An audio signal anomaly monitoring method based on data analysis

Abstract

The present application relates to the field of signal anomaly detection, and more particularly to an audio signal anomaly monitoring method based on data analysis, which comprises: collecting and preprocessing original audio signals to obtain an audio signal sequence and a candidate window length set; performing local signal-to-noise ratio and feature consistency fusion analysis on the audio signal sequence to obtain an initial time-frequency credibility weight; performing time-frequency structure optimization analysis on the initial time-frequency credibility weight to obtain an optimized time-frequency credibility weight; performing fusion calculation on the optimized time-frequency credibility weight and time-frequency spectrum energy distribution to obtain a weighted information entropy; and performing candidate window length evaluation and anomaly determination on the weighted information entropy to obtain an audio signal anomaly monitoring result, thereby solving the problem of decreased anomaly monitoring accuracy caused by information entropy evaluation distortion of traditional adaptive short-time Fourier transform window selection technology in a strong noise environment.

Description

Technical Field

[0001] This invention relates to the field of signal anomaly detection technology, and in particular to a method for monitoring audio signal anomalies based on data analysis. Background Technology

[0002] In the fields of industrial predictive maintenance, equipment health monitoring, security monitoring, and intelligent operation and maintenance, audio signal-based anomaly monitoring methods have become an important means of state perception due to their advantages such as non-contact operation, flexible deployment, rich information content, and sensitivity to early, minor faults. For pumps, fans, motors, compressors, bearing assemblies, and other continuously operating equipment, sound signals containing information such as mechanical friction, impact vibration, airflow disturbance, and structural resonance are continuously generated during operation. By analyzing these sound signals, abnormal changes in equipment operating status can be identified earlier, providing a basis for fault warning, maintenance decisions, and operational safety assurance. Therefore, how to accurately extract abnormal features from complex audio signals and achieve reliable monitoring has become an important issue of concern in related technical fields. In existing technologies, the analysis of non-stationary audio signals typically uses short-time Fourier transform to convert one-dimensional time-domain signals into two-dimensional time-frequency representations, so as to simultaneously observe the signal's variation characteristics in both the time and frequency dimensions. Since the analytical performance of the Short-Time Fourier Transform (SFT) is closely related to the choice of window length, shorter analysis windows are beneficial for improving temporal resolution and are more suitable for characterizing anomalous features such as transient impulses and short-term abrupt changes; while longer analysis windows are beneficial for improving frequency resolution and are more suitable for describing frequency domain features such as stable harmonics and continuous frequency band structures. Therefore, to meet the detection needs of different types of anomalous signals, existing technologies have gradually developed adaptive window analysis methods based on candidate window length traversal and evaluation criteria selection. This involves performing time-frequency analysis for different candidate window lengths, evaluating the analytical performance of each candidate window length using information entropy, energy concentration index, or other statistical indicators, and then selecting the optimal window length as the basis for the analysis of the current audio signal segment.

[0003] The aforementioned adaptive short-time Fourier transform window selection method typically achieves good analysis results in conventional environments with high signal-to-noise ratios and weak background interference. Especially in scenarios where the equipment operating environment is relatively stable, external noise is low, and abnormal characteristics are obvious, time-frequency clustering evaluation indicators such as information entropy can effectively reflect the concentration of the time spectrum under different candidate window lengths, allowing for the selection of a more suitable analysis window length for the current signal state, thereby improving the accuracy of subsequent feature extraction and anomaly detection. Therefore, this type of method has been widely applied in general industrial equipment condition monitoring, basic acoustic alarms, and routine fault identification scenarios. However, in actual industrial settings, the audio monitoring environment is often complex, especially in pump rooms, compressor rooms, production workshops, fan nacelles, and areas with concentrated deployment of large electromechanical equipment. These environments often simultaneously contain steady-state rumbling, broadband random noise, mechanical impact noise, pipeline fluid disturbance noise, and superimposed noise generated by the coupled operation of multiple devices. This strong background noise significantly raises the noise floor in the time spectrum, alters the local time-frequency energy distribution, and submerges previously weak abnormal impacts, local harmonic shifts, or short-term abrupt changes within the noise. In this context, traditional evaluation criteria, such as information entropy calculated directly from the original time-frequency energy distribution, are easily contaminated by noise, resulting in inaccurate reflections of the aggregation state of effective anomaly features in the time-frequency domain. When the evaluation criteria themselves are affected by strong noise, existing adaptive window selection mechanisms are prone to producing incorrect window length selections. For example, in high-noise scenarios, to reduce the randomness of the overall time-frequency spectrum, traditional methods may tend to choose a longer analysis window to average and smooth the noise. However, this may weaken or even obscure the weak impact characteristics, local narrowband anomalies, and transient structural changes corresponding to early equipment failures. On the other hand, choosing a shorter analysis window may lead to an overly dispersed frequency domain structure, making it difficult to clearly reveal stable anomaly components. Thus, while existing technologies possess a certain degree of adaptive analysis capability, their core evaluation criteria lack sufficient robustness in high-noise environments, making it difficult to balance noise suppression with the preservation of true anomaly features. This can easily lead to problems such as decreased monitoring sensitivity, increased false alarm rates, and increased missed alarm rates.

[0004] Therefore, how to address the problem of distorted evaluation criteria in traditional adaptive short-time Fourier transform window selection techniques under strong noise environments, and how to construct an audio signal anomaly monitoring method that can more accurately characterize the degree of aggregation of effective anomaly features and has stronger robustness to noise interference, has become an urgent technical problem to be solved in this field. Summary of the Invention

[0005] In view of this, the present invention aims to propose an audio signal anomaly monitoring method based on data analysis, so as to solve the problem of decreased anomaly monitoring accuracy caused by information entropy evaluation distortion in traditional adaptive short-time Fourier transform window selection technology under strong noise environment.

[0006] To achieve the above objectives, the technical solution of the present invention is implemented as follows: A method for detecting audio signal anomalies based on data analysis, the method comprising: Step S1: Obtain the audio signal sequence and candidate window length set by acquiring and preprocessing the original audio signal; Step S2: Obtain the initial time-frequency confidence weight by performing local signal-to-noise ratio and feature consistency fusion analysis on the audio signal sequence; Step S3: Obtain optimized time-frequency confidence weights by performing time-frequency structure optimization analysis on the initial time-frequency confidence weights; Step S4: Obtain the weighted information entropy by fusing the optimized time-frequency confidence weights with the time-spectrum energy distribution; Step S5: Obtain audio signal anomaly monitoring results by evaluating candidate window lengths and determining anomalies in the weighted information entropy.

[0007] Furthermore, the step of acquiring and preprocessing the original audio signal to obtain the audio signal sequence and candidate window length set includes: Audio acquisition nodes are deployed in the area where the monitoring target is located, and acoustic sensors for acquiring raw audio signals are installed on the audio acquisition nodes. The sampling frequency and quantization bit depth are set, and the sound signal during the operation of the monitoring target is continuously or triggered according to the sampling frequency and quantization bit depth to obtain the raw audio signal. The raw audio signal is filtered to remove low-frequency interference components and high-frequency noise components unrelated to the monitoring target to obtain a preprocessed audio signal. The preprocessed audio signal is amplitude normalized to obtain an audio signal sequence. According to the sampling frequency corresponding to the raw audio signal and the target frequency range, a candidate window length set consisting of multiple candidate analysis window lengths is set, and the window function type and window overlap rate are set.

[0008] Furthermore, the step of obtaining the initial time-frequency confidence weight by performing local signal-to-noise ratio and feature consistency fusion analysis on the audio signal sequence includes: The local signal-to-noise ratio coefficients are obtained by performing time-frequency transformation and local signal-to-noise ratio analysis on the audio signal sequence. The initial time-frequency confidence weights are obtained by performing feature consistency analysis and fusion processing on the audio signal sequence.

[0009] Furthermore, the step of obtaining local signal-to-noise ratio coefficients by performing time-frequency transformation and local signal-to-noise ratio analysis on the audio signal sequence includes: For any target candidate window length in the candidate window length set, extract the target candidate window length from the candidate window length set, and perform frame segmentation on the audio signal sequence based on the target candidate window length, window function type, and window overlap rate; perform short-time Fourier transform on the framed audio signal sequence to obtain the time-spectrum data corresponding to the target candidate window length; For any target time-frequency point in the time-spectrum data, local energy statistics are performed based on the time-spectrum amplitude data of the local neighborhood where the target time-frequency point is located to obtain the local signal power estimate corresponding to the target time-frequency point; based on the background noise tracking results of the local neighborhood where the target time-frequency point is located, the local background noise power estimate corresponding to the target time-frequency point is obtained. The ratio of the estimated local signal power and the estimated local background noise power at the target time and frequency point is calculated by adding them to a preset minimum constant. The obtained ratio is then logarithmically transformed and normalized to obtain the local signal-to-noise ratio coefficient at the target time and frequency point.

[0010] Furthermore, the step of obtaining the initial time-frequency confidence weight by performing feature consistency analysis and fusion processing on the audio signal sequence includes: For any target candidate window length in the candidate window length set, extract the time-frequency amplitude data of the local neighborhood where any target time-frequency point is located from the time-frequency data corresponding to the target candidate window length. Calculate the mean of the time-frequency amplitude data of the local neighborhood to obtain the local amplitude mean corresponding to the target time-frequency point. Calculate the variance of the time-frequency amplitude data of the local neighborhood to obtain the local amplitude variance corresponding to the target time-frequency point. The local amplitude variance corresponding to the target time-frequency point is used as the numerator, and the result of adding the local amplitude mean to the preset minimum constant is used as the denominator. The resulting fraction is used as the amplitude fluctuation evaluation corresponding to the target time-frequency point. The difference between the constant 1 and the amplitude fluctuation evaluation is used as the initial consistency evaluation corresponding to the target time-frequency point. When the initial consistency evaluation is less than the constant 0, the characteristic consistency coefficient corresponding to the target time-frequency point is set to the constant 0. When the initial consistency evaluation is greater than or equal to the constant 0, the initial consistency evaluation is used as the characteristic consistency coefficient corresponding to the target time-frequency point. The local signal-to-noise ratio coefficient and feature consistency coefficient corresponding to the target time-frequency point are weighted and summed to obtain the initial time-frequency confidence weight corresponding to the target time-frequency point.

[0011] Furthermore, the step of obtaining optimized time-frequency confidence weights by performing time-frequency structure optimization analysis on the initial time-frequency confidence weights includes: Directional response data is obtained by performing multi-directional structural response analysis on time-spectrum data; Structural similarity data is obtained by evaluating the structural similarity of the directional response data. The optimized time-frequency confidence weights are obtained by fusing and optimizing the initial time-frequency confidence weights with the structural similarity data.

[0012] Furthermore, the step of obtaining directional response data by performing multi-directional structural response analysis on the time-spectrum data includes: For any target candidate window length in the candidate window length set, extract the time-spectrum amplitude data from the time-spectrum data corresponding to the target candidate window length; set multiple structural response directions, and set corresponding directional filters for each structural response direction; For any target time-frequency point in the time-frequency data, based on the local time-frequency amplitude data at the location of the target time-frequency point, convolution operations are performed using the directional filters corresponding to each structural response direction to obtain the directional response values ​​of the target time-frequency point under each structural response direction. The set of directional response values ​​corresponding to the target time-frequency point under all structural response directions is taken as the directional response data corresponding to the target time-frequency point.

[0013] Furthermore, the step of obtaining structural similarity data by performing structural similarity assessment on the directional response data includes: For any target time-frequency point, extract the directional response values ​​corresponding to all structural response directions from the directional response data corresponding to the target time-frequency point, compare the magnitudes of all directional response values, and obtain the maximum directional response value corresponding to the target time-frequency point. For any target time-frequency point, the sum of all directional response values ​​corresponding to the target time-frequency point is divided by the number of structural response directions to obtain the average directional response value corresponding to the target time-frequency point. The maximum directional response value corresponding to the target time-frequency point is used as the numerator, and the larger of the average directional response value and the preset minimum constant corresponding to the target time-frequency point is used as the denominator. The resulting fraction is used as the directional concentration degree evaluation corresponding to the target time-frequency point. The constant 1 and the sum of the directional concentration degree evaluation are logarithmically transformed to obtain the structural similarity data corresponding to the target time-frequency point.

[0014] Furthermore, the step of obtaining optimized time-frequency confidence weights by fusing and optimizing the initial time-frequency confidence weights and structural similarity data includes: For any target time-frequency point, set a structural enhancement coefficient, multiply the structural enhancement coefficient by the structural similarity data corresponding to the target time-frequency point, and add the product result to a constant 1 to obtain the structural influence term corresponding to the target time-frequency point; The initial time-frequency confidence weight corresponding to the target time-frequency point is multiplicatively fused with the structural influence term to obtain the optimized time-frequency confidence weight corresponding to the target time-frequency point.

[0015] Furthermore, the step of obtaining the weighted information entropy by fusing the optimized time-frequency confidence weights and the time-spectrum energy distribution includes: For any target candidate window length in the candidate window length set, extract the time spectrum amplitude data corresponding to each target time frequency point from the time spectrum data corresponding to the target candidate window length, and take the square of the time spectrum amplitude data corresponding to each target time frequency point as the time spectrum energy data corresponding to each target time frequency point. For any target time-frequency point, the result of multiplying the optimized time-frequency confidence weight corresponding to the target time-frequency point with the time-spectrum energy data corresponding to the target time-frequency point is used as the weighted energy evaluation corresponding to the target time-frequency point; the result of adding the weighted energy evaluations of all target time-frequency points corresponding to the target candidate window length is used as the weighted energy sum corresponding to the target candidate window length. For any target time-frequency point, the weighted energy assessment corresponding to the target time-frequency point is used as the numerator, the sum of the weighted energy and the result of adding the preset minimum constant are used as the denominator, and the corresponding fraction is used as the weighted energy probability distribution data corresponding to the target time-frequency point. For any target time-frequency point, the product of the weighted energy probability distribution data and the corresponding logarithm of the weighted energy probability distribution data is used as the entropy component evaluation of the target time-frequency point; the negative number of the sum of the entropy component evaluations of all target time-frequency points is used to obtain the weighted information entropy corresponding to the target candidate window length.

[0016] Compared with the prior art, the present invention has the following advantages: This invention discloses an audio signal anomaly monitoring method based on data analysis. By quantifying the reliability of local signals in the time-frequency spectrum and jointly incorporating energy reliability and structural continuity characteristics into the window evaluation process, anomaly monitoring no longer relies solely on the original energy distribution susceptible to noise disturbance. Instead, it can more accurately focus on the time-frequency region where the true anomaly features are located. In complex acoustic environments such as pump rooms, compressor rooms, fan nacelles, and trackside equipment rooms, where steady-state rumbling, mechanical friction noise, and random impact interference are present for extended periods, this technique can effectively suppress the masking effect of high-energy background noise on anomaly feature extraction. It avoids the difficulty in identifying weak fault impacts, early wear harmonics, or short-term abnormal pulses due to being submerged in the noise substrate, thereby improving the detection capability of early anomalies under low signal-to-noise ratio conditions. Furthermore, this invention reconstructs the traditional information entropy evaluation method with reliability weighting, making the selection criteria of the adaptive analysis window more consistent with the aggregation state of the true signal features, rather than being dominated by random falsehoods caused by noise. Therefore, it can significantly reduce the time-frequency ambiguity problem caused by window misselection and improve the stability and robustness of anomaly monitoring results. In industrial predictive maintenance and equipment health monitoring scenarios, this technology helps reduce false alarms and missed alarms, enhances the applicability of the system in complex operating conditions, continuously operating equipment, and highly disturbed environments, makes the monitoring results more credible in engineering, and provides more reliable technical support for early equipment fault warning, operational status assessment, and maintenance decisions. Attached Figure Description

[0017] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a flowchart of an audio signal anomaly monitoring method based on data analysis, as described in an embodiment of the present invention. Detailed Implementation

[0018] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0019] See Figure 1 This is a flowchart of an audio signal anomaly monitoring method based on data analysis provided in Embodiment 1 of the present invention. Figure 1 As shown, a data analysis-based method for detecting audio signal anomalies may include: Step S1: Acquire and preprocess the original audio signal to obtain the audio signal sequence and candidate window length set.

[0020] First, audio acquisition nodes are deployed in the area where the monitoring target is located, and acoustic sensors for acquiring raw audio signals are installed on the audio acquisition nodes. The sampling frequency and quantization bit depth are set, and the sound signals during the operation of the monitoring target are continuously or triggered according to the sampling frequency and quantization bit depth to obtain the raw audio signal. In this embodiment, the sampling frequency is set to 48kHz and the quantization bit depth to 16 bits for continuous acquisition. The raw audio signal is filtered to remove low-frequency interference components and high-frequency noise components unrelated to the monitoring target, obtaining a pre-processed audio signal. The pre-processed audio signal is then amplitude normalized to obtain an audio signal sequence. Based on the sampling frequency corresponding to the raw audio signal and the target frequency range, a candidate window length set consisting of multiple candidate analysis window lengths is set, and the window function type and window overlap rate are also set.

[0021] It should be noted that, for the candidate window length set, it is preferable to set a set consisting of multiple discrete window lengths. In this embodiment of the invention, the candidate window length set is set to (256, 512, 1024, 2048) (unit: number of sampling points). For the window function type, a Hamming window or a Blackman window can be used; in this embodiment of the invention, a Hamming window is selected as the window function. For the bed overlap rate, it is set to 75% in this embodiment of the invention.

[0022] This completes the process of acquiring and preprocessing the original audio signal to obtain the audio signal sequence and the candidate window length set.

[0023] Step S2: Obtain the initial time-frequency confidence weight by performing local signal-to-noise ratio and feature consistency fusion analysis on the audio signal sequence.

[0024] In noisy environments, traditional audio anomaly monitoring methods based on adaptive short-time Fourier transform window selection typically use time-frequency clustering indices such as information entropy as the evaluation criteria for candidate window lengths. However, when steady-state booming, broadband random noise, or sudden impact interference exists at the monitoring site, noise energy significantly raises the time-frequency spectrum floor and alters the time-frequency energy distribution, making it difficult for traditional optimization criteria to accurately reflect the actual degree of clustering of anomaly features in the time-frequency domain. This leads to distortion in the candidate window length evaluation results, affecting the accuracy and stability of subsequent anomaly monitoring. Therefore, this step first analyzes the reliability of each time-frequency point in the time-frequency spectrum. By jointly evaluating the local signal-to-noise ratio and feature consistency of the time-frequency spectrum, an initial time-frequency reliability weight is constructed to quantify the degree of noise interference at each time-frequency point and the probability that it belongs to a real signal component. This approach prevents different regions in the time-frequency spectrum from being treated equally, instead providing differentiated reliability characterizations based on local energy significance and neighborhood variation patterns, thus laying the foundation for subsequent noise interference suppression and anomaly feature highlighting. Furthermore, while the initial time-frequency confidence weights based solely on local signal-to-noise ratio and feature consistency can distinguish between high-confidence signal regions and low-confidence noise regions to some extent, they still suffer from insufficient differentiation capabilities when complex background noise and weak anomaly features are statistically similar, as they are primarily based on local amplitude statistical characteristics. To improve the ability to identify true anomaly features, this step further introduces time-frequency structure optimization analysis. By identifying structural features such as continuous ridges, local directional consistency, and harmonic extension trends in the time spectrum, the initial time-frequency confidence weights are further modified so that they not only reflect the energy confidence level of a time-frequency point but also the rationality of that time-frequency point in terms of physical structure pattern. Based on this, the optimized time-frequency confidence weights are then introduced into the traditional time-frequency information entropy calculation process to reconstruct the time spectrum energy distribution using weighted methods, obtaining weighted information entropy. Unlike traditional information entropy, which directly characterizes the overall energy dispersion, the weighted information entropy obtained in this step focuses more on characterizing the clustering state of effective signal features within high-confidence regions. Therefore, it can serve as a more robust candidate window length evaluation criterion against strong noise. When comparing multiple candidate window lengths, the system can calculate the weighted information entropy corresponding to each candidate window length, and determine a better analysis window length based on the magnitude of the weighted information entropy, thereby providing a more robust time-frequency representation basis for subsequent feature extraction and anomaly detection.

[0025] In summary, this invention first obtains the local signal-to-noise ratio (SNR) coefficients by performing time-frequency transformation and local signal-to-noise ratio (SNR) analysis on the audio signal sequence. Specifically, for any target candidate window length in the candidate window length set, the target candidate window length is extracted from the set, and the audio signal sequence is framed based on the target candidate window length, window function type, and window overlap rate. The framed audio signal sequence is then subjected to short-time Fourier transform (SFT) to obtain the time-spectrum data corresponding to the target candidate window length. For any target time-frequency point in the time-spectrum data, local energy statistics are performed based on the time-spectrum amplitude data of the local neighborhood where the target time-frequency point is located to obtain the local signal power estimate corresponding to the target time-frequency point. It should be noted that in this embodiment of the invention, the local neighborhood is defined as centered on the target time-frequency point. Range. Based on the background noise tracking results of the local neighborhood where the target time-frequency point is located, the local background noise power estimate corresponding to the target time-frequency point is obtained. It should be noted that in this embodiment of the invention, the local background noise power estimate is performed by a recursive minimum value tracking method. The local signal power estimate and the local background noise power estimate corresponding to the target time-frequency point are added to a preset minimum constant, and the ratio is calculated. The obtained ratio is then logarithmically transformed and normalized to obtain the local signal-to-noise ratio coefficient corresponding to the target time-frequency point. In this embodiment of the invention, the minimum constant is set to... .

[0026] After obtaining the local signal-to-noise ratio coefficient corresponding to the target time-frequency point, the initial time-frequency confidence weight is obtained by performing feature consistency analysis and fusion processing on the audio signal sequence. Specifically, for any target candidate window length in the candidate window length set, the time-frequency amplitude data of the local neighborhood where any target time-frequency point is located is extracted from the time-frequency data corresponding to the target candidate window length. The mean of the time-frequency amplitude data of the local neighborhood is calculated to obtain the local amplitude mean corresponding to the target time-frequency point. The variance of the time-frequency amplitude data of the local neighborhood is calculated to obtain the local amplitude variance corresponding to the target time-frequency point. The local amplitude variance corresponding to the target time-frequency point is used as the numerator, and the result of adding the local amplitude mean to a preset minimum constant is used as the denominator. The resulting fraction is used as the amplitude fluctuation evaluation corresponding to the target time-frequency point. The difference between the constant 1 and the amplitude fluctuation evaluation is used as the initial consistency evaluation corresponding to the target time-frequency point. When the initial consistency evaluation is less than the constant 0, the characteristic consistency coefficient corresponding to the target time-frequency point is set to the constant 0. When the initial consistency evaluation is greater than or equal to the constant 0, the initial consistency evaluation is used as the characteristic consistency coefficient corresponding to the target time-frequency point. The local signal-to-noise ratio coefficient and the characteristic consistency coefficient corresponding to the target time-frequency point are weighted and summed to obtain the initial time-frequency confidence weight corresponding to the target time-frequency point.

[0027] In one implementation, assume time-frequency points The corresponding local signal power estimate is Time and frequency points The corresponding local background noise power estimate is The minimum constant is Time and frequency points The corresponding local amplitude variance is Time and frequency points The corresponding local amplitude mean is The weighting of the local signal-to-noise ratio coefficients is: The weighted average of the feature consistency coefficients is: Then the time frequency point The corresponding expression for calculating the initial time-frequency confidence weight is:

[0028] in, Representing time and frequency points The corresponding initial time-frequency confidence weights; The weights represent the local signal-to-noise ratio coefficients. The weighted weights representing the feature consistency coefficients are set as follows in this embodiment of the invention: ; Representing time and frequency points Corresponding local signal power estimation; Representing time and frequency points Corresponding local background noise power estimation; This represents a minimal constant. ; Representing time and frequency points The corresponding local amplitude variance; Representing time and frequency points The corresponding local amplitude mean; This indicates normalization processing, which in this embodiment of the invention performs maximum and minimum value normalization processing using the corresponding values ​​of all time and frequency points; This represents the ReLU activation function.

[0029] It should be noted that, This is the local signal-to-noise ratio term, where the ratio is the linear signal-to-noise ratio. If the ratio is close to 1, it indicates that the energy at that point is comparable to the noise floor, and it is very likely noise; if the ratio is much greater than 1, it indicates that there is significant energy exceeding the noise, and it may contain a signal. The reason for using logarithms is that in practice, the signal-to-noise ratio may span several orders of magnitude, and logarithmic transformation makes its changes smoother. This is to normalize the signal-to-noise ratio to [0,1]. This is the characteristic consistency term, where the ratio represents the amplitude fluctuation, reflected by the ratio of variance to mean. It is expressed through the constant 1 and... Subtraction transforms this into consistency. The smaller the fluctuation, the closer the value is to 1, indicating high consistency; the larger the fluctuation, the smaller the value is, or even negative, indicating low consistency. Finally, the lower bound is truncated using an activation function to keep its value range in [0,1].

[0030] Thus, the initial time-frequency confidence weights were obtained by performing local signal-to-noise ratio and feature consistency fusion analysis on the audio signal sequence.

[0031] Step S3: Obtain optimized time-frequency confidence weights by performing time-frequency structure optimization analysis on the initial time-frequency confidence weights.

[0032] Although the initial confidence weight obtained in step S2 can integrate energy and infrastructure information, it still relies on the characteristic consistency calculation of local statistics and is difficult to accurately distinguish points with similar statistical characteristics but different physical patterns, such as high-energy burst noise and weak but continuous real fault harmonics.

[0033] Therefore, this step introduces a deeper time-frequency structure analysis to identify inherent structural features (the continuity of time-frequency ridges) in the signal that conform to specific physical fault modes. By detecting these more discriminative patterns, the initial weights can be targeted for enhancement or suppression.

[0034] In summary, this invention first obtains directional response data by performing multi-directional structural response analysis on time-spectrum data. Specifically, for any target candidate window length in the candidate window length set, time-spectrum amplitude data is extracted from the time-spectrum data corresponding to the target candidate window length. Multiple structural response directions are defined, and corresponding directional filters are set for each structural response direction. In this embodiment of the invention, four structural response directions are defined, corresponding to… The preferred directional filter is a Gabor filter. For any target time-frequency point in the time-spectrum data, based on the local time-spectrum amplitude data at the location of the target time-frequency point, convolution operations are performed using the directional filters corresponding to each structural response direction to obtain the directional response values ​​of the target time-frequency point under each structural response direction. The set of directional response values ​​corresponding to the target time-frequency point under all structural response directions is taken as the directional response data corresponding to the target time-frequency point.

[0035] After obtaining the directional response data, the structural similarity assessment is performed to obtain structural similarity data. Specifically, for any target time-frequency point, the directional response values ​​corresponding to all structural response directions are extracted from the directional response data corresponding to the target time-frequency point. The magnitudes of all directional response values ​​are compared to obtain the maximum directional response value corresponding to the target time-frequency point. For any target time-frequency point, the sum of all directional response values ​​corresponding to the target time-frequency point is divided by the number of structural response directions to obtain the average directional response value corresponding to the target time-frequency point. The maximum directional response value corresponding to the target time-frequency point is used as the numerator, and the larger of the average directional response value and a preset minimum constant is used as the denominator. The resulting fraction is used as the directional concentration assessment for the target time-frequency point. A logarithmic transformation is performed on the constant 1 and the sum of the directional concentration assessments to obtain the structural similarity data corresponding to the target time-frequency point.

[0036] After obtaining the structural similarity data, the optimized time-frequency confidence weight is obtained by fusing and optimizing the initial time-frequency confidence weight with the structural similarity data. Specifically, for any target time-frequency point, a structural enhancement coefficient is set. In this embodiment, the structural enhancement coefficient is set to 0.5. The structural enhancement coefficient is multiplied by the structural similarity data corresponding to the target time-frequency point, and the product is added to a constant 1 to obtain the structural influence term corresponding to the target time-frequency point. The initial time-frequency confidence weight and the structural influence term corresponding to the target time-frequency point are then multiplicatively fused to obtain the optimized time-frequency confidence weight corresponding to the target time-frequency point.

[0037] In one embodiment, the structural reinforcement coefficient is assumed to be Time and frequency points In the The directional response value corresponding to each direction is Time and frequency points The maximum response amplitude in all directions is Then the time frequency point The corresponding expression for calculating the optimized time-frequency confidence weight is:

[0038] in, Representing time and frequency points The corresponding optimized time-frequency confidence weights; Representing time and frequency points The corresponding initial time-frequency confidence weights; This represents the structural enhancement coefficient. The larger the value of the structural enhancement coefficient, the greater the influence of the structural features on the credibility weight. Representing time and frequency points The maximum value of the response amplitude in all directions; Representing time and frequency points In the The directional response values ​​corresponding to each direction; This represents a minimal constant. ; Represents the absolute value function; This represents the maximum value function.

[0039] It should be noted that, This is a term representing the degree of structural similarity, where the denominator is the mean of the response amplitudes in all directions, and the numerator is the response in the direction of maximum magnitude. For a point on an ideal continuous ridge, its energy will be highly concentrated in a specific direction, resulting in a ratio much greater than 1. For an isolated noisy point or an isotropic region, its responses in all directions are similar, and the ratio is close to 1. This is the structural influence term. Taking the logarithm and adding 1 ensures that when the ratio is close to 1 (no significant structure), the optimization term is approximately equal to (1 + 0.35). The initial weighting of a point with significant continuity is only slightly increased; however, when the ratio is large (structurally significant), ln will give a significant increment. Multiplying by the strength parameter gives the degree of structural influence, and adding 1 gives the structural influence term on the initial confidence level. This means that for points with significant continuity, their initial weights will be significantly enhanced by a factor greater than 1. For isolated points without obvious structural features, their weights are only slightly adjusted or remain almost unchanged.

[0040] Thus, the optimized time-frequency confidence weights were obtained by performing time-frequency structure optimization analysis on the initial time-frequency confidence weights.

[0041] Step S4: Obtain the weighted information entropy by fusing the optimized time-frequency confidence weights and time-spectrum energy distribution.

[0042] Traditional information entropy is calculated directly based on the time-spectrum energy distribution. However, this distribution is severely distorted under strong noise, rendering the entropy value invalid. To address this issue, this step uses optimized confidence weights to perform weighted correction on the original time-spectrum energy, redefining a purified probability distribution, and calculating the final weighted information entropy based on this distribution. For a given candidate window length, the smaller the calculated weighted information entropy value, the more concentrated and prominent the true signal features are in the time-spectrum obtained using that window length, indicating better performance of that window length under the current noise environment.

[0043] Specifically, for any target candidate window length in the candidate window length set, the time-frequency amplitude data corresponding to each target time-frequency point is extracted from the time-frequency data corresponding to the target candidate window length, and the square of the time-frequency amplitude data corresponding to each target time-frequency point is taken as the time-frequency energy data corresponding to each target time-frequency point. For any target time-frequency point, the result of multiplying the optimized time-frequency confidence weight corresponding to the target time-frequency point by the time-frequency energy data corresponding to the target time-frequency point is taken as the weighted energy evaluation corresponding to the target time-frequency point; the result of summing the weighted energy evaluations of all target time-frequency points corresponding to the target candidate window length is taken as the weighted energy sum corresponding to the target candidate window length. For any target time-frequency point, the weighted energy evaluation corresponding to the target time-frequency point is taken as the numerator, and the result of adding the weighted energy sum to the preset minimum constant is taken as the denominator. The resulting fraction is taken as the weighted energy probability distribution data corresponding to the target time-frequency point. For any target time-frequency point, the product of the weighted energy probability distribution data and the corresponding logarithm of the weighted energy probability distribution data is used as the entropy component evaluation of the target time-frequency point; the negative number of the sum of the entropy component evaluations of all target time-frequency points is used to obtain the weighted information entropy corresponding to the target candidate window length.

[0044] Thus, the weighted information entropy is obtained by fusing the optimized time-frequency confidence weights with the time-spectrum energy distribution.

[0045] Step S5: By evaluating the candidate window length and determining anomalies in the weighted information entropy, the audio signal anomaly monitoring results are obtained.

[0046] After completing step S4, the corresponding weighted information entropy has been obtained for each candidate window length in the candidate window length set. Since the weighted information entropy is the evaluation result obtained after weighted reconstruction of the time-frequency energy distribution with the participation of optimized time-frequency confidence weights, its value can more realistically reflect the degree of aggregation of effective abnormal features in the time-frequency domain under the current candidate window length. Generally speaking, when the weighted information entropy corresponding to a certain candidate window length is small, it indicates that in the time-frequency representation obtained under that candidate window length, the energy distribution in the high-confidence region is more concentrated, the local structure corresponding to the real abnormal signal is more prominent, and the noise component has a relatively weak disturbance to the overall time-frequency representation. Therefore, this candidate window length is more suitable as the analysis window length of the current audio signal segment. Based on this, this step first compares the weighted information entropy corresponding to each candidate window length in the candidate window length set and selects the target candidate window length with the smallest weighted information entropy as the optimal analysis window length of the current audio signal segment.

[0047] In specific implementation, for any candidate window length in the candidate window length set, the processing flow from steps S2 to S4 is called sequentially to obtain the weighted information entropy corresponding to that candidate window length. The weighted information entropies corresponding to all candidate window lengths are sorted or a minimum value search is performed to determine the target candidate window length with the smallest weighted information entropy, and this target candidate window length is taken as the optimal analysis window length. Preferably, the candidate window length set can be set to {256, 512, 1024, 2048} sampling points. Of course, in practical applications, it can also be adjusted according to the operating characteristics of the monitored object, the target abnormal frequency band range, and computing resources. This approach avoids the problem of traditional methods erroneously favoring excessively long or short window lengths due to entropy distortion in strong noise environments, making the determination of the optimal analysis window length more consistent with the time-frequency distribution characteristics of the real abnormal signal.

[0048] After determining the optimal analysis window length, a short-time Fourier transform is performed on the current audio signal sequence using the optimal analysis window length, a preset window function type, and a window overlap rate to obtain the corresponding optimal time spectrum. Compared to time spectra obtained under other candidate window lengths, the optimal time spectrum achieves a more suitable balance between time domain resolution and frequency domain resolution for the current signal state, making anomalous features such as weak abnormal impulses, harmonic disturbances, intermittent abnormal pulses, or local frequency band energy rises clearer in the time and frequency domains. To further achieve automated anomaly detection, anomaly detection feature data reflecting the current audio state can be extracted based on the optimal time spectrum. The anomaly detection feature data can be statistical features, spectral morphology features, or compressed characterization features in the time spectrum. Preferably, it can include one or more of the following: Mel-frequency cepstral coefficients, spectral centroid, spectral band energy distribution, spectral flatness, and spectral roll-off points.

[0049] After acquiring the anomaly detection feature data, the anomaly detection feature data is input into a pre-constructed normal state reference model for deviation evaluation. The normal state reference model is used to characterize the audio feature distribution of the monitored target under normal operating conditions, and can be trained by collecting historical audio signals of the monitored target under normal operating conditions and extracting corresponding features. Preferably, the normal state reference model can be any one of a single-class support vector machine model, a Gaussian mixture model, an isolated forest model, or an autoencoder reconstruction model. When using a single-class support vector machine model or a Gaussian mixture model, the distance, probability, or score between the current anomaly detection feature data and the normal sample feature distribution can be used as the deviation evaluation result; when using an autoencoder reconstruction model, the reconstruction error of the current anomaly detection feature data can be used as the deviation evaluation result. Through the above processing, the anomaly deviation data corresponding to the current audio signal segment can be obtained.

[0050] Furthermore, the abnormal deviation data is compared with a preset abnormality judgment threshold to output the final audio signal abnormality monitoring result. When the abnormal deviation data is greater than the preset abnormality judgment threshold, the current audio signal segment is determined to be abnormal, an abnormality alarm signal is output, and the corresponding time position, optimal analysis window length, weighted information entropy, and abnormal deviation are recorded. When the abnormal deviation data is less than or equal to the preset abnormality judgment threshold, the current audio signal segment is determined to be in a normal state, and a normal monitoring result is output. Preferably, the abnormality judgment threshold can be set according to the statistical distribution of the normal state reference model on the verification sample, for example, set to the mean of normal deviation plus 2, 2.5, or 3 standard deviations, more preferably set to the mean of normal deviation plus 3 standard deviations, so as to achieve a better balance between false alarm rate and false negative rate.

[0051] Through the above processing, this step adaptively optimizes the candidate window length using weighted information entropy as the core evaluation criterion, and on this basis, completes the extraction of abnormal features and the determination of anomalies, thereby ultimately obtaining the audio signal anomaly monitoring results. Compared with traditional methods that directly analyze based on fixed window lengths or entropy criteria affected by noise, this step can provide a clearer and more stable time-frequency characterization basis for subsequent anomaly judgment in complex noise environments, thereby improving the accuracy, robustness, and detection capability of early weak anomalies in audio anomaly monitoring.

[0052] Thus, the process of evaluating candidate window lengths and determining anomalies in weighted information entropy to obtain audio signal anomaly monitoring results has been completed.

[0053] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for detecting audio signal anomalies based on data analysis, characterized in that, The method includes: Step S1: Obtain the audio signal sequence and candidate window length set by acquiring and preprocessing the original audio signal; Step S2: Obtain the initial time-frequency confidence weight by performing local signal-to-noise ratio and feature consistency fusion analysis on the audio signal sequence; Step S3: Obtain optimized time-frequency confidence weights by performing time-frequency structure optimization analysis on the initial time-frequency confidence weights; Step S4: Obtain the weighted information entropy by fusing the optimized time-frequency confidence weights with the time-spectrum energy distribution; Step S5: Obtain audio signal anomaly monitoring results by evaluating candidate window lengths and determining anomalies in the weighted information entropy.

2. The audio signal anomaly monitoring method based on data analysis according to claim 1, characterized in that... By acquiring and preprocessing the raw audio signal, an audio signal sequence and a set of candidate window lengths are obtained, including: Audio acquisition nodes are deployed in the area where the monitoring target is located, and acoustic sensors for acquiring raw audio signals are installed on the audio acquisition nodes. The sampling frequency and quantization bit depth are set, and the sound signal during the operation of the monitoring target is continuously or triggered according to the sampling frequency and quantization bit depth to obtain the raw audio signal. The raw audio signal is filtered to remove low-frequency interference components and high-frequency noise components unrelated to the monitoring target to obtain a preprocessed audio signal. The preprocessed audio signal is amplitude normalized to obtain an audio signal sequence. According to the sampling frequency corresponding to the raw audio signal and the target frequency range, a candidate window length set consisting of multiple candidate analysis window lengths is set, and the window function type and window overlap rate are set.

3. The audio signal anomaly monitoring method based on data analysis according to claim 1, characterized in that, The initial time-frequency confidence weight is obtained by performing local signal-to-noise ratio and feature consistency fusion analysis on the audio signal sequence, including: The local signal-to-noise ratio coefficients are obtained by performing time-frequency transformation and local signal-to-noise ratio analysis on the audio signal sequence. The initial time-frequency confidence weights are obtained by performing feature consistency analysis and fusion processing on the audio signal sequence.

4. The audio signal anomaly monitoring method based on data analysis according to claim 3, characterized in that, The process of obtaining local signal-to-noise ratio coefficients by performing time-frequency transformation and local signal-to-noise ratio analysis on the audio signal sequence includes: For any target candidate window length in the candidate window length set, extract the target candidate window length from the candidate window length set, and perform frame segmentation on the audio signal sequence based on the target candidate window length, window function type, and window overlap rate; perform short-time Fourier transform on the framed audio signal sequence to obtain the time-spectrum data corresponding to the target candidate window length; For any target time-frequency point in the time-spectrum data, local energy statistics are performed based on the time-spectrum amplitude data of the local neighborhood where the target time-frequency point is located to obtain the local signal power estimate corresponding to the target time-frequency point; based on the background noise tracking results of the local neighborhood where the target time-frequency point is located, the local background noise power estimate corresponding to the target time-frequency point is obtained. The ratio of the estimated local signal power and the estimated local background noise power at the target time and frequency point is calculated by adding them to a preset minimum constant. The obtained ratio is then logarithmically transformed and normalized to obtain the local signal-to-noise ratio coefficient at the target time and frequency point.

5. The audio signal anomaly monitoring method based on data analysis according to claim 3, characterized in that, The process of obtaining initial time-frequency confidence weights by performing feature consistency analysis and fusion processing on the audio signal sequence includes: For any target candidate window length in the candidate window length set, extract the time-frequency amplitude data of the local neighborhood where any target time-frequency point is located from the time-frequency data corresponding to the target candidate window length. Calculate the mean of the time-frequency amplitude data of the local neighborhood to obtain the local amplitude mean corresponding to the target time-frequency point. Calculate the variance of the time-frequency amplitude data of the local neighborhood to obtain the local amplitude variance corresponding to the target time-frequency point. The local amplitude variance corresponding to the target time-frequency point is used as the numerator, and the result of adding the local amplitude mean to the preset minimum constant is used as the denominator. The resulting fraction is used as the amplitude fluctuation evaluation corresponding to the target time-frequency point. The difference between the constant 1 and the amplitude fluctuation evaluation is used as the initial consistency evaluation corresponding to the target time-frequency point. When the initial consistency evaluation is less than the constant 0, the characteristic consistency coefficient corresponding to the target time-frequency point is set to the constant 0. When the initial consistency evaluation is greater than or equal to the constant 0, the initial consistency evaluation is used as the characteristic consistency coefficient corresponding to the target time-frequency point. The local signal-to-noise ratio coefficient and feature consistency coefficient corresponding to the target time-frequency point are weighted and summed to obtain the initial time-frequency confidence weight corresponding to the target time-frequency point.

6. The audio signal anomaly monitoring method based on data analysis according to claim 1, characterized in that, The step of obtaining optimized time-frequency confidence weights by performing time-frequency structure optimization analysis on the initial time-frequency confidence weights includes: Directional response data is obtained by performing multi-directional structural response analysis on time-spectrum data; Structural similarity data is obtained by evaluating the structural similarity of the directional response data. The optimized time-frequency confidence weights are obtained by fusing and optimizing the initial time-frequency confidence weights with the structural similarity data.

7. The audio signal anomaly monitoring method based on data analysis according to claim 6, characterized in that, The process of obtaining directional response data by performing multi-directional structural response analysis on time-spectrum data includes: For any target candidate window length in the candidate window length set, extract the time-spectrum amplitude data from the time-spectrum data corresponding to the target candidate window length; set multiple structural response directions, and set corresponding directional filters for each structural response direction; For any target time-frequency point in the time-frequency data, based on the local time-frequency amplitude data at the location of the target time-frequency point, convolution operations are performed using the directional filters corresponding to each structural response direction to obtain the directional response values ​​of the target time-frequency point under each structural response direction. The set of directional response values ​​corresponding to the target time-frequency point under all structural response directions is taken as the directional response data corresponding to the target time-frequency point.

8. The audio signal anomaly monitoring method based on data analysis according to claim 6, characterized in that, The process of evaluating the structural similarity of the directional response data to obtain structural similarity data includes: For any target time-frequency point, extract the directional response values ​​corresponding to all structural response directions from the directional response data corresponding to the target time-frequency point, compare the magnitudes of all directional response values, and obtain the maximum directional response value corresponding to the target time-frequency point. For any target time-frequency point, the sum of all directional response values ​​corresponding to the target time-frequency point is divided by the number of structural response directions to obtain the average directional response value corresponding to the target time-frequency point. The maximum directional response value corresponding to the target time-frequency point is used as the numerator, and the larger of the average directional response value and the preset minimum constant corresponding to the target time-frequency point is used as the denominator. The resulting fraction is used as the directional concentration degree evaluation corresponding to the target time-frequency point. The constant 1 and the sum of the directional concentration degree evaluation are logarithmically transformed to obtain the structural similarity data corresponding to the target time-frequency point.

9. The audio signal anomaly monitoring method based on data analysis according to claim 6, characterized in that, The process of fusing and optimizing the initial time-frequency confidence weights and structural similarity data to obtain optimized time-frequency confidence weights includes: For any target time-frequency point, set a structural enhancement coefficient, multiply the structural enhancement coefficient by the structural similarity data corresponding to the target time-frequency point, and add the product result to a constant 1 to obtain the structural influence term corresponding to the target time-frequency point; The initial time-frequency confidence weight corresponding to the target time-frequency point is multiplicatively fused with the structural influence term to obtain the optimized time-frequency confidence weight corresponding to the target time-frequency point.

10. The method for monitoring audio signal anomalies based on data analysis according to claim 1, characterized in that, The step of obtaining weighted information entropy by fusing the optimized time-frequency confidence weights and time-spectrum energy distribution includes: For any target candidate window length in the candidate window length set, extract the time spectrum amplitude data corresponding to each target time frequency point from the time spectrum data corresponding to the target candidate window length, and take the square of the time spectrum amplitude data corresponding to each target time frequency point as the time spectrum energy data corresponding to each target time frequency point. For any target time-frequency point, the result of multiplying the optimized time-frequency confidence weight corresponding to the target time-frequency point with the time-spectrum energy data corresponding to the target time-frequency point is used as the weighted energy evaluation corresponding to the target time-frequency point; the result of adding the weighted energy evaluations of all target time-frequency points corresponding to the target candidate window length is used as the weighted energy sum corresponding to the target candidate window length. For any target time-frequency point, the weighted energy assessment corresponding to the target time-frequency point is used as the numerator, and the result of adding the weighted energy sum to the preset minimum constant is used as the denominator. The resulting fraction is used as the weighted energy probability distribution data corresponding to the target time-frequency point. For any target time-frequency point, the product of the weighted energy probability distribution data and the corresponding logarithm is used as the entropy component assessment corresponding to the target time-frequency point. The negative of the result of adding the entropy component assessments corresponding to all target time-frequency points is used to obtain the weighted information entropy corresponding to the target candidate window length.

Images