Motor abnormal sound intelligent identification method based on time-frequency characteristics and auditory adaptation

Abstract

This invention relates to an intelligent method for identifying abnormal motor noises based on time-frequency features and auditory adaptation, belonging to the field of motor fault detection technology. The method includes: real-time acquisition of motor operating sound signals through a multi-directional acoustic signal acquisition array, followed by signal preprocessing and dynamic framing; generation of a time-frequency spectrum using short-time sliding window time-frequency conversion, establishing an auditory baseline for normal operation sound using auditory perception adaptation dynamic weighting technology, and performing hierarchical weighting processing on the time-frequency spectrum to generate fused time-frequency weighted features; construction of a time-frequency weighted feature-enhanced residual network model, achieving abnormal feature extraction and identification through cross-dimensional fusion mechanisms, fault feature transfer learning mechanisms, and fault feature attention mechanisms; and completion of status identification and alarm response using a dynamic adaptation inference mechanism and a hierarchical linkage alarm system. This invention effectively improves the accuracy and real-time performance of motor abnormal noise identification and is suitable for motor status monitoring in complex industrial environments.

Description

Technical Field

[0001] This invention belongs to the field of motor fault detection technology, specifically relating to an intelligent identification method for abnormal motor noises based on time-frequency characteristics and auditory adaptation. Background Technology

[0002] As a core power source in industrial automation, the operating status of electric motors directly affects the stability and reliability of the entire production system. During motor operation, abnormal noises often occur due to mechanical wear, component aging, or improper installation. These abnormal noises are usually early signs of potential faults such as bearing damage, rotor imbalance, or stator winding failure. If these abnormalities are not detected and addressed in a timely manner, they can lead to decreased equipment performance and affected product quality, or even serious equipment damage or safety accidents, resulting in production interruptions and economic losses.

[0003] Currently, in the field of motor condition monitoring, traditional fault detection methods mainly suffer from the following limitations: First, vibration sensor-based detection methods require the installation of sensors on the motor housing, which faces challenges in actual industrial production environments, such as limited installation locations and complex sensor deployment, especially in high-speed rotating or sealed motors. Second, the acquisition of vibration signals by these methods is easily affected by the mechanical transmission path; signal characteristics acquired at different installation locations may vary significantly, reducing the reliability of the detection. Furthermore, vibration analysis methods lack sensitivity to early, subtle faults, often only effectively identifying them when the fault has developed to a certain extent, thus losing the significance of early warning.

[0004] Another traditional method relies on experienced technicians to diagnose faults through manual listening. While this method directly utilizes the human ear's sensitivity to sound, it is susceptible to the influence of factors such as the technician's subjective experience, ambient noise, and fatigue from continuous work, making it difficult to guarantee the consistency and reliability of the test results. Especially in modern large-scale manufacturing enterprises, facing numerous motor devices, relying solely on manual inspection is not only inefficient but also makes it difficult to achieve real-time monitoring and timely early warning.

[0005] In recent years, with the development of acoustic signal processing technology, motor fault detection methods based on sound analysis have gradually attracted attention. These methods collect sound signals from the motor during operation, extract feature parameters using digital signal processing techniques, and then combine this with pattern recognition algorithms for state determination. Compared to vibration detection, sound signal acquisition is more convenient, requiring no direct contact with the equipment, and acoustic signals contain richer state information. However, existing acoustic detection methods still have significant shortcomings: most methods employ traditional spectrum analysis techniques, failing to fully consider the perceptual characteristics of the human auditory system, leading to discrepancies between feature extraction and human auditory experience; they perform poorly when dealing with background noise interference in complex industrial environments, especially in scenarios where multiple devices operate simultaneously, where noise aliasing severely affects recognition accuracy; furthermore, existing methods have poor adaptability to changes in motor operating conditions, easily leading to misjudgments when motor load and speed change.

[0006] In summary, existing motor noise detection technologies still have many limitations in terms of accuracy, reliability, and practicality. There is an urgent need for an intelligent identification method that can fully utilize the characteristics of sound signals, adapt to complex industrial environments, and be consistent with human auditory perception. Summary of the Invention

[0007] To address the aforementioned problems in the existing technology, this invention provides an intelligent identification method for abnormal motor noise based on time-frequency characteristics and auditory adaptation;

[0008] The objective of this invention can be achieved through the following technical solutions:

[0009] A method for intelligent identification of abnormal motor noise based on time-frequency characteristics and auditory adaptation includes:

[0010] S1: Based on a multi-directional acoustic signal acquisition array, the motor operation acoustic signal is acquired in real time; a bandpass filter passband is set based on the frequency band associated with the rated speed of the motor to filter out environmental noise and motor background noise; the acoustic signal is normalized, and the duration and interval of short time frames are dynamically set based on the real-time speed of the motor to divide the continuous acoustic signal into multiple short time frames.

[0011] S2: Short-time sliding window time-frequency conversion is adopted, and the window overlap rate is dynamically adjusted based on the amplitude fluctuation of the time domain of the acoustic signal frame to generate the time spectrum; simultaneously, auditory perception adaptation dynamic weighting technology is adopted to establish a normal operating acoustic baseline, and frequency band weighting coefficients are generated by the deviation between the real-time signal and the baseline. Layered weighting processing is performed on the time spectrum to generate fused time-frequency weighted features that enhance abnormal features.

[0012] S3: Construct a time-frequency weighted feature-enhanced residual network model, and perform multi-scale feature association on the fused time-frequency weighted features through a cross-dimensional fusion mechanism to strengthen the coupling features of abnormal sound signals in the time-frequency domain; introduce a fault feature transfer learning mechanism to dynamically transfer fault discrimination knowledge in the pre-trained model; embed a fault feature attention mechanism to focus on the time-frequency region related to abnormal noise and suppress the interference of normal operation sound features;

[0013] S4: Employ a dynamic adaptation reasoning mechanism based on operating conditions, and adjust the identification confidence threshold by combining the dynamic parameters of time-frequency conversion and weighting coefficients; establish a hierarchical linkage alarm system, and trigger corresponding levels of sound and light, system work orders and equipment linkage responses based on abnormal noise characteristics; execute full-link data closed-loop recording, synchronously store the original sound signal, fuse time-frequency weighted features and abnormal noise identification results, and conduct fault source analysis and model iteration optimization.

[0014] As a preferred technical solution of the present invention, the multi-directional acoustic signal acquisition array adopts an acoustic radiation characteristic adaptation method, divides the acquisition area along the motor housing according to the vibration mode distribution law, forms a network of acquisition points distributed in the axial and circumferential directions of the housing, dynamically determines the layout spacing based on the acoustic signal propagation characteristics during motor operation, densifies the layout in the near sound source area to capture detailed signals, and sparsely lays the layout in the far sound source area to cover and diffuse signals, and performs time alignment of the acoustic signals through a synchronous triggering mechanism to form a multi-dimensional acoustic signal acquisition matrix.

[0015] Specifically, the method for filtering out environmental noise and motor background noise is as follows:

[0016] By using bandpass filtering to retain the acoustic signal within the frequency band associated with the rated operation of the motor, an adaptive noise cancellation mechanism based on the acoustic pattern of normal motor operation is adopted. The continuous and stable background noise component within the frequency band is extracted as a reference signal. The filter coefficient is dynamically adjusted by the adaptive filter to eliminate noise components that match the characteristics of the reference signal, while retaining the transient acoustic signal components within the frequency band. The filtered signal is then smoothed to eliminate edge fluctuations introduced by the filtering.

[0017] Specifically, the segmented continuous acoustic signal is divided into multiple short time frames, and the specific method is as follows:

[0018] Based on the real-time speed signal of the motor, the frame length parameter is dynamically calculated through the speed-frame length mapping function to establish a nonlinear inverse proportional relationship between speed and frame length.

[0019] Simultaneously, the non-stationarity index is calculated based on the signal time-domain amplitude change rate, and the frame overlap rate is adaptively adjusted through the overlap rate adjustment function.

[0020] A frame parameter coordination mechanism is established, which uses an association function to constrain the frame length parameter and the overlap rate parameter, thereby forming a coordinated adjustment of the two-dimensional parameters.

[0021] Specifically, the short-time sliding window time-frequency conversion includes:

[0022] The analysis window length is dynamically adjusted based on the signal nonstationarity evaluation function. A nonlinear mapping relationship between the window length parameter and the signal nonstationarity is established. The window length is adaptively adjusted by continuously monitoring the changing trend of the signal's time-domain characteristics.

[0023] A dynamic adjustment mechanism for overlap rate is established based on the temporal amplitude fluctuation of the acoustic signal frame. The window overlap ratio is determined by analyzing the gradient of signal amplitude change, and a functional correspondence between overlap rate and amplitude fluctuation is constructed.

[0024] Multi-dimensional feature extraction is performed on the spectrum transformation results, and the spectrum amplitude is dynamically adjusted through a nonlinear compression function to form a multi-dimensional time-frequency feature expression that includes spectrum amplitude and dynamic change features.

[0025] Specifically, the method for establishing a normal operating auditory baseline is as follows:

[0026] Collect normal operating sound samples of motors under standard operating conditions, obtain the reference spectrum characteristics of each operating condition through time-frequency analysis, and construct a multi-operating-condition normal operating sound feature library.

[0027] An auditory perception weighting coefficient matrix is ​​established, and frequency band weights are assigned to steady-state spectral features based on the characteristics of human auditory perception. By analyzing the discriminative contribution of different frequency bands in abnormal noise identification, a frequency band weighting function is formed, and a weighted normal operation sound auditory feature template is generated.

[0028] A preset time-varying adaptive calibration mechanism is used to continuously collect normal sound signal samples and update the auditory feature template in real time using a sliding window statistical method combined with a time-varying attenuation factor.

[0029] Specifically, the method for performing hierarchical weighting processing on the time spectrum is as follows:

[0030] A hierarchical auditory perception weighted network is constructed, dividing time-spectrum processing into a basic auditory weighting layer, an anomaly feature enhancement layer, and a dynamic range adaptation layer. For the basic auditory weighting layer, a frequency weighting function simulating the auditory perception characteristics of the human ear is applied to perform preliminary weighting processing on the time-spectrum. For the anomaly feature enhancement layer, based on the spectral distribution characteristics of historical anomaly samples, the discrimination index of each frequency band in anomaly identification is calculated, and an anomaly sensitivity weighting matrix is ​​constructed. For the dynamic range adaptation layer, based on the amplitude distribution characteristics of the time-spectrum, a nonlinear compression function is used to dynamically adjust the frequency bands of different intensities.

[0031] Specifically, the cross-dimensional fusion mechanism includes:

[0032] A feature association analysis network is constructed. By calculating the covariance matrix of the time dimension feature sequence and the frequency dimension feature sequence, a cross-dimensional feature association map is established. A graph convolutional network is used to perform deep feature extraction on the feature association map to obtain cross-dimensional coupled feature representation.

[0033] A two-way feature interaction architecture is established, and the two-way interaction and information fusion of time-dimensional and frequency-dimensional features are realized through gated loop units;

[0034] A preset dynamic weight allocation mechanism is used. Based on the coupled feature representation output by the feature association analysis network, dynamic fusion weights related to the features are generated through a multilayer perceptron network. The weight coefficients are probabilistically processed using a normalized exponential function to achieve adaptive fusion of cross-dimensional features.

[0035] Specifically, the fault feature transfer learning mechanism includes:

[0036] A cross-domain feature mapping network is constructed. By establishing a feature alignment layer between the source domain pre-trained model and the target domain motor noise identification task, a domain adaptation transformation of deep features is performed, mapping the general acoustic feature space to the motor noise-specific feature space.

[0037] Establish a knowledge distillation transmission link, extract multi-scale feature response maps from the pre-trained model as teacher signals, and drive the target network to learn fault feature representations through an attention-guided feature matching loss function;

[0038] A preset dynamic fine-tuning strategy is used to adaptively adjust the learning rate of the network layers based on the distribution characteristics of the target domain samples, retaining general knowledge of the source domain while performing targeted optimization of the abnormal noise characteristics of the motor.

[0039] Specifically, the embedded fault feature attention mechanism includes:

[0040] A dual-path attention architecture is constructed. In the time dimension path, a multi-head self-attention mechanism is used to capture the temporal dependencies of abnormal sounds, and in the frequency dimension path, a graph attention network is used to model the topological relationships between spectral components, forming a spatiotemporally coordinated attention feature map.

[0041] A dynamic weight generation mechanism is established. By analyzing the fault discrimination contribution of each region in the time-frequency feature map, spatial location weights and channel weights are calculated to generate a three-dimensional attention weight matrix related to the saliency of fault features.

[0042] The preset attention-guided feature reconstruction module performs element-wise weighted fusion of the generated attention weight matrix and the original feature map, and reconstructs a deep feature representation with enhanced fault features through cross-channel information integration and spatial feature enhancement.

[0043] Specifically, the dynamic adaptation reasoning mechanism for operating conditions includes:

[0044] A multi-source operating condition parameter fusion mechanism is constructed, and an association mapping between operating condition parameters and acoustic features is established through a multi-modal data fusion network to dynamically adjust the confidence threshold of the recognition model.

[0045] Establish a working condition-feature mapping relationship library, construct acoustic feature benchmark models under different working conditions based on historical operating data, and retrieve the corresponding feature benchmarks through real-time working condition parameters to dynamically adapt the recognition standards.

[0046] An online feedback calibration mechanism is constructed to dynamically correct the mapping relationship between operating parameters and identification thresholds by comparing and analyzing continuously collected operating data with identification results.

[0047] Specifically, the hierarchical linkage alarm system uses the following method:

[0048] A multi-level alarm triggering model is constructed. Based on the anomaly identification confidence, the duration of the anomaly feature and the feature intensity, a three-level discrimination standard is established. The alarm threshold is dynamically adjusted in combination with the current operating condition parameters to form an adaptive alarm triggering mechanism.

[0049] A multi-dimensional linkage response strategy is established, and the corresponding response process is initiated according to the alarm level. The alarm level is automatically matched with a preset handling plan, and data is exchanged with the production execution system through the Internet of Things protocol.

[0050] By recording the processing and results of alarm events, an alarm effectiveness evaluation index system is established. Based on historical data, alarm thresholds and response strategies are continuously optimized to form an alarm system with continuous optimization capabilities.

[0051] The beneficial effects of this invention are as follows:

[0052] (1) By integrating the characteristics of human auditory perception with deep learning technology, an acoustic feature representation method conforming to human auditory experience was constructed. This method innovatively introduces an auditory perception weighting mechanism, using a frequency weighting function simulating the equal loudness curve of the human ear to process the time-frequency spectrum, and strengthening the expressive ability of abnormal features through a hierarchical weighting strategy. This feature representation method retains the physical characteristics of the sound signal while incorporating subjective perceptual factors of the human ear, enabling more accurate capture of abnormal sound features identifiable by the human ear. Compared with traditional spectrum analysis methods, this method shows certain advantages in processing transient abnormal sounds that are sensitive to the human ear, improving the accuracy and reliability of motor noise identification, and providing a new technical approach for motor condition monitoring in complex industrial environments.

[0053] (2) A dynamic adaptive processing mechanism was adopted to establish an intelligent analysis method that can automatically adjust according to the motor's operating conditions. By monitoring the motor's speed, load, and other operating parameters in real time, a mapping relationship between operating parameters and acoustic features was constructed, realizing the adaptive adjustment of analysis parameters. In the signal preprocessing stage, the frame parameters were dynamically adjusted based on the real-time speed; in the feature extraction stage, the analysis window parameters were adaptively adjusted based on the non-stationary characteristics of the signal; and in the identification and decision-making stage, the confidence threshold was dynamically adjusted in conjunction with the operating conditions. This comprehensive adaptive mechanism can effectively cope with complex operating conditions such as speed fluctuations and load changes, maintain stable detection performance in strong noise environments, reduce false alarm rates and false negative rates, and provide reliable protection for motor operating status monitoring.

[0054] (3) A complete intelligent monitoring scheme from signal acquisition to alarm response has been established, realizing full-process automation of motor status monitoring. The scheme covers multiple links such as signal acquisition, feature extraction, status recognition, alarm response, and data analysis, forming a complete monitoring closed loop. Multi-directional array acquisition ensures signal integrity, cross-dimensional fusion and attention mechanisms are used to deeply mine abnormal features, accurate early warning is achieved based on multi-level alarm models, and fault tracing and model iteration are supported by full-link data recording. This integrated design can not only monitor motor status in real time and discover potential faults in a timely manner, but also continuously improve recognition performance through continuous optimization, providing comprehensive technical support for predictive maintenance and full life cycle management of equipment. Attached Figure Description

[0055] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0056] Figure 1 This is a flowchart illustrating an intelligent identification method for abnormal motor noise based on time-frequency characteristics and auditory adaptation according to the present invention.

[0057] Figure 2 This is an architecture diagram of an intelligent identification method for abnormal motor noise based on time-frequency characteristics and auditory adaptation, as described in this invention. Detailed Implementation

[0058] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided.

[0059] Please see Figure 1-2 A method for intelligent identification of abnormal motor noise based on time-frequency characteristics and auditory adaptation, comprising:

[0060] S1: Based on a multi-directional acoustic signal acquisition array, the motor operation acoustic signal is acquired in real time; a bandpass filter passband is set based on the frequency band associated with the rated speed of the motor to filter out environmental noise and motor background noise; the acoustic signal is normalized, and the duration and interval of short time frames are dynamically set based on the real-time speed of the motor to divide the continuous acoustic signal into multiple short time frames.

[0061] S2: Short-time sliding window time-frequency conversion is adopted, and the window overlap rate is dynamically adjusted based on the amplitude fluctuation of the time domain of the acoustic signal frame to generate the time spectrum; simultaneously, auditory perception adaptation dynamic weighting technology is adopted to establish a normal operating acoustic baseline, and frequency band weighting coefficients are generated by the deviation between the real-time signal and the baseline. Layered weighting processing is performed on the time spectrum to generate fused time-frequency weighted features that enhance abnormal features.

[0062] S3: Construct a time-frequency weighted feature-enhanced residual network model, and perform multi-scale feature association on the fused time-frequency weighted features through a cross-dimensional fusion mechanism to strengthen the coupling features of abnormal sound signals in the time-frequency domain; introduce a fault feature transfer learning mechanism to dynamically transfer fault discrimination knowledge in the pre-trained model; embed a fault feature attention mechanism to focus on the time-frequency region related to abnormal noise and suppress the interference of normal operation sound features;

[0063] S4: Employ a dynamic adaptation reasoning mechanism based on operating conditions, and adjust the identification confidence threshold by combining the dynamic parameters of time-frequency conversion and weighting coefficients; establish a hierarchical linkage alarm system, and trigger corresponding levels of sound and light, system work orders and equipment linkage responses based on abnormal noise characteristics; execute full-link data closed-loop recording, synchronously store the original sound signal, fuse time-frequency weighted features and abnormal noise identification results, and conduct fault source analysis and model iteration optimization.

[0064] Specifically, the multi-directional acoustic signal acquisition array adopts an acoustic radiation characteristic adaptation method, divides the acquisition area along the motor housing according to the vibration mode distribution law, forms a network of acquisition points distributed in the axial and circumferential directions of the housing, dynamically determines the layout spacing based on the acoustic signal propagation characteristics during motor operation, densifies the layout in the near sound source area to capture detailed signals, and sparsely lays the layout in the far sound source area to cover and diffuse signals, and performs time alignment of the acoustic signals through a synchronous triggering mechanism to form a multi-dimensional acoustic signal acquisition matrix.

[0065] Specifically, the method for filtering out environmental noise and motor background noise is as follows:

[0066] By using bandpass filtering to retain the acoustic signal within the frequency band associated with the rated operation of the motor, an adaptive noise cancellation mechanism based on the acoustic pattern of normal motor operation is adopted. The continuous and stable background noise component within the frequency band is extracted as a reference signal. The filter coefficient is dynamically adjusted by the adaptive filter to eliminate noise components that match the characteristics of the reference signal, while retaining the transient acoustic signal components within the frequency band. The filtered signal is then smoothed to eliminate edge fluctuations introduced by the filtering.

[0067] Specifically, the segmented continuous acoustic signal is divided into multiple short time frames, and the specific method is as follows:

[0068] Based on the real-time speed signal of the motor, the frame length parameter is dynamically calculated through the speed-frame length mapping function to establish a nonlinear inverse proportional relationship between speed and frame length.

[0069] Simultaneously, the non-stationarity index is calculated based on the signal time-domain amplitude change rate, and the frame overlap rate is adaptively adjusted through the overlap rate adjustment function.

[0070] A frame parameter coordination mechanism is established, which uses an association function to constrain the frame length parameter and the overlap rate parameter, thereby forming a coordinated adjustment of the two-dimensional parameters.

[0071] Specifically, the short-time sliding window time-frequency conversion includes:

[0072] The analysis window length is dynamically adjusted based on the signal nonstationarity evaluation function. A nonlinear mapping relationship between the window length parameter and the signal nonstationarity is established. The window length is adaptively adjusted by continuously monitoring the changing trend of the signal's time-domain characteristics.

[0073] A dynamic adjustment mechanism for overlap rate is established based on the temporal amplitude fluctuation of the acoustic signal frame. The window overlap ratio is determined by analyzing the gradient of signal amplitude change, and a functional correspondence between overlap rate and amplitude fluctuation is constructed.

[0074] Multi-dimensional feature extraction is performed on the spectrum transformation results, and the spectrum amplitude is dynamically adjusted through a nonlinear compression function to form a multi-dimensional time-frequency feature expression that includes spectrum amplitude and dynamic change features.

[0075] In this embodiment, the short-time sliding window time-frequency conversion is used to extract the time-frequency features of the motor running sound signal in the motor running sound signal processing.

[0076] First, the real-time input motor sound signal frames are processed. By calculating the instantaneous rate of change and local variance of the time-domain amplitude within the frame, a signal non-stationarity evaluation function F is constructed. ns This quantifies the non-stationarity of the current signal—a larger function value indicates a more drastic change in signal characteristics, potentially corresponding to minor fluctuations in motor operating status or precursors to abnormal noise. Subsequently, window lengths L and F are established. nsThe nonlinear mapping model L=f(F) ns When F ns When the speed increases (e.g., a sudden change in motor speed or the appearance of slight abnormal noise), L adaptively shortens to accurately capture transient characteristics in the short time domain; when F... ns When the speed is reduced (e.g., when the motor is running at a stable, constant speed), L is appropriately lengthened to improve the spectral resolution and preserve the detailed features of the stable signal. Simultaneously, a continuous monitoring mechanism is implemented, requiring F to be recalculated after each frame of signal processing. ns The window length parameter is updated in real time to ensure that the analysis window always matches the non-stationary characteristics of the current signal.

[0077] Next, a dynamic adjustment of the overlap rate is performed to address signal feature breakage or redundancy issues. The signal processing module divides each frame of motor sound signal into sliding sub-windows. By statistically analyzing the maximum change gradient G of the amplitude within each sub-window, the module reflects the severity of local changes in the signal within the frame—a larger G indicates a faster transition of local signal features, potentially containing transient information related to abnormal sounds. Based on this, a functional relationship R=g(G) is constructed between the overlap rate R and G: when G exceeds a preset threshold G1, R is adjusted to a higher range to ensure that adjacent analysis windows can fully cover transient features and avoid the loss of key information; when G is below the threshold G2, R is adjusted to a lower range to reduce data redundancy and computational load. Furthermore, the frame parameter coordination module associates the window length L with the overlap rate R. For example, when L shortens, the upward adjustment of R is synchronously constrained to prevent excessive overlap from causing a surge in computational load, achieving coordinated adaptation between the two.

[0078] Finally, multi-dimensional feature extraction and dynamic range adjustment are completed. After performing a spectral transform on the windowed signal, the signal processing module extracts multi-dimensional features from the transform result: on the one hand, it extracts static features such as peak spectral amplitude, intra-frame spectral mean, and spectral variance to characterize the spectral distribution characteristics of the current frame; on the other hand, it calculates dynamic features such as the difference in spectral amplitude between adjacent frames and the spectral centroid shift to reflect the changing trend of time-frequency features. To address the problem of large spectral amplitude spans and the easy masking of weak features, a nonlinear compression function Y=k×log(X+c) (where k is the compression coefficient and c is the shift) is introduced to dynamically adjust the spectral amplitude—suppressing high-amplitude feature interference corresponding to strong background noise while amplifying weak-amplitude features related to abnormal noise. Finally, the static and dynamic features are integrated into a multi-dimensional time-frequency feature vector, which is output to the subsequent time-frequency weighted feature enhancement residual network, providing core feature support for the accurate identification of motor abnormal noise.

[0079] Specifically, the method for establishing a normal operating auditory baseline is as follows:

[0080] Collect normal operating sound samples of motors under standard operating conditions, obtain the reference spectrum characteristics of each operating condition through time-frequency analysis, and construct a multi-operating-condition normal operating sound feature library.

[0081] An auditory perception weighting coefficient matrix is ​​established, and frequency band weights are assigned to steady-state spectral features based on the characteristics of human auditory perception. By analyzing the discriminative contribution of different frequency bands in abnormal noise identification, a frequency band weighting function is formed, and a weighted normal operation sound auditory feature template is generated.

[0082] A preset time-varying adaptive calibration mechanism is used to continuously collect normal sound signal samples and update the auditory feature template in real time using a sliding window statistical method combined with a time-varying attenuation factor.

[0083] In this embodiment, based on the intelligent identification of abnormal motor noise, a dynamic normal operating sound baseline that conforms to human auditory perception is established and maintained.

[0084] First, a multi-condition normal operation sound feature library is constructed. After the motor is installed, debugged, and confirmed to be in good condition, a large number of normal operation sound samples are continuously collected under various standard operating conditions (such as no-load, rated load, and different stable speed ranges) using the multi-directional sound signal acquisition array. After preprocessing and dynamically framing each sample signal, a high-resolution short-time Fourier transform is performed to extract the complex spectrum of each frame and calculate its amplitude spectrum. Subsequently, for each standard operating condition, the amplitude spectra of all corresponding samples are statistically averaged (e.g., the median or mean is calculated), and their statistical fluctuation range (e.g., percentiles) is recorded, thereby forming the baseline spectral features under the stated operating condition. ,in Representative operating condition index. The baseline spectral characteristics and statistical information of all operating conditions are integrated to construct a multi-dimensional normal operating sound feature library, which serves as the "sound fingerprint" for initial cognition.

[0085] Secondly, an auditory perception-weighted normal operating sound feature template is generated. The core of this step is mapping the physical spectrum to a feature space that conforms to the subjective hearing of the human ear. An auditory perception weighting coefficient matrix is ​​then established. The matrix is ​​generated based on the equal loudness curve characteristics of the human ear, assigning higher weights to the mid-frequency band (the area sensitive to human hearing) and lower weights to the extreme low-frequency and high-frequency bands (the area insensitive to human hearing). Simultaneously, it analyzes historical abnormal sample data (if available) or prior knowledge to calculate the discrimination contribution index of each frequency band in distinguishing normal and abnormal sounds. High-discrimination frequency bands are further enhanced. Ultimately, the frequency band weighting function is determined by... = Determined comprehensively. This weighted matrix... The reference spectral features obtained in the first stage were applied. Above, perform frequency-by-frequency multiplication: = This generates This is the weighted auditory characteristic template of normal operation, which amplifies the frequency components that are easily detected by the human ear and sensitive to faults.

[0086] A time-varying adaptive calibration mechanism is implemented to ensure the timeliness of the baseline. During long-term operation, the motor itself or its environment may undergo slow changes (such as slight wear or changes in ambient temperature). Therefore, a continuous learning process is pre-defined: real-time acoustic signal samples identified as normal are included in a sliding window sample pool of fixed time length. Periodically (e.g., after processing N normal samples), new samples within the window are used to recalculate the spectral statistical characteristics under the current operating conditions. This updates the baseline characteristics. At this time, a time-varying decay factor α (0 < α < 1) is introduced, and an exponentially weighted moving average strategy is adopted: =α× +(1-α)× This update method gives newer data greater impact while retaining some historical information to prevent baseline abrupt changes. Subsequently, the updated data is used... Reapply auditory weighting and refresh the auditory feature template. This closed-loop process enables the "normal operating sound baseline" to adapt to the slow drift of the appliance's state, always maintaining the accuracy of its definition of the current "normal" state, thereby effectively reducing the false alarm rate.

[0087] Specifically, the method for performing hierarchical weighting processing on the time spectrum is as follows:

[0088] A hierarchical auditory perception weighted network is constructed, dividing time-spectrum processing into a basic auditory weighting layer, an anomaly feature enhancement layer, and a dynamic range adaptation layer. For the basic auditory weighting layer, a frequency weighting function simulating the auditory perception characteristics of the human ear is applied to perform preliminary weighting processing on the time-spectrum. For the anomaly feature enhancement layer, based on the spectral distribution characteristics of historical anomaly samples, the discrimination index of each frequency band in anomaly identification is calculated, and an anomaly sensitivity weighting matrix is ​​constructed. For the dynamic range adaptation layer, based on the amplitude distribution characteristics of the time-spectrum, a nonlinear compression function is used to dynamically adjust the frequency bands of different intensities.

[0089] Specifically, the cross-dimensional fusion mechanism includes:

[0090] A feature association analysis network is constructed. By calculating the covariance matrix of the time dimension feature sequence and the frequency dimension feature sequence, a cross-dimensional feature association map is established. A graph convolutional network is used to perform deep feature extraction on the feature association map to obtain cross-dimensional coupled feature representation.

[0091] A two-way feature interaction architecture is established, and the two-way interaction and information fusion of time-dimensional and frequency-dimensional features are realized through gated loop units;

[0092] A preset dynamic weight allocation mechanism is used. Based on the coupled feature representation output by the feature association analysis network, dynamic fusion weights related to the features are generated through a multilayer perceptron network. The weight coefficients are probabilistically processed using a normalized exponential function to achieve adaptive fusion of cross-dimensional features.

[0093] Specifically, the fault feature transfer learning mechanism includes:

[0094] A cross-domain feature mapping network is constructed. By establishing a feature alignment layer between the source domain pre-trained model and the target domain motor noise identification task, a domain adaptation transformation of deep features is performed, mapping the general acoustic feature space to the motor noise-specific feature space.

[0095] Establish a knowledge distillation transmission link, extract multi-scale feature response maps from the pre-trained model as teacher signals, and drive the target network to learn fault feature representations through an attention-guided feature matching loss function;

[0096] A preset dynamic fine-tuning strategy is used to adaptively adjust the learning rate of the network layers based on the distribution characteristics of the target domain samples, retaining general knowledge of the source domain while performing targeted optimization of the abnormal noise characteristics of the motor.

[0097] This embodiment aims to illustrate how to effectively adapt a model pre-trained on a general acoustic dataset to the specific task of motor noise recognition through transfer learning, thereby solving the problem of insufficient labeled samples in the target domain and improving model performance.

[0098] First, a cross-domain feature mapping network is constructed to achieve feature space transformation. Specifically, a deep neural network pre-trained on a large-scale general audio dataset is chosen as the source model. While this model possesses powerful general acoustic feature extraction capabilities, its feature space exhibits a distributional difference from the specific features of motor noise. Therefore, a trainable feature alignment layer is introduced at the end of the feature extractor of the pre-trained model. This layer typically consists of fully connected layers or adaptive batch normalization layers, and its core function is to learn a non-linear mapping function that projects the general high-level features extracted by the pre-trained model onto the specific feature space of motor noise. During training, by minimizing the difference in feature distribution between the source domain (general audio) and the target domain (motor sound), for example using metrics such as maximum mean difference, the feature alignment layer learns to filter out general acoustic features unrelated to motor faults, while retaining and strengthening key feature representations related to faults such as bearing wear and electromagnetic noise.

[0099] Secondly, a knowledge distillation transfer link is established to drive deep learning in the target network. In this step, we fix the parameters of the pre-trained source model, using it as the "teacher model." The target domain motor noise recognition network to be trained serves as the "student model." The knowledge transfer process is not simply imitating the final output of the teacher model, but rather delving into its internal structure. We extract multi-scale feature response maps from different depths of the teacher model; these feature maps contain the model's hierarchical understanding of sound signals from details to semantics. Subsequently, we design an attention-guided feature matching loss function, which calculates the difference between the corresponding layer feature maps of the student model and the teacher model feature maps, and uses the saliency information of the teacher model feature maps to generate attention weights, enabling the student model to focus more on learning feature regions that contribute significantly to fault detection. Through this deep, guided feature imitation, the student model can inherit the powerful feature representation capabilities of the teacher model and quickly focus on the discriminative patterns of motor noise.

[0100] Finally, a dynamic fine-tuning strategy was employed to optimize the network in a targeted manner. After initial knowledge distillation, we used all motor acoustic samples to fine-tune the entire target network (including the feature alignment layer and subsequent classification layer) end-to-end. To avoid catastrophic forgetting and balance general knowledge with task-specific knowledge, we preset a dynamic learning rate strategy: a lower learning rate was assigned to the lower-level, general feature extraction layers to retain their universal representations learned from massive amounts of data; a higher learning rate was assigned to the higher-level networks, especially the newly introduced feature alignment and classification layers, to incentivize them to make rapid and sufficient adaptive adjustments based on the distribution of motor noise data. This hierarchical, dynamic fine-tuning process ensured that the model achieved accurate targeted optimization of motor noise features based on a solid foundation of general acoustic knowledge.

[0101] Specifically, the embedded fault feature attention mechanism includes:

[0102] A dual-path attention architecture is constructed. In the time dimension path, a multi-head self-attention mechanism is used to capture the temporal dependencies of abnormal sounds, and in the frequency dimension path, a graph attention network is used to model the topological relationships between spectral components, forming a spatiotemporally coordinated attention feature map.

[0103] A dynamic weight generation mechanism is established. By analyzing the fault discrimination contribution of each region in the time-frequency feature map, spatial location weights and channel weights are calculated to generate a three-dimensional attention weight matrix related to the saliency of fault features.

[0104] The preset attention-guided feature reconstruction module performs element-wise weighted fusion of the generated attention weight matrix and the original feature map, and reconstructs a deep feature representation with enhanced fault features through cross-channel information integration and spatial feature enhancement.

[0105] Specifically, the dynamic adaptation reasoning mechanism for operating conditions includes:

[0106] A multi-source operating condition parameter fusion mechanism is constructed, and an association mapping between operating condition parameters and acoustic features is established through a multi-modal data fusion network to dynamically adjust the confidence threshold of the recognition model.

[0107] Establish a working condition-feature mapping relationship library, construct acoustic feature benchmark models under different working conditions based on historical operating data, and retrieve the corresponding feature benchmarks through real-time working condition parameters to dynamically adapt the recognition standards.

[0108] An online feedback calibration mechanism is constructed to dynamically correct the mapping relationship between operating parameters and identification thresholds by comparing and analyzing continuously collected operating data with identification results.

[0109] Specifically, the hierarchical linkage alarm system uses the following method:

[0110] A multi-level alarm triggering model is constructed. Based on the anomaly identification confidence, the duration of the anomaly feature and the feature intensity, a three-level discrimination standard is established. The alarm threshold is dynamically adjusted in combination with the current operating condition parameters to form an adaptive alarm triggering mechanism.

[0111] A multi-dimensional linkage response strategy is established, and the corresponding response process is initiated according to the alarm level. The alarm level is automatically matched with a preset handling plan, and data is exchanged with the production execution system through the Internet of Things protocol.

[0112] By recording the processing and results of alarm events, an alarm effectiveness evaluation index system is established. Based on historical data, alarm thresholds and response strategies are continuously optimized to form an alarm system with continuous optimization capabilities.

[0113] This embodiment describes the specific implementation process of alarm signal generation, response, and optimization closed loop in the intelligent identification method for abnormal motor noise.

[0114] First, an adaptive multi-level alarm triggering model is constructed. This model uses anomaly identification confidence, anomaly characteristic duration, and characteristic strength as core criteria to classify alarm levels into three levels: Attention, Warning, and Danger. For example, when the system detects an anomaly characteristic with moderate confidence but short duration, it may only trigger an "Attention" level alarm; while when an anomaly pattern with extremely high confidence, high characteristic strength, and persistent presence is identified, it will immediately trigger a "Danger" level alarm. Crucially, these alarm thresholds are not fixed but dynamically adjusted based on the motor's current real-time operating parameters (such as load and speed). The system's built-in operating condition-threshold mapping table ensures that under conditions prone to interference and noise, such as heavy load or high speed, the alarm threshold is appropriately increased, thereby effectively suppressing false alarms and achieving intelligent and contextualized alarm triggering mechanisms.

[0115] Secondly, the system executes a precise multi-dimensional linkage response strategy. Once the alarm level is determined, the system automatically activates the pre-set handling plan for that level. For example, a "Caution" level alarm may only provide a visual prompt on the monitoring interface and log the information; a "Warning" level alarm will activate an audible and visual alarm and automatically generate an inspection work order in the production execution system, notifying maintenance personnel to intervene; and the highest level "Danger" alarm, in addition to the above responses, will also send instructions to the control system via the Internet of Things protocol to trigger the equipment's load reduction operation or safe shutdown logic to prevent the situation from escalating. This tiered response ensures a strict match between the handling measures and the fault risk, avoiding overreaction and ensuring equipment safety.

[0116] Finally, the system forms a data-driven closed-loop optimization mechanism. The processing, results, and subsequent equipment diagnostic reports for each alarm event are fully recorded. Based on this historical data, the system establishes an alarm effectiveness evaluation index system, such as false alarm rate, missed alarm rate, and average response time. Through periodic data analysis, the system can automatically evaluate the effectiveness of existing alarm thresholds and response strategies, and fine-tune and optimize them using machine learning algorithms. This enables the entire alarm system to continuously learn and evolve, constantly improving itself over long-term operation, becoming increasingly accurate and efficient.

[0117] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A motor abnormal sound intelligent identification method based on time-frequency features and auditory adaptation, characterized in that, The method comprises the following steps: S1: Real-time acquisition of motor operation sound signals based on a multi-directional sound signal acquisition array; setting a band-pass filter passband based on the motor rated speed correlation frequency band to filter out environmental noise and motor background noise; Normalizing the sound signals, and dynamically setting the length and interval of the short-term frame based on the real-time speed of the motor to divide the continuous sound signals into multiple short-term frames; S2: Using short-time sliding window time-frequency conversion, dynamically adjusting the window overlap rate based on the amplitude fluctuation of the sound signal frame time domain to generate a time-frequency spectrum; simultaneously using an auditory perception adaptive dynamic weighting technology to establish a normal operation sound auditory baseline, generating a frequency band weighting coefficient through the deviation degree of the real-time signal and the baseline, and performing hierarchical weighting processing on the time-frequency spectrum to generate a fusion time-frequency weighted feature that enhances abnormal characteristics; The short-time sliding window time-frequency conversion comprises: Dynamically adjusting the analysis window length based on the signal non-stationarity evaluation function, establishing a nonlinear mapping relationship between the window length parameter and the signal non-stationarity, and adaptively adjusting the window length by continuously monitoring the change trend of the signal time domain characteristics; Establishing a dynamic adjustment mechanism for the overlap rate based on the amplitude fluctuation of the sound signal frame time domain, determining the window overlap ratio by analyzing the signal amplitude change gradient, and constructing a function corresponding relationship between the overlap ratio and the amplitude fluctuation amplitude; Extracting multi-dimensional features from the spectrum transformation result, and dynamically adjusting the spectrum amplitude through a nonlinear compression function to form a multi-dimensional time-frequency feature expression containing the spectrum amplitude and dynamic change characteristics; The method for establishing a normal operation sound auditory baseline comprises: Collecting normal operation sound samples of the motor under standard working conditions, obtaining reference spectrum characteristics under each working condition through time-frequency analysis, and constructing a multi-condition normal operation sound feature library; Establishing an auditory perception weighting coefficient matrix, distributing frequency band weights to the steady-state spectrum characteristics based on human ear auditory perception characteristics, forming a frequency band weight distribution function by analyzing the discriminant contribution of different frequency bands in abnormal sound identification, and generating a weighted normal operation sound auditory feature template; Pre-setting a time-varying adaptive calibration mechanism, continuously collecting normal sound signal samples, and using a sliding window statistical method combined with a time-varying attenuation factor to update the auditory feature template in real time; The hierarchical weighting processing of the time-frequency spectrum comprises: Constructing a hierarchical auditory perception weighting network to divide the time-frequency spectrum processing into a basic auditory weighting layer, an abnormal characteristic enhancement layer, and a dynamic range adaptation layer; for the basic auditory weighting layer, applying a frequency weighting function simulating human ear auditory perception characteristics to perform preliminary weighting processing on the time-frequency spectrum; for the abnormal characteristic enhancement layer, calculating the discriminant index of each frequency band in abnormal sound identification based on the spectrum distribution characteristics of historical abnormal samples to construct an abnormal sensitivity weighting matrix; for the dynamic range adaptation layer, using a nonlinear compression function to dynamically adjust the dynamic range of different intensity frequency bands based on the amplitude distribution characteristics of the time-frequency spectrum; S3: Construct a time-frequency weighted feature enhanced residual network model, and perform multi-scale feature correlation on the fused time-frequency weighted features through a cross-dimension fusion mechanism to strengthen the coupling features of abnormal sound signals in the time-frequency domain; introduce a fault feature transfer learning mechanism to dynamically transfer fault discrimination knowledge in a pre-trained model; embed a fault feature attention mechanism to focus on time-frequency regions related to abnormal sounds and suppress the interference of normal operation sound features; S4: Adopt a working condition dynamic adaptation reasoning mechanism, adjust the recognition confidence threshold by combining the time-frequency conversion dynamic parameters and the weighting coefficient; establish a hierarchical and linked alarm system, trigger corresponding levels of sound and light, system work orders and equipment linked responses based on abnormal sound features; perform full-link data closed-loop recording, synchronously store original sound signals, fused time-frequency weighted features and abnormal sound recognition results, and perform fault tracing analysis and model iterative optimization.

2. The method of claim 1, wherein, The multi-directional sound signal collection array adopts an acoustic radiation characteristic adaptation method, divides the collection area according to the vibration mode distribution law along the motor shell, forms a cross-distributed collection point network in the axial and circumferential directions of the shell, dynamically determines the layout spacing based on the sound signal propagation characteristics during motor operation, densely arranges preset capture detail signals in the near sound source area, sparsely arranges preset coverage diffusion signals in the far sound source area, and performs time alignment of sound signals through a synchronous triggering mechanism to form a multi-dimensional sound signal collection matrix.

3. The method of claim 1, wherein, The method for filtering out environmental noise and motor background noise is as follows: The sound signals in the frequency band associated with the rated operation of the motor are retained through band-pass filtering, an adaptive noise cancellation mechanism based on the motor normal operation soundprint is used, the continuous and stable background noise components in the frequency band are extracted as reference signals, the filter coefficients are dynamically adjusted through an adaptive filter to eliminate noise components matching the reference signal features, while retaining the transient sound signal components in the frequency band; the filtered signal is subjected to smoothing processing to eliminate the edge fluctuations introduced by filtering.

4. The method of claim 1, wherein, The method for dividing the continuous sound signal into multiple short-period frames is as follows: Based on the real-time motor speed signal, the frame length parameter is dynamically calculated through a speed-frame length mapping function, and a nonlinear inverse relationship between the speed and the frame length is established; At the same time, a non-stationarity index is calculated based on the signal time-domain amplitude change rate, and the frame overlap rate is adaptively adjusted through an overlap rate adjustment function; A frame parameter coordination mechanism is established, the frame length parameter and the overlap rate parameter are mutually constrained through an association function, and the coordinated adjustment of the two-dimensional parameters is formed.

5. The method of claim 1, wherein, The cross-dimension fusion mechanism includes: A feature correlation analysis network is constructed, a cross-dimension feature correlation graph is established by calculating the covariance matrix of the time dimension feature sequence and the frequency dimension feature sequence, a graph convolution network is used to perform deep feature extraction on the feature correlation graph to obtain a cross-dimension coupling feature representation; A bidirectional feature interaction architecture is established, a gated recurrent unit is used to realize bidirectional interaction and information fusion of the time dimension and the frequency dimension features; A preset dynamic weight distribution mechanism is established, based on the coupling feature representation output by the feature correlation analysis network, a multi-layer perception network is used to generate dynamic fusion weights related to the features, a normalization exponential function is used to probabilistic process the weight coefficients, and adaptive fusion of the cross-dimension features is performed.

6. The method of claim 1, wherein, The fault feature migration learning mechanism comprises: A cross-domain feature mapping network is constructed, a feature alignment layer is established between a pre-trained model in a source domain and a motor abnormal sound identification task in a target domain, domain adaptation conversion of deep features is performed, and a general acoustic feature space is mapped to a motor abnormal sound specific feature space; A knowledge distillation transmission link is established, multi-scale feature response graphs are extracted from the pre-trained model as teacher signals, an attention-guided feature matching loss function is used to drive the target network to learn fault feature representation; A preset dynamic fine-tuning strategy is used to adaptively adjust the learning rate of the network layer based on the distribution characteristics of the target domain samples, to retain the general knowledge of the source domain and to perform directional optimization on the motor abnormal sound features.

7. The method of claim 1, wherein, The embedded fault feature attention mechanism comprises: A double-path attention architecture is constructed, a multi-head self-attention mechanism is used in the time dimension path to capture the time sequence dependence of abnormal sound, a graph attention network is used in the frequency dimension path to model the topological association between spectral components, and a spatiotemporal collaborative attention feature map is formed; A dynamic weight generation mechanism is established, the fault discrimination contribution of each region in the time-frequency feature map is analyzed, spatial position weights and channel weights are calculated, and a three-dimensional attention weight matrix related to the fault feature saliency is generated; A preset attention-guided feature reconstruction module is used to perform element-wise weighted fusion of the generated attention weight matrix and the original feature map, to integrate cross-channel information and enhance spatial features, and to reconstruct deep feature representation with fault feature enhancement.

8. The method of claim 1, wherein, The working condition dynamic adaptation reasoning mechanism comprises: A multi-source working condition parameter fusion mechanism is constructed, a working condition parameter and acoustic feature association mapping is established through a multi-modal data fusion network, and the confidence threshold of the identification model is dynamically adjusted; A working condition-feature mapping relationship database is established, acoustic feature benchmark models under different working condition conditions are constructed based on historical operation data, corresponding feature benchmarks are retrieved through real-time working condition parameters, and the identification standard is dynamically adapted; An online feedback calibration mechanism is constructed, the mapping relationship between the working condition parameters and the identification threshold is dynamically corrected through comparative analysis of the continuously collected operation data and the identification results.

9. The method of claim 1, wherein, The hierarchical and linked alarm system specifically comprises: A multi-level alarm triggering model is constructed, three-level discrimination standards are established based on abnormal identification confidence, abnormal feature duration and feature intensity, alarm threshold is dynamically adjusted according to current working condition parameters, and an adaptive alarm triggering mechanism is formed; A multi-dimensional linked response strategy is established, corresponding response processes are started according to the alarm level, the alarm level is automatically matched with the preset disposal scheme, and data interaction is performed with the production execution system through the Internet of Things protocol; An alarm effect evaluation index system is established by recording the processing process and results of the alarm event, the alarm threshold and the response strategy are continuously optimized based on historical data, and an alarm system with continuous optimization capability is formed.

Images