GIS equipment anomaly detection method based on AI voiceprint recognition

Abstract

The invention discloses a GIS equipment anomaly detection method based on AI voiceprint recognition. The GIS equipment anomaly detection method comprises the following steps that 1, original sound wave signals are collected and preprocessed; step 2, performing short-time Fourier transform, and reconstructing phase information based on a minimum phase reconstruction algorithm and a Gr offset-Lim phase recovery algorithm; 3, extracting a target equipment sound source mask value by adopting a mask prediction network, and outputting a target voiceprint map tensor; 4, outputting a time sequence voiceprint feature tensor; 5, based on the improved DeepAR model, introducing a spectrum residual activation function to obtain a parameter set; step 6, constructing a prediction probability density function, and outputting an anomaly discrimination tag sequence; step 7, outputting an abnormal scoring curve; and 8, presenting the abnormal label sequence and the abnormal scoring curve on a visual interface. According to the invention, based on multi-channel voiceprint modeling and the improved DeepAR model, accurate identification of the abnormal state of the equipment is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power equipment monitoring, and particularly relates to a GIS device abnormality detection method based on AI voiceprint recognition. BACKGROUND

[0002] With the growing demand for intelligent sensing technology and equipment health management, acoustic anomaly detection and state trend evaluation during the operation of GIS (Gas Insulated Switchgear) equipment have become an important research direction in the field of power system operation and maintenance. Existing GIS equipment anomaly detection methods mainly rely on partial discharge, current oscillation and other electrical quantity monitoring, but the deep utilization of sound waves as an important physical carrier of internal structure vibration and discharge activity of the equipment is still limited, and the following problems exist.

[0003] Traditional sound wave analysis methods mostly use Fourier transform or envelope demodulation and other frequency domain methods to process sound signals, lack the modeling ability of nonlinear disturbance characteristics in voiceprint spectrum, and are difficult to capture potential abnormal behavior patterns; for the multi-source reverberation sound environment in complex industrial sites, the existing sound source separation technology is insufficient in suppressing non-target sound sources in the frequency spectrum space, resulting in the target voiceprint features being masked or distorted; in the voiceprint time series prediction link, the commonly used statistical regression method or shallow neural network model is slow in responding to abnormal changes, lacks the ability to model multi-modal uncertainty, and is prone to misjudgment or false reporting; in terms of abnormal trend output, most existing systems stay at the single-point abnormality labeling level and do not form a trend curve with continuous time evaluation capability, making it difficult to support the evolution perception and active early warning of the running state of GIS equipment.

[0004] Therefore, how to provide a GIS device abnormality detection method based on AI voiceprint recognition is a problem that those skilled in the art need to solve. SUMMARY

[0005] One object of the present application is to provide a GIS device abnormality detection method based on AI voiceprint recognition. The present application combines multi-channel voiceprint spectrum modeling, mask prediction and improved DeepAR model to realize accurate identification and trend analysis of abnormal states during the operation of GIS equipment. By introducing a spectral residual activation function to enhance the abnormality perception ability of the improved DeepAR model, and using multiple Gaussian distribution components to model the time series voiceprint features, the robustness, identification accuracy and real-time response ability of the system in complex sound field environments are effectively improved, and the system is suitable for high-reliability power equipment state monitoring and intelligent early warning scenarios.

[0006] According to the GIS device abnormality detection method based on AI voiceprint recognition of the present application, the following steps are included:

[0007] Step one: collect the original acoustic wave signal in the running process of the GIS equipment, and pre-process to obtain the pre-processed acoustic wave signal;

[0008] Step two: short-time Fourier transform is performed on the pre-processed acoustic wave signal to obtain a complex spectrum matrix, and the phase information is reconstructed based on the joint of the minimum phase reconstruction algorithm and the Griffin-Lim phase recovery algorithm to generate an initial voiceprint spectrum tensor;

[0009] Step three: based on the initial voiceprint spectrum tensor, a mask prediction network is used to extract the target device sound source mask value, and the non-target sound source is subjected to spectral attenuation processing to output a target voiceprint spectrum tensor;

[0010] Step four: using residual convolution structure and channel attention mechanism to extract features from the target voiceprint spectrum tensor, outputting a time series voiceprint feature tensor;

[0011] Step five: input the time series voiceprint feature tensor into the improved DeepAR model, the improvement of the improved DeepAR model is that the spectral residual activation function is introduced, and the parameter set of multiple Gaussian distribution components is obtained based on the activation result;

[0012] Step six: based on the parameter set of multiple Gaussian distribution components, a prediction probability density function is constructed, the confidence score of the current time step voiceprint observation value is calculated, the preset confidence interval threshold is used for anomaly judgment, and an abnormal label sequence is output;

[0013] Step seven: sliding window aggregation is performed on the abnormal label sequence to output an abnormal score curve reflecting the state change trend of the GIS equipment;

[0014] Step eight: the abnormal label sequence and the abnormal score curve are presented in a graphical form on a visualization interface.

[0015] Optionally, the pre-processing specifically includes:

[0016] The collected original acoustic wave signal is normalized by dividing the amplitude of each sampling point in the original acoustic wave signal by the maximum value of the absolute value of the amplitude of all sampling points in the original acoustic wave signal, so that the amplitude range of the normalized original acoustic wave signal is limited to -1 to 1;

[0017] First-order differential high-pass filtering is applied to the normalized original acoustic wave signal, and the difference between the amplitude of the current sampling point and the amplitude of the previous sampling point is calculated as the filtering output at the current time, and a high-pass filtered acoustic wave signal is obtained;

[0018] The high-pass filtered acoustic wave signal is dynamically gain-adjusted, the average value of the amplitude square of all sampling points in the current sliding window is calculated based on a sliding window of a set length, a gain coefficient is obtained by taking a set scaling factor as the numerator and the sum of the average value and a set non-zero smoothing term as the denominator, and the gain coefficient is used for scaling processing of the amplitude of the acoustic wave signal in the current sliding window.

[0019] Optionally, the step two specifically includes:

[0020] The preprocessed acoustic wave signal is subjected to short-time Fourier transform, the preprocessed acoustic wave signal is framed by using a sliding window of a set length, the frame length is set to 1024 sampling points, and the overlap length between adjacent frames is set to 256 sampling points, and a corresponding complex spectrum matrix is obtained;

[0021] The complex amplitude of each frequency point is extracted from the complex spectrum matrix, the amplitude spectrum of the frame is constructed based on the absolute value of the complex amplitude, and the amplitude spectrum is subjected to logarithmic transformation to obtain the logarithmic amplitude spectrum of each frame;

[0022] The logarithmic amplitude spectrum is input into a minimum phase reconstruction algorithm, the Hilbert transform is used to analytically continue the logarithmic amplitude spectrum in the frequency domain, and a minimum phase spectrum is constructed;

[0023] An initial complex spectrum estimate is constructed based on the minimum phase spectrum and the corresponding amplitude spectrum, and a Griffin-Lim phase recovery algorithm is input for iterative optimization, in each iteration, the current complex spectrum estimate is restored to a time domain signal by performing inverse short-time Fourier transform, the short-time Fourier transform is performed again on the time domain signal to extract phase information, the phase information is combined with the amplitude spectrum to construct a new complex spectrum estimate input for the next iteration, until a preset maximum iteration number is reached, and finally a reconstructed phase spectrum is obtained;

[0024] The reconstructed phase spectrum and the amplitude spectrum are combined into a final complex spectrum tensor, and a final logarithmic amplitude spectrum, a final phase difference spectrum and a final spectral envelope map are extracted therefrom, and an initial voiceprint spectrum tensor of three channels is generated by splicing.

[0025] Optionally, the mask prediction network adopts an encoder-decoder structure, the encoder is composed of multiple two-dimensional convolution layers, batch normalization layers and ReLU activation functions, is used for extracting time-frequency features in the initial voiceprint spectrum tensor, the decoder is composed of multiple two-dimensional deconvolution layers and Sigmoid activation functions, is used for generating a corresponding target device sound source mask value for each frequency point, and the target device sound source mask value ranges from 0 to 1;

[0026] The target device sound source mask value is multiplied element by element with the initial voiceprint spectrum tensor, and an energy suppression is performed on an amplitude value of a frequency point of the target device sound source mask value smaller than a set attenuation threshold by multiplying the amplitude value by a preset scaling factor, to generate a target voiceprint spectrum tensor.

[0027] Optionally, the step four is specifically:

[0028] The target voiceprint spectrum tensor is sequentially input into a plurality of residual convolution units for feature extraction, each residual convolution unit is composed of a first convolution layer, a second convolution layer and an identity skip connection path, the first convolution layer is used to extract local time-frequency features, and the second convolution layer performs linear feature transformation on an output of the first convolution layer.

[0029] When the input tensor of the residual convolution unit and the output tensor of the second convolution layer are consistent in the number of channels, the identity skip connection path is an identity mapping path directly connecting the input and the output;

[0030] When the number of channels of the input tensor of the residual convolution unit and the output tensor of the second convolution layer is inconsistent, the identity skip connection path includes a one-dimensional convolution layer, which is used to adjust the number of channels of the input tensor of the residual convolution unit to the number of channels of the output tensor of the second convolution layer;

[0031] The output tensor of the identity skip connection path and the output tensor of the second convolution layer are added element by element in the channel dimension to generate a residual output feature map;

[0032] A channel attention mechanism is connected after each residual convolution unit, the channel attention mechanism performs global average pooling on each channel in the time and frequency dimensions, outputs a channel response weight through a two-layer fully connected network and a Sigmoid activation function, and multiplies the channel response weight and the residual output feature map element by element in the channel dimension to obtain a weighted feature map;

[0033] The weighted feature map is stacked in a frame sequence order and kept consistent in the channel dimension to generate a time-series voiceprint feature tensor.

[0034] Optionally, the step five is specifically:

[0035] The time-series voiceprint feature tensor is input into an improved DeepAR model, the improved DeepAR model includes a group of recurrent neural network units unfolded in time steps, adopts a gated recurrent unit for sequence modeling, and introduces a spectral residual activation function in an output stage to obtain an activation result.

[0036] The calculation process of the spectrum residual activation function includes: performing fast Fourier transform on the gated recurrent unit output of the current time step to obtain a current frame spectrum; obtaining a spectrum mean of several frames before the current time step as a reference background spectrum with a preset window length; calculating a point-by-point difference between the current frame spectrum and the reference background spectrum to form a residual spectrum; and reducing the residual spectrum to a time domain through inverse Fourier transform as an activation result;

[0037] The activation result is input into a parameter generation unit to output a parameter set of multiple Gaussian distribution components, including a mean, a standard deviation and a weight coefficient of each Gaussian distribution component, wherein:

[0038] The mean is extracted from the activation result through a set of linear fully connected layers;

[0039] The standard deviation is obtained by performing exponential transformation on the activation result to ensure non-negativity;

[0040] The weight coefficient is obtained by performing a Softmax normalization operation on the activation result to ensure that the sum is 1.

[0041] Optionally, the recurrent neural network unit is composed of a gated recurrent unit, which is unfolded by time step and performs state transmission between adjacent time steps, receives an input of the current time step and a hidden state of the previous time step, controls the preservation and update of state information by using an update gate and a reset gate, outputs a hidden state of the current time step, and models a time sequence dependency relationship as a gated recurrent unit output of the current time step.

[0042] Optionally, the step six is specifically:

[0043] Based on the parameter set of multiple Gaussian distribution components, the mean, the standard deviation and the weight coefficient of each Gaussian distribution component are obtained, a prediction probability density function is constructed, and a confidence score of a voiceprint observation value of the current time step is calculated;

[0044] The prediction probability density function is defined as:

[0045]

[0046] wherein, p(x t | x <t ) represents a confidence score of a voiceprint observation value of time step t, x t represents a voiceprint feature observation value of time step t, x <t represents a historical voiceprint feature sequence before time step t, M represents a total number of Gaussian distribution components, represents a weight coefficient of the kth Gaussian distribution component at time step t, and satisfies and each represents the mean of the kth Gaussian distribution component at time step t, represents the standard deviation of the kth Gaussian distribution component at time step t, and needs to satisfy

[0047] represents the Gaussian distribution function:

[0048]

[0049] wherein exp represents an exponential function;

[0050] The confidence score is compared with a preset confidence interval threshold, if the confidence score is less than the preset confidence interval threshold, the current time step label value is recorded as 1, indicating an abnormal label, otherwise 0, indicating a normal label, and an abnormal judgment label sequence is output, the abnormal judgment label sequence is an array of 0 and 1 label values arranged in time step order, indicating the abnormal distribution in the GIS device running process.

[0051] Optionally, the step seven is specifically:

[0052] The abnormal judgment label sequence is subjected to a sliding window aggregation processing, a fixed length time sliding window is set, the time sliding window is slid on the abnormal judgment label sequence with a set step length, the label values in each time sliding window are counted, and the ratio of the number of label values being 1 in the time sliding window to the length of the time sliding window is taken as an abnormal score value of a center time step of the time sliding window;

[0053] The abnormal score values are arranged in time sequence of the center time steps of the time sliding windows, and an abnormal score curve is generated, the abnormal score curve is used for reflecting the intensive degree and trend change of local abnormalities in the GIS device running process.

[0054] The present application has the following beneficial effects:

[0055] The application constructs an initial acoustic fingerprint spectrum tensor, separates sound sources through a mask prediction network, extracts time-frequency joint features, and integrates a multi-level structure of an improved DeepAR model based on a spectral residual activation function, so as to solve the problems of non-target sound source interference in GIS device operation sound wave signals in a complex industrial noise field, difficulty in separating multi-band disturbances, and low abnormality detection sensitivity. An intelligent abnormality detection method with sound source perception, spectral reconstruction, and probability prediction capability is proposed. In the sound wave signal processing link, the minimum phase reconstruction algorithm and the Griffin-Lim phase recovery algorithm are combined to reconstruct the final complex spectrum tensor, construct a three-channel acoustic fingerprint spectrum tensor, and retain more frequency domain structure information. In the sound source identification link, a mask prediction network is used to learn the target device sound source mask, realize spectral-level attenuation separation of multi-source acoustic fingerprints, and effectively suppress non-device sound interference. In the time series modeling aspect, a feature extraction path integrating residual convolution structure and channel attention mechanism is constructed to extract the time-frequency evolution mode of the target acoustic fingerprint, and a spectral residual activation function is introduced to highlight the abnormal spectral disturbance features. Further, a multi-Gaussian mixture model is used to generate a time step prediction probability distribution, and an abnormality discrimination label sequence is generated by combining a confidence interval threshold. Finally, an abnormality score curve is output by sliding window aggregation to realize trend perception and accurate labeling of potential operation abnormalities of GIS devices, effectively improving the real-time performance, stability, and noise robustness of the detection system. BRIEF DESCRIPTION OF DRAWINGS

[0056] The accompanying drawings are included to provide a further understanding of the application, and constitute a part of the specification, illustrate the application, and are used to explain the application together with the embodiments of the application, and do not constitute a limitation on the application. In the drawings:

[0057] Figure 1 A flowchart of a GIS device abnormality detection method based on AI acoustic fingerprint recognition proposed by the application;

[0058] Figure 2 A flowchart of initial acoustic fingerprint spectrum tensor generation of a GIS device abnormality detection method based on AI acoustic fingerprint recognition proposed by the application;

[0059] Figure 3 A structure diagram of an improved DeepAR model of a GIS device abnormality detection method based on AI acoustic fingerprint recognition proposed by the application. DETAILED DESCRIPTION

[0060] The application will now be described in further detail with reference to the drawings. These drawings are simplified schematic diagrams that only schematically show the basic structure of the application, and therefore only show the components related to the application.

[0061] REFERENCE Figures 1-3 A GIS device abnormality detection method based on AI acoustic fingerprint recognition includes the following steps:

[0062] Step one: collect the original sound wave signal in the running process of GIS equipment and pre-process to obtain the pre-processed sound wave signal;

[0063] Step two: perform short-time Fourier transform on the pre-processed sound wave signal to obtain a complex spectrum matrix, and jointly reconstruct phase information based on a minimum phase reconstruction algorithm and a Griffin-Lim phase recovery algorithm to generate an initial voiceprint spectrum tensor;

[0064] Step three: based on the initial voiceprint spectrum tensor, a mask prediction network is used to extract target device sound source mask values, and non-target sound sources are subjected to spectral attenuation processing to output a target voiceprint spectrum tensor;

[0065] Step four: use residual convolution structure and channel attention mechanism to extract features from the target voiceprint spectrum tensor to output a time series voiceprint feature tensor;

[0066] Step five: input the time series voiceprint feature tensor into an improved DeepAR model, wherein the improvement of the improved DeepAR model is that a spectral residual activation function is introduced, and based on the activation result, a parameter set of multiple Gaussian distribution components is obtained;

[0067] Step six: based on the parameter set of multiple Gaussian distribution components, a prediction probability density function is constructed, the confidence score of the current time step voiceprint observation value is calculated, the preset confidence interval threshold is used for anomaly judgment, and an abnormal label sequence is output;

[0068] Step seven: sliding window aggregation is performed on the abnormal label sequence to output an abnormal score curve reflecting the state change trend of the GIS equipment;

[0069] Step eight: the abnormal label sequence and the abnormal score curve are presented in a graphical form on a visualization interface.

[0070] In the embodiment, the pre-processing is specifically:

[0071] The collected original sound wave signal is normalized by dividing the amplitude of each sampling point in the original sound wave signal by the maximum value of the absolute value of the amplitude of all sampling points in the original sound wave signal, so that the amplitude range of the normalized original sound wave signal is limited to -1 to 1;

[0072] First-order differential high-pass filtering is applied to the normalized original sound wave signal, and the difference between the amplitude of the current sampling point and the amplitude of the previous sampling point is calculated as the filtering output at the current time to obtain a high-pass filtered sound wave signal;

[0073] The high-pass filtered acoustic signal is dynamically gain-adjusted, the average value of the amplitude square of all sampling points in a current sliding window is calculated based on a sliding window of a set length, a gain coefficient is obtained by taking a set scaling factor as the numerator and the sum of the average value and a set non-zero smoothing term as the denominator, and the gain coefficient is used for scaling processing of the amplitude of the acoustic signal in the current sliding window.

[0074] To avoid the case that the denominator is zero when the amplitudes of all sampling points in the sliding window are zero, and to improve the numerical stability in the dynamic gain adjustment process, the smoothing term is set to a non-zero positive constant, which is used to ensure the numerical safety of the gain coefficient calculation process and the robustness of the system. In practical applications, the value of the smoothing term is generally set to a positive real number between 10 -6 and 10 -3 .

[0075] In the embodiment, the step two is specifically:

[0076] The preprocessed acoustic signal is subjected to short-time Fourier transform, the preprocessed acoustic signal is framed by using a sliding window of a set length, the frame length is set to 1024 sampling points, and the overlap length between adjacent frames is set to 256 sampling points, and a corresponding complex spectrum matrix is obtained;

[0077] The complex amplitude of each frequency point is extracted from the complex spectrum matrix, the amplitude spectrum of the frame is constructed based on the absolute value of the complex amplitude, and the amplitude spectrum is subjected to logarithmic transformation to obtain the logarithmic amplitude spectrum of each frame;

[0078] The logarithmic amplitude spectrum is input into a minimum phase reconstruction algorithm, and the Hilbert transform is used to analytically continue the logarithmic amplitude spectrum in the frequency domain to construct a minimum phase spectrum;

[0079] In the process of analytically continuing the logarithmic amplitude spectrum, the logarithmic amplitude spectrum of each frame is subjected to even continuation processing by using a frequency domain symmetric expansion method, so that it satisfies the conjugate symmetry condition in the negative frequency part to form a complete spectrum structure. On this basis, the logarithmic amplitude spectrum after continuation is analytically continued by using the Hilbert transform to obtain a minimum phase spectrum corresponding to the input logarithmic amplitude spectrum. The phase spectrum is the analytic signal phase part satisfying the minimum phase characteristic;

[0080] constructing an initial complex spectrum estimation based on the minimum phase spectrum and the corresponding amplitude spectrum, inputting a Griffin-Lim phase recovery algorithm for iterative optimization, in each iteration, performing an inverse short-time Fourier transform to restore the current complex spectrum estimation into a time-domain signal, and re-performing a short-time Fourier transform on the time-domain signal to extract phase information therefrom, recombining the phase information with the amplitude spectrum to construct a new complex spectrum estimation input for the next iteration, until a preset maximum number of iterations is reached, and finally obtaining a reconstructed phase spectrum;

[0081] In the Griffin-Lim algorithm, the initial complex spectrum estimation is an artificially synthesized spectrum estimation value, the amplitude of which comes from the amplitude spectrum extracted in the foregoing (obtained from the complex spectrum matrix), and the phase comes from the minimum phase spectrum;

[0082] combining the reconstructed phase spectrum and the amplitude spectrum into a final complex spectrum tensor, and extracting therefrom a final log amplitude spectrum, a final phase difference spectrum, and a final spectral envelope map, respectively, to splice an initial voiceprint spectrum tensor of three channels.

[0083] The final log amplitude spectrum refers to a two-dimensional spectrum obtained by taking the absolute value of the amplitude part of the reconstructed final complex spectrum tensor and performing a logarithmic transformation thereon, representing the energy distribution characteristics of each frame and each frequency point on a logarithmic scale; the final phase difference spectrum is a two-dimensional matrix calculated according to the phase difference values of the same frequency point between adjacent frames in the final complex spectrum tensor, used to depict the phase dynamic change information of the voiceprint signal on the time axis; and the final spectral envelope map is a contour map obtained by applying a Hilbert envelope extraction algorithm to the amplitude spectrum of each frame, reflecting the energy distribution trend of each frame of signal in the frequency domain. The above three maps are spliced in the channel dimension to form an initial voiceprint spectrum tensor of three channels.

[0084] In the embodiment, the mask prediction network adopts an encoder-decoder structure, the encoder is composed of multiple two-dimensional convolution layers, batch normalization layers, and ReLU activation functions, and is used to extract time-frequency features in the initial voiceprint spectrum tensor, the decoder is composed of multiple two-dimensional deconvolution layers and Sigmoid activation functions, and is used to generate a corresponding target device sound source mask value for each frequency point, the target device sound source mask value has a value range of 0 to 1;

[0085] The mask prediction network is trained in a supervised learning manner, training data is composed of an initial voiceprint spectrum tensor containing a target device sound source and a non-target sound source mixed signal, and label data is a target sound source mask at a corresponding frame frequency point. During network training, the input spectrum is forward propagated through an encoder-decoder structure, the output predicted mask map is compared with the real mask point by point, and the error is back propagated using a mean square error loss function to update the network parameters. After training is completed, the mask prediction network has the ability to identify the target device sound source region from the mixed sound source, and is used for spectral masking processing in the inference stage.

[0086] The target device sound source mask value is multiplied element by element with the initial voiceprint spectrum tensor, and the amplitude value of the frequency point whose target device sound source mask value is less than a set attenuation threshold is multiplied by a preset scaling factor for energy suppression to generate a target voiceprint spectrum tensor.

[0087] In the embodiment, the step four is specifically:

[0088] The target voiceprint spectrum tensor is sequentially input into a plurality of residual convolution units for feature extraction, each residual convolution unit is composed of a first convolution layer, a second convolution layer and an identity skip connection path, the first convolution layer is used to extract local time-frequency features, and the second convolution layer performs linear feature transformation on the output of the first convolution layer;

[0089] When the input tensor of the residual convolution unit and the output tensor of the second convolution layer are consistent in the number of channels, the identity skip connection path is an identity mapping path directly connecting the input and the output;

[0090] When the number of channels of the input tensor of the residual convolution unit and the output tensor of the second convolution layer is inconsistent, the identity skip connection path includes a one-dimensional convolution layer for adjusting the number of channels of the input tensor of the residual convolution unit to the number of channels of the output tensor of the second convolution layer;

[0091] The output tensor of the identity skip connection path and the output tensor of the second convolution layer are added element by element in the channel dimension to generate a residual output feature map;

[0092] A channel attention mechanism is connected after each residual convolution unit, the channel attention mechanism performs global average pooling on each channel in the time and frequency dimensions, outputs channel response weights through a two-layer fully connected network and a Sigmoid activation function, and multiplies the channel response weights and the residual output feature map element by element in the channel dimension to obtain a weighted feature map;

[0093] The weighted feature map is stacked in frame sequence order, and the channel dimension is kept consistent to generate a time sequence voiceprint feature tensor.

[0094] In this embodiment, step five is specifically:

[0095] The time sequence voiceprint feature tensor is input into an improved DeepAR model, the improved DeepAR model comprises a set of recurrent neural network units unfolded in time steps, adopts a gated recurrent unit for sequence modeling, and introduces a spectral residual activation function in an output stage to enhance the abnormal perception ability of the voiceprint high-frequency disturbance, and obtains an activation result;

[0096] The calculation process of the spectral residual activation function comprises: performing fast Fourier transform on the output of the gated recurrent unit at the current time step to obtain a current frame spectrum; a spectral mean of a plurality of frames before the current time step is obtained as a reference background spectrum with a preset window length; a point-by-point difference between the current frame spectrum and the reference background spectrum is calculated to form a residual spectrum; and the residual spectrum is restored to a time domain by inverse Fourier transform as an activation result;

[0097] The activation result is input into a parameter generation unit to output a parameter set of a plurality of Gaussian distribution components, including a mean, a standard deviation and a weight coefficient of each Gaussian distribution component, wherein:

[0098] The mean is extracted from the activation result through a set of linear fully connected layers;

[0099] The standard deviation is obtained by performing exponential transformation on the activation result, ensuring non-negativity;

[0100] The weight coefficient is obtained by performing a Softmax normalization operation on the activation result, ensuring that the sum is 1;

[0101] In the improved DeepAR model adopted in the application, the prediction output is modeled in the form of a mixed Gaussian distribution to more accurately represent the complex and multi-modal conditional probability distribution in the voiceprint time sequence data. The mixed Gaussian distribution is one of the main prediction distribution forms supported by DeepAR, and is suitable for modeling asymmetric, multi-peak or highly uncertain continuous variables. Compared with the traditional single Gaussian distribution, the mixed Gaussian distribution has stronger fitting ability and can more effectively capture the probability mode of the time sequence voiceprint feature in the abnormal state change process.

[0102] In the model output stage, the time sequence voiceprint feature tensor is first input into a spectral residual activation function to extract important prediction signals, and then the activation result thereof is input into a parameter generation unit to perform a plurality of linear transformation operations, thereby generating a parameter set of a plurality of Gaussian distribution components.

[0103] In the embodiment, the recurrent neural network unit is composed of a gated recurrent unit, which is expanded by time steps and performs state transmission between adjacent time steps, receives input of the current time step and hidden state of the previous time step, controls the reservation and update of state information by using an update gate and a reset gate, outputs the hidden state of the current time step, and models the time sequence dependency relationship as the output of the gated recurrent unit of the current time step.

[0104] In the embodiment, the step six is specifically:

[0105] Based on the parameter set of the plurality of Gaussian distribution components, the mean, the standard deviation and the weight coefficient of each Gaussian distribution component are obtained, the prediction probability density function is constructed, and the confidence score of the voiceprint observation value of the current time step is calculated.

[0106] The prediction probability density function is defined as:

[0107]

[0108] wherein, p(x t | x <t ) represents the confidence score of the voiceprint observation value of the time step t, x t represents the voiceprint feature observation value of the time step t, x <t represents the historical voiceprint feature sequence before the time step t, and M represents the total number of Gaussian distribution components, represents the weight coefficient of the kth Gaussian distribution component at the time step t, which is used to describe the contribution proportion of the kth Gaussian distribution component in the overall distribution, and satisfies and each represents the mean of the kth Gaussian distribution component at the time step t, which reflects the expected value of the voiceprint feature corresponding to the kth Gaussian distribution component; represents the standard deviation of the kth Gaussian distribution component at the time step t, which is used to describe the distribution discrete degree of the kth Gaussian distribution component, and needs to satisfy

[0109] represents the Gaussian distribution function:

[0110]

[0111] wherein, exp represents the exponential function;

[0112] The confidence score is compared with a preset confidence interval threshold, if the confidence score is less than the preset confidence interval threshold, the current time step label value is recorded as 1, indicating an abnormal label; otherwise, it is recorded as 0, indicating a normal label, and an abnormal judgment label sequence is output, the abnormal judgment label sequence is an array of 0 and 1 label values arranged in time step order, indicating the abnormal distribution in the GIS device running process.

[0113] In the embodiment, the step seven is specifically:

[0114] The sliding window aggregation processing is performed on the abnormal judgment label sequence, a fixed length time sliding window is set, the time sliding window is slid on the abnormal judgment label sequence with a set step, the label values in each time sliding window are counted, and the ratio of the number of label values being 1 in the time sliding window to the length of the time sliding window is calculated as an abnormal score value of a center time step of the time sliding window;

[0115] The abnormal score values are arranged in time order of the center time steps of the time sliding windows, and an abnormal score curve is generated, the abnormal score curve is used to reflect the density and trend change of local abnormalities in the GIS device running process.

[0116] Embodiment 1:

[0117] In order to verify the feasibility of the application in implementation, the application is applied to the intelligent abnormal detection task of the running state of GIS (gas insulated metal enclosed switchgear) of a certain substation. The station has been built for many years, and adopts double busbar wiring mode for operation, and the equipment is in long period operation state. In the previous operation process, bus discharge and intermittent noise abnormality problems occurred many times, and the traditional manual inspection and electrical detection means failed to identify and locate the abnormal source point in time. Therefore, the application is deployed and implemented to realize high-precision identification and trend monitoring of the equipment state.

[0118] In the implementation process, three high-sensitivity sound wave sensor arrays are arranged near the GIS device shell, the sampling frequency is set to 48 kHz, and the embedded acquisition module is used to continuously acquire the running sound wave signal. In order to improve the signal quality, the original sound wave signal is subjected to multi-stage preprocessing such as normalization, first-order difference high-pass filtering and sliding gain dynamic adjustment, so as to reduce background noise interference and enhance the weak soundprint change characteristics.

[0119] The preprocessed signal is modeled in the frequency domain by using short-time Fourier transform, and the phase information is restored by the minimum phase reconstruction algorithm combined with the Griffin-Lim algorithm iteration, and finally the three-channel initial voiceprint spectrum tensor is obtained. After the target device sound source is extracted and the interference component is suppressed by using the mask prediction network, the target voiceprint spectrum tensor is generated, and further through the residual convolution structure and the channel attention mechanism, the time sequence voiceprint feature tensor is extracted.

[0120] The extracted time sequence voiceprint feature tensor is input into the improved DeepAR model, which introduces a spectral residual activation function in the output stage to enhance the response ability to abnormal spectral structure changes. By multi-Gaussian mixture modeling, the prediction probability density distribution of each time step is obtained. The prediction probability density function is constructed, the confidence score of the current time step voiceprint observation value is calculated, and combined with the set confidence interval threshold, the abnormal judgment label sequence composed of 0 / 1 labels is generated, and further through the sliding window aggregation, the abnormal score curve is formed, which is used to evaluate the abnormal aggregation degree and trend.

[0121] The present application has carried out systematic training test on the collected samples, and the normal data samples are more than 80,000 segments, and the abnormal samples are nearly 2,000 segments. The abnormality is marked by an acoustic expert and verified by a local detector. In the training process, the Adam optimizer is used, and the appropriate batch size and learning rate are set to complete the model fitting. Compared with the basic model, it has significant performance advantage.

[0122] Table 1 abnormal detection model performance comparison table

[0123]

[0124] According to the data results of the above table 1 abnormal detection model performance comparison table, it can be seen that the improved DeepAR model (introducing spectral residual activation) is better than other comparison methods in each abnormal detection performance index, which shows its significant advantage in the GIS device voiceprint abnormal detection task.

[0125] From the accuracy index, the accuracy of the improved DeepAR model reached 96.84%, which is higher than that of the conventional DeepAR model (91.72%), the BiLSTM+attention mechanism model (89.56%) and the rule-based detection algorithm (76.88%), which shows that it can effectively distinguish between abnormal and normal states and reduce misjudgment.

[0126] In terms of recall rate, the improved DeepAR model achieved 94.02%, which is significantly higher than other models (the conventional DeepAR model is 86.57%, the BiLSTM+attention mechanism is 85.21%, and the rule algorithm is only 62.34%), which reflects its stronger recognition ability to real abnormal events and lower omission rate.

[0127] The F1 value integrates the accuracy and recall, and the F1 value of the improved DeepAR model reaches 95.41, showing excellent comprehensive detection performance, while the F1 values of the conventional DeepAR, BiLSTM+attention mechanism and rule method are 89.06, 87.33 and 68.78 respectively, all of which are inferior to the former.

[0128] In terms of the AUC (area under the curve) index, the improved model is 0.983, the conventional DeepAR model is 0.931, the BiLSTM+attention mechanism is 0.917, and the rule-based algorithm is 0.743, further verifying that the model after introducing the spectral residual activation mechanism has more stable and reliable identification performance in the entire discrimination range.

[0129] Further, the abnormal score curve key data of the detection situation of a certain running period is extracted as shown in Table 2.

[0130] Table 2 Abnormal score key point example table

[0131] Time point segment Confidence score Discrimination label Anomaly score (sliding window) Segment A 0.0062 Anomaly 1 0.80 Segment B 0.0094 Anomaly 1 0.83 Segment C 0.0115 Anomaly 1 0.87 Segment D 0.0571 Normal 0 0.67 Segment E 0.0612 Normal 0 0.42 Segment F 0.0083 Anomaly 1 0.70 Segment G 0.0054 Anomaly 1 0.76

[0132] By analyzing the data of each time point segment in Table 2 above, the determination basis of abnormal points and their corresponding sliding window abnormal score trends can be more clearly identified.

[0133] In segments A, B and C, the confidence scores are 0.0062, 0.0094 and 0.0115 respectively, all of which are much lower than the pre-set confidence interval threshold 0.02, so they are all marked as abnormal, with a discrimination label of 1, and the corresponding sliding window abnormal scores are 0.80, 0.83 and 0.87 respectively, showing a high abnormal density. Among them, the abnormal score of segment C is the highest, indicating that it has aggregated many abnormal points in the local time range, and the abnormal trend is the most obvious.

[0134] The confidence scores of segments D and E are 0.0571 and 0.0612 respectively, which are significantly higher than 0.02, so they are determined to be normal, with a label of 0, and their corresponding sliding window abnormal scores are 0.67 and 0.42, both of which are lower than those of other abnormal segments. This shows that although there may be some weak abnormal fluctuations locally, the overall confidence is within the safe interval, and the system can effectively avoid misjudgment.

[0135] The confidence scores of segments F and G are 0.0083 and 0.0054 respectively, which are still lower than 0.02, so they are marked as abnormal. The sliding window abnormal scores are 0.70 and 0.76 respectively, indicating that the time periods they are in also contain a certain density of abnormal labels, and the system's sliding window mechanism accurately reflects the local abnormal aggregation phenomenon caused by the change in confidence.

[0136] The embodiment fully integrates the time sequence modeling and the probability prediction mechanism, introduces a spectral residual activation function by improving a DeepAR model, and significantly enhances the response capability to weak abnormal features in the voiceprint signal. Compared with a traditional method, the application has advantages in modeling depth and prediction accuracy, realizes probability modeling of voiceprint features at each time step by constructing a multi-Gaussian mixed distribution, and makes subsequent confidence calculation more sensitive and robust. Meanwhile, the abnormal label is aggregated through a sliding window mechanism, further smooths the time fluctuation of the detection result, and effectively describes the trend of the abnormal evolution over time. The overall process has high robustness, high sensitivity and real-time performance, can adapt to the dynamic changes of the GIS device state in a complex field environment, and has important engineering application value.

[0137] The above merely describes a preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can make equivalent replacements or changes according to the technical solution and the inventive concept of the present application within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A method for anomaly detection in GIS equipment based on AI voiceprint recognition, characterized in that, Includes the following steps: Step 1: Collect the raw acoustic signals during the operation of the GIS equipment and preprocess them to obtain the preprocessed acoustic signals; Step 2: Perform a short-time Fourier transform on the preprocessed acoustic signal to obtain the complex spectrum matrix, and reconstruct the phase information based on the minimum phase reconstruction algorithm and the Griffin-Lim phase recovery algorithm to generate the initial acoustic spectrum tensor. Step 3: Based on the initial acoustic signature tensor, a mask prediction network is used to extract the target device sound source mask value, and non-target sound sources are subjected to spectral attenuation processing to output the target acoustic signature tensor. Step 4: Use residual convolution structure and channel attention mechanism to extract features from the target voiceprint spectrum tensor and output temporal voiceprint feature tensor; Step 5: Input the temporal voiceprint feature tensor into the improved DeepAR model. The improvement of the improved DeepAR model is that it introduces a spectral residual activation function and obtains the parameter set of multiple Gaussian distribution components based on the activation result. Step 6: Based on the parameter set of multiple Gaussian distribution components, construct the prediction probability density function, calculate the confidence score of the current time step voiceprint observation, use the preset confidence interval threshold to judge anomalies, and output the anomaly discrimination label sequence; Step 7: Perform sliding window aggregation on the anomaly identification label sequence and output the anomaly scoring curve; Step 8: Present the abnormal label sequence and the abnormal score curve in a graphical form on the visualization interface.

2. The GIS equipment anomaly detection method based on AI voiceprint recognition according to claim 1, characterized in that, The preprocessing specifically includes: The acquired raw acoustic signal is normalized by dividing the amplitude of each sampling point in the raw acoustic signal by the maximum absolute value of the amplitude of all sampling points in the raw acoustic signal, so that the amplitude range of the normalized raw acoustic signal is limited to between -1 and 1. The normalized original acoustic signal is processed by a first-order differential high-pass filter. The difference between the amplitude of the current sampling point and the amplitude of the previous sampling point is calculated and used as the filtered output at the current moment to obtain the high-pass filtered acoustic signal. The high-pass filtered acoustic signal is dynamically gain-adjusted. The average value of the squared amplitudes of all sampling points within the current sliding window is calculated based on a sliding window of a set length. A gain coefficient is obtained by using a set scaling factor as the numerator and the sum of the average value and a set non-zero smoothing term as the denominator. This gain coefficient is used to scale the amplitude of the acoustic signal within the current sliding window.

3. The GIS equipment anomaly detection method based on AI voiceprint recognition according to claim 1, characterized in that, Step two specifically involves: The preprocessed acoustic signal is subjected to a short-time Fourier transform. The preprocessed acoustic signal is divided into frames using a sliding window of a set length. The frame length is set to 1024 sampling points, and the overlap length between adjacent frames is set to 256 sampling points. The corresponding complex spectrum matrix is ​​then obtained. The complex amplitude of each frequency point is extracted from the complex spectrum matrix, the amplitude spectrum of the frame is constructed based on the absolute value of the complex amplitude, and the amplitude spectrum is logarithmically transformed to obtain the logarithmic amplitude spectrum of each frame. The logarithmic amplitude spectrum is input into the minimum phase reconstruction algorithm, and the minimum phase spectrum is constructed by analytically extending the logarithmic amplitude spectrum in the frequency domain using the Hilbert transform. Based on the minimum phase spectrum and the corresponding amplitude spectrum, an initial complex spectrum estimate is constructed. This estimate is then input into the Griffin-Lim phase retrieval algorithm for iterative optimization. In each iteration, an inverse short-time Fourier transform is performed to restore the current complex spectrum estimate to a time-domain signal. The time-domain signal is then re-performed with a short-time Fourier transform to extract phase information. This phase information is then combined with the amplitude spectrum to construct a new complex spectrum estimate for the next iteration. This process continues until the preset maximum number of iterations is reached, ultimately resulting in the reconstructed phase spectrum. The reconstructed phase spectrum and the amplitude spectrum are combined into a final complex spectrum tensor, and the final logarithmic amplitude spectrum, the final phase difference spectrum and the final spectrum envelope are extracted from it respectively. They are then spliced ​​together to generate a three-channel initial voiceprint spectrum tensor.

4. The GIS equipment anomaly detection method based on AI voiceprint recognition according to claim 1, characterized in that, The mask prediction network adopts an encoder-decoder structure. The encoder consists of multiple two-dimensional convolutional layers, batch normalization layers, and ReLU activation functions, which are used to extract the time-frequency features in the initial acoustic spectrum tensor. The decoder consists of multiple two-dimensional deconvolutional layers and Sigmoid activation functions, which are used to generate a corresponding target device sound source mask value for each frequency point. The target device sound source mask value ranges from 0 to 1. The target device sound source mask value is multiplied element-wise with the initial acoustic spectrum tensor, and the frequency amplitude values ​​of the target device sound source mask value that are less than the set attenuation threshold are multiplied by a preset scaling factor to suppress energy, thereby generating the target acoustic spectrum tensor.

5. The GIS equipment anomaly detection method based on AI voiceprint recognition according to claim 1, characterized in that, Step four specifically involves: The target voiceprint spectrum tensor is sequentially input into multiple residual convolutional units for feature extraction. Each residual convolutional unit consists of a first convolutional layer, a second convolutional layer, and an identity jump path. The first convolutional layer is used to extract local time-frequency features, and the second convolutional layer performs linear feature transformation on the output of the first convolutional layer. When the input tensor of the residual convolutional unit and the output tensor of the second convolutional layer have the same number of channels, the identity jump path is an identity mapping path that directly connects the input and the output. When the number of channels of the input tensor of the residual convolutional unit is inconsistent with the number of channels of the output tensor of the second convolutional layer, the identity jump path includes a one-dimensional convolutional layer, which is used to adjust the number of channels of the input tensor of the residual convolutional unit to the number of channels of the output tensor of the second convolutional layer. The output tensor of the identity jump path is added element-wise to the output tensor of the second convolutional layer along the channel dimension to generate a residual output feature map. A channel attention mechanism is connected after each residual convolutional unit. The channel attention mechanism performs global average pooling on each channel in the time and frequency dimensions, outputs channel response weights through two fully connected layers and a sigmoid activation function, and multiplies the channel response weights with the residual output feature map element-wise in the channel dimension to obtain a weighted feature map. The weighted feature maps are stacked in frame order, maintaining consistency in the channel dimension, to generate a temporal voiceprint feature tensor.

6. The GIS equipment anomaly detection method based on AI voiceprint recognition according to claim 1, characterized in that, Step five specifically involves: The temporal voiceprint feature tensor is input into the improved DeepAR model, which includes a set of recurrent neural network units expanded by time step, uses gated recurrent units for sequence modeling, and introduces a spectral residual activation function in the output stage to obtain the activation result. The calculation process of the spectral residual activation function includes: performing a fast Fourier transform on the output of the gated recurrent unit at the current time step to obtain the spectrum of the current frame; obtaining the mean spectral value of several frames before the current time step as a reference background spectrum with a preset window length; calculating the frequency-by-frequency difference between the spectrum of the current frame and the reference background spectrum to form a residual spectrum; and restoring the residual spectrum to the time domain through an inverse Fourier transform as the activation result. The activation result is input into the parameter generation unit, which outputs a set of parameters for multiple Gaussian distribution components, including the mean, standard deviation, and weighting coefficient of each Gaussian distribution component, wherein: The mean is extracted from the activation results through a set of linear fully connected layers; The standard deviation is obtained by performing an exponential transformation on the activation result to ensure that it is non-negative; The weighting coefficients are obtained by performing a Softmax normalization operation on the activation results to ensure that the sum is 1.

7. The GIS equipment anomaly detection method based on AI voiceprint recognition according to claim 6, characterized in that, The recurrent neural network unit is composed of gated recurrent units, which are expanded by time steps and perform state transfer between adjacent time steps. The gated recurrent unit receives the input of the current time step and the hidden state of the previous time step, uses update gates and reset gates to control the retention and update of state information, outputs the hidden state of the current time step, and models the temporal dependency relationship as the output of the gated recurrent unit of the current time step.

8. The GIS equipment anomaly detection method based on AI voiceprint recognition according to claim 1, characterized in that, Step six specifically involves: Based on the parameter set of multiple Gaussian distribution components, the mean, standard deviation and weight coefficient of each Gaussian distribution component are obtained, a prediction probability density function is constructed, and the confidence score of the current time step voiceprint observation is calculated. The prediction probability density function is defined as: Where p(x) t |x <t ) represents the confidence score of the voiceprint observation at time step t, x t x represents the voiceprint feature observation value at time step t. <t This represents the historical voiceprint feature sequence before time step t, where M represents the total number of Gaussian distribution components. This represents the weight coefficient of the k-th Gaussian distribution component at time step t, satisfying... And each Let represent the mean of the k-th Gaussian distribution component at time step t. Let represent the standard deviation of the k-th Gaussian distribution component at time step t, and it must satisfy . Representing the Gaussian distribution function: Where exp represents the exponential function; The confidence score is compared with a preset confidence interval threshold. If the confidence score is less than the preset confidence interval threshold, the current time step label value is recorded as 1, indicating an abnormal label; otherwise, it is recorded as 0, indicating a normal label. An abnormal label sequence is output, which is an array of 0 and 1 label values ​​arranged in time step order, representing the abnormal distribution during the operation of the GIS equipment.

9. The GIS equipment anomaly detection method based on AI voiceprint recognition according to claim 1, characterized in that, Step seven specifically involves: A sliding window aggregation process is performed on the anomaly identification label sequence. A fixed-length time sliding window is set, and the time sliding window is slid on the anomaly identification label sequence with a set step size. The label values ​​in each time sliding window are counted, and the ratio of the number of label values ​​of 1 in the time sliding window to the length of the time sliding window is calculated as the anomaly score value of the center time step of the time sliding window. The anomaly score values ​​are arranged in chronological order according to the center time step of the time sliding window to generate an anomaly score curve. The anomaly score curve is used to reflect the density and trend changes of local anomalies during the operation of GIS equipment.

Images