Sound anomaly detection method and system based on causal Gaussian process dynamic system

Abstract

The present invention relates to the field of audio detection, and in particular to a method and system for detecting anomalies in sound based on a causal Gaussian process dynamic system, the method comprising the following steps: obtaining an audio feature spectrum diagram of audio data; modeling the audio feature spectrum diagram to obtain low-dimensional latent variables; obtaining an audio causal relationship diagram based on the low-dimensional latent variables; obtaining a universal causal relationship diagram based on a normal audio causal relationship diagram; calculating the mean square error value between the normal audio causal relationship diagram and the universal causal relationship diagram, and using the mean square error value as an empirical threshold; calculating the difference between the audio causal relationship diagram to be tested and the universal causal relationship diagram; comparing the difference with the empirical threshold, and obtaining the detection result of the audio data to be tested based on the comparison result. The present invention introduces a causal discovery method into a Gaussian process dynamic system, solves the problem of difficulty in training anomaly detection of small sample data in the prior art, and provides an interpretability basis.

Description

Technical Field

[0001] The present invention relates to the field of audio detection, and in particular to a method and system for detecting sound anomalies based on a causal Gaussian process dynamic system. Background Art

[0002] As an important branch of anomaly detection technology, audio detection technology has its unique advantages. Audio signals contain rich information. Especially in the industrial context, detecting whether the audio signal is normal can provide early warning signals for workers, so that they can formulate preventive measures or emergency measures in time to minimize risks.

[0003] In the existing audio detection technology, after obtaining the audio feature spectrum graph, the audio detection is mostly realized based on the related methods of deep neural network. The method of the present invention uses the Gaussian process dynamic system to model the feature spectrum graph of audio data, and introduces the causal discovery method into the Gaussian process dynamic system, which not only improves the accuracy of audio data modeling, but also can quantify the uncertainty of prediction, and provides an interpretable basis for anomaly detection, solves the problem of the difficulty of deep neural network training for anomaly detection of small sample data, and provides strong support for practical applications.

[0004] The method of the present invention has fewer model parameters, so it has obvious advantages in the task of detecting anomalies in small sample audio data. The method of the present invention is particularly suitable for small sample audio data anomaly detection scenarios in industrial environments that require fast anomaly resolution and explainable anomaly solutions, and is of great significance to the development of the industrial production field. Summary of the invention

[0005] In view of the defects in the prior art, the present invention provides a sound anomaly detection method and system based on a causal Gaussian process dynamic system.

[0006] In order to achieve the above-mentioned purpose, in the first aspect, the present invention provides a sound anomaly detection method based on a causal Gaussian process dynamic system, the method comprising the following steps: obtaining an audio feature spectrum graph of audio data; modeling the audio feature spectrum graph to obtain a low-dimensional latent variable; obtaining an audio causal relationship graph based on the low-dimensional latent variable; obtaining a universal causal relationship graph based on a normal audio causal relationship graph; calculating the mean square error value between the normal audio causal relationship graph and the universal causal relationship graph, and using the mean square error value as an empirical threshold; calculating the difference between the audio causal relationship graph to be tested and the universal causal relationship graph; comparing the difference with the empirical threshold, and obtaining the detection result of the audio data to be tested according to the comparison result. The present invention uses Gaussian process dynamic system modeling to solve the problem of the difficulty of deep neural network training for small sample data anomaly detection, and at the same time introduces the causal discovery method into the Gaussian process dynamic system, effectively improving the accuracy of modeling, and providing an interpretable basis for audio anomaly detection. The method of the present invention performs better in the case of small sample data in an industrial environment.

[0007] Optionally, the audio feature spectrogram of the audio data is obtained, including: establishing an audio data set; performing a short-time Fourier transform on the audio signal of the audio data set to obtain a first spectrogram; performing a fast Fourier transform on the first spectrogram to obtain a second spectrogram; filtering the second spectrogram with a Mel filter group to obtain a Mel spectrogram; taking the logarithm of the amplitude of the Mel spectrogram to obtain a logarithmic Mel spectrogram. The present invention converts audio data into image data, enhances the expression of detailed features of audio data, improves the efficiency and accuracy of feature extraction of audio data, and effectively improves the operability and analysis accuracy of audio data.

[0008] Optionally, the audio feature spectrum graph is modeled to obtain low-dimensional latent variables, including: using a Gaussian process dynamic system to model the logarithmic Mel spectrum graph to obtain a reference model; based on the reference model, obtaining a mapping relationship between logarithmic Mel spectrum data and low-dimensional latent variables; based on the logarithmic Mel spectrum data, obtaining an initial low-dimensional latent variable; based on the initial low-dimensional latent variable, using the mapping relationship, obtaining logarithmic Mel spectrum prediction data; subtracting the logarithmic Mel spectrum prediction data from the real data to obtain a difference; adjusting the random parameters in the reference model to revise the initial low-dimensional latent variables; based on the revised initial low-dimensional latent variables, using the mapping relationship, revising the logarithmic Mel spectrum prediction data to minimize the difference; adjusting the random parameters to obtain the low-dimensional latent variables while obtaining the optimal solution of the random parameters. The present invention uses a Gaussian process dynamic system to model the audio feature spectrum graph, which not only improves the accuracy of the prediction results, but also can quantify the uncertainty of the prediction, solves the problem of the difficulty of deep neural network training for small sample data anomaly detection, and provides strong support for audio detection in practical applications.

[0009] Optionally, the audio causal relationship graph is obtained based on the low-dimensional latent variable, including: calculating the Pierre correlation coefficient between a certain element and other elements in the low-dimensional latent variable; based on the Pierre correlation coefficient, screening elements with strong correlation with the certain element to form a candidate parent node set of the certain element; combining the certain element with elements in the candidate parent node set to obtain a set of two-tuples; calculating the partial correlation between the certain element and the elements in the candidate parent node set; performing a significance test on the partial correlation to obtain a significance test result; based on the significance test result, obtaining a set of screened candidate parent nodes; judging whether all elements in the screened candidate parent node set are conditionally independent of the certain element; the obtained conditionally related elements form a parent node set of the certain element to obtain an audio causal relationship graph. The present invention introduces the causal discovery method into the Gaussian process dynamic system, effectively improving the accuracy of audio data modeling, and providing an interpretability basis for anomaly detection, making the constructed model more rigorous and reliable, and increasing the credibility of the audio data detection results.

[0010] Optionally, the partial correlation between an element and the elements in the candidate parent node set is calculated using the following calculation formula:

[0011] p=p(x i ,x n |z)=p(Res(x i ),Res(x n ))

[0012] Among them, p is the probability, x i and x nis the candidate parent node set P(x i ), p(x i ,x n |z) is x under z condition i and x n The probability of simultaneous occurrence, z is the candidate parent node set P(x i ) i and x n Element-wise and x i The element with the highest correlation, Res(x) is the first-order partial correlation coefficient of x. In the present invention, when controlling one variable, the calculation formula can accurately reflect whether there is a linear relationship between two variables, making the process steps more rigorous.

[0013] Optionally, obtaining a universal causal relationship graph based on a normal audio causal relationship graph includes: taking normal audio data in the audio data set; obtaining a normal audio causal relationship graph based on the normal audio data; taking the mean of the causal relationship values ​​in the normal audio causal relationship graph; and obtaining a universal causal relationship graph based on the mean. The present invention intuitively displays the mutual relationship and degree of dependence between various factors of the audio signal through a universal causal relationship graph, provides a clear causal relationship path, facilitates understanding of complex causal relationship networks, and helps to formulate effective optimization plans in practical applications.

[0014] Optionally, the mean square error value between the normal audio causal relationship graph and the universal causal relationship graph is calculated, and the mean square error value is used as an empirical threshold, and the calculation formula is as follows:

[0015]

[0016] Among them, ∈ ijτ is the mean square error of the (i, j) node value between the universal causal relationship graph and the normal audio causal relationship graph at time τ, T n is the duration of the nth audio. is the (i, j) node value of the nth normal audio causal relationship graph at time τ, is the (i, j) node value in the universal causal graph at time τ. The present invention uses the mean square error value as an empirical threshold, as an important reference indicator of the audio data feature value, which helps to improve the accuracy and reliability of abnormal detection and further improve the accuracy of the method of the present invention.

[0017] Optionally, the calculating the difference between the audio causal relationship graph to be tested and the universal causal relationship graph comprises: taking the audio data to be tested in the audio data set; obtaining the audio causal relationship graph to be tested based on the audio data to be tested; and calculating the difference between the audio causal relationship graph to be tested and the universal causal relationship graph. The present invention uses the universal causal graph as a reference standard for the audio causal relationship graph to be tested, which makes the steps of the method of the present invention more rigorous and efficient, while simplifying the process steps.

[0018] Optionally, the difference is compared with the empirical threshold, and based on the comparison result, the detection result of the audio data to be tested is obtained, including: comparing the difference with the empirical threshold; screening out the relationship tuples whose difference is greater than the empirical threshold; obtaining an abnormal chain based on the relationship tuple; combining the abnormal chain and the audio causal relationship graph to be tested to obtain all abnormal information of the audio data to be tested. The present invention combines the audio causal relationship graph to be tested with the obtained abnormal chain to present the abnormal information of the audio data to be tested, which improves the rigor of the method, and at the same time makes the method of the present invention more intuitive and efficient, and enhances the operability and practicality of the method.

[0019] In the second aspect, the present invention provides a sound anomaly detection system based on a causal Gaussian process dynamic system. The system uses the sound anomaly detection method based on a causal Gaussian process dynamic system provided by the present invention. The system includes a data acquisition device, a processor, an output device and a storage device. Its gain lies in: the hardware facilities provided by the present invention have excellent performance, the data acquisition device, processor, output device and storage device are interconnected, and information transmission between each component is smooth. Through the interaction of multiple hardware facilities, the detection of the audio data set to be tested is efficiently completed. The detection results provide a reference for the staff, so that the staff can formulate measures in time to minimize the risks.

[0020] In summary, the method of the present invention introduces Gaussian process dynamic system and causal discovery method into the field of audio anomaly detection, which not only solves the problem of difficulty in training deep neural network anomaly detection model for small sample audio data, but also effectively improves the accuracy of modeling and provides an explainable basis for audio data anomaly detection. The method of the present invention is more practical and further optimizes the audio detection method. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 This is a flow chart of a method for detecting sound anomalies based on a causal Gaussian process dynamic system according to an embodiment of the present invention;

[0022] Figure 2 The audio waveform diagram of the fan in normal and abnormal operation according to the embodiment of the present invention;

[0023] Figure 3The audio causal relationship graph is inferred by the causal discovery method of the embodiment of the present invention;

[0024] Figure 4 4 is a framework diagram of a sound anomaly detection system based on a causal Gaussian process dynamic system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0025] The specific embodiments of the present invention will be described in detail below. It should be noted that the embodiments described herein are only for illustration and are not intended to limit the present invention. In the following description, a large number of specific details are set forth in order to provide a thorough understanding of the present invention. However, it is obvious to those of ordinary skill in the art that these specific details do not need to be adopted to implement the present invention. In other examples, in order to avoid confusing the present invention, known circuits, software or methods are not specifically described.

[0026] Throughout the specification, references to "one embodiment," "an embodiment," "an example," or "an example" mean that a particular feature, structure, or characteristic described in conjunction with the embodiment or example is included in at least one embodiment of the present invention. Therefore, the phrases "in one embodiment," "in an embodiment," "an example," or "an example" appearing in various places throughout the specification do not necessarily all refer to the same embodiment or example. In addition, particular features, structures, or characteristics may be combined in one or more embodiments or examples in any suitable combination and / or subcombination. In addition, it should be understood by those of ordinary skill in the art that the figures provided herein are for illustrative purposes and that the figures are not necessarily drawn to scale.

[0027] See also Figure 1 An embodiment of the present invention provides a method for detecting sound anomalies based on a causal Gaussian process dynamic system, the method comprising the following steps:

[0028] S1. Obtain an audio feature spectrum diagram of audio data.

[0029] Among them, S1 specifically includes the following steps:

[0030] S11. Establish an audio dataset.

[0031] Specifically, a microphone is used to collect audio data, thereby establishing the audio data set.

[0032] S12: Perform short-time Fourier transform on the audio signal of the audio data set to obtain a first spectrum graph.

[0033] Specifically, in this embodiment, a processor is used to divide the audio signal into small time windows using short-time Fourier transform, and then Fourier transform is performed on each window to obtain an image representation of frequency changes over time. The horizontal axis of the spectrum graph represents time, and the vertical axis represents frequency. At the same time, the obtained first spectrum graph is saved in a storage device.

[0034] More specifically, for a continuous-time signal x(t), its spectrum x(f) can be expressed using the Fourier transform integral formula as follows:

[0035] x(f)=∫x(t)e -j2πft dt

[0036] Where f is frequency, t is time, and j is the imaginary unit.

[0037] S13. Perform fast Fourier transform on the first spectrum graph to obtain a second spectrum graph.

[0038] Specifically, based on the first spectrum diagram obtained by short-time Fourier transform, a processor is used to convert the first spectrum diagram into a frequency domain representation using fast Fourier transform to obtain a second spectrum diagram, wherein the second spectrum diagram is represented as x(k), where k is a frequency index, and the second spectrum diagram is stored in a memory.

[0039] S14. Filter the second frequency spectrogram using a Mel filter bank to obtain a Mel frequency spectrogram.

[0040] Specifically, a processor is used to filter the second spectrum graph through a group of Mel filters, each Mel filter has a different center frequency and bandwidth, and after filtering, the frequency represented by the vertical axis of the second spectrum graph is converted into a Mel scale, and the Mel spectrum graph is stored in a storage.

[0041] More specifically, filtering is done by dot-wise multiplication of the spectrum obtained by fast Fourier transform with the Mel filter bank, namely:

[0042] Mel spectrum =x(k)×Mel_Filter(k)

[0043] Among them, Mel spectrum is the Mel spectrum, x(k) is the second spectrum obtained by fast Fourier transform, and Mel_Filter(k) is the response of the Mel filter bank.

[0044] S15. Taking the logarithm of the amplitude of the Mel-spectrogram to obtain a logarithmic Mel-spectrogram.

[0045] Specifically, a processor is used to take the logarithm of the amplitude of the Mel spectrum, convert the Mel spectrum into a logarithmic domain, and obtain a logarithmic Mel spectrum and save it in a memory. The formula is as follows:

[0046] log(Mel_Spectrum(k))=log(x(k)*Mel_Filter(k))

[0047] Wherein, Mel_Spectrum(k) is the amplitude of the Mel spectrum, x(k) is the second spectrum obtained by fast Fourier transform, and Mel_Filter(k) is the response of the Mel filter bank.

[0048] Furthermore, a processor is used to perform discrete cosine transform on the logarithmic Mel spectrum, and the 2nd to 13th coefficients after discrete cosine transform are taken as Mel frequency cepstrum coefficients and stored in the memory. The calculation formula is as follows:

[0049]

[0050] Among them, MFCC is the Mel frequency cepstral coefficient, Mel_Spectrum(k) is the amplitude of the Mel spectrum, n is the index of the Mel frequency cepstral coefficient, and k is the length of the Mel spectrum.

[0051] Discrete cosine transform is used to remove the correlation of the Mel frequency spectrum, and some discrete cosine transform coefficients are selected as Mel frequency cepstrum coefficients to remove the high-frequency part of the spectrum.

[0052] S2. Modeling the audio feature spectrum to obtain low-dimensional latent variables.

[0053] Specifically, a processor is used to perform dimension reduction modeling on the logarithmic Mel spectrum graph through a Gaussian process dynamic system, and the constructed model is saved in a storage device. Specifically, the model is constructed through the following formula:

[0054]

[0055] Among them, X q (t) is the qth dimension of the low-dimensional latent variable X of the spectrum at time t, is a Gaussian process, K X (t, t′) is the covariance matrix of any two moments t and t′, Q is the dimension of the logarithmic Mel spectrum data after the Gaussian process dynamic system is reduced in dimension, p(f|X) is the probability of f under the condition of X, f is a Gaussian process after the low-dimensional latent variable X is upgraded in dimension, X is the low-dimensional latent variable, is a Gaussian process distribution, K X,X is the covariance matrix calculated between any low-dimensional latent variable X and itself, p(y d |f) is y under the condition of f d The probability of y d is the dth dimension of the spectrum graph, K f,f is the covariance matrix calculated between any high-dimensional variable f and itself, and D is the original dimension of the spectrum graph.

[0056] Furthermore, if the audio data and its low-dimensional latent variables are independent in all dimensions, then:

[0057]

[0058] Where p is the probability of X under the condition of t, X is a low-dimensional latent variable, t is the time, Q is the dimension of the logarithmic Mel spectrum data after the Gaussian process dynamic system is reduced in dimension, and X q is the qth dimension of the low-dimensional latent variable X of the spectrum graph, is a Gaussian process distribution, K t,t is the covariance matrix between any two moments.

[0059] In addition, the covariance calculation formula is as follows:

[0060]

[0061] Among them, K(x,x′) is the radial basis kernel function, θ is the adjustment parameter of the function value range, exp is the exponential function with the natural constant e as the base, η is the width parameter of the function, and x i is any element in the low-dimensional latent variable X, x′ i is any element in the low-dimensional latent variable X′.

[0062] Furthermore, the processor is used to obtain the optimal solution of the random parameter while obtaining the optimal low-dimensional latent variable, and the obtained low-dimensional latent variable is stored in the storage.

[0063] Among them, the low-dimensional latent variable X={X t＝1 ,X t＝2 ,…,X t＝T} has N dimensions at each moment, that is, X t＝τ ={x1,x2,…,x N}, where τ∈(1,N); x i ,i∈(1,N) corresponds to each element in the different dimension sets at each moment.

[0064] S3. Based on the low-dimensional latent variables, an audio causal relationship graph is obtained.

[0065] Among them, S3 specifically includes the following steps:

[0066] S31. Calculate the Pierre correlation coefficient between an element and other elements in the low-dimensional latent variable.

[0067] Specifically, the certain element refers to each element in the different dimensional sets at each moment in the low-dimensional latent variable, and the Pierre correlation coefficient is calculated using a processor and stored in a memory.

[0068] S32. Based on the Pierre correlation coefficient, select elements that have a strong correlation with the certain element to form a candidate parent node set of the certain element.

[0069] Specifically, each τ different candidate parent node is stored as a different element in the candidate parent node set, where τ is the delay time, and the candidate parent node set is stored in the storage.

[0070] S33. Combine the certain element with the elements in the candidate parent node set to obtain a set of two-tuples.

[0071] Specifically, the binary set is (x i ,x n ), i∈(1,N), n∈(1,N), and save the two-tuple set in the memory.

[0072] S34. Calculate the partial correlation between the certain element and the elements in the candidate parent node set.

[0073] Specifically, using the processor, first x i and x n Perform linear regression on z respectively, and then calculate the partial correlation between the residuals of the two, and save the results in the memory. The calculation formula is as follows:

[0074] p=p(x i ,x n |z)=p(Res(x i ),Res(x n ))

[0075] Among them, p is the probability, x i and x n is the candidate parent node set P(x i ), p(x i ,x n |z) is x under z condition i and x n The probability of simultaneous occurrence, z is the candidate parent node set P(x i ) i and x n Element-wise and x i The element with the highest correlation, Res(x) is the first-order partial correlation coefficient of x.

[0076] It should be noted that when x n When taking z, z is the candidate parent node in the set that matches x. i The second most relevant element.

[0077] S35. Perform a significance test on the partial correlation to obtain a significance test result.

[0078] Specifically, if x i and x n Independently under z conditions, at a certain confidence level α, use the processor to judge x i and x n Whether the partial correlation between them is still significant, the judgment result is saved in the storage.

[0079] S36: Based on the significance test result, a set of candidate parent nodes to be screened out is obtained.

[0080] Specifically, the processor is used to process the judgment result in the memory. If x i and x n The partial correlation between is still significant, then x i and x n For z conditional correlation, that is, x n is x i Candidate parent node of x n Not x i Candidate parent node of , then delete node x n , and save the screened candidate parent node set in the storage.

[0081] S37: Determine whether all elements in the screened candidate parent node set are conditionally independent from the certain element.

[0082] Specifically, for the remaining elements in the candidate parent node, under the parent node condition of the candidate parent node, a processor is used to determine whether the current node and the candidate parent node are conditionally independent, and the result is stored in a storage.

[0083] S38. The elements related to the obtained conditions constitute a parent node set of the element, and an audio causal relationship graph is obtained.

[0084] Specifically, if the conditions are independent, the candidate parent node is deleted; if the conditions are related, it means that the candidate parent node has a direct causal relationship with the current node, that is, the candidate parent node is the causal parent node of the current node, and all conditionally related causal parent nodes constitute an audio causal relationship graph, and the obtained audio causal relationship graph is saved in the storage.

[0085] S4. Based on the normal audio causal relationship graph, a universal causal relationship graph is obtained.

[0086] Specifically, a processor is used to obtain normal audio data in an audio data set to obtain a normal audio causal relationship graph, and then the mean of each causal relationship value in the normal audio causal relationship graph is taken. All mean values ​​constitute a universal causal relationship graph of the audio data and are stored in a storage device.

[0087] S5. Calculate the mean square error between the normal audio causal relationship graph and the universal causal relationship graph, and use the mean square error as an empirical threshold. The calculation formula is as follows:

[0088]

[0089] Among them, ∈ ijτ is the mean square error of the (i, j) node value between the universal causal relationship graph and the normal audio causal relationship graph at time τ, T n is the duration of the nth audio. is the (i, j) node value of the nth normal audio causal relationship graph at time τ, is the value of the (i, j) node in the universal causal graph at time τ.

[0090] S6. Calculate the difference between the audio causal relationship graph to be tested and the universal causal relationship graph.

[0091] Specifically, based on the audio data to be tested in the audio data set, an audio causal relationship graph of the audio data to be tested is obtained, and then a processor is used to calculate the difference between the audio data to be tested and the universal causal relationship graph, and the difference is stored in a storage device.

[0092] S7: Compare the difference with the empirical threshold, and obtain a detection result of the audio data to be tested according to the comparison result.

[0093] Specifically, a processor is used to compare the difference with an empirical threshold, and a relation tuple whose difference is greater than the empirical threshold is added to an abnormal chain of the audio data to be tested, and the abnormal chain is stored in a storage.

[0094] Furthermore, by combining the causal relationship diagram and the exception chain of the audio data to be tested, all abnormal information of the audio data to be tested can be obtained and saved in the storage device. At the same time, the abnormal information of the audio data to be tested can be displayed to the staff through the output device. The abnormal information includes the number of abnormal nodes, abnormal source nodes, future time points of abnormal occurrence, etc.

[0095] Another embodiment of the present invention provides an experiment, which selects the classic industrial audio dataset MIMII for experiment to evaluate the operability and practicality of the proposed sound anomaly detection method based on the causal Gaussian process dynamic system.

[0096] The MIMII is a mechanical fault detection audio data set collected by a microphone, which has four tags: audio ID, data collection period, audio data, and audio tag. Among them, the audio tag includes two forms: "normal" and "fault".

[0097] See also Figure 2In this embodiment, sub-figure (a) shows an audio waveform diagram when the fan is running normally, and sub-figure (b) shows an audio waveform diagram when the faulty fan is running.

[0098] Specifically, all audio data with a "normal" label are selected as training data for the Gaussian process dynamic system, and a number of low-dimensional latent variables of audio data with time series information are obtained and stored in a storage device.

[0099] Furthermore, a sliding window with a window size of 10 is selected to create multiple audio low-dimensional latent variable sequence data with a step size of 4.

[0100] Furthermore, for this embodiment, if there is a causal relationship between nodes within the delay range of only three moments, the PCMCI algorithm is used to perform causal reasoning on low-dimensional latent variables.

[0101] See also Figure 3 In this embodiment, a universal causal relationship graph is shown which is inferred based on normal audio data in the MIMII dataset and is stored in a storage device.

[0102] Specifically, an audio causal relationship graph of the audio data set to be tested is obtained by inference, and based on the specific implementation steps of the sound anomaly detection method based on the causal Gaussian process dynamic system described in the present invention, the detection results of the audio data set to be tested can be obtained and displayed through an output device.

[0103] See also Figure 4 In an optional embodiment, in order to efficiently execute the sound anomaly detection method based on a causal Gaussian process dynamic system provided by the present invention, the present invention provides a sound anomaly detection system based on a causal Gaussian process dynamic system, and the sound anomaly detection system based on a causal Gaussian process dynamic system includes a data acquisition device, a processor, an output device and a storage device, and the data acquisition device, processor, output device and storage device are interconnected, wherein the data acquisition device is used to collect audio data sets, the processor is used to implement the steps of the method of the present invention to process the audio data sets to be tested, the output device is used to display the test results of the audio data sets to be tested, and the storage device is used to store program instructions, model data, image data, calculation formulas, etc. involved in the data processing process.

[0104] To sum up, the sound anomaly detection method and system based on the causal Gaussian process dynamic system provided by the method of the present invention introduces the Gaussian process dynamic system and causal discovery method into the field of audio anomaly detection, so that the method and system have a complete structure, rigorous steps, objectivity and stability, which makes up for the defects of existing audio detection methods and deficiencies in practical applications, improves the accuracy of audio detection results, and at the same time enhances the overall applicability and practical application capabilities of the present invention.

[0105] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein by equivalents. These modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be included in the scope of the claims and specification of the present invention.

Claims

1. A sound anomaly detection method based on a causal Gaussian process dynamic system, characterized in that: The steps include: Get the audio feature spectrum of the audio data; Modeling the audio feature spectrum to obtain a low-dimensional latent variable; Based on the low-dimensional latent variables, an audio causal relationship graph is obtained; Based on the normal audio causal relationship graph, a universal causal relationship graph is obtained; Calculating a mean square error value between a normal audio causal relationship graph and the universal causal relationship graph, and using the mean square error value as an empirical threshold; Calculating the difference between the audio causal relationship graph to be tested and the universal causal relationship graph; The difference is compared with the empirical threshold, and a detection result of the audio data to be tested is obtained according to the comparison result.

2. The sound anomaly detection method based on causal Gaussian process dynamic system according to claim 1 is characterized in that: The step of obtaining an audio feature spectrum diagram of audio data comprises: Create an audio dataset; Performing short-time Fourier transform on the audio signal of the audio data set to obtain a first spectrogram; Performing a fast Fourier transform on the first spectrum graph to obtain a second spectrum graph; Filtering the second spectrogram using a Mel filter bank to obtain a Mel spectrogram; The logarithm of the amplitude of the mel-spectrogram is taken to obtain a logarithmic mel-spectrogram.

3. The sound anomaly detection method based on causal Gaussian process dynamic system according to claim 2 is characterized in that: The step of modeling the audio feature spectrum graph to obtain a low-dimensional latent variable includes: Modeling the logarithmic Mel-frequency spectrum graph using a Gaussian process dynamic system to obtain a reference model; Based on the reference model, a mapping relationship between the logarithmic Mel spectrum data and the low-dimensional latent variable is obtained; Based on the logarithmic Mel spectrum data, an initial low-dimensional latent variable is obtained; Based on the initial low-dimensional latent variable, using the mapping relationship, obtaining logarithmic Mel spectrum prediction data; Subtract the logarithmic Mel spectrum prediction data from the actual data to obtain a difference; Adjusting random parameters in the reference model to revise the initial low-dimensional latent variables; Based on the revised initial low-dimensional latent variable, the logarithmic Mel spectrum prediction data is revised using the mapping relationship to minimize the difference; By adjusting the random parameters, low-dimensional latent variables are obtained while obtaining the optimal solution of the random parameters.

4. The sound anomaly detection method based on causal Gaussian process dynamic system according to claim 3 is characterized in that: The step of obtaining an audio causal relationship graph based on the low-dimensional latent variable includes: Calculate the Pierre correlation coefficient between an element and other elements in the low-dimensional latent variable; Based on the Pierre correlation coefficient, elements having a strong correlation with the certain element are screened to form a candidate parent node set of the certain element; Combine the certain element with the elements in the candidate parent node set to obtain a set of two-tuples; Calculate the partial correlation between the certain element and the elements in the candidate parent node set; Performing a significance test on the partial correlation to obtain a significance test result; Based on the significance test result, a set of candidate parent nodes to be screened out is obtained; Determine whether all elements in the screened candidate parent node set are conditionally independent from the certain element; The elements related to the obtained conditions constitute a parent node set of the element, and an audio causal relationship graph is obtained.

5. The sound anomaly detection method based on causal Gaussian process dynamic system according to claim 4 is characterized in that: The partial correlation between an element and the elements in the candidate parent node set is calculated using the following formula: p=p(x i ,x n |z)=p(Res(x i ),Res(x n )) Among them, p is the probability, x i and x n is the candidate parent node set P(x i ), p(x i , x n |z) is x under z condition i and x n The probability of simultaneous occurrence, z is the candidate parent node set P(x i ) i and x n Element-wise and x i The element with the highest correlation, Res(x) is the first-order partial correlation coefficient of x.

6. The sound anomaly detection method based on causal Gaussian process dynamic system according to claim 4 is characterized in that: The universal causal relationship graph is obtained based on the normal audio causal relationship graph, including: Obtaining normal audio data from the audio data set; Obtain a normal audio causal relationship graph based on the normal audio data; Taking the mean of the causal relationship values ​​in the normal audio causal relationship graph; Based on the mean, a universal causal relationship diagram is obtained.

7. The sound anomaly detection method based on causal Gaussian process dynamic system according to claim 6 is characterized in that: The mean square error value between the normal audio causal relationship graph and the universal causal relationship graph is calculated, and the mean square error value is used as the empirical threshold. The calculation formula is as follows: Among them, ∈ ijτ is the mean square error of the node value (i, j) between the universal causal relationship graph and the normal audio causal relationship graph at time τ, T n is the duration of the nth audio. is the (i, j) node value of the nth normal audio causal relationship graph at time τ, is the value of the (i, j) node in the universal causal graph at time τ.

8. The sound anomaly detection method based on causal Gaussian process dynamic system according to claim 7 is characterized in that: The calculating the difference between the causal relationship graph of the audio to be tested and the universal causal relationship graph comprises: Obtaining the audio data to be tested from the audio data set; Obtaining a cause-effect relationship diagram of the audio data to be tested based on the audio data to be tested; The difference between the audio causal relationship graph to be tested and the universal causal graph is calculated.

9. The sound anomaly detection method based on causal Gaussian process dynamic system according to claim 8 is characterized in that: The step of comparing the difference with the empirical threshold and obtaining a detection result of the audio data to be tested according to the comparison result includes: comparing the difference to an empirical threshold; Filter out the relation tuples whose difference is greater than the empirical threshold; Based on the relation tuple, an exception chain is obtained; By combining the abnormal chain and the causal relationship diagram of the audio data to be tested, all abnormal information of the audio data to be tested is obtained.

10. A sound anomaly detection system based on a causal Gaussian process dynamic system, characterized by: The system includes a data acquisition device, a processor, an output device and a storage device, which are interconnected, wherein the data acquisition device is used to collect audio data, the processor is used to process the collected audio data, the output device is used to display the audio detection results to be tested, and the storage device is used to store data and data processing results, and executes the sound anomaly detection method based on a causal Gaussian process dynamic system as described in any one of claims 1 to 9.

Images