A Method and System for Diagnosing Abnormal Noises in Components Based on Nonlinear Dynamic Neural Networks

Abstract

The application provides a component abnormal sound diagnosis method based on a nonlinear dynamics neural network, and belongs to the technical field of acoustics, vibration signal processing and artificial intelligence, and comprises the following steps: constructing a multi-dimensional vibration-sound state feature vector according to collected three-axis acceleration of an excitation source and microphone sound pressure signals of a response source, and inputting the multi-dimensional vibration-sound state feature vector into a double-branch neural network, wherein the neural network comprises an abnormal sound mechanism classification branch neural network based on a deep nonlinear physical constraint loss function, which is used for analyzing the type of abnormal sound; and a material feature decoupling branch neural network based on fractional calculus, which is used for analyzing the material producing the abnormal sound.The application can solve the technical problem that the BSR abnormal sound AI automatic diagnosis scheme based on deep learning in the prior art has poor accuracy in abnormal sound identification and is difficult to guide component abnormal sound rectification work.

Description

Technical Field

[0001] This invention relates to the interdisciplinary technical fields of vehicle engineering, acoustics and vibration signal processing, artificial intelligence and nonlinear dynamics, and specifically to a method and system for diagnosing abnormal noises in components based on a nonlinear dynamics neural network. Background Technology

[0002] In bench testing of automotive component noises, the traditional approach to diagnosing BSR noises (a collective term for three types of non-functional noises: Buzz, Squeak, and Rattle) relies primarily on engineers using stethoscopes for subjective manual assessment. In recent years, with the increasing automation in the industry, AI-based automated diagnostic systems / products for BSR noises based on deep learning have gradually emerged in the market. The basic structure and functions of mainstream AI noise diagnosis products are as follows: Hardware structure: Primarily configured with single-dimensional acoustic acquisition hardware, i.e., unidirectional microphones or microphone arrays are arranged around the test bench, specifically for recording the sound signals generated by the vibration of components. Algorithm and software functions: After receiving pure sound data, the system's internal signal processing module first converts the one-dimensional audio into a two-dimensional acoustic feature map (for example, by generating a frequency-division spectrogram through short-time Fourier transform (STFT) or extracting Mel-frequency cepstral coefficients (MFCC)). Then, these image-based acoustic features are input into a conventional convolutional neural network (such as CNN, ResNet, and other image classification models), and finally outputs the macroscopic classification result of the abnormal sound, that is, informing the tester whether the current abnormal sound belongs to Buzz (resonance), Squeak (friction), or Rattle (tapping).

[0003] The aforementioned existing technologies have the following significant drawbacks and shortcomings in practical engineering applications: Defect 1: The "pure acoustic black box" misconception caused by the simple hardware structure results in extremely poor resistance to background noise interference. Existing technology relies solely on acoustic sensors, completely detached from the physical motion state (acceleration, velocity, displacement) applied to the sample by the test bench. This single-dimensional "pure data-driven" blind testing model often faces significant challenges in actual industrial settings. When the multi-degree-of-freedom excitation table itself generates high-frequency mechanical or fluid noise, relying solely on acoustic characteristics can easily misjudge the equipment noise as abnormal noise from components. Due to the lack of a comparison with the physical signal of the excitation source, the existing system has extremely low robustness under harsh background noise, reducing the accuracy of neural networks in identifying abnormal noise from components.

[0004] Defect 2: The algorithm design lacks constraints on physical causality, posing a risk of misjudgment that "violates common sense in physics." In classical mechanics, the generation of abnormal noises has extremely strict physical causal laws. However, existing conventional neural networks (such as CNNs) only perform probability fitting on the image features of sound waveforms, without incorporating Newtonian mechanics and tribological mechanisms as constraints into the model. This leads to current AI frequently making predictions that violate fundamental physical laws, resulting in weak model generalization ability; it also reduces the accuracy of neural networks in identifying abnormal noises from parts.

[0005] Defect 3: The diagnostic function is too coarse-grained, lacking "material-level" root cause analysis capabilities, making it difficult to directly guide rectification. The current technology's functional limit is only to outputting macroscopic "noise classification (B / S / R)," unable to identify the physical properties of the sound-generating interface through in-depth soundprint decoupling. Because it cannot output specific "sound-generating material combinations" (for example, the system cannot distinguish whether the current friction sound comes from "PP plastic and ABS plastic" or "metal clips and rubber"), noise engineers still need to spend a lot of time blindly disassembling parts to find the real root cause of the noise after receiving the diagnostic report, greatly reducing its engineering practical value and guiding significance. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention proposes a component noise diagnosis method and system based on nonlinear dynamic neural networks. This method is applicable to the extraction of multimodal features of BSR (Body Sound Resonance) noise in components under semi-anechoic chamber bench testing environments, classification based on complex mechanical mechanisms, and intelligent diagnosis at the material level. This addresses the technical problems of existing deep learning-based AI-based automatic BSR noise diagnosis schemes, which suffer from poor resistance to background noise interference, prediction results that violate fundamental physical laws in some cases, and an inability to distinguish the sound-generating interface of the noise, resulting in poor accuracy in noise identification and difficulty in guiding component noise rectification.

[0007] The technical solution adopted in this invention is as follows: Firstly, a method for diagnosing abnormal noises from components based on a nonlinear dynamic neural network is provided, comprising the following steps: A multidimensional vibration-acoustic state feature vector is constructed based on the triaxial acceleration signal of the excitation source and the microphone sound pressure signal of the response source. The physical state and acoustic feature map matrix are obtained by mapping based on the multidimensional vibration-acoustic state feature vector. The physical state and acoustic feature map matrix are input into a nonlinear dynamic physical information dual-branch neural network to obtain the types of abnormal noises and the materials that produce them.

[0008] Furthermore, a multidimensional vibration-acoustic state feature vector is constructed, including: Based on triaxial acceleration signals, the relative velocity and relative displacement between the two components of the potential abnormal noise-generating interface are obtained by time-domain integration. Based on the sound pressure signal from the microphone array, the high-frequency envelope of the sound pressure signal is extracted. The high-frequency envelope of the sound pressure signal is then hard-aligned in the time domain with the physical motion quantity excited in the same location to obtain a multidimensional vibration-acoustic state feature vector.

[0009] Furthermore, the physical state and acoustic feature map matrix is ​​obtained by mapping based on the multidimensional vibration-acoustic state feature vector, including: mapping the evolution trajectory of the multidimensional vibration-acoustic state feature vector in phase space into the physical state and acoustic feature map matrix.

[0010] Furthermore, the nonlinear dynamic physical information dual-branch neural network includes an abnormal noise mechanism classification branch neural network based on a deep nonlinear physical constraint loss function, used to analyze the types of abnormal noise; and a material feature decoupling branch neural network based on fractional calculus, used to analyze the materials that produce abnormal noise.

[0011] Furthermore, the total loss function of the abnormal sound mechanism classification branch neural network is: in For classification cross-entropy, Adaptive weights; constrained loss function The forced abnormal noise mechanism classification branch neural network follows the constraints on frictional sound based on the LuGre dynamic friction model, and further, the constraints on frictional sound based on the LuGre dynamic friction model include: The stick-slip effect of the micro-friction surface is abstracted into a differential equation that incorporates the deformation of elastic bristles; Calculate the dynamic interfacial friction force of micro-friction surfaces; The frictional noise constraint requires that the envelope of the frictional noise predicted by the neural network be constrained by the gradient of the frictional dissipation power density function.

[0012] Furthermore, based on the constraints of the nonlinear Hertz contact force model and transient sound radiation on the impact sound, including: Construct a Hertz-Kelvin nonlinear contact force model for local intrusion displacement; Based on the Rayleigh integral approximation, the characteristics of the impact sound pressure predicted by the neural network are required to be strictly positively correlated with the second derivative of the nonlinear contact force.

[0013] Furthermore, the material feature decoupled branch neural network based on fractional calculus introduces Caputo fractional calculus to characterize the complex viscoelastic acoustic attenuation memory effect of polymer materials.

[0014] Furthermore, the material feature decoupling branch neural network extracts the time-frequency domain sound pressure pole features and backfits the fractional damping order characterizing the material's constitutive properties. and relaxation time spectrum ; Including the spectral centroid The three-dimensional feature manifold space ( , , In this method, material combinations are defined using support vector boundaries to obtain material-level clustering 3D spatial scatter points.

[0015] Secondly, a component noise diagnosis system based on a nonlinear dynamic neural network is provided to implement the component noise diagnosis method based on a nonlinear dynamic neural network described in the first aspect, including: A multimodal data synchronous acquisition layer is used to acquire triaxial acceleration signals and microphone sound pressure signals; The spatiotemporal phase space reconstruction layer is used to construct multidimensional vibration-acoustic state feature vectors based on triaxial acceleration signals and microphone sound pressure signals. It is also used to obtain physical state and acoustic feature map matrices based on the multidimensional vibration-acoustic state feature vectors through mapping. The nonlinear dynamic physical information dual-branch neural network includes a branch neural network for classifying abnormal noise mechanisms based on a deep nonlinear physical constraint loss function, used to analyze the types of abnormal noises; and a branch neural network for decoupling material features based on fractional calculus, used to analyze the materials that produce abnormal noises.

[0016] As can be seen from the above technical solution, the beneficial technical effects of the present invention are as follows: 1. An impenetrable "white-box" defense barrier has been built: LuGre dynamic friction model, nonlinear Hertz contact and Caputo fractional calculus are introduced into acoustic AI diagnosis, which completely eliminates the "black box" uninterpretability of AI and eliminates false alarms that violate classical mechanics.

[0017] 2. The overall technical solution has extreme noise resistance: Through vibration-sound spatiotemporal phase space mapping, the system no longer listens to sounds blindly, but actively searches for legitimate sound sources based on the motion state. Even in harsh environments where the fluid noise or mechanical background noise of the excitation table is as high as 70dB, it can still accurately extract weak abnormal sound pulses.

[0018] 3. The technical solution also provides penetrating engineering guidance value: by using the viscoelastic acoustic memory effect to infer the constitutive parameters of the material, it achieves the ultimate decoupling of "what specific material is colliding / frictioning at the abnormal noise location"; it directly outputs rectification reports at the level of "friction between ABS plastic and spring steel sheet" for abnormal noise engineers, which greatly shortens the problem investigation cycle that traditionally relies mainly on human experience to identify problems by listening. Attached Figure Description

[0019] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0020] Figure 1 This is a structural diagram of the component noise diagnosis system based on a nonlinear dynamic neural network in an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the working principle of the abnormal noise mechanism classification branch neural network in an embodiment of the present invention; Figure 3 This is a clustered three-dimensional spatial scatter plot of materials that generate abnormal noises in an embodiment of the present invention. Detailed Implementation

[0021] The embodiments of the technical solution of the present invention will now be described in detail with reference to the accompanying drawings. These embodiments are merely illustrative of the technical solution of the present invention and are therefore intended to limit the scope of protection of the present invention.

[0022] It should be noted that, unless otherwise stated, the technical or scientific terms used in this application should have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0023] Example This application provides a component noise diagnosis method based on a nonlinear dynamic neural network, and also provides a BSR intelligent diagnosis system that integrates high-frequency physical quantities of the excitation table with multi-dimensional acoustic space and uses advanced nonlinear dynamic differential equations as the neural network loss function constraint to implement the method, thereby achieving high-precision physical mechanism classification and in-depth sound-generating material combination determination under strong interference. The diagnosis method includes the following steps: Step S1: Construct a multidimensional vibration-acoustic state feature vector based on the collected triaxial acceleration signal of the excitation source and the microphone sound pressure signal of the response source. Obtain the physical state and acoustic feature map matrix based on the multidimensional vibration-acoustic state feature vector through mapping. In this step, the triaxial acceleration signal of the test bench excitation source is acquired through high-frequency (e.g., to accurately capture the nonlinear hard collision characteristics of microsecond-level transient impacts and reduce time-domain integration loss error, the sampling frequency is preferably set to 20kHz~51.2kHz) synchronization. The sound pressure signal from the microphone array of the response source (such as the tested component that is prone to abnormal noise, such as the car door or dashboard) The method of data acquisition is not limited; it can be implemented in any feasible way using existing technology, such as acquiring triaxial acceleration signals through an accelerometer or acquiring sound pressure signals from a microphone array through an audio analyzer.

[0024] Based on triaxial acceleration signals The relative velocity between the two components of the potential abnormal noise emission interface is obtained by time-domain integration. and relative displacement : In the formula, At the current sampling time, For integration variables; The initial relative velocity constant of the system; Let be the initial relative displacement constant of the system.

[0025] Based on microphone array sound pressure signal Extract the high-frequency envelope of the sound pressure signal, and then combine the high-frequency envelope of the sound pressure signal with the physical motion quantity (i.e., relative velocity) excited in the same location. and relative displacement Temporal hard alignment is performed to construct a multidimensional vibration-acoustic state feature vector. : The evolution trajectory of the multidimensional vibration-acoustic state eigenvectors in phase space is mapped to a matrix of physical states and acoustic feature maps. Specifically, the feature map matrix It is composed of single-time multidimensional vibration-acoustic state feature vectors The two-dimensional feature matrix, constructed by splicing together sequences that evolve over time within an analysis window, is represented as: In the formula, This represents the total number of sampling points within a single analysis time window. These are continuous sampling times. This feature map matrix... As the input layer of the neural network, it cuts off the probability of misjudgment from the physical source of background noise, thus solving the technical problem of poor anti-background noise interference capability of AI automatic diagnosis system in traditional solutions.

[0026] Step S2: Input the feature map matrix obtained after step S1 into a nonlinear dynamic physical information dual-branch neural network to obtain the types of abnormal noises and the materials that produce them. The nonlinear dynamics neural network constructed in this application is a two-branch neural network, including a branch neural network for classifying abnormal noise mechanisms based on a deep nonlinear physical constraint loss function, used to analyze the types of abnormal noises; and a branch neural network for decoupling material features based on fractional calculus, used to analyze the materials that produce abnormal noises. Details are as follows: The inventors of this application discovered through research that Rattle knocking sounds invariably occur around extreme points or zero-crossing points of local acceleration, while Squeak friction sounds invariably occur during the stick-slip phase after the relative velocity exceeds the static friction threshold. Based on these findings, in this application, for the neural network classifying abnormal noise mechanisms, the data flow and computational topology of its constraint modules are as follows: Figure 2 As shown, its total loss function is defined as: in For classification cross-entropy, For adaptive weights; physical constraint loss function The forced abnormal noise mechanism classification neural network follows the following two major dynamic models: (1) Squeak (frictional sound) depth constraint based on LuGre dynamic friction model: The stick-slip effect of the micro-friction surface is abstracted as the introduction of elastic bristle deformation. Differential equation: In the formula, For relative velocity, The coefficient of stiffness of the micro-elastic bristles at the contact surface. Let be a nonlinear function of frictional velocity that incorporates the Stribeck effect.

[0027] Calculate the interfacial dynamic friction force of micro-friction surfaces : In the formula, The coefficient of stiffness of the micro-elastic bristles at the contact surface. For elastic bristle deformation, The microscopic brush bristle damping coefficient, The macroscopic coefficient of viscous friction. This refers to relative velocity.

[0028] Friction noise constraint term, requiring the constraint loss function of the neural network to predict friction noise. envelope Constrained by the gradient of the frictional dissipation power density function: In the formula, This represents the total number of time series sampling points. This is the current sampling point sequence number. For the gradient operator in the time dimension, For the predicted friction noise envelope, The acoustic-mechanical energy radiation conversion constant specific to the system, typically taking values ​​within a certain range. , This refers to the interfacial dynamic friction force of the microscopic friction surface. Denotes the Frobenius norm in the time-frequency manifold space; For the first Each sampling time.

[0029] (2) Depth constraint of Rattle sound based on nonlinear Hertz contact force model and transient acoustic radiation: Constructing local intrusion displacement Hertz-Kelvin nonlinear contact force model: In the formula, This is a local intrusive displacement. For the speed of intrusion, The nonlinear contact stiffness coefficient, The material contact damping coefficient, This is the nonlinearity index of the contact material (usually taken as 1 / 4 or 1 / 2). It is a nonlinear contact force.

[0030] Based on the Rayleigh integral approximation, the characteristics of the impact sound pressure predicted by the neural network are required to be strictly positively correlated with the second derivative of the nonlinear contact force: In the formula, The sound pressure level feature value predicted by the neural network; The acoustic radiation hysteresis coefficient; Represents the convolution operator; To test the structural acoustic transfer function of the component itself; The constraint loss function for the knocking noise. This represents the total number of time series sampling points. This is the current sampling point sequence number. For the gradient operator in the time dimension, For the first Each sampling time.

[0031] The material feature decoupling branch neural network is used to determine the "combination of sound-producing materials" and distinguish whether the current abnormal noise comes from "plastic-plastic" or "metal-rubber".

[0032] The material determination method in this application introduces Caputo fractional calculus to characterize the complex viscoelastic acoustic attenuation memory effect of polymer materials: In the formula, For generalized internal stress of materials, The elastic modulus of the material. For material deformation and strain, It is the viscoelastic coefficient; It is a fractional derivative calculus operator based on Caputo's definition; The fractional order constants that reflect the micro-damping characteristics of different materials ( ).

[0033] Material feature decoupling branch neural network extracts time-frequency domain sound pressure pole features and backfits the fractional damping order characterizing the constitutive properties of the material. and relaxation time spectrum ; Including the spectral centroid The three-dimensional feature manifold space ( , , In this study, material combinations are defined using support vector boundaries, and the material-level clustering 3D spatial scatter distribution is as follows: Figure 3 As shown in the figure, the boundary distinctions between different materials such as metal / metal and PP / rubber are clearly displayed, enabling root cause-level material diagnosis.

[0034] The feature map matrix obtained in step S1 is input into the dual-branch neural network constructed in this step for further processing, enabling intelligent diagnostic output. The output results include: Types of abnormal noises: friction / tapping / resonance; Materials that produce abnormal noise: plastic-plastic / metal-rubber.

[0035] In some embodiments, the technical solution of this application may further prioritize generating a 3D model based on the above output results to locate the abnormal noise with high brightness, and at the same time provide rectification suggestions.

[0036] The following example, using a 6-DOF (DoF) noise test on the instrument panel (IP) bench of a certain brand of vehicle, illustrates the above technical solution: 1. Fix the instrument panel to the vibration platform and arrange a three-component accelerometer and microphone array. Input a random road spectrum to drive the test bench operation, through... Figure 1 The multimodal data synchronous acquisition layer shown in the figure enables the system to synchronously acquire local acceleration at the microsecond level. Harmony and sound pressure .

[0037] 2. System integral generation relative velocity Extracting the time window of high-frequency sound pressure change, constructing a multi-dimensional state vector and mapping it to a feature map matrix. .

[0038] 3. The feature map matrix Input a pre-trained, nonlinear dynamic physical information bibranch neural network. During backpropagation, the network must satisfy the constraint loss function. minimize.

[0039] 4. During the calculation process, if Figure 2 The deep nonlinear physical constraint loss function calculation module shown in the network discovers the power dissipation extremum point of a certain continuous high-frequency sound pressure pulse that satisfies the dynamic friction surface stick-slip equation. The penalty value approaches zero (i.e., it fully conforms to the physical laws of LuGre dynamic friction), and the mechanism branch is confirmed as "Squeak friction sound", eliminating the interference of random high-frequency background noise.

[0040] 5. Simultaneously, the material decoupling branch extracts the attenuation envelope of the sound pressure signal segment and fits its Caputo fractional damping order. Combined with the corresponding relaxation time spectrum Locate the coordinate point in the feature space.

[0041] 6. For example Figure 3 The 3D spatial scatter plot mapping system compares the data with a pre-set material acoustic impedance constitutive database and outputs a high-precision diagnostic conclusion: "The current friction noise originates from the PP plastic and rubber sealing strip, with a confidence level of 94%." The system automatically highlights the corresponding material matching area in the 3D digital model on the host computer's UI interface, allowing engineers to directly apply grease or add a buffer pad to the affected area to resolve the problem.

[0042] This application also provides a component noise diagnosis system based on a nonlinear dynamic neural network, used to implement the component noise diagnosis method based on a nonlinear dynamic neural network described above, including: A multimodal data synchronous acquisition layer is used to acquire triaxial acceleration signals and microphone sound pressure signals; The spatiotemporal phase space reconstruction layer is used to construct multidimensional vibration-acoustic state feature vectors based on triaxial acceleration signals and microphone sound pressure signals. It is also used to obtain physical state and acoustic feature map matrices based on the multidimensional vibration-acoustic state feature vectors through mapping. The nonlinear dynamic physical information dual-branch neural network includes a branch neural network for classifying abnormal noise mechanisms based on a deep nonlinear physical constraint loss function, used to analyze the types of abnormal noises; and a branch neural network for decoupling material features based on fractional calculus, used to analyze the materials that produce abnormal noises.

[0043] By adopting the technical solution of this application, an impenetrable "white-box" defense barrier is constructed: for the first time, LuGre dynamic friction model, nonlinear Hertz contact, and Caputo fractional calculus are introduced into acoustic AI diagnosis, completely eliminating the "black box" uninterpretability of AI and preventing false alarms that violate classical mechanics. The overall technical solution has extreme noise resistance: through vibration-sound spatiotemporal phase space mapping, the system no longer blindly listens to sound, but actively searches for legitimate sound sources based on motion state. Even in harsh environments with fluid noise or mechanical background noise as high as 70dB on the vibration table, it can still accurately extract weak abnormal noise pulses. The technical solution also provides penetrating engineering guidance value: by using the viscoelastic acoustic memory effect to back-deduce the constitutive parameters of materials, it achieves extreme decoupling of "what specific material is colliding / frictioning at the abnormal noise location"; it directly outputs rectification reports at the level of "ABS plastic rubbing against spring steel sheet" for abnormal noise engineers, significantly shortening the problem-solving cycle that traditionally relies mainly on human experience for sound identification.

[0044] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.

Claims

1. A method for diagnosing abnormal noises in components based on a nonlinear dynamic neural network, characterized in that, Includes the following steps: A multidimensional vibration-acoustic state feature vector is constructed based on the triaxial acceleration signal of the excitation source and the microphone sound pressure signal of the response source. The physical state and acoustic feature map matrix are obtained by mapping based on the multidimensional vibration-acoustic state feature vector. The physical state and acoustic feature map matrix are input into a nonlinear dynamic physical information dual-branch neural network to obtain the types of abnormal noises and the materials that produce them.

2. The component noise diagnosis method based on nonlinear dynamic neural network according to claim 1, characterized in that, Constructing a multidimensional vibration-acoustic state feature vector includes: Based on triaxial acceleration signals, the relative velocity and relative displacement between the two components of the potential abnormal noise-generating interface are obtained by time-domain integration. Based on the sound pressure signal from the microphone array, the high-frequency envelope of the sound pressure signal is extracted. The high-frequency envelope of the sound pressure signal is then hard-aligned in the time domain with the physical motion quantity excited in the same location to obtain a multidimensional vibration-acoustic state feature vector.

3. The component noise diagnosis method based on nonlinear dynamic neural network according to claim 2, characterized in that, The physical state and acoustic feature map matrix is ​​obtained by mapping based on multidimensional vibration-acoustic state feature vectors, including mapping the evolution trajectory of multidimensional vibration-acoustic state feature vectors in phase space into physical state and acoustic feature map matrix.

4. The component noise diagnosis method based on nonlinear dynamic neural network according to claim 1, characterized in that, The nonlinear dynamic physical information dual-branch neural network includes an abnormal noise mechanism classification branch neural network based on a deep nonlinear physical constraint loss function, used to analyze the types of abnormal noise; and a material feature decoupling branch neural network based on fractional calculus, used to analyze the materials that produce abnormal noise.

5. The component noise diagnosis method based on nonlinear dynamic neural network according to claim 4, characterized in that, The total loss function of the abnormal noise mechanism classification branch neural network is: in For classification cross-entropy, For adaptive weights; physical constraint loss function The forced abnormal noise mechanism classification branch neural network follows the constraints on frictional sound based on the LuGre dynamic friction model, and the constraints on percussion sound based on the nonlinear Hertz contact force model and transient sound radiation.

6. The component noise diagnosis method based on nonlinear dynamic neural network according to claim 5, characterized in that, Constraints on frictional sound based on the LuGre dynamic friction model include: The stick-slip effect of the micro-friction surface is abstracted into a differential equation that incorporates the deformation of elastic bristles; Calculate the dynamic interfacial friction force of micro-friction surfaces; The frictional noise constraint requires that the envelope of the frictional noise predicted by the neural network be constrained by the gradient of the frictional dissipation power density function.

7. The component noise diagnosis method based on nonlinear dynamic neural network according to claim 5, characterized in that, Based on the constraints of the nonlinear Hertz contact force model and transient sound radiation on the impact sound, including: Construct a Hertz-Kelvin nonlinear contact force model for local intrusion displacement; Based on the Rayleigh integral approximation, the characteristics of the impact sound pressure predicted by the neural network are required to be strictly positively correlated with the second derivative of the nonlinear contact force.

8. The component noise diagnosis method based on nonlinear dynamic neural network according to claim 4, characterized in that, The material feature decoupling branch neural network based on fractional calculus introduces Caputo fractional calculus to characterize the complex viscoelastic acoustic attenuation memory effect of polymer materials.

9. The component noise diagnosis method based on nonlinear dynamic neural network according to claim 8, characterized in that, Material feature decoupling branch neural network extracts time-frequency domain sound pressure pole features and backfits the fractional damping order characterizing the constitutive properties of the material. and relaxation time spectrum ; Including the spectral centroid The three-dimensional feature manifold space ( , , In this method, material combinations are defined using support vector boundaries to obtain material-level clustering 3D spatial scatter points.

10. A component noise diagnosis system based on a nonlinear dynamic neural network, characterized in that, The component noise diagnosis method based on nonlinear dynamic neural network as described in claims 1-9 includes: A multimodal data synchronous acquisition layer is used to acquire triaxial acceleration signals and microphone sound pressure signals; The spatiotemporal phase space reconstruction layer is used to construct multidimensional vibration-acoustic state feature vectors based on triaxial acceleration signals and microphone sound pressure signals. It is also used to obtain physical state and acoustic feature map matrices based on the multidimensional vibration-acoustic state feature vectors through mapping. The nonlinear dynamic physical information dual-branch neural network includes a branch neural network for classifying abnormal noise mechanisms based on a deep nonlinear physical constraint loss function, used to analyze the types of abnormal noises; and a branch neural network for decoupling material features based on fractional calculus, used to analyze the materials that produce abnormal noises.

Images