into a three-dimensional spatiotemporal evolution model. The model utilizes the flywheel's finite element mesh, composed of hundreds of thousands of hexahedral elements, meticulously describing the flywheel's rim, spokes, and hub structure. Each element node stores multiple physical field variables, including temperature, stress, and displacement fields. The locations of the reporting temperature and vibration sensors within the mesh are precisely pinpointed, and their readings are assigned to corresponding nodes as initial observations for model evolution. The model's core spatiotemporal convolution kernel then begins operation. Spatially, the kernel defines an influence decay function, with mesh points closer to the anomalous node receiving greater influence weights. Temporally, the kernel incorporates a memory function, assigning higher weights to more recent observations. The parameters of the convolution kernel, including the spatial radius of influence, the time decay constant, and the coupling coefficients between field variables, constitute the set of parameters that need to be optimized. A particle swarm optimization algorithm is initiated to find the optimal convolution kernel parameters. The algorithm initializes hundreds of particles, each representing a random combination of convolution kernel parameters. Each particle uses its own parameter combination to drive the model to predict the overall temperature and vibration field distribution of the flywheel over a future period. The prediction results are compared with subsequent actual sensor readings, and the calculated prediction error is used as the particle's fitness value. Particles fly in the parameter space, each recording its historical best position, and the entire particle swarm shares the global best position. By iteratively updating the particle velocity and position, the particle swarm gradually converges towards the parameter region that minimizes the prediction error. After multiple iterations, the parameter combination corresponding to the global best position is identified as the current optimal solution and used to configure the spatiotemporal convolution kernel. A transfer learning framework plays a role in this process, loading a pre-trained material thermo-mechanical coupling response model obtained from fatigue thermal imaging experiments on various composite material samples. When the flywheel model identifies the current hotspot region as being located at the rim of the composite material, knowledge from the pre-trained model regarding heat generation, heat conduction, and hotspot evolution caused by damage in the composite material under cyclic stress is transferred. This transfer is not a simple copy; rather, it involves initializing the relevant sub-network parameters in the flywheel model and fine-tuning them on flywheel-specific data. This allows the model to more quickly capture subtle features of anomalous thermal behavior in composite material regions, avoiding the need to learn material properties from scratch. The spatiotemporal evolution model, configured with optimized parameters, begins its inference and prediction. Starting with the grid state of the currently injected sensor data, the model applies spatiotemporal convolution kernels at each time step.Convolutional kernel operations simulated the heat transfer process from the hot spot area at the rim edge towards the hub, as well as the effect of vibration energy transfer from the source point through the structure. Model simulations showed that, without intervention, the current small-scale temperature anomaly would gradually expand and affect the interface between the rim and the metal spokes within the next fifteen minutes, where the temperature difference could lead to thermal stress concentration. The vibration model predicted a trend of anomalous vibration components towards frequency modulation, suggesting a potential dynamic equilibrium problem. These simulation results—the expansion path of the temperature field and the evolution trajectory of the vibration modes—were synthesized into the anomalous evolution trajectories of each subsystem of the flywheel. These trajectories were visualized and overlaid on the 3D flywheel model, with key risk areas and predicted time points marked.

[0043] Example 4: The generation logic of the graded early warning instruction set is built upon a well-defined state space, which contains three discrete levels, corresponding to different health levels of the flywheel system. The minor anomaly level indicates that the system has detected parameter fluctuations deviating from the normal baseline, but the fluctuation amplitude has not yet constituted a functional risk, and the flywheel can still operate safely. This level is visually coded in yellow on the monitoring interface. The moderate risk level indicates that one or more key parameters continue to deviate and the trend is deteriorating, system performance may have been affected, and there is a potential possibility of failure development. Maintenance personnel need to pay attention and prepare intervention measures. This level is marked in orange. The severe fault level means that the monitoring data strongly predicts that a component functional failure is about to occur or has already occurred, such as bearing jamming or rotor friction. Immediate emergency measures such as shutdown and inspection must be taken. This level is displayed in red, the highest warning level.

[0044] Referring to Table 1, the rule-based decision tree is the core reasoning mechanism for automatically assessing early warning levels. The decision tree uses the abnormal evolution trajectory output by the three-dimensional spatiotemporal evolution model as input features for logical judgment. The root node of the decision tree determines the number of abnormal parameters. The left and right branches are two independent evaluation branches set in the rule-based decision tree for different scenarios with varying numbers of abnormal parameters. If it is a single-parameter anomaly, it enters the left branch for in-depth evaluation; if it is a multi-parameter correlated anomaly, it enters the right branch for comprehensive evaluation. In the single-parameter evaluation branch, the decision node sequentially checks whether the abnormal amplitude exceeds the first-level threshold, whether the abnormal duration is greater than the time window, and whether the rate of change of the abnormal growth slope is positive and accelerating. In the multi-parameter correlated evaluation branch, the decision node checks the spatial diffusion rate of the abnormal region and whether the covariance between different physical quantity anomaly indices is significantly higher than the normal correlation level. Each leaf node of the decision tree corresponds to a specific early warning level output. The judgment conditions on the decision path together constitute the mapping rule set from data features to early warning status.

[0045] Table 1: Mapping Relationship Between Warning Levels and Response Measures

[0046] The reliability of high-level early warnings is enhanced through a data verification mechanism using backup sensor channels. When the decision tree inference result triggers a medium-risk or severe fault warning, the warning generation module sends an instruction to the sensor management unit. The instruction includes the type of abnormal parameter, the approximate time range of the abnormality, and the identifier of the backup channel to be activated. Upon receiving the instruction, the sensor management unit immediately activates the backup sensor data acquisition link for the corresponding physical quantity. Backup sensors are typically installed at a physical distance from the main sensor or employ different measurement principles to obtain independent observation samples. The raw data acquired by the backup channel undergoes a simplified and rapid processing flow to extract key features and compare them with the abnormal features of the main channel. The verification consistency judgment logic requires that the backup channel data must also detect significant abnormal features, and the abnormal trend must be consistent with the main channel's judgment. Only when the consistency condition is met will the corresponding high-level early warning instruction be formally sent to the monitoring terminal, log database, and relevant alarm interfaces. If the verification result differs significantly from the main channel's judgment, the system generates a pending early warning state and triggers a diagnostic task, notifying technicians to check the sensor system. This mechanism effectively prevents false alarms caused by a single sensor failure or transient interference. The early warning instruction set itself is a structured data object, containing an early warning level code, a trigger timestamp, a list of associated abnormal parameters, a summary of key feature values ​​of the abnormal evolution trajectory, and the encoding of recommended handling measures. The instruction set is distributed via message queues to the monitoring interface for visualization, persistently stored in the database, and may be sent to the upper-level management system via communication interfaces. The entire tiered early warning process automates the transformation from data characteristics to operational actions. The clarity of the rules ensures the transparency and interpretability of early warning decisions, while the introduction of a review mechanism enhances the robustness of the system's decisions at critical moments.

[0047] Example 5: The physical and logical implementation of the bidirectional data channel relies on a message middleware system, with an Apache Kafka message queue cluster deployed in the system architecture. A dedicated topic channel named "human_feedback_topic" is established within the cluster. This topic channel is configured in persistent mode and has a message acknowledgment mechanism enabled to ensure reliable transmission of manual correction instructions. Interactive components on the monitoring interface, such as confirmation buttons, category dropdown menus, and threshold sliders, have their operation events monitored and encapsulated into structured JSON messages. The JSON message body includes the instruction type, operation timestamp, target data segment identifier, and corrected parameter value or category label. The encapsulated messages are continuously sent to the "human_feedback_topic" topic channel via the producer API, awaiting processing by the background consumer service. The incremental learning mechanism is the core algorithm module for digesting and absorbing human feedback. The design goal of the mechanism is to seamlessly integrate discrete human intervention knowledge into the existing spatiotemporal evolution model without needing to discard the original model for complete retraining. The consumer service pulls manual correction instruction messages from the message queue, parses them, and pairs them with the corresponding data snapshots to form new training sample pairs. The incremental learning process employs online sequence learning or mini-batch gradient descent algorithms. The parameter updates of the spatiotemporal evolution model are no longer based on the initial complete historical dataset, but rather on newly arrived samples and a potentially decaying regularization term for the old model parameters. For each new sample, the algorithm calculates the gradient of the loss function between the model's predicted output with the current parameters and the manually labeled true values, and fine-tunes the model parameters along the inverse direction of the gradient with a small learning rate. This learning approach allows the model to progressively adjust its internal representation, making its predictive behavior more consistent with the judgment of operations experts, while avoiding forgetting general patterns learned from massive amounts of data due to an excessively large learning rate.

[0048] The automatic generation logic for model retraining tasks is based on pattern recognition of the feedback instruction stream. The system maintains a sliding time window, the length of which is configurable, for example, 24 hours. Within the window, the system performs cluster analysis on each received manual correction instruction. Instructions are feature-extracted and vectorized based on the anomaly type they point to, the sensor parameters involved, and the type of correction operation. When the number of instructions in a cluster accumulates to a preset threshold within the time window, and these instructions all point to the same correction operation for the same type of anomaly, the system determines that the current model has a systematic bias in recognizing or predicting this type of anomaly. At this point, the task scheduler automatically creates a high-priority model retraining task in the background task queue. The retraining task is not a simple incremental learning process. The task retrieves all recent data related to this type of anomaly from the historical database, including raw sensor readings, automatic warning records, and all relevant manual correction labels, forming a targeted training dataset. The task initiates an independent training process, using this dataset to perform a centralized, deeper parameter optimization of the spatiotemporal evolution model. This process may last from several minutes to several hours, depending on the amount of data and the complexity of the model. After retraining, the generated new model version will undergo a simplified offline evaluation process. Once the evaluation is passed, it will be deployed online to replace the old model, thus completing an iterative upgrade of the model.

[0049] The human feedback processing flow and model update mechanism together form a closed-loop learning system. Message middleware ensures the asynchronous and reliable transmission of instructions, incremental learning enables daily fine-tuning and adaptation of the model, and trigger-based retraining tasks address systematic and persistent biases in the model. This design transforms the monitoring system from a static, rule-based executor into an intelligent agent capable of continuous learning and evolution from human expert experience. The two-way channel not only transmits instructions but also, in essence, transmits knowledge, which is encoded into the model update parameters. This allows the system's monitoring accuracy and early warning reliability to gradually improve with extended use.

[0050] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0051] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for real-time monitoring of flywheel data, characterized in that, include: Multidimensional dynamic parameters of the flywheel during operation are synchronously acquired through a multi-channel signal acquisition interface. These multidimensional dynamic parameters serve as the raw signals and include temperature gradient distribution, mechanical vibration spectrum, electromagnetic field fluctuation characteristics, and spatial attitude change. The acquired raw signal stream is processed by time-domain filtering through an adaptive noise suppression layer to generate purified standardized data frames. A parallel computing architecture is used to decouple the features of standardized data frames, separate the steady-state components and transient components, and establish a dynamic correlation matrix; A three-dimensional spatiotemporal evolution model is constructed based on a dynamic correlation matrix, and the abnormal evolution trajectory of each subsystem of the flywheel is predicted through model deduction. Based on the abnormal evolution trajectory and warning threshold, a hierarchical warning instruction set is generated and synchronously transmitted to the visual interactive interface. At the same time, a data snapshot mechanism is activated to save the current system status panorama and obtain full-dimensional data at the warning trigger time. Real-time data streams, early warning signals, and historical comparison curves are rendered synchronously in the visual interactive interface to form a multi-dimensional monitoring dashboard. Based on the warning trigger and full-dimensional data, a two-way data channel is established to receive manual correction instructions, and the warning threshold and model parameters are dynamically adjusted according to the information of the multi-dimensional monitoring dashboard.

2. The flywheel data real-time monitoring method according to claim 1, characterized in that, The method of synchronously acquiring multi-dimensional dynamic parameters of the flywheel during operation through a multi-channel signal acquisition interface includes: Connect the signal output terminal of the heterogeneous sensor array to the input terminal of the multi-channel signal acquisition interface, configure the sampling clock synchronization protocol, and enable the sensors to achieve nanosecond-level time alignment. The system receives raw sampling points from each sensor through a multi-channel signal acquisition interface. A circular buffer queue is designed to store the raw sampling points of each channel. When the queue reaches a preset depth, data packet reassembly is triggered. The data packet is a collection of raw sampling points of each channel after nanosecond-level time alignment. The reassembled data packets are then stamped with timestamps and device fingerprint information to form standardized data blocks with identification features.

3. The flywheel data real-time monitoring method according to claim 2, characterized in that, The step of performing time-domain filtering on the acquired raw signal stream through an adaptive noise suppression layer includes: A hybrid dictionary containing an environmental noise database and a device feature database is constructed, and interference components in the original signal are decomposed through sparse coding. The local signal-to-noise ratio is calculated using a sliding window mechanism, and frequency domain notch or wavelet threshold denoising strategy is dynamically selected based on the signal-to-noise ratio result to denoise the signal. Baseline drift compensation is performed on the denoised signal segment to eliminate the systematic error caused by the sensor zero-point offset and generate a purified standardized data frame.

4. The flywheel data real-time monitoring method according to claim 3, characterized in that, The feature decoupling of standardized data frames using a parallel computing architecture includes: Deploy a CPU-GPU collaborative computing pipeline, with the standardized data frame as the processing object. The CPU thread pool is responsible for extracting time-domain statistical features, while the GPU stream processor performs frequency-domain transformation and mode decomposition to initially separate potential steady-state components and transient components. Establishing a shared memory swapping area enables synchronization of intermediate results between heterogeneous computing units, providing data collaboration support for feature decoupling; The coupling strength between steady-state and transient components is evaluated by mutual information entropy. When the entropy value exceeds the critical threshold, the recalculation process is initiated to ensure accurate separation of steady-state and transient components.

5. The flywheel data real-time monitoring method according to claim 4, characterized in that, The three-dimensional spatiotemporal evolution model constructed based on the dynamic correlation matrix includes: The flywheel shell is discretized into a finite element mesh. The physical quantity observation values ​​of each mesh node are extracted from the dynamic correlation matrix and injected into the corresponding finite element mesh nodes to provide a spatialized data foundation for the three-dimensional spatiotemporal evolution model. Using finite element meshes of physical quantity observations as model input, the spatiotemporal convolution kernel parameters are trained using particle swarm optimization algorithm, enabling the model to capture the path characteristics of anomaly propagation. Based on the spatiotemporal convolution kernel parameters, a transfer learning framework is introduced to load a pre-trained material fatigue model, which accelerates the convergence speed of local hotspot regions.

6. The flywheel data real-time monitoring method according to claim 5, characterized in that, The generated hierarchical early warning instruction set includes: Define a three-level state space including minor anomalies, moderate risks, and serious failures. Each state corresponds to a different color code, and preset feature judgment thresholds for the abnormal evolution trajectory in each state to build the judgment basis for hierarchical early warning. Based on the three-level state space and feature judgment threshold, a rule engine-based decision tree is developed. The abnormal evolution trajectory output by the three-dimensional spatiotemporal evolution model is input into the decision tree. By calculating the slope change rate of the trajectory and comparing it with the preset threshold, the warning level is automatically matched and the initial graded warning instruction is generated. The generated initial graded early warning instructions are graded, and the initial graded early warning instructions are output as the final graded early warning instruction set.

7. The flywheel data real-time monitoring method according to claim 6, characterized in that, The establishment of the bidirectional data channel includes: A dedicated topic channel is established in the message middleware as the core transmission carrier of the two-way data channel. The dedicated topic channel is used to receive manual correction instructions, which include manually annotated anomaly samples. Using the received manual correction instructions as the input source, an incremental learning mechanism is designed to integrate manual feedback into the parameter update process of the spatiotemporal evolution model and dynamically adjust the model parameters. The received manual correction instructions are classified and statistically analyzed by type. When three consecutive correction instructions of the same type are received, a model retraining task is automatically generated to improve the model iteration closed loop of the two-way data channel.

8. A flywheel data real-time monitoring system, used to implement the flywheel data real-time monitoring method as described in any one of claims 1-7, characterized in that, The system includes: The multi-channel signal synchronous acquisition module is used to synchronously acquire multi-dimensional dynamic parameters of the flywheel during operation through the multi-channel signal acquisition interface. The multi-dimensional dynamic parameters are used as the raw signals, including temperature gradient distribution, mechanical vibration spectrum, electromagnetic field fluctuation characteristics and spatial attitude change. The adaptive noise suppression processing module is used to pass the acquired raw signal stream through the adaptive noise suppression layer for time-domain filtering to generate purified standardized data frames. The parallel computing and feature decoupling module is used to decouple features of standardized data frames using a parallel computing architecture, separating steady-state components from transient components and establishing a dynamic correlation matrix. The three-dimensional spatiotemporal evolution model construction module is used to construct a three-dimensional spatiotemporal evolution model based on a dynamic correlation matrix, and predict the abnormal evolution trajectory of each subsystem of the flywheel through model deduction; The graded early warning and data snapshot module is used to generate graded early warning instruction sets based on the abnormal evolution trajectory and early warning thresholds and transmit them synchronously to the visual interactive interface. At the same time, it activates the data snapshot mechanism to save the current system status panorama and obtain full-dimensional data at the time of early warning triggering. The multi-dimensional visualization and interaction module is used to synchronously render real-time data streams, early warning signals, and historical comparison curves in the visualization and interaction interface, forming a multi-dimensional monitoring dashboard. The two-way manual correction feedback module is used to establish a two-way data channel based on the full-dimensional data triggered by the early warning to receive manual correction instructions, and dynamically adjust the early warning threshold and model parameters according to the information of the multi-dimensional monitoring dashboard.

9. An electronic device, characterized in that, It includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The device stores computer instructions that, when executed by a processor, implement the entire process of the method according to any one of claims 1 to 7.

Images