Abnormal sound monitoring and state identification method for power equipment

Abstract

The invention provides an abnormal sound monitoring and state identification method for power equipment. The method comprises the following steps: acquiring a topological structure of the power equipment; initializing an audio receiving array topological parameter set based on a power equipment topological structure and constructing an initial array ant colony set; multi-index fitness calculation and conditional ant colony iteration are carried out, a target array topological parameter set is acquired, and an audio receiving array is constructed; and monitoring the power equipment based on the audio receiving array, obtaining an abnormal sound acoustic signal, performing state identification, and outputting a state identification result of the power equipment. According to the abnormal sound monitoring and state identification method for the power equipment, the audio receiving array is dynamically constructed according to the topological structure of the power equipment, abnormal sound detection is carried out on the power equipment, the integrity and quality of obtained abnormal sound acoustic signals can be improved by dynamically constructing the audio receiving array, and the reliability of the abnormal sound monitoring and state identification method is improved. And thus, accurate monitoring and state identification of abnormal sound of the power equipment are realized.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of power equipment state monitoring, and particularly relates to a power equipment abnormal sound monitoring and state identification method. BACKGROUND

[0002] In the field of power equipment state monitoring, abnormal sound generated during equipment operation is a key information carrier representing the running state of the equipment. For example, typical faults such as transformer winding loosening, circuit breaker mechanism jamming or insulator surface discharge can all cause characteristic acoustic responses. Traditional acoustic monitoring methods usually use regular array structures to collect equipment noise. However, these methods have significant limitations in complex power equipment monitoring scenarios: on the one hand, equipment is usually large in size, irregular in shape, and densely distributed in open or semi-enclosed substations, with diverse background noise and complex sound reflection paths, so fixed array structures are easily disturbed by reverberation, leading to signal distortion; on the other hand, various background noise sources in complex industrial environments can seriously pollute the target acoustic signal, thereby affecting the accurate monitoring and state identification of power equipment abnormal sound. SUMMARY

[0003] In view of the deficiencies of the prior art, the application provides a power equipment abnormal sound monitoring and state identification method to solve the above problems. The method dynamically constructs an audio receiving array according to the topology structure of the power equipment, detects the abnormal sound of the power equipment, and improves the integrity and quality of the acquired abnormal sound acoustic signals through dynamic construction of the audio receiving array, thereby realizing accurate monitoring and state identification of the abnormal sound of the power equipment.

[0004] To solve the above technical problems, the application provides a power equipment abnormal sound monitoring and state identification method, which comprises the following steps:

[0005] Initializing an audio receiving array topology parameter set based on the topology structure of the power equipment;

[0006] Constructing an initial array ant colony set based on the audio receiving array topology parameter set, and performing multi-index fitness calculation based on the initial array ant colony set to obtain an array fitness value set;

[0007] Performing conditional ant colony iteration based on the array fitness value set and the audio receiving array topology parameter set to obtain a target array topology parameter set;

[0008] Constructing an audio receiving array based on the target array topology parameter set;

[0009] Monitoring the power equipment based on the audio receiving array to obtain an abnormal sound acoustic signal;

[0010] Performing state identification based on the abnormal sound acoustic signal to output a power equipment state identification result.

[0011] In the above scheme, the power equipment topology structure in the actual application scene is acquired, and an audio receiving array is dynamically constructed based on different topology structures of the power equipment. In the process of constructing the audio array, an ant colony algorithm is adopted, conditional iteration is performed in the initialized audio receiving array topology parameter set, and finally a target array topology parameter set is obtained to match the topology structure of the corresponding power equipment, so that the audio receiving array constructed based on the target array topology parameter can effectively capture the key abnormal sound characteristics of the power equipment and suppress environmental noise, thereby improving the integrity and quality of the acquired abnormal sound acoustic signals. The high-quality abnormal sound acoustic signals obtained based on the preset array topology structure provide high-quality input for subsequent state recognition, and finally realize accurate monitoring and state recognition of the abnormal sound of the power equipment.

[0012] Further, an initial array ant colony set is constructed based on the audio receiving array topology parameter set, and multi-index fitness calculation is performed based on the initial array ant colony set to obtain an array fitness value set. The multi-index fitness calculation formula is:

[0013] Fitness=y1·MSL+y2·MBW

[0014] In the formula, Fitness is the array fitness value, y1 is the first weight coefficient, y2 is the second weight coefficient, MSL is the maximum side lobe level obtained based on the power equipment topology structure, and MBW is the main lobe width obtained based on the power equipment topology structure.

[0015] In the above scheme, the discrete audio receiving array topology parameter set is mapped to the initial array ant colony set of the ant colony algorithm, which provides an effective initial search starting point for subsequent optimization. A multi-index fitness function is constructed with the maximum side lobe level (MSL) as the main index and the main lobe width (MBW) as the auxiliary index, and the first weight coefficient and the second weight coefficient are used to realize differentiated optimization of the main and auxiliary indexes.

[0016] It should be noted that the following restriction conditions can be set in the above multi-index fitness function: y1+y2=1 and y1>y2. The above setting conditions aim to prioritize optimizing the maximum side lobe level (MSL), as it directly affects the array's anti-interference ability. Prioritizing reducing the MSL value can ensure the signal purity of the array in a complex electromagnetic environment. On the basis of optimizing MSL, the main lobe width (MBW) is also considered for optimization, as it determines the beam pointing accuracy, thereby comprehensively improving the beam forming performance. By adjusting the weight coefficients y1 and y2, different application scene requirements can be flexibly adapted: for example, in a medium-high frequency scene, MSL can be focused on reducing (increasing y1) to reduce interference, and in a low frequency scene, MBW can be focused on optimizing (relatively increasing y2) to improve target positioning accuracy.

[0017] Further, conditional ant colony iteration is performed based on the array fitness value set and the audio receiving array topology parameter set to obtain a target array topology parameter set; including:

[0018] Array pheromone updating and probability path selection are performed based on the array fitness value set and the audio receiving array topology parameter set to obtain a next generation array fitness value set and a next generation audio receiving array topology parameter set;

[0019] The above steps are repeated until when the iteration number is greater than or equal to a preset iteration number, or the fitness value set after iteration is greater than or equal to a preset fitness threshold, the target array topology parameter set is output.

[0020] In the above scheme, by setting the iteration number and the fitness threshold as double termination conditions, when the number of conditional ant colony iterations based on the array fitness value set and the audio receiving array topology parameter set reaches the preset iteration number, the search is terminated to ensure that the computing resources are controllable; if the fitness value set after iteration reaches or exceeds the preset fitness threshold, it indicates that the optimization result meeting the engineering requirements has been obtained, and the iteration is also terminated. Through the double condition judgment, it is effectively avoided to fall into invalid excessive iteration, and at the same time, it is ensured that the target array topology parameter set output meets the performance requirements, so as to balance between the calculation efficiency and the optimization effect, and improve the engineering practicability of the scheme. By using the swarm intelligence characteristics of the ant colony algorithm and under the constraint of the above double termination conditions, the above scheme realizes the automatic and intelligent search and optimization of the audio receiving array topology parameters, and finally outputs the highly optimized target array topology parameter set meeting the predetermined acoustic performance standards.

[0021] Further, state recognition is performed based on the abnormal sound acoustic signal to output a power equipment state recognition result; including:

[0022] Feature extraction is performed based on the abnormal sound acoustic signal to obtain an abnormal sound feature vector;

[0023] Transfer learning adaptation is performed based on the abnormal sound feature vector to obtain an adapted feature vector;

[0024] State recognition is performed based on the adapted feature vector to output a power equipment state recognition result.

[0025] In the above scheme, first, the abnormal sound acoustic signal is feature extracted, the original high-dimensional abnormal sound acoustic signal is converted into an abnormal sound feature vector with stronger physical representation, the key acoustic features related to the state of the power equipment are effectively screened out, and redundant information is removed. Subsequently, the abnormal sound feature vector is converted into an adaptive feature vector through transfer learning adaptation, the robustness of the abnormal sound feature vector to the field background noise, equipment individual difference and working condition change is improved, finally, the state recognition is performed based on the adaptive feature vector, the optimized abnormal sound feature vector, i.e. the adaptive feature vector, outputs the power equipment state recognition result, and provides a high-reliability acoustic diagnosis basis directly guiding the decision for equipment intelligent operation and maintenance and preventive maintenance.

[0026] It should be noted that the power equipment state recognition result can include a state classification result of the power equipment, an abnormal early warning signal and fault position information, etc.

[0027] Further, the abnormal sound feature vector is obtained based on feature extraction of the abnormal sound acoustic signal, and the feature extraction comprises:

[0028] The short-time Fourier transform is performed based on the abnormal sound acoustic signal to obtain an abnormal sound time-frequency matrix;

[0029] The abnormal sound time-frequency matrix is input into a preset residual neural network, so that the preset residual neural network performs deep semantic extraction based on the abnormal sound adaptive matrix to obtain a preliminary feature vector;

[0030] The abnormal sound feature vector is obtained based on attention weighting of the preliminary feature vector.

[0031] In the above scheme, first, the original one-dimensional abnormal sound acoustic signal is converted into a two-dimensional abnormal sound time-frequency matrix through short-time Fourier transform, and the time-varying spectral features of the non-stationary acoustic event are effectively captured. Subsequently, the abnormal sound time-frequency matrix is input into a preset residual neural network, and the deep nonlinear mapping and residual skip connection advantages of the preset residual neural network can adaptively learn and extract deep abstract features (i.e. the preliminary feature vector) containing power equipment state semantics from complex background noise, overcoming the subjectivity and limitations of artificial feature design. Finally, the preliminary feature vector is dynamically screened through the attention weighting mechanism, higher weights are given to the key time-frequency regions strongly related to the fault, and the contributions of irrelevant or interference features are suppressed, so that the abnormal sound feature vector highly focused on the essence of the abnormal sound and strongly discriminative is output, and the input quality and diagnosis reliability of the subsequent state recognition module are ensured. The deep feature extraction method based on the residual neural network and the attention mechanism can automatically learn the time-frequency dynamic features of the abnormal sound acoustic signal and the electromagnetic-mechanical composite noise mode. Compared with the traditional spectral analysis, the above scheme can improve the feature extraction accuracy of early faults such as component loosening and wear.

[0032] It should be noted that the preset residual neural network may include four residual blocks and a global average pooling layer.

[0033] Further, transfer learning adaptation is performed based on the aforementioned abnormal sound feature vector to obtain an adapted feature vector; including:

[0034] The abnormal noise feature vector is input into a preset source domain model so that the preset source domain model performs feature mapping based on the abnormal noise feature vector to obtain a high-dimensional feature vector.

[0035] The high-dimensional feature vector is input into a preset domain adaptation network, so that the preset domain adaptation network performs loss calculation based on the high-dimensional feature vector to obtain a joint optimized loss value.

[0036] Based on the joint optimization loss value and the abnormal noise feature vector, iterative optimization is performed to obtain the adaptive feature vector.

[0037] In the above scheme, the abnormal noise feature vector is first input into a pre-trained source domain model on abundant source domain data. Utilizing its learned high-order semantic mapping capabilities, the original features are transformed into a more discriminative high-dimensional feature vector, inheriting common fault knowledge from the source domain. Subsequently, the high-dimensional feature vector is input into a pre-trained domain adaptation network. By calculating a joint optimization loss value, including domain difference loss and task classification loss, the distribution difference between the target domain features and the source domain features is explicitly measured and reduced. Finally, based on this joint loss value, the feature extraction and adaptation parameters are iteratively optimized in reverse, forcing the feature space to align the data distribution of the source and target domains while retaining key fault information. The final output is an adapted feature vector that combines domain invariance and fault sensitivity, enabling the subsequent state identification module to maintain high-precision diagnostic capabilities even in real-world scenarios with scarce target device annotations and variable background noise.

[0038] Further, the high-dimensional feature vector is input into a preset domain adaptation network, so that the preset domain adaptation layer performs loss calculation based on the high-dimensional feature vector to obtain the joint loss value; the preset domain adaptation network includes an MMD loss calculation module, a center loss calculation module, a classification loss calculation module, and a joint module; wherein:

[0039] The high-dimensional feature vector is input into a preset domain adaptation network so that the MMD loss calculation module, the center loss calculation module, and the classification loss calculation module perform loss calculations based on the high-dimensional feature vector, and obtain the MMD loss value, the center loss value, and the classification loss value, respectively.

[0040] The MMD loss value, center loss value, and classification loss value are input into the joint module so that the joint module can perform joint optimization to obtain the joint optimized loss value.

[0041] In the above scheme, high-dimensional feature vectors are simultaneously input into a pre-defined domain adaptation network for feature adaptation. Through collaborative optimization of a triple loss function, cross-domain feature alignment is achieved while maintaining discriminative capability: the MMD loss calculation module calculates the maximum mean difference (MMD) statistic between the feature distributions of the source and target domains. By minimizing this loss, the global distribution alignment of features in the two domains is driven, effectively reducing the overall acoustic feature shift caused by cross-device or cross-scene conditions. The center loss module constrains feature vectors of similar faults to cluster towards their corresponding class centers in the latent space to maximize intra-class compactness and avoid a decrease in class separability caused by intra-class feature dispersion. Simultaneously, the classification loss module continuously optimizes the feature's ability to discriminate fault categories based on the source domain labeled data, ensuring that the semantic discriminativeness of the feature space is not weakened during domain alignment. The MMD loss value, center loss value, and classification loss value output by the above modules are input into the joint module to generate a unified joint optimized loss value. During backpropagation of the jointly optimized loss value, the pre-defined domain adaptation network is simultaneously constrained to meet three objectives: domain invariance, intra-class aggregation, and fault separability. Finally, based on the iterative optimization process of the jointly optimized loss value and the high-dimensional abnormal noise feature vector, an adapted feature vector with both high cross-domain adaptability and strong fault discrimination confidence is output. This scheme effectively overcomes the limitation of a single loss function in balancing domain alignment and task performance, providing core feature adaptation support for the reliable application of power equipment abnormal noise diagnosis in complex real-world scenarios. By minimizing inter-domain distribution differences through MMD and combining central loss to optimize intra-class aggregation, the system significantly reduces its dependence on target domain labeled data when applied across devices or scenarios. This alleviates the generalization challenge faced by traditional deep learning methods due to high sample dependence and improves state classification accuracy.

[0042] Further, based on the adapted feature vector, state identification is performed, and the power equipment state identification result is output; including:

[0043] The adapted feature vectors are weighted and fused using a preset multi-classifier to obtain the fused class probabilities.

[0044] Fault location is performed based on the fused category probability, and the power equipment status identification result is output.

[0045] Further, the adapted feature vectors are weighted and fused using a preset multi-classifier to obtain the fusion probability; including:

[0046] The adaptive feature vectors are weighted and fused through a preset multi-classifier to obtain an adaptive weight set and a probability output vector set.

[0047] The adaptive weight set and the probability output vector set are fused to obtain the fused probability:

[0048]

[0049] In the formula, S is the fusion probability, and w i Let C be the i-th adaptive weight in the adaptive weight set. i (x) is the i-th probability output vector in the probability output vector set.

[0050] In the above scheme, the adapted feature vector is first input into a preset multi-classifier. Complementary fault discrimination information is obtained using the classifier's decision preferences, generating an adaptive weight set that reflects the reliability assessment and probability output vector set of each classifier under specific input features. The adaptive weight set and the probability output vector set are then fused, dynamically strengthening the decision result of the classifier with the highest confidence level while suppressing interference from low-reliability models, thus outputting the fused category probability that best matches the actual fault mode. Finally, fault location is performed based on this fused probability, accurately distinguishing the health status, early faults, and severe faults of power equipment, outputting power equipment status identification results with clear fault type labeling and confidence quantification. This scheme effectively overcomes the shortcomings of insufficient generalization of single classifiers and significantly improves the robustness and engineering interpretability of diagnostic results in complex acoustic scenarios.

[0051] It should be noted that the preset multi-classifier may include the LightGBM classifier and the SoftMax classifier, etc.

[0052] Furthermore, the adapted feature vectors are weighted and fused using a preset multi-classifier to obtain an adaptive weight set. The weighted fusion calculation formula is as follows:

[0053]

[0054] In the formula, w i For the i-th adaptive weight, Accuracy i To preset the accuracy of the i-th classifier in the multi-classifier, Coverage i To preset the feature space coverage of the i-th classifier in the multi-classifier, ∑ j Accuracy j The sum of the accuracies of all classifiers, ∑ j Coverage j This is the sum of the feature space coverage of all classifiers.

[0055] In the above scheme, the weighted fusion calculation integrates classification results and introduces a two-dimensional index of accuracy and feature space coverage to solve the performance limitations of traditional single accuracy weights in scenarios with fluctuating operating conditions. iAs prior knowledge, it reflects the long-term statistical performance of the i-th classifier on the validation set, ensuring that a highly reliable classifier dominates the fusion process; Feature space coverage i The matching degree between the current adaptive feature vector and the classifier's advantageous decision region is dynamically quantified, so that the weight allocation responds to the input characteristics in real time. Through the above scheme, the advantages of different classifiers are adaptively integrated, thereby improving the accuracy and reliability of device status classification. Attached Figure Description

[0056] Figure 1 This is a schematic flowchart of a method for monitoring and identifying abnormal noises in power equipment according to an embodiment of the present invention;

[0057] Figure 2 This is a comparative schematic diagram of an audio receiving array topology optimized by the ant colony algorithm according to an embodiment of the present invention;

[0058] Figure 3 This is a schematic diagram comparing the simulation results of an audio receiving array optimized by the ant colony algorithm, provided in an embodiment of the present invention.

[0059] Figure 4 This is a schematic diagram of the iterative curve of an audio receiving array optimized by an ant colony algorithm according to an embodiment of the present invention;

[0060] Figure 5 This is a schematic diagram of an actual audio receiving array optimized by the ant colony algorithm, provided as an embodiment of the present invention. Detailed Implementation

[0061] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0062] Example 1

[0063] Please see Figure 1 This embodiment provides a method for monitoring and identifying abnormal noises in power equipment under the scenario of transformer abnormal noise monitoring, including the following steps:

[0064] Based on the power equipment topology, initialize the audio receiving array topology parameter set: set the maximum annular radius of the array to R meters, and the total number of array elements to be in the range [m]. min ,m max The range of the number of array rings is [n]. min, n max The radius of each ring is [r].min ,r max The rotation angle range of each ring [θ] min, θ max ].

[0065] An initial array ant colony set is constructed based on the audio receiver array topology parameter set. Multi-index fitness calculations are then performed on this initial array ant colony set to obtain an array fitness value set. Specifically, constructing the initial array ant colony set involves: for the total number of array elements m, within the interval [m...]. min ,m max Random sampling is performed within [n] to ensure that the total number of array elements for each individual ant satisfies the engineering constraints; for the number of array rings n, within [n] min, n max The values ​​within the circle are randomly selected, and the number of array elements in each ring is dynamically allocated based on m and n to avoid unreasonable element distribution; the radius of each ring is r. i In [r min ,r max Randomly generated within [ ], simultaneously satisfying r i +1>r i Hierarchical relationship; rotation angle θ i In [θ min, θ max Random value is selected from within ] .

[0066] Based on the array fitness value set and the audio receiving array topology parameter set, conditional ant colony iteration is performed to obtain the target array topology parameter set. The conditional ant colony iteration includes updating array pheromones and selecting probabilistic paths based on the array fitness value set to obtain the next generation array fitness value set. The probabilistic path selection formula is:

[0067]

[0068] In the formula, k is the array ant obtained based on the set of topology parameters of the audio receiving array. Let τ be the probability that ant k moves from node i to node j. ij (t) represents the pheromone concentration obtained based on the array pheromone update operation, η ij (t) represents the preset heuristic information, allowed. k Let be the set of unvisited nodes that array ant k can currently select, α be the pheromone heuristic factor, and β be the expected heuristic factor. This update strategy achieves the reinforcement marking of high-quality solution paths by positively mapping the pheromone concentration of the array fitness value set, guiding subsequent ants to prioritize the exploration of high-value areas. At the same time, the pheromone evaporation mechanism avoids premature convergence of the search and maintains the global search capability of the algorithm.

[0069] The preset heuristic information η in the probability path selection formulaij The formula for calculating (t) is:

[0070]

[0071] In the formula, MSL ij The maximum sidelobe level measures the array's anti-interference capability. A lower value indicates stronger suppression of ambient noise (such as background humming in substations); MBW ij The main lobe width determines the beam focusing accuracy. A smaller value results in more accurate spatial localization of the noise source; MSL target MBW target | represents the target performance threshold preset based on the power equipment topology, and ∈ represents the engineering fault tolerance factor, which serves as a numerical stability guarantee to prevent the denominator from being zero.

[0072] An audio receiving array is constructed based on the target array topology parameter set; the array setup and parameter adjustment are performed for transformer noise monitoring scenarios based on the audio receiving array: the array microphones adopt a high-density spiral array, arranged in a honeycomb topology, with a single high-precision analog microphone embedded next to them, forming a heterogeneous sensor array. The array is arranged around the transformer, 1-2 meters away from the device, with a sampling frequency of 200kHz, covering the 1kHz-100kHz frequency band.

[0073] Based on the audio receiving array, monitor the power equipment and acquire abnormal acoustic signals;

[0074] Based on the abnormal acoustic signals, the state is identified, and the state identification result of the power equipment is output.

[0075] Please see Figure 2 , Figure 3 , Figure 4 and Figure 5 The audio receiving array optimized using the ant colony algorithm exhibits better main lobe focusing and side lobe suppression performance than the unoptimized array under the same test frequency, test distance, and dynamic range. In this embodiment, the topology of power equipment in a real-world application scenario is obtained, and an audio receiving array is dynamically constructed based on different topologies of the power equipment. During the construction of the audio array, the ant colony algorithm is used to iterate conditions within the initial set of audio receiving array topology parameters, ultimately obtaining a target array topology parameter set to match the corresponding power equipment topology. This ensures that the audio receiving array constructed based on the target array topology parameters can effectively capture key abnormal noise characteristics of the power equipment and suppress environmental noise, thereby improving the integrity and signal-to-noise ratio of the abnormal noise acoustic signal. The high-quality abnormal noise acoustic signal obtained based on the preset array topology provides excellent input for subsequent state identification, ultimately achieving accurate acquisition of the power equipment topology for abnormal noises.

[0076] It should be noted that,Figure 2 (a) is a two-dimensional schematic diagram of a conventional array, where the vertical axis represents the x-axis in meters and the horizontal axis represents the y-axis in meters; Figure 2 (b) is a two-dimensional schematic diagram of the optimized array, where the vertical axis represents the x-axis in meters and the horizontal axis represents the y-axis in meters; Figure 3 (a) shows the single-source localization results of a conventional array at 2kHz. The vertical axis represents the x-axis in meters, and the horizontal axis represents the y-axis in meters. Figure 3 (b) The optimized array provides single-source localization results at 2kHz. The vertical axis represents the x-axis in meters, and the horizontal axis represents the y-axis in meters. Figure 3 (c) shows the four-source localization results of a conventional array at 100kHz. The vertical axis represents the x-axis in meters, and the horizontal axis represents the y-axis in meters. Figure 3 (d) shows the four-source localization results of the optimized array at 100kHz. The vertical axis represents the x-axis in meters, and the horizontal axis represents the y-axis in meters. Figure 4 The curve represents the iterative curve, with the vertical axis representing the sidelobe level and the horizontal axis representing the number of iterations.

[0077] Further, an initial array ant colony set is constructed based on the audio receiving array topology parameter set, and multi-index fitness calculation is performed based on the initial array ant colony set to obtain the array fitness value set; the multi-index fitness calculation formula is:

[0078] Fitness = y1·MSL + y2·MBW

[0079] In the formula, Fitness is the array fitness value, y1 is the first weight coefficient, y2 is the second weight coefficient, MSL is the maximum sidelobe level obtained based on the power equipment topology, and MBW is the main lobe width obtained based on the power equipment topology.

[0080] In this embodiment, by mapping the discrete set of audio receiver array topology parameters to the initial array ant colony set of the ant colony algorithm, an effective initial search starting point is provided for subsequent optimization. A multi-index fitness function is constructed, with maximum sidelobe level (MSL) as the primary index and main lobe width (MBW) as the secondary index, and differentiated optimization of the primary and secondary indices is achieved using the first and second weighting coefficients.

[0081] It should be noted that the following constraints can be set in the multi-index fitness function mentioned above: y1 + y2 = 1 and y1 > y2. These conditions aim to prioritize optimizing the maximum sidelobe level (MSL), as it directly affects the array's anti-interference capability. Prioritizing the reduction of the MSL value ensures the signal purity of the array in complex electromagnetic environments. While optimizing the MSL, the main lobe width (MBW) is also optimized, as it determines beam pointing accuracy, thus comprehensively improving beamforming performance. By adjusting the weighting coefficients y1 and y2, different application scenarios can be flexibly adapted: for example, in mid-to-high frequency scenarios, the focus can be on reducing MSL (increasing y1) to reduce interference, while in low-frequency scenarios, the focus is on optimizing MBW (relatively increasing y2) to improve target positioning accuracy.

[0082] Further, based on the array fitness value set and the audio receiving array topology parameter set, conditional ant colony iteration is performed to obtain the target array topology parameter set; including:

[0083] Based on the array fitness value set and the audio receiving array topology parameter set, array pheromone updates and probabilistic path selection are performed to obtain the next-generation array fitness value set and the next-generation audio receiving array topology parameter set.

[0084] Repeat the above steps until the number of iterations is greater than or equal to the preset number of iterations, or the set of fitness values ​​after iteration is greater than or equal to the preset fitness threshold, and then output the set of topology parameters of the target array.

[0085] In this embodiment, by setting dual termination conditions—the number of iterations and a fitness threshold—the search terminates when the number of conditional ant colony iterations based on the array fitness value set and the audio receiver array topology parameter set reaches a preset number of iterations to ensure controllable computational resources. If the fitness value set after iteration reaches or exceeds the preset fitness threshold, it indicates that an optimized result meeting engineering requirements has been obtained, and the iteration also terminates. This dual-condition judgment effectively avoids getting bogged down in ineffective over-iterations while ensuring that the output target array topology parameter set meets performance requirements, thus achieving a balance between computational efficiency and optimization effect, and improving the engineering practicality of this embodiment. Utilizing the swarm intelligence characteristics of the ant colony algorithm and under the aforementioned dual termination condition constraints, this embodiment achieves automated and intelligent search and optimization of the audio receiver array topology parameters, ultimately outputting a highly optimized target array topology parameter set that meets predetermined acoustic performance standards.

[0086] Furthermore, based on the abnormal acoustic signals, state identification is performed, and the power equipment state identification result is output; including:

[0087] Based on the aforementioned abnormal acoustic signal, feature extraction is performed to obtain the abnormal acoustic feature vector;

[0088] Based on the abnormal noise feature vector, transfer learning adaptation is performed to obtain the adaptation feature vector;

[0089] Based on the adapted feature vector, state identification is performed, and the state identification result of the power equipment is output.

[0090] In this embodiment, feature extraction is first performed on the abnormal noise acoustic signal, transforming the original high-dimensional abnormal noise acoustic signal into an abnormal noise feature vector with stronger physical representation. This effectively filters out key acoustic features related to the state of the power equipment and removes redundant information. Subsequently, the abnormal noise feature vector is transformed into an adapted feature vector through transfer learning, improving the robustness of the abnormal noise feature vector to on-site background noise, individual equipment differences, and changes in operating conditions. Finally, state identification is performed based on the adapted feature vector, and the optimized abnormal noise feature vector, i.e., the adapted feature vector, is used to output the power equipment state identification result, providing a highly reliable acoustic diagnostic basis that can directly guide decision-making for intelligent operation and maintenance and preventive maintenance of the equipment.

[0091] It should be noted that, in this embodiment, before feature extraction based on the abnormal noise acoustic signal, the abnormal noise acoustic signal is preprocessed and subjected to wavelet transform denoising: a 5-level decomposition is performed using the db4 wavelet, and a soft threshold is selected as the threshold function to remove white noise and impulse interference; Z-score normalization: the signal is normalized to a distribution X' with a mean of 0 and a standard deviation of 1.

[0092]

[0093] In the formula, X represents the original abnormal acoustic signal data points, μ represents the signal mean, and σ represents the signal standard deviation;

[0094] Integrated Empirical Mode Decomposition (EEMD): Add Gaussian white noise with a standard deviation of 0.2, set the number of decomposition layers to 10, and remove periodic interference (such as 100Hz and 200Hz components) generated by transformer core vibration.

[0095] Further, feature extraction is performed based on the abnormal noise acoustic signal to obtain an abnormal noise feature vector; including:

[0096] A short-time Fourier transform is performed on the aforementioned abnormal acoustic signal to obtain the abnormal time-frequency matrix;

[0097] The abnormal noise time-frequency matrix is ​​input into a preset residual neural network, so that the preset residual neural network performs deep semantic extraction based on the abnormal noise adaptation matrix to obtain a preliminary feature vector;

[0098] Based on the initial feature vector, attention weighting is performed to obtain the abnormal noise feature vector.

[0099] In this embodiment, the original one-dimensional abnormal noise acoustic signal is first converted into a two-dimensional abnormal noise time-frequency matrix using short-time Fourier transform. A Hanning window (window length N = 1024) with a 50% overlap is then used for frame segmentation. The m-th frame signal x... m (n) is:

[0100] x m (n)=x(n+m·N / 2)·w(n), (n=0,1,…,N-1)

[0101] In the formula, w(n) is the Hanning window function.

[0102] The abnormal noise time-frequency matrix is ​​calculated by Discrete Fourier Transform (DFT):

[0103]

[0104] In the formula, X m (k) represents the elements of the abnormal noise time-frequency matrix, where k represents the frequency point index and the corresponding frequency.

[0105] f s To achieve a high sampling rate, joint time-domain and frequency-domain feature capture is implemented, effectively capturing the time-varying spectral features of non-steady-state acoustic events. Subsequently, the abnormal noise time-frequency matrix is ​​input into a pre-defined residual neural network. Leveraging the deep nonlinear mapping and residual jump connections of this network, it adaptively learns from complex background noise and extracts deep abstract features (i.e., preliminary feature vectors) containing the state semantics of power equipment, overcoming the subjectivity and limitations of manual feature design. Finally, an attention-weighted mechanism dynamically filters the preliminary feature vectors, assigning higher weights to key time-frequency regions strongly correlated with faults while suppressing the contribution of irrelevant or interfering features. This outputs highly focused and discriminative abnormal noise feature vectors, ensuring the input quality and diagnostic reliability of the subsequent state identification module. The deep feature extraction method based on residual neural networks and attention mechanisms can automatically learn the time-frequency dynamic features and electromagnetic-mechanical composite noise patterns of abnormal noise acoustic signals. Compared to traditional spectrum analysis, this embodiment improves the accuracy of feature extraction for early faults such as component loosening and wear.

[0106] It should be noted that the preset residual neural network may include four residual blocks and one global average pooling layer. The structure of a single residual block can satisfy the following relationship:

[0107] y=F(x,{W i})+x

[0108] In the formula, x is the input feature map of the residual block; F(·) is the residual function; W iy represents the convolution kernel weights; y represents the residual block output. The above residual block structure directly transmits shallow features x to the output layer through identity skip connections, which can effectively alleviate the gradient vanishing problem in deep network training.

[0109] The four residual blocks are processed sequentially to extract deep features (such as frequency changes of abnormal equipment noise, periodic pulse patterns, etc.), and finally output a feature map (dimension: [1,H,W,C], where H and W are the feature map size and C is the number of channels).

[0110] The global average pooling layer compresses the spatial dimension of the feature map and calculates the feature mean of each channel C using the following formula:

[0111]

[0112] In the formula, the global feature is z g As the input for subsequent attention weighting, H and W are the feature map dimensions, and X(h,w) represents the feature value at the h-th row and w-th column position, reflecting the feature information at that position, which is used for subsequent operations such as global average pooling to extract more abstract features.

[0113] The attention weighting can compress global features using a multilayer perceptron (MLP), and channel attention weights a can be generated using a sigmoid activation function. c :

[0114] a c =σ(MLP(z) g ))

[0115] In the formula, σ(x) is the sigmoid activation function, which outputs weights from 0 to 1 to highlight the fault channel. Then, the weights are multiplied channel by channel, using attention weights to enhance fault features and suppress noise channels.

[0116] X att =X⊙a c

[0117] In the formula, ⊙ represents channel-by-channel multiplication, X att These are weighted features, which serve as input to the subsequent transfer learning adaptation module. The above operations can enhance the response of the fault feature channels while suppressing the contribution of noise-dominated channels, thereby amplifying the fault features.

[0118] Further, transfer learning adaptation is performed based on the aforementioned abnormal sound feature vector to obtain an adapted feature vector; including:

[0119] The abnormal noise feature vector is input into a preset source domain model so that the preset source domain model performs feature mapping based on the abnormal noise feature vector to obtain a high-dimensional feature vector.

[0120] The high-dimensional feature vector is input into a preset domain adaptation network, so that the preset domain adaptation network performs loss calculation based on the high-dimensional feature vector to obtain a joint optimized loss value.

[0121] Based on the joint optimization loss value and the abnormal noise feature vector, iterative optimization is performed to obtain the adaptive feature vector.

[0122] In this embodiment, the abnormal noise feature vector is first input into a pre-trained source domain model on abundant source domain data. Utilizing its learned high-order semantic mapping capabilities, the original features are transformed into a more discriminative high-dimensional feature vector, inheriting common fault knowledge from the source domain. Subsequently, the high-dimensional feature vector is input into a pre-trained domain adaptation network. By calculating a joint optimization loss value, including domain difference loss and task classification loss, the distribution difference between the target domain features and the source domain features is explicitly measured and reduced. Finally, based on this joint loss value, the feature extraction and adaptation parameters are iteratively optimized in reverse, forcing the feature space to align the data distribution of the source and target domains while retaining key fault information. The final output is an adapted feature vector that combines domain invariance and fault sensitivity, enabling the subsequent state identification module to maintain high-precision diagnostic capabilities even in real-world scenarios with scarce target device annotations and variable background noise.

[0123] It should be noted that, in this embodiment, the preset source domain model is pre-trained using normal / abnormal acoustic data from 1000 transformers of different models, and the source domain pre-training data features are extracted. (High-dimensional features after deep feature extraction, consistent with the feature dimension of the target domain), where n is the number of samples in the source domain.

[0124] Further, the high-dimensional feature vector is input into a preset domain adaptation network, so that the preset domain adaptation layer performs loss calculation based on the high-dimensional feature vector to obtain the joint loss value; the preset domain adaptation network includes an MMD loss calculation module, a center loss calculation module, a classification loss calculation module, and a joint module; wherein:

[0125] The high-dimensional feature vector is input into a preset domain adaptation network so that the MMD loss calculation module, the center loss calculation module, and the classification loss calculation module perform loss calculations based on the high-dimensional feature vector, and obtain the MMD loss value, the center loss value, and the classification loss value, respectively.

[0126] The specific calculation formula for the MMD loss calculation module is as follows:

[0127]

[0128] In the formula, z s,i Let z be the i-th high-dimensional eigenvector of the source domain. t,jLet be the high-dimensional feature vector of the j-th sample in the target domain, φ(·) be the Gaussian kernel function, and n and m be the number of high-dimensional feature vectors in the source and target domains, respectively. For the regenerated nucleus Hilbert space.

[0129] The calculation formula for the center loss calculation module is as follows:

[0130]

[0131] In the formula, z i This is the high-dimensional feature vector output by the MMD loss calculation module. For the yth i The central feature vector of the class, where N is the number of samples in the current batch.

[0132] The specific calculation formula for the classification loss calculation module is as follows:

[0133]

[0134] In the formula, y i,c For real labels, p i,c To predict probabilities, c represents the category index, and C represents the total number of device states;

[0135] The MMD loss value, center loss value, and classification loss value are input into the joint module to perform joint optimization, obtaining the joint optimized loss value. The total loss function is:

[0136]

[0137] In the formula, min To jointly optimize the loss value, This represents the MMD loss value. The center loss value, Let λ be the classification loss value, λ be the center loss weight coefficient, and γ be the classification loss weight coefficient. λ and γ are jointly optimized through backpropagation to achieve a triple balance.

[0138] In this embodiment, high-dimensional feature vectors are simultaneously input into a preset domain adaptation network for feature adaptation. Cross-domain feature alignment and discriminative capability are maintained through collaborative optimization of a triple loss function: the MMD loss calculation module calculates the maximum mean difference (MMD) statistic between the feature distributions of the source and target domains, minimizing this loss to drive global distribution alignment of features in both domains, effectively reducing the overall acoustic feature shift caused by cross-device or cross-scene conditions. The center loss module constrains feature vectors of similar faults to cluster towards their corresponding class centers in the latent space, maximizing intra-class compactness and avoiding a decrease in class separability caused by intra-class feature dispersion. Simultaneously, the classification loss module continuously optimizes the feature's discriminative ability for fault categories based on source domain labeled data, ensuring that the semantic discriminativeness of the feature space is not weakened during domain alignment. The MMD loss value, center loss value, and classification loss value output by the above modules are input into the joint module to generate a unified joint optimized loss value. During backpropagation of the jointly optimized loss value, the pre-defined domain adaptation network is simultaneously constrained to meet three objectives: domain invariance, intra-class aggregation, and fault separability. Finally, based on the iterative optimization process of the jointly optimized loss value and the high-dimensional abnormal noise feature vector, an adapted feature vector with both high cross-domain adaptability and strong fault discrimination confidence is output. This embodiment effectively overcomes the limitation of a single loss function in balancing domain alignment and task performance, providing core feature adaptation support for the reliable application of power equipment abnormal noise diagnosis in complex real-world scenarios. By minimizing inter-domain distribution differences through MMD and combining it with center loss to optimize intra-class aggregation, the dependence on target domain labeled data is significantly reduced when applying across devices or scenarios. This alleviates the generalization challenge faced by traditional deep learning methods due to high sample dependence and improves state classification accuracy.

[0139] Further, based on the adapted feature vector, state identification is performed, and the power equipment state identification result is output; including:

[0140] The adapted feature vectors are weighted and fused using a preset multi-classifier to obtain the fused class probabilities.

[0141] Fault location is performed based on the fused category probability, and the power equipment status identification result is output.

[0142] Further, the adapted feature vectors are weighted and fused using a preset multi-classifier to obtain the fusion probability; including:

[0143] The adaptive feature vectors are weighted and fused through a preset multi-classifier to obtain an adaptive weight set and a probability output vector set.

[0144] The adaptive weight set and the probability output vector set are fused to obtain the fused probability:

[0145]

[0146] In the formula, S is the fusion probability, and w i Let C be the i-th adaptive weight in the adaptive weight set. i (x) is the i-th probability output vector in the probability output vector set.

[0147] In this embodiment, the adapted feature vector is first input into a preset multi-classifier: a computationally efficient LightGBM classifier (tree depth 8, learning rate 0.01, 1000 samples) and a SoftMax classifier (10 fully connected neurons) that inherits deep features to extract semantic information. The classifiers' decision preferences are used to obtain complementary fault discrimination information, generating an adaptive weight set that reflects the reliability assessment and probability output vector set of each classifier under specific input features. The adaptive weight set and the probability output vector set are then fused, dynamically strengthening the decision result of the classifier with the highest confidence while suppressing interference from low-reliability models, thus outputting the fused category probability that best matches the actual fault mode. Finally, fault location is performed based on this fused probability, accurately distinguishing the health status, early faults, and severe faults of power equipment, outputting power equipment status identification results with clear fault type labeling and confidence quantification. This embodiment effectively overcomes the shortcomings of insufficient generalization of a single classifier and significantly improves the robustness and engineering interpretability of diagnostic results in complex acoustic scenarios.

[0148] Furthermore, the adapted feature vectors are weighted and fused using a preset multi-classifier to obtain an adaptive weight set. The weighted fusion calculation formula is as follows:

[0149]

[0150] In the formula, w i For the i-th adaptive weight, To preset the accuracy of the i-th classifier in the multi-classifier algorithm. To preset the feature space coverage of the i-th classifier in a multi-classifier system, i.e., the classifier's "confidence coverage capability" for samples, ∑ j Accuracy j The sum of the accuracies of all classifiers, ∑ j Coverage j This is the sum of the feature space coverage of all classifiers.

[0151] In one embodiment, on a certain transformer test set, LightGBM has an accuracy of 92% and a coverage of 85%; SoftMax has an accuracy of 88% and a coverage of 90%. Therefore, the LightGBM weight is:

[0152]

[0153] In this embodiment, the weighted fusion calculation integrates classification results and introduces a two-dimensional index of accuracy and feature space coverage to address the performance limitations of traditional single accuracy weights in scenarios with fluctuating operating conditions. i As prior knowledge, it reflects the long-term statistical performance of the i-th classifier on the validation set, ensuring that a highly reliable classifier dominates the fusion process; Feature space coverage i The matching degree between the current adaptive feature vector and the classifier's advantageous decision region is dynamically quantified, so that the weight allocation responds to the input characteristics in real time. This embodiment realizes the adaptive integration of the advantages of different classifiers, improving the accuracy and reliability of device status classification.

[0154] In another embodiment, samples with the same distribution as the source and target domains (such as historical fault data of the transformer to be monitored + simulated fault data of equipment under similar operating conditions) are selected to ensure that the accuracy and coverage indicators truly reflect the performance of the classifier in the power scenario. Accuracy and coverage are recalculated every 100 hours (or every 1000 new data points), and the weights wi are dynamically adjusted to adapt to feature drift caused by equipment aging and changes in operating conditions. The adapted feature vectors are weighted and fused through a preset multi-classifier, and then processed using a LightGBM classifier: the features are mapped to a decision tree space, and the class probability P is output. lgbm ; Using SoftMax processing: The features are passed through a fully connected layer + SoftMax, and the output class probability P is obtained. softmax .

[0155] The adaptive weight set and the probability output vector set are fused to obtain the fused probability:

[0156] P fusion =ω1·P lgbm +ω2·P softmax

[0157] Where ω2=1-ω1.

[0158] Fault location is performed based on the fused category probabilities, outputting power equipment status identification results, including fault type (e.g., "partial discharge, severity level 4"), confidence level (e.g., 0.92, further determined by a threshold to trigger an alarm), and abnormal noise location (determined using microphone array-based acoustic imaging technology). An alarm is triggered when the confidence level (threshold, freely adjustable) and the fault type is "severe" (e.g., partial discharge severity ≥ level 3). After each manual confirmation of a fault, new samples are fed back into the test set, accuracy and coverage are recalculated, and the model and classifier weights are updated, forming a data loop.

[0159] Example 2

[0160] This embodiment provides a method for monitoring and identifying abnormal noises in power equipment in a high-voltage switchgear noise monitoring scenario:

[0161] Obtain the topology of power equipment;

[0162] Initialize the audio receiving array topology parameter set based on the power equipment topology;

[0163] An initial array ant colony set is constructed based on the audio receiving array topology parameter set, and a multi-index fitness calculation is performed based on the initial array ant colony set to obtain an array fitness value set.

[0164] Based on the array fitness value set and the audio receiving array topology parameter set, conditional ant colony iteration is performed to obtain the target array topology parameter set.

[0165] Construct an audio receiving array based on the target array topology parameter set;

[0166] Based on the aforementioned audio receiving array, array settings and parameter adjustments are performed for high-voltage switchgear abnormal noise monitoring scenarios: a high-density, multi-channel spiral array microphone is used, arranged in front of the switchgear cabinet, with a sampling frequency of 200kHz (focusing on the high-frequency characteristics of switchgear discharge noise, primarily covering the 10kHz-50kHz frequency band).

[0167] Based on the audio receiving array, monitor the power equipment and acquire abnormal acoustic signals;

[0168] The abnormal acoustic signal is preprocessed: the wavelet transform uses the sym8 wavelet to remove transient interference caused by the switching operation; the EEMD decomposition layer is set to 8 layers to extract the non-periodic features generated by partial discharge.

[0169] Based on the abnormal acoustic signals, the state is identified, and the state identification result of the power equipment is output.

[0170] Based on the aforementioned abnormal acoustic signals, state identification is performed, and the power equipment state identification result is output; including:

[0171] Based on the aforementioned abnormal acoustic signal, feature extraction is performed to obtain the abnormal acoustic feature vector;

[0172] Based on the abnormal noise feature vector, transfer learning adaptation is performed to obtain the adaptation feature vector;

[0173] Based on the adapted feature vector, state identification is performed, and the state identification result of the power equipment is output.

[0174] In this embodiment, the topology of the power equipment in the high-voltage switchgear abnormal noise monitoring scenario is acquired, and an audio receiving array is dynamically constructed based on the different topologies of the power equipment. During the construction of the audio array, an ant colony algorithm is used to iterate conditions within the initial set of audio receiving array topology parameters, ultimately obtaining a target array topology parameter set to match the corresponding power equipment topology. This ensures that the audio receiving array constructed based on the target array topology parameters can effectively capture the key abnormal noise characteristics of the power equipment and suppress environmental noise, thereby improving the integrity and signal-to-noise ratio of the abnormal noise acoustic signal. The high-quality abnormal noise acoustic signal obtained based on the preset array topology provides excellent input for subsequent state identification, ultimately achieving accurate acquisition of the power equipment topology for abnormal noise.

[0175] Further, feature extraction is performed based on the abnormal noise acoustic signal to obtain an abnormal noise feature vector; including:

[0176] A short-time Fourier transform is performed on the aforementioned abnormal acoustic signal to obtain the abnormal time-frequency matrix;

[0177] The abnormal noise time-frequency matrix is ​​input into a preset residual neural network, so that the preset residual neural network performs deep semantic extraction based on the abnormal noise adaptation matrix to obtain a preliminary feature vector;

[0178] Based on the initial feature vector, attention weighting is performed to obtain the abnormal noise feature vector.

[0179] In this embodiment, the abnormal noise feature vector includes time-frequency (STFT, MFCC), spatial (beamforming power, spatial spectral entropy, transient (pulse rising edge), and energy centroid characteristics); firstly, the switchgear is divided into three areas: circuit breaker compartment, busbar compartment, and cable compartment, and a regional feature database is established:

[0180] 1. Typical fault characteristics of circuit breaker compartment: including circuit breaker body, operating mechanism and insulating bushing. Typical faults are arc discharge (characteristic frequency 20-50kHz) and mechanical vibration (1-10kHz periodic signal).

[0181] 2. Typical fault characteristics of busbar compartment: including busbars, insulators and wall bushings, the main faults are insulation aging (wideband noise 10-100kHz) and poor contact (pulse discharge signal);

[0182] 3. Typical fault characteristics of cable rooms: including cable joints, grounding devices and surge arresters. Typical faults are local discharge (30-300kHz narrowband signal) and discharge-induced temperature rise (accompanied by low-frequency vibration coupling).

[0183] Secondly, a priori feature sets of fault signals for each region are established, including: time-frequency domain features: Short-Time Fourier Transform (STFT) time-spectrum matrix.j (f,t), Mel frequency cepstral coefficients (MFCC) j The acoustic signal x(n) from the power equipment is processed by the STFT to obtain the time-frequency matrix STFT. j (f,t)), with dimension (time frame number T × frequency point number F). After Mel filtering, logarithmic compression, and DCT transform, the MFCC output dimension is (in this embodiment, considering the integrity of information, its order is set to M = 12). To construct a unified feature vector, global pooling is performed on the STFT time-frequency matrix to reduce its dimension to , and global pooling is performed on the MFCC to reduce its dimension to , and then they are concatenated:

[0184]

[0185] In the formula, FeatureFusion(·) represents the feature fusion function, which includes operations such as pooling, concatenation, and normalization; f t,f The STFT and MFCC features after splicing and pooling are represented by (1×(F+M)); in (j,t), j represents the region index, corresponding to three different regions: circuit breaker room, busbar room, and cable room, used to distinguish the acoustic signals and corresponding features collected in different regions; t is the time dimension index, used to identify the time sequence position of the signal in the spectrum matrix during short-time Fourier transform (STFT), and to accurately locate the acoustic signal features at different times.

[0186] Spatial spectrum characteristics: beamforming output power spectrum, spatial spectrum entropy value;

[0187] The beamforming output power spectrum is calculated as follows: For a microphone array consisting of K elements, the received signal x from the i-th channel... i (n) is:

[0188] x i (n)=s(n-τ i )+v i (n)s(n) is the sound source signal, τ i v is the propagation delay from the sound source to the i-th microphone. i (n) represents noise.

[0189] By applying a direction-dependent delay to the signals of each channel and then weighting and summing them:

[0190]

[0191] In the formula, For direction-related weights, w is typically chosen. i =1, K is the number of array elements.

[0192] Beamforming output Calculate short-time power:

[0193]

[0194] In the formula, L is the short-time analysis window length, and t is the time frame index. This reflects the location of the fault sound source (such as a partial discharge point) in space. Energy distribution in direction; for electrical equipment, the discharge point usually manifests as a high power peak in space (such as a bus discharge forming a significant power peak at a specific angle).

[0195] The spatial spectral entropy is calculated as follows, after normalizing the power spectrum:

[0196]

[0197] In the formula, H s The entropy value reflects the spatial concentration of fault energy. A low entropy value indicates that the fault sound source is concentrated in a few directions (e.g., partial discharge points exhibit sharp power peaks with highly concentrated spatial distribution); a high entropy value indicates that the fault sound source is dispersed (e.g., loose screws in multiple locations within a switchgear cabinet lead to a uniform distribution of noise energy). This allows for beamforming to output the power spectrum. and spatial spectral entropy value H s (Both eigenvalues ​​are one-dimensional vectors) concatenated as

[0198] Transient characteristics: Discharge pulse rise time τ r Energy center of gravity frequency f g =∑P(f)·f / ∑P(f), where ∑P(f) represents the summation of the power P(f) of the discharge pulse signal at each frequency point f, and represents the total power of the discharge pulse signal.

[0199] The rise time of the discharge pulse is calculated as follows: For the discharge pulse signal, the start point (at 10% of the amplitude) and the end point (at 90% of the amplitude) of the rise time are detected, and the rise time is calculated as follows:

[0200] τ r =t end -t start

[0201] In the formula, τ r The τ value reflects the type of discharge, such as arc discharge (fast-acting). r Less than insulation aging noise (slow time), t end This represents the time point when the amplitude of the discharge pulse signal reaches 90% during the rising edge of the discharge pulse signal, i.e., the end time of the discharge rising edge, t. start This indicates the time point when the amplitude reaches 10% during the rising edge of the discharge pulse signal, which is the starting time of the discharge rising edge.

[0202] The energy centroid frequency is calculated as follows: for the spectrum P(f) of the pulse signal, the energy centroid frequency is calculated:

[0203]

[0204] In the formula, f g This indicates the main frequency of the energy source used to locate faults.

[0205] The rise time of the discharge pulse and the energy centroid frequency (both of which are one-dimensional vectors) can then be concatenated to form f. transition ={τ r ;f g}

[0206] You will then receive:

[0207] F = FeatureFusion{f t,f ;f spatial ;f transitin}

[0208] In the formula, FeatureFusion(·) represents the feature fusion function, which includes operations such as pooling, concatenation, and normalization.

[0209] Next, a probabilistic mapping model based on Bayesian decision theory is constructed. First, the prior feature set Γ = {f} is defined. j,i}, where f j,i It is region R j (e.g., circuit breaker compartment) The feature vector of the i-th sample (integrating STFT time-frequency, MFCC, spatial spectrum, and transient features). Define the area fault probability function P(R). j |f), where R j ∈{R1,R2,R3} corresponds to three functional regions.

[0210] Using the prior feature set as training samples, kernel density estimation (KDE) is employed to establish the conditional probability density p(f|R) of each region's features. i ), combined with the region prior probability P(R j (Based on historical fault data statistics), construct a probability mapping model:

[0211]

[0212] in, represents the total probability of the eigenvectors.

[0213] By combining historical fault statistics with equipment operating status, a regional prior probability P(Rj) is defined, and dynamic adjustments based on equipment aging are considered:

[0214] P′(R j )=P(R j )·λj (t),

[0215] Where, μ j λ is the regional aging coefficient, t is the equipment service time, and λ is the aging coefficient. j (t) is the equipment aging impact factor used to dynamically correct the regional prior probability P(Rj), and is a function of the equipment service time t, expressed in exponential form. Reflecting the regional aging coefficient μ j The impact of service time t on equipment failure probability reflects the degree of correction for changes in failure probability due to aging as equipment service time increases.

[0216] Calculate the regional failure probability using Bayes' theorem:

[0217]

[0218] The feature vector is input into a Bayesian classifier, which outputs P(R1), P(R2), and P(R3). These are combined with beamforming to form the location coordinates and region probability distribution, and the maximum a posteriori probability criterion is applied.

[0219]

[0220] Wherein, δ(D) j ) is the distance weighting function, D j For the location point to region R j The Euclidean distance from the geometric center.

[0221] Further, transfer learning adaptation is performed based on the aforementioned abnormal sound feature vector to obtain an adapted feature vector; including:

[0222] The abnormal noise feature vector is input into a preset source domain model so that the preset source domain model performs feature mapping based on the abnormal noise feature vector to obtain a high-dimensional feature vector.

[0223] The high-dimensional feature vector is input into a preset domain adaptation network, so that the preset domain adaptation network performs loss calculation based on the high-dimensional feature vector to obtain a joint optimized loss value.

[0224] Based on the joint optimization loss value and the abnormal noise feature vector, iterative optimization is performed to obtain the adaptive feature vector.

[0225] In this embodiment, the preset source domain model includes the normal, discharge, and mechanical loosening status data of 500 switchgear cabinets.

[0226] Further, the high-dimensional feature vector is input into a preset domain adaptation network, so that the preset domain adaptation layer performs loss calculation based on the high-dimensional feature vector to obtain the joint loss value; the preset domain adaptation network includes an MMD loss calculation module, a center loss calculation module, a classification loss calculation module, and a joint module; wherein:

[0227] The high-dimensional feature vector is input into a preset domain adaptation network so that the MMD loss calculation module, the center loss calculation module, and the classification loss calculation module perform loss calculations based on the high-dimensional feature vector, and obtain the MMD loss value, the center loss value, and the classification loss value, respectively.

[0228] The MMD loss value, center loss value, and classification loss value are input into the joint module so that the joint module can perform joint optimization to obtain the joint optimized loss value.

[0229] In this embodiment, the differences in acoustic feature distribution between different switchgear models are further reduced by combining the MMD loss calculation module, the center loss calculation module, and the classification loss calculation module. The target domain (new switchgear model) only needs 20 labeled samples to complete the adaptation.

[0230] Further, based on the adapted feature vector, state identification is performed, and the power equipment state identification result is output; including:

[0231] The adapted feature vectors are weighted and fused using a preset multi-classifier to obtain the fused class probabilities.

[0232] Fault location is performed based on the fused category probability, and the power equipment status identification result is output.

[0233] In this embodiment, a Bayesian classifier is used to calculate the fault probability P(R) of the corresponding power equipment functional areas in different regions R1, R2, and R3. j |f), combined with the distance weighting function α is the attenuation coefficient (e.g., 0.1 / m, with higher weighting for closer distances) to achieve accurate identification of defect location and type, and output a comprehensive score S. j =P(R) j ∣f)·δ(D j ), where j = 1, 2, 3 correspond to three regions. Through The region with the highest overall probability is selected as the fault region, i.e., "the fault region is determined by the highest posterior probability (fusion distance weighted)," representing the overall score S corresponding to the three regions (j=1,2,3). j In the middle, find the one that makes S jThe largest region index is used to determine the region where the fault is most likely to occur. Finally, the joint feature parameters and beamforming results are comprehensively evaluated, and the fault type features in the feature vector f are matched against a predefined fault type library to output the result. Fault type = T match T match To match the fault type, the region, type, and beam are integrated to form the location coordinates, and the final result is output: "Region" Fault occurred, location coordinates An azimuth angle (such as a horizontal angle in space relative to a reference direction) indicating the location of the fault; The pitch angle (e.g., the vertical angle relative to the horizontal plane in space) indicates the location of the fault. These two parameters, combined, are used to accurately pinpoint the fault's coordinates in space. This enables a precise closed-loop identification of the "feature-location-type" sequence.

[0234] It should be noted that, in the process of identifying the state based on the aforementioned adaptive feature vector and outputting the state identification results of the power equipment, this embodiment can set the warning threshold based on the quantile method (95th percentile) of historical data. When the switch cabinet temperature rises or the load increases, the threshold is automatically adjusted upward by 5% to 10% to reduce the false alarm rate.

[0235] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for monitoring abnormal sound and identifying state of power equipment, characterized in that, The method comprises the following steps: obtaining a power equipment topology structure; initializing an audio receiving array topology parameter set based on the power equipment topology structure; constructing an initial array ant colony set based on the audio receiving array topology parameter set, and performing multi-index fitness calculation based on the initial array ant colony set to obtain an array fitness value set; performing conditional ant colony iteration based on the array fitness value set and the audio receiving array topology parameter set to obtain a target array topology parameter set; monitoring the power equipment based on the audio receiving array to obtain an abnormal sound acoustic signal; performing state recognition based on the abnormal sound acoustic signal to output a power equipment state recognition result. The method comprises the following steps:

2. The method of claim 1, wherein, constructing an initial array ant colony set based on the audio receiving array topology parameter set, and performing multi-index fitness calculation based on the initial array ant colony set to obtain an array fitness value set; the multi-index fitness calculation formula is: Fitness=y1·MSL+y2·MBW wherein Fitness is an array fitness value, y1 is a first weight coefficient, y2 is a second weight coefficient, MSL is a maximum side lobe level obtained based on the power equipment topology structure, and MBW is a main lobe width obtained based on the power equipment topology structure.

3. The method of claim 1, wherein the method further comprises: The method comprises the following steps: performing array pheromone update and probability path selection based on the array fitness value set and the audio receiving array topology parameter set to obtain a next-generation array fitness value set and a next-generation audio receiving array topology parameter set; repeating the above steps until the iteration number is greater than or equal to a preset iteration number, or the fitness value set after iteration is greater than or equal to a preset fitness threshold, and outputting the target array topology parameter set.

4. The method of claim 1, wherein, The method comprises the following steps: performing feature extraction based on the abnormal sound acoustic signal to obtain an abnormal sound feature vector; performing transfer learning adaptation based on the abnormal sound feature vector to obtain an adapted feature vector; performing state recognition based on the adapted feature vector to output a power equipment state recognition result.

5. The method of claim 4, wherein the method further comprises: The method comprises the following steps: performing short-time Fourier transform based on the abnormal sound acoustic signal to obtain an abnormal sound time-frequency matrix; inputting the abnormal sound time-frequency matrix into a preset residual neural network to enable the preset residual neural network to perform deep semantic extraction based on an abnormal sound adaptation matrix to obtain a preliminary feature vector; performing attention weighting based on the preliminary feature vector to obtain an abnormal sound feature vector.

6. The method of claim 4, wherein the method further comprises: The method comprises the following steps: inputting the abnormal sound feature vector into a preset source domain model to enable the preset source domain model to perform feature mapping based on the abnormal sound feature vector to obtain a high-dimensional feature vector; Input the high-dimensional feature vector into a preset domain adaptation network, so that the preset domain adaptation network performs loss calculation based on the high-dimensional feature vector to obtain a joint optimization loss value; Iterative optimization is performed based on the joint optimization loss value and the heteroresponse feature vector to obtain an adaptive feature vector.

7. The method of claim 6, wherein the method further comprises: Input the high-dimensional feature vector into a preset domain adaptation network, so that the preset domain adaptation network performs loss calculation based on the high-dimensional feature vector to obtain a joint loss value; the preset domain adaptation network comprises an MMD loss calculation module, a center loss calculation module, a classification loss calculation module and a joint module; wherein: Input the high-dimensional feature vector into a preset domain adaptation network, so that the MMD loss calculation module, the center loss calculation module and the classification loss calculation module perform loss calculation based on the high-dimensional feature vector respectively to obtain an MMD loss value, a center loss value and a classification loss value respectively; Input the MMD loss value, the center loss value and the classification loss value into the joint module, so that the joint module performs joint optimization to obtain a joint optimization loss value.

8. The method of claim 4, wherein the method further comprises: Based on the adaptive feature vector, state recognition is performed to output a power equipment state recognition result; comprising: The adaptive feature vector is weighted and fused through a preset multi-classifier to obtain a fusion class probability; Based on the fusion class probability, fault positioning is performed to output a power equipment state recognition result.

9. The method of claim 8, wherein the method further comprises: The adaptive feature vector is weighted and fused through a preset multi-classifier to obtain a fusion probability; comprising: The adaptive feature vector is weighted and fused through a preset multi-classifier to obtain an adaptive weight set and a probability output vector set; The adaptive feature vector is weighted and fused through a preset multi-classifier to obtain an adaptive weight set and a probability output vector set; where S is the fusion probability, w i is the ith adaptive weight in the set of adaptive weights, C i (x) is the ith probability output vector in the set of probability output vectors.

10. The method of claim 9, wherein the method further comprises: The adaptive feature vector is weighted and fused through a preset multi-classifier to obtain an adaptive weight set and a probability output vector set; The adaptive feature vector is weighted and fused through a preset multi-classifier to obtain an adaptive weight set and a probability output vector set; The adaptive feature vector is weighted and fused through a preset multi-classifier to obtain an adaptive weight set and a probability output vector set; where w i is the ith adaptive weight, Accuracy i is the accuracy of the ith classifier in the preset multi-classifier, Coverage i is the feature space coverage of the ith classifier in the preset multi-classifier, ∑jAccuracy j is the sum of the accuracies of all classifiers, ∑jCoverage j is the sum of the feature space coverages of all classifiers.

Images