Weak fault symptom enhancement method based on sound source induced excitation resonance
Through the sound source induced excitation resonance method, a multimodal signal acquisition model is constructed and a deep residual network is introduced for feature fusion. Combined with the genetic operator group optimization algorithm and the LSTM network, the problem of incomplete information acquisition in the detection of weak faults in mechanical equipment is solved, and the accurate description of complex systems and the effective capture of early fault signs are achieved.
Patent Information
- Application Number
- CN202511007707.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-22
- Publication Date
- 2025-10-24
AI Technical Summary
The existing technology for detecting weak faults in mechanical equipment has problems such as a single signal acquisition mode, limitations of multimodal feature fusion methods, inaccurate characterization of nonlinear coupling mechanisms, lack of effective evaluation criteria for parameter optimization, and weak dynamic control capabilities. These problems lead to incomplete acquisition of operating status information, poor feature fusion quality, and difficulty in capturing early signs of faults.
A method based on sound source induced excitation resonance is adopted. By collecting multimodal signals and constructing a nonlinear resonance response model, a deep residual network is introduced for feature fusion and quantification of nonlinear interactions. The genetic algorithm and particle swarm optimization algorithm are combined for parameter optimization. The LSTM network is used to adjust the delay time parameters in real time to achieve dynamic synchronization between the system resonance frequency and the fault characteristic frequency.
It achieves comprehensive acquisition of mechanical equipment operation information and high-quality fusion of multi-source features, accurately characterizes the nonlinear characteristics of complex systems, and improves the ability to capture weak fault signs and detection sensitivity.
Smart Images
Figure CN120831228A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of signal detection and signal processing, in particular to a weak fault symptom enhancement method based on sound source induced vibration resonance. BACKGROUND
[0002] The weak fault symptom enhancement technology actively applies a specific frequency of acoustic excitation to make the mechanical structure produce controlled resonance in the fault characteristic frequency band, thereby amplifying the weak fault signal that is originally covered by noise. This technology utilizes the acoustic-force coupling effect to convert high-frequency acoustic energy into structural vibration modal energy, significantly improves the signal-to-noise ratio by matching the resonance conditions of the fault sensitive frequency band, and is especially suitable for the detection of early bearing spalling, gear micro-pitting and other weak faults.
[0003] In the field of weak fault detection of mechanical equipment, the existing technology generally has the following key problems: first, the signal acquisition mode is single, relying on traditional vibration signals and ignoring the coordinated acquisition of multi-source physical field information such as sound field and temperature field, resulting in incomplete acquisition of running state information; second, the multi-modal feature fusion method has limitations, the traditional deep learning model is prone to gradient vanishing problem when processing multi-source heterogeneous data due to deepening of network layers, affecting the quality of feature fusion; third, the nonlinear coupling mechanism is not accurate, the existing method lacks quantitative analysis of the nonlinear interaction between sound field and vibration signal, and ignores the nonlinear modulation effect of temperature field on sound field-vibration coupling through thermoacoustic coupling, making it difficult to accurately describe the dynamic characteristics of complex systems; fourth, the parameter optimization lacks effective evaluation criteria and efficient algorithms, traditional optimization strategies either rely on fixed parameter initialization leading to blind optimization direction, or use single optimization algorithm leading to imbalance between global search ability and local convergence speed; fifth, the dynamic regulation and control ability is weak, the delay time parameters are mostly fixed settings, which cannot synchronize the system resonance frequency and fault characteristic frequency in real time, resulting in poor weak fault feature extraction effect and difficulty in effectively capturing early fault symptoms. These problems jointly restrict the sensitivity and accuracy of weak fault detection of mechanical equipment. SUMMARY
[0004] In view of the deficiencies of the prior art, the present application provides a weak fault symptom enhancement method based on sound source induced vibration resonance to solve the problems of incomplete acquisition of running state information, affecting the quality of feature fusion, difficulty in accurately describing the dynamic characteristics of complex systems, imbalance between global search ability and local convergence speed, and difficulty in effectively capturing early fault symptoms.
[0005] To achieve the above purpose, the present application provides the following technical scheme: a weak fault symptom enhancement method based on sound source induced vibration resonance, comprising the following steps:
[0006] S1 Multi-modal signal acquisition and model construction:
[0007] The original vibration signal, synchronous sound field signal and multi-source physical field signal of the measured mechanical equipment in the running state are collected, and a nonlinear resonance response model with multi-modal coupling excitation and dynamic delay resonance regulation characteristics is constructed. In the multi-modal coupling excitation mechanism, a deep residual network is introduced to fuse the features of the multi-source physical field signal, and the deep residual network relieves the gradient disappearance problem through a jump connection structure.
[0008] S2 nonlinear coupling mechanism design:
[0009] A sound field-vibration coupling term containing a quadratic term and a cross-modal coupling coefficient is introduced to quantify the strength of nonlinear interaction. A thermoacoustic coupling term is embedded in the potential function, and the nonlinear modulation effect of the temperature field on the sound field-vibration coupling effect is described by the Fourier heat conduction equation.
[0010] S3 parameter initialization and objective function construction:
[0011] The model parameters and variable scale factors are initialized, the adaptive step control algorithm is used to solve the system response, and the objective function is constructed based on the resonance response amplitude amplification factor to provide an evaluation standard for parameter optimization.
[0012] S4 hybrid optimization parameter optimization:
[0013] A hybrid optimization strategy is used to adaptively optimize the model parameters. The hybrid optimization strategy combines genetic algorithm and particle swarm optimization algorithm, and takes the synergistic optimization of global search ability and local convergence speed as the goal. The genetic algorithm is used for global parameter space exploration, and the particle swarm optimization algorithm is used for local optimal solution refinement. The two achieve iteration information sharing through cross-validation mechanism.
[0014] S5 resonance response reconstruction and dynamic regulation:
[0015] Based on the optimized parameters, the resonance response is reconstructed, the enhanced weak fault features are extracted, and the LSTM network is used to adjust the delay time parameter in real time, so that the system resonance frequency and the fault feature frequency remain dynamically synchronized.
[0016] Preferably, in step S1, the original vibration signal and synchronous sound field signal collection equipment all adopt anti-interference packaging structure, and the spatial layout positions of the vibration sensor and the sound field sensor are optimized and configured according to the structural characteristics of the measured equipment.
[0017] The collection device of the original vibration signal and the synchronous sound field signal needs to adopt a multi-layer anti-interference packaging structure: a high-conductivity metal shell is used for the outer layer, and a vibration absorption material is filled in the inner layer to suppress mechanical vibration conduction; the signal transmission line adopts a twisted pair or a coaxial cable, and a differential amplification circuit is configured to eliminate common-mode noise. For the spatial arrangement of the sensor, first, based on the three-dimensional structure model of the measured device, the vibration modal and sound pressure level distribution of the key components are simulated through finite element analysis to determine the main vibration direction and the sound field energy concentration area; then, combined with the experimental test data, the vibration sensor is arranged at the parts with high stiffness and clear vibration transmission path such as the bearing seat or the gear box shell, and the sound field sensor is arranged at the position far away from the reflective surface and in the peak value area of the sound pressure level; finally, the time domain synchronization and the frequency domain coupling characteristics of the multi-sensor signals are verified through correlation analysis, and the arrangement scheme is iteratively adjusted until the optimal signal-to-noise ratio is achieved.
[0018] Preferably, in the step S1, the multi-modal coupling excitation mechanism is realized through a feature concatenation fusion manner, and the fused feature vector is processed by a deep residual network, and the residual unit of the network adopts a pre-activation structure and contains cross-layer connections;
[0019] The collected original vibration signal, synchronous sound field signal and multi-source physical field signal are preprocessed to extract initial feature vectors of each modal; then, through a feature concatenation fusion manner, the feature vectors of different modes are spliced in the channel dimension to form a fused multi-modal feature tensor; the feature tensor is input into a deep residual network, each residual unit of the network adopts a pre-activation structure, that is, batch normalization and ReLU activation function are performed first, and then convolution operation is performed to alleviate the gradient vanishing problem in deep network training; at the same time, each residual unit directly transmits the input features to the output end through cross-layer connections, and element-wise addition is performed with the features after convolution to form a residual learning mechanism; the cooperative design of such pre-activation structure and cross-layer connection not only enhances the non-linear expression ability of the features, but also preserves the key information of the original features through the jump connection, and finally outputs the fused multi-modal coupling feature vector, providing a basis for constructing a nonlinear resonance response model with dynamic delay resonance regulation characteristics.
[0020] Preferably, in the step S1, the multi-modal coupling excitation and the dynamic delay resonance regulation characteristics are realized by a sound source modulation function for dynamic loading, and the function automatically adjusts the excitation characteristics according to the change of the device load, and triggers the excitation mode switching when the load rate changes significantly;
[0021] The key working condition data such as load rate, torque and power are collected in real time by the equipment operation parameter monitoring system, and the triggering condition of the sound source modulation function is constructed based on the preset load change threshold; the sound source modulation function adopts a combination of a piecewise linear model and fuzzy logic control, when the load rate is in a stable interval, the function outputs the basic excitation parameters, and when the load rate changes significantly, the table lookup method is used to switch to the high load mode or the low load mode; a gradual transition mechanism is introduced in the excitation mode switching process, the excitation parameters are adjusted smoothly through the linear interpolation algorithm, and the system instability caused by sudden change is avoided; finally, the dynamically adjusted excitation parameters are injected into the multi-modal coupling excitation mechanism in real time, the input layer weight matrix of the deep residual network and the scaling factor of the cross-layer connection are modified, the dynamic coupling characteristic matching of the sound field-vibration-physical field signal is realized, and it is ensured that the nonlinear resonance response model can adapt to the load change of the equipment.
[0022] Preferably, in the step S2, the sound field-vibration coupling term is realized by introducing a dynamic weight coefficient, which is automatically adjusted according to the equipment running state, and when the sound field intensity is detected to change significantly, the weight coefficient is adjusted synchronously according to the preset rule;
[0023] The sound pressure signal of the equipment during operation is collected in real time by the sound field sensor, and the time domain amplitude mean value and the frequency energy spectrum density are calculated as the sound field intensity evaluation index; when the sound pressure level instantaneous value exceeds the preset threshold, the weight coefficient adjustment mechanism is triggered; the dynamic weight coefficient is designed by using a piecewise nonlinear function, when the sound field intensity is in a normal interval, the coefficient remains a basic value, when the sound field intensity is significantly enhanced, the coefficient increases exponentially to strengthen the excitation effect of the sound field on the vibration, and when the sound field intensity is weakened, the coefficient decreases according to the logarithmic function to suppress unnecessary coupling noise; meanwhile, temperature field feedback adjustment is introduced, the temperature gradient of the equipment surface is calculated in real time by the Fourier heat conduction equation, and if the temperature exceeds the safety threshold, the weight coefficient is modified twice to avoid thermal-acoustic coupling overload; finally, the dynamically adjusted weight coefficient is embedded in the quadratic term and the cross term of the sound field-vibration coupling term, the multiplicative factor before the coupling term is modified to adjust the nonlinear interaction strength in real time, and it is ensured that the nonlinear resonance response model can adapt to the changes of the sound field intensity and the temperature field.
[0024] Preferably, in the step S3, the initialization model adopts a hybrid initialization strategy, part of the parameters are set to an initial range based on the prior knowledge of the equipment, and the other part of the parameters are extracted as initial values from the historical operation data by unsupervised learning method;
[0025] Prior knowledge of key physical parameters is extracted from equipment design documents and operation manuals, such as determining the natural frequency range of mechanical systems through finite element analysis, and obtaining the initial interval of the thermal diffusion coefficient in the heat conduction equation from the material manual. These initial ranges of parameters are injected into the model parameter space in the form of uniform distribution or Gaussian distribution; at the same time, multi-modal signals under normal working conditions of the equipment are extracted from the historical operation database, and unsupervised feature extraction is performed on the data using a variational autoencoder. The high-dimensional signal is compressed into a low-dimensional latent feature vector through the encoder network, and this vector is used as the initial value of the weight matrix of the deep residual network and the cross-modal coupling coefficient; for dynamic parameters that cannot be directly determined by prior knowledge, two initialization methods are combined: first, set the basic range according to the type of equipment, and then cluster the resonance response patterns in the historical data through unsupervised learning for cluster analysis, and the parameter combination corresponding to the cluster center is used as the local initial value; finally, the prior knowledge initial range and the unsupervised learning features are fused through the parameter mapping function, such as weighted average for physical parameters and cascading splicing for neural network weights, to ensure that the initialized parameters meet the physical laws and have data adaptability.
[0026] Preferably, the synergistic mechanism of the genetic algorithm and the particle swarm algorithm in step S4 is realized by the following method: periodically exchange population information during optimization, combining the global search ability of the genetic algorithm with the local convergence characteristics of the particle swarm algorithm;
[0027] Initialize two independent populations: the genetic algorithm population contains randomly generated parameter combinations, and the particle swarm algorithm population is initialized based on the optimal frequency band of historical operation data; set the iteration period, and at each exchange, convert the individual with the highest fitness in the current population of the genetic algorithm into the particle position of the particle swarm algorithm through the mapping function, replacing the particle with the lowest fitness in the PSO population, realizing the injection of global optimal solution to local search; at the same time, the global optimal solution found by the particle swarm algorithm is added to the population of the genetic algorithm as an elite individual, replacing the individual with the lowest fitness, and using the local convergence result of PSO to guide the evolution direction of GA; cross-validation is realized through the shared fitness evaluation function, that is, the individual of the genetic algorithm and the particle of the particle swarm algorithm are evaluated based on the same resonance response amplitude factor and fault feature extraction effect index, ensuring the consistency of the optimization target; finally, the genetic algorithm maintains population diversity through crossover and mutation operations, and the particle swarm algorithm realizes local fine search through velocity-position update, and the two form a synergistic optimization cycle of "global exploration-local deepening-feedback guidance" through periodic information exchange, balancing the global search ability and local convergence speed of the algorithm.
[0028] Preferably, in the step S4, the adaptive optimization strategy combines the genetic algorithm and the particle swarm optimization algorithm: when the global exploration process of the genetic algorithm and the local convergence speed of the particle swarm optimization algorithm differ significantly, the step size parameter is automatically adjusted by a dynamic weight coefficient, and the adjustment rule is optimized in real time according to the change rate of the fitness function of the two algorithms, to ensure the balance between global search and local refinement.
[0029] The optimization progress of the genetic algorithm and the particle swarm optimization algorithm is tracked in real time by a double-channel monitoring system: for the GA population, the improvement amplitude of the optimal solution of each generation is recorded; for the PSO population, the improvement rate of the overall fitness of the particle swarm is calculated. When the optimization progress of the two algorithms differs by more than a preset threshold, the step size adjustment mechanism is started: a dynamic weight coefficient is calculated according to the current progress difference, which amplifies the step size parameter of the algorithm with slower progress and suppresses the step size parameter of the algorithm with faster progress, to ensure that the search speeds of the two algorithms tend to be balanced; the adjusted step size parameter is applied to the algorithm iteration process in real time: the GA enhances the global exploration ability by increasing the step size, and the PSO refines the local search accuracy by reducing the step size; if the progress difference after adjustment is still not alleviated, the system will further fine-tune the weight coefficient, forming a dynamic closed loop of “monitoring-judgment-adjustment-verification”, to finally realize the collaborative optimization of global search and local refinement.
[0030] Preferably, in the step S5, the LSTM network uses multi-resolution analysis technology to decompose and reconstruct the optimized signal, and selects the frequency band components containing the main fault feature information for feature enhancement.
[0031] The optimized resonance response signal is decomposed by wavelet packet, mapped to different frequency band subspaces by multi-layer wavelet transform, and a tree structure containing detail coefficients and approximation coefficients is generated. Then, based on the prior knowledge of fault features, the correlation index of each frequency band and the target feature is calculated, and the main frequency band components with a correlation exceeding a preset threshold are selected. The selected frequency band components are denoised by soft threshold method, the non-feature noise is suppressed by adjusting the threshold parameter, and the frequency band coefficients with obvious features are nonlinearly amplified to enhance the fault signs. Finally, the processed frequency band components are combined into an enhanced signal by wavelet packet reconstruction algorithm, which is directly input into the LSTM network to predict the drift trend of the fault feature frequency in real time, dynamically adjust the delay time parameter, and ensure that the system resonance frequency always synchronizes with the enhanced fault feature frequency, realizing the directional enhancement and stable extraction of weak fault signs.
[0032] Preferably, in the step S5, the delay time parameter is adjusted in real time by an adaptive controller, and the adjustment strategy is automatically switched according to the change of the device operating state, and the response mode is enabled when the speed fluctuation is detected.
[0033] The speed sensor is used to collect the speed signal of the equipment in real time, and the instantaneous fluctuation rate is calculated; when the speed fluctuation rate exceeds the preset threshold, the adaptive controller automatically switches to the speed response mode, the input features of the LSTM network are increased by the speed change rate and the phase shift amount, and the delay time adjustment amount is output by the trained prediction model; the adjustment strategy adopts segmented control: when the speed fluctuates periodically, the periodic correction mode is enabled, and when the speed suddenly changes, the emergency correction mode is enabled; at the same time, a feedback verification mechanism is introduced, the resonance response amplitude amplification factor and the fault feature signal-to-noise ratio improvement amount before and after adjustment are compared, if the improvement amount is lower than the threshold, the hidden layer weight of the LSTM is fine-tuned, and the closed-loop control process of "state monitoring-mode switching-parameter adjustment-effect verification" is formed, so that the delay time parameter is always dynamically adapted to the actual running state of the equipment.
[0034] Compared with the prior art, the present application provides a weak fault symptom enhancement method based on sound source induced excitation resonance, which has the following beneficial effects:
[0035] Collecting multi-source signals and constructing a nonlinear resonance response model with multi-modal coupling excitation and dynamic delay resonance regulation characteristics, and introducing a deep residual network for feature fusion in the multi-modal coupling excitation mechanism, which can alleviate the gradient vanishing problem through a skip connection structure, can fully obtain mechanical equipment operation information and effectively fuse multi-source features, providing rich and high-quality data basis for subsequent accurate analysis;
[0036] Introducing a sound field-vibration coupling term containing a quadratic term and a cross-modal coupling coefficient and quantifying the nonlinear interaction strength, embedding a thermoacoustic coupling term in the potential function and describing the nonlinear modulation effect of the temperature field on the sound field-vibration coupling effect through the Fourier heat conduction equation, which can accurately depict the nonlinear characteristics of the system and improve the ability to grasp complex nonlinear relationships;
[0037] Initializing model parameters and variable scale factors and using an adaptive step control algorithm to solve the system response, and constructing a target function based on the resonance response amplitude amplification factor, which can provide a clear evaluation standard for parameter optimization and ensure the directionality and effectiveness of parameter optimization;
[0038] A hybrid optimization strategy combining genetic algorithm and particle swarm optimization algorithm is used to adaptively optimize the model parameters, the genetic algorithm realizes global parameter space exploration, the particle swarm optimization algorithm realizes local optimal solution refinement, and the cross-validation mechanism realizes iteration information sharing, which can fully utilize the advantages of the two algorithms and improve the efficiency and accuracy of parameter optimization;
[0039] The resonance response is reconstructed based on the optimized parameters, and the enhanced weak fault features are extracted, the LSTM network is used to adjust the delay time parameter in real time, so that the system resonance frequency and the fault feature frequency are dynamically synchronized, which can effectively strengthen the feature extraction effect and enhance the capture ability of weak fault signs. BRIEF DESCRIPTION OF DRAWINGS
[0040] Figure 1 The method flowchart of the present application. DETAILED DESCRIPTION
[0041] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.
[0042] Please refer to Figure 1 The present application provides a technical solution, a weak fault sign enhancement method based on sound source induced excitation resonance, comprising the following steps:
[0043] S1 multi-modal signal acquisition and model construction:
[0044] Acquire the original vibration signal, synchronous sound field signal and multi-source physical field signal under the running state of the measured mechanical equipment, and construct a nonlinear resonance response model with multi-modal coupling excitation and dynamic delay resonance regulation characteristics; in the multi-modal coupling excitation mechanism, a deep residual network is introduced to fuse the multi-source physical field signal, and the deep residual network relieves the gradient disappearance problem through a jump connection structure;
[0045] In the multi-modal signal acquisition stage, a high-sensitivity sensor array is used to realize synchronous capture of vibration, sound field and physical field signals, and a multi-channel data alignment algorithm is used to ensure time domain synchronization accuracy. The nonlinear resonance response model constructed adopts a dynamic delay resonance regulation mechanism, and realizes dynamic matching of the system natural frequency and the external excitation frequency by introducing a variable time lag factor. In the feature fusion process, the deep residual network uses its unique identity mapping structure to gradually fuse low-level features and high-level semantic information, effectively improving the representation ability of multi-source heterogeneous data;
[0046] S2 nonlinear coupling mechanism design:
[0047] The sound field-vibration coupling term containing the quadratic term and the cross-modal coupling coefficient is introduced to quantify the nonlinear interaction strength, and the thermoacoustic coupling term is embedded in the potential function, and the Fourier heat conduction equation is used to describe the nonlinear modulation effect of the temperature field on the sound field-vibration coupling effect;
[0048] The sound field-vibration coupling term is constructed by introducing quadratic nonlinear terms and cross-modal coupling coefficients, and a complex nonlinear model containing stiffness nonlinearity, damping nonlinearity and inertia nonlinearity is constructed. The thermoacoustic coupling term embedded in the potential function is divided into active and passive thermoacoustic effects by using hierarchical expansion. The spatial derivative term is introduced into the Fourier heat conduction equation to realize the spatio-temporal coupling modeling of temperature field distribution and sound pressure oscillation, and to provide physical constraint conditions for nonlinear modulation effect;
[0049] S3 parameter initialization and objective function construction:
[0050] The model parameters and variable scale factors are initialized, the adaptive step size control algorithm is used to solve the system response, and the objective function is constructed based on the resonance response amplitude amplification factor to provide evaluation criteria for parameter optimization;
[0051] The variable scale factor initialization adopts a hierarchical strategy, and the initial range of linear and nonlinear parameters in the model is set respectively. The adaptive step size control algorithm combines the trust region method and line search strategy, and dynamically adjusts the search step according to the gradient change of the objective function. When constructing the objective function, in addition to the resonance response amplitude amplification factor, the feature fidelity term is introduced as a regularization constraint to ensure that the enhanced fault features maintain the original physical characteristics;
[0052] S4 hybrid optimization parameter optimization:
[0053] A hybrid optimization strategy is used to adaptively optimize the model parameters. The hybrid optimization strategy combines genetic algorithm and particle swarm optimization algorithm, aiming to optimize the global search ability and local convergence speed. The genetic algorithm is used for global parameter space exploration, and the particle swarm optimization algorithm is used for local optimal solution refinement. The two methods realize iteration information sharing through cross-validation mechanism;
[0054] In the hybrid optimization strategy, the genetic algorithm uses real number coding to realize global parameter space exploration through tournament selection, uniform crossover and non-uniform mutation operations. The particle swarm optimization algorithm introduces an inertia weight dynamic adjustment mechanism to realize local optimal solution refinement combined with individual extreme value and group extreme value. The cross-validation mechanism combines the population diversity advantage of genetic algorithm with the fast convergence characteristics of particle swarm algorithm by sharing the position information of elite individuals;
[0055] S5 resonance response reconstruction and dynamic regulation:
[0056] Based on the optimized parameters, the resonance response is reconstructed, the enhanced weak fault features are extracted, and the LSTM network is used to adjust the delay time parameter in real time, so that the system resonance frequency and the fault feature frequency remain dynamic synchronization;
[0057] The resonance response reconstruction adopts a time-frequency analysis method based on wavelet packet transform to realize high-resolution extraction of fault features by selecting appropriate wavelet basis functions. In the delay time parameter adjustment process, the LSTM network utilizes its gating mechanism to model long-term dependencies and dynamically corrects the delay parameter based on real-time vibration signal characteristics. The dynamic synchronization mechanism ensures that the system resonance frequency always tracks the time-varying characteristics of the fault feature frequency by designing a phase-locked loop.
[0058] In step S1, the anti-interference packaging structure is adopted for the collection devices of the original vibration signal and the synchronous sound field signal, and the spatial layout positions of the vibration sensor and the sound field sensor are optimized and configured according to the structural characteristics of the measured device;
[0059] The collection devices of the original vibration signal and the synchronous sound field signal need to adopt a multi-layer anti-interference packaging structure: a high-conductivity metal shell is used for the outer layer, and vibration-absorbing materials are filled in the inner layer to suppress mechanical vibration conduction; the signal transmission line adopts twisted pair or coaxial cable, and a differential amplification circuit is configured to eliminate common-mode noise. For the spatial layout of the sensor, first, based on the three-dimensional structural model of the measured device, the vibration modal and sound pressure level distribution of the key components are simulated through finite element analysis to determine the main vibration direction and the sound field energy concentration area; then, combined with experimental test data, the vibration sensor is laid out at positions with high stiffness and clear vibration transmission path such as bearing seat or gear box shell, and the sound field sensor is deployed at positions far away from the reflection surface and in the peak value area of the sound pressure level; finally, the time domain synchronization and frequency domain coupling characteristics of the multi-sensor signals are verified through correlation analysis, and the layout scheme is iteratively adjusted until the optimal signal-to-noise ratio is achieved.
[0060] In step S1, the multi-modal coupling excitation mechanism is realized through feature concatenation fusion, and the fused feature vector is processed by a deep residual network. The residual unit of the network adopts a pre-activation structure and contains cross-layer connections.
[0061] The collected original vibration signal, synchronous sound field signal, and multi-source physical field signal are preprocessed to extract the initial feature vectors of each modality. Then, through feature concatenation fusion, the feature vectors of different modalities are spliced in the channel dimension to form a fused multi-modal feature tensor. This feature tensor is input into a deep residual network. Each residual unit of the network adopts a pre-activation structure, i.e., first performs batch normalization and ReLU activation function, and then performs convolution operation to alleviate the gradient vanishing problem in deep network training. At the same time, each residual unit transmits the input features directly to the output end through cross-layer connections, and adds them element by element with the convolved features to form a residual learning mechanism. This collaborative design of pre-activation structure and cross-layer connections not only enhances the non-linear expression ability of features, but also preserves the key information of original features through jump connections, finally outputting the fused multi-modal coupling feature vector, providing a basis for constructing a nonlinear resonance response model with dynamic delay resonance regulation characteristics.
[0062] In step S1, the multi-modal coupling excitation and dynamic delay resonance regulation characteristics are dynamically loaded through a sound source modulation function, which automatically adjusts the excitation characteristics according to the device load changes, and triggers the excitation mode switching when the load rate changes significantly;
[0063] The load rate, torque, power and other key working condition data are collected in real time by the device running parameter monitoring system, and the trigger conditions of the sound source modulation function are constructed based on the preset load change threshold; the sound source modulation function adopts a segmented linear model combined with fuzzy logic control, when the load rate is in the stable interval, the function outputs the basic excitation parameters, and when the load rate changes significantly, it switches to the high load mode or low load mode through the look-up table method; a gradual transition mechanism is introduced in the excitation mode switching process, and the excitation parameters are adjusted smoothly through linear interpolation algorithm to avoid system instability caused by sudden changes; finally, the dynamically adjusted excitation parameters are injected into the multi-modal coupling excitation mechanism, the input layer weight matrix and the scaling factor of cross-layer connection of the deep residual network are modified to realize the dynamic coupling characteristic matching of the sound field-vibration-physical field signal, and to ensure that the nonlinear resonance response model can adapt to the device load changes.
[0064] In step S2, the sound field-vibration coupling term is realized by introducing a dynamic weight coefficient, which is automatically adjusted according to the device running state, and the weight coefficient is adjusted synchronously according to the preset rules when the sound field intensity changes significantly;
[0065] The sound pressure signal during device operation is collected in real time by the sound field sensor, and its time domain amplitude mean and frequency energy spectrum density are calculated as the sound field intensity evaluation index; when the instantaneous value of sound pressure level exceeds the preset threshold, the weight coefficient adjustment mechanism is triggered; the dynamic weight coefficient adopts a segmented nonlinear function design, when the sound field intensity is in the normal interval, the coefficient remains the basic value, when the sound field intensity increases significantly, the coefficient increases exponentially to strengthen the excitation effect of the sound field on the vibration, and when the sound field intensity decreases, the coefficient decreases logarithmically to suppress unnecessary coupling noise; at the same time, temperature field feedback adjustment is introduced, the device surface temperature gradient is calculated in real time by Fourier heat conduction equation, if the temperature exceeds the safety threshold, the weight coefficient is modified twice to avoid thermal-acoustic coupling overload; finally, the dynamically adjusted weight coefficient is embedded in the quadratic term and cross term of the sound field-vibration coupling term, the multiplicative factor before the coupling term is modified to adjust the nonlinear interaction strength in real time, and to ensure that the nonlinear resonance response model can adapt to the changes of sound field intensity and temperature field.
[0066] In step S3, the initialization model adopts a hybrid initialization strategy, part of the parameters are set to the initial range based on the prior knowledge of the device, and the other part of the parameters are extracted from the historical running data as the initial value through unsupervised learning method;
[0067] Based on the prior knowledge of the key physical parameters extracted from the equipment design documents and operation manuals, such as determining the natural frequency range of the mechanical system through finite element analysis, obtaining the initial interval of the thermal diffusion coefficient in the heat conduction equation through the material manual, and injecting the initial range of these parameters into the model parameter space in the form of uniform distribution or Gaussian distribution; At the same time, the multi-modal signals of the equipment under normal working conditions are extracted from the historical operation database, and the variational autoencoder is used for unsupervised feature extraction of the data. The high-dimensional signal is compressed into a low-dimensional latent feature vector through the encoder network, and the vector is used as the initial value of the weight matrix of the deep residual network and the cross-modal coupling coefficient; For dynamic parameters that cannot be directly determined by prior knowledge, two initialization methods are combined: first, set the basic range according to the type of the equipment, and then cluster the resonance response mode in the historical data through unsupervised learning analysis, and the parameter combination corresponding to the cluster center is used as the local initial value; Finally, the prior knowledge initial range and the unsupervised learning features are fused through the parameter mapping function, such as weighted average for physical parameters and cascading splicing for neural network weights, to ensure that the initialized parameters meet the physical laws and have data adaptability.
[0068] The synergistic mechanism of genetic algorithm and particle swarm algorithm in step S4 is realized by the following methods: exchanging population information regularly during optimization, combining the global search ability of genetic algorithm with the local convergence characteristics of particle swarm algorithm;
[0069] Initialize two independent populations: the genetic algorithm population contains randomly generated parameter combinations, and the particle swarm algorithm population is initialized based on the optimal frequency band of historical operation data; Set the iteration period, and in each exchange, the individual with the highest fitness in the current population of the genetic algorithm is converted into the particle position of the particle swarm algorithm through the mapping function, replacing the particle with the lowest fitness in the PSO population, realizing the injection of global optimal solution to local search; At the same time, the global optimal solution found by the particle swarm algorithm is added to the population of the genetic algorithm as an elite individual, replacing the individual with the lowest fitness, and using the local convergence result of PSO to guide the evolution direction of GA; Cross-validation is realized through the shared fitness evaluation function, that is, the individual of the genetic algorithm and the particle of the particle swarm algorithm are evaluated based on the same resonance response amplitude factor and fault feature extraction effect index, ensuring the consistency of the optimization target; Finally, the genetic algorithm maintains population diversity through crossover and mutation operations, and the particle swarm algorithm realizes local fine search through velocity-position update, and the two form a synergistic optimization cycle of "global exploration-local deepening-feedback guidance" through periodic information exchange, balancing the global search ability and local convergence speed of the algorithm.
[0070] In step S4, the adaptive optimization strategy combines genetic algorithm and particle swarm optimization algorithm: when the global exploration process of genetic algorithm and the local convergence speed of particle swarm optimization algorithm differ significantly, the step size parameter is automatically adjusted by a dynamic weight coefficient, and the adjustment rule is optimized in real time according to the change rate of the fitness function of the two algorithms, to ensure the balance between global search and local refinement;
[0071] The optimization progress of genetic algorithm and particle swarm optimization algorithm is tracked in real time by a dual-channel monitoring system: for the GA population, the improvement amplitude of the optimal solution of each generation is recorded; for the PSO population, the improvement rate of the overall fitness of the particle swarm is calculated. When the optimization progress of the two algorithms differs by more than a preset threshold, the step size adjustment mechanism is started—according to the current progress difference, a dynamic weight coefficient is calculated, which amplifies the step size parameter of the algorithm with slower progress and suppresses the step size parameter of the algorithm with faster progress, to ensure that the search speeds of the two algorithms tend to be balanced; the adjusted step size parameter is applied to the algorithm iteration process in real time: GA enhances global exploration ability by increasing step size, and PSO refines local search accuracy by reducing step size; if the progress difference after adjustment is still not alleviated, the system will further fine-tune the weight coefficient, forming a dynamic closed loop of "monitoring-judgment-adjustment-verification", and finally realizing the collaborative optimization of global search and local refinement.
[0072] In step S5, the LSTM network uses multi-resolution analysis technology to decompose and reconstruct the optimized signal, and selects the frequency band components containing main fault feature information for feature enhancement;
[0073] The optimized resonance response signal is decomposed by wavelet packet, the signal is mapped to different frequency band subspaces by multi-layer wavelet transform, and a tree structure containing detail coefficients and approximation coefficients is generated; then, based on prior knowledge of fault features, the correlation index of each frequency band and the target feature is calculated, and the main frequency band components with correlation exceeding a preset threshold are selected; the selected frequency band components are denoised by soft threshold method, non-feature noise is suppressed by adjusting the threshold parameter, and the frequency band coefficients with obvious features are nonlinearly amplified to enhance the fault signs; finally, the processed frequency band components are combined into an enhanced signal by wavelet packet reconstruction algorithm, which is directly input into the LSTM network to predict the drift trend of the fault feature frequency in real time, dynamically adjust the delay time parameter, and ensure that the system resonance frequency always synchronizes with the enhanced fault feature frequency, realizing the directional enhancement and stable extraction of weak fault signs.
[0074] In step S5, the delay time parameter is adjusted in real time by an adaptive controller, and the adjustment strategy is automatically switched according to the change of the device operating state, and the response mode is enabled when the speed fluctuation is detected;
[0075] The instantaneous fluctuation rate of the rotating speed signal is calculated in real time by a rotating speed sensor when the equipment is running. When the rotating speed fluctuation rate exceeds a preset threshold, the adaptive controller automatically switches to a rotating speed response mode. In this mode, the input features of the LSTM network are increased by the rotating speed change rate and the phase shift amount, and the delay time adjustment amount is output by the trained prediction model. The adjustment strategy adopts segmented control: when the rotating speed fluctuates periodically, the periodic correction mode is enabled; when the rotating speed suddenly changes, the emergency correction mode is enabled. At the same time, a feedback verification mechanism is introduced. By comparing the resonance response amplitude amplification factor and the fault feature signal-to-noise ratio improvement before and after adjustment, if the improvement is less than the threshold, the hidden layer weight of the LSTM is fine-tuned. A closed-loop control process of "state monitoring-mode switching-parameter adjustment-effect verification" is formed to ensure that the delay time parameter is always dynamically adapted to the actual running state of the equipment.
[0076] In S1, multi-source signals are collected and a nonlinear resonance response model with multi-modal coupling excitation and dynamic delay resonance regulation characteristics is constructed. A deep residual network is introduced in the multi-modal coupling excitation mechanism for feature fusion, which can alleviate the gradient vanishing problem through a skip connection structure, fully acquire mechanical equipment operation information, and effectively fuse multi-source features to provide rich and high-quality data basis for subsequent accurate analysis.
[0077] In S2, a sound field-vibration coupling term containing a quadratic term and a cross-modal coupling coefficient is introduced and the nonlinear interaction strength is quantified. A thermoacoustic coupling term is embedded in the potential function, and the Fourier heat conduction equation is used to describe the nonlinear modulation effect of the temperature field on the sound field-vibration coupling effect, which can accurately depict the nonlinear characteristics of the system and improve the ability to grasp complex nonlinear relationships.
[0078] In S3, the model parameters and variable scale factors are initialized, and an adaptive step control algorithm is used to solve the system response. A target function is constructed based on the resonance response amplitude amplification factor, which can provide a clear evaluation standard for parameter optimization and ensure the directionality and effectiveness of parameter optimization.
[0079] In S4, a hybrid optimization strategy combining genetic algorithm and particle swarm optimization algorithm is used for adaptive optimization of model parameters. Genetic algorithm is used for global parameter space exploration, and particle swarm optimization algorithm is used for local optimal solution refinement. Through the cross-validation mechanism, the iteration information is shared, which can fully utilize the advantages of the two algorithms and improve the efficiency and accuracy of parameter optimization.
[0080] In S5, the resonance response is reconstructed based on the optimized parameters, and the enhanced weak fault features are extracted. The LSTM network is used to adjust the delay time parameter in real time to keep the system resonance frequency and the fault feature frequency dynamically synchronized, which can effectively enhance the feature extraction effect and enhance the ability to capture weak fault signs.
[0081] It is to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting; it is not intended to exclude myriad other embodiments of the present application that other inventors can develop based on the same general inventive concepts embodied by the described embodiments. That is, although the present application is described in terms of particular embodiments and implementations, it is to be understood that the terminology used is for the purpose of descriptive clarity and that it should be taken in its broadest possible sense. For example, the terms "a", "an", and "the" include both singular and plural referents unless the context clearly dictates otherwise. The terms "comprises", "comprising", "includes", "including" and the like can be used in conjunction with the term "consisting of to include the elements or steps listed after such conjunctive language, but not to the exclusion of other elements or steps. The singular forms "a", "an" and "the" include plural referents unless the context clearly dictates otherwise.
[0082] While the embodiments of the application have been shown and described herein, it is to be understood that the application is not limited to these embodiments. Rather, many modifications, changes and substitutions are intended to fall within the scope of the present application, which is limited only by the scope of the claims hereinafter appended.
Claims
1. A method for enhancing weak fault signatures based on sound source induced excitation resonance, characterized by, The method comprises the following steps: S1 multi-modal signal acquisition and model construction: Collect original vibration signals, synchronous sound field signals and multi-source physical field signals under the running state of the measured mechanical equipment, and construct a nonlinear resonance response model with multi-modal coupling excitation and dynamic delay resonance regulation characteristics; in the multi-modal coupling excitation mechanism, a deep residual network is introduced to fuse the features of the multi-source physical field signals, and the deep residual network alleviates the gradient disappearance problem through a jump connection structure; S2 non-linear coupling mechanism design: Introducing a sound field-vibration coupling term containing a quadratic term and a cross-modal coupling coefficient to quantify the strength of non-linear interaction, embedding a thermoacoustic coupling term in the potential function, and describing the non-linear modulation effect of the temperature field on the sound field-vibration coupling effect through the Fourier heat conduction equation; S3 parameter initialization and target function construction: Initialize the model parameters and the variable scale factor, solve the system response by using an adaptive step control algorithm, and construct a target function based on the resonance response amplitude amplification factor to provide an evaluation standard for parameter optimization; S4 hybrid optimization parameter optimization: Adopt a hybrid optimization strategy to adaptively optimize the model parameters, and the hybrid optimization strategy combines a genetic algorithm and a particle swarm optimization algorithm to achieve the cooperative optimization of global search capability and local convergence speed as the goal; wherein the genetic algorithm is used for global parameter space exploration, and the particle swarm optimization algorithm is used for local optimal solution refinement, and the two achieve iteration information sharing through a cross-validation mechanism; S5 resonance response reconstruction and dynamic regulation: Reconstruct the resonance response based on the optimized parameters, extract the enhanced weak fault features, and use an LSTM network to adjust the delay time parameter in real time, so that the system resonance frequency and the fault feature frequency remain dynamically synchronized.
2. The method of claim 1, wherein the method is characterized by: In the step S1, the anti-interference packaging structure is adopted for the acquisition devices of the original vibration signals and the synchronous sound field signals, and the spatial layout positions of the vibration sensors and the sound field sensors are optimized and configured according to the structural characteristics of the measured equipment.
3. The method of claim 1, wherein the method is characterized by: In the step S1, the multi-modal coupling excitation mechanism is realized through a feature cascading fusion mode, and the fused feature vectors are processed by a deep residual network, and the residual unit of the network adopts a pre-activation structure and contains a cross-layer connection.
4. The method of claim 1, wherein the method is characterized by: In the step S1, the multi-modal coupling excitation and dynamic delay resonance regulation characteristics are realized by dynamically loading a sound source modulation function, and the function automatically adjusts the excitation characteristics according to the load change of the equipment, and triggers the excitation mode switching when the load rate changes significantly.
5. The method of claim 1, wherein the method is characterized by: In the step S2, the sound field-vibration coupling term is realized by introducing a dynamic weight coefficient, and the coefficient is automatically adjusted according to the running state of the equipment, and when the sound field intensity changes significantly, the weight coefficient is adjusted synchronously according to a preset rule.
6. The method of claim 1, wherein the method is characterized by: In the step S3, the initialization model adopts a hybrid initialization strategy, part of the parameters are set to an initial range based on the prior knowledge of the equipment, and the other part of the parameters are extracted as initial values from historical running data through unsupervised learning.
7. The method of claim 1, wherein the method is characterized by: In the step S4, the cooperative mechanism of the genetic algorithm and the particle swarm algorithm is realized by the following way: periodically exchange population information during the optimization process, and combine the global search capability of the genetic algorithm with the local convergence characteristics of the particle swarm algorithm.
8. The method of claim 1, wherein the method is characterized by: In the step S4, the adaptive optimization strategy combines the genetic algorithm and the particle swarm optimization algorithm: when the global exploration process of the genetic algorithm and the local convergence speed of the particle swarm optimization algorithm differ significantly, the step parameter is automatically adjusted by a dynamic weight coefficient, the adjustment rule is optimized in real time according to the change rate of the fitness function of the two algorithms, and the balance between global search and local refinement is ensured.
9. The method of claim 1, wherein the method is characterized by: In the step S5, the LSTM network adopts a multi-resolution analysis technology to decompose and reconstruct the optimized signal, and selects a frequency band component containing main fault feature information for feature enhancement.
10. The method of claim 1, wherein the method is characterized by: In the step S5, the delay time parameter is adjusted in real time by an adaptive controller, and the adjustment strategy is automatically switched according to the change of the equipment operating state, and the response mode is enabled when the speed fluctuation is detected.