Signal detection method, system and device, readable storage medium and program product

Abstract

The invention relates to a signal detection method, system and device, a readable storage medium and a program product. The method comprises the following steps: acquiring an audio signal to be processed; determining a time-frequency graph corresponding to the audio signal, performing segmentation processing on the time-frequency graph to obtain a plurality of local windows, and determining local frequency domain features of each local window; performing power entropy calculation on the local frequency domain feature of each local window, and extracting a spectrum chaos feature of the audio signal; obtaining a hyper-sphere with learned feature distribution of normal audio, and determining a mahalanobis distance between the spectrum chaotic feature and the hyper-sphere; and determining an abnormal result of the audio signal corresponding to the spectrum chaos feature according to the mahalanobis distance. By adopting the method, normal audio signals and all abnormal audio signals can be effectively distinguished.

Description

Technical Field

[0001] The present application relates to the field of information security technology, and in particular to a signal processing method, system, device, readable storage medium and program product. Background Art

[0002] With the rapid development of smart devices and IoT technology, sensors are used as key devices in smart devices. These devices receive and process sensor signals through sensors to achieve functions such as voice control and abnormal sound detection. However, the openness and universality of sensors also provide opportunities for attackers. They can use signals beyond the normal range of sensors to carry out covert attacks. These attacks are called out-of-band injection attacks.

[0003] Due to its concealment and invasiveness, out-of-band injection attacks can bypass traditional security protection measures and pose a serious threat to smart devices and equipment. Therefore, a signal detection method that can defend against out-of-band injection attacks is needed. Summary of the invention

[0004] Based on this, it is necessary to provide a signal detection method, system, computer device, computer-readable storage medium and computer program product that can defend against out-of-band injection attacks in response to the above technical problems.

[0005] In a first aspect, the present application provides a signal detection method, comprising:

[0006] Obtaining the audio signal to be processed;

[0007] Determine a time-frequency graph corresponding to the audio signal, segment the time-frequency graph to obtain a plurality of local windows, and determine a local frequency domain feature of each local window;

[0008] Performing power entropy calculation on the local frequency domain features of each local window to extract the frequency spectrum chaos features of the audio signal;

[0009] Acquire a hypersphere of feature distribution of learned normal audio, and determine the Mahalanobis distance between the spectral chaos feature and the hypersphere;

[0010] The abnormal result of the audio signal corresponding to the spectral chaos feature is determined according to the Mahalanobis distance.

[0011] In one embodiment, segmenting the time-frequency graph to obtain a plurality of local windows and determining the local frequency domain features of each local window includes:

[0012] Determine a preset sliding window;

[0013] Based on the preset sliding window, the time-frequency diagram is segmented to obtain a plurality of local windows;

[0014] For each local window among the multiple local windows, data is collected by sliding each local window on the corresponding frequency spectrum to obtain local frequency domain features of each local window.

[0015] In one embodiment, the performing power entropy calculation on the local frequency domain features of each local window to extract the spectrum chaos features of the audio signal includes:

[0016] determining an entropy function for quantifying the degree of disorder of power energy in each of the local windows;

[0017] The entropy function is applied to each of the local windows, and the power entropy is calculated for the local frequency domain characteristics of each of the local windows to generate a spectrum chaos map; the spectrum chaos map includes spectrum chaos characteristics of the signal at different times and frequencies.

[0018] In one embodiment, before segmenting the time-frequency graph to obtain a plurality of local windows and determining the local frequency domain features of each local window, the method further includes:

[0019] Get the horizontal line template;

[0020] In the case where there is a frequency line matching the horizontal line template in the time-frequency graph, determining a slope value of the target frequency line, wherein the magnitude of the slope value is used to characterize the energy distribution of the corresponding candidate frequency line;

[0021] If the slope value is within the numerical range representing the characteristics of the carrier signal, it is determined that the audio signal corresponding to the target frequency line is an abnormal signal, and the abnormal signal is filtered out from the audio signal.

[0022] In one of the embodiments, when there is a candidate frequency line matching the horizontal line template in the time-frequency graph, determining the slope value of the target frequency line includes:

[0023] Determine the energy distribution correlation between each frequency line in the time-frequency diagram and the horizontal line template;

[0024] If the energy distribution correlation is within a preset range, determining that there is a candidate frequency line in the time-frequency graph that matches the horizontal line template;

[0025] A target frequency line is determined from the candidate frequency lines according to the numerical value of the energy distribution correlation.

[0026] In one embodiment, obtaining a hypersphere of feature distribution of learned normal audio includes:

[0027] Acquire a normal speech database, and extract normal frequency spectrum chaos features in the normal speech database;

[0028] Performing feature encoding on the chaotic features of the normal frequency spectrum to obtain a corresponding multi-scale feature map;

[0029] A hypersphere including normal audio features is constructed according to the multi-scale feature map, wherein the hypersphere is formed by minimizing the distance between all mapping representations and the center of the hypersphere so that the normal audio features are closely clustered at the center of the hypersphere.

[0030] In one embodiment, determining the Mahalanobis distance between the spectral chaotic feature and the hypersphere comprises:

[0031] Performing feature encoding on the chaotic features of the spectrum to obtain a corresponding multi-scale feature map;

[0032] Performing feature adaptation on the multi-scale feature map, and calculating the Mahalanobis distance between the multi-scale feature map and the center of the hypersphere;

[0033] The abnormal result of the audio signal corresponding to the chaotic feature of the spectrum is determined according to the Mahalanobis distance, comprising:

[0034] If the Mahalanobis distance is not within a preset threshold range, it is determined that the audio signal corresponding to the frequency spectrum chaos feature is an abnormal signal.

[0035] In a second aspect, the present application further provides a signal detection device, the signal detection device comprising an audio input device and a device, the device comprising a memory and a processor, characterized in that the audio input device is used to collect an audio signal to be processed, the memory stores a computer program, and characterized in that when the processor executes the computer program, the following steps are implemented:

[0036] Obtaining the audio signal to be processed;

[0037] Determine a time-frequency graph corresponding to the audio signal, segment the time-frequency graph to obtain a plurality of local windows, and determine a local frequency domain feature of each local window;

[0038] Performing power entropy calculation on the local frequency domain features of each local window to extract the frequency spectrum chaos features of the audio signal;

[0039] Acquire a hypersphere of feature distribution of learned normal audio, and determine the Mahalanobis distance between the spectral chaos feature and the hypersphere;

[0040] The abnormal result of the audio signal corresponding to the spectral chaos feature is determined according to the Mahalanobis distance.

[0041] In a third aspect, the present application further provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the following steps are implemented:

[0042] Obtaining the audio signal to be processed;

[0043] Determine a time-frequency graph corresponding to the audio signal, segment the time-frequency graph to obtain a plurality of local windows, and determine a local frequency domain feature of each local window;

[0044] Performing power entropy calculation on the local frequency domain features of each local window to extract the frequency spectrum chaos features of the audio signal;

[0045] Acquire a hypersphere of feature distribution of learned normal audio, and determine the Mahalanobis distance between the spectral chaos feature and the hypersphere;

[0046] The abnormal result of the audio signal corresponding to the spectral chaos feature is determined according to the Mahalanobis distance.

[0047] In a fourth aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the following steps are implemented:

[0048] Obtaining the audio signal to be processed;

[0049] Determine a time-frequency graph corresponding to the audio signal, segment the time-frequency graph to obtain a plurality of local windows, and determine a local frequency domain feature of each local window;

[0050] Performing power entropy calculation on the local frequency domain features of each local window to extract the frequency spectrum chaos features of the audio signal;

[0051] Acquire a hypersphere of feature distribution of learned normal audio, and determine the Mahalanobis distance between the spectral chaos feature and the hypersphere;

[0052] The abnormal result of the audio signal corresponding to the spectral chaos feature is determined according to the Mahalanobis distance.

[0053] In a fifth aspect, the present application further provides a computer program product, including a computer program, which implements the following steps when executed by a processor:

[0054] Obtaining the audio signal to be processed;

[0055] Determine a time-frequency graph corresponding to the audio signal, segment the time-frequency graph to obtain a plurality of local windows, and determine a local frequency domain feature of each local window;

[0056] Performing power entropy calculation on the local frequency domain features of each local window to extract the frequency spectrum chaos features of the audio signal;

[0057] Acquire a hypersphere of feature distribution of learned normal audio, and determine the Mahalanobis distance between the spectral chaos feature and the hypersphere;

[0058] The abnormal result of the audio signal corresponding to the spectral chaos feature is determined according to the Mahalanobis distance.

[0059] The above-mentioned signal detection method, device, computer equipment, computer-readable storage medium and computer program product determine the time-frequency diagram corresponding to the audio signal, segment the time-frequency diagram, obtain multiple local windows and determine the local frequency domain characteristics of each local window; perform power entropy calculation on the local frequency domain characteristics of each local window to extract the spectral chaos characteristics of the audio signal. This method calculates the power spectrum entropy of each window in a sliding window manner, thereby extracting the spectral chaos characteristics of the audio signal, quantifies the degree of chaos in the audio signal to be processed, and can determine the degree of spectral chaos of all abnormal signals. Moreover, this processing method is not affected by factors such as out-of-band signal changes and environmental changes. On this basis, a hypersphere of the feature distribution of the learned normal audio is obtained, and the Mahalanobis distance between the spectral chaos characteristics and the hypersphere is determined. The abnormal result of the audio signal corresponding to the spectral chaos characteristics is determined according to the Mahalanobis distance, thereby realizing accurate abnormal signal detection. In other words, this detection method does not need to rely on a large amount of attack sample data, and can effectively distinguish normal audio signals from all abnormal audio signals even in the absence of attack samples, thereby improving practicality and adaptability. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the drawings required for use in the embodiments of the present application or related technical descriptions will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0061] Figure 1 A diagram showing an application environment of a signal detection method in an embodiment;

[0062] Figure 2 is a flow chart of a signal detection method in one embodiment;

[0063] Figure 3 is a schematic flow chart of step 204 in one embodiment;

[0064] Figure 4 A schematic diagram of a flow chart of a spectrum chaos feature extraction method in one embodiment;

[0065] Figure 5 Schematic diagram of a flow chart of a carrier trace prediction method in an embodiment;

[0066] Figure 6 is a schematic flow chart of a method for determining a hypersphere in one embodiment;

[0067] Figure 7 is a flow chart of a signal detection method in another embodiment;

[0068] Figure 8 is a structural block diagram of a signal detection device in one embodiment;

[0069] Fig. 9 is a structural block diagram of a signal detection device in one embodiment;

[0070] Fig.10 FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0071] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0072] With the rapid development of smart devices and Internet of Things technology, sensors are used as key devices in smart devices. Taking microphones as an example, microphones are widely used as key input devices in smart phones, smart homes, industrial automation and vehicle-mounted devices. These devices receive and process audio signals through microphones to achieve functions such as voice control and abnormal sound detection. However, the openness and ubiquity of microphones also provide opportunities for attackers. They can use signals beyond the normal hearing range of microphones (such as ultrasound, lasers and electromagnetic waves) to carry out covert attacks. These attacks are called out-of-band injection attacks.

[0073] Out-of-band injection attacks are stealthy and invasive, and can bypass traditional security measures, posing a serious threat to smart devices and equipment. For example, attackers can inject voice commands into voice-controlled devices through ultrasonic or laser attacks, or manipulate microphone recordings through electromagnetic interference, thereby disrupting connected broadcasting devices or sensor fusion networks. Therefore, audio signal detection is required.

[0074] However, in the traditional way, ultrasonic signals can be identified and filtered through hardware modification (such as adding ultrasonic speakers) or ultrasonic anomaly detection algorithms, but these methods usually rely on additional hardware support or training data of attack samples. Or use microphone detection array technology, a technology that uses microphone arrays to distinguish and locate sound sources, and identify forged ultrasonic signals by analyzing changes in multi-channel acoustic signals. This method can improve the detection capability of ultrasonic attacks, but may require multiple microphones and complex signal processing algorithms. This type of method is not applicable to low-end devices without microphone arrays. Or use multi-sensor fusion detection technology: integrate information from multiple different sensors, such as visual, audio, and vibration sensors, and design intelligent integrated systems to correct conflicts and avoid misjudgments. This method can improve the robustness of the system, but may require complex data processing and fusion algorithms. Additional sensors such as speakers, inertial measurement units, or microphone arrays are also required, which may not be applicable to all devices, especially low-end devices. Or out-of-band attack detection technology based on hardware modification: Some research is committed to modifying existing hardware, such as adding bias voltage or replacing analog circuit modules consistent with prototype circuits to detect and correct abnormal analog signals caused by out-of-band electromagnetic interference. These methods may require hardware upgrades for existing equipment, increasing costs and complexity. Or deep learning-based anomaly detection mechanisms: Some existing detection systems rely on deep learning techniques and require a large amount of training data to identify attack samples. This may limit their feasibility in practical applications because collecting comprehensive attack samples is time-consuming and laborious.

[0075] Traditional detection technologies usually focus on one type of overclocking signal injection attack, and lack effective detection methods for ultrasonic attacks, lasers, electromagnetic waves and other types of attacks. This leads to obvious security vulnerabilities in the device when facing multiple types of attacks, making it difficult to provide comprehensive protection. A signal detection method is proposed. This method considers the common characteristics of all abnormal signals, namely, out-of-band injection attack signals, namely, spectral chaos, measures the abnormal intrusion of out-of-band signals by quantifying the spectral chaos entropy of normal speech signals and abnormal out-of-band injection signals, constructs a feature memory library based on the spectral chaos entropy extracted from the public normal audio data set, and filters out-of-band signals based on the different feature distributions of normal samples and abnormal out-of-band signals, thereby realizing the detection of different types of out-of-band injection attack signals. Moreover, this method does not require additional hardware, does not rely on the number of adversarial samples, requires low-cost data sets, and is suitable for all microphone types. It can ensure the effective detection of all out-of-band injection attacks, showing higher robustness, high versatility and strong security.

[0076] The signal detection method provided in the embodiment of the present application can be applied to Figure 1In the application environment shown. Among them, the audio input device 102 communicates with the terminal 104 through the network. The data storage device can store data that the terminal 104 needs to process. The data storage device can be integrated on the terminal 104, or it can be placed on the cloud or other network servers. The audio input device 102 collects the audio signal to be processed, sends the collected audio signal to be processed to the terminal, and obtains the audio signal to be processed; determines the time-frequency diagram corresponding to the audio signal, divides the time-frequency diagram, obtains multiple local windows and determines the local frequency domain features of each local window; performs power entropy calculation on the local frequency domain features of each local window, and extracts the spectral chaotic features of the audio signal; obtains a hypersphere of the feature distribution of the learned normal audio, and determines the Mahalanobis distance between the spectral chaotic features and the hypersphere; determines the abnormal result of the audio signal corresponding to the spectral chaotic features according to the Mahalanobis distance. The audio input device 102 may be provided on the terminal 104. The audio input device 102 may be different types of input devices, for example, a microphone. The terminal 104 may be, but is not limited to, various personal computers, laptops, smart phones, tablet computers, IoT devices, and portable wearable devices. The IoT devices may be smart speakers, smart TVs, smart air conditioners, smart vehicle-mounted devices, projection devices, etc. The portable wearable devices may be smart watches, smart bracelets, head-mounted devices, etc.

[0077] In an exemplary embodiment, Figure 2 As shown, a signal detection method is provided, which is applied to Figure 1 The terminal in is taken as an example to illustrate, including the following steps 202 to 210. Among them:

[0078] Step 202: Acquire an audio signal to be processed.

[0079] The audio signal to be processed may be an audio signal received by a sensor to be detected, for example, the sensor to be detected may be but is not limited to any one of an inertial measurement unit (IMU), a vibration sensor and a sound sensor. The audio signal to be processed may include an in-band signal and / or an out-of-band signal. The in-band signal may be understood as a normal voice signal, and the out-of-band signal may be understood as a signal of other non-speech audio bands / propagation paths.

[0080] Optionally, taking the sensor to be detected as a sound sensor and the sound sensor as a microphone as an example, the microphone collects the input audio signal to obtain the audio signal to be processed. The way in which the microphone collects and processes the signal can be implemented by existing methods and will not be elaborated here.

[0081] Step 204: determine the time-frequency diagram corresponding to the audio signal, segment the time-frequency diagram to obtain multiple local windows, and determine the local frequency domain features of each local window.

[0082] Among them, the method of determining the time-frequency diagram can be to convert the audio signal from the time domain to the time-frequency domain through short-time Fourier transform to obtain a two-dimensional time-frequency diagram. The time-frequency diagram can decompose the audio signal into components of different time and frequency and show their changes over time.

[0083] It should be noted that in order to effectively analyze the frequency components of non-stationary signals, reduce spectrum leakage and improve time resolution, the time-frequency graph can be smoothed and divided into multiple local windows. Furthermore, smoothing the audio signal can reduce the mutations at both ends of the signal, thereby reducing spectrum leakage. Dividing the time-frequency graph into multiple local windows helps to analyze the behavior of the signal in different frequency ranges. This can be used for a variety of applications such as feature extraction, signal classification, and sound synthesis.

[0084] When smoothing the time-frequency graph, it can be implemented by using, but not limited to, a Blackman window, a Hanning window or a Hamming window, etc. This embodiment uses a Blackman window as an example. Further, the size of the window determines the time resolution and frequency resolution of the STFT (Short-time Fourier Transform). A larger window provides better frequency resolution, but a lower time resolution. The size of the specific window can be determined according to the actual application scenario. For example, the window size can be 512 points, and 512 points can provide a balanced time and frequency resolution. Further, the overlapping part between the windows helps to improve the time resolution and can make the signal smoother at the window junction. Optionally, in an exemplary embodiment, the overlapping part can be 375 points.

[0085] It can be understood that the spectrum is the representation of the signal in the frequency domain. The spectrum is the basis of the spectrum matrix and the time-spectrum diagram. The spectrum matrix and the time-spectrum diagram are derived from the spectrum. The values ​​in the spectrum matrix are displayed in the form of an image through the time-spectrum diagram. By dividing the time-spectrum diagram into multiple windows and sliding the corresponding window on the corresponding spectrum matrix in each window, the time-frequency characteristics of the signal can be analyzed, and local information can be aggregated, that is, the local frequency domain characteristics of each local window can be obtained. Among them, the window N P (x, y) collects local information at position (x, y) in the spectrum S, which can be expressed as:

[0086]

[0087] Exemplarily, the audio signal is converted from the time domain to the time-frequency domain by short-time Fourier transform to obtain a two-dimensional time-frequency diagram, the time-frequency diagram is smoothed by using the Blackman window, the time-frequency diagram is divided into multiple windows, and each window is Fourier transformed to obtain the corresponding spectrum matrix, and the corresponding window is slid on the corresponding spectrum matrix in each window to collect local information and obtain the local frequency domain characteristics of each local window. It should be noted that this method can well analyze and understand the time-frequency characteristics of non-stationary signals, that is, it can retain the time-frequency resolution while avoiding the compression of information entropy, ensuring that the multi-level chaotic characteristics injected into the audio are captured.

[0088] Step 206: Perform power entropy calculation on the local frequency domain features of each local window to extract the frequency spectrum chaos features of the audio signal.

[0089] The spectral chaos feature may refer to the feature of a signal showing chaotic characteristics in the frequency domain. The power entropy calculation of the local frequency domain feature of each local window can be understood as the calculation of the power entropy of each local window N. p (x,y) applies the power entropy function. Power entropy is an indicator in the field of signal processing, which is used to measure the distribution of frequency components of a signal. It combines the concepts of power spectral density (PSD) and information entropy to evaluate the complexity and uncertainty of the signal.

[0090] The general way to calculate power entropy is to obtain signal data in the time domain, usually use fast Fourier transform (FFT) to convert the time domain signal to the frequency domain, and then calculate the power of each frequency component, usually the square of the frequency component. The calculation method can use the Welch method or other smoothing techniques to estimate PSD, which is more stable and reliable. On this basis, the power spectrum density can be normalized by dividing it by the total power to make it a probability distribution. The power of each frequency component represents the probability of the component appearing. Then, the entropy of the normalized power spectrum density is calculated using the discrete entropy formula. The entropy calculation formula is: H = -∑iP(fi)logP(fi); where (P(f_i)) represents the normalized power spectrum density corresponding to the frequency component (f_i).

[0091] It is understandable that the spectrum of in-band speech will not be chaotic or distorted, while out-of-band signals will introduce additional distortion. In order to measure the degree of "chaos" / "distortion" of the out-of-band injected signal on the injected audio, the power spectrum entropy can be selected as the entropy function.

[0092] Optionally, the power spectrum entropy is used to calculate the power entropy of the local frequency domain characteristics of each local window, and the spectrum chaos characteristics of the audio signal are determined according to the calculated power entropy value. The power spectrum entropy is used to quantify the degree of chaos of the power energy in the local window, and the specific formula is:

[0093] E p (x,y)=f entropy (N p (x,y));

[0094] Among them, N p (x,y) represents the local window, E p (x,y) is the power entropy value.

[0095] Step 208, obtaining a hypersphere of the learned feature distribution of normal audio, and determining the Mahalanobis distance between the spectral chaos feature and the hypersphere.

[0096] The hypersphere can be understood as describing the distribution of normal audio data in the feature space, which can be determined by the deep support vector data description (DSVDD) model based on the spectral chaos features extracted from a large public benign speech database. It can also be understood that the hypersphere is the feature space representation of normal audio data learned by the DSVDD model. The Mahalanobis distance between the spectral chaos features and the hypersphere The calculation method can be:

[0097]

[0098] Wherein, μ can be the mean value of the spectral chaos feature extracted from the preset benign speech database, ∑ is the covariance matrix, and t i It is a spectral chaos feature.

[0099] Step 210: determining the abnormal result of the audio signal corresponding to the chaotic feature of the spectrum according to the Mahalanobis distance.

[0100] It is understandable that the Mahalanobis distance As an abnormality score, that is, an abnormal result, the higher the score, the greater the possibility of abnormality. i If the anomaly score exceeds the preset threshold boundary, it is determined to be an out-of-band injection attack, that is, the corresponding audio signal is an abnormal signal.

[0101] In the above signal detection method, by determining the time-frequency diagram corresponding to the audio signal, the time-frequency diagram is segmented to obtain multiple local windows and the local frequency domain features of each local window are determined; the power entropy of the local frequency domain features of each local window is calculated to extract the spectral chaos features of the audio signal. This method calculates the power spectrum entropy of each window in a sliding window manner, thereby extracting the spectral chaos features of the audio signal, quantifying the degree of chaos in the audio signal to be processed, and being able to determine the degree of spectral chaos of all abnormal signals. Moreover, this processing method is not affected by factors such as out-of-band signal changes and environmental changes. On this basis, a hypersphere of the feature distribution of the learned normal audio is obtained, and the Mahalanobis distance between the spectral chaos features and the hypersphere is determined. The abnormal result of the audio signal corresponding to the spectral chaos features is determined according to the Mahalanobis distance, thereby achieving accurate abnormal signal detection. In other words, this detection method does not need to rely on a large amount of attack sample data, and can effectively distinguish normal audio signals from abnormal audio signals in the absence of attack samples, thereby improving practicality and adaptability.

[0102] It can be understood that the extraction of the chaotic features of the spectrum of the audio signal is achieved by calculating the power spectrum entropy of each window in a sliding window manner. In an exemplary embodiment, Figure 3 As shown, step 204 includes steps 302 to 306. Among them:

[0103] Step 302: determine a preset sliding window.

[0104] The preset sliding window may be a Blackman window, which is a weighted window function that can reduce the amplitude of the signal at both ends of the window, reduce the mutation of the signal, thereby reducing spectrum leakage and making the spectrum smoother. The Blackman window has a window size of 512 points and an overlap of 375 points.

[0105] Step 304: segment the time-frequency graph based on a preset sliding window to obtain a plurality of local windows.

[0106] Exemplarily, based on the Blackman window, the window size is 512 points, and the overlapping part is 375 points. The time-frequency diagram is segmented to obtain multiple local windows.

[0107] Step 306 , for each local window among the multiple local windows, collect data by sliding each local window on the corresponding frequency spectrum to obtain local frequency domain features of each local window.

[0108] In the above embodiment, based on the sliding window of the Blackman window, the time-frequency resolution can be retained while avoiding the compression of information entropy, thereby ensuring that the multi-level chaotic features injected into the audio are captured.

[0109] Traditional entropy measurement will compress information and has limited ability to capture chaotic features when extracting spectral chaos features. In order to effectively detect all out-of-band injection attacks, a spectral chaos feature extraction method is provided, such as Figure 4 As shown, it includes steps 402 to 404, wherein:

[0110] Step 402: determine an entropy function for quantifying the degree of disorder of power energy in each local window.

[0111] Among them, since power spectrum entropy performs best in distinguishing in-band and out-band overclocked signal categories, power spectrum entropy is selected as the entropy function, and power spectrum entropy is determined as the entropy function used to quantify the degree of disorder of power energy in each local window.

[0112] Step 404, applying the entropy function to each local window, performing power entropy calculation on the local frequency domain characteristics of each local window, and generating a spectrum chaos map; the spectrum chaos map includes spectrum chaos characteristics of the signal at different times and frequencies.

[0113] Among them, the power entropy calculation can also be understood as the power spectrum entropy calculation, and its specific calculation method can be realized by the calculation method of other entropy functions, which will not be repeated here. It can be understood that the spectrum chaos diagram provides a way to display the spectrum chaos characteristics, and the spectrum chaos characteristics are what the spectrum chaos diagram wants to express.

[0114] In the above embodiment, the power spectrum entropy is used for extraction to measure the degree of "chaos" and "distortion" of the out-of-band injected signal on the injected audio, avoiding the defects caused by information compression, capturing richer chaos features, and significantly improving the effect of out-of-band anomaly detection.

[0115] High-energy ultrasonic signals or TTL (Transistor-Transistor Logic, integrated circuit) modulated laser signals can induce carrier trace signals. Before extracting spectral chaos features, in order to reduce the amount of data processing, high-frequency attack signals that can induce carrier trace signals can be filtered. Therefore, in one embodiment, Figure 5 As shown, a carrier trace prediction method is provided, including steps 502 to 506, wherein:

[0116] Step 502, obtaining a horizontal line template.

[0117] It is understandable that the carrier trace is a horizontal line on the spectrum, so a horizontal line [:, f] across the time domain is constructed, that is, a straight line across all time points at a fixed frequency f on the spectrum graph. There is only one horizontal line template, and its correlation with each spectrum line is calculated line by line based on the horizontal line template. The spectrum line is related to the frequency resolution during short-time Fourier transform. The higher the resolution, the more spectrum lines, and the higher the correlation, indicating that a horizontal carrier trace is generated, indicating that an out-of-band attack has occurred. No matter how many out-of-band attacks occur, the difference between multiple or one carrier trace is that as long as it is judged once, it means that an out-of-band attack has occurred.

[0118] Exemplarily, the audio signal to be processed is converted into a time-frequency diagram using short-time Fourier transform (STFT), that is, the received audio signal is converted into a time-frequency diagram with time frames as columns and frequency frames as rows, and frequency lines in horizontal line template matching are constructed.

[0119] Step 504: when there is a frequency line matching the horizontal line template in the time-frequency diagram, determine the slope value of the target frequency line, and the magnitude of the slope value is used to characterize the energy distribution of the corresponding candidate frequency line.

[0120] The horizontal line template matching can be understood as calculating the correlation coefficient between the constructed horizontal line template and each row (ie, frequency line) in the time-frequency diagram. The correlation coefficient calculation can be achieved in an existing manner.

[0121] For example, since the higher the correlation coefficient is, the closer the energy distribution of the frequency line is to the horizontal line, that is, the more likely there is a carrier trace. In the case where there are frequency lines matching the horizontal line template in the time-frequency diagram, the frequency line with the highest correlation coefficient can be determined from the matching frequency lines as the target frequency line, and the slope value of the target frequency line can be calculated. The calculation method of the slope value of the target frequency line can be implemented by existing methods, which will not be repeated here.

[0122] Optionally, in an exemplary embodiment, when there are candidate frequency lines in the time-frequency graph that match the horizontal line template, the slope value of the target frequency line is determined, including: determining the energy distribution correlation between each frequency line in the time-frequency graph and the horizontal line template; if the energy distribution correlation is within a preset range, determining that there are candidate frequency lines in the time-frequency graph that match the horizontal line template; and determining the target frequency line from the candidate frequency lines according to the numerical value of the energy distribution correlation.

[0123] Exemplarily, in the time-frequency diagram, each row represents the amplitude of all frequency components at a specific time point, all frequency lines in the time-frequency diagram are extracted, and the correlation coefficients of the extracted frequency lines are calculated with the horizontal line template respectively. According to the obtained correlation value, the matching between each frequency line and the horizontal line template is determined. If the correlation value is within the preset range, that is, the higher the correlation coefficient, the closer the energy distribution of the frequency line is to the horizontal line, that is, the more likely there is a carrier trace, that is, it matches the horizontal line template, then the frequency line that matches the horizontal line template can be one or more. In the case of determining that there are frequency lines that match the horizontal line template in the time-frequency diagram, in order to further accurately filter out abnormal signals that can cause carrier traces in the audio signal to be processed, the candidate frequency line with the largest value of energy distribution correlation, that is, the largest correlation coefficient, is determined as the target frequency line.

[0124] Step 506: If the slope value is within the numerical range representing the characteristics of the carrier signal, it is determined that the audio signal corresponding to the target frequency line is an abnormal signal, and the abnormal signal is filtered out from the audio signal.

[0125] Among them, the slope value is within the numerical range that characterizes the characteristics of the carrier signal. It can be understood that if the slope is close to zero, it means that the energy distribution of the frequency line is uniform and highly correlated with the horizontal line template, so it is judged that there is a carrier trace, that is, the audio signal corresponding to the slope value is abnormal information, that is, an out-of-band injection attack. Once a carrier trace is detected, the device will immediately filter out the audio signal, thereby quickly identifying and blocking some high-frequency attack signals.

[0126] It is understandable that this part of the carrier trace pre-judgment is to remove part of the out-of-band injected audio in advance. However, there are still some out-of-band injected signals that will bypass the carrier trace pre-judgment module because they will not cause carrier traces, so we have to make another judgment / detection based on the spectrum chaos entropy, that is, to extract the spectrum chaos entropy based on the audio signal with the carrier traces filtered out.

[0127] In this embodiment, the audio signal is converted into a time-frequency diagram by utilizing short-time Fourier transform, and a horizontal line template is constructed to match the frequency line to detect whether there is a straight line consistent with the carrier frequency, thereby quickly identifying and filtering out some high-frequency attack signals. Furthermore, this method does not rely on prior information and data of the attacker and the victim device, and can be based on the traces generated by the carrier of the out-of-band injection signal observed on devices of different levels.

[0128] Modern detection technologies such as deep learning usually require a large amount of attack sample data for training, which is often difficult to achieve in practical applications because attack methods are constantly evolving and collecting comprehensive data requires huge resource investment. In order to reduce the dependence on attack sample data and effectively learn and identify abnormal signals without a large number of attack samples, detection can be performed based on feature memory libraries. To achieve this goal, a method for determining a hypersphere is provided, such as Figure 6 As shown, it includes steps 602 to 606, wherein:

[0129] Step 602: Acquire a normal speech database, and extract normal frequency spectrum chaos features in the normal speech database.

[0130] The normal speech database can be understood as a large public benign speech database, and the benign speech database can be understood as a speech database without abnormal speech. The normal spectrum chaos feature can be extracted by the above method, which will not be elaborated here.

[0131] Step 604: feature encoding is performed on the chaotic features of the normal spectrum to obtain a corresponding multi-scale feature map.

[0132] Among them, feature encoding can be to use a pre-trained convolutional neural network to further extract deep features from normal spectrum chaotic features.

[0133] Optionally, feature encoding can be done by using a pre-trained convolutional neural network (CNN) to extract high-level embedded representations. The convolutional neural network widens the network, increases the number of filters and regularizes, where each module network layer contains three convolutional layers and the number of channels is twice that of a normal convolutional network. This can capture a wider range of feature patterns and make the network better at understanding complex features and changes. The network consists of three module network layers and an average pooling layer Then, these four layers of feature maps are uniformly upsampled and concatenated along the channel axis to form a multi-scale feature map z i :

[0134] It can be seen that these feature maps are no longer the original chaotic maps, but their high-level compressed representations, which help to form data distribution in the network memory.

[0135] Step 606: construct a hypersphere including normal audio features according to the multi-scale feature map, wherein the hypersphere is constructed by minimizing the distance between all mapping representations and the center of the hypersphere so that the normal audio features are closely clustered at the center of the hypersphere.

[0136] It can be understood that in order to adapt to the spectral chaos characteristics in the audio, the Deep Support Vector Data Description (DSVDD) algorithm is used to perform feature distribution modeling and learning on the feature map.

[0137] Exemplarily, the deep support vector data description (DSVDD) model uses the extracted normal spectrum chaotic features to learn a hypersphere and construct a feature memory library. This hypersphere describes the distribution of normal audio data in the feature space, where the optimization goal is:

[0138]

[0139] Φ is the mapping function, Θ is an adjustable parameter, and c is the center of the hypersphere. That is, by minimizing the distance between all mapping representations and the center, normal data features are closely clustered at the center of the hypersphere, thus forming an adapted feature memory library.

[0140] Furthermore, based on the feature memory library obtained through the above training, in the actual application stage, the input sample is first extracted through spectral chaotic entropy to obtain a chaotic entropy spectrogram, and then these spectrograms are converted into deep feature vectors through the feature encoding process. Next, the DSVDD model adapts these feature vectors and calculates the Mahalanobis distance between the input audio sample and the center of the hypersphere. This distance is used as an anomaly score, and based on the anomaly score, out-of-band attack signals in the input audio sample can be identified.

[0141] In the above embodiment, by extracting spectral chaos features from a large public benign speech database, a pre-trained convolutional neural network is used to further extract deep features, i.e., feature encoding, from the spectral chaos, and a hypersphere is learned through deep support vector data description to learn the feature distribution of normal audio, thereby identifying various types of attack signals. That is, a feature memory library is constructed using only publicly available normal audio data sets, and these feature distributions are modeled and memorized through deep support vectors to identify abnormal signals. This method can effectively distinguish between normal audio and attack audio in the absence of attack samples, and can effectively identify various types of attack signals, and can maintain high accuracy even if the attack methods continue to evolve.

[0142] In the above embodiment, a feature memory library is constructed through training. On this basis, optionally, in an exemplary embodiment, the Mahalanobis distance between the spectral chaos feature and the hypersphere is determined, including: feature encoding the spectral chaos feature to obtain a corresponding multi-scale feature map; feature adaptation of the multi-scale feature map to calculate the Mahalanobis distance between the multi-scale feature map and the center of the hypersphere; if the Mahalanobis distance is not within a preset threshold range, it is determined that the audio signal corresponding to the spectral chaos feature is an abnormal signal.

[0143] Among them, feature coding can be implemented by the method in the above embodiment, which will not be described in detail here.

[0144] It can be understood that all the spectral chaotic features of each audio signal to be processed are i First, it is encoded and adapted, and then compared with the feature distribution stored in the feature memory M. In order to reduce the impact of noise caused by environmental fluctuations and unstable acquisition equipment, this module assumes that the amount and intensity of noise are less than normal data, and adopts multivariate Gaussian distribution modeling. The feature memory can be determined in the above way, which will not be repeated here. The joint probability density function of the memory is defined as:

[0145]

[0146] Among them, μ is the mean of the memory bank, ∑ is the covariance matrix. Then, all spectral chaos features t are calculated i Mahalanobis distance to memory distribution

[0147]

[0148] It should be noted that the Mahalanobis distance As the abnormality score, the higher the score, the greater the possibility of abnormality. i If the abnormal score exceeds the preset threshold, the corresponding audio signal is determined to be an out-of-band injection attack. In this way, out-of-band attack signals can be effectively detected and identified in practical applications, significantly improving the security and reliability of audio devices and other devices.

[0149] In the embodiments of this method, by constructing and utilizing the distribution characteristics of normal audio samples and combining deep learning and statistical models, efficient detection and identification of abnormal audio signals are achieved. It has high practicality and accuracy, and can maintain high accuracy even when the attack method changes.

[0150] In another exemplary embodiment, Figure 7 As shown, a signal detection method is provided, which is applied to Figure 1The terminal in is taken as an example to illustrate, including the following steps 702 to 716. Among them:

[0151] Step 702: Acquire the audio signal to be processed and determine the time-frequency diagram corresponding to the audio signal.

[0152] Step 704, determine the horizontal line template.

[0153] Step 706: if there is a frequency line matching the horizontal line template in the time-frequency diagram, determine the slope value of the target frequency line, and the magnitude of the slope value is used to characterize the energy distribution of the corresponding candidate frequency line.

[0154] Step 708: If the slope value is within the numerical range representing the characteristics of the carrier signal, it is determined that the audio signal corresponding to the target frequency line is an abnormal signal, and the abnormal signal is filtered out from the audio signal.

[0155] Step 710 , segment the time-frequency graph after filtering out abnormal signals to obtain multiple local windows and determine the local frequency domain features of each local window.

[0156] Step 712: Perform power entropy calculation on the local frequency domain features of each local window to extract the frequency spectrum chaos features of the audio signal.

[0157] Step 714, obtaining a hypersphere of the learned feature distribution of normal audio, and determining the Mahalanobis distance between the spectral chaos feature and the hypersphere.

[0158] Step 716: determine the abnormal result of the audio signal corresponding to the spectral chaos feature according to the Mahalanobis distance.

[0159] It should be noted that the specific implementation method of each step in this embodiment can be implemented by the method recorded in the above embodiment, which will not be repeated here.

[0160] The above method is applied to a microphone device for explanation. The microphone device receives an audio signal to be processed, and first performs carrier trace prediction on the audio signal, that is, converts the audio signal into a time-frequency diagram using a short-time Fourier transform, then constructs a horizontal line template, extracts frequency lines from the time-frequency diagram, calculates correlation coefficients between the horizontal line template and each frequency line in the time-frequency diagram, and performs horizontal line template matching. When there is a frequency line matching the horizontal line template in the time-frequency diagram, the slope value of the target frequency line is determined. If the slope value is within the numerical range that characterizes the characteristics of the carrier signal, the audio signal corresponding to the target frequency line is determined to be an abnormal signal, and the abnormal signal is filtered out from the audio signal, that is, some out-of-band attack signals are filtered out to obtain a time-frequency diagram with the abnormal signal filtered out.

[0161] The time-frequency graph that filters out abnormal signals is segmented to obtain multiple local windows and determine the local frequency domain features of each local window. The power entropy of the local frequency domain features of each local window is calculated to extract the spectrum chaos features of the audio signal corresponding to the time-frequency graph that filters out abnormal signals. Based on the normal spectrum chaos features extracted from a large public benign speech database, a pre-trained convolutional neural network is used to further extract deep features, i.e., feature encoding, from the normal spectrum chaos features, and a hypersphere is learned through deep support vector data description to learn the feature distribution of normal audio. After the above training, a feature memory library is constructed. On this basis, the extracted spectrum chaos features are feature encoded and DSVDD feature adapted, and then the Mahalanobis distance between the hypersphere is calculated. The abnormal score of the audio signal corresponding to the spectrum chaos feature is determined according to the Mahalanobis distance. If the abnormal score exceeds the threshold, the corresponding audio signal is judged to be an out-of-band attack and filtered out. If it does not exceed the threshold, the corresponding audio line signal is determined to be a normal signal.

[0162] In this method, abnormal out-of-band signals are distinguished from abnormal speech features, which is suitable for all high-end and low-end microphone devices. It does not rely on any additional hardware and equipment, and uses the signal of the microphone device itself for carrier trace detection and spectrum chaos entropy calculation. It does not require any additional hardware equipment, such as speakers, inertial sensors or microphone arrays, and can be widely used in microphone devices of various levels, such as low-cost sensors, recording microphones, smartphones and tablets. Moreover, it is based on the common phenomenon of out-of-band injection, namely carrier and chaotic spectrum entropy, to fully resist out-of-band injection attacks and improve the defense efficiency of the device.

[0163] In the above signal detection method, the carrier trace characteristics unique to out-of-band attacks are used to identify and filter out some high-frequency attack signals through carrier trace detection, which can identify and reject some high-energy injection signals in the first place, reduce computing costs and response time, and improve the defense efficiency of the device. And based on the different degrees of chaos caused by in-band and out-of-band signals on the voice spectrum, the spectrum chaos features are extracted through spectrum chaos entropy, and a feature memory library is built on this basis. The deep support vector is used to model and memorize the distribution of these spectrum chaos features to identify abnormal signals, effectively distinguish normal audio from various types of attack audio, and accurately detect attack audio. In addition, this method can also achieve efficient detection of various overclocked signal injection attacks such as ultrasound, lasers, and electromagnetic waves without relying on additional hardware, ensuring that the device can maintain efficient and reliable protection performance in various environments.

[0164] It should be understood that, although the various steps in the flowcharts involved in the above-mentioned embodiments are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps does not have a strict order restriction, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-mentioned embodiments can include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the steps or stages in other steps.

[0165] Based on the same inventive concept, the embodiment of the present application also provides a signal detection device for implementing the signal detection method involved above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above method, so the specific limitations in one or more signal detection device embodiments provided below can refer to the limitations on the signal detection method above, and will not be repeated here.

[0166] In an exemplary embodiment, Figure 8 As shown, a signal detection device is provided, including: a signal acquisition module 802, a segmentation processing module 804, a feature extraction module 806 and a signal detection module 808, wherein:

[0167] The signal acquisition module 802 is used to acquire the audio signal to be processed.

[0168] The segmentation process 804 is used to determine the time-frequency diagram corresponding to the audio signal, perform segmentation processing on the time-frequency diagram to obtain multiple local windows, and determine the local frequency domain features of each local window.

[0169] The feature extraction module 806 is used to perform power entropy calculation on the local frequency domain features of each local window to extract the frequency spectrum chaos features of the audio signal.

[0170] The signal detection module 808 is used to obtain the hypersphere of the learned normal audio feature distribution, determine the Mahalanobis distance between the spectral chaos feature and the hypersphere; and determine the abnormal result of the audio signal corresponding to the spectral chaos feature according to the Mahalanobis distance.

[0171] The above-mentioned signal detection device determines the time-frequency diagram corresponding to the audio signal, divides the time-frequency diagram, obtains multiple local windows and determines the local frequency domain characteristics of each local window; performs power entropy calculation on the local frequency domain characteristics of each local window, and extracts the spectral chaos characteristics of the audio signal. This method calculates the power spectrum entropy of each window in a sliding window manner, thereby extracting the spectral chaos characteristics of the audio signal, quantifies the degree of chaos in the audio signal to be processed, and can determine the degree of spectral chaos of all abnormal signals. Moreover, this processing method is not affected by factors such as out-of-band signal changes and environmental changes. On this basis, a hypersphere of the feature distribution of the learned normal audio is obtained, and the Mahalanobis distance between the spectral chaos characteristics and the hypersphere is determined. The abnormal result of the audio signal corresponding to the spectral chaos characteristics is determined according to the Mahalanobis distance, thereby realizing accurate abnormal signal detection. In other words, this detection method does not need to rely on a large amount of attack sample data, and can effectively distinguish normal audio signals from abnormal audio signals in the absence of attack samples, thereby improving practicality and adaptability.

[0172] Optionally, in an exemplary embodiment, the segmentation process 804 is also used to determine a preset sliding window; based on the preset sliding window, the time-frequency diagram is segmented to obtain multiple local windows; for each local window in the multiple local windows, data is collected by sliding each local window on the corresponding spectrum to obtain local frequency domain characteristics of each local window.

[0173] Optionally, in an exemplary embodiment, the feature extraction module 806 is further used to determine an entropy function for quantifying the degree of disorder of the power energy in each local window;

[0174] The entropy function is applied to each local window, and the power entropy is calculated for the local frequency domain characteristics of each local window to generate a spectrum chaos map; the spectrum chaos map includes the spectrum chaos characteristics of the signal at different times and frequencies.

[0175] Optionally, in an exemplary embodiment, the signal detection device further includes a carrier trace pre-judgment module, which is used to obtain a horizontal line template; when there is a frequency line matching the horizontal line template in the time-frequency diagram, a slope value of the target frequency line is determined, and the magnitude of the slope value is used to characterize the energy distribution of the corresponding candidate frequency line;

[0176] If the slope value is within the numerical range representing the characteristics of the carrier signal, the audio signal corresponding to the target frequency line is determined to be an abnormal signal, and the abnormal signal is filtered out from the audio signal.

[0177] Optionally, in an exemplary embodiment, the signal detection module 808 is further used to determine the energy distribution correlation between each frequency line in the time-frequency diagram and the horizontal line template;

[0178] If the energy distribution correlation is within a preset range, it is determined that there is a candidate frequency line in the time-frequency graph that matches the horizontal line template;

[0179] The target frequency line is determined from the candidate frequency lines according to the numerical value of the energy distribution correlation.

[0180] Optionally, in an exemplary embodiment, the feature extraction module 806 is further used to obtain a normal speech database, and extract normal spectrum chaos features in the normal speech database.

[0181] The feature encoding module is used to encode the chaotic features of the normal spectrum to obtain the corresponding multi-scale feature map.

[0182] Optionally, in an exemplary embodiment, the signal detection module 808 is further used to construct a hypersphere including normal audio features according to the multi-scale feature map, wherein the hypersphere is formed by minimizing the distance between all mapping representations and the center of the hypersphere so that the normal audio features are closely clustered at the center of the hypersphere.

[0183] A feature encoding module is used to perform feature encoding on the chaotic features of the spectrum to obtain a corresponding multi-scale feature map;

[0184] Optionally, in an exemplary embodiment, the signal detection module 808 is further used to perform feature adaptation on the multi-scale feature map and calculate the Mahalanobis distance between the multi-scale feature map and the center of the hypersphere;

[0185] The abnormal results of the audio signal corresponding to the chaotic characteristics of the spectrum determined by the Mahalanobis distance include:

[0186] If the Mahalanobis distance is not within the preset threshold range, the audio signal corresponding to the spectral chaos feature is determined to be an abnormal signal.

[0187] Each module in the above-mentioned signal detection device can be implemented in whole or in part by software, hardware and a combination thereof. Each module can be embedded in or independent of a processor in a computer device in the form of hardware, or can be stored in a memory in a computer device in the form of software, so that the processor can call and execute the operations corresponding to each module above.

[0188] Based on the same inventive concept, the embodiment of the present application also provides a signal detection system for implementing the signal detection method involved above. The implementation scheme for solving the problem provided by the system is similar to the implementation scheme recorded in the above method, so the specific limitations in one or more signal detection system embodiments provided below can refer to the limitations on the signal detection method above, and will not be repeated here.

[0189] In an exemplary embodiment, Fig. 9As shown, a signal detection system is provided, the signal detection system includes an audio input device, a memory and a processor, and the processor implements the steps in the above-mentioned method embodiments when executing a computer program. Among them, the audio input device can be a microphone, or a sensor for collecting audio signals, etc. This system can be a complete independent product, such as a microphone product integrating a processing chip and a memory chip, or a microphone product connected to a processing chip, a memory chip, etc. of other products.

[0190] This signal detection system uses the carrier trace characteristics unique to out-of-band attacks to identify and filter out some high-frequency attack signals through carrier trace detection. It can identify and reject some high-energy injection signals in the first place, reduce computing costs and response time, and improve the defense efficiency of the device. And based on the different degrees of chaos caused by in-band and out-of-band signals on the voice spectrum, the spectrum chaos features are extracted through spectrum chaos entropy, and a feature memory library is built on this basis. The deep support vector is used to model and memorize the distribution of these spectrum chaos features to identify abnormal signals, effectively distinguish normal audio from various types of attack audio, and accurately detect attack audio. In addition, this method can also achieve efficient detection of various overclocked signal injection attacks such as ultrasound, lasers, and electromagnetic waves without relying on additional hardware, ensuring that the device can maintain efficient and reliable protection performance in various environments.

[0191] In an exemplary embodiment, a signal detection device is provided, the signal detection device includes an audio input device, a memory and a processor, and the processor implements the steps in the above method embodiments when executing a computer program. The signal detection device can be an audio device, and the audio device can be but is not limited to a microphone device.

[0192] In an exemplary embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as shown in FIG. Fig.10As shown. The computer device includes a processor, a memory, an input / output interface, a communication interface, a display unit and an input device. Among them, the processor, the memory and the input / output interface are connected through a device bus, and the communication interface, the display unit and the input device are connected to the device bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating device and a computer program. The internal memory provides an environment for the operation of the operating device and the computer program in the non-volatile storage medium. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be implemented through WIFI, a mobile cellular network, near field communication (Near Field Communication, NFC) or other technologies. When the computer program is executed by the processor, a signal detection method is implemented. The display unit of the computer device is used to form a visually visible picture, which can be a display screen, a projection device or a virtual reality imaging device. The display screen can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covering the display screen, or a button, trackball or touchpad set on the computer device shell, or an external keyboard, touchpad or mouse.

[0193] Those skilled in the art will understand that Fig.10 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0194] In an exemplary embodiment, a computer device is further provided, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the steps in the above-mentioned method embodiments when executing the computer program.

[0195] In an exemplary embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0196] In an exemplary embodiment, a computer program product is provided, including a computer program, and when the computer program is executed by a processor, the steps in the above method embodiments are implemented.

[0197] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.

[0198] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in the present application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. Non-relational databases may include distributed databases based on blockchains, etc., but are not limited to this. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, an artificial intelligence (AI) processor, etc., but are not limited to this.

[0199] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0200] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A signal detection method, characterized in that: The method comprises: Obtaining the audio signal to be processed; Determine a time-frequency graph corresponding to the audio signal, segment the time-frequency graph to obtain a plurality of local windows, and determine a local frequency domain feature of each local window; Performing power entropy calculation on the local frequency domain features of each local window to extract the frequency spectrum chaos features of the audio signal; Acquire a hypersphere of feature distribution of learned normal audio, and determine the Mahalanobis distance between the spectral chaos feature and the hypersphere; The abnormal result of the audio signal corresponding to the spectral chaos feature is determined according to the Mahalanobis distance.

2. The method according to claim 1, characterized in that The segmenting process of the time-frequency graph to obtain a plurality of local windows and determining the local frequency domain features of each local window includes: Determine a preset sliding window; Based on the preset sliding window, the time-frequency diagram is segmented to obtain a plurality of local windows; For each local window among the multiple local windows, data is collected by sliding each local window on the corresponding frequency spectrum to obtain local frequency domain features of each local window.

3. The method according to claim 1, characterized in that The step of performing power entropy calculation on the local frequency domain features of each local window to extract the frequency spectrum chaos features of the audio signal comprises: determining an entropy function for quantifying the degree of disorder of power energy in each of the local windows; The entropy function is applied to each of the local windows, and the power entropy is calculated for the local frequency domain characteristics of each of the local windows to generate a spectrum chaos map; the spectrum chaos map includes spectrum chaos characteristics of the signal at different times and frequencies.

4. The method according to any one of claims 1 to 3, characterized in that: Before segmenting the time-frequency graph to obtain a plurality of local windows and determining the local frequency domain features of each local window, the method further includes: Get the horizontal line template; In the case where there is a frequency line matching the horizontal line template in the time-frequency graph, determining a slope value of the target frequency line, wherein the magnitude of the slope value is used to characterize the energy distribution of the corresponding candidate frequency line; If the slope value is within the numerical range representing the characteristics of the carrier signal, it is determined that the audio signal corresponding to the target frequency line is an abnormal signal, and the abnormal signal is filtered out from the audio signal.

5. The method according to claim 4, characterized in that The determining the slope value of the target frequency line when there is a candidate frequency line matching the horizontal line template in the time-frequency graph includes: Determine the energy distribution correlation between each frequency line in the time-frequency diagram and the horizontal line template; If the energy distribution correlation is within a preset range, determining that there is a candidate frequency line in the time-frequency diagram that matches the horizontal line template; A target frequency line is determined from the candidate frequency lines according to the numerical value of the energy distribution correlation.

6. The method according to claim 5, characterized in that The step of obtaining a hypersphere of feature distribution of learned normal audio includes: Acquire a normal speech database, and extract normal frequency spectrum chaos features in the normal speech database; Performing feature encoding on the chaotic features of the normal frequency spectrum to obtain a corresponding multi-scale feature map; A hypersphere including normal audio features is constructed according to the multi-scale feature map, wherein the hypersphere is formed by minimizing the distance between all mapping representations and the center of the hypersphere so that the normal audio features are closely clustered at the center of the hypersphere.

7. The method according to claim 5, characterized in that The determining of the Mahalanobis distance between the spectral chaotic feature and the hypersphere comprises: Performing feature encoding on the chaotic features of the spectrum to obtain a corresponding multi-scale feature map; Performing feature adaptation on the multi-scale feature map, and calculating the Mahalanobis distance between the multi-scale feature map and the center of the hypersphere; The abnormal result of the audio signal corresponding to the chaotic feature of the spectrum is determined according to the Mahalanobis distance, comprising: If the Mahalanobis distance is not within a preset threshold range, it is determined that the audio signal corresponding to the frequency spectrum chaos feature is an abnormal signal.

8. A signal detection system, characterized in that: The signal detection system comprises an audio input device, a memory and a processor, wherein the audio input device is used to collect audio signals to be processed, the memory stores a computer program, and the processor implements the steps of any one of claims 1 to 7 when executing the computer program.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images