Permanent magnet pump whole life cycle intelligent early warning and maintenance method driven by large model

Abstract

This invention belongs to the field of intelligent monitoring and maintenance technology for industrial equipment. It discloses a large-model-driven intelligent early warning and maintenance method for the entire lifecycle of permanent magnet pumps. The method involves acquiring the electrical and thermodynamic parameters of the permanent magnet pump during operation and constructing a temperature field distribution map. Then, based on the back electromotive force waveform and temperature field, a dynamic evolution model of the hysteresis loop is constructed to calculate the operating point offset and identify the critical demagnetization region. Historical temperature trajectory data is extracted to construct a thermo-magnetic coupling hysteresis feature map and calculate the reversible recovery coefficient and irreversible damage coefficient. A demagnetization state discrimination model embedded with physical boundary constraints is constructed, and a demagnetization state feature vector containing temperature attribution labels is generated through cross-domain feature fusion. The method achieves component decomposition of the magnetic flux decay signal, distinguishes between reversible thermal decay and irreversible demagnetization components, and outputs early warning levels and maintenance measures based on dynamic proportion coefficients. This invention improves the accuracy of permanent magnet pump demagnetization risk assessment and extends equipment service life.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring and maintenance technology for industrial equipment, and more specifically, to a large-scale model-driven intelligent early warning and maintenance method for the entire life cycle of permanent magnet pumps. Background Technology

[0002] Existing permanent magnet pump monitoring and maintenance technologies have blind spots when assessing the risk of permanent magnet demagnetization, particularly in identifying the critical state between "reversible thermal demagnetization" and "irreversible permanent demagnetization." As a core component of industrial pumps, permanent magnets are extremely sensitive to temperature fluctuations. When the ambient temperature rises, the arrangement of magnetic domains within the magnet undergoes thermal disturbance, leading to a decrease in magnetic flux and back electromotive force. However, this performance degradation is usually a recoverable physical process. However, when the temperature exceeds a certain critical point or remains at high temperatures for an extended period, the internal lattice structure of the permanent magnet undergoes an irreversible transformation, resulting in permanent energy loss. Existing data-driven models often lack the necessary physical constraints to accurately capture this subtle transition. In actual industrial environments, after detecting abnormal electrical parameters, the inability to accurately construct the three-dimensional temperature field distribution and thermo-magnetic coupling relationship of the permanent magnet often leads to the erroneous identification of temporary thermal effects as severe permanent demagnetization, triggering unnecessary shutdown warnings. This not only leads to unexpected production line interruptions and forced restarts of processes, but also results in a serious waste of maintenance resources. Especially in industries with continuous production, a single erroneous downtime can mean hours or even days of lost production. More seriously, due to a lack of in-depth understanding of the dynamic evolution of hysteresis loops, existing systems cannot identify high-risk states where surface parameters appear normal but are already on the verge of demagnetization. This leads to a failure to intervene in equipment that truly needs maintenance in a timely manner, ultimately causing sudden failures under critical operating conditions, resulting in chain reactions and major production accidents.

[0003] In view of this, the present invention proposes a large-scale model-driven intelligent early warning and maintenance method for the entire life cycle of permanent magnet pumps to solve the above problems. Summary of the Invention

[0004] To overcome the aforementioned shortcomings of the prior art and to achieve the above objectives, the present invention provides the following technical solution: a large-model-driven intelligent early warning and maintenance method for the entire life cycle of a permanent magnet pump, comprising:

[0005] The electrical and thermodynamic parameter data of the permanent magnet pump during operation are obtained, and a three-dimensional temperature field distribution map of the permanent magnet is constructed based on the thermodynamic parameter data.

[0006] Based on the electrical parameter data, the back electromotive force waveform is extracted, and combined with the three-dimensional temperature field distribution map, a dynamic evolution model of the hysteresis loop in each temperature gradient interval is established.

[0007] Based on the hysteresis loop dynamic evolution model, the operating point offset of each section of the permanent magnet under the current temperature condition is calculated, and the suspected demagnetization region at the critical demagnetization boundary is identified according to the relative position relationship between the operating point offset and the intrinsic demagnetization curve.

[0008] Extract historical temperature trajectory data of suspected demagnetization areas, and construct a thermal-magnetic coupling hysteresis feature map based on the peak temperature sequence and the corresponding magnetic flux response sequence in the historical temperature trajectory data.

[0009] Based on the asymmetry of the magnetic flux recovery path and attenuation path in the thermal-magnetic coupling hysteresis characteristic spectrum, the reversible recovery coefficient and irreversible damage coefficient are calculated, and the demagnetization critical discrimination factor is determined according to the ratio of the reversible recovery coefficient to the irreversible damage coefficient.

[0010] Based on the critical demagnetization discrimination factor and the inflection point characteristics of the intrinsic demagnetization curve, a demagnetization state discrimination model embedded with physical boundary constraints is constructed. The physical boundary constraints include Curie temperature constraints and coercive force attenuation constraints.

[0011] The output features of the hysteresis loop dynamic evolution model and the constraint output of the demagnetization state discrimination model are fused across domains to generate a demagnetization state fusion feature vector containing temperature attribution labels.

[0012] Based on the demagnetization state fusion feature vector, the magnetic flux decay signal is decomposed by temperature attribution label to separate the reversible thermal decay component and the irreversible demagnetization component, and the dynamic proportion coefficient between the reversible thermal decay component and the irreversible demagnetization component is calculated.

[0013] Based on the combined distribution characteristics of the dynamic proportion coefficient and the demagnetization critical discrimination factor, the early warning trigger threshold and maintenance strategy priority are dynamically adjusted to output the differentiated early warning level and corresponding graded maintenance measures for the permanent magnet pump.

[0014] The technical effects and advantages of the large-model-driven intelligent early warning and maintenance method for the entire life cycle of permanent magnet pumps in this invention are as follows:

[0015] This invention achieves precise differentiation between reversible thermal decay and irreversible permanent demagnetization, changing the scientific basis for industrial equipment maintenance decisions. Through this accurate identification capability, the system significantly reduces false alarm rates, avoiding unnecessary downtime caused by overly conservative judgments in traditional technologies, and saving continuous production enterprises from significant economic losses. For equipment truly in a critical demagnetization state, it can detect potential risks early, providing maintenance personnel with sufficient intervention windows and effectively preventing chain reactions caused by sudden failures. This differentiated early warning mechanism optimizes the allocation of maintenance resources, shifting from "periodic maintenance" to "condition-driven maintenance," avoiding resource waste caused by over-maintenance and eliminating safety hazards caused by under-maintenance. In actual industrial environments, this invention significantly extends the service life of permanent magnet pumps, reduces unplanned downtime, improves overall equipment reliability, and brings considerable long-term economic benefits to enterprises. More importantly, the application of this intelligent early warning and maintenance method promotes a paradigm shift in industrial equipment management from passive response to proactive prediction. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the intelligent early warning and maintenance method for the entire life cycle of a permanent magnet pump driven by a large model, as described in this invention. Detailed Implementation

[0017] 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.

[0018] This application provides a large-model-driven intelligent early warning and maintenance method for the entire lifecycle of permanent magnet pumps. The implementing entities of this method include, but are not limited to: a permanent magnet pump intelligent monitoring system, an industrial equipment health management platform, a predictive maintenance decision support system, and a thermomagnetic coupling dynamic analysis platform, which can be regarded as general computing nodes of this application. The intelligent early warning system includes, but is not limited to: at least one of a cloud-based demagnetization risk assessment engine, a distributed temperature monitoring system, and an intelligent operating point offset detector.

[0019] Please see Figure 1 In this embodiment of the invention, the specific implementation steps of the large-model-driven intelligent early warning and maintenance method for the entire life cycle of a permanent magnet pump include:

[0020] Electrical and thermodynamic parameter data during the operation of the permanent magnet pump are acquired, and a three-dimensional temperature field distribution map of the permanent magnet is constructed based on the thermodynamic parameter data. The electrical parameter data includes key information such as the three-phase current waveforms, voltage waveforms, and the extracted back electromotive force of the permanent magnet pump drive circuit. The thermodynamic parameter data includes time-series temperature data from multiple measuring points on the permanent magnet surface, acquired in real time through a distributed temperature sensor array. The copper and iron loss distributions calculated from the permanent magnet surface temperature data and the current waveforms can accurately invert the heat source intensity distribution inside the permanent magnet, providing fundamental data for the construction of the three-dimensional temperature field. This data provides comprehensive raw material for subsequent analysis, ensuring the completeness and accuracy of early warning and maintenance strategy formulation.

[0021] Based on electrical parameter data, the back electromotive force waveform is extracted. Combined with a three-dimensional temperature field distribution map, a dynamic evolution model of the hysteresis loop within each temperature gradient interval is established. First, a short-time Fourier transform is performed on the back electromotive force waveform to extract the fundamental amplitude sequence and harmonic distortion rate sequence. Then, based on the three-dimensional temperature field distribution map, the permanent magnet is divided into multiple temperature gradient intervals, and a spatial mapping relationship between each interval and the fundamental amplitude sequence is established. Based on the temperature coefficient curve of the permanent magnet material, the theoretical remanence and coercivity values ​​are calculated. By comparing these with measured values, the hysteresis loop shape parameters for each temperature gradient interval are fitted, forming a dynamic evolution model. These multidimensional features collectively constitute the "magnetic fingerprint" of the permanent magnet, providing a foundation for subsequent operating point analysis and demagnetization detection.

[0022] Based on the hysteresis loop dynamic evolution model, the operating point offset of each partition of the permanent magnet at the current temperature is calculated. Then, based on the relative positional relationship between the operating point offset and the intrinsic demagnetization curve, suspected demagnetization regions at the critical demagnetization boundary are identified. The dynamic operating point coordinates of each partition at the current temperature, including magnetic field strength and magnetic flux density values, are extracted from the hysteresis loop dynamic evolution model. Then, the intrinsic demagnetization curve of the permanent magnet material at a reference temperature is obtained, and the offset vector of the dynamic operating point relative to the corresponding reference operating point is calculated. By calculating the normalized distance between the dynamic operating point and the inflection point of the intrinsic demagnetization curve, the critical proximity is determined. Partitions with operating point offsets greater than a preset threshold and critical proximity less than a preset threshold are marked as suspected demagnetization regions, providing key input for the analysis of thermo-magnetic coupling characteristics.

[0023] Historical temperature trajectory data of suspected demagnetization areas are extracted. Based on the peak temperature sequences and corresponding magnetic flux response sequences in the historical temperature trajectory data, a thermo-magnetic coupling hysteresis feature map is constructed. Historical temperature trajectory data of suspected demagnetization areas are extracted from the historical operation database. Peak temperature events exceeding a preset temperature benchmark are identified, forming peak temperature sequences. Magnetic flux change data at the time of each peak temperature event and within a preset recovery period are extracted, forming a magnetic flux response sequence. A scatter plot of the thermo-magnetic response is drawn with the peak temperature sequence as the horizontal axis and the attenuation magnitude of the magnetic flux response sequence as the vertical axis. A thermo-magnetic coupling hysteresis feature map is generated through kernel density estimation, visually displaying the probability distribution of magnetic flux response under different temperature excitations, providing a spectral basis for demagnetization criticality determination.

[0024] Based on the asymmetry of the magnetic flux recovery path and attenuation path in the thermo-magnetic coupling hysteresis characteristic spectrum, the reversible recovery coefficient and irreversible damage coefficient are calculated, and the demagnetization critical discrimination factor is determined according to the ratio of the reversible recovery coefficient to the irreversible damage coefficient. The magnetic flux attenuation path curve during the heating stage and the magnetic flux recovery path curve during the cooling stage are extracted, and the area difference between the two paths is calculated as the hysteresis loss area. The reversible recovery coefficient is calculated by the ratio of the endpoint of the magnetic flux recovery path curve to the initial magnetic flux value, and the irreversible damage coefficient is calculated by the slope of the correlation between the hysteresis loss area and the cumulative temperature cycle number. Finally, the ratio of the two is logarithmically transformed to obtain the demagnetization critical discrimination factor, providing a quantitative basis for the construction of the state discrimination model.

[0025] Based on the demagnetization critical discrimination factor and the inflection point characteristics of the intrinsic demagnetization curve, a demagnetization state discrimination model embedded with physical boundary constraints is constructed. These physical boundary constraints include Curie temperature constraints and coercive force attenuation constraints. Inflection point coordinates are extracted from the intrinsic demagnetization curve to characterize the theoretical boundary between the reversible and irreversible demagnetization regions. A Curie temperature constraint function is constructed based on the Curie temperature parameter of the permanent magnet material, and a coercive force attenuation constraint function is constructed based on the relative position of the demagnetization critical discrimination factor and the inflection point coordinates. These physical constraints are then embedded into the physical constraint layer of a neural network to construct a demagnetization state discrimination model that comprehensively considers physical laws.

[0026] The output features of the hysteresis loop dynamic evolution model and the constraint output of the demagnetization state discrimination model are fused across domains to generate a demagnetization state fusion feature vector containing temperature attribution labels. Hysteresis loop shape feature vectors for each temperature gradient interval are extracted from the hysteresis loop dynamic evolution model as data-driven features, while Curie temperature margin and coercivity margin values ​​are extracted from the physical constraint layer of the demagnetization state discrimination model as physical constraint features. The cross-domain correlation weights between the two types of features are calculated using an attention mechanism, and then weighted and concatenated according to these weights, with temperature attribution labels added to generate the demagnetization state fusion feature vector, providing a feature basis for demagnetization component decomposition.

[0027] Based on the demagnetization state fusion feature vector, the magnetic flux decay signal is decomposed using temperature attribution labels to separate reversible thermal decay components and irreversible demagnetization components. The dynamic proportion coefficient between these two components is then calculated. According to the temperature attribution labels, the magnetic flux decay signal in the demagnetization state fusion feature vector is divided into temperature-dependent and temperature-independent components. Compensation calculations are performed on the temperature-dependent components to obtain the residual decay amount after eliminating temperature influence, which is then taken as the irreversible demagnetization component. The difference between the magnetic flux decay signal and the irreversible demagnetization component is labeled as the reversible thermal decay component. By calculating the ratio of the amplitudes of the two types of components and performing sliding window smoothing, the dynamic proportion coefficient is obtained, providing a quantitative basis for early warning strategy formulation.

[0028] Based on the combined distribution characteristics of the dynamic proportion coefficient and the demagnetization critical discrimination factor, the early warning trigger threshold and maintenance strategy priority are dynamically adjusted to output differentiated early warning levels and corresponding graded maintenance measures for permanent magnet pumps. A two-dimensional state space is constructed with the dynamic proportion coefficient as the first dimension and the demagnetization critical discrimination factor as the second dimension, dividing the space into a thermal attenuation-dominated zone, a critical transition zone, and a true demagnetization-dominated zone. Differentiated early warning trigger thresholds are configured for each zone. Based on the current state point's position in the two-dimensional state space and its moving velocity vector, the evolution trajectory of the state point is predicted. Based on the trajectory, corresponding maintenance strategy priorities and specific suggestions are output, achieving accurate early warning and differentiated maintenance for permanent magnet pump demagnetization risks.

[0029] In this embodiment of the invention, the detailed implementation steps for acquiring electrical and thermodynamic parameter data during the operation of the permanent magnet pump and constructing a three-dimensional temperature field distribution map of the permanent magnet based on the thermodynamic parameter data include:

[0030] Temperature time-series data from multiple measuring points on the surface of a permanent magnet is collected using a distributed temperature sensor array, serving as the first component of the thermodynamic parameter data. The temperature sensor array is the fundamental hardware infrastructure for acquiring thermodynamic data, enabling high-precision, multi-point real-time monitoring of the permanent magnet surface temperature. Sensor deployment follows a hotspot coverage principle, with sensors evenly spaced along the axial and radial directions of the permanent magnet at preset intervals, and densely concentrated in areas prone to hotspots, such as near the stator coils, bearings, and fluid inlet areas. High-precision thermocouples or PT100 platinum resistance thermometers are used, achieving a temperature measurement accuracy better than ±0.5℃. The sampling frequency is dynamically adjusted according to operating conditions, typically once every 1-5 minutes during steady-state operation, and can be increased to once every 10-30 seconds during dynamic changes. Temperature data is transmitted in real-time to the data acquisition unit via a fieldbus network, where preliminary filtering is performed to eliminate noise interference, forming a continuous temperature time-series data stream. This raw temperature data, as the first component of the thermodynamic parameters, provides key boundary conditions for subsequent temperature field construction.

[0031] The three-phase current waveforms and back electromotive force waveforms of the permanent magnet pump drive circuit are synchronously acquired using current and voltage sensors as electrical parameter data. Electrical parameter acquisition is the core method for obtaining the electromagnetic state of the permanent magnet pump, and high-precision sensors synchronously record the dynamic changes of electrical signals. A Hall effect sensor is used for current acquisition, and a high-impedance differential probe is used for voltage acquisition; both have high sampling rates and anti-interference capabilities. The sampling rate is set to 100-200 times the power supply frequency to ensure the capture of harmonic components and minute variations. Phase-synchronous sampling is implemented during acquisition to ensure time alignment of the three-phase data, providing accurate data for subsequent waveform analysis. The raw waveforms undergo digital filtering, scaling, and phase correction to form a standardized electrical parameter dataset, including three-phase current amplitudes, phase relationships, voltage waveforms, and the calculated back electromotive force waveform. These electrical parameters reflect the magnetic field state of the permanent magnet and the operating characteristics of the motor, providing fundamental data from an electromagnetic perspective for hysteresis loop construction and demagnetization analysis.

[0032] The copper and iron losses of the permanent magnet pump are calculated based on the three-phase current waveform, and the heat source intensity distribution inside the permanent magnet is inverted based on the copper and iron loss distribution. Loss calculation is a key step in determining the internal heat source, and the distribution law of heat generation is inverted through electrical parameters. The calculation process first calculates the copper loss distribution based on the three-phase current waveform and winding resistance, using the following formula:

[0033] ;in, For spatial points Copper loss density at the location, For the first RMS value of phase current For temperature The winding resistance value below, Let be the winding density function. Radial coordinates, Angular coordinates, For axial coordinates.

[0034] Iron loss calculation employs an improved Bertotti model, considering the combined effects of frequency, magnetic flux density, and temperature, and includes three parts: eddy current loss, hysteresis loss, and additional loss. The loss distribution is directly related to the electromagnetic field distribution. A spatial mapping relationship is established through finite element pre-calculation of the electromagnetic field, achieving efficient conversion from electrical parameters to heat source distribution. The heat source intensity distribution is the internal excitation condition for calculating the permanent magnet temperature field; its accuracy directly affects the accuracy of the temperature field construction and has a fundamental impact on subsequent demagnetization risk assessment.

[0035] Using time-series temperature data as boundary conditions and heat source intensity distribution as internal excitation, a three-dimensional temperature field distribution map is constructed through finite element heat conduction analysis. Temperature field construction is a crucial step in integrating surface measurement data and internal heat source information, achieving accurate mapping of the permanent magnet's global temperature through numerical solutions. The construction process first establishes a three-dimensional geometric model of the permanent magnet pump, including key components such as the permanent magnet, stator core, bearings, and fluid channels; then, material thermophysical parameters are set, including thermal conductivity, specific heat capacity, and density, which can be dynamically adjusted with temperature changes; next, the surface temperature time-series data is interpolated to form complete boundary conditions, and the heat source intensity distribution is used as the internal volume heat source; finally, the global temperature distribution is obtained by solving the heat conduction control equations. The heat conduction control equations are:

[0036] ;in, For density, For specific heat capacity, Thermal conductivity, The heat source intensity can be a function of temperature T. It is a time variable.

[0037] An implicit time-progression scheme is employed to ensure numerical stability, while the second-order finite element method is used for spatial discretization to balance accuracy and efficiency. The three-dimensional temperature field distribution map visually illustrates the spatial distribution and temporal evolution of the temperature inside the permanent magnet, providing a global perspective for identifying potential hotspots and assessing demagnetization risk. This serves as a crucial input for constructing the subsequent dynamic evolution model of the hysteresis loop.

[0038] In this embodiment of the invention, the detailed implementation steps for establishing the dynamic evolution model of the hysteresis loop within each temperature gradient interval include:

[0039] A short-time Fourier transform (STFT) was performed on the back electromotive force (EMF) waveform to extract the fundamental amplitude sequence and harmonic distortion rate sequence. Waveform analysis is a fundamental step in obtaining electromagnetic characteristics, and key features of the back EMF were extracted through frequency domain transformation. The analysis process employed a sliding window short-time Fourier transform (STFT), with a window length set to 1-2 electrical cycles. A Hanning window was selected as the window function to reduce spectral leakage. The transformation calculation formula is as follows:

[0040] ;in, The back electromotive force is the time-domain signal. For The window function is centered. Angular frequency, It is an imaginary number.

[0041] The amplitude of the fundamental frequency component is extracted from the transformation result to form a fundamental amplitude sequence, which directly reflects the changing trend of the permanent magnet's magnetic flux. Simultaneously, the amplitude ratio of each harmonic relative to the fundamental wave is calculated to form a harmonic distortion rate sequence. The harmonic distortion rate is sensitive to changes in magnetic circuit saturation and nonlinear characteristics, and can help identify changes in permanent magnet performance. Through time-frequency analysis of the STFT, dynamic features of the back electromotive force signal are extracted, providing fundamental data from a frequency domain perspective for hysteresis loop parameter estimation.

[0042] Based on the three-dimensional temperature field distribution map, the permanent magnet is divided into multiple temperature gradient intervals, and a spatial mapping relationship between each temperature gradient interval and the fundamental wave amplitude sequence is established. Interval division is a key strategy for handling temperature non-uniformity, and partition analysis improves the accuracy of characteristic evaluation. The division process first performs gradient analysis on the three-dimensional temperature field, calculating the amplitude distribution of the spatial derivative of the temperature. Then, based on the natural discontinuities of the temperature gradient, the permanent magnet space is divided into multiple temperature gradient intervals. The number of intervals is dynamically determined according to the complexity of the temperature distribution, typically 5-10. To ensure the rationality of the interval division, clustering verification indicators such as the silhouette coefficient are used to evaluate the division quality, and physical boundary alignment is used to ensure that the division boundaries are consistent with the physical structure of the permanent magnet. The spatial mapping relationship is established through coordinate transformation and spatial interpolation, mapping the fundamental wave amplitude data from the time domain to the spatial domain, establishing a one-to-one correspondence between each interval and the corresponding fundamental wave amplitude sequence segment. This differentiated approach effectively handles the local characteristic differences caused by the non-uniform temperature inside the permanent magnet, laying the foundation for subsequent refined hysteresis loop construction.

[0043] Based on the temperature coefficient curve of the permanent magnet material, the theoretical remanence and coercivity values ​​are calculated within each temperature gradient range. The theoretical parameter calculations form the theoretical foundation for model construction, deriving the magnetic parameters under ideal conditions through the relationship between material properties and temperature. The calculation process is based on the temperature coefficient curve of the permanent magnet material, which is typically provided by the material manufacturer or obtained through standard testing. The remanence temperature coefficient of a typical NdFeB material is shown below. The temperature coefficient of coercivity is approximately -0.08% / ℃ to -0.12% / ℃. It is approximately -0.4% / ℃ to -0.6% / ℃. The theoretical remanence value is calculated using the following formula:

[0044] ;in, For temperature The residual magnetism below, Reference temperature The residual magnetism below, is the temperature coefficient of remanence.

[0045] The theoretical coercivity value is calculated using a similar formula, but the nonlinear relationship between coercivity and temperature must be considered, especially the rapid decay characteristics near the Curie temperature. Through theoretical parameter calculations, a baseline relationship between temperature and magnetic properties is established, providing a theoretical reference for subsequent comparative analysis with measured values.

[0046] Based on the measured corresponding values ​​of theoretical remanence, theoretical coercivity, and fundamental amplitude sequences, the shape parameters of the hysteresis loop in each temperature gradient interval are fitted to construct a dynamic evolution model of the hysteresis loop. Parameter fitting is the core step in constructing an accurate model; the optimal hysteresis loop expression is determined by fusing theoretical and measured data. The fitting process uses an improved Jiles-Atherton (JA) model, which describes the hysteresis loop characteristics through five physical parameters: Ms (saturation magnetization), α (field parameter), k (hysteresis loss parameter), c (reversible magnetization parameter), and a (shape parameter). The fitting algorithm employs a hybrid strategy combining particle swarm optimization and the Levenberg-Marquardt method; the former provides good initial estimates, while the latter achieves fine-grained optimization. The objective function is designed to minimize the mean square error between the theoretical back electromotive force waveform and the measured waveform. To improve the model's adaptability to temperature changes, a temperature dependency term is introduced, making the JA model parameters functions of temperature.

[0047] ;in, Represents any JA model parameter. This is a temperature correlation function. This is the temperature sensitivity coefficient for the corresponding parameter.

[0048] By performing fitting processes in each temperature gradient interval, a set of JA model parameters varying with temperature distribution is obtained, constituting a complete dynamic evolution model of the hysteresis loop. This model can accurately describe the changes in the hysteresis characteristics of permanent magnets under different temperature conditions, providing a dynamic characteristic basis for operating point analysis and demagnetization risk assessment.

[0049] In this embodiment of the invention, based on the hysteresis loop dynamic evolution model, the detailed implementation steps for calculating the operating point offset of each partition of the permanent magnet at the current temperature state, and identifying the suspected demagnetization region at the critical demagnetization boundary according to the relative positional relationship between the operating point offset and the intrinsic demagnetization curve, include:

[0050] The dynamic operating point coordinates of each zone under the current temperature condition are extracted from the hysteresis loop dynamic evolution model. These coordinates include the magnetic field strength and magnetic flux density values. Operating point extraction is the starting point for analyzing demagnetization risk, and the real-time operating state of the permanent magnet is obtained through model calculations. The extraction process first updates the hysteresis loop model parameters of each zone based on the current temperature condition; then, the operating point of the permanent magnet is determined in the second quadrant, i.e., the coordinates of the intersection of the hysteresis loop and the demagnetization curve. For permanent magnet pump applications, the operating point is mainly determined by the external load line and temperature. The dynamic operating point coordinates are obtained by solving the following system of equations:

[0051] ;

[0052] ;in, The magnetic flux density at the operating point, The magnetic field strength at the operating point, For temperature The residual magnetization below , is the load factor, dimensionless, representing the degree of influence of the external load on the permanent magnet pump on the working state of the permanent magnet; Vacuum permeability is a physical constant with a value of (Henry / meter). is the relative permeability, dimensionless, representing the ratio of the material's permeability to the permeability of free space.

[0053] Through numerical iterative solutions, the precise operating point coordinates of each zone under the current temperature condition are obtained. These coordinates directly reflect the operating state of the permanent magnet and are the basic data for judging the risk of demagnetization, providing key input for subsequent comparative analysis with the intrinsic demagnetization curve.

[0054] The intrinsic demagnetization curve of the permanent magnet material at a reference temperature is obtained. The offset vector of the dynamic operating point coordinates relative to the corresponding reference operating point on the intrinsic demagnetization curve is calculated, and the magnitude of the offset vector is denoted as the operating point offset. Intrinsic curve comparison is a crucial step in operating point assessment, quantifying the demagnetization risk through difference analysis from the standard state. The comparison process first obtains the intrinsic demagnetization curve of the permanent magnet material at a reference temperature (typically 20℃ or 25℃), which is obtained from the material specification or standard test. Then, the dynamic operating point at the current temperature is mapped to the reference temperature space through temperature compensation transformation to obtain the equivalent operating point. Next, a reference operating point with the same magnetic field strength is found on the intrinsic demagnetization curve. Finally, the offset vector between the two points is calculated, and its magnitude is the operating point offset. The formula for calculating the offset is:

[0055] ;in, For dynamic working point coordinates, The coordinates are for the reference working point.

[0056] The operating point offset directly reflects the degree of performance deviation of the permanent magnet. The larger the offset, the closer the permanent magnet is to the demagnetization risk zone. It is an important quantitative indicator for demagnetization identification.

[0057] The normalized distance between the dynamic operating point coordinates and the inflection point of the intrinsic demagnetization curve is calculated and denoted as the critical proximity. Inflection point analysis is a key technique for accurately assessing the demagnetization boundary, evaluating the risk level by measuring the distance to key characteristic points of the demagnetization curve. The analysis process first identifies the inflection point on the intrinsic demagnetization curve, i.e., the position where the curve curvature is greatest. This point usually represents the critical point for transitioning from the reversible demagnetization region to the irreversible demagnetization region. Then, the Euclidean distance from the dynamic operating point to the inflection point is calculated. Finally, the distance is normalized using a reference length to obtain the dimensionless critical proximity index. The formula for calculating the critical proximity is:

[0058] ;in, The coordinates of the inflection point of the demagnetization curve. The reference length is usually taken as the intrinsic remanence value. .

[0059] Critical proximity and operating point offset are two complementary indicators. The former focuses on the absolute risk position, while the latter reflects the relative performance change. Together, they provide multi-dimensional quantitative basis for the identification of demagnetization areas.

[0060] Partitions with operating point offsets greater than a first preset threshold and critical proximity less than a second preset threshold are marked as suspected demagnetization areas. The dual-threshold judgment is a decision rule for identifying high-risk areas, filtering out the areas most likely to demagnetize by combining conditions. The judgment process sets two key thresholds: the first preset threshold controls the lower limit of the operating point offset, typically set to 3%-5% of the initial residual magnetism; the second preset threshold controls the upper limit of the critical proximity, typically set between 0.1 and 0.2. When a partition simultaneously meets the conditions of an offset greater than the first threshold and a proximity less than the second threshold, it indicates that the area has both experienced significant performance deviation and is close to the irreversible demagnetization boundary, and is marked as a suspected demagnetization area. To improve the reliability of the judgment, the system tracks the historical evolution trend of these areas; areas judged as high-risk at multiple consecutive time points will receive higher warning priority. Accurate identification of suspected demagnetization areas provides a focused target for subsequent in-depth analysis, avoiding the waste of resources in full-domain analysis and improving the efficiency and accuracy of the early warning system.

[0061] In this embodiment of the invention, the detailed implementation steps for extracting historical temperature trajectory data of suspected demagnetization regions and constructing a thermo-magnetic coupling hysteresis feature map based on the peak temperature sequence and the corresponding magnetic flux response sequence in the historical temperature trajectory data include:

[0062] Historical temperature trajectory data of suspected demagnetization areas is extracted from the historical operation database, and all peak temperature events exceeding a preset temperature benchmark are identified to form a peak temperature sequence. Historical data analysis is the foundation for building a thermal-magnetic correlation, extracting characteristic thermal shock samples through temperature peak events. The analysis process first connects to the historical operation database to query long-term temperature records of suspected demagnetization areas; then, a peak detection algorithm identifies local maxima in the temperature curves, ensuring that the interval between adjacent peaks is greater than a preset minimum separation; next, the peaks are compared with the preset temperature benchmark to filter out high-temperature events exceeding the benchmark; finally, they are arranged in chronological order to form a peak temperature sequence. The temperature benchmark is usually set to 80%-90% of the upper limit of the material's normal operating temperature. For common NdFeB materials, the benchmark value may be between 80℃ and 120℃. Peak detection uses an improved local extremum algorithm, combined with smoothing processing to reduce noise interference, ensuring accurate capture of real temperature peak events. The peak temperature sequence directly reflects the thermal shock history experienced by the permanent magnet, providing a time anchor for correlated magnetic performance responses.

[0063] Magnetic flux change data are extracted at the moment of each peak temperature event and within a preset recovery period to form a magnetic flux response sequence. The magnetic flux response is a direct reflection of the temperature effect, revealing the thermo-magnetic coupling relationship through synchronous analysis. The extraction process first determines the precise timestamp of each temperature peak event; then, it extends forward by a preset recovery period, typically 1.5-2 times the time required to cool to the reference temperature; next, magnetic flux data is extracted within the entire time window, with magnetic flux usually measured indirectly through the amplitude of the back electromotive force fundamental wave; finally, the data time axis is aligned and standardized to form the magnetic flux response sequence. To ensure data quality, outlier detection and missing value imputation techniques are used to eliminate the influence of measurement noise and data breakpoints. The magnetic flux response sequence corresponds one-to-one with the peak temperature sequence, together forming paired samples of temperature excitation-magnetic flux response, providing direct evidence for the analysis of thermo-magnetic coupling characteristics.

[0064] A scatter plot of the thermo-magnetic response is created, with the peak temperature sequence on the x-axis and the attenuation magnitude of the magnetic flux response sequence on the y-axis. Scatter analysis is an intuitive method for discovering thermo-magnetic patterns, revealing the inherent correlations in the data through a two-dimensional mapping. The plotting process first calculates the magnetic flux attenuation magnitude in each response sequence, i.e., the normalized value of the difference between the magnetic flux at the peak temperature and the magnetic flux at the recovery endpoint; then, the peak temperature value is used as the x-axis and the corresponding attenuation magnitude as the y-axis to plot scatter points on a two-dimensional plane; finally, temporal information is added as a third-dimensional attribute of the points, using color or size encoding to represent the chronological order of events. The scatter plot visually demonstrates the correlation pattern between temperature and magnetic flux changes. Higher temperatures typically correspond to larger attenuation magnitudes, and with increasing usage time, the attenuation magnitude at the same temperature may gradually increase, indicating cumulative damage to the permanent magnet's performance. This visualization analysis helps technicians intuitively understand thermo-magnetic coupling characteristics and discover nonlinear relationships and anomaly patterns.

[0065] Kernel density estimation is performed on the scatter distribution of the thermo-magnetic response to generate a thermo-magnetic coupling hysteresis characteristic map. The contour lines of the thermo-magnetic coupling hysteresis characteristic map represent the probability distribution of the magnetic flux response under different temperature excitations. Density estimation is a statistical method for generating continuous models from discrete samples, revealing the inherent distribution law of the data through the probability density function. The estimation process uses the kernel density estimation (KDE) method, which uses a two-dimensional Gaussian kernel function to smooth the scatter distribution, and the bandwidth parameter is automatically optimized through cross-validation. The KDE calculation formula is:

[0066] ;in, For the estimated probability density function, For kernel function, For bandwidth parameters, For the first The coordinates of the scattered points This represents the total number of scatter points.

[0067] The estimation results are presented as contour plots, forming a thermo-magnetic coupling hysteresis characteristic map. The contour lines in the map represent the probability distribution of magnetic flux response under different temperature excitations; high-density areas indicate the most common response modes, while outliers may indicate abnormal states or degradation trends. The hysteresis characteristic is mainly reflected in the inconsistency of magnetic flux change paths during heating and cooling; this degree of hysteresis reflects the health status of the permanent magnet. The thermo-magnetic coupling hysteresis characteristic map is an important tool for demagnetization analysis, providing a probability distribution basis for subsequent calculations of the restitution coefficient and damage coefficient, helping to identify the degradation modes of permanent magnets and predict future performance trends.

[0068] In this embodiment of the invention, the detailed implementation steps for calculating the reversible recovery coefficient and the irreversible damage coefficient based on the asymmetry of the magnetic flux recovery path and the attenuation path in the thermo-magnetic coupling hysteresis characteristic spectrum, and determining the demagnetization critical discrimination factor according to the ratio of the reversible recovery coefficient to the irreversible damage coefficient, include:

[0069] The magnetic flux decay path curve during the heating phase and the magnetic flux recovery path curve during the cooling phase are extracted from the thermo-magnetic coupling hysteresis characteristic spectrum. Path extraction is a fundamental step in analyzing hysteresis characteristics, revealing dynamic changes through temperature-magnetic flux relationship curves. The extraction process first identifies key inflection points of temperature change in the thermo-magnetic coupling hysteresis characteristic spectrum, dividing the entire process into heating and cooling phases; then, magnetic flux change data corresponding to the two phases are extracted along the temperature change trajectory; finally, continuous decay and recovery path curves are generated through spline interpolation. To ensure the representativeness of the extracted paths, the system analyzes multiple temperature cycle events, prioritizing typical samples with high signal-to-noise ratios and significant temperature changes. The decay path curve reflects the decreasing trend of magnetic properties of the permanent magnet during temperature increase, while the recovery path curve shows the recovery of magnetic properties after temperature decrease. The shape and relative position of the two paths contain rich information, effectively distinguishing between reversible thermal effects and irreversible demagnetization phenomena, providing basic data for subsequent quantitative analysis.

[0070] The area difference between the magnetic flux recovery path curve and the magnetic flux decay path curve is calculated as the hysteresis loss area. The hysteresis area is a direct indicator of path asymmetry and is used to assess the degree of irreversible loss through integration. The calculation process first ensures that the two path curves are defined within the same temperature range, performing interpolation if necessary; then, the area of ​​the closed region between the two curves is calculated, i.e., the hysteresis loss area. The calculation formula is:

[0071] ;in, This represents the hysteresis loss area. For temperature The flux value of the attenuation path under the condition, For temperature The recovery path flux value.

[0072] Integration is typically achieved using numerical methods such as the trapezoidal rule. For discrete data points, the calculation formula can be simplified to:

[0073] ;

[0074] in, For the first Temperature values ​​at each temperature sampling point For the first The temperature values ​​at each temperature sampling point, and A temperature interval is determined together.

[0075] The hysteresis loss area directly reflects the irreversible energy lost by the permanent magnet in one temperature cycle. The larger the area, the more severe the irreversible damage. It is a key quantitative indicator for assessing the degree of demagnetization.

[0076] The reversible recovery coefficient is calculated based on the ratio of the final magnetic flux value to the initial magnetic flux value on the magnetic flux recovery path curve. The recovery coefficient is a key parameter for evaluating reversible performance, quantifying the degree of performance recovery after temperature cycling by comparing the endpoints. The calculation process first determines the initial point (usually the magnetic flux corresponding to the highest temperature) and the endpoint (the magnetic flux after the temperature recovers to the reference value) of the recovery path curve; then, the ratio of the endpoint magnetic flux to the original magnetic flux at the initial temperature is calculated to obtain the reversible recovery coefficient. The calculation formula is:

[0077] ;in, The reversible recovery coefficient is... To recover the magnetic flux value at the end of the path, This represents the initial magnetic flux value before the temperature cycle.

[0078] The reversible recovery coefficient typically ranges from [0,1]. A value closer to 1 indicates that the permanent magnet can recover to near its original state after temperature cycling, primarily due to reversible thermal demagnetization. A smaller value indicates significant irreversible performance loss. This coefficient is a crucial indicator for distinguishing between reversible thermal demagnetization and irreversible permanent demagnetization, directly impacting maintenance decisions.

[0079] The irreversible damage coefficient is calculated based on the slope of the correlation between the hysteresis loss area and the cumulative number of temperature cycles. The damage coefficient is a long-term indicator for assessing cumulative effects, and it predicts degradation trends through correlation analysis of historical data. The calculation process first involves statistically analyzing the temperature cycle history of the permanent magnet, recording the hysteresis loss area and time for each cycle; then analyzing the trend of the hysteresis loss area with the cumulative number of cycles, which typically exhibits a linear or power-law relationship; finally, regression analysis is used to extract the trend slope as the irreversible damage coefficient. The calculation formula is:

[0080] ;in, The irreversible damage coefficient, This represents the hysteresis loss area. This is the cumulative number of temperature cycles; The derivative symbol represents the rate of change of the hysteresis loss area A with respect to the number of cycles N, reflecting the degree of marginal damage to the permanent magnet caused by each additional temperature cycle as the service time increases.

[0081] To handle nonlinear relationships, linear regression after logarithmic transformation is typically used to improve fitting accuracy. The irreversible damage coefficient reflects the rate of performance degradation of permanent magnets over time and is an important basis for predicting remaining service life, providing key guidance for developing preventative maintenance strategies.

[0082] The ratio of the reversible recovery coefficient to the irreversible damage coefficient is logarithmically transformed to obtain the demagnetization critical discriminant factor. The discriminant factor is a composite index that comprehensively evaluates the demagnetization state, balancing the effects of short-term recovery and long-term degradation through the ratio relationship. The calculation process first involves adjusting the reversible recovery coefficient... With irreversible damage coefficient Taking the logarithm of the ratio yields the critical demagnetization factor. The calculation formula is:

[0083] ;in, As the critical discrimination factor for demagnetization, The reversible recovery coefficient is... The irreversible damage coefficient.

[0084] The purpose of logarithmic transformation is to map a wide range of ratios to a more manageable linear scale, facilitating the setting of discrimination thresholds. The demagnetization critical discrimination factor comprehensively considers both the immediate recovery capability and long-term degradation trend of permanent magnets, serving as a comprehensive indicator of demagnetization status. A higher factor value indicates that reversible components dominate, suggesting a lower risk of demagnetization; a lower value indicates significant accumulation of irreversible damage, requiring close monitoring. The demagnetization critical discrimination factor provides a quantitative foundation for subsequent construction of physical constraint discrimination models, serving as a crucial bridge connecting data analysis and physical models.

[0085] In this embodiment of the invention, the detailed implementation steps for constructing a demagnetization state discrimination model embedded with physical boundary constraints based on the demagnetization critical discrimination factor and the inflection point characteristics of the intrinsic demagnetization curve include:

[0086] Inflection point coordinates are extracted from the intrinsic demagnetization curve. These coordinates represent the theoretical boundary between the reversible and irreversible demagnetization regions. Inflection point extraction is fundamental to physical model construction, determining the theoretical boundary of demagnetization through curve feature points. The extraction process first obtains the intrinsic demagnetization curve under standard test conditions, typically from material specifications or experimental tests; then, the second derivative of the curve is calculated, identifying the location of the most significant derivative change; finally, the inflection point coordinates are precisely located through interpolation. For most permanent magnet materials, the demagnetization curve exhibits a clear "inflection point" in the second quadrant. The region before this point is primarily reversible demagnetization, where magnetic properties can recover with external fields or temperature, while the region after this point enters irreversible demagnetization, resulting in permanent loss of magnetic properties. The inflection point coordinates (Bk, Hk) are key parameters for constructing physical constraints, directly determining the physical boundary of the discrimination model and ensuring that the model output is consistent with the actual physical properties of the material.

[0087] Based on the Curie temperature parameters of the permanent magnet material, a Curie temperature constraint function is constructed. This function restricts the demagnetization state output to be forcibly classified as irreversible demagnetization when approaching the Curie temperature. Temperature constraint is the primary condition for the physical boundary, ensuring the model's correct classification under extreme conditions through the Curie temperature characteristics. The construction process first obtains the Curie temperature Tc of the permanent magnet material, which is the critical point at which the material completely loses its ferromagnetism; then, an S-shaped decay function is designed to rapidly increase the weight of irreversible demagnetization as the temperature approaches the Curie temperature. The constraint function formula is:

[0088] ;in, The value of the Curie temperature constraint function. The current temperature. Curie temperature, For steepness parameters, For safety margin.

[0089] The function approaches 0 when the temperature is much below the Curie temperature, indicating that the influence of temperature factors is relatively small; when the temperature is close to... At this point, the function value rapidly rises to near 1, forcing the model to classify the state as high-risk irreversible demagnetization. This hard constraint based on a physical mechanism ensures that the model's judgment under extreme temperature conditions conforms to the principles of materials science, avoiding the physical inconsistencies that may arise from purely data-driven approaches.

[0090] Based on the relative position of the demagnetization critical discriminant factor and the inflection point coordinates, a coercivity attenuation constraint function is constructed. This function limits the demagnetization discrimination weight after the operating point crosses the inflection point. The operating point constraint is the core condition for demagnetization discrimination, ensuring accurate identification of the boundary state by the model through the magnetic field position relationship. The construction process first calculates the relative positional relationship between the current operating point and the inflection point of the demagnetization curve; then, a distance-based weighting function is designed to apply an increasing irreversible weight to the operating point that crosses the inflection point boundary. The constraint function formula is:

[0091] ;in, Let coercivity attenuation constraint function be used. The magnetic field strength at the current operating point. The magnetic field strength at the inflection point. As the critical discrimination factor for demagnetization, For sensitivity parameters, and These are the step function and the sigmoid function, respectively.

[0092] This function works when the magnetic field strength at the operating point is less than that at the inflection point ( A value close to 0 indicates a safe zone; when the inflection point is exceeded and the critical demagnetization discrimination factor is low, the function value increases significantly, strengthening the discrimination weight for irreversible demagnetization. This composite constraint, combining theoretical boundaries and measured indicators, ensures the model's high sensitivity and accurate discrimination capability for boundary states.

[0093] A demagnetization state discrimination model is constructed by embedding the Curie temperature constraint function and the coercive decay constraint function into the physical constraint layer of a neural network. Physical embedding is an innovative step in model construction, fusing data features and physical laws through the constraint layer. The construction process first designs a multi-layer neural network architecture, including an input layer, hidden layers, and an output layer; then, a dedicated physical constraint layer is inserted between the regular layers, integrating the Curie temperature constraint and the coercive decay constraint; finally, the overall model parameters are optimized through end-to-end training. The forward propagation formula for the physical constraint layer is:

[0094] ;in, For the constraint layer output, For the constraint layer input, and These are two constraint functions, and These are the weight parameters.

[0095] The network training employs a hybrid loss function, considering both prediction accuracy and physical consistency to ensure the model fits historical data while adhering to physical laws. The final output of the demagnetization state discrimination model is the probability distribution of the demagnetization state in each region of the permanent magnet, including three main states: "safe," "reversible thermal demagnetization," and "irreversible permanent demagnetization." This hybrid model architecture, embedded with physical constraints, combines the advantages of data-driven and mechanism-driven approaches, maintaining flexibility while ensuring physical rationality, providing a reliable tool for accurately assessing the demagnetization risk of permanent magnet pumps.

[0096] In this embodiment of the invention, the detailed implementation steps for cross-domain feature fusion of the output features of the hysteresis loop dynamic evolution model and the constraint output of the demagnetization state discrimination model to generate a demagnetization state fusion feature vector containing temperature attribution labels include:

[0097] The shape feature vectors of the hysteresis loops for each temperature gradient interval are extracted from the dynamic evolution model of the hysteresis loop as data-driven features. Shape features are microscopic indicators describing magnetic properties, quantifying the detailed changes in hysteresis behavior through loop characteristic parameters. The extraction process first obtains the hysteresis loop parameters for each interval under the current temperature condition from the dynamic evolution model; then, it calculates the shape feature indicators of the loop, including coercivity. ,remanence Maximum magnetic energy product Rectangular ratio , loop area Finally, these indicators are combined to form a fixed-dimensional feature vector. To improve the expressive power of the features, the temperature sensitivity coefficient and time derivative of each indicator are also calculated to capture dynamic trends. These data-driven features directly reflect the microscopic magnetic behavior of the permanent magnet, contain rich performance state information, and serve as the fundamental data source for feature fusion.

[0098] Curie temperature margin and coercivity margin values ​​are extracted from the physical constraint layer of the demagnetization state discrimination model as physical constraint features. Physical margin is a quantitative representation of the safety boundary, and the safety redundancy of system operation is assessed through critical distance. The extraction process first obtains the intermediate output values ​​of the constraint functions from the physical constraint layer; then, it calculates the margin between the current state and the physical critical point, where the Curie temperature margin represents the safe distance between the current temperature and the Curie temperature, and the coercivity margin represents the safe distance between the operating point and the inflection point boundary; finally, the margin values ​​are standardized to form a physical constraint feature vector. The physical margin calculation formula is:

[0099] ;

[0100] ;in, For Curie temperature margin, Curie temperature, This is the maximum operating temperature; For coercivity margin, The magnetic field strength at the inflection point. The minimum operating point magnetic field strength, For coercivity.

[0101] Physical constraint features directly reflect the safety boundary of system operation and are advanced indicators based on physical laws, providing a mechanistic perspective to complement feature fusion.

[0102] This paper calculates the cross-domain correlation weights between data-driven features and physically constrained features based on an attention mechanism. Attention computation is a key technology for achieving intelligent fusion, optimizing the combination of features from different domains through dynamic weight allocation. The calculation process first projects the two types of features into a shared feature space; then, it calculates the correlation matrix between feature vectors; finally, it converts the features into attention weights using a softmax function. The formula for calculating cross-domain correlation is as follows:

[0103] ;in, This is the attention weight matrix. For the query matrix (generated from data-driven features), The key matrix (generated from physical constraint features) For feature dimensions.

[0104] The attention mechanism allows the model to dynamically adjust the importance of different features, intelligently focusing on the most relevant feature combinations based on the current state, thus improving the adaptability and accuracy of feature fusion. For different operating conditions and demagnetization stages, optimized attention allocation ensures that the model focuses on the most relevant metrics, improving the relevance and accuracy of state assessment.

[0105] Based on cross-domain correlation weights, data-driven features and physical constraint features are weighted and concatenated, and temperature attribution labels are added to generate a demagnetization state fusion feature vector. The temperature attribution label identifies the dominant temperature factor in the current magnetic flux change. Feature fusion is an integrated step in comprehensive analysis, forming a unified state representation by integrating multi-source information. The fusion process first weights data-driven features and physical constraint features according to attention weights; then, the weighted feature vectors are concatenated to form a combined feature; next, the correlation pattern between temperature and magnetic flux changes is analyzed to generate temperature attribution labels; finally, the labels are integrated with the combined features to form a complete demagnetization state fusion feature vector. The temperature attribution label is determined through causal analysis, assessing the contribution of temperature changes to magnetic flux decay and distinguishing between different types such as direct temperature effects, cumulative effects, and non-temperature factors. The fusion feature vector is a high-dimensional comprehensive state representation that simultaneously contains microscopic magnetic features, macroscopic physical constraints, and causal attribution information, providing a comprehensive feature foundation for subsequent demagnetization component decomposition. This enables the system to accurately distinguish performance changes caused by different reasons, improving the targeting of early warning and maintenance.

[0106] In this embodiment of the invention, the detailed implementation steps of decomposing the magnetic flux decay signal based on the demagnetization state fusion feature vector, separating the reversible thermal decay component and the irreversible demagnetization component through temperature attribution labels, and calculating the dynamic proportion coefficient between the reversible thermal decay component and the irreversible demagnetization component include:

[0107] Based on temperature attribution labels, the magnetic flux decay signal in the demagnetization state fusion feature vector is divided into temperature-dependent and temperature-independent components. Signal segmentation is the initial step in component decomposition, separating the mixed signal into components from different sources through causal correlation. The segmentation process first analyzes the distribution characteristics of temperature attribution labels to determine the dominant direction and intensity of temperature influence; then, a temperature-based regression model is constructed to fit the relationship between temperature and magnetic flux changes; finally, by comparing the model predictions with the actual decay signals, the signal is divided into components directly related to temperature and components caused by other factors. Temperature correlation analysis employs statistical methods such as hysteresis correlation and Granger causality tests to ensure the accuracy of the segmentation. The two types of signal components initially segmented lay the foundation for subsequent refined decomposition, helping to distinguish between true permanent demagnetization and recoverable thermal effects, and improving the accuracy of state assessment.

[0108] Temperature compensation calculations are performed on the temperature-dependent components to obtain the residual attenuation after eliminating the temperature effect. This residual attenuation is then labeled as the irreversible demagnetization component. Temperature compensation is a key technology for identifying permanent damage, separating temperature-independent performance losses through theoretical correction. The compensation process first establishes a standard temperature-magnetic flux response model based on the temperature coefficient curve of the permanent magnet material; then, it calculates the theoretical reversible change at the current temperature; finally, it subtracts the theoretical reversible change from the temperature-dependent components to obtain the residual attenuation that will still exist even after temperature recovery, i.e., the irreversible demagnetization component. The temperature compensation calculation formula is as follows:

[0109] ;in, This is an irreversible demagnetization component. For temperature-dependent components, This is a theoretically reversible change.

[0110] Theoretical reversible changes are typically calculated based on the material's temperature coefficient and historical response characteristics. Considering the material's nonlinear properties, piecewise linear or polynomial models are used to improve fitting accuracy. The irreversible demagnetization component directly reflects the permanent energy loss of the permanent magnet and is a core indicator for assessing the severity of demagnetization and predicting remaining lifetime.

[0111] The difference between the magnetic flux decay signal and the irreversible demagnetization component is labeled as the reversible thermal decay component. Difference calculation is a direct method to determine the recoverable portion, quantitatively assessing the reversible component through the difference between the total amount and the permanent loss. The calculation process involves subtracting the identified irreversible demagnetization component from the original magnetic flux decay signal; the resulting difference is the reversible thermal decay component. The calculation formula is:

[0112] ;in, It is a reversible thermal decay component. For the total attenuation signal, This is an irreversible demagnetization component.

[0113] The reversible thermal decay component represents the portion of performance that can be recovered as temperature decreases. It mainly originates from the reversible effect of temperature on the magnetic moment alignment and does not lead to permanent energy loss. This signal disappears automatically after the temperature returns to normal and does not require special maintenance intervention. However, its magnitude and trend help assess the sensitivity of permanent magnets to temperature changes and provide a reference for optimizing operating strategies.

[0114] The ratio of the amplitude of the reversible thermal decay component to the amplitude of the irreversible demagnetization component is calculated, and a sliding window smoothing process is applied to obtain the dynamic proportion coefficient. Proportion analysis is a relative indicator for evaluating the demagnetization state, intuitively reflecting the dominance of different mechanisms through the proportional relationship. The calculation process first determines the amplitudes of the two components, usually using signal energy or peak value as a measure; then, their ratio is calculated to form an initial proportion sequence; finally, a sliding window smoothing process is applied to eliminate the influence of short-term fluctuations and obtain a stable dynamic proportion coefficient. The calculation formula is:

[0115] ;in, For a moment The dynamic proportion coefficient, and These are the amplitudes of the reversible and irreversible components, respectively.

[0116] Sliding window smoothing typically employs an exponentially weighted moving average method, which preserves trend information while reducing noise interference. The dynamic proportion coefficient is an important state indicator; a high value indicates that reversible thermal effects dominate, and the system is mainly affected by temperature fluctuations; a low value indicates significant irreversible demagnetization, suggesting substantial damage to the system. The trend of this coefficient directly guides the formulation of early warning levels and maintenance strategies, serving as a crucial bridge connecting technical analysis and decision-making.

[0117] In this embodiment of the invention, based on the combined distribution characteristics of the dynamic proportion coefficient and the demagnetization critical discrimination factor, the early warning trigger threshold and maintenance strategy priority are dynamically adjusted, and the detailed implementation steps for outputting differentiated early warning levels and corresponding graded maintenance measures for permanent magnet pumps include:

[0118] A two-dimensional state space is constructed, with the dynamic proportion coefficient as the first dimension and the demagnetization critical discriminant factor as the second dimension. State space construction is a fundamental step in visualized decision-making, showcasing the overall system operating state through multi-dimensional mapping. The construction process first defines the coordinate axes: the horizontal axis represents the dynamic proportion coefficient R, reflecting the relative intensity of reversible and irreversible components; the vertical axis represents the demagnetization critical discriminant factor F, reflecting the degree to which the system approaches the critical state. Then, historical operating data is mapped onto this space to form a state distribution cloud map. Finally, density analysis and trajectory tracking are used to identify typical state clusters and evolution paths. The two-dimensional state space provides an intuitive representation of states, allowing complex multi-parameter states to be presented in a single view, facilitating pattern recognition and trend analysis. Each point in the space represents the health state of the permanent magnet pump at a specific moment, and the trajectory of the point reflects the evolution of the state over time, providing an intuitive basis for partitioning and threshold setting.

[0119] In a two-dimensional state space, the system is divided into a thermal attenuation-dominant zone, a critical transition zone, and a true demagnetization-dominant zone, with differentiated warning trigger thresholds configured for each zone. Zone division is the core strategy for risk grading, achieving fine-grained classification of states through spatial segmentation. The division process first analyzes the distribution characteristics of historical data in the state space to identify natural clusters and boundaries; then, combining expert knowledge and physical constraints, the boundary lines of the three main functional zones are determined; finally, differentiated warning trigger thresholds are set for each zone to reflect the risk level and response strategy of different zones. The thermal attenuation-dominant zone typically corresponds to high R-value and high F-value areas, indicating that the system is mainly affected by temperature fluctuations and the demagnetization risk is controllable; the critical transition zone is located in the middle zone, indicating that the system is in a critical state of reversible and irreversible change, requiring close monitoring; the true demagnetization-dominant zone corresponds to low R-value and low F-value areas, indicating that the system has suffered significant irreversible damage and requires immediate intervention. The warning trigger thresholds are set according to the characteristics of each zone, gradually increasing the warning intensity from low to high to ensure that the warning system is both sensitive and does not over-alarm, providing accurate risk indications for maintenance decisions.

[0120] Based on the current state point's position in the two-dimensional state space and its velocity vector, the evolution trajectory of the state point is predicted. Trajectory prediction is the technical foundation for forward-looking decision-making, assessing the future direction of risk development through trend extrapolation. The prediction process first calculates the velocity and acceleration vector of the current state point within the most recent time window; then, a state evolution model is constructed based on physical models and statistical methods, considering factors such as system inertia, external disturbances, and natural decay; finally, the possible trajectories over a future period are extrapolated through the model, and the arrival times and probabilities of key nodes are calculated. The prediction model employs an improved Kalman filter or recurrent neural network, balancing physical consistency and learning ability, and can adapt to the evolution characteristics under different operating conditions. The prediction results not only include the most probable trajectory path but also include uncertainty assessments, presenting the possibilities of different development paths in the form of probability distributions, providing a comprehensive forward-looking perspective for risk assessment.

[0121] When the trajectory points to the true demagnetization dominance zone, the maintenance strategy priority is increased and a permanent magnet replacement recommendation is output. When the trajectory remains in the thermal decay dominance zone, routine monitoring is maintained and a cooling operation recommendation is output. Decision-making is the final output of the early warning system, generating specific action recommendations through status assessment and trend prediction. The decision-making process first analyzes the direction and speed of the predicted trajectory to determine the urgency of risk development; then, based on the intersection relationship between the trajectory and various functional areas, it determines the main areas of future status; finally, based on preset decision rules, it outputs the corresponding early warning level and maintenance measures. When the trajectory points to the true demagnetization dominance zone and moves at a relatively high speed, the system triggers a high-level early warning, increases the maintenance strategy priority, and provides a clear permanent magnet replacement recommendation, including the recommended replacement time window and replacement range. When the trajectory stabilizes within the thermal decay dominance zone, the system maintains routine monitoring, periodically outputs operational status reports, and provides optimization suggestions such as cooling operation strategies and load adjustment schemes. For situations in the critical transition zone or with an unclear trajectory, the system recommends increasing the monitoring frequency and provides a risk assessment report to help decision-makers make balanced choices under uncertain conditions. This trajectory prediction-based differentiated decision-making mechanism ensures efficient allocation of maintenance resources and precise implementation of risk control, maximizing the safe operation and service life of permanent magnet pumps.

[0122] This invention achieves intelligent early warning and maintenance of permanent magnet pumps throughout their entire lifecycle through multi-source data acquisition, dynamic evolution modeling of hysteresis loops, operating point offset analysis, thermo-magnetic coupling characteristic evaluation, and differentiated early warning strategies. The thermo-magnetic coupling analysis method of this invention can accurately distinguish between reversible thermal decay and irreversible demagnetization, effectively identifying changes in the health status of permanent magnets, and providing a systematic solution for predictive maintenance of industrial equipment.

[0123] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0124] It should be noted that all formulas in this manual are calculated by removing dimensions and taking their numerical values. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.

[0125] Although embodiments of the invention have been shown and described, those skilled in the art will understand 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 claims and their equivalents.

Claims

1. A large-model-driven intelligent early warning and maintenance method for the entire life cycle of permanent magnet pumps, characterized in that, include: The electrical and thermodynamic parameter data of the permanent magnet pump during operation are obtained, and a three-dimensional temperature field distribution map of the permanent magnet is constructed based on the thermodynamic parameter data. Based on the electrical parameter data, the back electromotive force waveform is extracted, and combined with the three-dimensional temperature field distribution map, a dynamic evolution model of the hysteresis loop in each temperature gradient interval is established. Based on the hysteresis loop dynamic evolution model, the operating point offset of each partition of the permanent magnet under the current temperature condition is calculated, and the suspected demagnetization region at the critical demagnetization boundary is identified according to the relative position relationship between the operating point offset and the intrinsic demagnetization curve. Extract the historical temperature trajectory data of the suspected demagnetization area, and construct a thermal-magnetic coupling hysteresis feature map based on the peak temperature sequence and the corresponding magnetic flux response sequence in the historical temperature trajectory data. Based on the asymmetry of the magnetic flux recovery path and attenuation path in the thermo-magnetic coupling hysteresis characteristic spectrum, the reversible recovery coefficient and the irreversible damage coefficient are calculated, and the demagnetization critical discrimination factor is determined according to the ratio of the reversible recovery coefficient to the irreversible damage coefficient, including: Extract the magnetic flux decay path curve during the heating stage and the magnetic flux recovery path curve during the cooling stage from the thermo-magnetic coupling hysteresis feature map. Calculate the area difference between the magnetic flux recovery path curve and the magnetic flux decay path curve, and use it as the hysteresis loss area. The reversible recovery coefficient is calculated based on the ratio of the final magnetic flux value to the initial magnetic flux value of the magnetic flux recovery path curve. The irreversible damage coefficient is calculated based on the slope of the correlation between the hysteresis loss area and the cumulative temperature cycle number. The ratio of the reversible recovery coefficient to the irreversible damage coefficient is logarithmically transformed to obtain the demagnetization critical discrimination factor; Based on the demagnetization critical discrimination factor and the inflection point characteristics of the intrinsic demagnetization curve, a demagnetization state discrimination model embedded with physical boundary constraints is constructed. The output features of the hysteresis loop dynamic evolution model and the constraint output of the demagnetization state discrimination model are fused across domains to generate a demagnetization state fusion feature vector containing temperature attribution labels. Based on the demagnetization state fusion feature vector, the magnetic flux decay signal is decomposed by the temperature attribution label to separate the reversible thermal decay component and the irreversible demagnetization component, and the dynamic proportion coefficient between the reversible thermal decay component and the irreversible demagnetization component is calculated. Based on the combined distribution characteristics of the dynamic proportion coefficient and the demagnetization critical discrimination factor, the early warning trigger threshold and maintenance strategy priority are dynamically adjusted to output the differentiated early warning level and corresponding graded maintenance measures for the permanent magnet pump.

2. The method according to claim 1, characterized in that, The process of acquiring electrical and thermodynamic parameter data during the operation of the permanent magnet pump, and constructing a three-dimensional temperature field distribution map of the permanent magnet based on the thermodynamic parameter data, includes: Temperature time-series data from multiple measuring points on the surface of the permanent magnet are collected by a distributed temperature sensor array and used as the first component of the thermodynamic parameter data. The three-phase current waveform and back electromotive force waveform of the permanent magnet pump drive circuit are synchronously acquired by current sensor and voltage sensor, and used as the electrical parameter data. The copper and iron loss distributions of the permanent magnet pump are calculated based on the three-phase current waveforms, and the heat source intensity distribution inside the permanent magnet is inverted based on the copper and iron loss distributions. Using the temperature time series data as boundary conditions and the heat source intensity distribution as internal excitation, the three-dimensional temperature field distribution map is constructed by solving the heat conduction problem using the finite element method.

3. The method according to claim 1, characterized in that, The establishment of the dynamic evolution model of the hysteresis loop within each temperature gradient interval includes: Perform a short-time Fourier transform on the back electromotive force waveform to extract the fundamental amplitude sequence and harmonic distortion rate sequence; Based on the three-dimensional temperature field distribution map, the permanent magnet is divided into multiple temperature gradient intervals, and a spatial mapping relationship between each temperature gradient interval and the fundamental wave amplitude sequence is established. Based on the temperature coefficient curve of permanent magnet materials, the theoretical remanence and theoretical coercivity values ​​are calculated in each temperature gradient range. Based on the theoretical remanence value, the theoretical coercivity value, and the measured corresponding values ​​of the fundamental amplitude sequence, the shape parameters of the hysteresis loop in each temperature gradient interval are fitted to construct the dynamic evolution model of the hysteresis loop.

4. The method according to claim 1, characterized in that, Based on the hysteresis loop dynamic evolution model, the operating point offset of each partition of the permanent magnet at the current temperature is calculated, and the suspected demagnetization region at the critical demagnetization boundary is identified according to the relative positional relationship between the operating point offset and the intrinsic demagnetization curve, including: Extract the dynamic operating point coordinates of each partition under the current temperature state from the hysteresis loop dynamic evolution model; Obtain the intrinsic demagnetization curve of the permanent magnet material at a reference temperature, calculate the offset vector of the dynamic operating point coordinates relative to the corresponding reference operating point on the intrinsic demagnetization curve, and record the magnitude of the offset vector as the operating point offset. Calculate the normalized distance between the coordinates of the dynamic operating point and the inflection point of the intrinsic demagnetization curve, and denot it as the critical proximity. The partitions whose working point offset is greater than a first preset threshold and whose critical proximity is less than a second preset threshold are marked as the suspected demagnetization regions.

5. The method according to claim 1, characterized in that, The step involves extracting historical temperature trajectory data from the suspected demagnetization region and constructing a thermo-magnetic coupling hysteresis feature map based on the peak temperature sequence and corresponding magnetic flux response sequence in the historical temperature trajectory data, including: Extract historical temperature trajectory data of the suspected demagnetization area from the historical operation database, identify all peak temperature events that exceed the preset temperature benchmark, and form the peak temperature sequence. Extract the magnetic flux change data at the time of occurrence of each peak temperature event and during the preset recovery period thereafter to form the magnetic flux response sequence; Plot the thermal-magnetic response scatter distribution with the peak temperature sequence as the horizontal axis and the attenuation amplitude of the magnetic flux response sequence as the vertical axis. Kernel density estimation is performed on the scatter distribution of the thermo-magnetic response to generate the thermo-magnetic coupling hysteresis characteristic spectrum.

6. The method according to claim 1, characterized in that, The step of constructing a demagnetization state discrimination model embedded with physical boundary constraints based on the demagnetization critical discrimination factor and the inflection point characteristics of the intrinsic demagnetization curve includes: Extract the inflection point coordinates from the intrinsic demagnetization curve; Based on the Curie temperature parameters of permanent magnet materials, a Curie temperature constraint function is constructed; Based on the relative position of the demagnetization critical discrimination factor and the inflection point coordinates, a coercive force attenuation constraint function is constructed; The Curie temperature constraint function and the coercivity decay constraint function are embedded into the physical constraint layer of the neural network to construct the demagnetization state discrimination model.

7. The method according to claim 1, characterized in that, The step of fusing the output features of the hysteresis loop dynamic evolution model with the constraint output of the demagnetization state discrimination model across domains to generate a demagnetization state fusion feature vector containing temperature attribution labels includes: The hysteresis loop shape feature vectors for each temperature gradient interval are extracted from the dynamic evolution model of the hysteresis loop and used as data-driven features. Curie temperature margin and coercivity margin are extracted from the physical constraint layer of the demagnetization state discrimination model as physical constraint features; The cross-domain correlation weight between the data-driven features and the physical constraint features is calculated based on an attention mechanism. The data-driven features and physical constraint features are weighted and concatenated according to the cross-domain correlation weights, and a temperature attribution label is added to generate the demagnetization state fusion feature vector.

8. The method according to claim 1, characterized in that, The step of decomposing the magnetic flux decay signal based on the demagnetization state fusion feature vector, using the temperature attribution label, to separate the reversible thermal decay component and the irreversible demagnetization component, and calculating the dynamic proportion coefficient between the reversible thermal decay component and the irreversible demagnetization component, includes: Based on the temperature attribution label, the magnetic flux attenuation signal in the demagnetization state fusion feature vector is divided into temperature-dependent components and temperature-independent components. Temperature compensation calculations are performed on the temperature-related components to obtain the residual attenuation after eliminating the temperature effect, and the residual attenuation is marked as the irreversible demagnetization component. The difference between the magnetic flux attenuation signal and the irreversible demagnetization component is marked as the reversible thermal attenuation component; The ratio of the amplitude of the reversible thermal decay component to the amplitude of the irreversible demagnetization component is calculated, and a sliding window smoothing process is performed to obtain the dynamic proportion coefficient.

9. The method according to claim 1, characterized in that, The method involves dynamically adjusting the early warning trigger threshold and maintenance strategy priority based on the combined distribution characteristics of the dynamic proportion coefficient and the demagnetization critical discrimination factor, and outputting differentiated early warning levels and corresponding graded maintenance measures for the permanent magnet pump, including: Construct a two-dimensional state space with the dynamic proportion coefficient as the first dimension and the demagnetization critical discrimination factor as the second dimension; The two-dimensional state space is divided into a thermal decay-dominated region, a critical transition region, and a true demagnetization-dominated region, and a differentiated early warning trigger threshold is configured for each region. Based on the current position of the state point in the two-dimensional state space and its moving velocity vector, predict the evolution trajectory of the state point; When the evolution trajectory points to the true demagnetization-dominant region, the maintenance strategy priority is increased and a permanent magnet replacement suggestion is output; when the evolution trajectory remains in the thermal decay-dominant region, routine monitoring is maintained and a cooling operation suggestion is output.

Images