A gear performance precise evaluation method based on voiceprint analysis

Abstract

The application discloses a kind of gear performance precision evaluation methods based on voiceprint analysis, it is related to gear performance evaluation field;Different gears determine the optimal arrangement position of acoustic sensor, according to the real-time adjustment sampling frequency and time length collection voiceprint data according to gear operating state change;Compound denoising network based on deep learning is used to denoise, and data is processed using a new normalization method;Extract a variety of innovative features to construct a feature vector, and use a deep forest combination model to establish an evaluation model;The feature vector of the gear to be evaluated is input into the model to evaluate the performance, and the evaluation result is verified by virtual verification and other methods, and the model is optimized based on feedback.The application accurately evaluates the performance of the gear, intelligently collects voiceprint data, and improves the evaluation accuracy through feature extraction and model construction;Real-time monitoring and early warning, suitable for different gears;Blockchain ensures data reliability, evaluation model continuously evolves, reduces maintenance costs, and ensures stable operation of equipment.

Description

Technical Field

[0001] This invention relates to the field of gear performance evaluation technology, and in particular to a precise gear performance evaluation method based on acoustic pattern analysis. Background Technology

[0002] In modern industrial production, gears, as key transmission components of mechanical equipment, are widely used in numerous fields such as automobiles, aerospace, shipbuilding, and machine tools. The performance of gears directly affects the operational stability, reliability, and service life of the entire equipment. Gear failure can not only lead to equipment downtime and reduced production efficiency but may also cause serious safety accidents. Therefore, accurate evaluation of gear performance, timely detection of potential problems, and implementation of corresponding measures are crucial for ensuring the safe and efficient operation of industrial production.

[0003] Traditional gear performance evaluation methods have several limitations. Physical testing methods, such as directly observing gear wear and measuring tooth profile accuracy, often require stopping the machine and disassembling the equipment, which is cumbersome and destructive, and cannot achieve real-time online monitoring. While methods relying on vibration signal analysis can reflect the gear's operating status to some extent, they are easily affected by overall equipment vibration and noise interference, leading to low accuracy in the evaluation results. Furthermore, these methods typically only detect obvious gear faults and are insufficient for effectively diagnosing early potential faults and predicting performance.

[0004] With the development of industrial intelligence, higher demands are being placed on the accuracy, real-time performance, and intelligence of gear performance evaluation. Acoustic fingerprint analysis technology, as a non-contact, real-time monitoring method, is gradually gaining attention. However, gear performance evaluation technology based on acoustic fingerprint analysis is still in its developmental stage. Existing acoustic fingerprint analysis methods struggle to comprehensively and accurately acquire acoustic fingerprint signals that reflect the actual operating state of gears, and the acquisition process is easily affected by environmental noise. Furthermore, in terms of feature extraction and model building, they lack sufficient specificity and innovation, failing to fully exploit the rich information contained in the acoustic fingerprint signals. This results in insufficient accuracy and generalization ability of the evaluation model, making it difficult to meet the needs of complex and ever-changing industrial application scenarios. Therefore, developing a more efficient and accurate gear performance evaluation method based on acoustic fingerprint analysis is urgently needed. Summary of the Invention

[0005] The present invention proposes a method for accurate evaluation of gear performance based on acoustic pattern analysis to solve the problems mentioned in the prior art.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a method for accurate evaluation of gear performance based on acoustic pattern analysis, comprising:

[0007] Voiceprint data acquisition steps: For different gears, an intelligent sensor layout algorithm is used to determine the optimal position of the sound sensor. An adaptive sampling strategy is introduced in the acquisition process to adjust the sampling frequency and duration in real time according to the gear running status.

[0008] Voiceprint data preprocessing steps: A convolutional neural network (CNN) combined with a recurrent neural network (RNN) is used for noise reduction. The network is trained with noisy and noise-free data to adaptively identify and remove noise. An improved normalization method is proposed, and voiceprint data is processed using a quantile-based normalization formula.

[0009] Voiceprint feature extraction steps: extract time domain, frequency domain, time-frequency domain and phase space reconstruction features, reconstruct the voiceprint time series through delayed embedding method, extract features to capture early fault information, introduce multi-scale entropy features, and decompose the voiceprint signal at multiple scales to calculate the entropy value.

[0010] The steps for constructing the feature vector are as follows: using a feature importance ranking method, we calculate the correlation and mutual information between features and gear performance indicators to determine the score, select high-scoring features to form the feature vector, and remove redundant and irrelevant features;

[0011] The steps to establish an evaluation model are as follows: Build an evaluation model based on deep forest, combine decision tree and deep learning ideas, train the model using an adaptive forest growth strategy, introduce a model fusion mechanism to integrate deep forest and traditional machine learning models, and use a cross-validation-based optimization algorithm to dynamically adjust the weights.

[0012] Gear performance evaluation steps: Before inputting feature vectors into the model, compare them with historical feature vectors to calculate similarity. Evaluate the gear performance status, remaining life, and confidence interval of the model output gear. Quantify the uncertainty of the evaluation results through Monte Carlo simulation and Bayesian inference methods.

[0013] Results verification and feedback steps: Introduce virtual verification technology, compare simulation and actual voiceprint features, establish an adaptive optimization mechanism based on feedback control theory, and adjust the evaluation model parameters and structure according to the deviation.

[0014] Furthermore, the voiceprint data acquisition step employs acoustic metamaterial-assisted acquisition technology. By optimizing the acoustic metamaterial structure and parameters to match the characteristic frequencies of the target gear's voiceprint, and combining this with sound, vibration, temperature, and stress sensor fusion technology, the data is processed using an information entropy-based fusion algorithm. Weights are determined based on the magnitude of the sensor data's information entropy, and the fusion formula is as follows: S fused For the fused signal, S i For the i-th sensor signal, H i Let be the entropy of the data from the i-th sensor.

[0015] Furthermore, the voiceprint feature extraction step employs a sparse representation-based feature extraction method. By solving a sparse optimization problem, the sparse representation coefficients of the signal are obtained. Then, higher-order cumulants are calculated using higher-order cumulant features. The formula for calculating higher-order cumulants is as follows: E is the expected value, and k is the order of the cumulant.

[0016] Furthermore, the feature vector construction step adopts a graph neural network (GNN) based feature fusion method. The GNN learns and processes the feature map, aggregates the nodes through information transmission, mines the potential relationships between features, introduces a dynamic feature update mechanism, regularly updates and optimizes the feature vector, calculates the feature importance score based on the collected data, adds valuable features, and removes irrelevant features.

[0017] Furthermore, the evaluation model establishment process employs a generative adversarial network (GAN) model enhancement technique. The GAN consists of a generator and a discriminator. The generator generates simulated voiceprint feature vectors and performance indicators, while the discriminator distinguishes between real and generated data. The training dataset is expanded by training the generator and discriminator, and the evaluation model is trained using the dataset. A meta-learning mechanism is introduced to optimize the strategy through training.

[0018] Furthermore, the gear performance evaluation step adopts a multi-dimensional evaluation strategy to assess the overall performance status, remaining life, tooth surface contact strength, tooth root bending strength, and lubrication status of the gear. A corresponding sub-evaluation model is established for each dimension, and the evaluation results of the sub-evaluation models are comprehensively analyzed. Trend analysis and prediction techniques are introduced, and time series analysis methods are used to predict the trend of gear performance changes based on historical evaluation data and current evaluation results.

[0019] Furthermore, the result verification and feedback steps adopt a blockchain-based verification and feedback system. Voiceprint data, feature vectors, evaluation results, and actual detection results are recorded on the blockchain. The evaluation results are verified and fed back using the blockchain smart contract function. If there is a deviation between the evaluation results and the actual detection results, the smart contract automatically triggers the evaluation model optimization process. Through the blockchain network, evaluation data and experience are shared and exchanged.

[0020] Furthermore, the formula for adjusting the sampling frequency in the voiceprint data acquisition step is as follows: f s (t) is the sampling frequency at time t, f s0 Where is the initial sampling frequency, k is the adjustment coefficient, Δv is the change in speed or load, and v0 is the initial speed or load;

[0021] A distributed acoustic signature acquisition network is adopted, in which multiple acoustic sensors are arranged in different parts of the gear system to form a distributed acquisition network. The data collected by each sensor is transmitted to the central processing unit through wireless communication technology. Synchronous acquisition technology is used to collect data simultaneously, and the data is encrypted using an encryption algorithm during the data transmission process.

[0022] Furthermore, the voiceprint feature extraction step employs an adaptive spectral clustering feature classification method. This method divides data points using an adaptive spectral clustering algorithm, determines the number of clusters and cluster centers, and dynamically adjusts them based on feature distribution. By analyzing category features, it identifies the gear's operating state and fault type. Combined with a fuzzy feature extraction method, a fuzzy membership function is introduced to map voiceprint feature values ​​to different fuzzy sets, resulting in fuzzy feature vectors. The evaluation model establishment step uses a federated learning approach to train the model. The federated learning algorithm aggregates local model parameters to obtain a global evaluation model, and model compression technology is introduced to compress the evaluation model.

[0023] Furthermore, when changing gear suppliers, acoustic signature data is collected after the new gears are installed using acoustic signature analysis. Feature vectors are extracted in the time domain, frequency domain, time-frequency domain, and phase space, and then compared with the feature vectors of the original gears in the historical database. The similarity calculation uses the cosine distance formula. , For the new gear feature vector, The original gear feature vector is used; if the similarity is higher than the set threshold and the performance status and remaining life index of the evaluation model are not significantly different from the original gear, it is determined that the new gear will replace the original gear, supporting supplier switching decisions.

[0024] Compared with existing technologies, the beneficial effects of this invention are:

[0025] In terms of assessment accuracy, intelligent sensor deployment and adaptive sampling strategies enable the collection of more representative and accurate voiceprint data. Simultaneously, innovative feature extraction methods, such as combining phase space reconstruction, multi-scale entropy, and sparse representation-based feature extraction, can deeply mine the nonlinear, non-Gaussian, and complex dynamic features of voiceprint signals, comprehensively reflecting the gear's operating status. These features, combined with assessment models built using deep forest and model fusion technologies, significantly improve assessment accuracy, enabling precise judgment of gear health, prediction of remaining lifespan, and identification of potential faults, reducing equipment failures and production interruptions caused by misjudgments.

[0026] In terms of real-time monitoring and early warning capabilities, the use of a distributed acoustic signature acquisition network and real-time monitoring technology enables comprehensive and real-time acquisition of gear acoustic signature signals. Once an anomaly is detected in the acoustic signature signal, the system can quickly issue an early warning, reminding staff to take timely measures. This transforms equipment maintenance from traditional periodic maintenance to precise maintenance based on actual operating conditions, not only reducing maintenance costs but also effectively avoiding serious consequences caused by untimely fault detection, thus ensuring the safe and stable operation of the equipment.

[0027] In terms of versatility and adaptability, the evaluation method of this invention, by introducing technologies such as transfer learning and meta-learning, can quickly adapt to the gear performance evaluation needs under different types and operating conditions. Whether it is new gear equipment or existing equipment with changing operating conditions, its performance can be accurately evaluated, improving the versatility of the evaluation method and reducing the technical threshold and cost investment for enterprises in equipment monitoring and maintenance.

[0028] Furthermore, utilizing blockchain technology for result verification and feedback ensures the authenticity and traceability of data, enhancing the credibility of the assessment. By establishing a historical database and employing an adaptive optimization mechanism, the assessment model can continuously learn and evolve. As data accumulates, the accuracy and reliability of the assessment will continuously improve, providing long-term and stable technical support for gear performance evaluation in industrial production. Attached Figure Description

[0029] Figure 1 This is a schematic block diagram of a gear performance accuracy evaluation method based on acoustic text analysis proposed in this invention;

[0030] Figure 2 This is a schematic diagram comparing voiceprint features;

[0031] Figure 3 A diagram illustrating the comparison of model accuracy;

[0032] Figure 4 A schematic diagram illustrating the prediction of remaining gear life;

[0033] Figure 5 This is a schematic diagram showing the change in sample entropy of voiceprint signals at different scales. Detailed Implementation

[0034] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0035] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," and "counterclockwise," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0036] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified. Furthermore, the terms "installed," "connected," and "linked" should be interpreted broadly; for example, they may refer to a fixed connection, a detachable connection, or an integral connection; they may refer to a mechanical connection or an electrical connection; they may refer to a direct connection or an indirect connection through an intermediate medium; and they may refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances. The invention will now be described in further detail with reference to the accompanying drawings.

[0037] Reference Figures 1 to 5 A method for accurate evaluation of gear performance based on acoustic signature analysis, comprising:

[0038] Voiceprint data acquisition steps: For gears of different types and application scenarios, an intelligent sensor layout algorithm is used to determine the optimal placement of the sound sensors. This algorithm comprehensively considers factors such as the gear's geometry, transmission method, sound propagation path, and surrounding environmental noise distribution. By establishing a sound propagation model and a noise interference model, it calculates the signal-to-noise ratio and feature integrity of the voiceprint signals acquired at each potential location, thereby selecting the location that can acquire the most representative voiceprint signal. During the acquisition process, an adaptive sampling strategy is introduced. Based on the dynamic changes in the gear's operating state, such as fluctuations in speed and load, the sampling frequency and sampling duration are adjusted in real time. When the gear's operating state is relatively stable, the sampling frequency is appropriately reduced to decrease the amount of data; when the operating state changes drastically, the sampling frequency is increased to capture more detailed information. The sampling frequency adjustment formula is as follows: , where f s (t) is the sampling frequency at time t, f s0 is the initial sampling frequency, k is the adjustment coefficient, Δv is the change in speed or load, and v0 is the initial speed or load.

[0039] Voiceprint data preprocessing steps: A deep learning-based composite denoising network is used for denoising. This network combines the advantages of Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs). CNNs are used to extract local features of the voiceprint signal, while RNNs are used to process the temporal information of the signal. The network is trained with a large amount of noisy and noiseless voiceprint data, enabling it to adaptively identify and remove different types of noise. An improved normalization method is proposed, considering not only the amplitude range of the voiceprint signal but also the statistical distribution characteristics of the signal. A quantile-based normalization formula is adopted: Where Q1 and Q3 are the first and third quartiles of the voiceprint data, respectively, and x is the original voiceprint data. norm This is the normalized voiceprint data. This method is better able to handle voiceprint data with outliers.

[0040] Voiceprint feature extraction steps: In addition to traditional time-domain, frequency-domain, and time-frequency-domain features, features based on phase space reconstruction are also extracted. Phase space reconstruction theory is applied to process the voiceprint signal. The one-dimensional voiceprint time series is reconstructed into a high-dimensional phase space using a delayed embedding method. Features such as correlation dimension and maximum Lyapunov exponent are extracted from the phase space. These features reflect the nonlinear dynamic characteristics of the voiceprint signal and sensitively capture early gear fault information. Multi-scale entropy features are introduced to decompose the voiceprint signal into multiple scales, calculating the sample entropy or fuzzy entropy at each scale. Multi-scale entropy describes the complexity and irregularity of the voiceprint signal at different scales, comprehensively reflecting the gear's operating state. The multi-scale entropy calculation formula is as follows: , , where m is the embedding dimension, r is the similarity tolerance, T is the delay time, k is the scaling factor, and SampEn is the sample entropy function.

[0041] The feature vector construction process involves: In the feature extraction stage, various features are obtained from multi-source data on gear operation, including vibration signals collected by vibration sensors, temperature change data recorded by temperature sensors, and torque values ​​output by torque sensors. Extracted signal features cover time-domain features (such as mean, variance, and peak value), frequency-domain features (such as frequency components and power spectral density), and time-frequency-domain features (such as wavelet transform coefficients). The correlation index is calculated using the Pearson correlation coefficient, which accurately measures the degree of linear correlation between each feature and gear performance indicators (such as wear level and transmission efficiency). Values ​​closer to 1 or -1 indicate a stronger correlation; values ​​closer to 0 indicate a weaker correlation. The mutual information index is calculated based on information theory principles. By calculating the information entropy and conditional entropy between features and performance indicators, the amount of information a feature provides for the performance indicator is determined. A higher mutual information value means that the feature contains more effective information in describing gear performance. After evaluation, each feature is assigned an importance score based on the calculated correlation and mutual information indices and is rigorously ranked. When selecting features to construct the final feature vector, a reasonable threshold is set to filter out features with importance scores higher than the threshold. Simultaneously, cluster analysis and other techniques are used to further eliminate features that, while important, contain redundant information. This approach removes redundant features irrelevant to gear performance, avoiding the "curse of dimensionality" during model training, and significantly improves the quality of the feature vector, thereby greatly enhancing the training efficiency and prediction accuracy of the evaluation model.

[0042] The evaluation model is built by constructing a deep forest-based evaluation model. Deep forest is an ensemble learning model that combines the ideas of decision trees and deep learning. The initial number of forest layers and decision trees is determined based on the dimension and complexity of the gear feature vectors. Each layer of the decision tree forest acts as a feature processor, progressively refining the analysis of the input feature vectors. During training, the deep forest employs an adaptive forest growth strategy, dynamically adjusting the forest structure by calculating the training error. If the training error decreases slowly at a certain stage, the system automatically increases the number of forest layers to further uncover hidden information in the feature vectors. Simultaneously, the number of decision trees is increased in a timely manner based on data distribution and feature correlation to ensure comprehensive capture of the complex mapping relationship between feature vectors and gear performance indicators. This adaptive adjustment mechanism greatly improves the model's generalization ability and accuracy, enabling it to better handle the performance evaluation of gears under different working conditions and of different types. To further enhance the reliability of the evaluation results, a model fusion mechanism is introduced. The deep forest model is organically integrated with traditional machine learning models, such as support vector machines and neural networks. Support Vector Machines (SVMs) excel at finding the optimal classification hyperplane in high-dimensional spaces, performing exceptionally well with linear data or data that can be mapped to high-dimensional linearly separable data via kernel functions. Neural networks, on the other hand, leverage their powerful non-linear fitting capabilities to learn deep-seated abstract features within the data. During the fusion process, weighted voting or weighted averaging is used to synthesize the prediction results of multiple models. Cross-validation-based optimization algorithms play a crucial role in determining the weights. Specifically, the training dataset is divided into multiple subsets, with one subset used as the validation set and the remaining subsets as the training set. The performance metrics (such as accuracy, recall, and mean squared error) of each model on different validation sets are evaluated. Based on the evaluation results, the weights of each model in the fusion process are dynamically adjusted, giving greater weight to the model that performs better on the validation set in the final evaluation results. This leads to more accurate and stable final evaluation results, providing a more reliable basis for gear performance evaluation.

[0043] Gear performance evaluation steps: Using classic metrics such as Euclidean distance or cosine similarity, the feature vector of the gear to be evaluated is compared one by one with the feature vectors of known performance states in the historical database. Euclidean distance, by calculating the straight-line distance between two vectors in multidimensional space, intuitively reflects the degree of difference between feature vectors; cosine similarity measures the similarity of the directions of two vectors, focusing more on the direction of the feature vectors than their length. During the calculation, each set of feature vectors in the historical database is traversed to obtain a similarity score with the feature vector to be evaluated. A reasonable threshold is pre-set; if the similarity score falls below this threshold, it means that the operating state of the gear to be evaluated deviates from the normal range and may be abnormal. Further in-depth analysis is then conducted using fault tree analysis, expert systems, or more complex and refined evaluation strategies, such as deep neural network evaluation models, to accurately determine the actual operating condition of the gear. The richness and reliability of the evaluation model's output results are crucial to ensuring the scientific nature of gear maintenance decisions. The evaluation model, using Monte Carlo simulation and Bayesian inference methods, can not only provide the performance state and remaining life of the gear but also determine the confidence interval for each performance indicator. Monte Carlo simulation quantifies uncertainty by simulating various possible variations in gear performance indicators through extensive random sampling; Bayesian inference, on the other hand, continuously updates the probability estimate of gear performance states based on prior knowledge and new observational data. The combination of these two methods provides a more comprehensive and reliable basis for gear maintenance decisions, making maintenance plans more targeted and forward-looking, effectively reducing equipment failure risks, and improving the stability and economy of production operations.

[0044] Results Verification and Feedback Steps: The verification process breaks through the limitations of traditional single physical testing by introducing virtual verification technology. Using advanced Computer-Aided Engineering (CAE) software, a highly realistic virtual simulation model of the gear is established based on its precise geometric parameters (such as module, number of teeth, pressure angle, etc.), material properties (elastic modulus, Poisson's ratio, density, etc.), and manufacturing process information. This model accurately simulates the gear's operating state under different working conditions, such as different speeds, loads, and lubrication conditions. During the simulation, an acoustic simulation module is used to acquire acoustic signature data generated by the gear during operation. This data includes sound pressure level, frequency components, and sound intensity distribution. Simultaneously, high-precision acoustic sensors collect acoustic signature data of the gear under actual working conditions. Signal processing and pattern recognition technologies are used to meticulously compare the acoustic signature features obtained from the virtual simulation with those collected in reality. A feature matching algorithm is employed to calculate the similarity between the two in multiple domains, including the time domain, frequency domain, and cepstral domain, to verify the accuracy of the evaluation results. Based on feedback control theory, an adaptive optimization mechanism for the evaluation model is constructed. The deviation between the verification results and the evaluation results is used as a feedback signal. When the deviation exceeds a preset range, the system automatically triggers the optimization process. Regarding parameter adjustment, intelligent optimization algorithms such as gradient descent and genetic algorithms are used to iteratively update the weight parameters and threshold parameters in the evaluation model. If model structure optimization is involved, pruning algorithms may be used to remove redundant neurons or connections, or new network layers may be added or the network topology adjusted. This allows the evaluation model to continuously learn and adapt to changes in gear performance over time and under different operating conditions, thereby continuously improving the accuracy and reliability of the evaluation and providing a solid guarantee for the efficient operation and maintenance of the gears.

[0045] In this invention, acoustic metamaterial-assisted acquisition technology is employed during the acoustic fingerprint data acquisition stage. Acoustic metamaterials possess unique acoustic properties, such as negative refraction and acoustic focusing. Arranging them around the acoustic sensor enhances the acoustic fingerprint signal within a specific frequency range while suppressing signals at other interfering frequencies. By optimizing the structure and parameters of the acoustic metamaterial to match the characteristic acoustic fingerprint frequencies of the target gear, the acquisition quality of the acoustic fingerprint signal is improved. Simultaneously, multi-sensor fusion technology is combined, introducing stress and strain sensors in addition to acoustic, vibration, and temperature sensors. The data from these sensors are fused using an information entropy-based fusion algorithm. The weight of each sensor's data in the fusion process is determined based on its information entropy, resulting in more comprehensive and accurate information about the gear's operating status. The fusion formula is as follows: S fused For the fused signal, S i For the signal of the i-th sensor, H i Let be the information entropy of the i-th sensor data.

[0046] In this invention, a sparse representation-based feature extraction method is proposed for the voiceprint feature extraction step. The voiceprint signal is represented as a linear combination of an overcomplete dictionary. By solving a sparse optimization problem, the sparse representation coefficients of the signal are obtained. Features such as sparsity and the distribution of non-zero coefficients are extracted from these coefficients. Sparse representation can highlight important features in the voiceprint signal and reduce the influence of noise and redundant information. Simultaneously, in conjunction with higher-order cumulants, in addition to the third and fourth moments, higher-order cumulants are calculated. Higher-order cumulants more comprehensively describe the non-Gaussian nature of the voiceprint signal, improving the sensitivity to gear fault characteristics. The formula for calculating higher-order cumulants is as follows: , where E is the mathematical expectation and k is the order of the cumulant.

[0047] In this invention, the feature vector construction step employs a feature fusion method based on Graph Neural Networks (GNNs). The feature fusion process based on GNNs first treats different types of features extracted from the gear acoustic signature signal, such as time-domain features (e.g., root mean square value, peak value), frequency-domain features (e.g., frequency components, harmonic amplitude), and joint time-frequency features, as nodes in a graph. These nodes represent different aspects of the acoustic signature features, and their correlations are represented by edges in the graph. For example, certain frequency components may be associated with specific time-domain peak values; this association is represented by edges in the graph. By constructing such a feature graph, the GNN can perform in-depth learning and processing. During the GNN's computation, information is passed and aggregated between nodes. Each node not only considers its own feature information but also receives information from neighboring nodes, uncovering potential complex relationships between features through a series of convolution or aggregation operations. For example, it may be discovered that changes in certain frequency components occur simultaneously with changes in specific time-domain features and are related to a certain operating state or fault mode of the gear. After processing by the GNN, the originally scattered features are fused into a more representative feature vector. This feature vector can more comprehensively and accurately reflect the essential characteristics of the gear soundprint signal, providing a higher quality data foundation for subsequent performance evaluation. Furthermore, a dynamic feature update mechanism is introduced to adapt to changes in the gear's operating state. As the gear continues to run, its soundprint characteristics will change due to factors such as wear and load variations. The dynamic feature update mechanism periodically updates and optimizes the feature vector based on the increase in running time and the accumulation of new data. Specifically, the system recalculates the importance score of each feature using a specific algorithm based on newly acquired soundprint data. Features that show a higher correlation with the gear's current operating state in the new data have their weight increased; while features that are no longer relevant or have reduced correlation have their weight decreased or even removed. Simultaneously, if valuable features reflecting new gear states or new fault modes appear in the newly acquired data, they will be added to the feature vector. In this way, the feature vector can reflect changes in the gear's operating state in real time and accurately, ensuring that gear performance evaluation based on soundprint analysis always has high accuracy and timeliness, providing a more reliable basis for gear maintenance and management.

[0048] In this invention, the evaluation model establishment step employs model enhancement technology based on Generative Adversarial Networks (GANs). A GAN consists of two core components: a generator and a discriminator. In the gear acoustic signature analysis scenario, the generator's task is to generate simulated acoustic signature feature vectors and corresponding performance indicators. It learns the distribution patterns of real acoustic signature data and, using a complex neural network structure, attempts to generate samples similar to real data. For example, it generates corresponding simulated data based on the characteristics of acoustic signatures in frequency, amplitude, and phase under different gear states such as normal operation, wear, and malfunction. The discriminator acts as a "data identification expert," responsible for distinguishing between the simulated data generated by the generator and the real acoustic signature data collected. It continuously improves its discrimination ability by learning from a large amount of real and simulated data. During training, the generator and discriminator engage in intense "adversarial" competition. The generator strives to generate more realistic data to deceive the discriminator, while the discriminator continuously improves its discrimination ability to accurately identify simulated data. As this adversarial training continues, the generator gradually becomes able to generate extremely realistic simulated data, which can effectively expand the training dataset. Training the evaluation model using the expanded dataset allows it to learn more diverse relationships between acoustic signature features and performance indicators, significantly improving its generalization ability and robustness. This enables it to more accurately evaluate gear performance under different operating conditions and noise interference. Furthermore, a meta-learning mechanism is introduced to further optimize the evaluation model. Meta-learning aims to teach the model how to learn. In gear performance evaluation, by training on multiple different tasks (i.e., different gear types or operating conditions), the model can learn general initialization parameters and optimization strategies. For example, for gears of different materials and modules, and operating conditions under different loads and speeds, the meta-learning process allows the model to understand the commonalities and differences between these tasks. When encountering a new gear type or operating condition, the model can quickly adjust its parameters based on the learned initialization parameters and optimization strategies, achieving rapid convergence and quickly adapting to the new task. This allows for accurate evaluation of gear performance under new conditions, greatly improving the flexibility and practicality of the evaluation model and providing stronger technical support for accurate gear performance evaluation based on acoustic signature analysis.

[0049] In this invention, the gear performance evaluation process employs a multi-dimensional evaluation strategy. In terms of evaluation dimensions, it breaks through the limitations of traditional methods that only focus on overall performance and remaining life, delving into multiple key aspects of gear operation. Regarding tooth surface contact strength, the distribution of contact stress between the tooth surfaces during gear operation is considered. By establishing sub-evaluation models and utilizing techniques such as finite element analysis to simulate the tooth surface contact process, the magnitude and variation of contact stress under different operating conditions are analyzed to assess the tooth surface's ability to resist fatigue wear during long-term operation. For tooth root bending strength, based on the principles of mechanics of materials and combined with the gear's geometric parameters and load conditions, a corresponding model is established to calculate the bending stress on the tooth root. The influence of factors such as the tooth root fillet radius and stress concentration factor on bending strength is also considered to accurately assess the risk of tooth root fracture under alternating loads. Lubrication condition evaluation focuses on the viscosity of the lubricating oil, oil film thickness, and lubrication method. Using oil analysis techniques and tribological theory, it is determined whether the lubricating oil can effectively reduce friction and wear between the tooth surfaces and maintain a good lubrication environment. For each dimension's sub-evaluation model, a large amount of relevant data is collected for training and validation. For example, data such as tooth surface temperature and wear under different operating conditions are collected for the tooth surface contact strength model; stress and strain data of the tooth root are obtained to optimize the tooth root bending strength model; and the physicochemical properties of lubricating oil and lubrication system parameters are monitored to improve the lubrication condition assessment model. Then, a comprehensive analysis algorithm is used to integrate the results of these sub-assessment models. Through weight allocation and data fusion, a more comprehensive and detailed gear performance assessment report is obtained, clearly showing the gear's performance status in various aspects. Furthermore, trend analysis and prediction techniques are introduced. Time series analysis methods, such as the ARIMA model, are used to deeply mine historical assessment data. This historical data contains various performance indicators and operating condition information of the gear at different periods. Through steps such as checking the stationarity of the data series and parameter estimation, a suitable prediction model is established. Combined with the current assessment results, this model can accurately predict the changing trend of gear performance over a future period. For example, it can predict the growth rate of tooth surface wear and the decline in tooth root bending strength, providing a reliable basis for equipment maintenance personnel to formulate scientific and reasonable maintenance plans in advance, avoiding equipment downtime due to gear failures, and ensuring the continuity and stability of production.

[0050] In this invention, the result verification and feedback steps establish a blockchain-based verification and feedback system. Blockchain, with its unique distributed ledger technology, provides a solid guarantee for the data security and trustworthiness of the evaluation process. Each time a gear undergoes voiceprint analysis and evaluation, the voiceprint data, extracted feature vectors, final evaluation results, and actual detection results are accurately recorded on the blockchain. The immutability of blockchain stems from its encryption algorithm and chain-like data structure. Each data block contains the hash value of the previous data block. Once data is recorded, any attempt to modify its content will result in a change in the hash value, which will then be detected by the entire blockchain network, ensuring the authenticity and reliability of the data. Simultaneously, data traceability allows users to review the evaluation history at any time and understand every detail of the evaluation process. Smart contracts play a crucial role in this system. Essentially, they are automatically executed code deployed on the blockchain. When the evaluation results are compared with the actual detection results, if a discrepancy is found, the smart contract will automatically trigger the optimization process of the evaluation model according to pre-set rules. For example, it will call relevant algorithms to adjust the parameters of the evaluation model or retrain the model. This process requires no human intervention, significantly improving the timeliness and accuracy of feedback, enabling the evaluation model to continuously adapt to real-world conditions and optimize performance. Furthermore, the blockchain network provides an open platform where different users and organizations can share and exchange evaluation data and experience. Through a consensus mechanism, nodes can securely exchange information without needing to trust third parties. This sharing of data and experience breaks down information barriers, promoting exchange and cooperation across the industry in gear performance evaluation technology. Different organizations can learn from each other's successful cases to jointly explore solutions to challenges encountered in the evaluation process, driving continuous technological progress throughout the industry and improving the overall level of accurate gear performance evaluation methods based on acoustic fingerprint analysis.

[0051] In this invention, the voiceprint data acquisition step employs a distributed voiceprint acquisition network. In the gear system, considering the differences in voiceprint signals generated by different parts of the gear during operation, multiple sound sensors are arranged at different locations in the gearbox, such as near the gear meshing point and bearing housing. These sensors possess high sensitivity and a wide frequency response range, enabling them to accurately capture various voiceprint signals generated during gear operation. The sensors are interconnected via wireless communication technology and establish a communication link with the central processing unit. For example, Low Power Wide Area Network (LPWAN) technology is used to ensure stable and reliable data transmission in complex industrial environments. Each sensor transmits the acquired voiceprint signals to the central processing unit in real time, achieving comprehensive, multi-angle acquisition of gear voiceprint signals, essentially establishing a 360-degree "sound monitoring network" for the gear's operating state, capturing no subtle sound changes. Synchronous acquisition technology plays a crucial role. To ensure the consistency of data acquired by each sensor in time, the system employs a high-precision clock synchronization mechanism. The clocks of all sensors are precisely calibrated using technologies such as the Global Positioning System (GPS) or Precise Time Protocol (PTP). This ensures the accuracy and consistency of the timestamps regardless of where the voiceprint signal is collected within the gear system, providing a solid foundation for subsequent time-series-based voiceprint signal analysis and significantly improving the accuracy of the analysis results. During data transmission, data security and privacy are paramount. Advanced encryption algorithms, such as Advanced Encryption Standard (AES) or Elliptic Curve Cryptography (ECC), are employed to encrypt the collected voiceprint data. The data is converted to ciphertext before transmission and can only be restored to its original form after decryption by the central processing unit. This encryption effectively prevents data theft or tampering during transmission, ensuring the security and privacy of the voiceprint data and providing strong support for the reliable application of voiceprint-based methods for accurate gear performance evaluation.

[0052] In this invention, the voiceprint feature extraction step employs an adaptive spectral clustering feature classification method. Numerous features, such as frequency and amplitude features, are extracted from the voiceprint signal generated by gear operation. The adaptive spectral clustering algorithm acts as an intelligent "classifier," overcoming the limitations of traditional clustering algorithms that require pre-setting the number of clusters and their centers. Based on the similarity measure between these voiceprint feature data points, the algorithm constructs a similarity graph. In this graph, the connection weights between data points reflect their degree of similarity. By analyzing the Laplacian matrix of the graph and utilizing its eigenvalues ​​and eigenvectors, the optimal number of clusters and cluster centers are automatically determined. During the clustering process, the clustering strategy is dynamically adjusted according to the actual distribution characteristics of the voiceprint features. For example, when the gear is in different wear stages or load states, the distribution of voiceprint features changes. The adaptive spectral clustering algorithm can keenly capture these changes and accurately classify the voiceprint features into different categories. Each category corresponds to a specific gear operating state or fault mode, such as normal operation, slight wear, and severe wear. By deeply analyzing the commonalities and differences in features within each category, the current operating state and fault type of the gear can be identified more accurately. Furthermore, a fuzzy feature extraction method was incorporated, introducing a fuzzy membership function. Due to the inherent uncertainty and fuzziness of voiceprint features—for example, during the transition of a gear from a normal to a faulty state, the voiceprint features do not exhibit clear boundaries—the fuzzy membership function maps the values ​​of voiceprint features to different fuzzy sets. For a given voiceprint feature value, it may simultaneously belong to both the "normal" and "slightly worn" fuzzy sets with a certain membership degree. The fuzzy feature vector obtained in this way can more flexibly and comprehensively describe the characteristics of voiceprint features, effectively improving the ability to identify complex gear operating states, and enabling more accurate judgments even when feature boundaries are unclear. The evaluation model was established using a federated learning approach for training. In a distributed data environment, each participant can train a local model locally. Through the federated learning algorithm, the parameters of each local model are safely and efficiently aggregated to obtain a global evaluation model. To further improve the model's practicality and deployment efficiency, model compression technology was also introduced. For example, methods such as pruning and quantization can be used to remove redundant parameters and connections in the model, reducing the model's storage space and computational overhead without significantly affecting its performance. This allows the evaluation model to be applied more quickly and conveniently to real-world gear performance evaluation scenarios.

[0053] In this invention, when equipment faces the need to change gear suppliers, acoustic signature analysis plays a crucial role in determining whether a new gear can perfectly replace the original one. After completing traditional quality checks such as hardness and dimensions, acoustic signature data of the equipment operation after the new gear is installed is first collected. High-precision acoustic sensors are used to ensure that even the most subtle sound changes during equipment operation are captured. Advanced signal processing techniques are employed to analyze the signal from multiple dimensions, including the time domain, frequency domain, time-frequency domain, and phase space reconstruction. In the time domain analysis, the characteristics of the acoustic signature signal's amplitude and period changing over time are focused on; frequency domain analysis focuses on the energy distribution of the signal at different frequencies; time-frequency domain analysis integrates time and frequency dimensions, revealing the frequency characteristics of the signal at different times more comprehensively; phase space reconstruction, through processing the acoustic signature signal, uncovers its deeper dynamic characteristics. Through these analyses, a feature vector containing multi-dimensional information is obtained. The cosine distance formula is used. Calculate the feature vector of the new gear Compared with the original gear feature vector The similarity is calculated using a formula based on the vector dot product and vector magnitude, scientifically measuring the angular relationship between two feature vectors in space. The smaller the angle, the higher the similarity. If the calculated similarity exceeds a set threshold (e.g., 0.9), further evaluation of the new gear's performance status, remaining lifespan, and other indicators using an evaluation model is required. Only when these indicators differ from the original gear by less than 5% can the new gear be deemed an effective replacement, providing strong support for supplier switching decisions. The entire process utilizes multi-dimensional technical means to ensure the scientific rigor and accuracy of gear replacement decisions.

[0054] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for precise evaluation of gear performance based on voiceprint analysis, characterized in that, The method comprises the following steps: Voiceprint data acquisition step: adopt intelligent sensor layout algorithm to determine the optimal position of acoustic sensor for different gears, introduce adaptive sampling strategy in the collection process, and adjust the sampling frequency and duration in real time according to the running state of the gear; Voiceprint data preprocessing step: adopt convolutional neural network (CNN) combined with recurrent neural network (RNN) for denoising, train the network with noisy and noise-free data, adaptively identify and remove noise, and propose an improved normalization method to process voiceprint data based on quantile normalization formula; Voiceprint feature extraction step: extract time domain, frequency domain, time-frequency domain and phase space reconstruction features, reconstruct the voiceprint time series by delay embedding method, extract features to capture early fault information, and introduce multi-scale entropy feature to calculate the entropy value of the multi-scale decomposed voiceprint signal; Feature vector construction step: adopt feature importance sorting method to calculate the correlation and mutual information between features and gear performance indicators to determine the score, select high-score features to form a feature vector, and remove redundant and irrelevant features; Establishing evaluation model step: build a deep forest evaluation model based on the combination of decision tree and deep learning idea, adopt adaptive forest growth strategy to train the model, introduce model fusion mechanism to fuse deep forest and traditional machine learning model, and adopt cross-validation optimization algorithm to dynamically adjust the weight; Gear performance evaluation step: compare the similarity of the feature vector with the historical feature vector before inputting the feature vector into the model, and the evaluation model outputs the gear performance state, remaining life and confidence interval. The uncertainty of the evaluation result is quantified by Monte Carlo simulation and Bayesian inference method; Result verification and feedback step: introduce virtual verification technology to compare the simulation and actual voiceprint features, establish an adaptive optimization mechanism based on feedback control theory, and adjust the evaluation model parameters and structure according to the deviation.

2. The precise evaluation method of gear performance based on voiceprint analysis according to claim 1, characterized in that, The voiceprint data collection step adopts an acoustic metamaterial auxiliary collection technology, optimizes acoustic metamaterial structure and parameter matching target gear voiceprint characteristic frequency, combines acoustic, vibration, temperature and stress sensor fusion technology for data fusion processing, adopts an information entropy fusion algorithm, determines the weight according to the information entropy size of the sensor data, and the fusion formula is , S fused is the fused signal, S i is the i-th sensor signal, , H i is the information entropy of the i-th sensor data.

3. The precise evaluation method of gear performance based on voiceprint analysis according to claim 1, characterized in that, The voiceprint feature extraction step adopts a feature extraction method based on sparse representation, and obtains signal sparse representation coefficients by solving a sparse optimization problem, and calculates a higher order cumulant in combination with a high-order cumulant feature , E is a mathematical expectation, and k is a cumulant order.

4. The precise evaluation method of gear performance based on voiceprint analysis according to claim 1, characterized in that, The feature vector construction step adopts a graph neural network (GNN) feature fusion method to learn and process feature maps using GNN, aggregates node information through information transmission, mines the potential relationship between features, introduces a dynamic feature updating mechanism to periodically update and optimize the feature vector, calculates the feature importance score based on the collected data, adds valuable features, and removes irrelevant features.

5. The precise evaluation method of gear performance based on voiceprint analysis according to claim 1, characterized in that, The evaluation model establishment step adopts a generative adversarial network (GAN) model enhancement technology. The GAN consists of a generator and a discriminator. The generator generates simulated voiceprint feature vectors and performance indicators, and the discriminator discriminates between real data and generated data. The training data set is expanded by training the generator and discriminator, the evaluation model is trained using the data set, and a meta-learning mechanism is introduced to optimize the training strategy.

6. The precise evaluation method of gear performance based on voiceprint analysis according to claim 1, characterized in that, The gear performance evaluation step adopts a multi-dimensional evaluation strategy to evaluate the overall performance state, remaining life, tooth surface contact strength, tooth root bending strength and lubrication state of the gear. For each dimension, a corresponding sub-evaluation model is established, the evaluation results of the sub-evaluation models are comprehensively analyzed, trend analysis and prediction technology is introduced, and the time series analysis method is used to predict the performance change trend of the gear based on the historical evaluation data and the current evaluation result.

7. The precise evaluation method of gear performance based on voiceprint analysis according to claim 1, characterized in that, The result verification and feedback step adopts a blockchain-based verification and feedback system, records the voiceprint data, feature vectors, evaluation results and actual detection results in the blockchain, verifies the feedback evaluation results using the smart contract function of the blockchain, and automatically triggers the evaluation model optimization process when the evaluation results deviate from the actual detection results, and shares and exchanges evaluation data and experience through the blockchain network.

8. The precise evaluation method of gear performance based on voiceprint analysis according to claim 1, characterized in that, The sampling frequency adjustment formula in the voiceprint data collection step is , f s (t) is the sampling frequency at time t, f s0 is the initial sampling frequency, k is the adjustment coefficient, Δv is the change amount of the rotation speed or load, and v0 is the initial rotation speed or load. A distributed voiceprint collection network is adopted, multiple acoustic sensors are arranged at different positions of the gear system to form a distributed collection network, data collected by each sensor is transmitted to a central processing unit through wireless communication technology, synchronous collection technology is adopted to collect data simultaneously, and an encryption algorithm is used to encrypt data during the data transmission process.

9. The precise evaluation method of gear performance based on voiceprint analysis according to claim 1, characterized in that, The voiceprint feature extraction step adopts an adaptive spectral clustering feature classification method, divides data points through an adaptive spectral clustering algorithm, determines the number of clusters and cluster centers, dynamically adjusts according to feature distribution, identifies the running state and fault type of the gear through analysis of category features, introduces a fuzzy membership function to map the voiceprint feature values to different fuzzy sets to obtain a fuzzy feature vector by combining a fuzzy feature extraction method, and the evaluation model is established by training the model in a federated learning manner, aggregating local model parameters to obtain a global evaluation model, and introducing model compression technology to compress the evaluation model.

10. The precise evaluation method of gear performance based on voiceprint analysis according to claim 1, characterized in that, The device collects the sound print data of the new gear after installation through sound print analysis when changing the gear supplier, extracts the feature vectors of time domain, frequency domain, time-frequency domain and phase space reconstruction, and performs similarity matching with the original gear feature vectors in the historical database, and the similarity calculation adopts the cosine distance formula , is the feature vector of the new gear, is the feature vector of the original gear If the similarity is higher than the set threshold and the performance state and remaining life indicators output by the evaluation model have no significant difference from the original gear, the new gear is determined to replace the original gear, and the supplier switching decision is supported.

Images