A substation equipment state monitoring method based on voiceprint features

Abstract

The application discloses a kind of based on voiceprint feature substation equipment state monitoring method and system, it is related to data processing technical field;Method includes obtaining the voiceprint signal and electrical signal of substation equipment;Voiceprint signal and electrical signal are extracted to obtain voiceprint feature and electrical feature;After the fusion of voiceprint feature and electrical feature, into fault detection model, and according to state information in fault detection model whether to judge whether fault detection model needs to update;State evaluation coefficient is obtained according to state information calculation;If state evaluation coefficient is greater than or equal to preset state threshold, then normal output fault result;If state evaluation coefficient is less than preset state threshold, then update fault detection model, and obtain fault result after the fault detection model of updating.The application determines the fault of substation equipment from multidimensional angle, improves the accuracy of fault detection, and updates fault detection model, improves the new fault identification ability of substation.

Description

Technical Field

[0001] This invention belongs to the field of data processing technology, specifically relating to a method for monitoring the status of substation equipment based on voiceprint features. Background Technology

[0002] As substation equipment is continuously updated, the acoustic signals generated during its operation are collected and analyzed. Through a series of data processing methods such as acoustic data preprocessing, feature extraction, and pattern recognition, the acoustic information of the equipment is transformed into effective indicators that can be used for status assessment. This constructs a monitoring model that can accurately identify equipment anomalies, thereby achieving real-time and efficient monitoring of the operating status of substation equipment and providing data support for the operation of power facilities.

[0003] The prior art (publication number: CN119249854B) discloses a method and system for monitoring the operating status of substation equipment based on digital twins. This method acquires substation detection data, analyzes the data, and filters out feature data. First, a 3D digital model of the substation is established based on its actual physical objects and 3D geometric data, and a feature database is constructed. Then, feature matching is performed on the 3D digital model based on the substation's operating status and real-time parameters to form a digital twin of the substation. Next, the operating data in the digital twin is analyzed to output fault feature data. Fault analysis is then performed on the substation equipment based on the fault feature data. By modeling and simulating the entire substation data, a digital twin of the substation is established to simulate the substation's operating environment, enabling real-time monitoring of substation operating information and data, and significantly improving the reliability and real-time performance of substation fault handling.

[0004] The aforementioned patent uses digital twins to simulate substation scenarios, improving the real-time performance of substation fault handling. However, in practical applications, the core of the digital twin is the intelligent model, which cannot identify faults when new samples are not available. Therefore, the model becomes unstable and is prone to misjudging unknown abnormal faults, reducing the accuracy of substation fault identification. Summary of the Invention

[0005] The purpose of this invention is to solve the problem that the above-mentioned model cannot identify faults when no new samples appear, which leads to model instability, easy misjudgment of unknown abnormal faults, and reduced accuracy of substation fault identification. Therefore, this invention proposes a substation equipment status monitoring method based on voiceprint features.

[0006] In a first aspect of this invention, a method for monitoring the condition of substation equipment based on voiceprint features is first proposed, the method comprising: Acquire acoustic and electrical signals from substation equipment; The voiceprint signal and electrical signal are extracted separately to obtain voiceprint features and electrical features; The fused features are obtained by fusing voiceprint features and electrical features; The fused features are substituted into the fault detection model, and the fault detection model is judged whether it needs to be updated based on the state information in the fault detection model; the state evaluation coefficient is calculated based on the state information. If the state evaluation coefficient is greater than or equal to the preset state threshold, the fused features are substituted into the fault detection model to output the fault result normally. If the state evaluation coefficient is less than the preset state threshold, the fault detection model is updated, and the fused features are substituted into the updated fault detection model to obtain the fault result.

[0007] Optionally, the step of extracting features from the voiceprint signal and the electrical signal to obtain voiceprint features and electrical features respectively includes: The first sampled signal is obtained by resampling the voiceprint signal; The first sampled signal is segmented to obtain multiple sample frames, and the multiple sample frames are sequentially divided to obtain multiple frame signals; the multiple frame signals are pre-emphasized to obtain multiple high-frequency signals; Voiceprint features are obtained by frequency domain feature extraction from multiple high-frequency signals; The electrical signal is preprocessed to obtain the first electrical signal; Electrical features are obtained by extracting frequency domain features from the first electrical signal; The preprocessing specifically includes: Extract an integer number of fundamental frequency cycles from the electrical signal. If the length of the electrical signal is not an integer number of fundamental frequency cycles, extend it to the nearest number by padding with zeros. length; The intercepted signal is input into a leakage suppression function to obtain a first electrical signal; the leakage suppression function includes: ,in, This represents a time-domain smoothed signal, where N represents the total length of the truncated signal, and n represents the total number of sampling points, with values ​​ranging from [0, N-1].

[0008] Optionally, the feature fusion of voiceprint features and electrical features to obtain fused features includes: The first feature and the second feature are obtained by performing 1×1 convolutional layers on the voiceprint features and electrical features respectively; Apply the HardSwish activation function to the first feature to obtain the activated feature; The activation features are then subjected to global average pooling to obtain the global average features; After performing a fully connected layer on the global average features, the HardSwish activation function is then applied to obtain the feature weights. The target feature is obtained by multiplying the activation feature and feature weights element by element. The target feature and the second feature are concatenated to obtain the concatenated feature, and the concatenated feature is transposed to obtain the transposed feature; the concatenated feature and the transposed feature are multiplied by an outer product to obtain the feature matrix; The feature matrix is ​​transformed into fused features.

[0009] Optionally, the calculation of the state evaluation coefficient based on the state information includes: The status information includes the percentage of time lost and the CPU computation ratio. The process involves obtaining the time required for the fault detection model to process fused features; obtaining the CPU computation required for the fault detection model to process fused features; calculating the percentage of time lost by comparing the time lost with a preset time lost; calculating the CPU computation ratio by comparing the CPU computation with a preset CPU computation; and weighted summing the percentage of time lost and the CPU computation ratio to obtain the state evaluation coefficient.

[0010] Optionally, the update process of the fault detection model includes: Step 1: Obtain multiple historical fusion feature data, derive multiple feature samples based on the fusion features, and analyze the multiple historical fusion feature data to obtain the corresponding state types; the state types include normal type and known fault type; after extracting features from the multiple historical fusion feature data and multiple feature samples, construct a fusion training set; randomly select a fusion feature from the fusion training set, analyze the fusion feature to obtain the number of fault categories, randomly select k samples from each fault category to form the support set of the fusion feature; then randomly select q samples from each fault category to form the query set of the fusion feature, where k is less than q; Step 2: Copy the fault detection model to obtain a new feature model; input the support set into the new feature model and calculate the first loss value; obtain the parameters of the new feature model through the first update; The first update formula is: ,in, The parameters represent the new feature model. The parameters represent the fault detection model. This indicates that the first loss function is applied to the parameters. The gradient of Cz1_i() represents the first loss function for the i-th fused feature; It is the fault detection model definition function for the i-th fused feature.

[0011] Step 3: Input the query set into the new feature model and calculate the second loss value; obtain the optimized parameters of the fault detection model through the second update; The second update formula is: ;in, This represents the parameters after the fault detection model has been optimized. The second loss function represents the parameters. The gradient of Cz2_i; Cz2_i() represents the second loss function for the i-th fused feature; It is the new feature model definition function for the i-th fused feature; Step 4: Repeat steps 1 to 3 until the parameter values ​​of the fault detection model converge.

[0012] In a second aspect of this invention, a substation equipment condition monitoring system based on voiceprint features is proposed, the system comprising: Data acquisition module: Acquires acoustic and electrical signals from substation equipment; Feature extraction module: Extracts features from the voiceprint signal and electrical signal respectively to obtain voiceprint features and electrical features; Fusion module: Combines voiceprint features and electrical features to obtain fused features; Fault detection module: Substitutes the fused features into the fault detection model and determines whether the fault detection model needs to be updated based on the state information in the fault detection model; calculates the state evaluation coefficient based on the state information; if the state evaluation coefficient is greater than or equal to the preset state threshold, the fused features are substituted into the fault detection model to output the fault result normally; if the state evaluation coefficient is less than the preset state threshold, the fault detection model is updated, and the fused features are substituted into the updated fault detection model to obtain the fault result.

[0013] Optionally, the feature extraction module includes: a resampling module, a high-frequency signal module, a voiceprint feature module, and an electrical feature module. The resampling module is used to resample the voiceprint signal to obtain a first sampled signal. The high-frequency signal module is used to segment the first sampled signal to obtain multiple sample frames, divide the multiple sample frames sequentially to obtain multiple frame signals, and pre-emphasize the multiple frame signals to obtain multiple high-frequency signals. The voiceprint feature module is used to extract voiceprint features from multiple high-frequency signals in the frequency domain. The electrical feature module is used to preprocess the electrical signal to obtain a first electrical signal; and to extract electrical features from the frequency domain features of the first electrical signal. The preprocessing specifically includes: Extract an integer number of fundamental frequency cycles from the electrical signal. If the length of the electrical signal is not an integer number of fundamental frequency cycles, extend it to the nearest number by padding with zeros. length; The intercepted signal is input into a leakage suppression function to obtain a first electrical signal; the leakage suppression function includes: ,in, This represents a time-domain smoothed signal, where N represents the total length of the truncated signal, and n represents the total number of sampling points, with values ​​ranging from [0, N-1].

[0014] Optionally, the fusion module is further configured to fuse voiceprint features and electrical features to obtain fused features, including: The method of fusing voiceprint features and electrical features to obtain fused features includes: The first feature and the second feature are obtained by performing 1×1 convolutional layers on the voiceprint features and electrical features respectively; Apply the HardSwish activation function to the first feature to obtain the activated feature; The activation features are then subjected to global average pooling to obtain the global average features; After performing a fully connected layer on the global average features, the HardSwish activation function is then applied to obtain the feature weights. The target feature is obtained by multiplying the activation feature and feature weights element by element. The target feature and the second feature are concatenated to obtain the concatenated feature, and the concatenated feature is transposed to obtain the transposed feature; the concatenated feature and the transposed feature are multiplied by an outer product to obtain the feature matrix; The feature matrix is ​​transformed into fused features.

[0015] Optionally, the fault detection module includes: a status information acquisition module and a status evaluation coefficient module. The status information acquisition module is used to obtain the status information, which includes the percentage of time lost and the CPU computation ratio. The state evaluation coefficient module is used to obtain the time required by the fault detection model to process the fused features; obtain the CPU computation amount required by the fault detection model to process the fused features; calculate the proportion of time lost by the ratio of the time lost to the preset time lost; calculate the CPU computation amount ratio by the ratio of the CPU computation amount to the preset CPU computation amount; and obtain the state evaluation coefficient by weighted summation of the proportion of time lost and the CPU computation amount ratio.

[0016] Optionally, the update process of the fault detection model includes: Step 1: Obtain multiple historical fusion feature data, derive multiple feature samples based on the fusion features, and analyze the multiple historical fusion feature data to obtain the corresponding state types; the state types include normal type and known fault type; after extracting features from the multiple historical fusion feature data and multiple feature samples, construct a fusion training set; randomly select a fusion feature from the fusion training set, analyze the fusion feature to obtain the number of fault categories, randomly select k samples from each fault category to form the support set of the fusion feature; then randomly select q samples from each fault category to form the query set of the fusion feature, where k is less than q; Step 2: Copy the fault detection model to obtain a new feature model; input the support set into the new feature model and calculate the first loss value; obtain the parameters of the new feature model through the first update; The first update formula is: ,in, The parameters represent the new feature model. The parameters represent the fault detection model. This indicates that the first loss function is applied to the parameters. The gradient of Cz1_i() represents the first loss function for the i-th fused feature; It is the fault detection model definition function for the i-th fused feature.

[0017] Step 3: Input the query set into the new feature model and calculate the second loss value; obtain the optimized parameters of the fault detection model through the second update; The second update formula is: ;in, This represents the parameters after the fault detection model has been optimized. The second loss function represents the parameters. The gradient of Cz2_i; Cz2_i() represents the second loss function for the i-th fused feature; It is the new feature model definition function for the i-th fused feature; Step 4: Repeat steps 1 to 3 until the parameter values ​​of the fault detection model converge.

[0018] The beneficial effects of this invention are: This invention proposes a substation equipment status monitoring method based on voiceprint features. By fusing voiceprint signals and electrical signals, it comprehensively judges substation equipment faults from a multi-dimensional perspective, effectively improving the accuracy of fault detection. Simultaneously, it intelligently identifies new samples based on equipment status information and, through a dynamic update mechanism for the fault detection model, significantly reduces the model's computational overhead and greatly enhances the system's ability to identify and adapt to new fault types in substations, ensuring that the fault detection system always maintains a highly efficient and accurate operating state. Attached Figure Description

[0019] The invention will now be further described with reference to the accompanying drawings.

[0020] Figure 1 A flowchart illustrating a substation equipment status monitoring method based on voiceprint features, provided as an embodiment of the present invention; Figure 2 This is a framework diagram of a substation equipment condition monitoring system based on voiceprint features, provided for an embodiment of the present invention. Detailed Implementation

[0021] 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. The term "and / or" in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and B can represent: A alone, A and B simultaneously, and B alone. Furthermore, descriptions involving "first," "second," etc., in this invention are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, features defined with "first" or "second" can explicitly or implicitly include at least one of those features. Additionally, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0022] 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.

[0023] This invention provides a method for monitoring the condition of substation equipment based on voiceprint features. See also... Figure 1 , Figure 1 A flowchart illustrating a substation equipment condition monitoring method based on voiceprint features, provided as an embodiment of the present invention. The method includes the following steps: Acquire acoustic and electrical signals from substation equipment; The voiceprint signal and electrical signal are extracted separately to obtain voiceprint features and electrical features; The fused features are obtained by fusing voiceprint features and electrical features; The fused features are substituted into the fault detection model, and the state information in the fault detection model is used to determine whether the fault detection model needs to be updated; the state evaluation coefficient is calculated based on the state information. If the state evaluation coefficient is greater than or equal to the preset state threshold, the fused features are substituted into the fault detection model to output the fault result normally. If the state evaluation coefficient is less than the preset state threshold, the fault detection model is updated, and the fused features are substituted into the updated fault detection model to obtain the fault result.

[0024] The present invention provides a method for monitoring the status of substation equipment based on voiceprint features. By fusing voiceprint signals and electrical signals, the method determines the faults of substation equipment from multiple dimensions, thereby improving the accuracy of fault detection. The method also judges new samples based on status information and updates the fault detection model, reducing the computational load of the model and improving the ability to identify new faults in substations.

[0025] In one implementation, the fault detection model can be a combination of a deep neural network (DNN) and an attention mechanism model. The DNN and attention mechanism model take the fused feature dimension as input, containing 3 fully connected layers (256, 128, and 64 neurons) and 4 attention layers. The output layer corresponds to the fault category. The model uses cross-entropy + L2 regularization loss, Adam optimizer, and training parameters such as 32 batches and 100 rounds.

[0026] In one implementation, feature extraction is performed on the voiceprint signal and the electrical signal to obtain voiceprint features and electrical features, respectively, including: The first sampled signal is obtained by resampling the voiceprint signal; The first sampled signal is segmented to obtain multiple sample frames, and the multiple sample frames are sequentially divided to obtain multiple frame signals; the multiple frame signals are pre-emphasized to obtain multiple high-frequency signals; Voiceprint features are obtained by frequency domain feature extraction from multiple high-frequency signals; The electrical signal is preprocessed to obtain the first electrical signal; Electrical features are obtained by extracting frequency domain features from the first electrical signal; Preprocessing specifically includes: Extract an integer number of fundamental frequency cycles from the electrical signal. If the length of the electrical signal is not an integer number of fundamental frequency cycles, extend it to the nearest number by padding with zeros. length; The intercepted signal is input into a leakage suppression function to obtain the first electrical signal; the leakage suppression function includes: ,in, This represents a time-domain smoothed signal, where N represents the total length of the truncated signal, and n represents the total number of sampling points, with values ​​ranging from [0, N-1].

[0027] In one implementation, because different voiceprint signals may have different sampling rates (e.g., three datasets may have sampling rates of 43.1Hz or 90Hz), all signals need to be resampled to a uniform frequency to ensure consistent model input dimensions and avoid feature corruption due to sampling rate differences. Multiple high-frequency signals are cut into fixed-length short frames (e.g., 2.5 seconds / segment), with each segment serving as an independent input sample. Simultaneously, to avoid inter-frame information loss, a frame shift (e.g., 10 milliseconds) is set between adjacent frames to ensure the continuity of signal features. Multiple high-frequency signals are obtained by high-frequency enhancement using a first-order high-pass filter. The specific implementation of frequency domain feature extraction is as follows: first, the multiple high-frequency signals are subjected to a short-time Fourier transform to obtain the amplitude spectrum; then, a Mel filter bank (typically 40-128 filters) is constructed; finally, the amplitude spectrum is used to extract voiceprint features.

[0028] In one implementation, the preprocessing operation introduces a time-domain smoothing suppression leakage function (cosine window function) to further weaken the impact of spectral leakage, reduce signal energy dispersion, and enhance the accuracy of feature extraction. The first electrical signal obtained after preprocessing improves the recognizability of frequency domain features (such as harmonics, amplitude, etc.), providing basic support for subsequent fault diagnosis and effectively reducing the risk of misjudgment.

[0029] In one implementation, the fused features obtained by fusing voiceprint features and electrical features include: The first feature and the second feature are obtained by performing 1×1 convolutional layers on the voiceprint features and electrical features respectively; Apply the HardSwish activation function to the first feature to obtain the activated feature; Global average pooling is performed on the activation features to obtain the global average features; After performing a fully connected layer on the global average features, the HardSwish activation function is then applied to obtain the feature weights. The target feature is obtained by multiplying the activation feature and feature weights element by element. The target feature and the second feature are concatenated to obtain the concatenated feature. The concatenated feature is then transposed to obtain the transposed feature. The concatenated feature and the transposed feature are then multiplied by an outer product to obtain the feature matrix. The feature matrix is ​​transformed into fused features.

[0030] In one implementation, the traditional Swish function suffers from high computational cost due to its exponential operations (Sigmoid component). HardSwish, however, approximates Swish using a piecewise linear function, avoiding complex operations like exponentiation and division. It can be optimized through integer operations, reducing computational cost without significantly sacrificing performance. HardSwish's advantage lies in improving network accuracy. It concatenates the enhanced target feature (voiceprint) with the second feature (electrical signature) to construct a concatenated feature containing complete information from both modalities, achieving initial integration of heterogeneous features. Transposed outer product feature interaction: Transposing the concatenated feature yields a transposed feature. The outer product of the concatenated and transposed features constructs a high-dimensional feature matrix, strengthening the interaction between the two modal features. This makes the fusion of the two types of features more reasonable.

[0031] In one implementation, the state evaluation coefficients calculated based on the state information include: Status information includes the percentage of time lost and the percentage of CPU computation. The process involves obtaining the time required for the fault detection model to process fused features; obtaining the CPU computation required for the fault detection model to process fused features; calculating the percentage of time lost by comparing the time lost with a preset time lost; calculating the CPU computation ratio by comparing the CPU computation with a preset CPU computation; and weighted summing the percentage of time lost and the CPU computation ratio to obtain the state evaluation coefficient.

[0032] Specifically, the state evaluation coefficient is a state indicator used to measure the state information of the fused feature input into the model; the preset state threshold is obtained from statistical analysis of historical state information in the database.

[0033] In one implementation, the state evaluation coefficient is obtained by weighted summation of the proportion of time lost and the ratio of CPU computation, thus constructing a quantitative and comprehensive standard for judging the model state, avoiding the one-sidedness of a single indicator and improving the accuracy of state recognition. Secondly, by calculating the average time lost and the average CPU computation of new samples and comparing them with preset values ​​to obtain the ratio, the state evaluation becomes more targeted and objective, effectively avoiding subjective judgment errors.

[0034] In one implementation, the update process of the fault detection model includes: Step 1: Obtain multiple historical fusion feature data, derive multiple feature samples based on the fusion features, and analyze the multiple historical fusion feature data to obtain the corresponding state types; the state types include normal type and known fault type; after extracting features from multiple historical fusion feature data and multiple feature samples, construct a fusion training set; randomly select a fusion feature from the fusion training set, analyze the fusion feature to obtain the number of fault categories, randomly select k samples from each fault category to form the support set of the fusion feature; then randomly select q samples from each fault category to form the query set of the fusion feature, where k is less than q; Step 2: Copy the fault detection model to obtain a new feature model; input the support set into the new feature model and calculate the first loss value; obtain the parameters of the new feature model through the first update; The first update formula is: ,in, The parameters represent the new feature model. The parameters represent the fault detection model. This indicates that the first loss function is applied to the parameters. The gradient of Cz1_i() represents the first loss function for the i-th fused feature; It is the fault detection model definition function for the i-th fused feature.

[0035] Step 3: Input the query set into the new feature model and calculate the second loss value; obtain the optimized parameters of the fault detection model through the second update; The second updated formula is: ;in, This represents the parameters after the fault detection model has been optimized. The second loss function represents the parameters. The gradient of Cz2_i; Cz2_i() represents the second loss function for the i-th fused feature; It is the new feature model definition function for the i-th fused feature; Step 4: Repeat steps 1 to 3 until the parameter values ​​of the fault detection model converge.

[0036] In one implementation, the fault type is known (e.g., mechanical loosening, partial discharge, winding deformation, etc.); multiple feature samples are derived based on the fusion characteristics, wherein the derivation method includes: first determining the fluctuation amplitude (which can be set as a fixed proportion or fixed value of the fusion feature value), performing "amplitude increase" and "amplitude decrease" operations on each fusion feature value respectively to generate positive fluctuation samples and negative fluctuation samples; if more samples are needed, multiple sets of different small amplitudes (e.g., 1%, 2%, 3%) can be set, and the above addition and subtraction operations can be repeated to obtain multiple sets of fixed fluctuation feature samples.

[0037] The specific formula for the fault detection model definition function is as follows: Among them, represents the fault detection model definition function, i represents the i-th fusion feature, , and are the weights of each fully connected layer, , and are the bias terms, () represents the activation function, Attn() represents the multi-head attention layer, () represents the prediction probability distribution function; the specific formula of the new feature model definition function is consistent with the structure of the fault detection model definition function.

[0038] In one implementation, the specific process of the parameter values of the fault detection model converging is that when the parameter values tend to a certain stable value and remain unchanged, it is considered convergence; both the first loss function and the second loss function are derived by the staff based on the model operation experience.

[0039] In one implementation, by dividing the support set and the query set and limiting the sample ratio of k < q, not only can the new feature model parameters be quickly initialized using a small amount of support set, but also more query sets can be used for precise optimization, realizing efficient training with small samples, greatly reducing the dependence on large-scale labeled data, and improving the utilization rate of historical fusion feature data. The "two-update" mechanism is adopted. First, the new feature model adapted to the current fusion feature is obtained by fine-tuning with the support set, and then the original model parameters are optimized backward based on the query set, enabling the model to dynamically adapt to the feature distributions of different fault types and improving the recognition flexibility of known faults. Through multiple rounds of iteration until the parameters converge, the loss function is continuously optimized, the prediction deviation of the model is reduced, the distinction degree of fault categories is enhanced, the misjudgment rate and omission rate of known faults are effectively reduced, and at the same time, a stable model foundation is laid for subsequent unknown fault detection, significantly improving the reliability and accuracy of fault detection.

[0040] Based on the same inventive concept, the embodiment of the present invention also provides a method for monitoring the state of substation equipment based on voiceprint features. Refer to Figure 2 , Figure 2 is the framework diagram of a method for monitoring the state of substation equipment based on voiceprint features provided by the embodiment of the present invention, including: Data acquisition module: acquiring the voiceprint signal and electrical signal of the substation equipment; Feature extraction module: respectively extracting features from the voiceprint signal and electrical signal to obtain voiceprint features and electrical features; Fusion module: fusing the voiceprint features and electrical features to obtain fusion features; Fault Detection Module: Substitutes the fused features into the fault detection model and determines whether the fault detection model needs to be updated based on the state information in the fault detection model; calculates the state evaluation coefficient based on the state information; if the state evaluation coefficient is greater than or equal to the preset state threshold, the fused features are substituted into the fault detection model and the fault result is output normally; if the state evaluation coefficient is less than the preset state threshold, the fault detection model is updated, and the fused features are substituted into the updated fault detection model to obtain the fault result.

[0041] The substation equipment status monitoring system based on voiceprint features provided in this embodiment of the invention fuses voiceprint signals and electrical signals to determine faults in substation equipment from multiple dimensions, thereby improving the accuracy of fault detection. It also judges new samples based on status information and updates the fault detection model, reducing the computational load of the model and improving the ability to identify new faults in substations.

[0042] The foregoing has described one embodiment of the present invention in detail, but this content is merely a preferred embodiment and should not be considered as limiting the scope of the present invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the scope of the claims of this invention.

Claims

1. A method for monitoring the condition of substation equipment based on voiceprint features, characterized in that, The method includes: Acquire acoustic and electrical signals from substation equipment; The voiceprint signal and electrical signal are extracted separately to obtain voiceprint features and electrical features; The fused features are obtained by fusing voiceprint features and electrical features; The fused features are substituted into the fault detection model, and the fault detection model is judged whether it needs to be updated based on the state information in the fault detection model; the state evaluation coefficient is calculated based on the state information. If the state evaluation coefficient is greater than or equal to the preset state threshold, the fused features are substituted into the fault detection model to output the fault result normally. If the state evaluation coefficient is less than the preset state threshold, the fault detection model is updated, and the fused features are substituted into the updated fault detection model to obtain the fault result.

2. The method for monitoring the condition of substation equipment based on voiceprint features according to claim 1, characterized in that, The step of extracting features from the voiceprint signal and the electrical signal to obtain voiceprint features and electrical features includes: The first sampled signal is obtained by resampling the voiceprint signal; The first sampled signal is segmented to obtain multiple sample frames, and the multiple sample frames are sequentially divided to obtain multiple frame signals; the multiple frame signals are pre-emphasized to obtain multiple high-frequency signals; Voiceprint features are obtained by frequency domain feature extraction from multiple high-frequency signals; The electrical signal is preprocessed to obtain the first electrical signal; Electrical features are obtained by extracting frequency domain features from the first electrical signal; The preprocessing specifically includes: Extract an integer number of fundamental frequency cycles from the electrical signal. If the length of the electrical signal is not an integer number of fundamental frequency cycles, extend it to the nearest number by padding with zeros. length; The intercepted signal is input into a leakage suppression function to obtain a first electrical signal; the leakage suppression function includes: ,in, This represents a time-domain smoothed signal, where N represents the total length of the truncated signal, and n represents the total number of sampling points, with values ​​ranging from [0, N-1].

3. The method for monitoring the condition of substation equipment based on voiceprint features according to claim 1, characterized in that, The method of fusing voiceprint features and electrical features to obtain fused features includes: The first feature and the second feature are obtained by performing 1×1 convolutional layers on the voiceprint features and electrical features respectively; Apply the HardSwish activation function to the first feature to obtain the activated feature; The activation features are then subjected to global average pooling to obtain the global average features; After performing a fully connected layer on the global average features, the HardSwish activation function is then applied to obtain the feature weights. The target feature is obtained by multiplying the activation feature and feature weights element by element. The target feature and the second feature are concatenated to obtain the concatenated feature, and the concatenated feature is transposed to obtain the transposed feature; the concatenated feature and the transposed feature are multiplied by an outer product to obtain the feature matrix; The feature matrix is ​​transformed into fused features.

4. The method for monitoring the condition of substation equipment based on voiceprint features according to claim 1, characterized in that, The state evaluation coefficient calculated based on the state information includes: The status information includes the percentage of time lost and the CPU computation ratio. The process involves obtaining the time required for the fault detection model to process fused features; obtaining the CPU computation required for the fault detection model to process fused features; calculating the percentage of time lost by comparing the time lost with a preset time lost; calculating the CPU computation ratio by comparing the CPU computation with a preset CPU computation; and weighted summing the percentage of time lost and the CPU computation ratio to obtain the state evaluation coefficient.

5. The method for monitoring the condition of substation equipment based on voiceprint features according to claim 1, characterized in that, The update process of the fault detection model includes: Step 1: Obtain multiple historical fusion feature data, derive multiple feature samples based on the fusion features, and analyze the multiple historical fusion feature data to obtain the corresponding state types; the state types include normal type and known fault type; after extracting features from the multiple historical fusion feature data and multiple feature samples, construct a fusion training set; randomly select a fusion feature from the fusion training set, analyze the fusion feature to obtain the number of fault categories, randomly select k samples from each fault category to form the support set of the fusion feature; then randomly select q samples from each fault category to form the query set of the fusion feature, where k is less than q; Step 2: Copy the fault detection model to obtain a new feature model; input the support set into the new feature model and calculate the first loss value; obtain the parameters of the new feature model through the first update; The first update formula is: ,in, The parameters represent the new feature model. The parameters represent the fault detection model. This indicates that the first loss function is applied to the parameters. The gradient of Cz1_i() represents the first loss function for the i-th fused feature; It is the fault detection model definition function for the i-th fused feature; Step 3: Input the query set into the new feature model and calculate the second loss value; obtain the optimized parameters of the fault detection model through the second update; The second update formula is: ;in, This represents the parameters after the fault detection model has been optimized. The second loss function represents the parameters. The gradient of Cz2_i; Cz2_i() represents the second loss function for the i-th fused feature; It is the new feature model definition function for the i-th fused feature; Step 4: Repeat steps 1 to 3 until the parameter values ​​of the fault detection model converge.

6. A substation equipment condition monitoring system based on voiceprint features, characterized in that, The system includes: Data acquisition module: Acquires acoustic and electrical signals from substation equipment; Feature extraction module: Extracts features from the voiceprint signal and electrical signal respectively to obtain voiceprint features and electrical features; Fusion module: Combines voiceprint features and electrical features to obtain fused features; Fault detection module: Substitutes the fused features into the fault detection model and determines whether the fault detection model needs to be updated based on the state information in the fault detection model; calculates the state evaluation coefficient based on the state information; if the state evaluation coefficient is greater than or equal to the preset state threshold, the fused features are substituted into the fault detection model and the fault result is output normally; if the state evaluation coefficient is less than the preset state threshold, the fault detection model is updated, and the fused features are substituted into the updated fault detection model to obtain the fault result.

7. A substation equipment condition monitoring system based on voiceprint features according to claim 6, characterized in that, The feature extraction module includes: a resampling module, a high-frequency signal module, a voiceprint feature module, and an electrical feature module. The resampling module is used to resample the voiceprint signal to obtain a first sampled signal. The high-frequency signal module is used to segment the first sampled signal to obtain multiple sample frames, divide the multiple sample frames sequentially to obtain multiple frame signals, and pre-emphasize the multiple frame signals to obtain multiple high-frequency signals. The voiceprint feature module is used to extract voiceprint features from multiple high-frequency signals in the frequency domain. The electrical feature module is used to preprocess the electrical signal to obtain a first electrical signal; and to extract electrical features from the frequency domain features of the first electrical signal. The preprocessing specifically includes: Extract an integer number of fundamental frequency cycles from the electrical signal. If the length of the electrical signal is not an integer number of fundamental frequency cycles, extend it to the nearest number by padding with zeros. length; The intercepted signal is input into a leakage suppression function to obtain a first electrical signal; the leakage suppression function includes: ,in, This represents a time-domain smoothed signal, where N represents the total length of the truncated signal, and n represents the total number of sampling points, with values ​​ranging from [0, N-1].

8. A substation equipment condition monitoring system based on voiceprint features according to claim 6, characterized in that, The fusion module is further configured to fuse voiceprint features and electrical features to obtain fused features, including: The method of fusing voiceprint features and electrical features to obtain fused features includes: The first feature and the second feature are obtained by performing 1×1 convolutional layers on the voiceprint features and electrical features respectively; Apply the HardSwish activation function to the first feature to obtain the activated feature; The activation features are then subjected to global average pooling to obtain the global average features; After performing a fully connected layer on the global average features, the HardSwish activation function is then applied to obtain the feature weights. The target feature is obtained by multiplying the activation feature and feature weights element by element. The target feature and the second feature are concatenated to obtain the concatenated feature, and the concatenated feature is transposed to obtain the transposed feature; the concatenated feature and the transposed feature are multiplied by an outer product to obtain the feature matrix; The feature matrix is ​​transformed into fused features.

9. A substation equipment condition monitoring system based on voiceprint features according to claim 6, characterized in that, The fault detection module includes: a status information acquisition module and a status evaluation coefficient module. The status information acquisition module is used to obtain the status information, which includes the percentage of time lost and the CPU computation ratio. The state evaluation coefficient module is used to obtain the time required by the fault detection model to process the fused features; obtain the CPU computation amount required by the fault detection model to process the fused features; calculate the proportion of time lost by the ratio of the time lost to the preset time lost; calculate the CPU computation amount ratio by the ratio of the CPU computation amount to the preset CPU computation amount; and obtain the state evaluation coefficient by weighted summation of the proportion of time lost and the CPU computation amount ratio.

10. A substation equipment condition monitoring system based on voiceprint features according to claim 6, characterized in that, The update process of the fault detection model includes: Step 1: Obtain multiple historical fusion feature data, derive multiple feature samples based on the fusion features, and analyze the multiple historical fusion feature data to obtain the corresponding state types; the state types include normal type and known fault type; after extracting features from the multiple historical fusion feature data and multiple feature samples, construct a fusion training set; randomly select a fusion feature from the fusion training set, analyze the fusion feature to obtain the number of fault categories, randomly select k samples from each fault category to form the support set of the fusion feature; then randomly select q samples from each fault category to form the query set of the fusion feature, where k is less than q; Step 2: Copy the fault detection model to obtain a new feature model; input the support set into the new feature model and calculate the first loss value; obtain the parameters of the new feature model through the first update; The first update formula is: ,in, The parameters represent the new feature model. The parameters represent the fault detection model. This indicates that the first loss function is applied to the parameters. The gradient of Cz1_i() represents the first loss function for the i-th fused feature; It is the fault detection model definition function for the i-th fused feature; Step 3: Input the query set into the new feature model and calculate the second loss value; obtain the optimized parameters of the fault detection model through the second update; The second update formula is: ;in, This represents the parameters after the fault detection model has been optimized. The second loss function represents the parameters. The gradient of Cz2_i; Cz2_i() represents the second loss function for the i-th fused feature; It is the new feature model definition function for the i-th fused feature; Step 4: Repeat steps 1 to 3 until the parameter values ​​of the fault detection model converge.

Images