Acoustic sensor array data anomaly detection method based on multi-auditory-angle adaptive sparse semi-nonnegative matrix factorization
An anomaly detection method for acoustic sensor array data based on multi-angle adaptive sparse semi-nonnegative matrix factorization and Gaussian distribution estimation solves the problems of shallow information fusion and redundancy in multi-angle data processing, and achieves high-precision and robust anomaly detection.
Patent Information
- Application Number
- CN202511641339.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-11
- Publication Date
- 2026-02-17
AI Technical Summary
Existing acoustic detection methods suffer from insufficient depth of multi-angle information fusion and lack of active reduction mechanism for information redundancy in multi-angle data processing, resulting in inadequate detection accuracy and robustness.
An anomaly detection method for acoustic sensor array data based on multi-angle adaptive sparse semi-nonnegative matrix factorization is adopted. Through an anomaly detection network based on sparse semi-nonnegative matrix factorization and Gaussian distribution estimation, deep feature fusion and redundancy reduction are achieved, and anomaly detection is performed using sparse shared features and Gaussian distribution model.
It significantly improves detection accuracy and robustness, achieves deep end-to-end fusion and redundancy reduction, and enhances the accuracy and reliability of detection.
Smart Images
Figure CN121542948A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of anomaly detection technology for multi-angle data, specifically to an anomaly detection method for acoustic sensor array data based on multi-angle adaptive sparse semi-nonnegative matrix decomposition. Background Technology
[0002] Acoustic detection, as a non-contact technology that can simulate manual listening, has broad application prospects in industrial anomaly detection. Although the scheme based on a single acoustic sensor and deep learning has achieved good results under laboratory conditions, its detection accuracy and robustness under real-world complex working conditions still need to be improved.
[0003] The inherent limitations of a single sensor layout are a key constraint on performance. Sound signals attenuate during propagation due to physical effects such as obstruction, scattering, and medium absorption, and the degree of attenuation varies with frequency. This can easily lead to incomplete or biased information captured by a single listening angle. To address this issue, using acoustic sensor arrays to collaboratively acquire signals from multiple listening angles has become an effective technical approach. This method not only compensates for the signal loss of a single channel through spatial diversity, but the spatial information of the sound sources it contains also helps distinguish target sound sources within the array's enclosed area from external environmental noise, thereby fundamentally improving the signal-to-noise ratio.
[0004] However, the introduction of multi-angle data also brings new technical challenges: how to effectively reduce inherent data redundancy while deeply fusing multi-source information. Current mainstream methods mostly employ a "voting"-style post-fusion strategy, where each angle of hearing data is detected and judged independently before integrating multiple local results. This method has fundamental shortcomings: first, its fusion level remains at the decision-making stage, failing to achieve deep fusion at the data or feature level and thus unable to fully explore the complementary and synergistic relationships between multiple perspectives; second, it completely ignores the potentially high correlation between different angles of hearing data and lacks an active reduction mechanism for information redundancy. These two deficiencies severely limit further performance improvements of existing methods.
[0005] Therefore, improvements are needed. Summary of the Invention
[0006] To address the technical problems of multi-angle information fusion and information redundancy in current acoustic detection methods for multi-angle data processing, this application provides an acoustic sensor array data anomaly detection method based on multi-angle adaptive sparse semi-nonnegative matrix decomposition.
[0007] The objective of this invention is achieved through the following technical solution:
[0008] An acoustic sensor array data anomaly detection method based on multi-angle adaptive sparse semi-nonnegative matrix factorization includes:
[0009] S1, Assumption It is composed of n v A normal array of data is collected by an acoustic sensing array consisting of several acoustic sensors. This represents the v-th listening angle data, 1≤v≤n v , The dimension is 1×m, using batch normal array data. An anomaly detection network W based on multi-angle sparse semi-nonnegative matrix factorization and Gaussian distribution estimation is trained to obtain an anomaly detection model M;
[0010] The structure of the anomaly detection network W described in S1, based on multi-angle sparse semi-nonnegative matrix factorization and Gaussian distribution estimation, is as follows:
[0011] S11, Input Layer: Input n normal array data ;
[0012] S12. Extracting Feature Data Layer: Calculating Normal Array Data The features of each listening angle data are used to obtain the feature array data c, where , This indicates that the data originates from the v-th hearing angle. Feature data, The dimension is 1×p;
[0013] S13, Feature Data Decomposition and Sparse Shared Feature Computation Layer:
[0014] S131, When using a diagonal matrix (1) When sparse constraints are applied, the formula is used. (2) For each feature data in the feature array data c Perform sparse semi-nonnegative matrix decomposition to obtain the corresponding coefficient matrix. and basis matrix ,in The dimension is n×q. The dimension is q×p;
[0015] S132, Using the formula (3) Calculate feature data The sparse shared features of the feature array data c are obtained by using these features. ;
[0016] S133, In formula (1), It is the identity matrix. It is a column vector with all elements equal to 1, and the sparsity constraint parameters are... and Controlling the coefficient matrix and basis matrix The sparsity, and is limited during network training. , , ;
[0017] S134. In formula (3), Represents the data of the v-th hearing angle. The weight, It is a diagonal matrix, according to the formula (4) Perform calculations to obtain the sparse shared feature matrix. sparse shared feature matrix The size is n×q, and the row vector is... The dimension is 1×q;
[0018] S13 further includes the following steps:
[0019] S135. Initialize all basis matrices for multi-angle sparse semi-nonnegative matrix decomposition. , The initial values are in the range (-1, 1), and all coefficient matrices of the multi-angle sparse semi-nonnegative matrix decomposition are initialized. , The initial values are in the range (0,1), and the sparse shared feature matrix of each hearing angle data is initialized. , The initial value is in the range (0,1);
[0020] S136. Using the formula (13) and formula (3), for the coefficient matrix Sparse shared feature matrix Perform several alternating iterative updates to achieve the goal of updating the feature data. Decomposed into coefficient matrix and basis matrix And obtain the sparse shared feature matrix of the feature array data. ;
[0021] S137. In formula (13), express Positive elements in express Positive elements in;
[0022] S14, Gaussian distribution parameter estimation and Gaussian density calculation layer:
[0023] S141. Using the formula (5) Calculate the mean parameter of the Gaussian distribution model. ;
[0024] S142, Using the formula (6) Calculate the covariance parameter of the Gaussian distribution model. ;
[0025] S143, Using the formula (7) Calculate the logarithmic probability density value L;
[0026] S144. In formulas (5), (6) and (7), The dimension is 1×q. Indicates calculation The value of the determinant, express The inverse matrix, The expression asks for the logarithm to base 2. yes transpose;
[0027] The loss function used in training the anomaly detection network W, which is based on multi-angle sparse semi-nonnegative matrix factorization and Gaussian distribution estimation, as described in SA and S1, is... for: (8), In formula (8), E is derived from formula (9) The calculation yields, where This indicates that Frobenius norm operations are being performed. From the formula (10) Calculations show that From the formula (11) The calculation shows that, From the formula (12) The calculation shows that , and These are weight control parameters;
[0028] The training process of the anomaly detection network W based on multi-angle sparse semi-nonnegative matrix factorization and Gaussian distribution estimation described in S1 is as follows:
[0029] SE1. Randomly select n normal array data from the batch of normal data. Input network W, according to the loss function The weights of network W are trained using the principle of minimization;
[0030] SE2, after repeating the training process SE1 several times, the anomaly detection model M is obtained;
[0031] S2. Input the array data y to be detected into the anomaly detection model M to obtain the anomaly detection result R, where This represents the array data to be detected collected by the sensor array. The dimension is 1×m1, representing the data to be inspected collected by the v-th sensor, 1≤v≤n. v；
[0032] In the anomaly detection network W described in S3, S1, and S2, for each hearing angle data... Basis matrix for sparse semi-nonnegative matrix decomposition sparse constraint parameters and and the mean parameter of the Gaussian distribution model. Covariance parameter These are the network weights that need to be trained and optimized as the anomaly detection network is trained.
[0033] In a preferred embodiment, step S12, extracting the feature data layer, involves calculating the normal array data. The features of each listening angle data are used to obtain the feature array data c, where , This indicates that the data originates from the v-th hearing angle. Feature data, The dimension is 1×p, including:
[0034] S121, the feature data It can be from the data of the vth listening angle. The features calculated in the middle can also be used to calculate the v-th listening angle data. As feature data .
[0035] In a preferred embodiment, step S2 involves inputting the array data y to be detected into the anomaly detection model M to obtain the anomaly detection result R, wherein... This represents the array data to be detected collected by the sensor array. The dimension is 1×m1, representing the data to be detected collected by the v-th sensor, 1≤v≤n. v， include:
[0036] S21. Input the array data y to be detected into the anomaly detection model M to obtain the Gaussian log probability density value L;
[0037] S22. Based on the relationship between the logarithmic probability density value L of the Gaussian distribution and the threshold δ, the anomaly detection result R is obtained. If L≤δ, then the anomaly detection result R is the array data y to be detected, indicating anomaly. If p>δ, then the anomaly detection result R is the array data y to be detected, indicating normality.
[0038] In a preferred embodiment, step S22, obtaining the anomaly detection result R based on the relationship between the Gaussian distribution log probability density value L and the threshold δ, wherein if L≤δ, the anomaly detection result R indicates that the array data y to be detected is abnormal; and if L>δ, the anomaly detection result R indicates that the array data y to be detected is normal, includes:
[0039] The steps for obtaining the threshold δ are as follows:
[0040] S221. Input all normal array data x into the anomaly detection model M in sequence to obtain the corresponding log probability density value L. j ;
[0041] S222, Find all L j In Quantiles are used as the threshold δ.
[0042] In summary, this application includes at least one of the following beneficial technical effects:
[0043] This method achieves deep fusion of multi-angle data by performing adaptive sparse semi-nonnegative matrix factorization at the feature level, fundamentally changing the shallow mode of existing "voting" fusion, realizing end-to-end deep fusion and redundancy reduction, and significantly improving the detection accuracy and robustness. Attached Figure Description
[0044] Figure 1 This is a flowchart of an embodiment of step S13 of this application;
[0045] Figure 2 This is a flowchart of an embodiment of step S2 of this application. Detailed Implementation
[0046] The following is in conjunction with the appendix Figure 1-2 This application will be described in further detail.
[0047] In one embodiment, this application discloses a method for detecting anomalies in acoustic sensor array data based on multi-angle adaptive sparse semi-nonnegative matrix factorization, which specifically includes the following steps:
[0048] S1, Assumption It is composed of n v A normal array of data is collected by an acoustic sensing array consisting of several acoustic sensors. This represents the v-th listening angle data, 1≤v≤n v , The dimension is 1×m, using batch normal array data. An anomaly detection network W based on multi-angle sparse semi-nonnegative matrix factorization and Gaussian distribution estimation is trained to obtain an anomaly detection model M;
[0049] The structure of the anomaly detection network W described in S1, based on multi-angle sparse semi-nonnegative matrix factorization and Gaussian distribution estimation, is as follows:
[0050] S11, Input Layer: Input n normal array data ;
[0051] S12. Extracting Feature Data Layer: Calculating Normal Array Data The features of each listening angle data are used to obtain the feature array data c, where , This indicates that the data originates from the v-th hearing angle. Feature data, The dimension is 1×p;
[0052] like Figure 1 As shown, S13, the feature data decomposition and sparse shared feature computation layer:
[0053] S131, When using a diagonal matrix (1) When sparse constraints are applied, the formula is used. (2) For each feature data in the feature array data c Perform sparse semi-nonnegative matrix decomposition to obtain the corresponding coefficient matrix. and basis matrix ,in The dimension is n×q. The dimension is q×p;
[0054] S132, Using the formula (3) Calculate feature data The sparse shared features of the feature array data c are obtained by using these features. ;
[0055] S133, In formula (1), It is the identity matrix. It is a column vector with all elements equal to 1, and the sparsity constraint parameters are... and Controlling the coefficient matrix and basis matrix The sparsity, and is limited during network training. , , ;
[0056] S134. In formula (3), Represents the data of the v-th hearing angle. The weight, It is a diagonal matrix, according to the formula (4) Perform calculations to obtain the sparse shared feature matrix. sparse shared feature matrix The size is n×q, and the row vector is... The dimension is 1×q;
[0057] S13 further includes the following steps:
[0058] S135. Initialize all basis matrices for multi-angle sparse semi-nonnegative matrix decomposition. , The initial values are in the range (-1, 1), and all coefficient matrices of the multi-angle sparse semi-nonnegative matrix decomposition are initialized. , The initial values are in the range (0,1), and the sparse shared feature matrix of each hearing angle data is initialized. , The initial value is in the range (0,1);
[0059] S136. Using the formula (13) and formula (3), for the coefficient matrix Sparse shared feature matrix Perform several alternating iterative updates to achieve the goal of updating the feature data. Decomposed into coefficient matrix and basis matrix And obtain the sparse shared feature matrix of the feature array data. ;
[0060] S137. In formula (13), express Positive elements in express Positive elements in;
[0061] S14, Gaussian distribution parameter estimation and Gaussian density calculation layer:
[0062] S141. Using the formula (5) Calculate the mean parameter of the Gaussian distribution model. ;
[0063] S142, Using the formula (6) Calculate the covariance parameter of the Gaussian distribution model. ;
[0064] S143, Using the formula (7) Calculate the logarithmic probability density value L;
[0065] S144. In formulas (5), (6) and (7), The dimension is 1×q. Indicates calculation The value of the determinant, express The inverse matrix, The expression asks for the logarithm to base 2. yes transpose;
[0066] The loss function used in training the anomaly detection network W, which is based on multi-angle sparse semi-nonnegative matrix factorization and Gaussian distribution estimation, as described in SA and S1, is... for:
[0067] (8), In formula (8), E is derived from formula (9) The calculation yields, where This indicates that Frobenius norm operations are being performed. From the formula (10) Calculations show that From the formula (11) The calculation shows that, From the formula (12) The calculation shows that , and These are weight control parameters;
[0068] The training process of the anomaly detection network W based on multi-angle sparse semi-nonnegative matrix factorization and Gaussian distribution estimation described in S1 is as follows:
[0069] SE1. Randomly select n normal array data from the batch of normal data. Input network W, according to the loss function The weights of network W are trained using the principle of minimization;
[0070] SE2, after repeating the training process SE1 several times, the anomaly detection model M is obtained;
[0071] S2. Input the array data y to be detected into the anomaly detection model M to obtain the anomaly detection result R, where This represents the array data to be detected collected by the sensor array. The dimension is 1×m1, representing the data to be detected collected by the v-th sensor, 1≤v≤n. v ;
[0072] In the anomaly detection network W described in S3, S1, and S2, for each hearing angle data... Basis matrix for sparse semi-nonnegative matrix decomposition sparse constraint parameters and and the mean parameter of the Gaussian distribution model. Covariance parameter These are the network weights that need to be trained and optimized as the anomaly detection network is trained.
[0073] In this embodiment, the method achieves deep fusion of multi-angle listening data through adaptive sparse semi-nonnegative matrix factorization at the feature level. Its core principle lies in decomposing feature data from different listening angles into a shared sparse feature matrix and an angle-specific basis matrix, thereby automatically learning and extracting common features that are sensitive to anomalies and resistant to information redundancy. Subsequently, a Gaussian distribution is used to model the sparse shared features under normal conditions, providing a stable and reliable probabilistic criterion for anomaly detection. This method fundamentally changes the shallow "voting" fusion mode of existing methods, achieving end-to-end deep fusion and redundancy reduction, significantly improving detection accuracy and robustness.
[0074] S12 includes:
[0075] S121, the feature data It can be from the data of the vth listening angle. The features calculated in the middle can also be used to calculate the v-th listening angle data. As feature data .
[0076] This embodiment provides a flexible feature definition approach. The principle is to allow both the use of computed, more representative high-level features and the direct use of raw auditory angle data as feature input. This design endows the model with strong adaptability, enabling it to select the most suitable feature input based on the needs and data characteristics of different detection tasks, thereby achieving optimal detection performance while ensuring the method's versatility.
[0077] like Figure 2 S2 includes:
[0078] S21. Input the array data y to be detected into the anomaly detection model M to obtain the Gaussian log probability density value L;
[0079] S22. Based on the relationship between the logarithmic probability density value L of the Gaussian distribution and the threshold δ, the anomaly detection result R is obtained. If L≤δ, then the anomaly detection result R is the array data y to be detected, indicating anomaly. If L>δ, then the anomaly detection result R is the array data y to be detected, indicating normality.
[0080] In this embodiment, probabilistic and interpretable anomaly decision-making is implemented. The principle is to use a trained Gaussian distribution model to calculate the log probability density of sparsely shared features in the data to be tested. This probability density value intuitively reflects the degree of agreement between the tested data and normal data patterns. By comparing this probability density with a probability threshold, abstract high-dimensional feature differences can be transformed into explicit "normal / abnormal" binary judgments, making the detection results not only accurate but also statistically interpretable.
[0081] S22, including:
[0082] The steps for obtaining the threshold δ are as follows:
[0083] S221. Input all normal array data x into the anomaly detection model M in sequence to obtain the corresponding log probability density value L. j ;
[0084] S222, Find all L j In Quantiles are used as the threshold δ.
[0085] This embodiment provides a data-driven, objective method for automatically determining thresholds. Its principle is based on the probability density distribution of all normal training data on the model, selecting an appropriate quantile as the threshold. This ensures that the threshold setting relies entirely on the statistical characteristics of normal data, rather than subjective assumptions, and can adaptively match the normal fluctuation range under different devices and environments, effectively reducing the false alarm rate and improving the automation and reliability of the detection system.
[0086] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.
[0087] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.
Claims
1. An acoustic sensor array data anomaly detection method based on multi-listening-angle adaptive sparse semi-non-negative matrix factorization, characterized in that, Comprising the following steps: S1, hypothesis is a normal array data collected by an acoustic sensor array composed of n v acoustic sensors, represents the vth listening angle data, 1≤v≤n v , The dimension of is 1×m, and the batch normal array data is used to train the abnormal detection network W based on multi-listening angle sparse semi-non-negative matrix decomposition and Gaussian distribution estimation, and an abnormal detection model M is obtained. The structure of the abnormality detection network W based on multi-listen-angle sparse semi-non-negative matrix factorization and Gaussian distribution estimation in S1 is: S11, input layer: input n normal array data ; S12, extracting feature data layer: calculating normal array data characteristics of each angle data, thereby obtaining feature array data c, wherein , represents the feature data derived from the vth angle data , The dimension of is 1xp; S13, feature data decomposition and sparse shared feature calculation layer: S131、in the case of using a diagonal matrix (1) In the case of sparse constraint, use the formula (2) For each feature data in the feature array data c Sparse semi-non-negative matrix decomposition is performed to obtain the corresponding coefficient matrix and the base matrix , wherein The dimension of is n×q, The dimension of is q×p; S132, using the formula (3) calculating the sparse shared features of the characteristic data , thereby obtaining the sparse shared feature matrix of the characteristic array data c ; S133, in formula (1), is a unit matrix, is a column vector with all elements being 1, and a sparse constraint parameter and control the sparsity of the coefficient matrix and the basis matrix , and limit , , during network training. S134, in formula (3), represents the weight of the vth listening angle data is a diagonal matrix, calculated according to formula (4), the sparse shared feature matrix , the sparse shared feature matrix has a size of n x q, and the row vector has a dimension of 1 x q; The S13 further comprises the following steps: S135, initializing all basis matrices of multi-listener-angle sparse semi-non-negative matrix factorization , initial values of all coefficient matrices of multi-listener-angle sparse semi-non-negative matrix factorization are in the interval (-1, 1) , initial values of the sparse shared feature matrices of each listener-angle data are in the interval (0, 1) , initial values of the sparse shared feature matrices of each listener-angle data are in the interval (0, 1) S136、Utilizing formula (13) and formula (3), the coefficient matrix and the sparse shared feature matrix are updated for several times of alternating iterations to realize the decomposition of the feature data into the coefficient matrix and the basis matrix , and the sparse shared feature matrix of the feature array data is obtained; S137、In formula (13), denotes a positive value element in denotes a positive value element in S14, Gaussian distribution parameter estimation and Gaussian density calculation layer: S141, utilize the formula (5) Calculate the mean parameter of the Gaussian distribution model ; S142, using the formula (6) Calculate the covariance parameter of the Gaussian distribution model ; S143, using the formula (7) Calculate the log probability density value L; S144, in the formulas (5), (6) and (7), is 1 x q, represents a value of a determinant of represents an inverse matrix of represents a logarithm with base 2, is a transpose of ; The loss function used in training of the anomaly detection network W based on multi-listening-angle sparse semi-non-negative matrix decomposition and Gaussian distribution estimation in SA and S1 is: (8) in formula (8), E is calculated by formula (9), wherein The Frobenius norm operation is represented by is calculated by formula (10), is calculated by formula (11), is calculated by formula (12), , and is a weight control parameter; The training process of the abnormality detection network W based on multi-listen-angle sparse semi-non-negative matrix factorization and Gaussian distribution estimation in S1 is: SE1, in batch normal data, randomly select n normal array data input network W, according to the loss function The principle of minimization trains the weights of network W; SE2, after repeating the training process SE1 for several times, an abnormality detection model M is obtained; S2, input the array data to be detected y into the anomaly detection model M, and obtain an anomaly detection result R, wherein represents an array data to be detected collected by a sensor array, The dimension of is 1×m1, which represents the detected data collected by the vth sensor, 1≤v≤n v ; The abnormality detection network W in S3, S1 and S2 is trained for each listening angle data The basis matrix for sparse semi-non-negative matrix factorization The sparse constraint parameter And The mean parameter of the Gaussian distribution model And the covariance parameter The network weights that need to be trained are optimized following the training of the abnormality detection network.
2. The acoustic sensor array data anomaly detection method based on multi- listening angle adaptive sparse semi-non-negative matrix factorization according to claim 1, characterized in that, The S12 comprises: S121、 the feature data may be a feature calculated from the vth listening angle data may also take the vth listening angle data as feature data .
3. The acoustic sensor array data anomaly detection method based on multi- listening angle adaptive sparse semi-non-negative matrix factorization according to claim 1, characterized in that, The S2 comprises: S21, inputting the array data y to be detected into the abnormality detection model M to obtain a Gaussian distribution logarithmic probability density value L; S22, obtaining an abnormality detection result R according to the size relationship between the Gaussian distribution logarithmic probability density value L and a threshold value δ, if L≤δ, the abnormality detection result R is that the array data y to be detected represents abnormality, if L>δ, the abnormality detection result R is that the array data y to be detected represents normality.
4. The acoustic sensor array data anomaly detection method based on multi- listening angle adaptive sparse semi-non-negative matrix factorization according to claim 3, characterized in that, The S22 comprises: The obtaining step of the threshold value δ is: S221. Input all normal array data x into the anomaly detection model M in sequence to obtain the corresponding log probability density value L. j ; S222, find all L j Quantiles as thresholds δ.