Micro-abrasive particle cluster recognition method, system, medium and device based on magnetic coupling characteristics

Abstract

The present application relates to the field of mechanical equipment health monitoring and intelligent diagnosis, and discloses a micro abrasive particle cluster identification method and system based on magnetic coupling characteristics, a medium and equipment, which comprises: taking the morphology, material and agglomeration structure of the micro abrasive particle cluster and the preset variable working condition parameters as inputs, establishing a multi-physical field fully coupled abrasive particle cluster finite element model for simulation calculation, and extracting a magnetic coupling feature set representing the magnetic coupling response of the micro abrasive particle cluster; a double-branch sparse auto-encoding network containing a classification branch and a regression branch is constructed as an abrasive particle cluster end-to-end intelligent identification model, the magnetic coupling feature set is used as training data, and a joint loss function integrating physical mechanisms is used for training until the abrasive particle cluster end-to-end intelligent identification model converges; the magnetic coupling feature set of the micro abrasive particle cluster to be measured is input into the trained abrasive particle cluster end-to-end intelligent identification, and the agglomeration type identification result and the agglomeration key parameter quantitative prediction result of the micro abrasive particle cluster are output.

Description

Technical Field

[0001] This invention relates to the field of mechanical equipment health monitoring and intelligent diagnosis technology, and in particular to a method, system, medium and device for identifying micro-abrasive particles based on magnetic coupling characteristics. Background Technology

[0002] For large-scale precision mechanical equipment, such as aerospace equipment and space devices, accurate assessment of their service capabilities and operational status is crucial for ensuring safe and reliable operation, improving information technology levels, and promoting the development of health management systems towards intelligent operation and maintenance. Early failures in large-scale precision mechanical equipment are closely related to abnormal wear of key components. Statistics show that approximately 70% to 80% of failures in large-scale precision mechanical equipment are caused by abnormal wear of components. Once such equipment fails, it inevitably causes enormous losses and even catastrophic consequences. Therefore, real-time monitoring of the wear condition of large-scale precision mechanical equipment is of significant practical importance for early detection of early wear failures and for preventing major equipment accidents.

[0003] Currently, the severe lack of high-precision, high-accuracy wear detection technologies remains a key bottleneck hindering the early wear fault detection of large precision mechanical equipment. Electromagnetic online oil abrasive particle detection technology has advantages such as strong anti-electromagnetic interference capability, good temperature stability, large allowable flow rate, and detection results unaffected by the cleanliness of the fluid medium, and has been gradually applied in the field of mechanical equipment wear monitoring. However, during wear monitoring, abrasive particles are easily formed into abrasive particle clusters due to the combined effects of fluid movement and electromagnetic fields. The differences in the size distribution of abrasive particle clusters, the randomness of fluid movement, changes in oil temperature, and the effect of alternating magnetic fields make the degree and state of aggregation of abrasive particle clusters in the detection area exhibit significant uncertainty. Furthermore, the abrasive particle clusters in different aggregation states cause diverse and complex disturbances to the sensor's magnetic field, significantly increasing the difficulty of identifying micro abrasive particle clusters and seriously affecting the accuracy and consistency of the next generation of high-precision online oil abrasive particle monitoring. Summary of the Invention

[0004] To address the aforementioned problems, the present invention aims to provide a method, system, medium, and device for identifying micro-abrasive aggregates based on magnetic coupling characteristics. This method enables accurate identification of abrasive aggregate types and quantitative prediction of key structural parameters of abrasive aggregates, thereby achieving accurate and intelligent identification of micro-abrasive aggregates under complex working conditions.

[0005] To achieve the above objectives, in a first aspect, the technical solution adopted by the present invention is as follows: a method for identifying micro-abrasive aggregates based on magnetic coupling characteristics, comprising: acquiring the morphology, material, and agglomeration structure of micro-abrasive aggregates under typical aggregation states; using the morphology, material, and agglomeration structure of the micro-abrasive aggregates, as well as preset variable working condition parameters, as input, establishing a multi-physics field fully coupled finite element model of the abrasive aggregates, and performing simulation calculations on the multi-physics field fully coupled finite element model of the abrasive aggregates to extract a magnetic coupling feature set characterizing the magnetic coupling response of the micro-abrasive aggregates; constructing a bi-branch sparse autoencoder network including classification and regression branches as an end-to-end intelligent identification model for abrasive aggregates, using the magnetic coupling feature set as training data, and training it using a joint loss function that integrates physical mechanisms until the end-to-end intelligent identification model of abrasive aggregates converges; inputting the magnetic coupling feature set of the micro-abrasive aggregates to be tested into the trained end-to-end intelligent identification model of abrasive aggregates, and outputting the agglomeration type identification result and the quantitative prediction result of the key agglomeration parameters of the micro-abrasive aggregates.

[0006] In some embodiments, the multiphysics fully coupled finite element model is a magnetic-fluid-solid-thermal four-field fully coupled model; establishing the multiphysics fully coupled finite element model includes: A magnetic field module was established, and the Maxwell control equations were solved using the magnetic vector potential method. A high-frequency eddy current loss equivalent correction + temperature coupling mechanism was introduced. The eddy current loss and temperature effect were equivalently integrated into the permeability parameter through the complex permeability correction formula to characterize the magnetic response under variable temperature alternating magnetic field. A fluid module was established, employing laminar flow control equations that combine flow boundary drive, temperature-dependent viscosity correction, and wall shear coupling to simulate the fluid motion environment of abrasive particles. A heat conduction module was established, with magnetic field eddy current loss and fluid viscosity dissipation as internal heat sources. The transient heat conduction control equation of oil-abrasive-pipe wall was established to realize the bidirectional coupling of temperature field with magnetic field and flow field, so as to characterize the heat transfer law under variable temperature conditions. A solid module was established, and an aggregated structure was constructed based on the measured parameters of abrasive particles in a typical aggregated state. The motion and micro-deformation of the abrasive particles were simulated by a dynamic mesh and adaptive time step coupling method. The magnetic field module, fluid module, heat conduction module, and solid module are coupled to form the fully coupled four-field magnetic-fluid-solid-thermal model.

[0007] In some embodiments, the magnetic coupling feature set includes: magnetic field time-domain steady-state fluctuation coefficient, frequency-domain eddy current frequency offset, spatial magnetic field disturbance non-uniformity, magnetohydrodynamic coupling dynamic response delay, and thermomagnetic linkage sensitivity coefficient. The magnetic coupling feature set extracted to characterize the magnetic coupling response of micro-abrasive aggregates includes: After performing simulation calculations on the multi-physics fully coupled abrasive particle finite model, the time series of magnetic induction intensity and the spatial magnetic field distribution cloud map of the sensor measurement points are output. Based on the magnetic induction intensity time series, the steady-state fluctuation coefficient of the magnetic field in the time domain is calculated to characterize the steady-state stability of the overall magnetic response of the abrasive particle cluster; The magnetic induction intensity time series is transformed in the frequency domain, the main response frequency is extracted and compared with the pure oil reference frequency, and the frequency domain eddy current frequency offset is calculated to correlate the difference between abrasive material and agglomeration density. Based on the spatial magnetic field distribution cloud map, the global magnetic field gradient and the standard deviation of its mean are calculated to obtain the spatial magnetic field disturbance non-uniformity, which reflects the overall ability of the aggregate to spatially aggregate and distort external magnetic fields. Based on the rising edge of the magnetic induction intensity time series, the characteristic moment when the calibration signal reaches the set steady-state amplitude is calculated to characterize the magnetization response speed of the agglomerates under the combined action of magnetic field excitation and fluid shear. At least two sets of typical service conditions at different temperatures are selected, and the average magnetic field value under the corresponding conditions is combined to calculate the thermomagnetic linkage sensitivity coefficient, so as to quantify the change of the magnetic response of micro-abrasive particles with temperature variation.

[0008] In some embodiments, a bi-branch sparse autoencoder network comprising a classification branch and a regression branch is constructed as an end-to-end intelligent identification model for abrasive particle clusters, including: A shared sparse coding backbone is constructed, which adopts a three-layer fully connected sparse coding shared layer and uses the ReLU activation function to uniformly compress the input features and extract the magnetic coupling low-level correlation features. KL divergence sparsity constraints are applied to the entire domain of the shared sparse coding backbone to strengthen the binding relationship between the extracted magnetic coupling low-level correlation features and the abrasive particle physical mechanism. A classification branch is constructed that is connected to the shared sparse coding backbone. The classification branch is a fully connected inference subnetwork and its end is connected to the Softmax classification output layer to output the qualitative identification result of the clustering type. A regression branch is constructed that is connected to the shared sparse coding backbone. The regression branch is structurally symmetrical and independent of the classification branch. Based on the extracted magnetic coupling low-level association features, the regression branch learns the nonlinear mapping relationship between the magnetic coupling features and the clustering key parameters. The end of the regression branch adopts linear activation output to output the continuous values ​​of the clustering key parameters for quantitative regression prediction. The clustering key parameters include equivalent particle size, cluster density, and arrangement dispersion.

[0009] In some embodiments, the KL divergence sparsity constraint is: ; In the formula, To preset sparse activation prior values, This represents the actual average activation level of neurons.

[0010] In some embodiments, the joint loss function of the fused physical mechanism is: ; In the formula, The combined total loss is constrained by category conditions; These are the weight hyperparameters for classification loss, regression loss, and physical constraint loss, respectively. For classification loss, cross-entropy loss is adopted and embedded with class conditional regularization constraints; For regression loss, mean squared error is used, plus a category conditional bias correction; The physical constraint loss is constructed based on the coupling mechanism of multiple physics fields of magneto-fluid-solid-thermal. The prediction results of the constraint model conform to the physical laws of magnetic response, kinematics and mechanics of abrasive particles.

[0011] In some embodiments, the method for training the dual-branch sparse autoencoder network is as follows: The joint loss function of the fused physical mechanism is used as the total loss, and a stratified random sampling method is used to divide the training set and the validation set according to the abrasive material, agglomeration structure type, agglomeration size, and temperature and flow rate conditions. The AdamW optimizer is used in conjunction with cosine annealing learning rate scheduling for parameter updates, and an early stopping mechanism is set up to terminate training early when the classification accuracy and regression average error on the validation set have not been optimized for several rounds.

[0012] Secondly, the technical solution adopted by the present invention is as follows: a micro-abrasive particle cluster identification system based on magnetic coupling characteristics, comprising: a micro-abrasive particle cluster morphology acquisition module, which acquires the morphology, material, and agglomeration structure of micro-abrasive particles under typical aggregation states; and a magnetic coupling feature extraction module, which, using the morphology, material, and agglomeration structure of the micro-abrasive particles and preset variable working condition parameters as input, establishes a multi-physics field fully coupled finite element model of the abrasive particles, and performs simulation calculations on the multi-physics field fully coupled finite element model of the abrasive particles to extract magnetic coupling features characterizing the magnetic coupling response of the micro-abrasive particles. The model consists of a coupled feature set and a model building and training module. The module constructs a bi-branch sparse autoencoder network (BCA) containing classification and regression branches as an end-to-end intelligent identification model for abrasive particles. The magnetic coupling feature set is used as training data, and the model is trained using a joint loss function that integrates physical mechanisms until the end-to-end intelligent identification model for abrasive particles converges. The identification module inputs the magnetic coupling feature set of the micro-abrasive particles to be tested into the trained end-to-end intelligent identification model for abrasive particles, and outputs the agglomeration type identification result and the quantitative prediction result of key agglomeration parameters for the micro-abrasive particles.

[0013] Thirdly, the technical solution adopted by the present invention is: a computer-readable storage medium for storing one or more programs, wherein the one or more programs include instructions, and when the instructions are executed by a computing device, the computing device performs any of the methods described above.

[0014] Fourthly, the technical solution adopted by the present invention is: a computing device comprising: one or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the one or more programs include instructions for performing any of the methods described above.

[0015] The present invention has the following advantages due to the adoption of the above technical solutions: 1. This invention can break through the limitations of traditional oil abrasive monitoring particle detection, fully consider the abrasive agglomeration effect and the actual service conditions of variable temperature and variable flow rate, and accurately restore the eddy current loss, magnetic disturbance and fluid-structure coupling characteristics of abrasive agglomerates through multi-field full coupling simulation, effectively eliminate the detection deviation caused by agglomeration interference and operating noise, and greatly improve the accuracy of characterizing the magnetic coupling characteristics of micro abrasive agglomerates under complex operating conditions.

[0016] 2. This invention adopts a method of identifying abrasive particles by actual measurement of abrasive particles, feature extraction by full-domain simulation of thermal-fluid-solid-magnetic coupling, and deep learning of physical constraints. It abandons the drawbacks of pure data-driven approaches, strengthens the correlation between features and physical mechanisms, and makes abrasive particle identification more interpretable and generalizable across working conditions. It can effectively solve the problems of poor interpretability of traditional models, easy failure of identification under varying working conditions, resulting in low early weak wear identification and inaccurate wear fault judgment.

[0017] 3. This invention simultaneously achieves qualitative classification of agglomeration types and quantitative inversion of agglomeration parameters through a dual-branch end-to-end network, enabling precise identification of equipment wear patterns and tracing of wear causes, thus achieving accurate early warning of minor wear. It can provide core technical support for health assessment, condition-based maintenance, and intelligent operation and maintenance throughout the entire lifecycle of lubrication systems in aerospace and other equipment. Attached Figure Description

[0018] Figure 1 This is a flowchart of the micro-abrasive particle cluster identification method based on magnetic coupling characteristics in an embodiment of the present invention; Figure 2 This is a flowchart illustrating the measurement process of the morphology of micro-abrasive particles in a typical state of polymerization in an embodiment of the present invention. Figure 3 This is a flowchart of the extraction of magnetic coupling characteristics of micro-abrasive particles under typical working conditions in an embodiment of the present invention; Figure 4 This is a typical end-to-end intelligent identification flowchart of abrasive particles based on a multi-physical-constrained bi-branch sparse autoencoder network in an embodiment of the present invention. Detailed Implementation

[0019] To address the shortcomings of existing online abrasive monitoring technologies, such as neglecting abrasive agglomeration effects and failing to consider the synergistic influence of temperature and flow rate variations in micro-abrasive particle identification, this invention proposes a micro-abrasive particle identification method, system, medium, and equipment based on magnetic coupling characteristics. Through a closed-loop technology encompassing typical aggregated micro-abrasive particle morphology measurement, extraction of magnetic coupling characteristics under typical operating conditions, and end-to-end intelligent identification of typical abrasive particles based on a multi-physical constraint bi-branch sparse autoencoder network, this invention achieves high-precision and robust identification of the physical characteristics of micro-abrasive particles under varying operating conditions. This invention solves the technical challenges of micro-abrasive particle identification under varying operating conditions, significantly improving the early, subtle wear identification capability of power lubrication systems in aerospace and other equipment. It facilitates precise wear level classification and accurate fault cause tracing, preventing the evolution of hidden early wear into major structural failures from the source, and ensuring the long-term stable, reliable, and safe operation of high-end precision combat equipment.

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention are within the scope of protection of the present invention.

[0021] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0022] In one embodiment of the present invention, a method for identifying micro-wear particles based on magnetic coupling characteristics is provided for real-time online monitoring of early wear conditions in large precision machinery under varying operating conditions. In this embodiment, as... Figure 1 As shown, the method includes the following steps: 1) Obtain the morphology, material and agglomeration structure of micro-abrasive particles under typical polymerization conditions.

[0023] 2) Using the morphology, material and agglomeration structure of the micro-abrasive particles and the preset variable working condition parameters as inputs, a multi-physics field fully coupled finite element model of the abrasive particles is established, and the multi-physics field fully coupled finite element model of the abrasive particles is simulated and calculated to extract the magnetic coupling feature set characterizing the magnetic coupling response of the micro-abrasive particles.

[0024] 3) Construct a bi-branch sparse autoencoder network containing classification and regression branches as an end-to-end intelligent identification model for abrasive particles. Use magnetic coupling feature set as training data and train it using a joint loss function that integrates physical mechanisms until the end-to-end intelligent identification model for abrasive particles converges.

[0025] 4) Input the magnetic coupling feature set of the micro-abrasive particles to be tested into the trained end-to-end intelligent identification of abrasive particles, and output the agglomeration type identification result and the quantitative prediction result of the key agglomeration parameters of the micro-abrasive particles.

[0026] In use, this invention uses microscopic observation and energy dispersive spectroscopy to quantitatively measure key parameters such as material, agglomeration structure, and geometric dimensions of micro-abrasive particles; it establishes a fully coupled finite element model of abrasive particles considering the influence of temperature-flow rate variations, and extracts the core features of magnetic coupling; it constructs a bi-branch sparse autoencoder end-to-end network that integrates electromagnetic physical constraints, and establishes a highly robust and interpretable nonlinear mapping relationship between magnetic coupling features and the real physical properties of abrasive particles, thereby achieving accurate identification of abrasive particle agglomeration types and quantitative prediction of key structural parameters of agglomeration, and completing accurate and intelligent identification of micro-abrasive particles under complex working conditions.

[0027] In one possible implementation, step 1) above involves obtaining the morphology, material composition, and agglomeration structure of micro-abrasive particles in a typical aggregated state, such as... Figure 2 As shown, the specific implementation method is as follows: In the lubricating medium, abrasive particles move with the fluid and collide with each other, agglomerating under the action of van der Waals forces and electrostatics. As the intensity of the action increases, the agglomerates successively exhibit three typical structures: linear chain-like (initial adhesion), clustered loose (intermediate aggregation), and dense spherical (completely dense), corresponding to different agglomeration states of abrasive particles from weak to strong. The morphology, material, and agglomeration structure of abrasive particles are the core physical basis for detecting magnetic coupling response and intelligent identification. This embodiment aims to obtain quantifiable micro-abrasive particle benchmark input parameters that can be directly imported into finite element simulation and intelligent identification models through precise measurement of the morphology of micro-abrasive particles under three typical aggregation structures.

[0028] In this embodiment, the material discrimination of micro-abrasive particles is specifically as follows: for standard samples of micro-abrasive particles with three typical agglomeration morphologies, a high-precision microtweezers microscopic observation system is used to quantitatively measure the structural parameters of typical agglomerated micro-abrasive particles formed by the aggregation of multiple particles.

[0029] Microscopic energy dispersive spectroscopy (EDS) is a mainstream technique for micro-area material analysis. It can rapidly and non-destructively obtain elemental composition, enabling qualitative identification of the composition of micro-abrasive particle clusters. Let... ∈{ The three typical abrasive particle aggregation structures—linear chain-like, clustered loose, and dense spherical aggregated—were respectively identified using a high-precision microtweezers microscopic observation system. Equivalent particle size... It is a quantitative parameter that transforms irregular, non-spherical abrasive aggregates into comparable equivalent spherical sizes, and has clear physical and engineering significance for understanding wear states and fault evolution.

[0030] The formula for calculating equivalent particle size is: (1) in, The total pixel area of ​​the aggregate.

[0031] The aspect ratio of aggregates is defined as the ratio of the longest axis to the shortest axis (or equivalent width) of abrasive agglomerates. It quantitatively characterizes the morphological evolution of agglomerates from linear chains and loose clusters to dense spherical structures, reflecting the anisotropy, density, and shear stability of the agglomerate structure. It provides a key morphological parameter for distinguishing wear stages and identifying agglomeration types. The calculation formula is: (2) in, The length of the major axis of the minimum circumscribed rectangle of the aggregate. This is the length of the minor axis.

[0032] Aggregation compactness coefficient It can characterize the effective filling ratio inside abrasive grains and effectively distinguish between loose / dense agglomerates. The calculation formula is as follows: (3) in, To identify the entire connected region of abrasive particles in the preprocessed microscopic image, the total number of effective pixels of all abrasive materials within the connected region is counted, and the sum of the actual pixel areas of the abrasive region inside the abrasive particle cluster is calculated by combining the pixel-size calibration coefficient. This indicates the total number of abrasive particles contained within a micro-abrasive agglomerate; The total area of ​​the minimum circumscribed contour of the abrasive grain cluster; The larger the agglomeration, the more compact the agglomeration, and the stronger the classification and differentiation of micro-abrasive particles.

[0033] Spatial arrangement dispersion It can quantify the degree of disorder in particle arrangement and is specifically used to distinguish linear chains. With disordered clusters The calculation formula is: (4) in, Taking the abrasive aggregate as the whole as the analysis object, the center coordinates of each local feature region inside it are extracted, and the polar angle of each region center relative to the centroid of the abrasive aggregate as a whole is calculated to obtain the arrangement azimuth angle; This is the average azimuth angle of all particle clusters.

[0034] Equivalent diameter of aggregated pores It can characterize the average gap size inside abrasive particles, directly correlate with oil penetration, magnetic flux penetration, and flow field resistance, and can be used to connect to the boundary of subsequent magnetic-fluid-solid simulations. The calculation formula is: (5) in, Aggregation equivalent particle size; The above-mentioned aggregation compactness coefficient.

[0035] In one possible implementation, the multiphysics fully coupled finite element model in step 2) above is a four-field fully coupled model of magneto-fluid-solid-thermal. Due to the influence of magnetic coupling effect, the magnetic field disturbance characteristics caused by micro-abrasive particles are closely related to their aggregate structure and scale. The dynamic changes in the aggregate state of micro-abrasive particles also make their magnetic properties exhibit complex characteristics. Based on the real physical parameters of abrasive particles measured by microscopy, this embodiment constructs a fully coupled magneto-fluid-solid-thermal model that considers high-frequency eddy current loss, wall shear coupling, dynamic agglomeration deformation, and temperature-flow multi-condition coupling. This model realizes the extraction of magnetic coupling characteristics of micro-abrasive particles under typical working conditions, and overcomes the inherent defects of traditional magnetic coupling characteristic extraction, which only uses single or two-field coupling, ignores eddy current effect, flow shear, and the cross-influence of temperature and flow, and is difficult to characterize the dynamic magnetic response of micro-abrasive particles.

[0036] A full-factor orthogonal combination simulation method for quantifying the parameters of micro-abrasive particles was adopted, with the abrasive material ( ), three types of aggregation structures ( ), equivalent particle size Polymer aspect ratio Aggregation compactness coefficient Spatial arrangement dispersion Equivalent diameter of aggregated pores For structural parameters, select This represents a typical oil flow rate condition. Compared to sparse sampling and univariate progressive simulation, full-factor operating conditions, which represent typical oil temperature conditions, can fully cover the actual service parameter space, fill in boundary conditions and weakly coupled sensitive conditions, reduce parameter cross-coupling interference, strengthen the mapping relationship between magnetic coupling characteristics and micro-wear particle morphology and operating conditions, reduce the difficulty of model fitting, and significantly improve the physical consistency and cross-condition generalization recognition ability of subsequent networks.

[0037] Specifically, using measured parameters of typical aggregated abrasive particles (abrasive material, equivalent particle size, agglomeration arrangement, particle spacing, and agglomeration density) and flow rate and temperature parameters as boundary inputs, a three-dimensional transient "magnetic-fluid-solid-thermal" bidirectional coupled solution model adapted to three types of agglomeration morphologies is established. This model incorporates high-frequency complex permeability eddy current equivalent correction, wall shear fluid-solid coupling constraints, a dynamic mesh adaptive dynamic agglomeration drift solution mechanism, temperature-dependent property correction, and a flow rate-driven flow field iteration mechanism to extract amplitude coupling characteristics, energy coupling characteristics, axial magnetic field gradient characteristics, magnetic pulse timing characteristics, frequency domain response characteristics, and magnetic field spatial distribution characteristics. The establishment of the multi-physics fully coupled finite element model includes the following steps: 2.1) Establish a magnetic field module, use the magnetic vector potential method to solve Maxwell's control equations, introduce a high-frequency eddy current loss equivalent correction + temperature coupling mechanism, and use the complex permeability correction formula to integrate eddy current loss and temperature influence into the permeability parameter to characterize the magnetic response under variable temperature alternating magnetic field.

[0038] In this embodiment, specifically addressing the significant abrasive eddy current effect under high-frequency alternating magnetic fields and the large errors in traditional constant magnetic permeability modeling caused by temperature-induced changes in material permeability and conductivity, the magnetic vector potential method is used to solve Maxwell's governing equations, achieving accurate solutions for the spatial distribution and transient changes of the magnetic field. The core governing equations are: (6) in, It is a magnetic vector potential; Temperature-dependent permeability; Temperature-dependent conductivity; For excitation current density; yes The Hamiltonian operator (also known as the Hamiltonian operator) is used in electromagnetism to represent spatial partial derivative operations. This formula can directly characterize the curl of a magnetic field, is suitable for multi-medium and variable-temperature coupling scenarios, and accurately solves the conduction and coupling laws of magnetic fields in oil and abrasive particles.

[0039] A high-frequency eddy current loss equivalent correction and temperature coupling mechanism are introduced. The eddy current loss and temperature influence are equivalently incorporated into the permeability parameter through a complex permeability correction formula, simplifying the solution process while ensuring accuracy. The correction formula is as follows: (7) in, The complex permeability after temperature-frequency coupling correction; The temperature-dependent static real permeability of the abrasive particles; To excite the magnetic field angular frequency; The imaginary unit; The characteristic angular frequency of the eddy current is given by the formula, which can effectively characterize the magnetic response hysteresis and energy loss characteristics of abrasive particles under high-frequency magnetic fields and variable temperature conditions, and correct the calculation errors of traditional models.

[0040] , From step 1), the measured abrasive material ( Determined by combining temperature experimental curves; The actual detection system parameters are set; Substitute into formula (6) Complete the modeling of the magnetic field module.

[0041] 2.2) Establish a fluid module and adopt laminar flow control equations with flow boundary drive, temperature-dependent viscosity correction and wall shear coupling to simulate the fluid motion environment of abrasive particles.

[0042] In this embodiment, industrial standard hydraulic oil is used as the simulation medium, and laminar flow is employed, characterized by flow boundary drive, temperature-dependent viscosity correction, and wall shear coupling. The governing equations closely match the actual conditions of low-speed laminar flow and variable temperature and flow rate of oil, accurately simulating the motion environment of abrasive particles. The core governing equations are: (8) in, For fluid density; The fluid velocity vector (based on inlet flow rate) (Calculated) For fluid pressure; Temperature-dependent dynamic viscosity; This is the wall shear stress correction term. By introducing temperature viscosity correction and flow-driven boundary conditions, this formula accurately characterizes the shearing action between the pipe wall, oil, and abrasive particles under varying temperature and flow conditions, thus improving the accuracy of flow field simulation.

[0043] fluid density The value is determined by the industrial standard hydraulic oil model; Fitting the oil viscosity-temperature experimental curve; Determined by pipe wall roughness and pipe inner diameter; inlet flow rate Set the gradient according to the actual operating conditions of the lubrication system; combine the measured abrasive particle parameters and temperature from step 1). Substitute into formula (8) to complete the flow field module modeling.

[0044] 2.3) Establish a heat conduction module, using magnetic field eddy current loss and fluid viscosity dissipation as internal heat sources, and establish transient heat conduction control equations for oil-abrasive particles-pipe wall to achieve bidirectional coupling between temperature field and magnetic field and flow field, so as to characterize the heat transfer law under variable temperature conditions.

[0045] In this embodiment, specifically, a transient heat conduction control equation is established for oil-abrasive particles-pipe wall. Magnetic field eddy current loss and fluid viscous dissipation are used as internal heat sources to achieve bidirectional coupling between the temperature field and the magnetic and flow fields. The core control equation is: (9) in, Specific heat capacity at constant pressure; The thermal conductivity is temperature-dependent; T is the temperature. The heat source for magnetic field eddy current loss is obtained by solving formula (6); It is a heat source for fluid viscosity heat dissipation (obtained by solving formula (8)); this equation realizes real-time bidirectional coupling of temperature field and magnetic and flow field, and accurately characterizes the heat transfer law under variable temperature conditions.

[0046] Isobaric specific heat capacity Temperature-dependent thermal conductivity Determined based on experimental data of oil and abrasive materials; , The output is provided in real time by the magnetic and flow field modules; the temperature boundary is set by the actual working conditions to complete the modeling of the heat conduction module.

[0047] 2.4) Establish a solid module, construct an agglomeration structure based on the measured parameters of abrasive particles in a typical aggregated state, and use a dynamic mesh and adaptive time step coupling method to simulate the motion and micro-deformation of abrasive particles.

[0048] In this embodiment, specifically, the modeling is based on the microscopically measured parameters of typical aggregated abrasive particles, and the agglomeration structure... Represents a linear chain-like aggregation type. Represents a clustered, loosely aggregated type. Representing a dense, spherical aggregate type, the abrasive material is ferrous Fe and nonferrous NFe. Therefore, the agglomerated particle sizes of the three typical aggregate structures are as follows: The aggregation scales of the three typical aggregated abrasive particles are as follows: The agglomeration spacing of the three typical polymeric abrasive agglomerates are as follows: The agglomeration density of three typical polymeric abrasive particles is divided into: .

[0049] Traffic flow is categorized based on its numerical value. ;in Represents high traffic volume. Representing medium flow, Represents low flow rate; temperature is categorized based on numerical value. .in This represents a high oil temperature. Represents the temperature of the oil in the middle layer. This indicates a low oil temperature.

[0050] A coupled method of dynamic mesh and adaptive time step is used to simulate the motion and micro-deformation of abrasive particles. The stability of the solution is ensured by displacement constraints, which are as follows: (10) in, This represents the displacement of abrasive particles in a single time step. for The threshold. During modeling. Set a threshold according to the stability requirements of the solution, and select a value based on the stability requirements of multiphysics coupling solutions. This enables real-time bidirectional coupled solution of magnetic field, flow field, and solid motion.

[0051] 2.5) Couple the magnetic field module, fluid module, heat conduction module and solid module to form a fully coupled four-field magnetic-fluid-solid-thermal model.

[0052] In one possible implementation, the magnetic coupling feature set in step 2) above includes: magnetic field time-domain steady-state fluctuation coefficient, frequency domain eddy current frequency offset, spatial magnetic field disturbance non-uniformity, magnetohydrodynamic coupling dynamic response delay, and thermomagnetic linkage sensitivity coefficient.

[0053] Based on the inherent parameters of the abrasive particles obtained by microscopic measurement, an additional oil flow rate was added. Oil temperature Two types of actual service variable condition parameters are used to construct a full-factor simulation variable set; a four-dimensional fully coupled finite element model of "magnetism-fluid-solid-thermal" is established to carry out full-coverage variable condition simulation calculations. After each set of conditions completes transient coupled iterative convergence, a unified output of the magnetic induction intensity time series at the sensor measurement points is generated. Space magnetic field distribution cloud map Two types of core simulation data serve as the original data sources for feature extraction.

[0054] To overcome the limitations of traditional abrasive feature extraction techniques that rely on static single-condition analysis, extract only time-domain features, and ignore the coupling interference between temperature and fluid velocity, this paper summarizes five standardized, reproducible, and algorithm-adaptive core magnetic coupling features from five dimensions: time-domain steady-state fluctuations, frequency-domain eddy current losses, spatial magnetic field disturbances, magnetofluid dynamic response, and thermomagnetic correlation characteristics. In this embodiment, as shown... Figure 3 As shown, the magnetic coupling feature set extracted to characterize the magnetic coupling response of micro-abrasive aggregates includes: (1) After performing simulation calculations on the finite model of abrasive particles with full coupling of multiple physics fields, output the time series of magnetic induction intensity and the spatial magnetic field distribution cloud map of the sensor measurement points.

[0055] (2) Based on the magnetic induction intensity time series, the steady-state fluctuation coefficient of the magnetic field in the time domain is calculated to characterize the steady-state stability of the overall magnetic response of the abrasive particle cluster.

[0056] Specifically, based on the magnetic induction intensity time-series signal Calculate the steady-state fluctuation coefficient of the magnetic field in the time domain : (11) in, The magnetic flux density at a single moment. The time series mean This represents the number of sampling points. This feature characterizes the steady-state stability of the overall magnetic response of abrasive aggregates under the effects of flow velocity shearing and temperature softening. Higher flow velocities and temperatures increase the likelihood of micro-deformation and localized loosening of the aggregates, leading to increased signal fluctuations. The synchronous increase can intuitively distinguish the resistance to flow and temperature disturbance of the three types of aggregate structures.

[0057] (3) Perform frequency domain transformation on the magnetic induction intensity time series, extract the main response frequency and compare it with the pure oil reference frequency to calculate the frequency domain eddy current frequency offset, so as to correlate the difference between abrasive material and agglomeration density.

[0058] Specifically, for the timing signal of magnetic induction intensity conduct Transformation to extract the main response frequency Compare with the reference frequency of pure oil. Calculate the frequency offset of the eddy current in the frequency domain : (12) in, To obtain the inherent response frequency of the sensor under corresponding temperature and flow conditions in pure oil without abrasive particles, through multiphysics simulation or field calibration. This feature focuses on the temperature-induced changes in magnetic permeability and conductivity, as well as the differences in eddy current losses caused by flow velocity disturbances. Increased temperature enhances eddy current heating within the agglomerates, and increased flow velocity exacerbates magnetic field timing distortion, ultimately manifesting as an amplified dominant frequency shift. This can be correlated with both the abrasive material and the agglomerate density differences.

[0059] (4) Based on the spatial magnetic field distribution cloud map, calculate the global magnetic field gradient and the standard deviation of its mean to obtain the spatial magnetic field disturbance inhomogeneity, so as to reflect the overall ability of the aggregate to spatially aggregate and distort the external magnetic field.

[0060] Specifically, the inhomogeneity of space magnetic field disturbance is solved by relying on the global gradient data of the space magnetic field cloud map. : (13) in, For the magnetic field gradient of each spatial grid, The mean of the global gradient. This represents the number of spatial sampling grids. This feature reflects the overall spatial aggregation and distortion capability of the aggregate to an external magnetic field under different temperature and flow rate conditions; the denser and larger the aggregate, the stronger the local magnetic field disturbance. The higher the value, the less affected by temporal noise, and the better suited for static spatial feature identification.

[0061] (5) Based on the rising edge of the magnetic induction intensity time series, the characteristic moment when the calibration signal reaches the steady-state amplitude setting is determined, and the magnetofluid coupling dynamic response delay is calculated to characterize the magnetization response speed of the agglomerate under the combined action of magnetic field excitation and fluid shear.

[0062] Specifically, the dynamic response delay of magnetofluid coupling is calibrated based on the characteristics of the rising edge of magnetic induction intensity. : (14) in, , These represent the characteristic moments when the signal reaches 10% and 90% of its steady-state amplitude, respectively. This characteristic characterizes the speed of the overall magnetization response of the aggregates under the combined effect of magnetic field excitation and fluid shear. High flow velocity accelerates particle transport, while high temperature weakens magnetic binding force; both of these shorten the response delay, accurately mapping the regulation law of dynamic magnetic response under flow velocity conditions.

[0063] (6) Select at least two sets of typical service conditions at different temperatures, and calculate the thermomagnetic linkage sensitivity coefficient in combination with the mean magnetic field value under the corresponding conditions, so as to quantify the change of magnetic response of micro-abrasive particles with temperature.

[0064] Specifically, two sets of typical service temperatures were selected. The thermomagnetic linkage sensitivity coefficient is calculated by combining the mean magnetic field value under the corresponding operating conditions. : (15) In the formula, The thermomagnetic coupling response coefficient is used to quantify the variation of the magnetic response of micro-abrasive particles with temperature-dependent working conditions, serving as a core indicator for identifying the degree of temperature sensitivity. The temperature is The average value of the magnetic response signal of the micro-abrasive particles, corresponding to the magnetic signal intensity at the initial reference temperature; The temperature is The average value of the magnetic response signal of the micro-abrasive particles, corresponding to the magnetic signal intensity under strain temperature conditions; The initial reference temperature is set to a value typical of low-temperature operating conditions and is used to calibrate the initial state of thermomagnetic coupling. The test temperature is a variable temperature value, taken from a typical high-temperature operating condition, and is... This constitutes a temperature gradient variable operating condition. This feature exclusively characterizes the quantitative impact of temperature variation on the magnetic signal of agglomerates, making up for the deficiency of conventional simulations that ignore thermomagnetic coupling; ferromagnetic agglomerates are much more sensitive to temperature changes than non-ferromagnetic agglomerates, and their dense structure has stronger thermomagnetic stability, which is a key unique feature for material and structure identification in variable temperature scenarios.

[0065] In this embodiment, the above five types of features are... Synchronously bind aggregate structure tags, material tags, and flow rate and temperature condition tags to construct a complete sample set that takes into account both inherent physical properties and variable operating condition disturbances. Its mathematical expression is as follows: (16) In the formula, the input feature vector Magnetic response feature set extracted from multiphysics coupling simulation of magneto-fluid-solid-thermal fields: (17) In the formula, For local extrema of the magnetic field, For magnetic pole position, For magnetic coupling coefficient, Core features include magnetic pulse width.

[0066] Category Tags Corresponding to the three typical polymerization states: linear chain-like, clustered loose, and dense spherical. ∈{ }; Regression label vector Covering the multi-gradient physical parameters of abrasive particles: (18) In the formula, The particle size of the aggregated particles (corresponding to) ); For the aggregation scale (corresponding to) ); Cluster spacing (corresponding to) ); For the density of aggregate (corresponding to) Simultaneously, the sample set embedding temperature (correspond ),flow (correspond By analyzing the interference characteristics of various working conditions, the system directly connects to the subsequent dual-branch sparse autoencoder network to achieve accurate and intelligent identification of abrasive particles under all working conditions.

[0067] In one possible implementation, step 3) above uses standardized magnetic coupling features as a unified input and employs a heterogeneous dual-branch decoupled architecture of classification branch + regression branch to build an end-to-end inference link. The classification branch achieves qualitative identification of cluster types, while the regression branch completes quantitative prediction of key cluster parameters. Both branches share underlying sparse coding features, with upper-level task-specific decoupled learning, balancing feature commonality mining with task-specific adaptation. This achieves accurate mapping from magnetic coupling features to the physical properties of abrasive particles, ensuring high recognition accuracy for cluster structure, material, and other category features in the classification task, while also satisfying high-precision inversion of continuous parameters such as equivalent particle size and density in the regression task. Ultimately, this completes integrated intelligent identification from magnetic signals to the full-dimensional physical features of abrasive particles, providing reliable quantitative basis for refined assessment of equipment wear status and early fault warning. Figure 4 As shown. Specifically, a bi-branch sparse autoencoder network containing classification and regression branches is constructed as an end-to-end intelligent identification model for abrasive particle clusters, including the following steps: 3.1) Construct a shared sparse coding backbone. The shared sparse coding backbone adopts a three-layer fully connected sparse coding shared layer and uses the ReLU activation function to uniformly compress the input features and extract the magnetic coupling low-level correlation features. Apply KL divergence sparsity constraints to the entire domain of the shared sparse coding backbone to strengthen the binding relationship between the extracted magnetic coupling low-level correlation features and the abrasive physical mechanism.

[0068] in, The function is: (19) In the formula, Input weighted values ​​into the neurons; Computational lightweighting can alleviate gradient vanishing in deep networks and is suitable for subsequent engineering deployments on the edge; global application of shared layers The divergence sparsity constraint filters redundant noise and strengthens the binding relationship between features and abrasive physical mechanisms.

[0069] Wherein, the KL divergence sparsity constraint is: ; In the formula, To preset sparse activation prior values, This represents the actual average activation level of neurons. This constraint forces the network to retain only core features strongly correlated with electromagnetic and thermal coupling mechanisms, suppresses redundant activation, and improves the robustness and physical interpretability of the dual-task model.

[0070] 3.2) Construct a classification branch connected to the shared sparse coding backbone. The classification branch is a fully connected inference subnetwork, and its end is connected to the Softmax classification output layer to adapt to multi-class probability distribution inference and to output qualitative identification results of clustering type.

[0071] Specifically, the classification branch builds a dedicated fully connected inference subnetwork based on shared coding features, focusing on clustering structures. Material Category identification.

[0072] 3.3) Construct a regression branch connected to the shared sparse coding backbone. The regression branch and the classification branch are structurally symmetrical and independent. The regression branch is based on the extracted magnetic coupling low-level association features and learns the nonlinear mapping relationship between the magnetic coupling features and the clustering key parameters. The end of the regression branch uses linear activation output to output the continuous values ​​of the clustering key parameters for quantitative regression prediction. Among them, the clustering key parameters include equivalent particle size, cluster density and arrangement dispersion.

[0073] In this embodiment, a heterogeneous dual-branch sparse autoencoder network integrating multiple constraints is constructed to achieve end-to-end intelligent identification of typical abrasive aggregates. This network can perform unsupervised deep feature learning on high-dimensional, multi-condition coupled magnetic coupling features through sparsity constraints, automatically filtering out operating noise and redundant information, achieving effective dimensionality reduction and physical correlation enhancement of magnetic features, and providing robust and interpretable core feature inputs for dual-branch tasks. Its encoder-decoder shared structure naturally adapts to the heterogeneous output of classification and regression dual tasks, sharing the underlying representation of magnetic features in the encoder stage, and completing qualitative identification of aggregate types and quantitative inversion of key parameters in the decoder branch, respectively, achieving end-to-end integrated collaborative modeling and avoiding the problems of feature fragmentation and asynchronous training in traditional multi-task models. Simultaneously, sparse constraints can effectively suppress model overfitting, improve the generalization ability to boundary conditions, small samples, and measured noise data, and, combined with category-conditional constraint joint loss optimization, further ensure the data fitting accuracy and physical consistency of the model output, enabling the network output to closely correlate with the real material, aggregate structure, and service condition parameters of the abrasive aggregates, meeting the accuracy and reliability requirements of intelligent equipment operation and maintenance monitoring.

[0074] Before step 3.1) above, a data preprocessing step is also included: For the acquired variable operating condition magnetic coupling feature set, effective samples are first screened based on the signal-to-noise ratio (SNR) threshold, eliminating abnormal feature samples with low SNR and strong operating condition interference to avoid the influence of dirty data on the classification boundary and regression fitting. The time-domain signal-to-noise ratio formula is used to determine the validity of the samples. The data preprocessing SNR calculation formula is as follows: (20) In the formula, For the first Signal-to-noise ratio of each magnetic coupling feature sample (unit: ), For the first The average power of the effective magnetic response signal in each sample corresponds to the true characteristic component generated by the magnetic coupling of abrasive particles; For the first The average power of the variable operating condition interference noise in each sample corresponds to irrelevant interference components such as oil turbulence, temperature fluctuation, and electromagnetic environment. For the first The first sample magnetic coupling characteristic signal Amplitude of each sampling point ( This corresponds to the magnetic disturbance signal sequence acquired through multiphysics simulation and experiments; This represents the total number of sampling points for the effective magnetic response signal. For the first The first of the variable operating condition interference noise in the sample amplitude at each sampling point The signal was extracted from the blank oil (without abrasive particles) under varying operating conditions and used to calibrate the power of pure interference noise. This represents the total number of sampling points for interference noise under varying operating conditions. This invention sets a signal-to-noise ratio threshold. ,when When a sample is deemed valid, it is retained for subsequent model training; when Samples identified as abnormal are removed.

[0075] After sample selection, quantile robust standardization is used to perform dimensionless unification on the high-dimensional magnetic coupling feature vectors, eliminating differences in the dimensions and magnitudes of different features, while suppressing the interference of extreme outliers on the standardization results and improving the stability of model training. The core formula is: (twenty one) In the formula, For the first The first valid sample, the first The standardized eigenvalues ​​of the 3D magnetic coupling characteristics; For the corresponding original feature values; For the first The 50th percentile (median) of the feature in the entire effective sample is less sensitive to extreme outliers than the traditional mean, thus avoiding dirty data from interfering with the standardized benchmark. , The first The difference between the 25th and 75th percentiles of the dimensional feature in the entire effective sample. Interquartile range It is used to replace the traditional standard deviation, and only uses the middle 50% of the data to estimate the scale, completely avoiding the impact of extreme outliers and greatly improving the robustness of scale estimation.

[0076] This formula maps each feature dimension to a standard distribution with a median of 0 and interquartile range normalization, balancing the weight contribution of different features in the backpropagation of the network, avoiding high-amplitude features from dominating the training gradient, and providing high-quality and robust input data for subsequent two-branch sparse autoencoder networks.

[0077] In one possible implementation, step 3) above addresses the problems of discrete parameter fitting for similar samples and blurred classification boundaries in the process of simultaneously performing micro-abrasive particle agglomeration classification and key physical parameter regression in a dual-branch sparse autoencoder network. This is achieved by embedding the inherent electromagnetic coupling mechanism between the agglomeration structure and physical parameters into the model optimization process, thereby improving the stability of the dual-task collaborative convergence and the generalization ability under varying working conditions. Category constraints are introduced to achieve collaborative convergence of qualitative identification and quantitative prediction. The joint loss function integrating the physical mechanism is: ; In the formula, The category-conditional constraint joint total loss is used for training the bi-branch sparse autoencoder network, simultaneously optimizing the dual tasks of micro-abrasive particle aggregation classification and key physical parameter regression, achieving coordinated convergence of qualitative identification and quantitative prediction. These are the weight hyperparameters for classification loss, regression loss, and physical constraint loss, respectively. They are used to balance the gradient magnitudes of the three types of losses, ensuring the stability and convergence of dual-task training. They can be adaptively adjusted according to the model training effect. For classification loss, cross-entropy loss is adopted and category conditional regularization constraints are embedded. Relying on the prior physical boundary of the cluster category, the cluster compactness of the features of the same sample is constrained, ensuring clear cluster category identification boundaries and adapting to multi-class classification tasks. For regression loss, mean squared error (MSE) is used. ), fitting key aggregation parameters ( , By combining the continuous true values ​​of (etc.) with the category conditional bias correction, the accuracy of parameter prediction within the same cluster category is improved. To address the physical constraint loss, a multi-physics coupling mechanism based on magneto-fluid-solid-thermal is constructed. The constraint model prediction results conform to the physical laws of magnetic response, kinematics, and mechanics of abrasive particles, avoiding physical inconsistencies in data-driven models and improving the reliability and generalization ability of identification results under all working conditions.

[0078] In one possible implementation, the method for training the dual-branch sparse autoencoder network in step 3) above is as follows: A joint loss function incorporating physical mechanisms is used as the total loss. A stratified random sampling method is employed to divide the training and validation sets according to abrasive material, agglomeration structure type, agglomeration scale, and temperature and flow rate conditions. The AdamW optimizer combined with cosine annealing learning rate scheduling is used for parameter updates, and an early stopping mechanism is implemented to terminate training prematurely when the classification accuracy and regression average error of the validation set have not improved for several rounds.

[0079] Specifically, the stratified dataset is divided strictly according to the abrasive material, aggregate structure category, aggregate scale, and temperature-flow rate conditions. The training set and validation set are divided proportionally to ensure balanced classification samples, continuous gradient of regression parameters, and to prevent class bias and parameter gaps.

[0080] A combined training strategy of AdamW optimizer, cosine annealing learning rate scheduling, and early stopping mechanism is adopted. AdamW dynamically adjusts the parameter update step size and constrains the parameter scale, cosine annealing achieves smooth decay of the learning rate, and the early stopping mechanism avoids overfitting. This approach adapts to the convergence requirements of dual-task end-to-end models and ensures generalization ability and engineering accuracy.

[0081] This training strategy is adaptable to high-dimensional, multi-condition magnetic coupling features (temperature, flow rate, material, and agglomeration density coupling) inputs, dual-branch heterogeneous tasks (agglomeration type classification and key parameter regression), and small-sample, noisy real-world testing scenarios. It addresses issues such as gradient oscillations, slow convergence, and overfitting that easily occur in simple training, while balancing the accuracy of both tasks. It enhances the model's generalization ability under different typical temperature and flow rate conditions, ensuring stable model migration from simulation data to engineering measurement data. The optimizer selected... The optimizer completes parameter iteration and updates, and the core momentum update rules are as follows: (twenty two) in, For first-order and second-order momentum estimation, For gradient, The momentum decay coefficient, For learning rate, It is a minimal zero constant; The optimizer optimizes the weight decay logic to effectively suppress overfitting in the dual-task model; the learning rate incorporates cosine annealing for dynamic scheduling. (twenty three) in, For real-time learning rate, , These are the upper and lower limits of the learning rate. To maximize the number of training rounds and ensure rapid convergence in the early stages and fine-grained fitting in the later stages, an early stopping mechanism is added: real-time monitoring of the classification accuracy and regression average error on the validation set, terminating training early when both metrics show no improvement after multiple rounds. For typical micro-abrasive particle cluster categories and small-parameter gradient samples, dynamic adjustment of category conditional loss weights combined with incremental sample enhancement for targeted optimization ultimately achieves end-to-end intelligent identification results with accurate micro-cluster type classification and high-precision quantitative prediction of key cluster parameters.

[0082] In one embodiment of the present invention, a micro-abrasive particle cluster identification system based on magnetic coupling characteristics is provided, comprising: The micro-abrasive particle morphology acquisition module acquires the morphology, material, and agglomeration structure of micro-abrasive particles in a typical aggregated state; The magnetic coupling feature extraction module takes the morphology, material and agglomeration structure of the micro-abrasive particles and the preset variable working condition parameters as inputs to establish a multi-physics field fully coupled finite element model of the abrasive particles, and performs simulation calculations on the multi-physics field fully coupled abrasive particles finite element model to extract a set of magnetic coupling features characterizing the magnetic coupling response of the micro-abrasive particles. The model building and training module constructs a bi-branch sparse autoencoder network containing classification and regression branches as an end-to-end intelligent identification model for abrasive particles. The magnetic coupling feature set is used as training data, and the joint loss function that integrates physical mechanisms is used for training until the end-to-end intelligent identification model for abrasive particles converges. The identification module inputs the magnetic coupling feature set of the micro-abrasive particles to be tested into the trained end-to-end intelligent identification of abrasive particles, and outputs the agglomeration type identification result and the quantitative prediction result of the key agglomeration parameters of the micro-abrasive particles.

[0083] In one possible implementation, the multiphysics fully coupled finite element model is a four-field fully coupled model of magnetism, fluidity, solid-structure interaction, and thermal motion; establishing the multiphysics fully coupled finite element model includes: A magnetic field module was established, and the Maxwell control equations were solved using the magnetic vector potential method. A high-frequency eddy current loss equivalent correction + temperature coupling mechanism was introduced. The eddy current loss and temperature effect were equivalently integrated into the permeability parameter through the complex permeability correction formula to characterize the magnetic response under variable temperature alternating magnetic field. A fluid module was established, employing laminar flow control equations that combine flow boundary drive, temperature-dependent viscosity correction, and wall shear coupling to simulate the fluid motion environment of abrasive particles. A heat conduction module was established, with magnetic field eddy current loss and fluid viscosity dissipation as internal heat sources. The transient heat conduction control equation of oil-abrasive-pipe wall was established to realize the bidirectional coupling of temperature field with magnetic field and flow field, so as to characterize the heat transfer law under variable temperature conditions. A solid module was established, and an aggregated structure was constructed based on the measured parameters of abrasive particles in a typical aggregated state. The motion and micro-deformation of the abrasive particles were simulated by a dynamic mesh and adaptive time step coupling method. The magnetic field module, fluid module, heat conduction module, and solid module are coupled to form a fully coupled four-field model of magnetic-fluid-solid-thermal.

[0084] In one possible implementation, the magnetic coupling feature set includes: magnetic field time-domain steady-state fluctuation coefficient, frequency-domain eddy current frequency offset, spatial magnetic field disturbance inhomogeneity, magnetohydrodynamic coupling dynamic response delay, and thermomagnetic linkage sensitivity coefficient. The magnetic coupling feature set extracted to characterize the magnetic coupling response of micro-abrasive aggregates includes: After simulating the finite model of abrasive particles with full multi-physics coupling, the time series of magnetic induction intensity and the spatial magnetic field distribution cloud map of the sensor measurement points are output. Based on the time series of magnetic induction intensity, the steady-state fluctuation coefficient of the magnetic field in the time domain is calculated to characterize the steady-state stability of the overall magnetic response of the abrasive aggregate; The magnetic induction intensity time series was transformed in the frequency domain, the main response frequency was extracted and compared with the pure oil reference frequency, and the frequency domain eddy current frequency offset was calculated to correlate the difference between abrasive material and agglomeration density. Based on the spatial magnetic field distribution cloud map, the global magnetic field gradient and the standard deviation of its mean are calculated to obtain the spatial magnetic field disturbance inhomogeneity, which reflects the overall ability of the aggregate to spatially aggregate and distort the external magnetic field. Based on the rising edge of the magnetic induction intensity time series, the characteristic moment when the calibration signal reaches the steady-state amplitude setting is determined, and the magnetofluid coupling dynamic response delay is calculated to characterize the magnetization response speed of the agglomerates under the combined action of magnetic field excitation and fluid shear. At least two sets of typical service conditions at different temperatures are selected, and the average magnetic field value under the corresponding conditions is combined to calculate the thermomagnetic linkage sensitivity coefficient, so as to quantify the change of magnetic response of micro-abrasive particles with temperature variation.

[0085] In one possible implementation, a bi-branch sparse autoencoder network containing classification and regression branches is constructed as an end-to-end intelligent identification model for abrasive particle clusters, including: A shared sparse coding backbone is constructed, which adopts a three-layer fully connected sparse coding shared layer and uses the ReLU activation function to uniformly compress the input features and extract the magnetic coupling low-level correlation features. KL divergence sparsity constraints are applied to the entire domain of the shared sparse coding backbone to strengthen the binding relationship between the extracted magnetic coupling low-level correlation features and the physical mechanism of abrasive particles. A classification branch is constructed and connected to the sparse coding backbone. The classification branch is a fully connected inference subnetwork, and its end is connected to the Softmax classification output layer to output the qualitative identification results of the clustering type. A regression branch is constructed and connected to the shared sparse coding backbone. The regression branch and the classification branch are structurally symmetrical and independent. The regression branch is based on the extracted magnetic coupling low-level association features and learns the nonlinear mapping relationship between the magnetic coupling features and the clustering key parameters in a directional manner. The end of the regression branch adopts linear activation output to output the continuous values ​​of the clustering key parameters for quantitative regression prediction. Among them, the clustering key parameters include equivalent particle size, cluster density and arrangement dispersion.

[0086] In one possible implementation, the KL divergence sparsity constraint is: ; In the formula, To preset sparse activation prior values, This represents the actual average activation level of neurons.

[0087] In one possible implementation, the joint loss function that incorporates the physical mechanisms is: ; In the formula, The combined total loss is constrained by category conditions; These are the weight hyperparameters for classification loss, regression loss, and physical constraint loss, respectively. For classification loss, cross-entropy loss is adopted and embedded with class conditional regularization constraints; For regression loss, mean squared error is used, plus a category conditional bias correction; The physical constraint loss is constructed based on the coupling mechanism of multiple physics fields of magneto-fluid-solid-thermal. The prediction results of the constraint model conform to the physical laws of magnetic response, kinematics and mechanics of abrasive particles.

[0088] In one possible implementation, the method for training a dual-branch sparse autoencoder network is as follows: A joint loss function incorporating physical mechanisms is used as the total loss. A stratified random sampling method is employed to divide the training set and validation set according to the abrasive material, agglomeration structure type, agglomeration size, and temperature and flow rate conditions. The AdamW optimizer is used in conjunction with cosine annealing learning rate scheduling for parameter updates, and an early stopping mechanism is set up to terminate training early when the classification accuracy and regression average error on the validation set have not been optimized for several rounds.

[0089] The system provided in this embodiment is used to execute the above-described method embodiments. For specific processes and details, please refer to the above embodiments, which will not be repeated here.

[0090] In summary, traditional electromagnetic detection of oil abrasive particles is mostly based on the analysis of single-particle magnetic properties, ignoring the engineering reality that abrasive particles easily agglomerate under the coupling of fluid shear, temperature changes, and electromagnetic fields in actual service conditions. This lack of characterization of abrasive particle agglomeration effects and failure to consider their significant impact on detection results leads to inconsistencies in the accuracy and consistency of online oil abrasive particle detection. Furthermore, existing oil abrasive particle monitoring technologies do not consider abrasive particle eddy current losses and hysteresis effects under high-frequency magnetic fields, nor the magnetic response deviations caused by varying operating conditions such as oil flow rate and temperature. This makes it impossible to accurately characterize the magnetic coupling disturbance laws under different abrasive particle aggregate structures, sizes, and materials, resulting in poor adaptability of existing oil abrasive particle detection technologies under variable temperature and flow conditions. Conventional abrasive particle identification models mostly adopt a purely data-driven approach, lacking constraints from electromagnetic physical mechanisms. This leads to problems such as weak feature correlation, poor generalization ability, and easy identification failure under varying operating conditions, easily resulting in low accuracy in early minor wear identification and inaccurate wear fault assessment. This invention achieves efficient purification of magnetic coupling characteristics and accurate mapping of physical properties of micro-abrasive aggregates through precise measurement of the microstructure and material parameters of typical aggregated abrasive particles, feature extraction via fully coupled magnetic-fluid-solid-thermal simulation, and intelligent identification using a dual-branch sparse autoencoder network. It enables qualitative identification of agglomeration types and quantitative inversion of key parameters under varying operating conditions, providing reliable technical support for wear condition monitoring and early fault warning in lubrication systems of aerospace and other equipment. This invention effectively solves the problems of poor adaptability to varying temperature and flow conditions, lack of agglomeration effect characterization, insufficient modeling of magnetic coupling mechanisms, and weak generalization ability of intelligent identification in existing oil abrasive detection technologies. It can be widely applied in online abrasive detection of hydraulic lubrication systems, health monitoring of mechanical equipment, and intelligent operation and maintenance.

[0091] In one embodiment of the present invention, a computing device is also provided. This computing device can be a terminal, and may include: a processor, a communication interface, memory, a display screen, and an input device. The processor, communication interface, and memory communicate with each other via a communication bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and computer programs. When the computer programs are executed by the processor, they implement the methods described in the above embodiments. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The communication interface is used for wired or wireless communication with external terminals. Wireless communication can be achieved through Wi-Fi, a network management system, NFC (Near Field Communication), or other technologies. The display screen can be a liquid crystal display or an e-ink display. The input device can be a touch layer covering the display screen, or buttons, a trackball, or a touchpad mounted on the casing of the computing device, or an external keyboard, touchpad, or mouse. The processor can call logical instructions in the memory to execute the methods described in the above embodiments.

[0092] In one embodiment of the present invention, a computer program product including instructions is also provided. This computer program product may be a software or program product including instructions and capable of running on a computing device or stored on any available medium. When the computer program product is run on at least one computing device, it causes the at least one computing device to perform the methods and functions of any of the foregoing embodiments.

[0093] In one embodiment of the invention, a computer-readable storage medium is also provided. The computer-readable storage medium can be any available medium accessible to a computing device, or a data storage device (e.g., a data center) that includes one or more available media. The available media can be magnetic media (e.g., floppy disks, hard disks, or magnetic tapes), optical media (e.g., DVDs), semiconductor media (e.g., solid-state drives), etc. The computer-readable storage medium includes instructions, and the instructions instruct the computing device to perform the methods and functions in any of the foregoing embodiments.

[0094] Computer-readable media can be any tangible medium that includes or stores a program for executing instructions on or associated with a system, apparatus, or device, or a data storage device, such as a data center, including one or more available media. Computer-readable media can be computer-readable signal media or computer-readable storage media. Computer-readable media can include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, any suitable combination thereof. More detailed examples of computer-readable storage media include electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0095] Furthermore, although the operations of the methods disclosed in this invention are described in a specific order in the accompanying drawings, this does not require or imply that these operations need to be performed in a specific order, or that all the operations shown need to be performed to achieve the desired result. Rather, the order of execution of the operations shown in the flowcharts can be changed. Additionally or alternatively, some operations may be omitted, multiple operations may be combined into one operation for execution, and / or one operation may be decomposed into multiple operations for execution. It should also be noted that the features and functions of two or more devices disclosed in this invention may be specific in one device. Conversely, the features and functions of the aforementioned one device may be further specific in multiple devices.

[0096] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for identifying micro-abrasive particle clusters based on magnetic coupling characteristics, characterized in that, include: To obtain the morphology, material composition, and agglomeration structure of micro-abrasive particles in a typical aggregated state; Using the morphology, material and agglomeration structure of the micro-abrasive particles, as well as preset variable working condition parameters as input, a multi-physics field fully coupled finite element model of the abrasive particles is established, and the multi-physics field fully coupled finite element model of the abrasive particles is simulated and calculated to extract a magnetic coupling feature set characterizing the magnetic coupling response of the micro-abrasive particles. A bi-branch sparse autoencoder network containing classification and regression branches is constructed as an end-to-end intelligent identification model for abrasive particles. The magnetic coupling feature set is used as training data, and the model is trained using a joint loss function that integrates physical mechanisms until the end-to-end intelligent identification model for abrasive particles converges. The magnetic coupling feature set of the micro-abrasive particles to be tested is input into the trained end-to-end intelligent identification of the abrasive particles, and the agglomeration type identification result and the quantitative prediction result of the key agglomeration parameters of the micro-abrasive particles are output.

2. The method of claim 1, wherein the magnetic coupling characteristic is a magnetic field strength. The multiphysics fully coupled finite element model is a four-field fully coupled model of magnetism, fluidity, solid-structure interaction, and thermal motion; establishing the multiphysics fully coupled finite element model includes: A magnetic field module was established, and the Maxwell control equations were solved using the magnetic vector potential method. A high-frequency eddy current loss equivalent correction + temperature coupling mechanism was introduced. The eddy current loss and temperature effect were equivalently integrated into the permeability parameter through the complex permeability correction formula to characterize the magnetic response under variable temperature alternating magnetic field. A fluid module was established, employing laminar flow control equations that combine flow boundary drive, temperature-dependent viscosity correction, and wall shear coupling to simulate the fluid motion environment of abrasive particles. A heat conduction module was established, with magnetic field eddy current loss and fluid viscosity dissipation as internal heat sources. The transient heat conduction control equation of oil-abrasive-pipe wall was established to realize the bidirectional coupling of temperature field with magnetic field and flow field, so as to characterize the heat transfer law under variable temperature conditions. A solid module was established, and an aggregated structure was constructed based on the measured parameters of abrasive particles in a typical aggregated state. The motion and micro-deformation of the abrasive particles were simulated by a dynamic mesh and adaptive time step coupling method. The magnetic field module, fluid module, heat conduction module, and solid module are coupled to form the fully coupled four-field magnetic-fluid-solid-thermal model.

3. The method for identifying micro-abrasive aggregates based on magnetic coupling characteristics as described in claim 1, characterized in that, The magnetic coupling feature set includes: magnetic field time-domain steady-state fluctuation coefficient, frequency-domain eddy current frequency offset, spatial magnetic field disturbance non-uniformity, magnetohydrodynamic coupling dynamic response delay, and thermomagnetic linkage sensitivity coefficient. The magnetic coupling feature set extracted to characterize the magnetic coupling response of micro-abrasive aggregates includes: After performing simulation calculations on the multi-physics fully coupled abrasive particle finite model, the time series of magnetic induction intensity and the spatial magnetic field distribution cloud map of the sensor measurement points are output. Based on the magnetic induction intensity time series, the steady-state fluctuation coefficient of the magnetic field in the time domain is calculated to characterize the steady-state stability of the overall magnetic response of the abrasive particle cluster; The magnetic induction intensity time series is transformed in the frequency domain, the main response frequency is extracted and compared with the pure oil reference frequency, and the frequency domain eddy current frequency offset is calculated to correlate the difference between abrasive material and agglomeration density. Based on the spatial magnetic field distribution cloud map, the global magnetic field gradient and the standard deviation of its mean are calculated to obtain the spatial magnetic field disturbance non-uniformity, which reflects the overall ability of the aggregate to spatially aggregate and distort external magnetic fields. Based on the rising edge of the magnetic induction intensity time series, the characteristic moment when the calibration signal reaches the set steady-state amplitude is calculated to characterize the magnetization response speed of the agglomerates under the combined action of magnetic field excitation and fluid shear. At least two sets of typical service conditions at different temperatures are selected, and the average magnetic field value under the corresponding conditions is combined to calculate the thermomagnetic linkage sensitivity coefficient, so as to quantify the change of the magnetic response of micro-abrasive particles with temperature variation.

4. The method for identifying micro-abrasive particles based on magnetic coupling characteristics as described in claim 1, characterized in that, A bi-branch sparse autoencoder network containing classification and regression branches is constructed as an end-to-end intelligent identification model for abrasive particle clusters, including: A shared sparse coding backbone is constructed, which adopts a three-layer fully connected sparse coding shared layer and uses the ReLU activation function to uniformly compress the input features and extract the magnetic coupling low-level correlation features. KL divergence sparsity constraints are applied to the entire domain of the shared sparse coding backbone to strengthen the binding relationship between the extracted magnetic coupling low-level correlation features and the abrasive particle physical mechanism. A classification branch is constructed that is connected to the shared sparse coding backbone. The classification branch is a fully connected inference subnetwork and its end is connected to the Softmax classification output layer to output the qualitative identification result of the clustering type. A regression branch is constructed that is connected to the shared sparse coding backbone. The regression branch is structurally symmetrical and independent of the classification branch. Based on the extracted magnetic coupling low-level association features, the regression branch learns the nonlinear mapping relationship between the magnetic coupling features and the clustering key parameters. The end of the regression branch adopts linear activation output to output the continuous values ​​of the clustering key parameters for quantitative regression prediction. The clustering key parameters include equivalent particle size, cluster density, and arrangement dispersion.

5. The method for identifying micro-abrasive aggregates based on magnetic coupling characteristics as described in claim 4, characterized in that, The KL divergence sparsity constraint is: ； In the formula, To preset sparse activation prior values, This represents the actual average activation level of neurons.

6. The method for identifying micro-abrasive aggregates based on magnetic coupling characteristics as described in claim 1, characterized in that, The joint loss function of the fused physical mechanism is: ； In the formula, The combined total loss is constrained by category conditions; These are the weight hyperparameters for classification loss, regression loss, and physical constraint loss, respectively. For classification loss, cross-entropy loss is adopted and embedded with class conditional regularization constraints; For regression loss, mean squared error is used, plus a category conditional bias correction; The physical constraint loss is constructed based on the multi-physics coupling mechanism of magneto-fluid-solid-thermal fields. The prediction results of the constraint model conform to the physical laws of magnetic response, kinematics and mechanics of abrasive particles.

7. The method for identifying micro-abrasive aggregates based on magnetic coupling characteristics as described in claim 1, characterized in that, The method for training the dual-branch sparse autoencoder network is as follows: The joint loss function of the fused physical mechanism is used as the total loss, and a stratified random sampling method is used to divide the training set and the validation set according to the abrasive material, agglomeration structure type, agglomeration size, and temperature and flow rate conditions. The AdamW optimizer is used in conjunction with cosine annealing learning rate scheduling for parameter updates, and an early stopping mechanism is set up to terminate training early when the classification accuracy and regression average error on the validation set have not been optimized for several rounds.

8. A micro-abrasive particle cluster identification system based on magnetic coupling characteristics, characterized in that, include: The micro-abrasive particle morphology acquisition module acquires the morphology, material, and aggregation structure of micro-abrasive particles in a typical aggregated state. The magnetic coupling feature extraction module takes the morphology, material and agglomeration structure of the micro-abrasive particles and preset variable working condition parameters as inputs to establish a multi-physics field fully coupled finite element model of the abrasive particles, and performs simulation calculations on the multi-physics field fully coupled abrasive particles finite element model to extract a set of magnetic coupling features characterizing the magnetic coupling response of the micro-abrasive particles. The model building and training module constructs a bi-branch sparse autoencoder network containing classification and regression branches as an end-to-end intelligent identification model for abrasive particles. The magnetic coupling feature set is used as training data, and the model is trained using a joint loss function that integrates physical mechanisms until the end-to-end intelligent identification model for abrasive particles converges. The identification module inputs the magnetic coupling feature set of the micro-abrasive particles to be tested into the trained end-to-end intelligent identification of the abrasive particles, and outputs the agglomeration type identification result and the quantitative prediction result of the key agglomeration parameters of the micro-abrasive particles.

9. A computer-readable storage medium for storing one or more programs, characterized in that, The one or more programs include instructions that, when executed by a computing device, cause the computing device to perform any of the methods described in claims 1 to 7.

10. A computing device, characterized in that, include: One or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the one or more programs including instructions for performing any of the methods described in claims 1 to 7.

Images