Distribution network equipment on-line monitoring method, system, equipment and medium

Abstract

The invention discloses a distribution network equipment on-line monitoring method, system, equipment and medium, and relates to the technical field of distribution network equipment monitoring, and the method comprises the steps: collecting an audio signal and image data of distribution network equipment; performing dynamic differential semantic segmentation on the image data to extract image features, and performing anti-interference processing based on wavelet transform on the audio signal; voiceprint features are extracted from the audio signals subjected to anti-interference processing, wherein the voiceprint features comprise Mel-frequency cepstrum coefficients; carrying out hierarchical fusion on the image features and the voiceprint features, wherein the hierarchical fusion comprises splicing in a feature layer, and introducing an attention mechanism to a decision layer to dynamically distribute weights of different modal features; and based on the fused features, synchronously executing equipment defect identification and voiceprint forgery detection through a multi-task learning framework, and triggering an alarm according to an identification result. According to the invention, the problems of insufficient robustness, insufficient anti-interference capability and poor safety in the prior art are solved.

Description

Technical Field

[0001] This invention relates to the field of power distribution network equipment monitoring technology, and in particular to an online monitoring method, system, equipment and medium for power distribution network equipment. Background Technology

[0002] Against the backdrop of the rapid development of smart grids, the stable operation of equipment in the distribution network, as a crucial link connecting the main grid and users, is of paramount importance. Online monitoring of distribution network equipment is a core technological means to ensure grid security and achieve condition-based maintenance and intelligent operation and maintenance. With the expansion of power system scale and the surge in the number of devices, monitoring systems need to process massive amounts of heterogeneous data from various sensors (such as acoustic, optical, and temperature sensors). How to efficiently, accurately, and reliably integrate this multi-source information to achieve comprehensive perception and precise diagnosis of equipment status has become a key research focus and challenge in this field.

[0003] Currently, research has focused on improving monitoring capabilities through the fusion of multiple information sources. For example, the invention patent CN107390628A, entitled "Method and System for Monitoring and Early Warning of Distribution Network Status," proposes a method for fusing and analyzing data from multiple distribution network application systems. This method improves equipment management efficiency by associating monitoring information with equipment identifiers and generating alarms when equipment parameters exceed set thresholds. However, this method has the following main shortcomings: First, insufficient robustness and anti-interference capability: This method relies on traditional threshold judgment and relatively simple information fusion strategies. When facing the complex electromagnetic interference environment of the distribution network, signals such as sound patterns and current are easily affected by noise, leading to decreased monitoring accuracy and a high risk of false alarms or missed alarms. Second, poor security: This scheme typically assumes that the collected data is authentic and reliable, without fully considering malicious forgery or deception. This allows attackers to potentially bypass the audio analysis-based monitoring system by playing pre-recorded normal equipment sounds to forge a "normal operating" state, posing a potential risk to power grid security. Summary of the Invention

[0004] This invention provides a method, system, device, and medium for online monitoring of distribution network equipment, which addresses the problems of insufficient robustness and anti-interference ability, as well as poor security in existing technologies.

[0005] In view of the above, the first aspect of the present invention provides a method for online monitoring of distribution network equipment, the method comprising:

[0006] Collect audio signals and image data from power distribution network equipment;

[0007] The image data is subjected to dynamic differential semantic segmentation to extract image features, and the audio signal is subjected to anti-interference processing based on wavelet transform.

[0008] Voiceprint features are extracted from the audio signal after it has undergone anti-interference processing, the voiceprint features including Mel-spectral coefficients;

[0009] The image features and the voiceprint features are fused in a layered manner. The layered fusion includes splicing at the feature layer and introducing an attention mechanism at the decision layer to dynamically allocate the weights of different modal features.

[0010] Based on the fused features, the system simultaneously performs equipment defect identification and voiceprint forgery detection through a multi-task learning framework, and triggers an alarm based on the identification results.

[0011] Optionally, the dynamic differential semantic segmentation includes:

[0012] For key components pre-partitioned by the device, a PSPNet model based on pyramid pooling modules is used for high-resolution semantic segmentation.

[0013] For non-critical regions pre-divided by the device, a lightweight semantic segmentation model based on the MobileNetV2 encoder is used for segmentation.

[0014] Optionally, the anti-interference processing of the audio signal based on wavelet transform includes:

[0015] The audio signal is decomposed into signal components of multiple scales using discrete wavelet transform, including low-frequency and high-frequency components;

[0016] The high-frequency component obtained after decomposition is subjected to threshold processing to remove noise, resulting in a denoised audio signal.

[0017] The denoised audio signal is reconstructed using inverse wavelet transform.

[0018] Optionally, the extraction of voiceprint features from the interference-resistant audio signal, wherein the voiceprint features include Mel-spectral coefficients, including:

[0019] Mel-cepstral coefficients are extracted from the interference-resistant audio signal based on the Mel-cepstral coefficient formula.

[0020] The formula for the Mel-frequency cepstral coefficients is:

[0021] ;

[0022] In the formula, For Mel-frequency cepstral coefficients, It is a short-time Fourier transform. This is the Discrete Cosine Transform operation.

[0023] Optionally, the step of introducing an attention mechanism at the decision layer to dynamically allocate weights for different modal features includes:

[0024] The features fused from the feature layer are used as inputs to the query matrix, key matrix, and value matrix;

[0025] The attention weights of each modality feature are calculated based on the attention mechanism formula, and the features are then weighted and fused based on the attention weights of each modality feature.

[0026] The expression for the attention mechanism formula is as follows:

[0027] ;

[0028] In the formula, Q, K, and V are the query matrix, key matrix, and value matrix, respectively. This is the normalization factor.

[0029] Optionally, the optimization objective of the multi-task learning framework is to minimize the total loss function, and the expression for minimizing the total loss function is:

[0030]

[0031] In the formula, Losses due to equipment defect identification tasks. The loss due to the voiceprint forgery detection task and These are the weighting factors for the equipment defect identification task and the voiceprint forgery detection task, respectively;

[0032] in, ;

[0033] In the formula, Let i be the actual label of the i-th sample. The model predicts the label for the i-th sample. The total of the sample;

[0034] ;

[0035] In the formula, For actual labels, The predicted probability represents the probability that the voiceprint is fake.

[0036] Optionally, after performing the dynamic differential semantic segmentation, the method further includes:

[0037] Based on the results of dynamic differential semantic segmentation, high bitrates are assigned to pre-divided key component regions, and low bitrates are assigned to pre-divided non-key regions.

[0038] A second aspect of the present invention provides an online monitoring system for power distribution network equipment, the system comprising:

[0039] The acquisition unit is used to acquire audio signals and image data from the distribution network equipment;

[0040] The first extraction unit is used to perform dynamic differential semantic segmentation on the image data to extract image features, and to perform anti-interference processing on the audio signal based on wavelet transform.

[0041] The second extraction unit is used to extract voiceprint features from the audio signal that has undergone anti-interference processing, the voiceprint features including Mel-spectral coefficients;

[0042] The fusion unit is used to perform hierarchical fusion of the image features and the voiceprint features. The hierarchical fusion includes splicing at the feature layer and introducing an attention mechanism at the decision layer to dynamically allocate the weights of different modal features.

[0043] The monitoring unit is used to simultaneously perform equipment defect identification and voiceprint forgery detection based on the fused features through a multi-task learning framework, and trigger alarms according to the identification results.

[0044] A third aspect of the present invention provides an online monitoring device for power distribution network equipment, the device comprising a processor and a memory:

[0045] The memory is used to store program code and transmit the program code to the processor;

[0046] The processor is used to execute the steps of the online monitoring method for distribution network equipment as described in the first aspect above, according to the instructions in the program code.

[0047] A fourth aspect of the present invention provides a computer-readable storage medium for storing program code for executing the online monitoring method for distribution network equipment described in the first aspect above.

[0048] As can be seen from the above technical solutions, the present invention has the following advantages:

[0049] 1) Improve equipment monitoring accuracy: By combining voiceprint features (MFCC) and image features (semantic segmentation) for multimodal feature fusion, the audio and visual information of the equipment can be comprehensively analyzed, thereby providing a more accurate and comprehensive equipment status assessment.

[0050] 2) Enhanced robustness and anti-interference capability: This invention designs an anti-interference module based on wavelet transform, which effectively removes the influence of electromagnetic interference on the acoustic signal and improves the stability and reliability of equipment monitoring in complex environments.

[0051] 3) Intelligent anti-spoofing mechanism: Through a multi-task learning framework, the simultaneous processing of equipment defect identification and voiceprint forgery detection is achieved, which can effectively prevent the influence of forged voiceprints on equipment status judgment and improve the security of the monitoring system.

[0052] 4) Optimize the use of computing resources: Introduce dynamic differentiated semantic segmentation and priority encoding strategies to ensure high-precision processing of critical regions, while reducing the computational burden on non-critical regions through lightweight processing and real-time compression technology, thereby improving the system's real-time response capability and computing efficiency.

[0053] 5) Strong adaptability: By dynamically adjusting the semantic segmentation model and priority coding strategy, the system can flexibly adapt to different device types and working states, ensuring high monitoring accuracy under various working conditions. Attached Figure Description

[0054] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0055] Figure 1 A flowchart illustrating an online monitoring method for distribution network equipment provided in an embodiment of the present invention;

[0056] Figure 2 This is a schematic diagram of the structure of an online monitoring system for distribution network equipment provided in an embodiment of the present invention. Detailed Implementation

[0057] To make the objectives, features, and advantages of this invention more apparent and understandable, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described below are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0058] Please see Figure 1 An online monitoring method for distribution network equipment provided in this embodiment of the invention includes:

[0059] Step 101: Collect audio signals and image data from the distribution network equipment.

[0060] It should be noted that, firstly, the monitoring terminal collects audio signals and image data from the power distribution network equipment via sensors. The audio signals are used in subsequent steps to extract voiceprint features, and the image data is used in subsequent steps to extract the equipment's visual features. Through acoustic sensors and high-definition cameras, the system can acquire real-time information about the equipment's operating status.

[0061] Step 102: Perform dynamic differential semantic segmentation on the image data to extract image features, and perform anti-interference processing on the audio signal based on wavelet transform.

[0062] In one embodiment, the dynamic differential semantic segmentation in step 102 includes: for key components pre-divided by the device, high-resolution semantic segmentation is performed using a PSPNet model based on a pyramid pooling module; for non-key regions pre-divided by the device, segmentation is performed using a lightweight semantic segmentation model based on a MobileNetV2 encoder.

[0063] It should be noted that the goal of the multimodal feature fusion mechanism is to effectively combine the feature information of voiceprints and images. Since images are a crucial input source, the accuracy of their processing directly affects the fusion effect. To improve the processing efficiency and accuracy of image data, this invention proposes a dynamic differentiated semantic segmentation strategy based on device type, enabling targeted and efficient image analysis in device monitoring. The specific steps are as follows:

[0064] 1) High-resolution semantic segmentation:

[0065] For pre-defined key components of the equipment (such as contacts in switchgear and insulators in energy storage devices), high-resolution segmentation accuracy is required. A pyramid pooling module (PSPNet) is employed for high-resolution semantic segmentation. PSPNet extracts contextual information at different scales through multi-scale pooling, enabling better capture of image details. The pyramid pooling module in PSPNet captures information from different spatial scales through multi-scale pooling operations, and then fuses this information using the following fusion formula to improve image segmentation accuracy.

[0066] ;

[0067] In the formula, Representing pooling operations at different scales, concat concatenates features from different scales. It is the input feature map. It is the output fused feature map.

[0068] By using pyramid pooling, PSPNet can effectively capture global and local information in an image, making it possible to accurately segment the details of key components (such as device contacts, insulators, etc.).

[0069] 2) Lightweight semantic segmentation:

[0070] For pre-defined non-critical regions of the device (such as the external background), this invention employs a lightweight semantic segmentation method to improve computational efficiency. To reduce computational resource consumption, MobileNetV2 is used as the encoder for image segmentation. MobileNetV2 is a network architecture based on depthwise separable convolutions, which can significantly reduce computational resource consumption while maintaining high accuracy. MobileNetV2 further reduces computational complexity by introducing inverted residuals and a linear bottleneck.

[0071] ;

[0072] In the formula, ConvDW represents a depthwise separable convolution operation, Conv1x1 is a 1x1 convolution, and ReLU6 is the activation function. and These are the input and output feature maps, respectively. This structure enables MobileNetV2 to achieve image segmentation with low computational cost, making it suitable for processing non-critical regions.

[0073] Furthermore, in one embodiment, after dynamic differential semantic segmentation, the method further includes: allocating a high bitrate to the pre-divided key component regions and a low bitrate to the pre-divided non-key regions based on the result of dynamic differential semantic segmentation.

[0074] It should be noted that the image compression ratio is dynamically adjusted for both critical and non-critical areas of the device. The compression ratio can be adjusted based on the region's priority, resolution requirements, and computing resources. Understandably, high-priority areas receive lower compression to maintain higher data quality, while low-priority areas receive higher compression to reduce computational resource consumption.

[0075] ;

[0076] In the formula, It is the size of the original image data. This refers to the size of the compressed data.

[0077] In practical applications, device image data is processed according to a priority encoding strategy. Critical areas are allocated a higher bitrate, while non-critical areas are compressed using a lower bitrate. Through a smart chip, the system can process image data in real time and compress the image data stream, effectively reducing data transmission bandwidth requirements and improving system response speed.

[0078] In one embodiment, the anti-interference processing of the audio signal based on wavelet transform in step 102 includes: decomposing the audio signal into signal components of multiple scales using discrete wavelet transform, including: low-frequency components and high-frequency components; performing threshold processing on the high-frequency components obtained after decomposition to remove noise, thereby obtaining a denoised audio signal; and reconstructing the denoised audio signal using inverse wavelet transform.

[0079] It should be noted that in the multimodal feature fusion mechanism, the Mel-frequency cepstral coefficients (MFCC) of the acoustic print and the semantic segmentation features of the image are fused to provide comprehensive equipment status information. However, in electromagnetic interference environments, a single acoustic print feature may be affected by noise, leading to reduced monitoring accuracy. To improve the reliability and accuracy of the power distribution network equipment monitoring system in complex electromagnetic environments, this invention designs an acoustic print denoising algorithm based on wavelet transform and dynamically selects the acoustic print frequency band by combining the image semantic segmentation results. Electromagnetic interference (EMI) is common in field environments, especially near high-voltage power equipment. The interference of electromagnetic waves on the acoustic print signal can seriously affect the accuracy of equipment status monitoring. Therefore, developing an effective anti-interference technology is crucial for improving the robustness of the system. This invention uses wavelet transform to denoise the acoustic print signal. Wavelet transform can decompose the signal in a multi-scale manner, processing signal components of different frequencies independently, thereby removing noise while preserving the main features of the signal. The specific denoising process is as follows:

[0080] 1) Wavelet Transform: First, the original acoustic signature signal is decomposed into signal components of multiple scales using Discrete Wavelet Transform (DWT). These components contain information about the signal in different frequency ranges and can be divided into low-frequency components (representing the overall trend of the signal) and high-frequency components (representing the detailed parts of the signal).

[0081] ;

[0082] It is the input signal. It is a wavelet basis function. These are wavelet transform coefficients.

[0083] 2) Denoising: Thresholding is performed on the high-frequency components after wavelet transform, removing coefficients below a certain threshold to denoise the signal. This step effectively suppresses noise signals.

[0084] ;

[0085] These are the denoised coefficients, and λ is the threshold. The symbol representing the coefficient.

[0086] 3) Signal reconstruction: The denoised signal components at each scale are reconstructed into denoised voiceprint signals by inverse wavelet transform (IDWT).

[0087] ;

[0088] This process ensures that the removed noise does not affect the core features of the signal, thereby effectively improving the signal quality.

[0089] Step 103: Extract voiceprint features from the anti-interference processed audio signal. Voiceprint features include Mel-frequency cepstral coefficients.

[0090] In one embodiment, step 103 includes: extracting Mel-Cepstral Coefficients from the interference-resistant audio signal based on the Mel-Cepstral Coefficients formula.

[0091] It should be noted that, since this invention requires hierarchical fusion of image features and voiceprint features (including feature-level fusion and decision-level fusion, see subsequent step 104), it is necessary to first extract Mel-spectral coefficients (MFCCs) from the voiceprint data as audio features during the feature-level fusion stage. Specifically, in the feature-level fusion stage, the MFCC features of the voiceprint are initially fused with the semantic segmentation features of the image. To achieve this goal, the voiceprint signal is first processed to extract its Mel-spectral coefficients (MFCCs):

[0092] ;

[0093] In the formula, It is a short-time Fourier transform. The discrete cosine transform operation ultimately yields the Mel frequency cepstral coefficients (MFCCs) of the audio signal.

[0094] Step 104: Perform layered fusion of image features and voiceprint features. Layered fusion includes splicing at the feature layer and introducing an attention mechanism at the decision layer to dynamically allocate the weights of different modal features.

[0095] It should be noted that after obtaining the image features and voiceprint features, feature-level concatenation is used to connect the voiceprint and image features for further learning and fusion. The concatenated features are then fed into the subsequent decision layer. In the decision layer fusion stage, the result of feature layer fusion is first input into a classifier (such as a fully connected layer or convolutional layer) for analysis and decision-making. To further optimize the decision-making process, this invention introduces an attention mechanism to dynamically allocate weights to different modal features. This mechanism can automatically learn which features are most important for equipment status monitoring based on the different properties of the actual data and assign them greater weight in the decision-making process.

[0096] In one embodiment, step 104 introduces an attention mechanism at the decision layer to dynamically allocate the weights of different modal features, including: using the features fused at the feature layer as input to the query matrix, key matrix, and value matrix; calculating the attention weights of each modal feature based on the attention mechanism formula; and performing weighted fusion of the features based on the attention weights of each modal feature.

[0097] It should be noted that the formula for the attention mechanism is as follows:

[0098] ;

[0099] In the formula, Q, K, and V are the query matrix, key matrix, and value matrix, respectively. This is the normalization factor.

[0100] This formula allows for the adaptive assignment of a weight to each input feature, emphasizing key features and thus improving the ability to identify important information.

[0101] Step 105: Based on the fused features, simultaneously perform equipment defect identification and voiceprint forgery detection through a multi-task learning framework, and trigger an alarm based on the identification results.

[0102] In one embodiment, the optimization objective of the multi-task learning framework in step 105 is to minimize the total loss function.

[0103] It should be noted that in the multimodal feature fusion mechanism, voiceprint features and image features are combined through a fusion strategy to improve the monitoring accuracy of device status. However, in practical applications, deceptive behaviors such as voiceprint spoofing exist, which may mislead the monitoring system's judgment. To address this challenge, this invention employs a multi-task learning framework to simultaneously perform device defect identification and voiceprint spoofing detection. By sharing features and optimizing the learning process, the system's ability to handle different tasks is enhanced.

[0104] In this invention, equipment defect identification and voiceprint spoofing detection are treated as two tasks, learned within a unified model. Specifically, for the equipment defect identification task, this invention uses a traditional classification task (such as determining whether a device is faulty through images or voiceprint signals); while for the voiceprint spoofing detection task, it determines whether the voiceprint signal is spoofed. Both share the underlying feature network, enabling the model to learn more general features and effectively combine equipment appearance and sound information, thereby improving the overall performance of the system.

[0105] In a multi-task learning network architecture, the underlying feature extraction module processes the input voiceprint signal and device image using a convolutional neural network (CNN) or other deep learning model to extract shared feature representations. These shared features are then fed into two task branches: one for device defect identification and the other for voiceprint forgery detection.

[0106] During training, the loss functions of the two tasks are weighted and fused to ensure that both tasks are optimized. The overall loss function can be expressed as:

[0107] ;

[0108] This is a loss in the equipment defect identification task. This is a loss for the voiceprint forgery detection task. and These are the weighting factors for the two tasks. By adjusting the values ​​of these two factors, the learning contributions of the two tasks can be balanced, ensuring that the optimization of the two tasks does not conflict.

[0109] For equipment defect identification tasks, the cross-entropy loss function is used for classification optimization:

[0110] ;

[0111] It is the actual label of the i-th sample. It is the model-predicted label (or probability) of the i-th sample. It represents the total number of samples.

[0112] For voiceprint forgery detection, a binary classification cross-entropy loss function is used to distinguish between forged and genuine voiceprints:

[0113] ;

[0114] In the formula, For actual labels, The predicted probability represents the probability that the voiceprint is fake.

[0115] Multi-task learning frameworks can effectively utilize common information across different tasks by sharing feature extraction modules. For example, equipment defects and voiceprint spoofing are often related to changes in equipment status and operational anomalies, and this information can be shared through the voiceprint and image feature extraction parts. By sharing underlying features, equipment defect identification and voiceprint spoofing detection not only improve their own accuracy but also complement each other, achieving better results in the optimization of multiple tasks.

[0116] In practical applications, the above processing steps are used to monitor the equipment status in real time and generate alarm information based on the analysis results. Specifically, if a device malfunction or spoofed voiceprint is detected, an alarm is immediately triggered and the device's status information is sent to the control center. Furthermore, the monitoring results are displayed in real time on a screen for maintenance personnel to facilitate timely decision-making and handling.

[0117] This invention provides an online monitoring method for distribution network equipment. First, it performs dynamic differentiated semantic segmentation and priority encoding on the image data of the distribution network equipment, including: dynamically adjusting the semantic segmentation model according to the equipment type, using high-resolution segmentation for key components, and using lightweight segmentation to optimize non-critical areas; and combining this with a priority encoding strategy for real-time image data stream compression to optimize computational resource utilization. Further, it employs wavelet transform for voiceprint denoising and dynamically selects the voiceprint frequency band based on the image semantic segmentation results, thereby effectively addressing on-site electromagnetic interference and improving system stability. Next, it extracts Mel-frequency cepstral coefficients (MFCCs) as voiceprint features from the anti-interference processed audio signal. Then, it combines the MFCCs of the voiceprint with the semantic segmentation features of the image, employing a hierarchical fusion strategy (feature layer + decision layer) and introducing an attention mechanism to dynamically allocate weights to improve the accuracy and robustness of equipment status monitoring. Finally, it integrates a multi-task learning framework to achieve simultaneous processing of equipment defect identification and voiceprint forgery detection, and triggers alarms based on the identification results, thereby improving the ability to identify deceptive behavior and the accuracy of monitoring.

[0118] The above is a method for online monitoring of distribution network equipment provided in the embodiments of the present invention. The following is a system for online monitoring of distribution network equipment provided in the embodiments of the present invention.

[0119] Please see Figure 2 An online monitoring system for distribution network equipment provided in this embodiment of the invention includes:

[0120] The acquisition unit 201 is used to acquire audio signals and image data from the distribution network equipment.

[0121] The first extraction unit 202 is used to perform dynamic differential semantic segmentation on image data to extract image features and to perform anti-interference processing on audio signals based on wavelet transform.

[0122] The second extraction unit 203 is used to extract voiceprint features from the audio signal that has undergone anti-interference processing. The voiceprint features include Mel-frequency cepstral coefficients.

[0123] The fusion unit 204 is used to perform hierarchical fusion of image features and voiceprint features. The hierarchical fusion includes splicing at the feature layer and introducing an attention mechanism at the decision layer to dynamically allocate the weights of different modal features.

[0124] The monitoring unit 205 is used to simultaneously perform equipment defect identification and voiceprint forgery detection based on the fused features through a multi-task learning framework, and trigger an alarm based on the identification results.

[0125] Furthermore, this embodiment of the invention also provides an online monitoring device for power distribution network equipment, the device including a processor and a memory:

[0126] The memory is used to store program code and transmit the program code to the processor;

[0127] The processor is used to execute the steps of the online monitoring method for distribution network equipment as described in the above method embodiments, according to the instructions in the program code.

[0128] Furthermore, this embodiment of the invention also provides a computer-readable storage medium for storing program code, which is used to execute the online monitoring method for distribution network equipment described in the above method embodiments.

[0129] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0130] In the embodiments provided by this invention, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, indirect coupling or communication connection between apparatuses or units, and may be electrical, mechanical, or other forms.

[0131] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0132] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0133] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0134] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for online monitoring of power distribution network equipment, characterized in that, include: Collect audio signals and image data from power distribution network equipment; The image data is subjected to dynamic differential semantic segmentation to extract image features, and the audio signal is subjected to anti-interference processing based on wavelet transform. Voiceprint features are extracted from the audio signal after it has undergone anti-interference processing, the voiceprint features including Mel-spectral coefficients; The image features and the voiceprint features are fused in a layered manner. The layered fusion includes splicing at the feature layer and introducing an attention mechanism at the decision layer to dynamically allocate the weights of different modal features. Based on the fused features, the system simultaneously performs equipment defect identification and voiceprint forgery detection through a multi-task learning framework, and triggers an alarm based on the identification results.

2. The online monitoring method for distribution network equipment according to claim 1, characterized in that, The dynamic differential semantic segmentation includes: For key components pre-partitioned by the device, a PSPNet model based on pyramid pooling modules is used for high-resolution semantic segmentation. For non-critical regions pre-divided by the device, a lightweight semantic segmentation model based on the MobileNetV2 encoder is used for segmentation.

3. The online monitoring method for distribution network equipment according to claim 1, characterized in that, The anti-interference processing of the audio signal based on wavelet transform includes: The audio signal is decomposed into signal components of multiple scales using discrete wavelet transform, including low-frequency and high-frequency components; The high-frequency component obtained after decomposition is subjected to threshold processing to remove noise, resulting in a denoised audio signal. The denoised audio signal is reconstructed using inverse wavelet transform.

4. The online monitoring method for distribution network equipment according to claim 1, characterized in that, The process of extracting voiceprint features from the interference-resistant audio signal, wherein the voiceprint features include Mel-spectral coefficients, including: Mel-cepstral coefficients are extracted from the interference-resistant audio signal based on the Mel-cepstral coefficient formula. The formula for the Mel-frequency cepstral coefficients is: ； In the formula, For Mel-frequency cepstral coefficients, It is a short-time Fourier transform. This is the Discrete Cosine Transform operation.

5. The online monitoring method for distribution network equipment according to claim 1, characterized in that, The introduction of an attention mechanism at the decision-making level to dynamically allocate weights for different modal features includes: The features fused from the feature layer are used as inputs to the query matrix, key matrix, and value matrix; The attention weights of each modality feature are calculated based on the attention mechanism formula, and the features are then weighted and fused based on the attention weights of each modality feature. The expression for the attention mechanism formula is as follows: ； In the formula, Q, K, and V are the query matrix, key matrix, and value matrix, respectively. This is the normalization factor.

6. The online monitoring method for distribution network equipment according to claim 1, characterized in that, The optimization objective of the multi-task learning framework is to minimize the total loss function, and the expression for minimizing the total loss function is: ； In the formula, Losses due to equipment defect identification tasks. The loss due to the voiceprint forgery detection task and These are the weighting factors for the equipment defect identification task and the voiceprint forgery detection task, respectively; in, ; In the formula, Let i be the actual label of the i-th sample. The model predicts the label for the i-th sample. The total number of samples; ； In the formula, For actual labels; The predicted probability represents the probability that the voiceprint is fake.

7. The online monitoring method for distribution network equipment according to claim 2, characterized in that, After performing the dynamic differential semantic segmentation, the method further includes: Based on the results of dynamic differential semantic segmentation, high bitrates are assigned to pre-divided key component regions, and low bitrates are assigned to pre-divided non-key regions.

8. An online monitoring system for power distribution network equipment, characterized in that, include: The acquisition unit is used to acquire audio signals and image data from the distribution network equipment; The first extraction unit is used to perform dynamic differential semantic segmentation on the image data to extract image features, and to perform anti-interference processing on the audio signal based on wavelet transform. The second extraction unit is used to extract voiceprint features from the audio signal that has undergone anti-interference processing, the voiceprint features including Mel-spectral coefficients; The fusion unit is used to perform hierarchical fusion of the image features and the voiceprint features. The hierarchical fusion includes splicing at the feature layer and introducing an attention mechanism at the decision layer to dynamically allocate the weights of different modal features. The monitoring unit is used to simultaneously perform equipment defect identification and voiceprint forgery detection based on the fused features through a multi-task learning framework, and trigger alarms according to the identification results.

9. An online monitoring device for power distribution network equipment, characterized in that, The device includes a processor and a memory: The memory is used to store program code and transmit the program code to the processor; The processor is used to execute the online monitoring method for distribution network equipment according to any one of claims 1-7 according to the instructions in the program code.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium is used to store program code for executing the online monitoring method for distribution network equipment according to any one of claims 1-7.

Images