Tunnel power distribution cabinet acoustic fingerprint monitoring and early fault warning system

Abstract

The present application relates to the technical field of smart grid, disclose a tunnel power distribution cabinet acoustic fingerprint monitoring and early fault warning system, including: first, acoustic signal is collected and power supply fundamental frequency is estimated, the target sound source attribution is positioned and confirmed by using spatial coherence. Subsequently, the logarithmic coupling propagation kernel representing the combined propagation mode of the tunnel and the cabinet is established and stripped for the target sound source, and the in-phase comb residual is constructed in combination with the power supply phase. Next, the acoustic signal is input into a double-flow network to extract multi-dimensional parameters and fuse the residual, and a generalized cosine algorithm introducing a domain variance ratio is used for model optimization to eliminate cross-domain errors. Finally, a Gaussian probability baseline is established, the specific fault tendency alarm instruction is output by comparing the current likelihood risk score with the dynamic alarm threshold. The present application effectively improves the early weak abnormal sound discrimination ability of the equipment under complex working conditions.

Description

Technical Field

[0001] This invention relates to the field of smart grid technology, and more specifically, to an acoustic fingerprint monitoring and early fault warning system for tunnel distribution cabinets. Background Technology

[0002] In power system operation and maintenance, monitoring the working status of equipment such as distribution cabinets and providing early warnings of potential faults through audio signal analysis is an important non-contact detection method. Industry-standard engineering practices show that using microphone arrays combined with spatial positioning algorithms can, to some extent, capture and locate partial discharge or mechanical vibration sounds inside electrical equipment. However, when such monitoring methods are deployed on a large scale in underground tunnel distribution rooms or utility tunnel distribution hubs, the engineering site faces extremely severe acoustic environment challenges. Tunnels are essentially long-distance, enclosed tubular spaces with highly reflective surfaces. In this environment, sound wave energy is difficult to dissipate rapidly, and the direct sound emitted by equipment and the reflected sound from the pipe walls and ground will undergo strong coherent superposition, forming severe multipath reflections and comb-like spectral modulation. Simultaneously, the metal compartments, openings, and door gaps inside the distribution cabinet will further exert a local cavity filtering effect on the sound waves. During actual commissioning, weak abnormal sounds such as early partial discharge, small arcs, or mechanical loosening will be highly superimposed with the normal electromagnetic hum of the distribution cabinet, the sound of the fan, and the broadband background noise from vehicles shuttling through the tunnel. Because the complex acoustic propagation environment alters the original spectral characteristics of fault sounds, traditional detection methods relying on simple extraction of high-frequency energy thresholds or single Mel-frequency patterns often fail. Existing monitoring systems frequently misinterpret tunnel reflection fluctuations, sudden changes in ambient noise, or the sound of workers opening and closing cabinet doors as equipment malfunctions. In particular, due to the significant differences in reflection conditions and cabinet models across different tunnel sections, acoustic baselines trained using normal recordings in a single power distribution room are difficult to reliably and stably migrate to other tunnel areas. Summary of the Invention

[0003] Purpose of the invention: In view of the shortcomings of the prior art mentioned in the background, the present invention aims to provide an acoustic fingerprint monitoring and early fault warning system for tunnel power distribution cabinets.

[0004] Technical Solution: An acoustic fingerprint monitoring and early fault warning system for tunnel power distribution cabinets, applied in power distribution cabinet environments containing microphone arrays, configured to execute: The microphone array is used to collect sound and generate digital signals. The power supply fundamental frequency, which represents the power frequency phase, is obtained by self-estimating the stable spectrum line of the digital signal. The spatial coherence of the digital signal is extracted, and multipath reflection interference is suppressed by spatial geometric constraints to locate candidate compartments and determine the target sound source. A log-coupled propagation kernel characterizing the joint propagation characteristics of the tunnel and the cabinet is established for the historical normal samples of the target sound source, and propagation stability weights characterizing the spectral reliability are extracted. The instantaneous residual is obtained by stripping the log-coupled propagation kernel from the energy spectrum of the digital signal, and phase alignment is performed based on the power supply base frequency to construct the in-phase comb residual characterizing the power frequency synchronization characteristics; The digital signal is fed into a two-stream network to extract multidimensional features and fused with the in-phase comb residual to generate an acoustic fingerprint characterizing the device status. The acoustic fingerprint was trained using the generalized cosine method, and a domain variance ratio constraint was introduced to suppress cross-domain generalization error. A Gaussian line is established for the normal acoustic fingerprint. The likelihood risk score of the current acoustic fingerprint deviating from the Gaussian line is calculated. An alarm command is output when the likelihood risk score exceeds the dynamic alarm threshold generated based on the historical normal samples.

[0005] The beneficial effects of this invention include: by locating and separating the coupled propagation reference generated by the combined effect of the long enclosed space of the tunnel and the metal cabinet through spatial coherence, environmental interference caused by multipath reflection and comb reverberation is effectively eliminated. Based on this, a power frequency phase alignment mechanism is introduced to extract abnormal residuals synchronized with the power supply cycle. Combined with the multidimensional information extraction capability of the dual-stream network and the generalized cosine training model incorporating domain variance ratio, the model's adaptability across environments and cabinet types under strong noise conditions is significantly improved. This scheme enables the system to sensitively and robustly capture faint early fault sounds such as partial discharge, arcing, or mechanical loosening on its dedicated dynamic baseline, significantly reducing the false alarm and false negative rates of conventional monitoring methods in complex substation environments. Attached Figure Description

[0006] Figure 1 This is a flowchart of the acoustic fingerprint monitoring and early fault warning system for tunnel power distribution cabinets of the present invention. Detailed Implementation

[0007] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, features described in some examples may be combined in other examples.

[0008] like Figure 1 As shown, the acoustic fingerprint monitoring and early fault warning system for tunnel power distribution cabinets is applied in power distribution cabinet environments containing microphone arrays and is configured to execute: The microphone array is used to collect sound and generate digital signals. The power supply fundamental frequency, which represents the power frequency phase, is obtained by self-estimating the stable spectrum line of the digital signal. The spatial coherence of the digital signal is extracted, and multipath reflection interference is suppressed by spatial geometric constraints to locate candidate compartments and determine the target sound source. A log-coupled propagation kernel characterizing the joint propagation characteristics of the tunnel and the cabinet is established for the historical normal samples of the target sound source, and propagation stability weights characterizing the spectral reliability are extracted. The instantaneous residual is obtained by stripping the log-coupled propagation kernel from the energy spectrum of the digital signal, and phase alignment is performed based on the power supply base frequency to construct the in-phase comb residual characterizing the power frequency synchronization characteristics; The digital signal is fed into a two-stream network to extract multidimensional features and fused with the in-phase comb residual to generate an acoustic fingerprint characterizing the device status. The acoustic fingerprint was trained using the generalized cosine method, and a domain variance ratio constraint was introduced to suppress cross-domain generalization error. A Gaussian line is established for the normal acoustic fingerprint. The likelihood risk score of the current acoustic fingerprint deviating from the Gaussian line is calculated. An alarm command is output when the likelihood risk score exceeds the dynamic alarm threshold generated based on the historical normal samples.

[0009] The acoustic fingerprint monitoring and early fault warning system for tunnel power distribution cabinets provided in this embodiment is applied to power distribution cabinet environments containing microphone arrays. The microphone array typically consists of 4 to 16 omnidirectional microphones, evenly distributed on the front or top of the power distribution cabinet, with a spacing of 3cm to 5cm between adjacent microphones, used to collect sound signals generated during equipment operation. The system processes the collected sound signals in real time through edge computing nodes. These edge computing nodes employ industrial-grade processors with ARM or x86 architectures, possessing at least 4GB of memory and at least 32GB of storage capacity, enabling equipment status monitoring and early fault warning.

[0010] S201: Generates a digital signal from the discrete signal acquired by the microphone array through analog-to-digital conversion and quantization. The calculation formula is:

[0011] in, For analog-to-digital conversion and quantization operations, The target number of distribution cabinets, For the first The total number of compartments in the distribution cabinet. For the distribution cabinet number index, For compartment number index, For microphone channel indexing, For discrete time step index, For tunnel propagation impulse response, For rack-borne impulse response, For real sound sources, Background noise is considered. Analog-to-digital conversion and quantization convert analog audio signals into 16-bit or 24-bit digital signals. The sampling frequency is typically set between 16kHz and 48kHz; in this embodiment, a 48kHz sampling frequency is used. The tunnel propagation impulse response describes the reflection, diffraction, and attenuation characteristics of sound waves in the long, enclosed space of a tunnel, while the cabinet propagation impulse response describes the cavity filtering and reflection characteristics of sound waves within the metal compartment of the distribution cabinet. Background noise includes environmental noise from vehicle operation, fan operation, cable vibration, etc., within the tunnel.

[0012] The microphone array adopts a linear array geometry, evenly arranged horizontally along the front of the power distribution cabinet, with the array center aligned with the center of the cabinet. Edge computing nodes are deployed in dedicated mounting boxes on the side of the power distribution cabinet, connected to the microphone array via shielded twisted-pair cables, with a transmission distance not exceeding 10 meters. The analog-to-digital conversion uses 16-bit linear quantization, with a quantization range of -1V to +1V, corresponding to a digital signal value range of -32768 to 32767.

[0013] S202: Perform a short-time Fourier transform on the digital signal to generate a frequency domain signal. The calculation formula is:

[0014] in, For time frame indexing, For frequency micro-unit index, For window functions, The frame shift step size, The number of transform points is specified. The Short-Time Fourier Transform (SFT) divides the time-domain signal into multiple overlapping time frames, performing a Fourier Transform on each frame to obtain the time-frequency domain representation. The window function typically uses a Hanning or Hamming window, with a window length generally set to 256 or 512 points. This embodiment uses a 512-point Hanning window to reduce spectral leakage. The frame shift step size is typically half the window length, i.e., 128 or 256 points. This embodiment uses a 256-point frame shift step size to ensure temporal continuity. The number of transform points is typically a power of 2, such as 512 or 1024 points. This embodiment uses 1024 transform points to improve the computational efficiency and frequency resolution of the Fast Fourier Transform.

[0015] The overlap rate of the Short Time Fourier Transform (SFT) is 50%, meaning that half of the samples overlap between adjacent time frames. The Fast Fourier Transform (FFT) is implemented using the radix-2 FFT algorithm, with a computational efficiency of [missing information]. ,in The number of transformation points.

[0016] S203: Within the frequency domain signal's higher-than-zero frequency band, a stable spectral line is extracted. A search is conducted to find the frequency whose weighted sum of the squares of the difference between the ratio of the stable spectral line's frequency to its corresponding harmonic order and a preset candidate frequency point is minimized. This self-estimates the power supply's fundamental frequency. To eliminate frequency drift interference, the calculation formula is as follows:

[0017] in, For the set of stable spectral lines extracted, For spectral line weights, The frequency of the spectral line. The harmonic order is... It is a set of positive integers. The preset candidate frequencies are defined as follows: Stable spectral lines refer to spectral components that persist for more than 10 consecutive time frames with amplitude fluctuations of less than 3dB, typically corresponding to the power supply frequency and its harmonics. Spectral line weights are determined based on the duration and amplitude stability of the spectral line; the longer the duration and the more stable the amplitude, the greater the weight. The weight range is 0 to 1. The preset candidate frequencies are set around the standard power supply frequency. For a 50Hz power grid, the search range is 49Hz to 51Hz with a frequency step of 0.01Hz; for a 60Hz power grid, the search range is 59Hz to 61Hz with a frequency step of 0.01Hz. By minimizing the weighted sum of squares, the actual power supply fundamental frequency can be accurately estimated, eliminating interference caused by power grid frequency fluctuations.

[0018] The specific formula for calculating spectral line weights is as follows: ,in This is the ratio of the spectral line duration to the total observation time. The standard deviation of the spectral line amplitude. This represents the average amplitude of the spectral lines. Amplitude fluctuations of stable spectral lines are calculated using relative fluctuations, i.e. .

[0019] S301: Calculate the cross-power spectrum of any two channels of the frequency domain signal, divide by the sum of the cross-power spectrum amplitude and a small constant to perform phase normalization, and obtain the generalized cross-correlation. To eliminate amplitude fluctuation interference, the calculation formula is as follows:

[0020] in, and The first channel The frequency domain signal and the second channel Complex conjugate of frequency domain signal, Indexing of micro-units along frequency The inverse Fourier transform operation is performed. It is a tiny constant. This represents the total number of microphone channels. The time delay variable is represented by the cross-power spectrum, which reflects the correlation between the two channel signals at different frequencies. Phase normalization eliminates the influence of gain differences between different channels on time delay estimation, retaining only phase information. A small constant is also included. Usually taken This is used to avoid cases where the denominator is zero. The inverse Fourier transform converts the cross-power spectrum in the frequency domain into a generalized cross-correlation function in the time domain, with its peak position corresponding to the signal propagation delay between the two channels.

[0021] During cross-power spectrum calculation, a Hanning window is pre-applied to the frequency domain signal of each channel to reduce spectral leakage. The number of points in the inverse Fourier transform is the same as the number of points in the short-time Fourier transform, which is 1024 points in this embodiment.

[0022] S302: Calculate the delay time by combining the distance difference between the spatial geometric center of the candidate compartment and the two channels, as well as the on-site sound velocity. Then, sum the generalized cross-correlation projections to each candidate compartment to obtain the spatial coherence. The calculation formula is:

[0023] in, The coordinates of the spatial geometric center of the candidate compartment are: and Channels With channel The microphone's physical coordinates The on-site sound velocity. A candidate compartment refers to an independent compartment within a distribution cabinet containing electrical components. Its spatial geometric center coordinates are pre-determined during system deployment using a 3D laser scanner or total station, with a measurement accuracy of no less than 1 cm. The on-site sound velocity is calculated based on the temperature and humidity within the tunnel, using the following formula: ,in Temperature in Celsius Relative humidity, temperature, and humidity are collected in real time by temperature and humidity sensors deployed inside the tunnel, with a collection period of 1 minute. For each candidate compartment, the theoretical time delay between all microphone channel pairs is calculated, and then the values ​​of the generalized cross-correlation function at that theoretical time delay are summed to obtain the spatial coherence of the candidate compartment. The higher the spatial coherence, the greater the probability that the compartment will generate a sound source.

[0024] Temperature and humidity sensors are deployed on the tunnel wall within 1 meter of the power distribution cabinet, at a height of 1.5 to 2 meters. Three-dimensional measurement uses a right-handed Cartesian coordinate system, with the origin set at the center of the ground at the tunnel entrance. The x-axis runs along the tunnel's extension direction, the y-axis is perpendicular to the tunnel wall and points towards the power distribution cabinet, and the z-axis is vertically upward.

[0025] S303: Select the candidate compartment that maximizes the spatial coherence response as the target sound source. The calculation formula is:

[0026] By comparing the spatial coherence of all candidate compartments, the compartment with the highest coherence is selected as the target sound source for the current time frame. This method utilizes spatial geometric constraints to suppress interference from multipath reflections and other irrelevant sound sources, enabling precise localization of sound sources inside the distribution cabinet with a positioning error typically less than 10 cm.

[0027] S401: Perform beamforming on the frequency domain signal focused on the target sound source to acquire the array signal. The calculation formula is:

[0028] in, For beam weighting parameters, Index of frequency micro-units The corresponding physical frequency, To compensate for the array channel delay. Beamforming enhances the signal from the direction of the target sound source and suppresses interference from other directions by performing phase compensation and weighted summation on the signals from each microphone channel. The array channel delay compensation is calculated based on the distance difference between the target sound source and each microphone, using the following formula: The beam weighting parameters employ a delay-summation beamforming method, with all channels having a weight of 1, which is simple to implement and computationally efficient.

[0029] The beamforming frequency range is 20Hz to 20kHz, covering the main frequency range of electrical equipment fault sounds. The array channel delay compensation calculation accuracy is 1 microsecond, and subsampling precision phase compensation is achieved through linear interpolation.

[0030] S402: After adding a small constant to the energy spectrum of the array signal corresponding to the historical normal sample, take the logarithm and calculate the median to construct a log-coupled propagation kernel characterizing the static background environment. The calculation formula is:

[0031] in, This is a set of historical normal sample frames. The median operation is performed along a time frame set. Historical normal samples refer to sound samples continuously collected for more than 72 hours under fault-free operating conditions, covering 10% to 100% of different load conditions and different ambient temperature conditions from 0℃ to 40℃. The logarithmic transformation of the energy spectrum converts the multiplicative propagation effect into an additive one, facilitating subsequent stripping operations. The median operation exhibits strong robustness and can effectively suppress the impact of occasional outliers on background estimation. The logarithmically coupled propagation kernel characterizes the sound wave propagation characteristics under the combined action of the tunnel and the cabinet, including effects such as multipath reflection, cavity filtering, and frequency-selective attenuation.

[0032] The sampling interval for historical normal samples is 1 second, meaning features are extracted from one time frame per second. The median is calculated using a fast selection algorithm, with a time complexity of O(n log n). ,in This represents the number of samples.

[0033] S403: Calculate the absolute median difference between the logarithmic value of the energy spectrum of the array signal and the logarithmically coupled propagation kernel, add a small constant, and take the reciprocal to convert it into propagation stability weights. The calculation formula is:

[0034] in, This is an operation to calculate the absolute median. The absolute median is a robust measure of dispersion, and its calculation formula is: The propagation stability weight is inversely proportional to the fluctuation of the logarithmic value of the energy spectrum. The smaller the fluctuation at a frequency point, the larger the propagation stability weight, indicating that the signal at that frequency point is less affected by environmental interference and has higher reliability. In subsequent processing, the propagation stability weight is used to weight the residuals at different frequency points, improving the accuracy of anomaly detection.

[0035] After calculating the absolute median, it needs to be multiplied by a scaling factor of 1.4826 to make it a consistent and robust estimate of the standard deviation. The modified formula for calculating the propagation stability weights is as follows: .

[0036] S501: Subtract the logarithmic coupling propagation kernel from the logarithmic value of the current array signal's energy spectrum to obtain the instantaneous residual. To remove spatial reverberation interference, the calculation formula is:

[0037] Instantaneous residuals represent the difference between the current signal and the normal background environment. By subtracting the logarithmic coupling propagation kernel, the static propagation effects caused by tunnels and cabinets can be effectively eliminated, highlighting the sound changes generated by the equipment itself. For normally operating equipment, the instantaneous residual should be close to zero; when the equipment malfunctions, the fault sound will cause the instantaneous residual to show a significant positive value at the corresponding frequency point.

[0038] S502: Calculate the current power phase based on the power supply base frequency and the sampling parameters of the acquired digital signal. The calculation formula is:

[0039] in, The sampling frequency is [value]. The power supply phase represents the position of the current time frame relative to the power supply frequency cycle, and its value range is [value range]. By accurately estimating the power supply fundamental frequency and sampling parameters, the power supply phase corresponding to each time frame can be precisely calculated, providing a basis for subsequent phase alignment.

[0040] S503: Generating phase smoothing width based on the phase fluctuation variance and small constants of historical normal samples. The in-phase baseline is obtained by performing cosine weighting according to the power supply phase. The calculation formula is:

[0041] in, Historical normal sample frames Corresponding power phase, To calculate the variance, the phase fluctuation variance reflects the degree of random fluctuation in the power supply phase in historical normal samples, and the phase smoothing width is used to control the range of cosine weighting. The in-phase baseline is obtained by cosine-weighted averaging of the instantaneous residuals in historical normal samples that are close to the current phase, characterizing the sound characteristics synchronized with the power supply phase under normal conditions. Cosine weighting increases the weight of samples with closer phases, thus better capturing power frequency synchronization characteristics.

[0042] Before calculating the phase fluctuation variance of historical normal samples, the phase sequence is first detrended to eliminate the influence of slow power frequency drift. Cosine weighting uses full-range weighting, that is, it uses all historical normal samples for weighted averaging without setting a phase difference cutoff range.

[0043] S504: The square of the positive instantaneous residuals with amplitudes higher than the in-phase baseline is multiplied by the propagation stability weight, summed, and divided by the total weight to generate the in-phase comb residuals. The calculation formula is:

[0044] in, The positive truncation operation retains only the portion of the instantaneous residual greater than the in-phase baseline, as equipment faults typically manifest as an increase in sound energy. The squaring operation amplifies larger residuals, improving sensitivity to subtle anomalies. Propagation stability weights are used to weight the residuals at different frequency points, with higher-reliability frequencies receiving greater weight. The in-phase comb residual characterizes anomalous components in the current signal that are in phase with the power supply, exhibiting high sensitivity for detecting power frequency-related faults such as partial discharge and electric arcs.

[0045] S601: Converts the energy spectrum of the array signal into a log-Mel plot using a Mel filter bank. The coordinate attention mechanism is used to aggregate along the time and frequency dimensions to generate a coordinate attention map that retains time and frequency location information. The calculation formula is:

[0046] in, For Mel filter banks, For Mel band indexing, For the total number of Mel bands, For time step index, This represents the total number of time steps. Indexed by channel dimension, The intermediate feature tensor extracted by the network. and For one-dimensional aggregation parameters, and The time-frequency direction attention weights are used. The Mel filter bank converts the linear frequency scale into a Mel frequency scale that simulates the auditory characteristics of the human ear. Typically, 40 or 64 Mel bands are set; this embodiment uses 64 Mel bands, covering a frequency range of 0Hz to 24kHz. The logarithmic Mel diagram is a commonly used time-frequency representation method for sound signals, effectively capturing the perceptual features of sound. The coordinate attention mechanism first inputs the logarithmic Mel diagram into a feature extraction network consisting of two convolutional layers, obtaining a dimension of... intermediate feature tensor ,in The number of channels is set to 64 in this embodiment. Then, global average pooling is performed on the features along both the time and frequency dimensions, resulting in a dimension of... Time dimension aggregation parameters and dimension are Frequency dimension aggregation parameter Next, the time dimension aggregation parameter is input into the first fully connected layer, whose input dimension is... The output dimension is The activation function is ReLU, with a bias term; the output is then fed into a second fully connected layer, whose input dimension is... The output dimension is The activation function is Sigmoid, and a bias term is used to obtain the attention weights in the temporal direction. The same structure is used to process the frequency dimension aggregation parameters to obtain the frequency direction attention weights. Finally, the attention weights are multiplied by the original features to obtain a coordinate attention map that preserves time-frequency location information. This mechanism can effectively capture local time-frequency features in sound signals and improve the accuracy of feature extraction.

[0047] The first convolutional layer of the feature extraction network has a 3×3 kernel, a stride of 1, padding of 1, 32 output channels, and a ReLU activation function. The second convolutional layer has a 3×3 kernel, a stride of 1, padding of 1, 64 output channels, and a ReLU activation function. The input log-Melplot of the coordinate attention mechanism is pre-normalized to zero mean and unit variance.

[0048] S602: Simultaneously take the logarithm of the sum of squares of the amplitudes of the frequency domain signal to obtain a linear spectrum characterizing the fine-grained harmonic structure. The calculation formula is:

[0049] in, For the corresponding time frame The time-domain truncated segment is a segment of the signal. A linear spectrogram preserves the original frequency resolution and clearly displays the harmonic structure of the signal. For electrical equipment, mechanical faults typically exhibit energy spikes at integer multiples of the power frequency; therefore, linear spectrograms are crucial for detecting mechanical vibration-related faults. The time-domain truncated segment is the corresponding time frame extracted from the time-domain signal after beamforming. The signal segment has the same length as the window length of the short-time Fourier transform, which is 512 points in this embodiment.

[0050] The linear spectrogram has a frequency range of 0 Hz to 24 kHz, which is half the sampling frequency. The time-domain truncated segment is pre-windowed with a Hanning window before the Fourier transform to reduce spectral leakage.

[0051] S603: Perform average pooling dimensionality reduction on the coordinate attention map and the linear spectrogram, concatenate the in-phase comb residuals, and output a fixed-dimensional acoustic fingerprint after matrix fusion and scale normalization. The calculation formula is:

[0052] in, For coordinate attention diagram The characteristic tensor constituted, For the linear spectrum diagram The characteristic tensor constituted, This is a global average pooling operation. For standardized operation, To fuse the projection matrices, global average pooling transforms the two-dimensional time-frequency feature map into a one-dimensional feature vector, achieving dimensionality reduction. Note that the coordinate attention graph has a dimension of... After global average pooling, a feature vector of dimension 64 is obtained. The dimension of the linear spectrogram is... After global average pooling, the resulting dimension is The feature vector, in this embodiment The value is 1024. The in-phase comb residual is a scalar with a dimension of 1. The concatenation operation combines features from three different branches to form a single feature with a dimension of 1. The comprehensive feature vector. The fused projection matrix is ​​a learnable parameter matrix with dimensions of . The concatenated high-dimensional feature vectors are mapped to a 128-dimensional space. The scaling normalization operation uses L2 normalization to normalize the norm of the feature vectors to 1, making the features between different samples comparable. Acoustic fingerprints are comprehensive feature vectors characterizing the operating status of equipment, including time-frequency sensing features, fine-grained harmonic structure features, and power frequency synchronization anomaly features.

[0053] Global average pooling is performed separately for each channel, meaning the average of the feature values ​​at all time and frequency positions for each channel is taken. The specific formula for L2 normalization is as follows: ,in is the dimension of the feature vector.

[0054] S701: Calculate the spatial cosine distance from the acoustic fingerprint to the center of the preset normal distribution sub-center, and truncate the maximum value as the class similarity characterizing the intra-class compactness. The calculation formula is:

[0055] in, Acoustic fingerprint , For the first The first category The pre-defined normal distribution subcenters were obtained by performing K-means clustering on the acoustic fingerprints of a large number of normal samples. K-means clustering used Euclidean distance as the distance metric, with a maximum number of iterations set to 300 and a convergence threshold set to... The initialization method uses k-means++. Five to ten sub-centers are set for each category; in this embodiment, eight sub-centers are set for each category to cover different variations of the normal state. Cosine distance is used to measure the similarity between two vectors, with a value ranging from [-1, 1]. A larger value indicates greater similarity between the two vectors. Class similarity is the maximum cosine distance from the acoustic fingerprint to all sub-centers of that category, representing the degree of similarity between the acoustic fingerprint and the normal state of that category.

[0056] The number of clusters in K-means clustering is determined by the elbow rule, which selects the number of clusters at which the rate of decrease in clustering error suddenly slows down. Normal distribution subcenters are updated every 3 months, and clustering is re-performed using the latest normal samples.

[0057] S702: Constructing an adaptive probability with multi-class normalization logic from adaptive scaling and class similarity. The calculation formula is:

[0058] in, Number of training iterations for the model Adaptive scaling at time The total number of categories, This represents the median angular distance of the training batch samples. The median angular distance of the training batch samples is the median of the angular distances between all samples in the training batch and their respective nearest centers of the normal distribution. The angular distance is calculated using the cosine distance formula: Adaptive scaling is used to adjust the temperature of the softmax function, controlling the sharpness of the probability distribution. In the early stages of training, a larger scaling scale is used to make the probability distribution more uniform, promoting the model's learning of differences between categories. In the later stages of training, the scaling scale is dynamically adjusted based on the median distance of the training batch samples, making the probability distribution sharper and improving classification accuracy. The adaptive probability is obtained by scaling and softmax normalizing the class similarity, representing the probability that the acoustic fingerprint belongs to each category.

[0059] The training and validation sets were split in an 8:2 ratio, meaning 80% of the samples were used for training and 20% for validation. No data augmentation methods were used during training to preserve the features of the original audio signals.

[0060] S703: The ratio of the conditional expected variance to the overall expected variance of class similarity under different monitoring domain labels is superimposed on the generalization loss function as the domain variance ratio. Decoupled optimization training is then performed to generate the generalization loss function. The calculation formula is:

[0061] in, For adaptive probability In actual category labels The mapping value at that location, To monitor domain labels, This is a mathematical expectation operation. Monitoring domain labels are used to identify different tunnel sections or cabinet models from which samples originate. Domain variance ratio measures the degree of difference in class similarity between different monitoring domains; a larger domain variance ratio indicates a greater difference between different monitoring domains, and a poorer generalization ability of the model. By adding a domain variance ratio constraint to the loss function, the model can learn cross-domain invariant features, suppress cross-domain generalization errors, and improve the model's transferability between different monitoring domains. Decoupled optimization training uses the Adam optimizer, with an initial learning rate set to... The learning rate decay strategy uses cosine annealing, with an annealing period of 100, a batch size of 32, and a maximum number of iterations of 100. During training, the classification loss and domain variance ratio constraint are optimized simultaneously, with both loss terms having a weight of 1, eliminating the need for alternating optimization. Model parameters are saved every 10 iterations, and the model with the minimum loss on the validation set is selected as the final model after training.

[0062] The evaluation metrics for the validation set are classification accuracy and generalization loss function value. An early stopping strategy is adopted: training is terminated early if the validation set loss does not decrease for 20 consecutive iterations.

[0063] S801: Establish a unique Gaussian line for the normal acoustic fingerprint corresponding to the target sound source according to the Bayesian information criterion, and determine the number of clusters. With probability distribution model The calculation formula is:

[0064] in, Set the number of clusters to Gaussian density function at time, For historical normal sample set The sample fingerprints, This represents the total number of normal samples in history. , , These are the Gaussian component weights, the mean vector, and the covariance matrix, respectively. The penalty coefficient is used to select the optimal number of clusters by adding a penalty to the likelihood function for model complexity. This is equal to the number of parameters in the Gaussian mixture model. The Gaussian mixture model is trained using the EM algorithm, with a maximum number of iterations set to 100 and a convergence threshold set to... The initialization method uses k-means, employing a diagonal covariance matrix to reduce computational cost. The Gaussian line is a mixture of Gaussian models composed of multiple Gaussian components, used to characterize the probability distribution of a normal acoustic fingerprint. Each Gaussian component corresponds to a sub-mode in the normal state; the Gaussian component weights represent the probability of that sub-mode occurring, and the mean vector and covariance matrix represent the center and dispersion of that sub-mode, respectively. Dedicated Gaussian lines are individually established for each target sound source, better adapting to individual differences between different devices.

[0065] The Bayesian information criterion searches for the number of clusters from 1 to 10, meaning it attempts mixture models with 1 to 10 Gaussian components. The parameters of the Gaussian mixture model are initialized using the results of k-means clustering, with the initial mean of each Gaussian component being the cluster center and the initial covariance being the covariance of the samples within the cluster.

[0066] S802: Calculate the probability density generation value of the current acoustic fingerprint under the Gaussian line, and take the negative logarithm as the likelihood risk score representing the degree of anomaly significance. The calculation formula is:

[0067] The likelihood risk score indicates the degree to which the current acoustic fingerprint deviates from the normal state. The smaller the probability density value, the more abnormal the current state, and the higher the likelihood risk score. The negative logarithmic transformation converts the probability value into a positive risk score, which facilitates subsequent threshold comparison.

[0068] S803: Establish a dynamic alarm threshold by extracting the maximum probability deviation of negative logarithm from historical normal samples. The calculation formula is:

[0069] in, up to the current time frame An accumulated set of historical normal samples. The dynamic alarm threshold is determined based on the maximum likelihood risk score of these historical normal samples. As these historical normal samples accumulate, the dynamic alarm threshold is automatically updated. Using a dynamic alarm threshold can adapt to the slow changes in equipment operating status and reduce the false alarm rate. When the likelihood risk score is greater than the determined dynamic alarm threshold, it indicates that the current operating status of the equipment is abnormal.

[0070] The dynamic alarm threshold is updated every 24 hours, recalculated using all historical normal samples up to the current time. The historical normal samples employ a sliding window mechanism with a window size of 30 days, meaning only the most recent 30 days of normal samples are retained for calculating the dynamic alarm threshold.

[0071] S804: An alarm command is triggered when the likelihood risk score exceeds the dynamic alarm threshold. Using the residuals of the in-phase comb respectively Mechanical frequency doubling abrupt change characteristics measured in combination with frequency domain signals and characteristics of electric arc mutation Extract the attribute with the largest tendency value. The alarm command is output together, and the calculation formula is:

[0072] in, It is a truth indicator function. For integer harmonic orders, Physical frequency The corresponding discrete frequency micro-unit index. The truth indicator function outputs 1 when the condition is met, and 0 otherwise, used to trigger alarm commands. Partial discharge tendency is characterized by the in-phase comb residual, because partial discharge usually occurs near the peak of the power supply voltage and is synchronized with the power supply phase. Mechanical vibration tendency is characterized by mechanical harmonic abrupt change characteristics. Mechanical faults such as bearing wear and loose screws usually show energy abrupt changes at integer multiples of the power frequency, integer harmonic orders. The value range is from 1 to 20. Arc tendency is characterized by arc abrupt changes; arc faults typically manifest as broadband abrupt changes in the spectrum, with significant spectral differences between adjacent time frames. By comparing the values ​​of the three fault tendencies, the attribute with the largest value is extracted as the fault type, providing accurate fault diagnosis information for maintenance personnel.

[0073] The alarm command output format is JSON, which includes alarm time, target sound source location, likelihood risk score, fault type, and confidence level. Fault tendency values ​​are normalized before comparison by dividing each tendency value by the historical maximum tendency value for that type, so that the value ranges from 0 to 1.

[0074] The embodiments of this example have been described above. However, this example is not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms based on the guidance of this example, and all of them are within the protection scope of this example.

Claims

1. An acoustic fingerprint monitoring and early fault warning system for tunnel power distribution cabinets, applied in power distribution cabinet environments containing microphone arrays, characterized in that, Configured for execution: The microphone array is used to collect sound and generate digital signals. The power supply fundamental frequency, which represents the power frequency phase, is obtained by self-estimating the stable spectrum line of the digital signal. The spatial coherence of the digital signal is extracted, and multipath reflection interference is suppressed by spatial geometric constraints to locate candidate compartments and determine the target sound source. A log-coupled propagation kernel characterizing the joint propagation characteristics of the tunnel and the cabinet is established for the historical normal samples of the target sound source, and propagation stability weights characterizing the spectral reliability are extracted. The instantaneous residual is obtained by stripping the log-coupled propagation kernel from the energy spectrum of the digital signal, and phase alignment is performed based on the power supply base frequency to construct the in-phase comb residual characterizing the power frequency synchronization characteristics; The digital signal is fed into a two-stream network to extract multidimensional features and fused with the in-phase comb residual to generate an acoustic fingerprint characterizing the device status. The acoustic fingerprint was trained using the generalized cosine method, and a domain variance ratio constraint was introduced to suppress cross-domain generalization error. A Gaussian line is established for the normal acoustic fingerprint. The likelihood risk score of the current acoustic fingerprint deviating from the Gaussian line is calculated. An alarm command is output when the likelihood risk score exceeds the dynamic alarm threshold generated based on the historical normal samples.

2. The acoustic fingerprint monitoring and early fault warning system for tunnel power distribution cabinets according to claim 1, characterized in that, The step of obtaining the power supply base frequency is as follows: The digital signal is subjected to a short-time Fourier transform to generate a frequency domain signal; The stable spectrum line is extracted from the frequency band greater than zero of the frequency domain signal. By searching, the weighted sum of the squares of the difference between the frequency of the stable spectrum line and its corresponding harmonic order and the preset candidate frequency point is minimized, and the fundamental frequency of the power supply is obtained by self-estimation, thereby eliminating frequency drift interference.

3. The acoustic fingerprint monitoring and early fault warning system for tunnel power distribution cabinets according to claim 2, characterized in that, The step of determining the target sound source is as follows: The cross-power spectrum of any two channels of the frequency domain signal is calculated and divided by the sum of the cross-power spectrum amplitude and a small constant to perform phase normalization processing, thereby obtaining a generalized cross-correlation to eliminate amplitude fluctuation interference. The spatial coherence is obtained by combining the distance difference between the spatial geometric center of the candidate compartment and the two channels and the on-site sound velocity to calculate the delay time, and then summing the generalized cross-correlation projection to each candidate compartment. The candidate compartment that maximizes the spatial coherence response is selected as the target sound source.

4. The acoustic fingerprint monitoring and early fault warning system for tunnel power distribution cabinets according to claim 3, characterized in that, The steps for establishing the logarithmically coupled propagation kernel and propagation stability weights are as follows: Beamforming is performed on the frequency domain signal focused on the target sound source to acquire an array signal; The logarithmic coupling propagation kernel characterizing the static background environment is constructed by adding the small constant to the energy spectrum of the array signal corresponding to the historical normal sample and taking the logarithm and calculating the median. The absolute median difference between the logarithmic value of the energy spectrum of the array signal and the logarithmic coupled propagation kernel is calculated, and the inverse of the result after adding the small constant is converted into the propagation stability weight.

5. The acoustic fingerprint monitoring and early fault warning system for tunnel power distribution cabinets according to claim 4, characterized in that, The steps for constructing the in-phase comb residual are as follows: The instantaneous residual is obtained by subtracting the logarithmic coupled propagation kernel from the current logarithmic value of the energy spectrum of the array signal to remove spatial reverberation interference; The current power phase is calculated based on the power supply base frequency and the sampling parameters of the acquired digital signal; Based on the phase fluctuation variance of the historical normal samples and the small constant, a phase smoothing width is generated, and a cosine weighting is performed according to the power supply phase to obtain the in-phase baseline; The square of the positive instantaneous residual with an amplitude higher than the in-phase baseline is multiplied by the propagation stability weight, summed, and divided by the total weight to generate the in-phase comb residual.

6. The acoustic fingerprint monitoring and early fault warning system for tunnel power distribution cabinets according to claim 5, characterized in that, The steps for generating an acoustic fingerprint are as follows: The energy spectrum of the array signal is converted into a log-Mel plot using a Mel filter bank, and then aggregated along the time and frequency dimensions using a coordinate attention mechanism to generate a coordinate attention plot that retains the time and frequency position information. Simultaneously, the sum of squares and logarithms of the amplitudes of the frequency domain signal are taken to obtain a linear spectrum characterizing the fine-grained harmonic structure; The coordinate attention map and the linear spectrogram are subjected to average pooling dimensionality reduction. After splicing the in-phase comb residuals, the acoustic fingerprint with fixed dimensions is output through matrix fusion and scale normalization.

7. The acoustic fingerprint monitoring and early fault warning system for tunnel power distribution cabinets according to claim 6, characterized in that, The steps for training the acoustic fingerprint network using the generalized cosine method are as follows: Calculate the spatial cosine distance from the acoustic fingerprint to the preset normal distribution subcenter, and take the maximum value as the class similarity characterizing the intra-class compactness; An adaptive probability containing multi-class normalization logic is constructed by using an adaptive scaling scale and the class similarity. The ratio of the conditional expected variance to the overall expected variance of the class similarity under different monitoring domain labels is superimposed on the generalization loss function as the domain variance ratio, and decoupled optimization training is performed.

8. The acoustic fingerprint monitoring and early fault warning system for tunnel power distribution cabinets according to claim 7, characterized in that, The steps for outputting the alarm command are as follows: According to the Bayesian information criterion, a unique Gaussian line is established for the normal acoustic fingerprint corresponding to the target sound source; Calculate the probability density generation value of the current acoustic fingerprint under the Gaussian line, and take the negative logarithm as the likelihood risk score representing the degree of anomaly significance; The dynamic alarm threshold is established by extracting the maximum probability deviation of the negative logarithm value from the historical normal samples. When the likelihood risk score exceeds the dynamic alarm threshold, the partial discharge, mechanical vibration and arc tendency are calculated by using the abrupt change characteristics of the in-phase comb residual and the frequency domain signal, respectively. The attribute with the largest tendency value is extracted and output together with the alarm command.

Images