Abnormal monitoring method for computer room based on sound recognition

Abstract

The present invention relates to the technical field of computer room management, and in particular to a computer room abnormality monitoring method based on sound recognition, the method comprising: step 1: collecting sound data in a computer room; step 2: in the sound data, using a preset component decomposition model to extract multiple categories of sound data respectively, step 3: based on each category of sound data, using a preset time series prediction model to generate predicted sound data; step 4: based on the waveform of the sound data or the predicted sound data on a time scale, converting the sound data or the predicted sound data into a corresponding matrix; after feature extraction of the matrix, calculating the eigenvalue, comparing the eigenvalue with the preset value to determine whether an abnormality occurs. The present invention can comprehensively monitor the computer room environment and accurately identify various abnormalities, thereby improving the monitoring accuracy.

Description

Technical Field

[0001] The present disclosure relates to, but is not limited to, the technical field of computer room management, and in particular to a method for abnormal monitoring of a computer room based on sound recognition. Background Art

[0002] With the rapid development of information technology, a large number of computing devices and servers are concentrated in computer rooms. The normal operation of these devices is crucial to ensuring the stability and reliability of information systems. However, due to the complexity and diversity of the operating status of equipment in the computer room, the occurrence of abnormal situations may lead to serious consequences such as system crashes and data loss. Therefore, computer room abnormality monitoring technology has become the key to solving this problem. As a common physical signal, sound carries rich information about the operating status of equipment. Therefore, sound recognition technology has broad application prospects in the field of computer room abnormality monitoring.

[0003] In the past few decades, sound recognition technology has made great progress and achieved remarkable results in many fields. However, in the field of computer room anomaly monitoring, traditional sound recognition methods are often constrained by some limitations, which hinder their effectiveness in practical applications. Traditional sound recognition methods usually rely on some manually designed sound features, such as MFCC (Mel Frequency Cepstral Coefficients), energy, etc. These feature extraction methods perform well in some scenarios, but it is difficult to capture the subtle differences in the operating status of multiple devices in a complex computer room environment. Traditional sound feature extraction methods often lose some useful information during the extraction process, resulting in insufficient understanding of the integrity of sound data. This is particularly important in computer room anomaly monitoring, because slight sound changes may indicate abnormal conditions of equipment. There are many types of equipment in the computer room, and each device has its own unique sound pattern in its operating state. Traditional methods have difficulty in processing multiple sound patterns at the same time, resulting in reduced recognition performance in a diverse environment. The computer room environment is complex, with multiple sounds, such as fans and air conditioners. There is still no better way to separate different sounds from all the sounds to achieve more accurate computer room anomaly monitoring. Summary of the invention

[0004] The present disclosure provides a method for monitoring anomalies in a computer room based on sound recognition. The present invention can comprehensively monitor the computer room environment and accurately identify various anomalies, thereby improving monitoring accuracy.

[0005] In order to solve the above problems, the technical solution of the present invention is achieved as follows:

[0006] A method for monitoring abnormalities in a computer room based on sound recognition, the method comprising:

[0007] Step 1: Collect sound data in the computer room;

[0008] Step 2: From the sound data, use a preset component decomposition model to extract multiple categories of sound data, including: fan sound data, air conditioning sound data, equipment operation sound data, UPS sound data, power equipment sound data, and space reflection sound data;

[0009] Step 3: Based on each category of sound data, a preset time series prediction model is used to generate predicted sound data, including: fan sound prediction data, air conditioning sound prediction data, equipment operation sound prediction data, UPS sound prediction data, power equipment sound prediction data, and space reflection sound prediction data;

[0010] Step 4: Based on the waveform of the sound data or the predicted sound data on the time scale, convert the sound data or the predicted sound data into a corresponding matrix; after extracting features from the matrix, calculate the eigenvalue, and compare the eigenvalue with the preset value to determine whether an abnormality occurs.

[0011] Furthermore, the step 4 specifically includes:

[0012] Step 4.1: Based on the waveform of the sound data on the time scale, the sound data is converted into corresponding matrices, including: fan sound matrix, air conditioning sound matrix, equipment operation sound matrix, UPS sound matrix, power equipment sound matrix and space reflection sound matrix; the coordinates of each element in the matrix correspond to the time step, and the element value of each element corresponds to the time amplitude value under the time step;

[0013] Step 4.2: Based on the waveform of the predicted sound data on the time scale, the predicted sound data is converted into corresponding matrices, including: fan sound prediction matrix, air conditioning sound prediction matrix, equipment operation sound prediction matrix, UPS sound prediction matrix, power equipment sound prediction matrix and space reflection sound prediction matrix; the coordinates of each element in the matrix correspond to the time step, and the element value of each element corresponds to the predicted amplitude value at the time step;

[0014] Step 4.3: After feature extraction of the matrix obtained in step 4.1, calculate the eigenvalues, compare each eigenvalue with its corresponding preset value to determine whether an abnormality occurs, and if an abnormality occurs, record the result; after feature extraction of the matrix obtained in step 4.2, calculate the eigenvalues, compare each eigenvalue with its corresponding preset value to determine whether an abnormality occurs, and if an abnormality occurs, record the result.

[0015] Furthermore, the step 2 specifically includes:

[0016] Step 2.1: Decompose the sound data x(t) into a linear combination of N components, expressed as:

[0017]

[0018] Among them, a i is the coefficient, d i (t) is the component basis function, t is time; i is a subscript, and its value is an integer from 1 to 6, corresponding to fan sound data, air conditioning sound data, equipment operation sound data, UPS sound data, power equipment sound data and space reflection sound data respectively;

[0019] Step 2.2: Select a set of component basis functions with different frequency, phase, and time-frequency localization characteristics;

[0020] Step 2.3: Calculate the coefficient a by minimizing the reconstruction error i , use L1 norm regularization to increase the sparsity of components;

[0021] Step 2.4: Update the coefficient a by alternating minimization method i ;

[0022] Step 2.5: Update the component basis function parameters by minimizing the reconstruction error;

[0023] Step 2.6: Adaptively adjust λ to achieve better sparse representation according to the current reconstruction error and sparsity;

[0024] Step 2.7: Use the updated coefficient a i and component basis functions d i (t) Reconstruct the sound data, expressed as:

[0025]

[0026] in, To reconstruct sound data;

[0027] Step 2.8: Calculate the error between the reconstructed sound data and the sound data to obtain the reconstruction error; determine whether the reconstruction error reaches the convergence threshold, if not, return to step 2.5; if yes, output a i ·d i (t) and obtain different types of sound data.

[0028] Furthermore, the component basis function is expressed using the following formula:

[0029]

[0030] Among them, μ i is the window center time, σ i is the window width, f i is the frequency, φ iis the phase, which is the component basis function parameter; i is the subscript, which is used to distinguish the window center time μ corresponding to different types of sound data i , window width σ i , window frequency f i and phase φ i .

[0031] Furthermore, the coefficient a is calculated by minimizing the reconstruction error using the following formula: i , using L1 norm regularization to increase the sparsity of the components:

[0032]

[0033] Among them, λ is the sparsity control parameter, ||a i ||1 means for a i Perform L1 norm regularization.

[0034] Update the coefficient a by alternating minimization method using the following formula i :

[0035]

[0036] Here, k is the number of iterations.

[0037] Furthermore, the following formula is used to update the component basis function parameters by minimizing the reconstruction error:

[0038]

[0039] Here, ||||2 represents L2 norm regularization.

[0040] Furthermore, the step 3 specifically includes:

[0041] Step 3.1: Based on each category of sound data a i ·d i (t), extract the time-frequency features and obtain their corresponding time-frequency feature expressions Where F is the number of features and T is the number of time steps;

[0042] Step 3.2: Construct an audio conversion network G, which is a generative adversarial network consisting of a generator and a discriminator; the input of the generator is the random noise z time-frequency feature X i , the output is the predicted sound feature Y i =G(z,X i );

[0043] Step 3.3: Predict the sound feature Y i With the time-frequency feature X i Splice along the time dimension to form an extended feature matrix

[0044] Step 3.4: Align Z i Noise injection is performed to simulate the randomness in the real world; then, the temporal attention mechanism is used to i The time series information in is modeled to obtain the enhanced time series feature matrix A i ;

[0045] Step 3.5: Enhance the time series feature matrix A i Input the spectrogram recovery model and use the inverse transform operation to convert the enhanced time series feature matrix back to the predicted spectrogram S i ;

[0046] Step 3.6: By adding S i Input the sound synthesis engine to restore the spectrogram to the predicted sound data W i .

[0047] Furthermore, the step 3.1 specifically includes: using the following formula, for each type of sound data a i ·d i (t) Perform pre-emphasis processing to enhance high-frequency components and weaken low-frequency components:

[0048] y(t)=a i ·d i (t)―α·a i ·d i (t―1);

[0049] y(t) is the pre-emphasized sound data, and α is the pre-emphasis coefficient.

[0050] The pre-emphasized sound data is divided into short frames; a window function is applied to each frame to reduce the spectrum leakage effect; FFT is applied to each frame to convert the time domain signal into a frequency domain signal; in the frequency domain, a set of Mel filters are applied to the frequency domain signal to obtain the Mel filter output, and the logarithm of the Mel filter output is taken; discrete cosine transform is applied to the Mel filter output after the logarithm operation to obtain the time-frequency feature.

[0051] Furthermore, the step 3.6 specifically includes: initializing the spectrum graph using random amplitude and phase information to obtain an initial estimate S init ; Perform multiple iterations, and in each iteration, first set the initial estimate S init Perform short-time Fourier transform to obtain phase information, and then transform the predicted spectrum S i The amplitude information is combined with the phase information to obtain a new spectrum estimate S new ; Gradually optimize the spectrum estimation s through multiple iterations new, and finally obtain the synthesized spectrogram; convert the synthesized spectrogram into sound data to obtain the predicted sound data W i .

[0052] Furthermore, the synthesized spectrogram is converted into sound data using inverse short-time Fourier transform to obtain the predicted sound data W i .

[0053] The computer room abnormality monitoring method based on sound recognition of the present invention has the following beneficial effects: Different from the traditional single feature extraction method, the present invention introduces component decomposition technology to decompose the sound data generated by different devices in the computer room into multiple categories of sound components, such as fan sound, air conditioning sound, equipment operation sound, etc. This innovative method enables each device sound to be extracted and analyzed independently, avoiding the information loss that may be caused by the single feature extraction method. Through the decomposed sound components, the operating status of different devices can be more accurately reflected, and comprehensive monitoring of various devices in the computer room can be achieved. After decomposing and obtaining sound components of different categories, the present invention further uses a time series prediction model to predict the time change trend of the sound data. This technical means helps to capture the changes in the operating status of the equipment on the time scale, so as to more accurately determine whether the equipment is abnormal. By predicting the time change of the sound data, abnormal situations can be discovered early, and measures can be taken in time to avoid risks caused by abnormal situations. Through the multiple technical means of component decomposition and time series prediction, the present invention realizes comprehensive monitoring and analysis of the sound data of different devices in the computer room. Once an abnormal situation occurs, the type and degree of the abnormality can be accurately determined by comparing with the predicted sound data and the time change trend. For example, abnormal fan sounds, abnormal equipment operation sounds, etc. can be identified by analyzing different components of sound data, so that more targeted countermeasures can be taken. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 A method flow chart of a method for monitoring abnormalities in a computer room based on sound recognition provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0055] In order to make the technical problems, technical solutions and beneficial effects to be solved by the present disclosure clearer and more understandable, the present disclosure 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 disclosure and are not used to limit the present disclosure.

[0056] Example 1: Reference Figure 1 , a method for monitoring abnormalities in a computer room based on sound recognition, the method comprising:

[0057] Step 1: Collect sound data in the computer room;

[0058] Step 2: From the sound data, use a preset component decomposition model to extract multiple categories of sound data, including: fan sound data, air conditioning sound data, equipment operation sound data, UPS sound data, power equipment sound data, and space reflection sound data;

[0059] Step 3: Based on each category of sound data, a preset time series prediction model is used to generate predicted sound data, including: fan sound prediction data, air conditioning sound prediction data, equipment operation sound prediction data, UPS sound prediction data, power equipment sound prediction data, and space reflection sound prediction data;

[0060] Step 4: Based on the waveform of the sound data or the predicted sound data on the time scale, convert the sound data or the predicted sound data into a corresponding matrix; after extracting features from the matrix, calculate the eigenvalue, and compare the eigenvalue with the preset value to determine whether an abnormality occurs.

[0061] Specifically, the sound data is processed using a preset component decomposition model to separate different sound categories, such as fan sound, air conditioner sound, and equipment operation sound. The component decomposition model can decompose complex sound signals into different frequency components through signal processing techniques, such as wavelet transform and spectral analysis, and then separate different sound categories. This method helps to improve the interpretability and classification accuracy of sound data. A preset time series prediction model is introduced. Through these models, the system can predict the sound in the future based on historical sound data. This prediction capability helps to capture the periodic changes and trends of sound signals, and then more accurately judge the state inside the computer room. The sound data or predicted sound data is converted into a waveform graph on a time scale and represented as a matrix. Then, these matrices are feature extracted and eigenvalues ​​are calculated. These eigenvalues ​​can include information such as sound frequency, amplitude, energy, etc., reflecting different aspects of the sound. Then, the calculated eigenvalues ​​are compared with the preset normal values. If the eigenvalues ​​deviate significantly from the preset values, it can be determined that an abnormal situation has occurred.

[0062] Fans are common heat dissipation devices in computer rooms, and their operation will produce specific noise. By decomposing fan sound data, the operating status, speed, and possible fault conditions of the fans can be monitored to ensure the normal heat dissipation effect of the computer room. The air conditioning system is a key device to maintain a stable temperature and humidity in the computer room, and its working state will produce specific sound characteristics. Decomposing air conditioning sound data can help detect the operation of the air conditioning system, including cooling or heating effects, fan operation, etc., to ensure that the environmental conditions of the computer room meet the requirements. Different types of equipment will produce unique sound characteristics when they are running. By decomposing the equipment operation sound data, the working status of various devices (such as servers, switches, etc.) can be monitored to detect abnormalities or failures in time. UPS (uninterruptible power supply) is an important device to ensure stable power supply in the computer room, and its working status and battery status will produce specific sound patterns. Decomposing UPS sound data can help detect the battery health of UPS and ensure power supply reliability. Power equipment (such as power supplies, transformers, etc.) in the computer room also produces sound when working. Decomposing power equipment sound data helps monitor the operating status of the power system and detect power supply problems in time. The sound inside the computer room reflects and propagates on different surfaces, forming a complex acoustic environment. Decomposing spatially reflected sound data can help analyze the propagation path of sound in the room, thereby better understanding the source and propagation characteristics of sound signals.

[0063] Fan noise: Fan noise usually generates energy in the low-frequency to medium-frequency range, and its spectrum presents a continuous and uniform feature. On the spectrum graph, fan noise may appear as a relatively smooth energy distribution, concentrated in the lower frequency range. Equipment operation noise: The spectrum characteristics of equipment operation noise depend on the nature of the specific equipment. For example, the sound of hard disk reading and writing may appear as intermittent spikes on the spectrum, and its frequency is related to the hard disk rotation speed. Air conditioning noise: Air conditioning noise may generate more energy in the medium-frequency to high-frequency range, and may include air flow surging sound and air conditioning mechanical operation sound. The spectrum graph may show some characteristics of energy concentrated in the medium frequency range. UPS noise: UPS switching may cause short-term high-frequency electronic noise. On the spectrum graph, these noises may appear as momentary spikes, and their frequencies may be in the higher range. Power equipment noise: Low-frequency vibration and mechanical noise generated by power equipment such as generators and transformers may show lower-frequency energy concentration on the spectrum.

[0064] Reflection noise: Reflection noise can cause attenuation and interference on the frequency spectrum, especially if there are reflection paths in the frequency range.

[0065] Embodiment 2: Based on the previous embodiment, step 4 specifically includes:

[0066] Step 4.1: Based on the waveform of the sound data on the time scale, the sound data is converted into corresponding matrices, including: fan sound matrix, air conditioning sound matrix, equipment operation sound matrix, UPS sound matrix, power equipment sound matrix and space reflection sound matrix; the coordinates of each element in the matrix correspond to the time step, and the element value of each element corresponds to the time amplitude value under the time step; in this step, for each sound data category (such as fan sound, air conditioning sound, etc.), the original sound data is converted into a waveform on the time scale, and the waveform is converted into a corresponding matrix. In the matrix, each row represents a different time step, and each column represents a different time amplitude value of the sound data. In this way, each sound category will have a corresponding sound matrix.

[0067] Next, feature extraction is performed from these matrices to calculate a set of eigenvalues. These eigenvalues ​​may include spectral characteristics (such as frequency distribution), time domain characteristics (such as mean, standard deviation), energy characteristics, etc. These eigenvalues ​​can capture the changes and characteristics of sound data in different aspects.

[0068] Step 4.2: Based on the waveform of the predicted sound data on the time scale, the predicted sound data is converted into corresponding matrices, including: fan sound prediction matrix, air conditioning sound prediction matrix, equipment operation sound prediction matrix, UPS sound prediction matrix, power equipment sound prediction matrix and space reflection sound prediction matrix; the coordinates of each element in the matrix correspond to the time step, and the element value of each element corresponds to the predicted amplitude value at the time step; similar to step 4.1, this step performs the same operation on the predicted sound data. Convert the predicted sound data into a waveform on the time scale and represent it as a corresponding matrix. Each predicted sound data category will have a corresponding predicted sound matrix.

[0069] Likewise, eigenvalues ​​are extracted from these predicted sound matrices to calculate a set of eigenvalues ​​associated with the characteristics of the sound data.

[0070] Step 4.3: After feature extraction of the matrix obtained in step 4.1, calculate the eigenvalues, compare each eigenvalue with its corresponding preset value to determine whether an abnormality occurs, and if an abnormality occurs, record the result; after feature extraction of the matrix obtained in step 4.2, calculate the eigenvalues, compare each eigenvalue with its corresponding preset value to determine whether an abnormality occurs, and if an abnormality occurs, record the result. For the sound data matrix and predicted sound matrix obtained in steps 4.1 and 4.2, further feature extraction is performed. This can include calculating spectral characteristics, time domain characteristics, energy characteristics, etc. to obtain more representative eigenvalues.

[0071] Next, each characteristic value is compared with its corresponding preset value. The preset value is obtained through the sound data under normal operating conditions and represents the sound characteristics under normal conditions.

[0072] If any characteristic value deviates significantly from its corresponding preset value, it can be determined that an abnormal situation has occurred. These abnormal situations will be recorded, so that the system can promptly detect and report abnormal conditions inside the computer room.

[0073] Embodiment 3: Based on the previous embodiment, step 2 specifically includes:

[0074] Step 2.1: Decompose the sound data x(t) into a linear combination of N components, expressed as:

[0075]

[0076] Among them, a i is the coefficient, d i (t) is the component basis function, t is time; i is a subscript, which takes an integer from 1 to 6, corresponding to fan sound data, air conditioning sound data, equipment operation sound data, UPS sound data, power equipment sound data and space reflection sound data respectively; by decomposing the sound data into different components, each component can be associated with a specific sound category, and the characteristics of different sound categories can be extracted. This can better understand the composition and meaning of the sound data. After decomposing the sound data into multiple components, each component can be regarded as a representative of a specific sound category. This makes subsequent sound classification tasks more accurate because each component has stronger category distinguishability. Components of different sound categories have a certain stability under normal operating conditions. By analyzing the coefficient a of the component i and basis function d i (t), it is easier to detect anomalies in sound data, such as unusual sound patterns or changes. Decomposing sound data can achieve sound source separation, that is, separating sounds from different sources into their respective components. This is particularly important in multi-source environments, helping to identify and analyze the characteristics of different sound sources.

[0077] Step 2.2: Select a set of component basis functions with different frequency, phase, and time-frequency localization characteristics; different sound categories have different characteristics in the frequency, time domain, and time-frequency domain. Selecting basis functions with different frequency resolutions and time-frequency localization characteristics can better represent these characteristics in the sound data, making the decomposed components more meaningful. Selecting diverse basis functions can enhance the expressiveness of the decomposition. The diversity of basis functions can capture the diverse characteristics of different sound categories, making the decomposition more comprehensive and accurate. Appropriate basis functions can adapt to the characteristics of different sound data. The choice of basis functions should be able to adapt to the frequency range and feature changes of the sound data to ensure that the decomposition effect is not restricted. The choice of basis functions can also affect the feature extraction ability of the decomposed components. Selecting appropriate basis functions can make it easier to extract meaningful sound features from the components, such as the rhythm and frequency distribution of the sound.

[0078] Step 2.3: Calculate the coefficient a by minimizing the reconstruction error i , L1 norm regularization is used to increase the sparsity of the components; by minimizing the reconstruction error, that is, the difference between the original sound data and the reconstructed data of the linear combination of the components, the component coefficients closest to the original data can be obtained. This helps to ensure that the decomposed components can accurately represent the sound data. The introduction of L1 norm regularization can make some component coefficients zero, thereby achieving sparse representation. Sparse representation can emphasize the unique characteristics of each component and highlight its contribution to the sound data. Sparse representation can be regarded as a method of feature selection. By making some component coefficients zero, unimportant components can be excluded and only components that have an important impact on the sound data can be retained, thereby achieving more refined feature extraction. Sparse representation can achieve dimensionality reduction. Since only a few coefficients are non-zero, the dimension of the component is actually reduced to a smaller number, which helps to reduce computational costs and storage overhead. Sparse representation can enhance the ability of anomaly detection. Abnormal conditions may cause significant changes in some component coefficients, resulting in instability of sparse representation, making it easier to detect abnormal conditions.

[0079] Step 2.4: Update the coefficient a by alternating minimization method i ; Through the alternating minimization method, each coefficient a is gradually optimized i , so that the linear combination of the components can best reconstruct the original sound data. This helps to make the component representation closer to the real sound characteristics. By constantly adjusting the coefficient a i , making the reconstructed sound data closer to the original data. This helps to reduce the reconstruction error, thereby achieving more accurate sound decomposition and reconstruction. The alternating minimization method can gradually optimize each coefficient, optimizing one coefficient at a time while keeping the other coefficients unchanged. This step-by-step optimization strategy can effectively avoid falling into a local optimal solution, thereby better optimizing the overall component representation.

[0080] Step 2.5: Update the component basis function parameters by minimizing the reconstruction error;

[0081] Step 2.6: Adaptively adjust λ to achieve better sparse representation according to the current reconstruction error and sparsity;

[0082] Step 2.7: Use the updated coefficient a i and component basis functions d i (t) Reconstruct the sound data, expressed as:

[0083]

[0084] in, To reconstruct sound data;

[0085] Step 2.8: Calculate the error between the reconstructed sound data and the sound data to obtain the reconstruction error; determine whether the reconstruction error reaches the convergence threshold, if not, return to step 2.5; if yes, output a i ·d i (t) and obtain different types of sound data.

[0086] Embodiment 4: Based on the previous embodiment, the component basis function is expressed using the following formula:

[0087]

[0088] Among them, μ i is the window center time, σ i is the window width, f i is the frequency, φ i is the phase, which is the component basis function parameter; i is the subscript, which is used to distinguish the window center time μ corresponding to different types of sound data i , window width σ i , window frequency f i and phase φ i .

[0089] Specifically, μ i (Window Center Time) This parameter indicates the center position of the component basis function in time. It determines the start and end time of the component basis function, affecting the position of the component on the time axis. The window width parameter determines the range of the component basis function in time, that is, the width of the window. A larger window width can cover more time range, while a smaller window width is more concentrated around a certain time point. The window frequency parameter controls the frequency component of the component basis function. The frequency parameter f iIt determines the periodic oscillation of the component basis function, that is, its fluctuation frequency on the time axis. The phase parameter represents the initial phase of the component basis function on the time axis. It affects the starting phase angle of the component basis function, thereby affecting its shape and phase characteristics.

[0090] The first term of the component basis function It is a Gaussian function with time-domain localization characteristics, which can help the component basis function to express the local characteristics of the sound data in a certain time window. The second term of the component basis function cos(2πf i t+φ i ) is a cosine function with frequency domain oscillation characteristics, which can simulate the periodic oscillation part in sound data, such as the oscillation of sound waveform. By adjusting the parameter μ i , σ i 、f i and φ i , the position, width, frequency and phase of the component basis functions can be controlled, so as to flexibly adapt to different types of sound data characteristics.

[0091] Embodiment 5: Based on the previous embodiment, the coefficient a is calculated by minimizing the reconstruction error using the following formula: i , using L1 norm regularization to increase the sparsity of the components:

[0092]

[0093] Among them, λ is the sparsity control parameter, ||a i ||1 means for a i Perform L1 norm regularization.

[0094] Update the coefficient a by alternating minimization method using the following formula i :

[0095]

[0096] Here, k is the number of iterations.

[0097] Specifically, calculate the coefficient a i The principle of the formula is based on optimization theory, which aims to find the best coefficient a i , so that the component basis function d i (t) and its coefficient a i The sound data reconstructed by the linear combination of is as close to the original sound data as possible. At the same time, by introducing the L1 norm regularization term, the coefficients are made sparse, that is, most of the coefficients are close to zero, so as to achieve the effect of feature selection and dimensionality reduction.

[0098] Part I The goal is to minimize the Euclidean distance (mean square error) between the original sound data and the reconstructed data from the linear combination of the components. This helps ensure that the linear combination of the components can best reconstruct the original sound data, preserving its key features. i |1 is the L1 norm regularization term, which penalizes the absolute value of the coefficients and encourages most of the coefficients to be zero, thereby achieving coefficient sparsity. This sparsity helps to highlight the importance of each component, remove unimportant components, and extract key features. The sparsity constraint of the coefficients has the effect of feature selection and dimensionality reduction. By making some coefficients approach zero, components that contribute less to the sound data can be excluded, thereby achieving a more refined and compact sound feature representation. Sparsity constraints can avoid overfitting. When the sound data is of high dimension or noisy, the sparsity constraint can prevent the coefficients from over-adapting to the noise or local features, thereby improving the generalization ability of the model.

[0099] Update coefficient a i The principle of the formula is based on the alternating minimization optimization method, whose goal is to find a new coefficient value a that minimizes the data reconstruction error given other coefficients and component basis functions. i By iteratively updating each coefficient, the component representation is gradually optimized to more accurately reconstruct the original sound data.

[0100] Part I The role of is to minimize the Euclidean distance (mean square error) between the original sound data and the reconstructed data of the linear combination of the components. By optimizing the coefficients, the linear combination of the component basis function and its coefficients is as close to the original sound data as possible. The second part λ|a i |1 is the L1 norm regularization term, which penalizes the absolute value of the coefficients to make most of the coefficients zero, thus achieving coefficient sparsity. Sparsity helps to exclude components that contribute less to the sound data and extract key features. Update coefficient a i The process uses an alternating minimization method, that is, optimizing one coefficient while fixing other coefficients. Through multiple iterations of alternating updates, all coefficients are gradually optimized so that the component representation approaches the optimal solution. The alternating minimization method can avoid falling into a local optimal solution because only one coefficient is optimized in each iteration. Such a step-by-step optimization strategy helps to approach the overall optimal solution. Update coefficient a i The data reconstruction error is minimized, while the importance of components is emphasized through sparsity constraints. This helps to improve the quality of sound decomposition, making the decomposed components more accurately represent the sound data.

[0101] Calculation coefficient a iThe goal is to minimize the Euclidean distance (mean square error) between the original sound data and the reconstructed data of the linear combination of the components. By optimizing the coefficients, the reconstructed data is made as close to the original data as possible. Introducing L1 norm regularization can make the coefficient a i It has sparseness, that is, most coefficients are zero, thus emphasizing the importance of components and excluding unnecessary components. The sparsity control parameter λ controls the degree of sparsity. A larger λ value will cause more coefficients to tend to zero, thus achieving higher sparsity. By alternating minimization method, each coefficient a is gradually updated i , find the optimal solution while keeping other coefficients unchanged. This helps to gradually approach the solution of the minimization problem.

[0102] Embodiment 6: Based on the previous embodiment, the following formula is used to update the component basis function parameters by minimizing the reconstruction error:

[0103]

[0104] Here, ||||2 represents L2 norm regularization.

[0105] Specifically, The role of the part is to minimize the Euclidean distance (mean square error) between the original sound data and the reconstructed data of the linear combination of the components. By optimizing the component basis function parameters, the linear combination of the component basis function and its coefficients can better reconstruct the original sound data. By updating μ i ,σ i ,f i ,φ i These parameters are actually adjusting the shape, position, width and frequency characteristics of the component basis function. This helps the component basis function to better capture the local features and periodic oscillations in the sound data. The process of optimizing the component basis function parameters is trying to find the optimal component basis function shape so that the component basis function can best fit the local characteristics of the sound data. This helps to improve the precision and accuracy of the decomposition. By optimizing the component basis function parameters, the component basis function can be made more adaptable to the sound data, thereby reducing data reconstruction errors. This helps to make the decomposition of the sound data closer to the original data.

[0106] Embodiment 7: Based on the previous embodiment, step 3 specifically includes:

[0107] Step 3.1: Based on each category of sound data a i ·d i (t), extract the time-frequency features and obtain their corresponding time-frequency feature expressions Where F is the number of features and T is the number of time steps;

[0108] Step 3.2: Construct an audio conversion network G, which is a generative adversarial network consisting of a generator and a discriminator; the input of the generator is random noise z and time-frequency features X i , the output is the predicted sound feature Y i =G(z,X i );

[0109] Step 3.3: Predict the sound feature Y i With the time-frequency feature X i Splice along the time dimension to form an extended feature matrix

[0110] Step 3.4: Align Z i Noise injection is performed to simulate the randomness in the real world; then, the temporal attention mechanism is used to i The time series information in is modeled to obtain the enhanced time series feature matrix A i ;

[0111] Step 3.5: Enhance the time series feature matrix A i Input the spectrogram recovery model and use the inverse transform operation to convert the enhanced time series feature matrix back to the predicted spectrogram S i ;

[0112] Step 3.6: By adding S i Input the sound synthesis engine to restore the spectrogram to the predicted sound data W i .

[0113] Specifically, the predicted sound feature Y i and the time-frequency feature X i Splice along the time dimension to form a larger extended feature matrix Z i :

[0114]

[0115] Where [X i ; Y i ] represents the matrix concatenation operation, which converts the time-frequency feature X i and predicted sound features Y i Connected in time dimension.

[0116] By transforming the time-frequency feature X i The predicted sound features are spliced ​​in the time dimension, so that the model can consider both the historical information of the input and the generated information when predicting the sound features, thereby better capturing the temporal evolution of the sound. This temporal information modeling can enhance the coherence and realism of the generated sound.

[0117] In the expanded feature matrix Z iA noise injection layer is introduced on top to simulate the randomness and variation in the real world. Noise can help the generated sound more natural and realistic.

[0118] Noise injection formula: Z ′i =Z i +∈

[0119] where Z′ i is the feature matrix after noise injection, and ∈ is a noise vector sampled from a certain distribution.

[0120] Temporal Attention Mechanism:

[0121] In order to better model the temporal information of sound features, an attention mechanism can be used to enable the model to focus on the importance of different time steps.

[0122] Temporal Attention Formula:

[0123] A i [t] = Softmax(W att ·Z′ i [t]+b att );

[0124] Among them A i [t] is the attention weight calculated at time step t, W att and b att is the learned parameter, and Softmax is the soft maximization function.

[0125] Then, the attention weights can be applied to the feature matrix to obtain the enhanced temporal feature matrix A i Finally, the enhanced time series feature matrix A i It contains attention-adjusted sound feature information while taking into account noise and temporal relationships, thereby enhancing the realism and temporal coherence of the generated sound.

[0126] Embodiment 8: Based on the previous embodiment, the step 3.1 specifically includes: using the following formula, for each type of sound data a i ·d i (t) Perform pre-emphasis processing to enhance high-frequency components and weaken low-frequency components:

[0127] y(t)=a i ·d i (t)―α·a i ·d i (t―1);

[0128] y(t) is the pre-emphasized sound data, and α is the pre-emphasis coefficient.

[0129] The pre-emphasized sound data is divided into short frames; a window function is applied to each frame to reduce the spectrum leakage effect; FFT is applied to each frame to convert the time domain signal into a frequency domain signal; in the frequency domain, a set of Mel filters are applied to the frequency domain signal to obtain the Mel filter output, and the logarithm of the Mel filter output is taken; discrete cosine transform is applied to the Mel filter output after the logarithm operation to obtain the time-frequency feature.

[0130] Specifically, discrete cosine transform is applied to the Mel filter output after logarithmic operation to obtain time-frequency features using the following formula:

[0131]

[0132] Among them, Spectrum j is the output of the jth Mel filter, F is the number of filters, and i is the index of the MFCC coefficient.

[0133] Pre-emphasis is an audio processing technique that aims to balance the spectral characteristics in sound data, increase the strength of high-frequency components, and weaken the influence of low-frequency components. This is achieved by applying a differential filter to the original sound data. In practical applications, pre-emphasis helps to enhance the high-frequency characteristics in the speech signal, making it easier to analyze and decompose in subsequent processing. Pre-emphasis can make the high-frequency components of the current time step relatively stronger by reducing the low-frequency components of the previous time step. This helps to enhance the high-frequency characteristics in the sound signal, making it easier to detect and analyze. By reducing the low-frequency components, pre-emphasis can weaken the influence of low-frequency components. This helps to reduce the impact of low-frequency environmental noise in some applications and improve the clarity and distinguishability of the signal. Pre-emphasis can improve the characteristics of the sound signal in the frequency domain, making the high-frequency components more prominent. This helps in subsequent feature extraction, analysis and processing, such as spectrum analysis and sound decomposition. Pre-emphasis can enhance the high-frequency characteristics of the sound signal, helping the decomposition process to more accurately capture the details and characteristics of the sound data.

[0134] Embodiment 9: Based on the previous embodiment, step 3.6 specifically includes: using random amplitude and phase information to initialize the spectrum graph to obtain an initial estimate S init ; Perform multiple iterations, and in each iteration, first set the initial estimate S init Perform short-time Fourier transform to obtain phase information, and then transform the predicted spectrum S i The amplitude information is combined with the phase information to obtain a new spectrum estimate S new ; Gradually optimize the spectrum map estimation S through multiple iterations new , and finally obtain the synthesized spectrogram; convert the synthesized spectrogram into sound data to obtain the predicted sound data W i .

[0135] Specifically, the audio conversion network is a generative adversarial network, which specifically includes:

[0136] Generator G:

[0137] The generator accepts random noise z and input time-frequency features X i , and then try to generate the predicted sound features Y i .

[0138] Generator input: random noise z and time-frequency features X i .

[0139] Generator output: predicted sound feature Y i .

[0140] The goal of the generator is to make the discriminator unable to distinguish the generated predicted sound features Y i Is it true.

[0141] Discriminator D:

[0142] The discriminator accepts the real time-frequency feature X i And the predicted sound features Y generated by the generator i , and then try to distinguish which one is the real sound signature.

[0143] Input of the discriminator: time-frequency feature X i and predicted sound features Y i .

[0144] Output of the discriminator: The discriminator estimates that the input is the true time-frequency feature.

[0145] The goal of the discriminator is to be able to accurately distinguish between real time-frequency features and generated time-frequency features.

[0146] The game process between the generator and the discriminator can be expressed as a minimum-maximum optimization problem:

[0147]

[0148] in is the distribution of the processed historical sound feature data, z is the random noise, and Noise is the distribution of the random noise.

[0149] The loss function of the generator can be defined as the negative logarithm of the probability that the discriminator identifies the generated predicted sound features as the true sound feature matrix: The goal of the generator is to minimize the generator loss so that the generated sound feature matrix is ​​more realistic.

[0150] A iUsing the inverse attention mechanism, the attention weights are reversed to reduce the influence of attention. This helps to better restore the temporal information when performing the inverse transformation.

[0151] The feature matrix A after inverse attention adjustment i Input the inverse short-time Fourier transform to get the predicted spectrum S i .

[0152] Embodiment 10: Based on the previous embodiment, the synthesized spectrogram is converted into sound data using inverse short-time Fourier transform to obtain predicted sound data W i .

[0153] Those skilled in the art will appreciate that all or some of the steps in the methods disclosed above, and the functional modules / units in the systems and devices may be implemented as software, firmware, hardware, or a suitable combination thereof.

[0154] The preferred embodiments of the present disclosure are described above with reference to the accompanying drawings, but the scope of the present disclosure is not limited thereby. Any modification, equivalent substitution and improvement made by those skilled in the art without departing from the scope and essence of the present disclosure shall be within the scope of the present disclosure.

Claims

1. A method for monitoring abnormalities in a computer room based on sound recognition, characterized in that: The method comprises: Step 1: Collect sound data in the computer room; Step 2: From the sound data, use a preset component decomposition model to extract multiple categories of sound data, including: fan sound data, air conditioning sound data, equipment operation sound data, UPS sound data, power equipment sound data, and space reflection sound data; Step 3: Based on each category of sound data, a preset time series prediction model is used to generate predicted sound data, including: fan sound prediction data, air conditioning sound prediction data, equipment operation sound prediction data, UPS sound prediction data, power equipment sound prediction data, and space reflection sound prediction data; Step 4: Based on the waveform of the sound data or the predicted sound data on the time scale, the sound data or the predicted sound data is converted into a corresponding matrix; after extracting features from the matrix, the eigenvalue is calculated, and the eigenvalue is compared with the preset value to determine whether an abnormality occurs; The step 2 specifically includes: Step 2.1: Convert the sound data Decompose into The linear combination of components is expressed as: ； in, is the coefficient, are the component basis functions, For time; is a subscript, with a value of an integer from 1 to 6, corresponding to fan sound data, air conditioning sound data, equipment operation sound data, UPS sound data, power equipment sound data, and space reflection sound data, respectively; Step 2.2: Select a set of component basis functions with different frequency, phase, and time-frequency localization characteristics; Step 2.3: Calculate the coefficients by minimizing the reconstruction error ,use Norm regularization to increase the sparsity of components; Step 2.4: Update coefficients by alternating minimization method ; Step 2.5: Update the component basis function parameters by minimizing the reconstruction error; Step 2.6: Adaptively adjust according to the current reconstruction error and sparsity To achieve better sparse representation; Step 2.7: Use the updated coefficients and component basis functions Reconstruct the sound data, expressed as: ； in, To reconstruct sound data; Step 2.8: Calculate the error between the reconstructed sound data and the sound data to obtain the reconstruction error; determine whether the reconstruction error reaches the convergence threshold, if not, return to step 2.5; if yes, output , get different kinds of sound data; The step 3 specifically includes: Step 3.1: Based on the sound data of each category , extract the time-frequency features and obtain the corresponding time-frequency feature expression ,in is the number of features, is the number of time steps; Step 3.2: Build the Audio Conversion Network , the audio conversion network It is a generative adversarial network consisting of a generator and a discriminator; the input of the generator is random noise Time-frequency characteristics , the output is the predicted sound feature ; Step 3.3: Predict the sound features With time-frequency characteristics Splice along the time dimension to form an extended feature matrix ; Step 3.4: Noise injection is performed to simulate the randomness in the real world; then, the temporal attention mechanism is used to The time series information in is modeled to obtain the enhanced time series feature matrix ; Step 3.5: Enhance the time series feature matrix Input the spectrogram recovery model and use the inverse transform operation to convert the enhanced time series feature matrix back to the predicted spectrogram ; Step 3.6: By Input to the sound synthesis engine to restore the spectrogram to the predicted sound data ; The step 3.6 specifically includes: using random amplitude and phase information to initialize the spectrum graph to obtain an initial estimate ; Perform multiple iterations, and in each iteration, first set the initial estimate Perform short-time Fourier transform to obtain phase information, and then transform the predicted spectrum into The amplitude information is combined with the phase information to obtain a new spectrum estimate ; Gradually optimize the spectrogram estimation through multiple iterations , and finally obtain a synthesized spectrogram; convert the synthesized spectrogram into sound data to obtain the predicted sound data .

2. The method for monitoring abnormalities in a computer room based on sound recognition according to claim 1, characterized in that: The step 4 specifically includes: Step 4.1: Based on the waveform of the sound data on the time scale, the sound data is converted into corresponding matrices, including: fan sound matrix, air conditioning sound matrix, equipment operation sound matrix, UPS sound matrix, power equipment sound matrix and space reflection sound matrix; the coordinates of each element in the matrix correspond to the time step, and the element value of each element corresponds to the time amplitude value under the time step; Step 4.2: Based on the waveform of the predicted sound data on the time scale, the predicted sound data is converted into corresponding matrices, including: fan sound prediction matrix, air conditioning sound prediction matrix, equipment operation sound prediction matrix, UPS sound prediction matrix, power equipment sound prediction matrix and space reflection sound prediction matrix; the coordinates of each element in the matrix correspond to the time step, and the element value of each element corresponds to the predicted amplitude value at the time step; Step 4.3: After feature extraction of the matrix obtained in step 4.1, calculate the eigenvalues, compare each eigenvalue with its corresponding preset value to determine whether an abnormality occurs, and if an abnormality occurs, record the result; after feature extraction of the matrix obtained in step 4.2, calculate the eigenvalues, compare each eigenvalue with its corresponding preset value to determine whether an abnormality occurs, and if an abnormality occurs, record the result.

3. The method for monitoring abnormalities in a computer room based on sound recognition according to claim 2, characterized in that: The component basis function is expressed using the following formula: ； in, is the window center time, is the window width, is the frequency, is the phase, and both are component basis function parameters; Subscript, used to distinguish the window center time corresponding to different types of sound data , Window Width , window frequency and Phase .

4. The method for monitoring abnormalities in a computer room based on sound recognition according to claim 3, characterized in that: The coefficients are calculated by minimizing the reconstruction error using the following formula: ,use Norm regularization to increase component sparsity: ； in, is the sparsity control parameter, Express conduct Norm regularization; Update the coefficients by alternating minimization using the following formula : ； in, is the number of iterations.

5. The method for monitoring abnormalities in a computer room based on sound recognition according to claim 4, characterized in that: The component basis function parameters are updated by minimizing the reconstruction error using the following formula: ； in, express Norm regularization.

6. The method for monitoring abnormalities in a computer room based on sound recognition according to claim 5, characterized in that: The step 3.1 specifically includes: using the following formula for each type of sound data Perform pre-emphasis to enhance high-frequency components and weaken low-frequency components: ； It is the sound data after pre-emphasis. is the pre-emphasis coefficient; The pre-emphasized sound data is divided into short frames; a window function is applied to each frame to reduce the spectrum leakage effect; FFT is applied to each frame to convert the time domain signal into a frequency domain signal; in the frequency domain, a set of Mel filters are applied to the frequency domain signal to obtain the Mel filter output, and the logarithm of the Mel filter output is taken; discrete cosine transform is applied to the Mel filter output after the logarithm operation to obtain the time-frequency feature.

7. The method for monitoring abnormalities in a computer room based on sound recognition according to claim 6, characterized in that: Use inverse short-time Fourier transform to convert the synthesized spectrogram into sound data to obtain the predicted sound data .

Images