A substation acoustic sensor frequency response calibration and fault location method

Abstract

The application discloses a substation acoustic sensor frequency response calibration and fault positioning method, comprising processing signals collected by each sensor in an acoustic sensor array to obtain frequency domain signals; applying a filter bank to the frequency domain signals obtained by each sensor to construct a complex domain deep residual calibration network, taking the frequency domain signals of all sensors as input, and outputting frequency domain filter coefficients corresponding to each sensor; training the complex domain deep residual calibration network according to a constructed loss function, obtaining calibrated frequency domain signals through the trained complex domain deep residual calibration network; and performing time domain reconstruction on sensor output according to the calibrated frequency domain signals to obtain calibrated signals. The method uses a complex domain deep residual network to replace a traditional linear adaptive FIR filter, can learn and compensate complex nonlinear frequency response mismatch between sensors, and has much higher calibration accuracy than existing linear methods.

Description

Technical Field

[0001] This invention belongs to the field of array signal processing technology, and relates to a method for frequency response calibration and fault location of acoustic sensors in substations. Background Technology

[0002] Distributed acoustic sensor arrays are widely used in substation fault location, calculating the fault location using the time difference of arrival (TDOA) of acoustic signals. Current technology assumes that all sensors in the array have the same frequency response. However, in reality, differences in sensors manufactured by different manufacturers / batches, tolerances in circuit components, and strong electromagnetic interference in substations can lead to frequency response mismatches in the distributed sensors, severely affecting fault location results.

[0003] The paper "Madhu N., Martin R. Low-complexity, robust algorithm for sensoranomaly detection and self-calibration of microphone arrays[J]. IET SignalProcessing, 2011, 5(1): 97-103." proposes a gain calibration method based on time-averaged power spectrum (TAPS). This method relies on a reference sensor, cannot correct phase deviations, and its performance degrades by 30% in reverberant environments.

[0004] Existing substation fault location technologies based on acoustic sensor arrays generally suffer from the following technical problems: (1) The amplitude and phase differences between sensors affect the accuracy of observation quantities used for fault location, such as sound wave intensity and arrival time, leading to an increase in fault location error in substations.

[0005] (2) Existing methods only consider the impact of amplitude differences between sensors on positioning, ignoring the interference of phase differences on positioning.

[0006] (3) The reference signal or the precise location of the sound source must be known, which is difficult to achieve on site. Summary of the Invention

[0007] The purpose of this invention is to provide a method for frequency response calibration and fault location of acoustic sensors in substations. This method uses a complex domain deep residual network to replace the traditional linear adaptive FIR filter, which can learn and compensate for complex nonlinear frequency response mismatches between sensors, and the calibration accuracy far exceeds that of existing linear methods.

[0008] The technical solution to achieve the purpose of this invention is as follows: A method for calibrating the frequency response of an acoustic sensor includes the following steps: The signals collected by each sensor in the acoustic sensor array are processed to obtain frequency domain signals; A filter bank is applied to the frequency domain signal obtained from each sensor to construct a complex domain deep residual calibration network. The network takes the frequency domain signals of all sensors as input and outputs the frequency domain filter coefficients corresponding to each sensor. The complex domain deep residual calibration network maps the input features to a high-dimensional space through an initial complex convolutional layer to obtain an initial complex feature map. After passing through multiple cascaded complex residual blocks, the feature map is subjected to global average pooling to aggregate global frequency domain information. The frequency domain filter coefficients of each sensor are generated through a complex fully connected layer. A loss function is constructed, and the complex domain deep residual calibration network is trained based on the constructed loss function to obtain the trained complex domain deep residual calibration network. The frequency domain filter coefficients are obtained through the trained complex domain deep residual calibration network. The sensor output is reconstructed in the time domain based on the obtained frequency domain filter coefficients to obtain the calibrated signal.

[0009] In the preferred embodiment, before processing the signals collected by each sensor in the acoustic sensor array, the following steps are also included: When distributed acoustic sensors are clustered together and a broadband white noise signal is played through a loudspeaker, its power spectral density satisfies:

[0010] in, N 0 is a constant. f For frequency, f L The lower limit frequency, f H This is the upper limit frequency.

[0011] In the preferred technical solution, the frequency domain signal obtained by each sensor is:

[0012] in, m = 0, 1, ..., M -1 is the frame index. M Total number of frames k = 0, 1, …, K -1 is the frequency index. K Number of effective frequency points w ( n ) is a rectangular window function. N Δ is the number of points in the Discrete Fourier Transform. N For frame shift.

[0013] In the preferred technical solution, obtaining the initial complex feature map includes: WillM The frequency domain signal of each training frame is organized into a complex tensor form. X ∈ M×I×K ,in Represents the complex field, any element in the tensor X. X ( m , i , k ) is represented as:

[0014] The input features are mapped to a high-dimensional space through an initial complex convolutional layer to obtain the initial complex feature map:

[0015] in, This represents the complex number convolution operation. For complex convolution kernels, C 0 represents the number of output channels. L 0 represents the kernel length. For complex bias, BN C (·) represents the normalization of complex numbers. s C (·) is the complex activation function.

[0016] In the preferred technical solution, the frequency domain filter coefficients for generating each sensor include the frequency domain filter coefficients: No. p The structure of a complex residual block is as follows:

[0017] in, For complex ReLU activation functions, , and The first p The complex convolution kernels of the first, second, and adjusted convolutional layers in each residual block. , and The first p The complex bias vectors of the first and second convolutional layers and the adjustment convolutional layer in each residual block; AdjustChannel(·) is the channel number adjustment operation. When the input and output channel numbers are the same, this operation is an identity mapping; when the channel numbers are different, channel number matching is performed through 1×1 convolution. H (p-0.5) , As an intermediate layer, For the first p The output layer of complex residual blocksK ' is the frequency domain feature dimension after pooling; go through P After a number of complex residual blocks, the feature map is subjected to global average pooling to aggregate global information in the frequency domain:

[0018] in, [:, :, k ] indicates taking medium depth k Elements in all rows and columns; The frequency domain filter coefficients for each sensor are generated using a complex fully connected layer:

[0019] in, Let be the output frequency domain filter coefficient matrix, the th i The row corresponds to the first i Frequency domain filter coefficients of each sensor , For the output layer activation function, and These represent the weight matrix and bias vector of the output fully connected layer, respectively.

[0020] In the preferred technical solution, the loss function is constructed as follows: The core loss function is designed as the sum of squares of the differences in the frequency domain outputs of all paired sensor combinations after calibration:

[0021] in, C ( I , 2) = I ( I -1) / 2 is the number of combinations of sensor pairs. I Number of sensors As the normalization factor, M Total number of frames K Number of effective frequency points; Introducing regularization constraints:

[0022] in, These are the frequency domain filter coefficients. l The regularization coefficient is used. The total loss function is: .

[0023] This invention also discloses a fault location method for a substation acoustic sensor, comprising: The calibrated signal is obtained using the aforementioned acoustic sensor frequency response calibration method. Obtain the observations from all acoustic sensors, construct an equation for the location of the substation fault, and solve the equation to obtain the coordinates of the substation fault point.

[0024] This invention further discloses an acoustic sensor frequency response calibration system for implementing the aforementioned acoustic sensor frequency response calibration method, comprising: The signal processing module processes the signals collected by each sensor in the acoustic sensor array to obtain frequency domain signals; The filter application and calibration network construction module applies a filter bank to the frequency domain signal obtained from each sensor, constructs a complex domain deep residual calibration network, takes the frequency domain signals of all sensors as input, and outputs the frequency domain filter coefficients corresponding to each sensor; the complex domain deep residual calibration network maps the input features to a high-dimensional space through an initial complex convolutional layer to obtain an initial complex feature map, and after passing through multiple cascaded complex residual blocks, performs global average pooling on the feature map to aggregate global frequency domain information, and generates the frequency domain filter coefficients of each sensor through a complex fully connected layer; The network training and calibration module constructs a loss function and trains the complex domain deep residual calibration network based on the constructed loss function to obtain the trained complex domain deep residual calibration network. The frequency domain filter coefficients are obtained through the trained complex domain deep residual calibration network. The signal reconstruction module reconstructs the sensor output in the time domain based on the obtained frequency domain filter coefficients to obtain the calibrated signal.

[0025] The present invention also discloses a computer storage medium storing a computer program thereon, wherein the computer executes the computer program to implement the aforementioned acoustic sensor frequency response calibration method.

[0026] The present invention also discloses an electronic device, including a memory and a processor, wherein the memory stores a computer program, the processor runs the computer program stored in the memory, and the computer program, when executed, implements the aforementioned acoustic sensor frequency response calibration method.

[0027] Compared with the prior art, the significant advantages of this invention are: 1. By using a complex domain deep residual network to replace the traditional linear adaptive FIR filter, it is possible to learn and compensate for the complex nonlinear frequency response mismatch between sensors, and the calibration accuracy far exceeds that of existing linear methods.

[0028] 2. The signal after calibration by the complex domain deep residual network z i ( nAs a DOA input, it eliminates DOA estimation errors caused by amplitude and phase mismatch between acoustic sensors, thereby improving the fault location accuracy of substations. Attached Figure Description

[0029] Figure 1 This is a flowchart of the acoustic sensor frequency response calibration method in this embodiment; Figure 2(a) shows the output signal waveforms of four acoustic sensors of the same manufacturer and model before calibration; Figure 2(b) shows the output signal waveform after calibration; Figure 3 The graph shows the DOA results of MUSIC power spectrum estimation before and after calibration. Figure 4 The diagram shows the results of locating fault points in the substation before and after calibration. Detailed Implementation

[0030] The principle of this invention is as follows: Frequency domain filters are used to calibrate the frequency response mismatch of acoustic sensors, and the frequency domain filters are solved based on complex domain deep residual networks; the frequency-domain calibrated signal is then processed. z i ( n As a DOA input, it eliminates DOA estimation errors caused by amplitude and phase mismatch between acoustic sensors, thereby improving the fault location effect in substations.

[0031] Example: like Figure 1 As shown, a method for calibrating the frequency response of an acoustic sensor includes the following steps: The signals collected by each sensor in the acoustic sensor array are processed to obtain frequency domain signals; A filter bank is applied to the frequency domain signal obtained from each sensor to construct a complex domain deep residual calibration network. The network takes the frequency domain signals of all sensors as input and outputs the frequency domain filter coefficients corresponding to each sensor. The complex domain deep residual calibration network maps the input features to a high-dimensional space through an initial complex convolutional layer to obtain an initial complex feature map. After passing through multiple cascaded complex residual blocks, the feature map is subjected to global average pooling to aggregate global frequency domain information. The frequency domain filter coefficients of each sensor are generated through a complex fully connected layer. A loss function is constructed, and the complex domain deep residual calibration network is trained based on the constructed loss function to obtain the trained complex domain deep residual calibration network. The frequency domain filter coefficients are obtained through the trained complex domain deep residual calibration network, and the calibrated frequency domain signal is obtained based on the obtained frequency domain filter coefficients. The sensor output is reconstructed in the time domain based on the obtained frequency domain filter coefficients to obtain the calibrated signal.

[0032] In a preferred embodiment, the method further includes processing the signals acquired by each sensor in the acoustic sensor array before proceeding with the following: When distributed acoustic sensors are clustered together and a broadband white noise signal is played through a loudspeaker, its power spectral density satisfies:

[0033] in, N 0 is a constant. f For frequency, f L The lower limit frequency, f H This is the upper limit frequency.

[0034] In a preferred embodiment, the frequency domain signal obtained by each sensor is:

[0035] in, m = 0, 1, ..., M -1 is the frame index. M Total number of frames k = 0, 1, …, K -1 is the frequency index. K Number of effective frequency points w ( n ) is a rectangular window function. N Δ is the number of points in the Discrete Fourier Transform. N For frame shift.

[0036] In a preferred embodiment, obtaining the initial complex feature map includes: Will M The frequency domain signal of each training frame is organized into a complex tensor form. X ∈ M×I×K ,in Represents the complex field, any element in the tensor X. X ( m , i , k ) is represented as:

[0037] The input features are mapped to a high-dimensional space through an initial complex convolutional layer to obtain the initial complex feature map:

[0038] in, This represents the complex number convolution operation. For complex convolution kernels, C 0 represents the number of output channels. L 0 represents the kernel length. For complex bias, BN C (·) represents the normalization of complex numbers. s C (·) is the complex activation function.

[0039] In a preferred embodiment, generating the frequency domain filter coefficients for each sensor includes the following: No. p The structure of a complex residual block is as follows:

[0040] in, For complex ReLU activation functions, , and The first p The complex convolution kernels of the first, second, and adjusted convolutional layers in each residual block. , and The first p The complex bias vectors of the first and second convolutional layers and the adjustment convolutional layer in each residual block; AdjustChannel(·) is the channel number adjustment operation. When the input and output channel numbers are the same, this operation is an identity mapping; when the channel numbers are different, channel number matching is performed through 1×1 convolution. H (p-0.5) , As an intermediate layer, For the first p The output layer of complex residual blocks K ' is the frequency domain feature dimension after pooling; go through P After a number of complex residual blocks, the feature map is subjected to global average pooling to aggregate global information in the frequency domain:

[0041] in, [:, :, k ] indicates taking medium depth k Elements in all rows and columns; The frequency domain filter coefficients for each sensor are generated using a complex fully connected layer:

[0042] in, Let be the output frequency domain filter coefficient matrix, the th i The row corresponds to the first iFrequency domain filter coefficients of each sensor , For the output layer activation function, and These represent the weight matrix and bias vector of the output fully connected layer, respectively.

[0043] In a preferred embodiment, constructing the loss function includes: The core loss function is designed as the sum of squares of the differences in the frequency domain outputs of all paired sensor combinations after calibration:

[0044] in, C ( I , 2) = I ( I -1) / 2 is the number of combinations of sensor pairs. I Number of sensors As the normalization factor, M Total number of frames K Number of effective frequency points; Introducing regularization constraints:

[0045] in, These are the frequency domain filter coefficients. l The regularization coefficient is used. The total loss function is: .

[0046] Another embodiment, a fault location method for a substation acoustic sensor, includes: The calibrated signal is obtained using the aforementioned acoustic sensor frequency response calibration method. Obtain the observations from all acoustic sensors, construct an equation for the location of the substation fault, and solve the equation to obtain the coordinates of the substation fault point.

[0047] It should be noted that using calibrated signals to estimate DOA using the MUSIC algorithm improves DOA estimation accuracy and thus enhances positioning performance. In fact, the accuracy of estimating any observation using calibrated signals should be improved. Therefore, this is not limited to Time Difference of Arrival (TDOA), DOA, energy, power, etc. The estimation algorithm for observations is not limited to the MUSIC algorithm; it can also be the GCC-PHAT algorithm for TDOA estimation or the MVDR algorithm for DOA estimation, both of which improve performance when using calibrated signals. Furthermore, the positioning algorithm is not limited to solving a system of simultaneous DOA equations.

[0048] Another embodiment provides an acoustic sensor frequency response calibration system for implementing the aforementioned acoustic sensor frequency response calibration method, comprising: The signal processing module processes the signals collected by each sensor in the acoustic sensor array to obtain frequency domain signals; The filter application and calibration network construction module applies a filter bank to the frequency domain signal obtained from each sensor, constructs a complex domain deep residual calibration network, takes the frequency domain signals of all sensors as input, and outputs the frequency domain filter coefficients corresponding to each sensor; the complex domain deep residual calibration network maps the input features to a high-dimensional space through an initial complex convolutional layer to obtain an initial complex feature map, and after passing through multiple cascaded complex residual blocks, performs global average pooling on the feature map to aggregate global frequency domain information, and generates the frequency domain filter coefficients of each sensor through a complex fully connected layer; The network training and calibration module constructs a loss function and trains the complex domain deep residual calibration network based on the constructed loss function to obtain the trained complex domain deep residual calibration network. The frequency domain filter coefficients are obtained through the trained complex domain deep residual calibration network. The signal reconstruction module reconstructs the sensor output in the time domain based on the obtained frequency domain filter coefficients to obtain the calibrated signal.

[0049] The following specific example illustrates the implementation of this invention, which includes the following steps: First, during the substation maintenance window, the distributed acoustic sensors are grouped together to ensure that the signals received by all sensors are as similar as possible. A broadband white noise signal is then played through a loudspeaker. s ( t Its power spectral density satisfies:

[0050] in, N 0 is a constant. f For frequency, f L The lower limit frequency, f H The upper limit frequency covers the operating frequency band of the sensor to be calibrated.

[0051] Signals collected by each sensor y i ( n The signal is then framed and windowed, and a Short-Time Fourier Transform (STFT) is performed on the windowed signal to obtain the frequency domain signal.

[0052] in, m= 0, 1, ..., M -1 is the frame index ( M (Total number of frames) k = 0, 1, …, K -1 is the frequency index. w ( n ) is a rectangular window function. N Δ is the number of points in the Discrete Fourier Transform. N For frame shift, Δ is usually taken. N = N / 2.

[0053] The complex domain depth residual calibration network designed in this invention uses all I The network takes the frequency domain signal from each sensor as input and outputs the complex frequency domain filter coefficients corresponding to each sensor. The network structure is as follows: Will M The frequency domain signal of each training frame is organized into a complex tensor form. X ∈ M×I×K ,in Represents the complex field. Any element in the tensor X X ( m , i , k ) is represented as:

[0054] First, an initial complex convolutional layer is used to map the input features to a high-dimensional space, obtaining the initial complex feature map:

[0055] in, This represents the complex number convolution operation. For complex convolution kernels, C 0 represents the number of output channels. L 0 represents the kernel length. For complex bias, BN C (·) represents the normalization of complex numbers. s C (·) is the complex activation function.

[0056] The detailed definition of complex convolution operation in the initial complex feature map is as follows: Let complex convolution kernel W = W r + j· W i Complex input X = X r + j·X i The output of complex convolution is:

[0057] in, This represents the real-valued convolution operation.

[0058] Complex batch normalization in initial complex feature maps BN C (·) performs joint normalization on the real and imaginary parts of a complex number. Let the complex number characteristic graph... H = H r + j· H i Its covariance matrix is:

[0059] Here, Cov represents the covariance operation. Complex batch normalization is achieved by normalizing this covariance matrix to an identity matrix:

[0060] in, E {·} represents the expected operation.

[0061] Initial complex feature map s C (·) The complex activation function uses the complex ReLU activation function. The magnitude and phase of the complex signal are processed separately. Let an element in the complex feature map be a complex number. z ,but:

[0062] in, t For a learnable threshold parameter, | z | represents the modulus of the complex number, arg( z () represents the complex phase. When the input modulus is less than... t The output is zero when it is greater than zero. t The phase information is retained and the threshold is subtracted.

[0063] The core network consists of P It consists of a series of complex residual blocks. p The structure of a complex residual block is as follows:

[0064] in, For complex ReLU activation functions, , and The first p The complex convolution kernels of the first, second, and adjusted convolutional layers in each residual block. , and The first p The complex bias vectors of the first and second convolutional layers and the adjusted convolutional layer in each residual block. , and The first p Input, output, and intermediate layer of each residual block H (p-0.5) The number of output channels, L is the kernel length.

[0065] AdjustChannel(·) is a channel number adjustment operation. When the number of input and output channels are the same, this operation is an identity mapping; when the number of channels is different, the channel number is matched by a 1×1 convolution.

[0066] No. p The dimension of the output of each complex residual block is , K ' is the frequency domain feature dimension after pooling.

[0067] go through P After each residual block, the feature map is subjected to global average pooling to aggregate global information in the frequency domain:

[0068] in, [:, :, k ] indicates taking medium depth k Elements in all rows and columns; The output layer is a complex fully connected layer that generates complex frequency domain filter coefficients for each sensor.

[0069] in, Let be the output frequency domain filter coefficient matrix, the th i The row corresponds to the first i Frequency domain filter coefficients of each sensor , . and These represent the weight matrix and bias vector of the output fully connected layer, respectively. For the output layer activation function, a variant of the complex sigmoid function is chosen:

[0070] in, .

[0071] For the i Each sensor outputs frequency domain filter coefficients via a network. The calibrated frequency domain output is:

[0072] in, These are intermediate parameters used during network training. They are continuously updated as the network trains, and the final result is obtained after the training is completed. As .

[0073] The core loss function is designed as the sum of squares of the differences in the frequency domain outputs of all paired sensor combinations after calibration:

[0074] in, C ( I , 2) = I ( I -1) / 2 is the number of sensor pair combinations, normalization factor. This is used to eliminate the influence of batch size, number of sensor logs, and number of frequency points on the magnitude of the loss. By minimizing this loss function, the network autonomously learns frequency domain filter coefficients that make the frequency domain outputs of all sensors tend to be consistent. To prevent the network from learning trivial solutions (such as pushing all frequency domain filter coefficients to zero), a regularization constraint is introduced:

[0075] in, l This is the regularization coefficient, which is usually set to a positive real number in the range [0.01, 1] to ensure calibration accuracy while avoiding excessively strong regularization constraints from affecting the calibration effect.

[0076] The total loss function is:

[0077] The complex-domain Adam optimizer is used for training, and its update rule is as follows:

[0078] in, t The number of iterations. i The network parameters (including all complex weights and biases to be optimized) m and v These are the exponential moving averages of the gradient and the squared magnitude of the gradient, respectively. β 1 and β 2 are respectively m and v The attenuation rate, and These are the deviation correction results. m and v , or For learning rate, This is a smoothing term. The total loss function is applied to the network parameters. The complex gradient of a number is defined as the combination of the gradients of its real and imaginary parts, i.e.

[0079] in, i r and i i Representing network parameters i The real and imaginary parts, Representation function f right a Find the partial derivative. f For about a and b Functions of variables This indicates taking the partial derivative with respect to θ.

[0080] Training employs an early stopping strategy: when the validation set loss is continuous... If the weights do not decrease, stop training and restore the optimal model weights.

[0081] The complex frequency domain filter coefficients obtained after training the complex domain deep residual calibration network are used as the optimized frequency domain filter coefficients. .

[0082] For frequency domain filter coefficients Perform inverse discrete Fourier transform to obtain the time-domain FIR filter coefficients. ,Right now:

[0083] in, l This is the index of the filter coefficients.

[0084] Will Once programmed to the sensor node, the node... i The output real-time signal is:

[0085] in, n To provide the sampling point index for the calibrated time-domain signal, Q The number of taps for the filter coefficients is typically set to... Q < N .

[0086] Based on the calibrated signal, beamforming algorithms such as Multiple Signal Classification (MUSIC) are used to obtain the Direction of Arrival (DOA). The DOA of the substation fault sound source arriving at the sound sensor array is then obtained. The equation for the substation fault location is constructed as follows:

[0087] in,( x s , y s ) represents the coordinates of the fault point, ( x i , y i )and α i The first i The coordinates of the center point of each sound sensor and their relationship to the map coordinate system. x The angle of clockwise rotation of the axis. The measured DOA (the angle between the line connecting the center point of the acoustic sensor and the fault point and the array detection direction; clockwise angles are positive, counterclockwise angles are negative) is used. Combining the equations of all acoustic sensors, a system of equations is established as follows: Axe = b in,

[0088] The solution for the coordinates of the substation fault point is: x = ( A T A ) -1 A T b .

[0089] The performance of this invention in calibrating the amplitude of acoustic sensors is shown in Figure 2. Comparing Figure 2(a) and Figure 2(b), it can be seen that the amplitude difference caused by manufacturing tolerances is very significant, and the proposed method can effectively calibrate the amplitude differences of each acoustic sensor.

[0090] This invention uses the output signal of a calibrated acoustic sensor to estimate the direction of arrival (DOA) of a substation fault sound source, such as... Figure 3 The figure shows the power spectrum of DOA estimated using the MUSIC algorithm before and after calibration. The actual angle of the fault sound relative to the acoustic sensor array is 40 degrees. The horizontal axis value corresponding to the peak value in the figure is the estimated value of DOA.

[0091] contrast Figure 3The MUSIC power spectra before and after calibration, without mismatch, show that the MUSIC algorithm can accurately estimate the DOA value. The amplitude and phase deviations between acoustic sensors have a significant impact on the DOA estimation results. Before calibration (i.e., ... Figure 3 The estimated DOA (Difference of Appearance) for the "mismatch" was 43.3 degrees, which differed from the true value by 3.3 degrees. After calibration using this method, the estimated DOA was 39.6 degrees, which differed from the true value by only 0.4 degrees, resulting in a significant improvement in accuracy.

[0092] This invention uses calibrated acoustic sensors to acquire DOA (Depth of Area) and then uses it to estimate the location of the substation fault sound source. In the map coordinate system, the substation fault point is located at (13.24m, 9.82m). Three acoustic sensor arrays are placed at (5m, 0m), (0m, 5m), and (14m, 7m), respectively, with rotation angles of 0 degrees, 90 degrees, and -45 degrees relative to the horizontal coordinate of the map coordinate system. Figure 4 The image shows the fault location results before and after calibration.

[0093] Depend on Figure 4 It can be seen that without calibration, directly using an acoustic sensor with amplitude and phase deviations to locate the fault point results in poor accuracy, with a distance error of 0.7384m. After calibration, the distance error for fault point location is reduced to 0.0484m, significantly improving accuracy.

[0094] In another embodiment, a computer storage medium stores a computer program thereon, wherein when a computer executes the computer program, it implements the acoustic sensor frequency response calibration method described in any of the preceding embodiments.

[0095] The specific implementation method is the one described above, and will not be repeated here.

[0096] In another embodiment, an electronic device includes a memory and a processor, wherein the memory stores a computer program, and the processor runs the computer program stored in the memory, wherein the computer program, when executed, implements the acoustic sensor frequency response calibration method described in any of the preceding embodiments.

[0097] The specific implementation method is the one described above, and will not be repeated here.

[0098] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A method for calibrating the frequency response of an acoustic sensor, characterized in that, Includes the following steps: The signals collected by each sensor in the acoustic sensor array are processed to obtain frequency domain signals; A filter bank is applied to the frequency domain signal obtained from each sensor to construct a complex domain deep residual calibration network. The network takes the frequency domain signals of all sensors as input and outputs the frequency domain filter coefficients corresponding to each sensor. The complex domain deep residual calibration network maps the input features to a high-dimensional space through an initial complex convolutional layer to obtain an initial complex feature map. After passing through multiple cascaded complex residual blocks, the feature map is subjected to global average pooling to aggregate global frequency domain information. The frequency domain filter coefficients of each sensor are generated through a complex fully connected layer. A loss function is constructed, and the complex domain deep residual calibration network is trained based on the constructed loss function to obtain the trained complex domain deep residual calibration network. The frequency domain filter coefficients are obtained through the trained complex domain deep residual calibration network. The construction of the loss function includes: The core loss function is designed as the sum of squares of the differences in the frequency domain outputs of all paired sensor combinations after calibration: in, C ( I , 2) = I ( I -1) / 2 is the number of combinations of sensor pairs. I Number of sensors As the normalization factor, M Total number of frames K Number of effective frequency points; Introducing regularization constraints: in, These are the frequency domain filter coefficients. λ The regularization coefficient is used. The total loss function is: ； The sensor output is reconstructed in the time domain based on the obtained frequency domain filter coefficients to obtain the calibrated signal.

2. The acoustic sensor frequency response calibration method according to claim 1, characterized in that, Before processing the signals acquired by each sensor in the acoustic sensor array, the following steps are also included: When distributed acoustic sensors are clustered together and a broadband white noise signal is played through a loudspeaker, its power spectral density satisfies: in, N 0 is a constant. f For frequency, f L The lower limit frequency, f H This is the upper limit frequency.

3. The acoustic sensor frequency response calibration method according to claim 1, characterized in that, The frequency domain signal obtained by each sensor is: in, m = 0, 1, ..., M -1 is the frame index. M Total number of frames k = 0, 1, …, K -1 is the frequency index. K Number of effective frequency points w ( n ) is a rectangular window function. N Δ is the number of points in the Discrete Fourier Transform. N For frame shift.

4. The acoustic sensor frequency response calibration method according to claim 3, characterized in that, Obtaining the initial complex feature map includes: Will M The frequency domain signal of each training frame is organized into a complex tensor form. X ∈ M×I×K ,in Represents any element in the tensor X, which is the field of complex numbers. X ( m , i , k ) is represented as: The input features are mapped to a high-dimensional space through an initial complex convolutional layer to obtain the initial complex feature map: in, This represents the complex number convolution operation. For complex convolution kernels, C 0 represents the number of output channels. L 0 represents the kernel length. For complex bias, BN C (·) represents the normalization of complex numbers. σ C (·) is the complex activation function.

5. The acoustic sensor frequency response calibration method according to claim 4, characterized in that, The frequency domain filter coefficients for each sensor include: No. p The structure of a complex residual block is as follows: in, For complex ReLU activation functions, , and The first p The complex convolution kernels of the first, second, and adjusted convolutional layers in each residual block. , and The first p The complex bias vectors of the first and second convolutional layers and the adjustment convolutional layer in each residual block; AdjustChannel(·) is the channel number adjustment operation. When the input and output channel numbers are the same, this operation is an identity mapping; when the channel numbers are different, channel number matching is performed through 1×1 convolution. H (p-0.5) , As an intermediate layer, For the first p The output layer of complex residual blocks K ' is the frequency domain feature dimension after pooling; go through P After a number of complex residual blocks, the feature map is subjected to global average pooling to aggregate global information in the frequency domain: in, [:, :, k ] indicates taking medium depth k Elements in all rows and columns; The frequency domain filter coefficients for each sensor are generated using a complex fully connected layer: in, Let be the output frequency domain filter coefficient matrix, the th i The row corresponds to the first i Frequency domain filter coefficients of each sensor , For the output layer activation function, and These represent the weight matrix and bias vector of the output fully connected layer, respectively.

6. A method for fault location of a substation acoustic sensor, characterized in that, include: The calibrated signal is obtained using the acoustic sensor frequency response calibration method according to any one of claims 1-5; Obtain the observations from all acoustic sensors, construct an equation for the location of the substation fault, and solve the equation to obtain the coordinates of the substation fault point.

7. A frequency response calibration system for an acoustic sensor, characterized in that, A method for implementing the acoustic sensor frequency response calibration method as described in any one of claims 1-5, comprising: The signal processing module processes the signals collected by each sensor in the acoustic sensor array to obtain frequency domain signals; The filter application and calibration network construction module applies a filter bank to the frequency domain signal obtained from each sensor, constructs a complex domain deep residual calibration network, takes the frequency domain signals of all sensors as input, and outputs the frequency domain filter coefficients corresponding to each sensor; the complex domain deep residual calibration network maps the input features to a high-dimensional space through an initial complex convolutional layer to obtain an initial complex feature map, and after passing through multiple cascaded complex residual blocks, performs global average pooling on the feature map to aggregate global frequency domain information, and generates the frequency domain filter coefficients of each sensor through a complex fully connected layer; The network training and calibration module constructs a loss function and trains the complex domain deep residual calibration network based on the constructed loss function to obtain the trained complex domain deep residual calibration network. The frequency domain filter coefficients are obtained through the trained complex domain deep residual calibration network. The signal reconstruction module reconstructs the sensor output in the time domain based on the obtained frequency domain filter coefficients to obtain the calibrated signal.

8. A computer storage medium having a computer program stored thereon, characterized in that, When the computer executes the computer program, it implements the acoustic sensor frequency response calibration method according to any one of claims 1-5.

9. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor runs the computer program stored in the memory. When the computer program is executed, it implements the acoustic sensor frequency response calibration method according to any one of claims 1-5.

Images