Cable gallery working condition abnormity monitoring method based on sound recognition

Abstract

The invention discloses a cable gallery working condition abnormity monitoring method based on sound recognition, and the method comprises the steps: collecting voiceprint recognition related data through a collection unit, and the voiceprint recognition related data comprises audio data, infrared temperature measurement image data, a smoke sensor signal, temperature data and humidity data; the collected voiceprint recognition related data is transmitted to a cloud data platform of a server through a data line and a switch, and feature value extraction is performed on the audio data through a voiceprint feature value data edge calculation gateway of the cloud data platform; performing voiceprint data processing on the audio data through a voiceprint recognition algorithm; carrying out comprehensive fault judgment; the method solves the problems that in the prior art, due to the fact that the judgment condition is single, comprehensive judgment cannot be carried out through multi-modal data, only part of faults can be recognized, and the fault recognition rate and accuracy are low, and has the advantages that comprehensive fault judgment can be carried out through combination of the multi-modal data and voiceprint scoring, and the fault recognition efficiency is improved. And the recognition rate and the accuracy of cable gallery abnormity or fault judgment are greatly improved.

Description

Technical Field

[0001] The invention belongs to the technical field of fault diagnosis of hydropower stations, and in particular relates to a method for monitoring abnormal working conditions of a cable corridor based on sound recognition. Background Art

[0002] The geographical environment inside the dam and powerhouse of a hydropower station is complex, and cable corridors are widely distributed. Manual inspections take a long time and are inefficient, so automatic online monitoring of safety risks in cable corridors is particularly important. Optical cables, power cables, and control cables laid in cable corridors are important components for maintaining normal operation of power plants. The flammability of cables and their ductility during fires have made cable safety work highly valued by power plants. For hydropower plants, cable accidents also seriously endanger the safe operation of dams and shipping facilities. In view of the serious consequences of cable safety accidents, it is necessary to conduct an in-depth analysis of the safety measures for cable corridors where cables are highly concentrated. The prior art generally adopts a method for distributed optical cable detection in tunnels, which belongs to the field of optical fiber distributed detection technology, including an optical cable body, the optical cable body is laid in the tunnel, the end of the optical cable body is connected to a real-time monitoring terminal, the real-time monitoring terminal is connected to a monitoring cloud platform and an alarm module, a temperature strain sensing optical core and an acoustic wave sensing optical core are arranged in the optical cable body, and an optical fiber signal processing module and an optical fiber voiceprint monitoring module are arranged outside the optical cable body, and the optical fiber signal processing module includes a laser. The present invention utilizes distributed optical fiber sensing technology and a special optical cable, combined with a neural network and a big data analysis algorithm, to accurately detect and calculate the stress change, vibration velocity and temperature change of the corridor wall of subway tunnels, high-speed tunnels, urban pipe corridors, etc. The disadvantage of this method is that the existing cable corridor is monitored by placing a temperature sensing cable in the cable corridor for possible fire alarms, but the prior art often has various defects and has not formed an overall multi-modal system to achieve comprehensive judgment, resulting in only partial faults being able to be identified, and the fault recognition rate and accuracy are very low; therefore, it is necessary to design a cable corridor working condition abnormality monitoring method based on sound recognition to solve the above problems. Summary of the invention

[0003] The technical problem to be solved by the present invention is to provide a method for monitoring abnormal working conditions of cable corridors based on sound recognition. The method is used to solve the problem that the prior art can only identify partial faults due to the relatively single judgment conditions and inability to make comprehensive judgments through multimodal data, and the fault recognition rate and accuracy are both low. The method has the characteristics of being able to make comprehensive fault judgments through multimodal data combined with voiceprint scoring, thereby greatly improving the recognition rate and accuracy of cable corridor abnormalities or fault judgments.

[0004] In order to achieve the above technical effects, the technical solution adopted by the present invention is: A method for monitoring abnormal working conditions of a cable corridor based on sound recognition comprises the following steps: S1, collecting voiceprint recognition related data through a collection unit, the collection unit includes a bone conduction voiceprint sensor, a microphone voiceprint sensor, an infrared temperature measurement camera, a smoke sensor, a temperature sensor and a humidity meter; the voiceprint recognition related data includes audio data, infrared temperature measurement image data, smoke sensor signal, temperature data and humidity data; S2, transmit the collected voiceprint recognition related data to the cloud data platform of the server through the data line and the switch, and extract the feature value of the audio data through the voiceprint feature value data edge computing gateway of the cloud data platform; S3, performing voiceprint data processing on the audio data through a voiceprint recognition algorithm; S4, conduct comprehensive fault judgment: S401, constructing a fault model library based on historical faults and events, wherein the fault model library includes fault types and event types and corresponding audio features; S402: The collected audio data is subjected to feature value extraction through the edge computing gateway and compared with the fault model library. If the fault type or event type is matched, the audio data is subjected to abnormality judgment through the voiceprint recognition algorithm. If it is judged to be normal, the event is output; if it is judged to be abnormal, the process proceeds to step S403 to make a comprehensive judgment in combination with the smoke sensor signal, temperature data and humidity data; if the fault type or event type is not matched, the process directly proceeds to step S403 for comprehensive judgment; S403, setting the thresholds of temperature data and humidity data. If at least one of the three conditions, namely, the temperature data exceeds the threshold, the humidity data exceeds the threshold, or the smoke sensor signal is actuated, is determined to be abnormal; otherwise, it is determined to be normal; if the comprehensive determination is abnormal, a fault is output and the process proceeds to S404; if the comprehensive determination is normal, the process directly proceeds to S404; S404, calling the infrared temperature measurement image data for manual screening, determining the fault location and type, and updating the parameters of the voiceprint recognition algorithm; S405, combining the fault type and event type output in step S403 with the corresponding audio data feature value to construct a fault model, and add it to the fault model library.

[0005] Preferably, in step S1, the bone conduction voiceprint sensor is installed on the cable tray, the microphone voiceprint sensor is installed in the cable corridor, and the infrared temperature measurement camera, smoke sensor, temperature sensor and humidity meter are all installed in the cable corridor.

[0006] Preferably, in step S3, performing voiceprint data processing on the audio data by using a voiceprint recognition algorithm includes performing MFCC extraction on the audio data, and the specific method is: Selecting audio data within a predetermined time period as a continuous sound sample; The audio data is passed through a high-pass filter to enhance the high-frequency part, making the spectrum of the signal flat and keeping it in the entire frequency band from low frequency to high frequency, so that the spectrum can be obtained with the same signal-to-noise ratio; Execute the framing operation, first group N sampling points into one frame, the value of N is 256 or 512, and the time covered is 20~30ms; in order to avoid too large changes between two adjacent frames, set an overlapping area between two adjacent frames, the overlapping area contains M sampling points, and the value of M is 1 / 3 of N; Perform windowing operation and multiply each frame by a Hamming window to increase the continuity of the left and right ends of the frame; assume that the framed signal is S(n), n=0,1,…,N-1, N is the frame size; Performing fast Fourier transform on each frame signal after frame division and windowing to obtain the frequency spectrum of each frame, and taking the modulus square of the frequency spectrum of the audio signal to obtain the power spectrum of the audio signal; Perform Mel filter calculation.

[0007] Furthermore, the method for performing Mel filter calculation is: Define a Mel-scale triangular filter bank with M filters, pass the power spectrum of the audio signal through the triangular filter bank, and calculate the logarithmic energy of each filter bank output by logarithmic operation; The MFCC coefficients are obtained by discrete cosine transform: Substitute the calculated logarithmic energy into the discrete cosine transform to find the L-order Mel-scale Cepstrum parameter; L-order refers to the MFCC coefficient order; Extract the dynamic difference parameters of audio data, including first-order difference and second-order difference: ; In the formula, represents the tth first-order difference; C t represents the tth cepstral coefficient; Q represents the order of the cepstral coefficient, and K represents the time difference of the first-order derivative, which is 1 or 2.

[0008] Preferably, in step S3, performing voiceprint data processing on the audio data by using a voiceprint recognition algorithm includes calculating UBM, and the specific method is: The collected audio data is divided into a development set and a test set in proportion, and the UBM-GMM model is trained using the development set; The test set data is substituted into the UBM-GMM model, and the parameters of the UBM are adjusted through the adaptive algorithm to obtain the target model parameters and complete the model training.

[0009] Preferably, in step S3, performing voiceprint data processing on the audio data by using a voiceprint recognition algorithm includes calculating an i-vector, and the specific method is: i-vector defines a low-dimensional vector To represent an audio segment; ; M represents the ideal feature super vector corresponding to a device being modeled. represents i-vector, T represents the T matrix, m represents the UBM mean supervector, and assumes that the UBM contains C Gaussian mixture components g, and m is the mean vector m of all mixture components. c The combination of , c=1,...,C; the dimension of m is c*F, the dimension of T is C*F×R, and F represents the dimension of MFCC features; Further split the T matrix into C combinations of Vc : ; In the formula, m c is the mean vector of the mixed components, and the dimension of Vc is F×R; Define a piece of audio feature data X, whose feature dimension is F and time sequence is T, that is, X = X1, . . . , X T ; The subset of X that belongs to the cth Gaussian component is X c , a frame in the subset is X t c ,but ; ; In the formula, Σ c is the covariance matrix; Combining the above formulas, we get: ; The calculation formula of i-vector is: ; Among them, N(u) dimension is a diagonal matrix of CF × CF, the diagonal block is NcI,(c=1,...,C), and F~(u) is a supervector of CF ×1, consisting of all first-order BW statistics F~c.

[0010] Preferably, in step S3, performing voiceprint data processing on the audio data by using a voiceprint recognition algorithm includes linear probability discriminant analysis, and the specific method is: Assume that the audio data used for training consists of audio from I devices, where each device has J different audio segments; the i-vector of the j-th audio segment of the i-th device is denoted by D ij ,Linear Probability Discriminant Analysis PLDA is defined as: ; The device information part is , which is only related to device i and describes the difference between devices; the noise part is ,Describe the differences between devices; represents the mean of all training data; F is the identity space, which contains information that can be used to represent various devices; It is the identity of a specific device or the position of the device in the identity space; G is the error space, which contains information that can be used to represent different audio changes of the same device; represents the position in G space; is the final residual noise term; this term is zero-mean Gaussian with variance Σ; Assume that each latent variable conforms to the following distribution: ; Training set , the model parameters are , the maximum likelihood estimation MLE criterion is used for optimization, that is, ; Initialization parameters ; E-step estimates latent variables; M-step updates parameters based on the estimated latent variables ; Repeat E-step and M-step N times; Use the likelihood ratio strategy for device identification: Assuming that the model describes the relationship between the identity factor h and the input feature i-vector, the test process is to determine ivert x p Whether to register ivivor x i Share the same device identity h; hypothesis M0 represents x i and x p From different identity latent variables, hypothesis M1 represents x i and x p From the same identity latent variable, the likelihood ratio calculates the score value: ; Where core represents the score value; n1 and n2 are the I-Vector vectors of the two audios respectively. The assumption that the two audios come from the same space is M1, and the assumption that they come from different spaces is M0. Among them, p(x1, xp|M1) is the likelihood function that the two audios come from the same space; p(x1, xp|M0) is the likelihood function that x1 and xp come from different spaces respectively; the similarity of the two audios is calculated by calculating the log-likelihood ratio; the normal range of similarity is set, and the audio data falling outside the normal range is judged as abnormal data.

[0011] The beneficial effects of the present invention are as follows: 1. The present invention realizes multi-modal online real-time monitoring of cable corridor equipment areas by installing various types of sensors. It uses bone conduction voiceprint sensors to collect audio data, and can collect almost all abnormal conditions in cable corridor equipment areas. At the same time, multiple auxiliary sensors collect related auxiliary data, which can identify all faults in the cable corridor equipment area, greatly improving the accuracy of abnormal monitoring of cable corridor equipment conditions. Through voiceprint recognition technology, and through comprehensive judgment and analysis of multi-modal data such as temperature, humidity, and smoke alarms, the accuracy of abnormal monitoring of cable corridor equipment conditions can be greatly improved, and human resources can be saved while reducing the time required for fault identification, providing an opportunity for emergency handling of power system accidents and greatly improving the stability of the power system.

[0012] 2. The invention introduces a voiceprint sensor to reduce the false alarm rate caused by the failure of smoke and temperature sensors in the cable corridor. The voiceprint sensor can collect the footsteps in the corridor, monitor whether there is a water leakage in the corridor, and solve the blind spot of water leakage monitoring in the cable corridor. The introduction of the voiceprint sensor can assist in solving the problem of locating the alarm in the corridor. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 It is a schematic diagram of the process of the present invention; Figure 2 Schematic diagram of the process of extracting MFCC in an embodiment of the present invention; Figure 3 It is a flow chart of the voiceprint recognition algorithm in an embodiment of the present invention; Figure 4 The figure is a flow chart of comprehensive determination by using a determination model in an embodiment of the present invention. DETAILED DESCRIPTION

[0014] Embodiment 1: like Figure 1 As shown, a method for monitoring abnormal working conditions of a cable corridor based on sound recognition comprises the following steps: S1, collecting voiceprint recognition related data through a collection unit, the collection unit includes a bone conduction voiceprint sensor, a microphone voiceprint sensor, an infrared temperature measurement camera, a smoke sensor, a temperature sensor and a humidity meter; the voiceprint recognition related data includes audio data, infrared temperature measurement image data, smoke sensor signal, temperature data and humidity data; S2, transmit the collected voiceprint recognition related data to the cloud data platform of the server through the data line and the switch, and extract the feature value of the audio data through the voiceprint feature value data edge computing gateway of the cloud data platform; S3, performing voiceprint data processing on the audio data through a voiceprint recognition algorithm; S4, perform comprehensive fault determination.

[0015] Preferably, in step S1, the bone conduction voiceprint sensor is installed on the cable tray, the microphone voiceprint sensor is installed in the cable corridor, and the infrared temperature measurement camera, smoke sensor, temperature sensor and humidity meter are all installed in the cable corridor.

[0016] like Figure 2 As shown, in step S3, the voiceprint data processing of the audio data by the voiceprint recognition algorithm includes MFCC extraction of the audio data, and the specific method is: Selecting audio data within a predetermined time period as a continuous sound sample; The audio data is passed through a high-pass filter to enhance the high-frequency part, making the spectrum of the signal flat and keeping it in the entire frequency band from low frequency to high frequency, so that the spectrum can be obtained with the same signal-to-noise ratio; Execute the framing operation, first group N sampling points into one frame, the value of N is 256 or 512, and the time covered is 20~30ms; in order to avoid too large changes between two adjacent frames, set an overlapping area between two adjacent frames, the overlapping area contains M sampling points, and the value of M is 1 / 3 of N; Perform windowing operation and multiply each frame by a Hamming window to increase the continuity of the left and right ends of the frame; assume that the framed signal is S(n), n=0,1,…,N-1, N is the frame size; Performing fast Fourier transform on each frame signal after frame division and windowing to obtain the frequency spectrum of each frame, and taking the modulus square of the frequency spectrum of the audio signal to obtain the power spectrum of the audio signal; Perform Mel filter calculation.

[0017] Furthermore, the method for performing Mel filter calculation is: Define a Mel-scale triangular filter bank with M filters, pass the power spectrum of the audio signal through the triangular filter bank, and calculate the logarithmic energy of each filter bank output by logarithmic operation; The MFCC coefficients are obtained by discrete cosine transform: Substitute the calculated logarithmic energy into the discrete cosine transform to find the L-order Mel-scale Cepstrum parameter; L-order refers to the MFCC coefficient order; Extract the dynamic difference parameters of audio data, including first-order difference and second-order difference: ; In the formula, represents the tth first-order difference; C t represents the tth cepstral coefficient; Q represents the order of the cepstral coefficient, and K represents the time difference of the first-order derivative, which is 1 or 2.

[0018] like Figure 3As shown, in step S3, the voiceprint data processing of the audio data by the voiceprint recognition algorithm includes calculating the UBM, and the specific method is: The collected audio data is divided into a development set and a test set in proportion, and the UBM-GMM model is trained using the development set; The test set data is substituted into the UBM-GMM model, and the parameters of the UBM are adjusted through the adaptive algorithm to obtain the target model parameters and complete the model training.

[0019] Preferably, in step S3, performing voiceprint data processing on the audio data by using a voiceprint recognition algorithm includes calculating an i-vector, and the specific method is: i-vector defines a low-dimensional vector To represent an audio segment; ; M represents the ideal feature super vector corresponding to a device being modeled. represents i-vector, T represents the T matrix, m represents the UBM mean supervector, and assumes that the UBM contains C Gaussian mixture components g, and m is the mean vector m of all mixture components. c The combination of , c=1,...,C; the dimension of m is c*F, the dimension of T is C*F×R, and F represents the dimension of MFCC features; Further split the T matrix into C combinations of Vc : ; In the formula, m c is the mean vector of the mixed components, and the dimension of Vc is F×R; Define a piece of audio feature data X, whose feature dimension is F and time sequence is T, that is, X = X1, . . . , X T ; The subset of X that belongs to the cth Gaussian component is X c , a frame in the subset is X t c ,but ; ; In the formula, Σ c is the covariance matrix; Combining the above formulas, we get: ; The calculation formula of i-vector is: ; Among them, N(u) dimension is a diagonal matrix of CF × CF, the diagonal block is NcI,(c=1,...,C), and F~(u) is a supervector of CF ×1, consisting of all first-order BW statistics F~c.

[0020] Preferably, in step S3, performing voiceprint data processing on the audio data by using a voiceprint recognition algorithm includes linear probability discriminant analysis, and the specific method is: Assume that the audio data used for training consists of audio from I devices, where each device has J different audio segments; the i-vector of the j-th audio segment of the i-th device is denoted by D ij ,Linear Probability Discriminant Analysis PLDA is defined as: ; The device information part is , which is only related to device i and describes the difference between devices; the noise part is ,Describe the differences between devices; represents the mean of all training data; F is the identity space, which contains information that can be used to represent various devices; It is the identity of a specific device or the position of the device in the identity space; G is the error space, which contains information that can be used to represent different audio changes of the same device; represents the position in G space; is the final residual noise term; this term is zero-mean Gaussian with variance Σ; Assume that each latent variable conforms to the following distribution: ; Training set , the model parameters are , the maximum likelihood estimation MLE criterion is used for optimization, that is, ; Initialization parameters ; E-step estimates latent variables; M-step updates parameters based on the estimated latent variables ; Repeat E-step and M-step N times; Use the likelihood ratio strategy for device identification: Assuming that the model describes the relationship between the identity factor h and the input feature i-vector, the test process is to determine ivert x p Whether to register ivivor x i Share the same device identity h; hypothesis M0 represents x i and x p From different identity latent variables, hypothesis M1 represents x iand x p From the same identity latent variable, the likelihood ratio calculates the score value: ; Where core represents the score value; n1 and n2 are the I-Vector vectors of the two audios respectively. The assumption that the two audios come from the same space is M1, and the assumption that they come from different spaces is M0. Among them, p(x1, xp|M1) is the likelihood function that the two audios come from the same space; p(x1, xp|M0) is the likelihood function that x1 and xp come from different spaces respectively; the similarity of the two audios is calculated by calculating the log-likelihood ratio; the normal range of similarity is set, and the audio data falling outside the normal range is judged as abnormal data.

[0021] Embodiment 2: like Figure 4 As shown in the figure, the method for comprehensive fault judgment is: S401, constructing a fault model library based on historical faults and events, wherein the fault model library includes fault types and event types and corresponding audio features; S402: The collected audio data is subjected to feature value extraction through the edge computing gateway and compared with the fault model library. If the fault type or event type is matched, the audio data is subjected to abnormality judgment through the voiceprint recognition algorithm. If it is judged to be normal, the event is output; if it is judged to be abnormal, the process proceeds to step S403 to make a comprehensive judgment in combination with the smoke sensor signal, temperature data and humidity data; if the fault type or event type is not matched, the process directly proceeds to step S403 for comprehensive judgment; S403, setting the thresholds of temperature data and humidity data. If at least one of the three conditions, namely, the temperature data exceeds the threshold, the humidity data exceeds the threshold, or the smoke sensor signal is actuated, is determined to be abnormal; otherwise, it is determined to be normal; if the comprehensive determination is abnormal, a fault is output and the process proceeds to S404; if the comprehensive determination is normal, the process directly proceeds to S404; S404, calling the infrared temperature measurement image data for manual screening, determining the fault location and type, and updating the parameters of the voiceprint recognition algorithm; S405, combining the fault type and event type output in step S403 with the corresponding audio data feature value to construct a fault model, and add it to the fault model library.

Claims

1. A method for monitoring abnormal working conditions of cable corridors based on sound recognition, characterized in that: The following steps are involved: S1, collecting voiceprint recognition related data through a collection unit, the collection unit includes a bone conduction voiceprint sensor, a microphone voiceprint sensor, an infrared temperature measurement camera, a smoke sensor, a temperature sensor and a humidity meter; the voiceprint recognition related data includes audio data, infrared temperature measurement image data, smoke sensor signal, temperature data and humidity data; S2, transmit the collected voiceprint recognition related data to the cloud data platform of the server through the data line and the switch, and extract the feature value of the audio data through the voiceprint feature value data edge computing gateway of the cloud data platform; S3, performing voiceprint data processing on the audio data through a voiceprint recognition algorithm; S4, conduct comprehensive fault judgment: S401, constructing a fault model library based on historical faults and events, wherein the fault model library includes fault types and event types and corresponding audio features; S402: The collected audio data is subjected to feature value extraction through the edge computing gateway and compared with the fault model library. If the fault type or event type is matched, the audio data is subjected to abnormality determination through the voiceprint recognition algorithm. If it is determined to be normal, the event is output; If it is determined to be abnormal, the process proceeds to step S403 to make a comprehensive determination based on the smoke sensor signal, temperature data and humidity data; If no fault type or event type is matched, the process directly proceeds to step S403 for comprehensive determination; S403, setting thresholds for temperature data and humidity data, if at least one of the following three conditions is met: the temperature data exceeds the threshold, the humidity data exceeds the threshold, or the smoke sensor signal is actuated, it is determined to be abnormal; Otherwise it is judged as normal; If the comprehensive judgment is abnormal, a fault is output and the process goes to S404; if the comprehensive judgment is normal, the process goes directly to S404; S404, calling the infrared temperature measurement image data for manual screening, determining the fault location and type, and updating the parameters of the voiceprint recognition algorithm; S405, combining the fault type and event type output in step S403 with the corresponding audio data feature value to construct a fault model, and add it to the fault model library.

2. A method for monitoring abnormal working conditions of a cable corridor based on sound recognition according to claim 1, characterized in that: In step S1, the bone conduction voiceprint sensor is installed on the cable tray, the microphone voiceprint sensor is installed in the cable corridor, and the infrared temperature measurement camera, smoke sensor, temperature sensor and humidity meter are all installed in the cable corridor.

3. The method for monitoring abnormal working conditions of a cable corridor based on sound recognition according to claim 1 is characterized in that: In step S3, the voiceprint data processing of the audio data by the voiceprint recognition algorithm includes MFCC extraction of the audio data, and the specific method is as follows: Selecting audio data within a predetermined time period as a continuous sound sample; The audio data is passed through a high-pass filter to enhance the high-frequency part, making the spectrum of the signal flat and keeping it in the entire frequency band from low frequency to high frequency, so that the spectrum can be obtained with the same signal-to-noise ratio; Execute the framing operation, first group N sampling points into one frame, the value of N is 256 or 512, and the time covered is 20~30ms; in order to avoid too large changes between two adjacent frames, set an overlapping area between two adjacent frames, the overlapping area contains M sampling points, and the value of M is 1 / 3 of N; Perform windowing operation and multiply each frame by a Hamming window to increase the continuity of the left and right ends of the frame; assume that the framed signal is S(n), n=0,1,…,N-1, N is the frame size; Performing fast Fourier transform on each frame signal after frame division and windowing to obtain the frequency spectrum of each frame, and taking the modulus square of the frequency spectrum of the audio signal to obtain the power spectrum of the audio signal; Perform Mel filter calculation.

4. A method for monitoring abnormal working conditions of a cable corridor based on sound recognition according to claim 3, characterized in that: In step S3, the voiceprint data processing of the audio data includes calculating the UBM by using the voiceprint recognition algorithm, and the specific method is as follows: The collected audio data is divided into a development set and a test set in proportion, and the UBM-GMM model is trained using the development set; The test set data is substituted into the UBM-GMM model, and the parameters of the UBM are adjusted through the adaptive algorithm to obtain the target model parameters and complete the model training.

5. A method for monitoring abnormal working conditions of a cable corridor based on sound recognition according to claim 4, characterized in that: In step S3, the voiceprint data processing of the audio data by the voiceprint recognition algorithm includes calculating the i-vector, and the specific method is: i-vector defines a low-dimensional vector To represent an audio segment; ； M represents the ideal feature super vector corresponding to a device being modeled. represents i-vector, T represents the T matrix, m represents the UBM mean supervector, and assumes that the UBM contains C Gaussian mixture components g, and m is the mean vector m of all mixture components. c The combination of , c=1,...,C; the dimension of m is c*F, the dimension of T is C*F×R, and F represents the dimension of MFCC features; Further split the T matrix into C combinations of Vc : ； In the formula, m c is the mean vector of the mixed components, and the dimension of Vc is F×R; Define a piece of audio feature data X, whose feature dimension is F and time sequence is T, that is, X = X1, . . ., X T ; The subset of X that belongs to the cth Gaussian component is X c , a frame in the subset is X t c ,but ; ； In the formula, Σ c is the covariance matrix; Combining the above formulas, we get: ； The calculation formula of i-vector is: ； Among them, N(u) dimension is a diagonal matrix of CF × CF, the diagonal block is NcI,(c=1,...,C), and F~(u) is a CF×1 supervector consisting of all first-order BW statistics F~c.

6. A method for monitoring abnormal working conditions of a cable corridor based on sound recognition according to claim 5, characterized in that: In step S3, the voiceprint data processing includes linear probability discriminant analysis on the audio data through the voiceprint recognition algorithm, and the specific method is as follows: Assume that the audio data used for training consists of audio from I devices, where each device has J different audio segments; the i-vector of the j-th audio segment of the i-th device is denoted by D ij ,Linear Probability Discriminant Analysis PLDA is defined as: ； The device information part is , which is only related to device i and describes the difference between devices; the noise part is , describing the differences between devices; represents the mean of all training data; F is the identity space, which contains information that can be used to represent various devices; It is the identity of a specific device or the position of the device in the identity space; G is the error space, which contains information that can be used to represent different audio changes of the same device; represents the position in G space; is the final residual noise term; this term is zero-mean Gaussian with variance Σ; Assume that each latent variable conforms to the following distribution: ； Training set ; The model parameters are ; The maximum likelihood estimation MLE criterion is used for optimization, that is, ; Initialization parameters ; E-step estimates latent variables; M-step updates parameters based on estimated latent variables ; Repeat E-step and M-step N times; Use likelihood ratio strategy for device identification: Assuming that the model describes the relationship between the identity factor h and the input feature i-vector, the test process is to determine ivert x p Whether to register ivivor x i Share the same device identity h; hypothesis M0 represents x i and x p From different identity latent variables, hypothesis M1 represents x i and x p From the same identity latent variable, the likelihood ratio calculates the score value: ； Where core represents the score value; n1 and n2 are the I-Vector vectors of the two audios respectively. The assumption that the two audios come from the same space is M1, and the assumption that they come from different spaces is M0. Among them, p(x1, xp|M1) is the likelihood function that the two audios come from the same space; p(x1, xp|M0) is the likelihood function that x1 and xp come from different spaces respectively; the similarity of the two audios is calculated by calculating the log-likelihood ratio; the normal range of similarity is set, and the audio data falling outside the normal range is judged as abnormal data.

Images