Acoustic sensor array data anomaly detection method based on multi-auditory-angle semi-nonnegative matrix factorization

Abstract

The invention relates to the technical field of anomaly detection of multi-auditory-angle data, in particular to an acoustic sensor array data anomaly detection method based on multi-auditory-angle semi-nonnegative matrix factorization. The method comprises the following steps: performing joint training on multi-auditory-angle semi-nonnegative matrix factorization and Gaussian distribution estimation by using a batch of normal array data to obtain an anomaly detection model; and inputting the to-be-detected array data into the anomaly detection model to obtain a detection result representing whether the to-be-detected data is abnormal or not. The anomaly detection network designed by the invention can map each auditory angle data to the same low-dimensional shared feature through a multi-auditory angle semi-nonnegative matrix factorization model, realizes multi-auditory angle information fusion and reduces information redundancy, and then inputs the low-dimensional shared feature into a Gaussian distribution model for anomaly discrimination. According to the method, the problems of multi-auditory-angle information fusion and information redundancy in multi-auditory-angle data processing of a current acoustic detection method can be effectively solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of anomaly detection of multi-listening-angle data, and particularly relates to an acoustic sensor array data anomaly detection method based on multi-listening-angle semi-non-negative matrix factorization. BACKGROUND

[0002] The acoustic detection method can completely imitate artificial listening detection, and belongs to non-contact detection, and has high application potential in the field of anomaly detection. In recent years, many scholars have carried out related research based on a single acoustic sensor (single listening angle) and deep learning, and good results have been achieved in the laboratory stage. However, the promotion and application of the acoustic detection method still needs to further improve the accuracy and robustness.

[0003] During the propagation of sound signals, signal attenuation may occur due to shielding, medium absorption and the like, and the attenuation of different frequency components is inconsistent, which may lead to insufficient information due to poor positioning of a single acoustic sensor. Using an acoustic sensor array to listen to sound information from multiple angles can increase the amount of information obtained, which is an effective method to solve the above problems. Moreover, multi-listening-angle data also contains spatial position information of the sound source, which helps to resist the influence of environmental noise, because environmental noise is generally outside the area surrounded by the acoustic sensor array. However, multi-listening-angle data processing needs to solve how to realize the fusion of multi-listening-angle information while reducing information redundancy.

[0004] At present, the method for performing anomaly detection using multi-listening-angle data generally fuses the detection results of single-listening-angle data in a "voting" manner to obtain the final detection result. The related method does not place deep fusion of multi-listening-angle information in the process of multi-listening-angle data processing, nor does it remove the information redundancy existing in multi-listening-angle data.

[0005] Therefore, improvement is needed. SUMMARY

[0006] To solve the above technical problems, the present application provides an acoustic sensor array data anomaly detection method based on multi-listening-angle semi-non-negative matrix factorization.

[0007] To achieve the above purpose, the present application adopts the following technical solutions:

[0008] An acoustic sensor array data anomaly detection method based on multi-listening-angle semi-non-negative matrix factorization comprises the following steps:

[0009] S1, assuming is a normal array data collected by an acoustic sensor array composed of n v acoustic sensors, wherein x (v) represents the vth listening-angle data, 1≤v≤n v , x (v)The dimension is 1×m. The anomaly detection network W based on multi-angle semi-nonnegative matrix factorization and Gaussian distribution estimation is trained using batch normal array data x to obtain the anomaly detection model M.

[0010] S2. Input y into the anomaly detection model M to obtain the anomaly detection result R, where Indicates that n v An acoustic sensing array composed of several acoustic sensors collects data from an array under test, where y (v) This represents the v-th listening angle data to be detected, where 1 ≤ v ≤ n. v y (v) The dimension is 1×m1.

[0011] In a preferred embodiment, S1 includes:

[0012] The structure of the anomaly detection network W is as follows:

[0013] S11, Input Layer: Input n normal array data x;

[0014] S12. Extract feature data layer: Calculate the auditory angle data x for each auditory angle in the normal array data x. (v) The features are used to obtain feature array data. Where c (v) This indicates that the data originates from the v-th hearing angle. (v) Feature data, 1≤v≤n v c (v) The dimension is 1×p;

[0015] S13, Feature Data Decomposition and Shared Feature Calculation Layer:

[0016] S131, Using formula c (v) =G (v) F (v) (1) For each feature data c in the feature array data c (v) Perform semi-nonnegative matrix decomposition to obtain the corresponding coefficient matrix G. (v) and basis matrix F (v) G (v) The dimension is n×q, F (v) The dimension is q×p;

[0017] S132, Using the formula The shared features of normal array data x are calculated, thus obtaining the shared feature matrix G of normal array data x. * ;

[0018] S133, In formula (2), λ v Q represents the weight of the v-th listening angle data. (v) It is a diagonal matrix, according to the formula Perform calculations, matrix Matrix G * The size is n×q, and the row vector is g. k The dimension is 1×q;

[0019] S14, Gaussian distribution parameter estimation and Gaussian density calculation layer:

[0020] S141. Using the formula Calculate the mean parameter μ of the Gaussian distribution model;

[0021] S142, Using the formula Calculate the covariance parameter ε of the Gaussian distribution model;

[0022] S143, Using the formula Calculate the logarithmic probability density value L;

[0023] S144. In formulas (4), (5), and (6), the dimension of μ is 1×q, |ε| represents the value of the determinant of ε, and ε -1 Let G denote the inverse matrix of ε, log(·) represents the base-2 logarithm, (G * -μ) T Yes (G) * Transpose of -μ).

[0024] In a preferred embodiment, S1 includes:

[0025] The training process of the anomaly detection network W is as follows:

[0026] SA1. In a batch of normal array data, randomly select n normal array data x and input them into the anomaly detection network W. Train the weights of the anomaly detection network W according to the principle of minimizing the loss function Ω.

[0027] SA2, after repeating the training process SA1 several times, the anomaly detection model M is obtained.

[0028] In a preferred embodiment, the SA1 includes:

[0029] SB1, The loss function Ω used for training the anomaly detection network W is: Ω=E-λ1L1+λ2L2(7);

[0030] SB2, where E is given by formula The calculation yields that ||·|| F This indicates that the Frobenius norm operation is performed, and L1 is given by the formula. Calculations show that L2 is obtained from the formula The calculations show that λ1 and λ2 are the weight control parameters.

[0031] In a preferred embodiment, S2 includes:

[0032] S21. Input the array data y to be inspected into the anomaly detection model M to obtain the Gaussian log probability density value L;

[0033] S22. Based on the relationship between the logarithmic probability density value L of the Gaussian distribution and the threshold δ, the abnormal detection result R is obtained. If L≤δ, the result R indicates that the array data y to be inspected is abnormal. If L>δ, the result R indicates that the array data y to be inspected is normal.

[0034] In a preferred embodiment, S12 includes:

[0035] S121, the feature data c (v) Including from the hearing angle data x (v) The features calculated in the middle and the hearing angle data x (v) .

[0036] In a preferred embodiment, S13 further includes:

[0037] SE1, Initialize all basis matrices F of the multi-angle semi-nonnegative matrix decomposition. (v) F (v) The initial value is in the range (-1, 1);

[0038] SE2, Initialize all coefficient matrices G of the multi-angle semi-nonnegative matrix decomposition. (v) G (v) The initial value is in the range (0,1);

[0039] SE3, Initialize the shared feature matrix G for each hearing angle data. * G * The initial value is in the range (0,1);

[0040] SE4. Use formula And Equation (2), for the coefficient matrix G (v) and shared feature matrix G * Perform several alternating iterative updates to decompose the feature data c into matrix G. (v) sum matrix F (v) And obtain the shared feature matrix G of the sensor array data. * ;

[0041] SE5. In formula (11), [·] + This represents the positive elements within the brackets [·]. - This represents the positive elements in [·].

[0042] In the anomaly detection network W described in SE6, S1, and S2, for each hearing angle data x(v) The basis matrix F for semi-nonnegative matrix decomposition (v) The mean parameter μ and covariance parameter ε of the Gaussian distribution model are the network weights that need to be trained and are optimized as the anomaly detection network is trained.

[0043] In a preferred embodiment, S22 includes:

[0044] The steps for obtaining the threshold δ are as follows:

[0045] 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 ;

[0046] S222, Find all L j n L Quantiles are used as the threshold δ.

[0047] In summary, this application includes at least one of the following beneficial technical effects:

[0048] The anomaly detection network designed in this application maps data from various auditory angles to the same low-dimensional shared feature using a multi-angle semi-nonnegative matrix factorization (MNF) model. This achieves multi-angle information fusion and reduces information redundancy. The low-dimensional shared feature is then input into a Gaussian distribution model for anomaly detection. Furthermore, this network achieves joint training of MNF and Gaussian distribution estimation, enabling simultaneous optimization of MNF shared feature extraction and Gaussian density estimation. This ensures that more data information beneficial for anomaly diagnosis is retained in the low-dimensional shared feature, improving the effectiveness of unsupervised anomaly diagnosis. This effectively addresses the problems of multi-angle information fusion and information redundancy in current acoustic detection methods for multi-angle data processing. Attached Figure Description

[0049] Figure 1 This is a flowchart illustrating the workflow of an acoustic sensor array data anomaly detection method based on multi-angle semi-nonnegative matrix decomposition according to this application.

[0050] Figure 2 This is a schematic diagram of the structure and process of the anomaly detection network W involved in this application. Detailed Implementation

[0051] like Figures 1-2 This application will be described in further detail.

[0052] In one embodiment, such as Figure 1 As shown, this application discloses a method for detecting anomalies in acoustic sensor array data based on multi-angle semi-nonnegative matrix decomposition, including the following steps:

[0053] S1, Assumption It is composed of nv A normal array of data is collected by an acoustic sensing array consisting of several acoustic sensors, where x (v) This represents the v-th listening angle data, 1≤v≤n v x (v) The dimension is 1×m. The anomaly detection network W based on multi-angle semi-nonnegative matrix factorization and Gaussian distribution estimation is trained using batch normal array data x to obtain the anomaly detection model M.

[0054] S2. Input y into the anomaly detection model M to obtain the anomaly detection result R, where Indicates that n v An acoustic sensing array composed of several acoustic sensors collects data from an array under test, where y (v) This represents the v-th listening angle data to be detected, where 1 ≤ v ≤ n. v y (v) The dimension is 1×m1.

[0055] In this embodiment, an anomaly detection model driven solely by normal data is constructed. Its core is a carefully designed anomaly detection network W that learns the essential characteristics and distribution patterns of normal acoustic sensor array data. In the S1 training phase, network W utilizes multi-angle semi-nonnegative matrix factorization to extract "shared features" representing the commonalities of all sensors from a large amount of normal data. Based on these features, a Gaussian probability model is built to describe the clustering of normal data in the feature space. The final trained anomaly detection model M encapsulates this normal feature extraction and probability evaluation capability. In the S2 detection phase, the data to be detected, y, is input into model M, and its logarithmic probability density value relative to the normal model is calculated. If this value is significantly lower (below the threshold δ obtained from normal data statistics), it indicates that the data to be detected deviates from the normal pattern and is thus judged as an anomaly. This method achieves efficient and automated anomaly detection. Since the training process does not require any abnormal samples, it is highly suitable for situations where abnormal samples are scarce or difficult to obtain in real-world industrial scenarios. Furthermore, by setting a threshold using probabilistic methods, the robustness and reliability of the detection are improved.

[0056] like Figure 2 S1 includes:

[0057] The structure of the anomaly detection network W is as follows:

[0058] S11, Input Layer: Input n normal array data x;

[0059] S12. Extract feature data layer: Calculate the auditory angle data x for each auditory angle in the normal array data x. (v) The features are used to obtain feature array data. Where c (v)This indicates that the data originates from the v-th hearing angle. (v) Feature data, 1≤v≤n v c (v) The dimension is 1×p;

[0060] S13, Feature Data Decomposition and Shared Feature Calculation Layer:

[0061] S131, Using formula c (v) =G (v) F (v) (1) For each feature data c in the feature array data c (v) Perform semi-nonnegative matrix decomposition to obtain the corresponding coefficient matrix G. (v) and basis matrix F (v) G (v) The dimension is n×q, F (v) The dimension is q×p;

[0062] S132, Using the formula The shared features of normal array data x are calculated, thus obtaining the shared feature matrix G of normal array data x. * ;

[0063] S133, In formula (2), λ v Q represents the weight of the v-th listening angle data. (v) It is a diagonal matrix, according to the formula Perform calculations, matrix Matrix G * The size is n×q, and the row vector is g. k The dimension is 1×q;

[0064] S14, Gaussian distribution parameter estimation and Gaussian density calculation layer:

[0065] S141. Using the formula Calculate the mean parameter μ of the Gaussian distribution model;

[0066] S142, Using the formula Calculate the covariance parameter ε of the Gaussian distribution model;

[0067] S143, Using the formula Calculate the logarithmic probability density value L;

[0068] S144. In formulas (4), (5), and (6), the dimension of μ is 1×q, |ε| represents the value of the determinant of ε, and ε -1 Let G denote the inverse matrix of ε, log(·) represents the base-2 logarithm, (G * -μ) T Yes (G) * Transpose of -μ).

[0069] In this embodiment, an end-to-end learning framework is constructed from multi-angle array data to probability assessment. The network first enhances the data representation capability through a feature extraction layer (S12), and then utilizes the core step of semi-nonnegative matrix factorization (S131) ​​to decompose the feature data of each angle into a basis matrix and coefficient matrix with partially nonnegative constraints. This captures the inherent structure of the data and is more flexible than fully nonnegative decomposition. Subsequently, weighted fusion (S132, S133) is used to calculate a shared feature matrix representing the global state of the entire array, achieving effective fusion of multi-source sensor information. Finally, the network assumes that the shared features of normal data follow a multivariate Gaussian distribution and quantifies the probability that any data sample belongs to the normal pattern by estimating its parameters (S141, S142) and calculating the log probability density (S143). This achieves a seamless connection between feature learning, information fusion, and probabilistic modeling. End-to-end training enables the shared features to optimally represent the essence of normal data, thus laying a stable and reliable foundation for subsequent probability-based anomaly decision-making.

[0070] S1 also includes:

[0071] The training process of the anomaly detection network W is as follows:

[0072] SA1. In a batch of normal array data, randomly select n normal array data x and input them into the anomaly detection network W. Train the weights of the anomaly detection network W according to the principle of minimizing the loss function Ω.

[0073] SA2, after repeating the training process SA1 several times, the anomaly detection model M is obtained.

[0074] In this embodiment, an unsupervised optimization strategy based on minimizing reconstruction and distribution losses is employed to shape the anomaly detection network W. Its core is to drive the network weights (including the basis matrix of the semi-nonnegative matrix factorization and the Gaussian distribution parameters) to iteratively update towards a common goal through multiple random sampling of normal data batches (SA1): that is, to enable the network to reconstruct normal input features with minimal error, while ensuring that the shared features of all normal samples have the highest joint probability density under the Gaussian distribution defined by the network. By repeatedly executing this process (SA2), the network is forced to learn the most representative and consistent intrinsic patterns in normal data. This allows for the construction of a highly specialized detection model M using only normal data. This model is inclusive of normal, common fluctuations, but exhibits significant reconstruction errors or low probability densities for anomaly data that deviates from the learned "normal pattern" and did not appear during training. This achieves powerful and stable anomaly recognition capabilities while effectively avoiding the training difficulties caused by a lack of anomaly samples.

[0075] SA1 includes:

[0076] SB1, The loss function Ω used for training the anomaly detection network W is: Ω=E-λ1L1+λ2L2(7);

[0077] SB2, where E is given by formula The calculation yields that ||·|| F This indicates that the Frobenius norm operation is performed, and L1 is given by the formula. Calculations show that L2 is obtained from the formula The calculations show that λ1 and λ2 are the weight control parameters.

[0078] In this embodiment, the principle of the composite loss function design lies in guiding the network to learn a feature space with both high-precision reconstruction capability and strong generalization through a multi-objective collaborative optimization mechanism. Its core consists of three parts: the reconstruction error E (based on the Frobenius norm) calculated by formula (8), which serves as the main optimization objective, forcing the base matrix and coefficient matrix learned by the network to recover the original input features as accurately as possible, ensuring the model's fidelity to the basic patterns of normal data; the regularization terms L1 and L2 defined by formulas (9) and (10), which respectively impose additional constraints (such as sparsity and smoothness) on the coefficient matrix or shared features to prevent the model from overfitting noise and irrelevant details in the training data, thereby improving the model's generalization capability; and the hyperparameters λ1 and λ2, which serve as precise control valves to dynamically balance the trade-off between reconstruction accuracy and model complexity. By minimizing this joint loss function (formula 7), the training process is guided to an optimal solution, which not only guarantees the accuracy of feature decomposition but also ensures that the learned feature representation is concise and robust. This design significantly improves the robustness and reliability of the final anomaly detection model. It effectively prevents the network from merely memorizing training data, instead prompting it to uncover more fundamental and generalized normal patterns hidden within the data. This allows the model to maintain a stable high probability density output when faced with new normal data, while exhibiting a significantly low probability density for anomaly data, thereby greatly improving the accuracy and generalization performance of anomaly detection.

[0079] S2 includes:

[0080] S21. Input the array data y to be inspected into the anomaly detection model M to obtain the Gaussian log probability density value L;

[0081] S22. Based on the relationship between the logarithmic probability density value L of the Gaussian distribution and the threshold δ, the abnormal detection result R is obtained. If L≤δ, the result R indicates that the array data y to be inspected is abnormal. If L>δ, the result R indicates that the array data y to be inspected is normal.

[0082] In this embodiment, a core statistical assumption is used: normal data exhibits a high probability density under a trained Gaussian model, while abnormal data corresponds to a low probability density. Specifically, in step S21, the data to be detected, y, is processed by the anomaly detection model M, with a process similar to that during training: first, multi-angle feature extraction and semi-nonnegative matrix factorization are performed to obtain its shared features. Then, these features are substituted into the Gaussian distribution model estimated during the training phase, ultimately calculating a quantized Gaussian distribution log probability density value L. This L value intuitively reflects the degree of fit between the data to be detected, y, and the "normal pattern." In the decision-making step S22, by comparing the L value with an unbiased threshold δ pre-determined based on normal data statistics, a reliable conversion from continuous probability values ​​to discrete judgments (normal / abnormal) is achieved. This enables the entire system to possess objective and automated decision-making capabilities. It abandons the method of subjectively setting thresholds and instead scientifically defines the boundary between normal and abnormal based on the distribution of normal data itself (such as through quantiles). This not only greatly improves the accuracy and consistency of detection, but also significantly reduces false alarms and false negatives caused by human factors, ensuring the interpretability and reliability of the detection results.

[0083] S12 includes:

[0084] S121, the feature data c (v) Including from the hearing angle data x (v) The features calculated in the middle and the hearing angle data x (v) .

[0085] In this embodiment, the feature construction scheme employs a strategy that integrates raw data with derived features, aiming to construct an input space with higher information density and stronger representational capabilities. Its core is to use unprocessed raw listening angle data along with various features calculated from it (such as time-domain statistics and frequency-domain components) as feature data. This method not only fully preserves all the true structure and subtle fluctuations contained in the original signal, preventing information loss during preprocessing, but also actively reveals and strengthens the potential, discriminative physical laws or statistical patterns in the data through feature calculation. The effect of this design is to provide an optimal data foundation for subsequent decomposition and modeling. It cleverly achieves a balance between information fidelity and feature enhancement, enabling the model to learn from both the macroscopic morphology and microscopic characteristics of the signal simultaneously, thereby greatly improving the richness and robustness of the extracted shared features and providing a fundamental guarantee for ultimately achieving accurate and reliable anomaly detection.

[0086] S13 also includes:

[0087] SE1, Initialize all basis matrices F of the multi-angle semi-nonnegative matrix decomposition. (v) F (v)The initial value is in the range (-1, 1);

[0088] SE2, Initialize all coefficient matrices G of the multi-angle semi-nonnegative matrix decomposition. (v) G (v) The initial value is in the range (0,1);

[0089] SE3, Initialize the shared feature matrix G for each hearing angle data. * G * The initial value is in the range (0,1);

[0090] SE4. Use formula And Equation (2), for the coefficient matrix G (v) and shared feature matrix G * Perform several alternating iterative updates to decompose the feature data c into matrix G. (v) sum matrix F (v) And obtain the shared feature matrix G of the sensor array data. * ;

[0091] SE5. In formula (11), [·] + This represents the positive elements within the brackets [·]. - This represents the positive elements in [·].

[0092] In the anomaly detection network W described in SE6, S1, and S2, for each hearing angle data x (v) The basis matrix F for semi-nonnegative matrix decomposition (v) The mean parameter μ and covariance parameter ε of the Gaussian distribution model are the network weights that need to be trained and are optimized as the anomaly detection network is trained.

[0093] In this embodiment, the principle of the initialization and optimization mechanism of the semi-nonnegative matrix factorization lies in laying a solid foundation for model training through a strategy that combines constrained initialization with alternating directional optimization. Its core is to differentially initialize various matrices: allowing the base matrix to contain negative values ​​(range (-1, 1)) to retain its full ability to express complex feature patterns (including inhibitory parts); while initializing the coefficient matrix and shared feature matrix to positive values ​​(range (0, 1)), which conforms to their physical meaning as "contribution" or "activation intensity" in the model, ensuring the physical interpretability of the decomposition. Subsequently, through alternating iterative updates (SE4) using formulas (11) and (2), and utilizing a special rule that operates only on positive elements (SE5), the coefficients and shared features are dynamically and constrainedly updated during the solution process, ensuring that the decomposition process converges stably while satisfying the semi-nonnegative requirement. Most importantly, all these locally optimized parameters (basis matrices) and global model parameters (μ and ε of a Gaussian distribution) are defined together as the network's trainable weights (SE6). This makes the entire decomposition process not an isolated preprocessing step, but an organic component capable of end-to-end global optimization through backpropagation based on the final anomaly detection task objective. This design achieves two effects: First, the refined initialization and iteration rules ensure stability and rapid convergence in the early stages of training, effectively avoiding getting trapped in local optima due to improper random initialization. Second, the end-to-end training mechanism ensures that feature decomposition and probabilistic modeling are no longer isolated but rather mutually reinforcing and co-evolving. Ultimately, this allows the learned basis vectors and Gaussian model to be optimally adapted to the anomaly detection task, thereby significantly improving the representational ability of shared features and the overall detection accuracy and generalization ability of the model.

[0094] S22, including:

[0095] The steps for obtaining the threshold δ are as follows:

[0096] 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 ;

[0097] S222, Find all L j n L Quantiles are used as the threshold δ.

[0098] In this embodiment, the principle of the threshold determination method is based on the hypothesis testing concept in statistical learning. That is, it uses a trained model to "self-verify" known normal data, objectively defining the boundary between normal and abnormal based on the distribution pattern of normal samples in their own constructed probability space. Specifically, in step S221, all normal data used for training is input into the converged model M to obtain a series of log probability density values ​​{L}. jThis process aims to accurately characterize the probability distribution range of "pure" normal data in the model. Subsequently, in step S222, {L} is selected. j A specific quantile (e.g., the 5th or 1st percentile) is used as the final threshold δ. This is equivalent to setting a statistical confidence lower bound for the "normal" state.

[0099] This scheme achieves data-driven and objective decision thresholding. It relies entirely on the statistical characteristics of normal data, completely avoiding the subjectivity and arbitrariness of manually setting thresholds. By selecting different quantiles, users can flexibly and directly control the model's sensitivity to anomalies (or false positive rate): for example, choosing a lower quantile (e.g., 1%) means the model is more sensitive to anomalies, but may also increase false positives; while choosing a higher quantile (e.g., 5%) makes the model more stringent and reduces the risk of missed detections. This method ensures that the threshold δ perfectly matches the "normal" concept learned by the model M and the characteristics of the specific dataset, thereby maximizing the normal data recognition rate while providing a stable, scientific, and interpretable criterion for reliably identifying anomalies, significantly improving the practicality and reliability of the detection system.

[0100] 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.

[0101] 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. A method for detecting anomalies in acoustic sensor array data based on multi-angle semi-nonnegative matrix decomposition, characterized in that, Includes the following steps: 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, where x (v) This represents the v-th listening angle data, 1≤v≤n v x (v) The dimension is 1×m. The anomaly detection network W based on multi-angle semi-nonnegative matrix factorization and Gaussian distribution estimation is trained using batch normal array data x to obtain the anomaly detection model M. S2. Input y into the anomaly detection model M to obtain the anomaly detection result R, where Indicates that n v An acoustic sensing array composed of several acoustic sensors collects data from an array under test, where y (v) This represents the v-th listening angle data to be detected, where 1 ≤ v ≤ n. v y (v) The dimension is 1×m1.

2. The method for detecting anomalies in acoustic sensor array data based on multi-angle semi-nonnegative matrix decomposition according to claim 1, characterized in that, S1 includes: The structure of the anomaly detection network W is as follows: S11, Input Layer: Input n normal array data x; S12. Extract feature data layer: Calculate the auditory angle data x for each auditory angle in the normal array data x. (v) The features are used to obtain feature array data. Where c (v) This indicates that the data originates from the v-th hearing angle. (v) Feature data, 1≤v≤n v c (v) The dimension is 1×p; S13, Feature Data Decomposition and Shared Feature Calculation Layer: S131, Using formula c (v) =G (v) F (v) (1) For each feature data c in the feature array data c (v) Perform semi-nonnegative matrix decomposition to obtain the corresponding coefficient matrix G. (v) and basis matrix F (v) G (v) The dimension is n×q, F (v) The dimension is q×p; S132, Using the formula The shared features of normal array data x are calculated, thus obtaining the shared feature matrix G of normal array data x. * ; S133, In formula (2), λ v Q represents the weight of the v-th listening angle data. (v) It is a diagonal matrix, according to the formula Perform calculations, matrix Matrix G * The size is n×q, and the row vector is g. k The dimension is 1×q; S14, Gaussian distribution parameter estimation and Gaussian density calculation layer: S141. Using the formula Calculate the mean parameter μ of the Gaussian distribution model; S142, Using the formula Calculate the covariance parameter ε of the Gaussian distribution model; S143, Using the formula Calculate the logarithmic probability density value L; S144. In formulas (4), (5), and (6), the dimension of μ is 1×q, |ε| represents the value of the determinant of ε, and ε -1 Let G denote the inverse matrix of ε, log(·) represents the base-2 logarithm, (G * -μ) T Yes (G) * Transpose of -μ).

3. The method for detecting anomalies in acoustic sensor array data based on multi-angle semi-nonnegative matrix decomposition according to claim 1, characterized in that, S1 further includes: The training process of the anomaly detection network W is as follows: SA1. In a batch of normal array data, randomly select n normal array data x and input them into the anomaly detection network W. Train the weights of the anomaly detection network W according to the principle of minimizing the loss function Ω. SA2, after repeating the training process SA1 several times, the anomaly detection model M is obtained.

4. The method for detecting anomalies in acoustic sensor array data based on multi-angle semi-nonnegative matrix decomposition according to claim 3, characterized in that, The SA1 includes: SB1, The loss function Ω used for training the anomaly detection network W is: Ω=E-λ1L1+λ2L2(7); SB2, where E is given by formula The calculation yields that ||·|| F This indicates that the Frobenius norm operation is performed, and L1 is given by the formula. Calculations show that L2 is obtained from the formula The calculations show that λ1 and λ2 are the weight control parameters.

5. The method for detecting anomalies in acoustic sensor array data based on multi-angle semi-nonnegative matrix decomposition according to claim 1, characterized in that, The S2 includes: S21. Input the array data y to be inspected into the anomaly detection model M to obtain the Gaussian log probability density value L; S22. Based on the relationship between the logarithmic probability density value L of the Gaussian distribution and the threshold δ, the abnormal detection result R is obtained. If L≤δ, the result R indicates that the array data y to be inspected is abnormal. If L>δ, the result R indicates that the array data y to be inspected is normal.

6. The method for detecting anomalies in acoustic sensor array data based on multi-angle semi-nonnegative matrix decomposition according to claim 2, characterized in that, S12 includes: S121, the feature data c (v) Including from the hearing angle data x (v) The features calculated in the middle and the hearing angle data x (v) .

7. The method for detecting anomalies in acoustic sensor array data based on multi-angle semi-nonnegative matrix decomposition according to claim 2, characterized in that, S13 further includes: SE1, Initialize all basis matrices F of the multi-angle semi-nonnegative matrix decomposition. (v) F (v) The initial value is in the range (-1, 1); SE2, Initialize all coefficient matrices G of the multi-angle semi-nonnegative matrix decomposition. (v) G (v) The initial value is in the range (0,1); SE3, Initialize the shared feature matrix G for each hearing angle data. * G * The initial value is in the range (0,1); SE4. Use formula And Equation (2), for the coefficient matrix G (v) and shared feature matrix G * Perform several alternating iterative updates to decompose the feature data c into matrix G. (v) sum matrix F (v) And obtain the shared feature matrix G of the sensor array data. * ; SE5. In formula (11), [·] + This represents the positive elements within the brackets [·]. - This represents the positive elements in [·]. In the anomaly detection network W described in SE6, S1, and S2, for each hearing angle data x (v) The basis matrix F for semi-nonnegative matrix decomposition (v) The mean parameter μ and covariance parameter ε of the Gaussian distribution model are the network weights that need to be trained and are optimized as the anomaly detection network is trained.

8. The method for detecting anomalies in acoustic sensor array data based on multi-angle semi-nonnegative matrix decomposition according to claim 5, characterized in that, S22 includes: The steps for obtaining the threshold δ are as follows: 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 n L Quantiles are used as the threshold δ.

Images