Distribution line hidden danger alarm intelligent noise reduction method based on multi-source data fusion, electronic equipment and storage medium

Abstract

The invention belongs to the technical field of intelligent operation inspection of distribution lines, and particularly relates to a distribution line hidden danger alarm intelligent noise reduction method based on multi-source data fusion, electronic equipment and a storage medium. The method comprises the following steps: realizing data acquisition by adopting an intelligent sensor and a visual monitoring and shooting device; constructing a space-time alignment engine to perform multi-source data fusion on the data acquired on the distribution line; the improved YOLOv5s model is used for counting, analyzing and cleaning the hidden danger alarm data of the distribution line; based on an XGBoost algorithm, constructing label data for the hidden danger alarm features of the distribution line; carrying out alarm feature model training, and constructing a multi-dimensional feature vector of the hidden danger alarm; first-stage alarm identification is carried out, and effective alarm and invalid alarm noise data of hidden dangers of the distribution line are identified; and carrying out second-stage repeated warning, noise reduction and filtering treatment on the distribution line. The problems that the number of invalid alarms of hidden dangers of the power distribution line is large, the operation and maintenance efficiency of power distribution operation and maintenance personnel is low and the effective alarm processing efficiency is low are solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent operation and inspection of distribution lines, and more specifically, relates to an intelligent noise reduction method for hidden danger alarms of distribution lines based on multi-source data fusion, an electronic device, and a storage medium. Background Art

[0002] The intelligent monitoring device for distribution lines based on artificial intelligence identification has the ability to intelligently analyze on-site hidden dangers and push alarms. As society's requirements for the quality of domestic electricity consumption continue to increase, a large number of intelligent monitoring devices are deployed on distribution lines to comprehensively monitor the operating conditions of the distribution lines. When the monitoring device identifies a safety hazard, it will promptly send an alarm to the distribution inspection personnel for confirmation and processing. The large number of intelligent monitoring devices deployed on-site will generate massive alarm images every day. At present, the mainstream application of artificial intelligence technology in distribution operation and maintenance is to analyze all scenarios through a unified model. During this period, no further difference analysis is performed on the analysis results, and no detailed classification of on-site scenarios is performed. All hidden danger information is pushed to monitoring personnel without taking any strategy. A large number of invalid alarms and false alarms bring additional workload to operation and maintenance personnel. It is very easy for effective alarm information to be handled in a timely manner. As a result, key hidden dangers are often submerged in a large number of ordinary hidden dangers or invalid alarms, making it difficult to detect and properly handle them in time. This seriously affects the accuracy and timeliness of operation and maintenance work, and thus increases the risk to the safe operation of distribution lines. For example, Chinese patent document CN118229067A discloses a method and system for identifying and judging electricity-related public safety hazards in distribution lines. The method includes the following steps: using image acquisition equipment to acquire images along the distribution lines to obtain original images of multiple locations, and assigning a location coordinate label to each original image; performing image processing on the original images to enhance the feature explicitness of the original images to obtain a first image sample; marking the risk levels of multiple locations in the distribution lines, and setting the number of image samples that need to be expanded for each location according to the risk level of each location; determining the number of second image samples that need to be expanded based on the number of image samples that need to be expanded, and applying a generative adversarial network to generate the required number of second image samples; using the first image samples and the second image samples to train a recognition model, and finally applying the recognition model to identify electricity-related public safety hazards in distribution lines.

[0003] In view of this, the present invention designs an intelligent noise reduction method for distribution line hidden danger alarm based on multi-source data fusion. Summary of the Invention

[0004] The present invention aims to overcome at least one defect of the above-mentioned prior art and provide an intelligent noise reduction method for distribution line hidden danger alarm based on multi-source data fusion to solve the problems of a large number of invalid alarms of distribution line hidden dangers, low operation and maintenance efficiency of distribution operation and maintenance personnel, and low efficiency of effective alarm handling.

[0005] The detailed technical solutions of the present invention are as follows: An intelligent noise reduction method for distribution line hidden danger alarm based on multi-source data fusion, the method comprising: S1. The design uses intelligent sensors and visual monitoring devices to collect time series data and image data on distribution lines; S2. Build a spatiotemporal alignment engine to perform multi-source data fusion on the data collected on the distribution lines; S3. Improve the YOLOv5s model and use the improved YOLOv5s model to count, analyze, and clean the distribution line hidden danger alarm data; S4. Construct label data for distribution line hidden danger alarm features based on the XGBoost algorithm. Specifically, first, randomly extract samples from the collected distribution line hidden danger data to ensure coverage of different distribution line scenarios; then, distribution line operation and maintenance professionals manually label the initial samples, and then use the XGBoost algorithm to train the classification model. The model then predicts new data, and finally, distribution line operation and maintenance professionals perform verification and iterative optimization to obtain the constructed label data. The label data is specifically defined as 0 = invalid alarm and 1 = valid alarm; S5. Use labeled data to train an alarm feature model and construct a multi-dimensional feature vector for hidden danger alarms in distribution line scenarios, using space, time, historical repetition, and object dynamics as dimensions. S6. The extracted multi-dimensional feature vector is used as the input of the CNN-LSTM model to perform the first stage of alarm recognition, identifying the valid alarms and invalid alarm noise data of the hidden dangers of the distribution line; S7. Perform the second stage of filtering on the valid alarms identified in the first stage, that is, perform noise reduction filtering on the repeated alarms: first, for the valid alarms, identify and analyze the alarm potential targets in the focus area of the pictures captured by the intelligent monitoring device of the distribution line, and extract the repeated alarms; then perform noise reduction filtering on the repeated alarms.

[0006] Preferably, according to the present invention, the spatiotemporal alignment engine includes a time synchronization module, a spatial coordinate mapping module, a data index and storage module, a calculation engine module, and a data verification module, and is specifically designed as follows: Time Synchronization Module: This module achieves nanosecond-level time synchronization through a dual-mode GPS / Beidou timing device, eliminating clock drift between devices. Furthermore, the design utilizes a hardware-level timestamp engine, NTP protocol conversion, and a four-level electromagnetic shielding design to ensure stable timing accuracy within ±15 nanoseconds even in strong interference environments.

[0007] Spatial coordinate mapping module: maps multi-source data to a unified three-dimensional geographic coordinate system. Further design uses spatial grid coding to establish a mapping relationship between pixel coordinates and physical locations with an error of <0.5 meters.

[0008] Data indexing and storage module: Build a spatiotemporal joint index to achieve millisecond-level data retrieval. Further design uses R-tree index to accelerate spatial range queries and B+ tree index to optimize time series retrieval. The storage and computing integrated architecture design places computing nodes close to the storage location to reduce network transmission delays.

[0009] Computing engine module: performs spatiotemporal correlation analysis and interpolation compensation complex calculations, further designs and integrates the streaming computing framework to process real-time data, and the batch processing engine to analyze historical data.

[0010] Data verification module: Verify the integrity and consistency of spatiotemporal data, further design and detect the benchmark deviation of the coordinate system and dynamically correct it, identify non-stationary data through ADF test and perform differential processing.

[0011] According to the preferred embodiment of the present invention, in step S2, the specific processing steps for multi-source data fusion are as follows: S21. Data preprocessing based on the time synchronization module: removing invalid timestamps from the collected distribution line data and converting data of different coordinate references into the CGCS2000 coordinate system; S22. Perform spatiotemporal grid division based on the spatial coordinate mapping module: Divide the monitoring area into 100m×100m grid cells, bind each cell to a unique spatiotemporal ID, slice the time dimension at 1-minute intervals, and establish a spatiotemporal cube data model; S23, extracting and associating features based on the data index and storage module: performing a spatiotemporal connection operation to associate the second-level current data collected by the smart sensor with the minute-level data collected by the visual monitoring device within the spatiotemporal grid; S24. Dynamic weight allocation based on the calculation engine module and data verification module: Build a feature importance evaluation model. When the image recognition confidence level is greater than 90%, visual features are preferentially used for spatial positioning. S25. Output fused data based on the data verification module: Generate spatiotemporal feature vectors in the format of [timestamp, longitude, latitude, current value, image feature code] and output them through Kafka streaming.

[0012] According to the preferred embodiment of the present invention, in step S3, the improvement of the YOLOv5s model refers to adding a small target detection weight to the loss function, which is divided into classification loss weighting and bounding box regression loss optimization, as follows: The classification loss weighting is designed as follows: In terms of dynamic weight allocation, a category weight factor is introduced into the original binary cross entropy loss. The calculation formula is: (1) in, is the classification loss, is the weight factor of category C, ,in, and is the adjustment coefficient, the default value =0.7, =0.3, is the category frequency; Indicates the area ratio of hidden danger targets in the image; represents the true label of category C, Represents the probability that the model predicts category C; The bounding box regression loss optimization design is as follows: Perform weighted IoU loss design and use dynamic IoU weights for special targets: (2) in, is the bounding box regression loss, is a constant factor, Indicates the number of target objects, represents the weighted intersection-over-union of the i-th target, is the category weight of the i-th target; The improved formula is as follows: (3) Among them, WIoU is the weighted intersection over union, A is the target box, B is the real box, is the intersection area of the target box A and the real box B, is the union area of the target box A and the real box B, is the scaling factor (the default value is 0.5), is the average object size of the dataset.

[0013] According to a preferred embodiment of the present invention, in step S3, the use of the improved YOLOv5s model to perform statistics and analysis on the distribution line hidden danger alarm data is specifically as follows: Statistics and analysis of hidden danger data of distribution lines, and design of category distribution statistics and target size distribution statistics; The specific design of category distribution statistics is as follows: Perform category frequency analysis on the labeled data set to identify special targets that need to be weighted. The frequency calculation formula for each category is: (4) in, is the number of samples of category C, K is the total number of categories, is the category frequency, is the number of samples in the K-th category; The specific design of target size distribution statistics is as follows: Count the area ratio of hidden danger targets in the image and identify small targets. The calculation formula is as follows: (5) in, is the target box size, is the image resolution, is the area ratio of the hidden danger target in the image.

[0014] Preferably, according to the present invention, in step S5, the alarm feature model training is specifically as follows: First, an incremental training strategy is performed, followed by feature importance screening, and finally adversarial verification; The incremental training strategy is designed as follows: Initial stage: The baseline model is trained based on 10,000 samples annotated by distribution line operation and maintenance professionals; Iteration phase: a. Use the model to predict unlabeled data and select samples with prediction probabilities in [0.3, 0.7]; b. Manually label 500 items per batch and update the training set; c. Update the model in each round of iteration until the PR-AUC improvement is less than 0.5%; The feature importance screening is designed to screen the top 20 features by SHAP value; The adversarial validation design is to build a two-classification model to distinguish the data distribution of the training set and the validation set, and eliminate features with large distribution differences, that is, PSI>0.1.

[0015] According to a preferred embodiment of the present invention, the S7 is specifically as follows: Analyze the hidden danger targets identified in the continuous capture scene and design a filtering strategy based on the three factors of the type and number of hidden danger objects remaining unchanged and the location not changing significantly. When all the above three factors are met, the hidden danger alarm is judged as a repeated alarm. When repeated alarms are found for the first time, noise reduction and filtering are performed on such hidden danger alarms; The key point of the repeated alarm filtering strategy is to judge the movement status of the alarm hidden danger target: first, the intersection-and-union ratio of the hidden danger target in the current captured image and the hidden danger target in the previous alarm image is calculated. When the calculated intersection-and-union ratio is greater than the preset judgment threshold, it is designed to determine that the hidden danger target has moved, otherwise it is determined that the hidden danger target has not moved. The intersection-and-union ratio of the hidden danger target is defined as the intersection area of the two rectangular boxes divided by the union area, which is used to describe the overlap of the two hidden danger target identification frames. The judgment threshold of the intersection-and-union ratio can be further adjusted according to actual conditions.

[0016] In another aspect of the present invention, an electronic device is provided, comprising: at least one processor; and A memory storing instructions, which, when executed by the at least one processor, enables the at least one processor to execute the intelligent noise reduction method for distribution line hidden danger alarm based on multi-source data fusion as described above.

[0017] In another aspect of the present invention, a machine-readable storage medium is provided, which stores executable instructions. When the instructions are executed, the machine executes the intelligent noise reduction method for distribution line hidden danger alarm based on multi-source data fusion as described above.

[0018] Compared with the prior art, the present invention has the following beneficial effects: (1) The proposed method constructs a spatiotemporal alignment engine to achieve spatiotemporal coordinate calibration of time series data and image data with different sampling frequencies. Alignment is performed based on GPS timestamp technology, mapping image pixels to line physical coordinates to achieve spatial grid mapping of distribution lines. This eliminates misalignment caused by multi-source data acquisition delays, such as the timing misalignment between tree barrier images and current mutations.

[0019] (2) The method of the present invention further completes the statistics, analysis and cleaning of the distribution line hidden danger alarm data through the improved design of the YOLOv5s algorithm; the specific design is to add a special target detection layer to the YOLOv5s algorithm model, such as hidden dangers such as mechanical construction damage around the wires and plastic bags hanging on the wires, so as to improve the accuracy of the distribution line image hidden danger target recognition and reduce the false alarm rate of image-based alarms.

[0020] (3) Based on the XGBoost algorithm, label data is constructed for the hidden danger alarm features of distribution lines, and the alarm feature model is trained to construct a multi-dimensional feature vector of hidden danger alarm in the distribution line scenario. Specifically, the multi-dimensional feature vector of hidden danger events is constructed based on the above-mentioned time series confidence score and image hidden danger recognition results as feature input (such as {current mutation probability 0.92, image foreign body confidence 0.85}). The first stage of identification of effective and invalid alarms of distribution line hidden dangers is completed through the above-mentioned feature combination design method.

[0021] (4) The method of the present invention realizes the second stage of noise reduction processing of the "invalid alarm" data in the identified valid alarms by designing a repeated alarm noise reduction strategy mechanism, that is, by designing a repeated alarm noise reduction filtering strategy, so as to further improve the accuracy of hidden danger alarms. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 It is a flow chart of the intelligent noise reduction method for hidden danger alarms described in the present invention.

[0023] Figure 2 This is a block diagram of the intersection and parallel ratio of distribution line hidden danger targets in Example 1 of the present invention.

[0024] Figure 3 This is an example diagram of a hidden danger alarm when the distribution line remains in a static state in Example 1 of the present invention.

[0025] Figure 4 This is an example diagram of the filtering effect of the first alarm and repeated alarm of hidden dangers of the distribution line in Example 1 of the present invention. DETAILED DESCRIPTION

[0026] The present disclosure will be further described below with reference to the accompanying drawings and embodiments.

[0027] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of the present disclosure. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present disclosure belongs.

[0028] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present disclosure. As used herein, unless the context clearly indicates otherwise, the singular form is intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.

[0029] In the absence of conflict, the embodiments of the present disclosure and the features thereof may be combined with each other.

[0030] Example 1 Ginseng Figure 1 ,This embodiment provides an intelligent noise reduction method for distribution line hidden danger alarm based on multi-source data fusion. The specific design of this method is divided into four stages. The first stage: multi-source data fusion and feature enhancement, that is, by designing and adopting intelligent sensors and visual monitoring devices to realize multi-source data collection of time series data and image data on distribution lines, and building a spatiotemporal alignment engine to perform multi-source data fusion on the data collected on distribution lines; the second stage: data cleaning and feature engineering, that is, using a time series model to perform statistics, analysis and cleaning on the collected data, and using a supervised classification model to construct label data for hidden danger alarm features. Through the improved design of the YOLOv5s algorithm, the distribution line hidden danger alarm data is further completed. Statistics, analysis and cleaning; the third stage: multimodal feature fusion, classification and alarm, that is, based on the XGBoost algorithm, label data is constructed for the distribution line hidden danger alarm features, alarm feature model training is carried out, and a multi-dimensional feature vector of hidden danger alarms in the distribution line scenario is constructed. Based on this, the first stage of identification of valid and invalid alarms of distribution line hidden dangers is completed; the fourth stage: repeated alarm noise reduction strategy mechanism design, that is, by designing a repeated alarm noise reduction filtering strategy, the second stage of noise reduction processing of the "invalid alarm" data in the identified valid alarms is realized. This method can effectively solve the problems of a large number of invalid alarms of distribution line hidden dangers, low operation and maintenance efficiency of distribution operation and maintenance personnel, and low efficiency in handling effective alarms.

[0031] The method comprises: S1. Deploy smart sensors and visual monitoring devices on distribution lines to collect time series and image data. This data is then cleaned using an LSTM time series model. Specifically, the LSTM model performs missing value processing, denoising, and outlier detection on the current time series data collected from the distribution lines. It extracts periodic and trend features to eliminate transient sensor interference, such as brief current fluctuations caused by bird contact. S2, build a spatiotemporal alignment engine to perform multi-source data fusion on the cleaned data in S1; The spatiotemporal alignment engine includes a time synchronization module, a spatial coordinate mapping module, a data indexing and storage module, a calculation engine module, and a data verification module. The specific design is as follows: Time Synchronization Module: This module achieves nanosecond-level time synchronization through a dual-mode GPS / Beidou timing device, eliminating clock drift between devices. Furthermore, the design utilizes a hardware-level timestamp engine, NTP protocol conversion, and a four-level electromagnetic shielding design to ensure stable timing accuracy within ±15 nanoseconds even in strong interference environments.

[0032] Spatial coordinate mapping module: maps multi-source data to a unified three-dimensional geographic coordinate system. Further design uses spatial grid coding to establish a mapping relationship between pixel coordinates and physical locations with an error of <0.5 meters.

[0033] Data indexing and storage module: Build a spatiotemporal joint index to achieve millisecond-level data retrieval. Further design uses R-tree index to accelerate spatial range queries and B+ tree index to optimize time series retrieval. The storage and computing integrated architecture design places computing nodes close to the storage location to reduce network transmission delays.

[0034] Computing engine module: performs spatiotemporal correlation analysis and interpolation compensation complex calculations, further designs and integrates the streaming computing framework to process real-time data, and the batch processing engine to analyze historical data.

[0035] Data verification module: Verify the integrity and consistency of spatiotemporal data, further design and detect the benchmark deviation of the coordinate system and dynamically correct it, identify non-stationary data through ADF test and perform differential processing.

[0036] Specific processing steps for multi-source data fusion: Step 1: Data preprocessing based on the time synchronization module: Remove invalid timestamps from the collected distribution line data and convert data with different coordinate references into the CGCS2000 coordinate system.

[0037] Step 2: Perform spatiotemporal grid division based on the spatial coordinate mapping module: Divide the monitoring area into 100m×100m grid cells, bind each cell to a unique spatiotemporal ID, slice it at 1-minute intervals in the time dimension, and establish a spatiotemporal cube data model.

[0038] Step 3: Feature extraction and association based on the data index and storage module: Perform a spatiotemporal connection operation to associate the second-level current data collected by the smart sensor with the minute-level data collected by the visual monitoring device within the spatiotemporal grid.

[0039] Step 4: Dynamically allocate weights based on the computing engine module and the data verification module: Build a feature importance evaluation model. When the image recognition confidence is greater than 90%, visual features are preferentially used for spatial positioning.

[0040] Step 5: Output the fused data based on the data verification module: Generate a spatiotemporal feature vector in the format of [timestamp, longitude, latitude, current value, image feature code] and output it through Kafka streaming.

[0041] The construction of the spatiotemporal alignment engine is based on GPS timestamps to enable multi-source acquisition devices to operate at the same timestamp. That is, the GPS module obtains nanosecond-level precision timestamps, performs hardware clock correction on distributed intelligent sensors, visual monitoring devices, and other equipment, and eliminates timing misalignment caused by local clock drift of the equipment. The design also combines GPS timestamps with geographic coordinate information to achieve three-dimensional spatial positioning of hidden danger events, mapping the pixel coordinates in the conductor hidden danger image to the power tower GIS coordinate system. The spatiotemporal alignment engine is constructed to solve the synchronization problem of data collected at different frequencies on the distribution line. The specific construction steps of the space-time alignment engine are designed as follows: Hardware layer deployment: Deploy GPS / Beidou dual-mode timing modules on smart sensors and visual monitoring devices to achieve device-level time synchronization.

[0042] Software layer design: Define the spatiotemporal data model, adopt the ISO 8601 standard timestamp format for the time dimension, and establish the spatial coding rules for power equipment based on tower ID + longitude and latitude + elevation for the spatial dimension.

[0043] Data layer construction: Build a spatiotemporal database and use the GanosBase engine to achieve unified storage of vector, raster, and three-dimensional data.

[0044] Algorithm layer development: Design a spatiotemporal calibration algorithm, and develop a visual-sensor data dynamic compensation algorithm for the time lag of image data.

[0045] Optimization layer iteration: Deploy an incremental learning mechanism to automatically update the spatial coordinate mapping rules when new monitoring devices are connected.

[0046] S3. Improve the YOLOv5s model and further complete the statistics, analysis and cleaning of distribution line hidden danger alarm data.

[0047] First, the loss function in the YOLOv5s algorithm model is improved, and the small target detection weight is added to the loss function of the YOLOv5s algorithm model. Specifically, the classification loss is weighted and the bounding box regression loss is optimized.

[0048] The specific design of classification loss weighting is as follows: In terms of dynamic weight allocation, a category weight factor is introduced into the original binary cross entropy loss. The calculation formula is: (1) is the classification loss, is the weight factor of category C, ,in, and is the adjustment coefficient, the default value =0.7, =0.3, is the category frequency; Indicates the area ratio of hidden danger targets in the image; represents the true label of category C, Represents the probability that the model predicts category C.

[0049] The specific design of bounding box regression loss optimization is as follows: Perform weighted IoU loss design and use dynamic IoU weights for special targets, such as foreign objects hanging on wires: (2) in, is the bounding box regression loss, is a constant factor, Indicates the number of target objects, represents the weighted intersection-over-union of the i-th target, is the category weight of the i-th target; The improved formula is as follows: (3) Among them, WIoU is the weighted intersection over union, A is the target box, B is the real box, is the intersection area of the target box A and the real box B, is the union area of the target box A and the real box B, is the scaling factor, the default value is 0.5, is the average object size of the dataset.

[0050] Further statistics and analysis of distribution line hidden danger data are carried out, and category distribution statistics and target size distribution statistics are designed.

[0051] The specific design of category distribution statistics is as follows: Perform category frequency analysis on the labeled data set to identify special targets that need to be weighted, such as broken mechanical vehicles and damaged wire insulators. The frequency calculation formula for each category is: (4) in, is the number of samples of category C, K is the total number of categories, is the category frequency, is the number of samples in the K-th category.

[0052] Specific example: In a distribution line hidden danger dataset, if insulator damage accounts for 5%, the weight needs to be increased.

[0053] The specific design of target size distribution statistics is as follows: Count the area ratio of hidden danger targets in the image and identify small targets, such as foreign objects hanging from wires that are less than 0.5% of the area. The calculation formula is: (5) in, is the target box size, is the image resolution, is the area ratio of the hidden danger target in the image.

[0054] The specific cleaning design is as follows: perform noise data filtering and small target enhancement; for noise data filtering, remove abnormal data of the annotation box size: if the target box area or , which is considered invalid annotation; in addition, K-means++ is used to re-cluster the prior box size and remove the ones that deviate from the cluster center by more than Abnormal samples; for small target enhancement, copy / paste enhancement is performed on special targets: small targets are randomly copied to blank areas of the image to improve the model's sensitivity to small targets; when using the Mosaic algorithm for data enhancement, a higher sampling probability is set for small target images.

[0055] S4. Construct label data for distribution line hidden danger alarm features based on the XGBoost algorithm. Specifically, first, randomly extract samples from the collected distribution line hidden danger data to ensure coverage of different distribution line scenarios; then, distribution line operation and maintenance professionals manually label the initial samples, and then use the XGBoost algorithm to train the classification model. The model then predicts new data, and finally, distribution line operation and maintenance professionals perform verification and iterative optimization to obtain the constructed label data. The label data is specifically defined as 0 = invalid alarm and 1 = valid alarm; S5. Train a feature model for hidden danger warnings on distribution lines, constructing a multidimensional feature vector for hidden danger warnings in distribution line scenarios. Specifically, feature extraction, modeling, training, and optimization are performed using spatial, temporal, historical repetition, and object dynamics as dimensions. The dataset is divided by time: the past year is the training set, the last three months is the validation set, and the most recent month is the test set.

[0056] The specific training of the alarm feature model is as follows: First, an incremental training strategy is performed, followed by feature importance screening, and finally adversarial verification; The incremental training strategy is designed as follows: Initial stage: The baseline model is trained based on 10,000 samples annotated by distribution line operation and maintenance professionals; Iteration phase: a. Use the model to predict unlabeled data and select samples with prediction probabilities in [0.3, 0.7]; b. Manually label 500 items per batch and update the training set; c. Update the model in each round of iteration until the PR-AUC improvement is less than 0.5%; The feature importance screening is designed to screen the top 20 features by SHAP value; The adversarial validation design is to build a two-classification model to distinguish the data distribution of the training set and the validation set, and eliminate features with large distribution differences, that is, PSI>0.1.

[0057] S6. The extracted multi-dimensional feature vector is used as the input of the CNN-LSTM model to perform the first stage of alarm recognition, identifying the valid alarms and invalid alarm noise data of the hidden dangers of the distribution line; S7, perform the second stage of effective alarm filtering again, specifically, perform multiple noise reduction filtering of repeated alarm filtering again, specifically for the hidden danger object alarm that remains in a static state in the protection zone, such as Figure 3 shown.

[0058] S8. Perform denoising again on the repeated alarm filtering strategy: Analyze the hidden danger targets identified in the continuous capture scenes of the visual monitoring device on the distribution line, and filter the three-dimensional factors of the type and quantity of the hidden danger objects and the lack of obvious changes in location. When all the above three-dimensional factors are met, the hidden danger alarm is determined to be a repeated alarm.

[0059] When a hidden danger alarm is judged to be a repeated alarm, the intersection-and-union ratio of the hidden danger target in the current snapshot image and the hidden danger target in the previous alarm image is first calculated. When the calculated intersection-and-union ratio is greater than the preset judgment threshold, it is determined that the hidden danger target has moved, otherwise it is determined that the hidden danger target has not moved. The intersection-and-union ratio of the hidden danger target is defined as the intersection area of two rectangular boxes divided by the union area, which is used to describe the overlap of the two hidden danger target identification frames. The default judgment threshold of the intersection-and-union ratio is designed to be 50%, which can be further adjusted according to actual conditions, such as Figure 2 shown.

[0060] When a new hidden danger alarm occurs, it is compared with the previous five pictures of the point. If the overlap range with the alarm frame of the five pictures meets the "no movement" judgment condition, the new alarm of the point is selected for push. After pushing twice, it will no longer be pushed to achieve filtering and noise reduction of such hidden danger alarms. 37 vehicle static parking points are set up on the distribution line site. After testing, after the noise reduction method of the distribution line hidden danger alarm is used, the average number of invalid alarms will be reduced by about 700 per day. Invalid alarms, for example Figure 4 shown.

[0061] Example 2 This embodiment further provides an electronic device, including: at least one processor; and A memory storing instructions, which, when executed by the at least one processor, enables the at least one processor to execute the intelligent noise reduction method for distribution line hidden danger alarm based on multi-source data fusion as described above.

[0062] In this embodiment, electronic devices may include but are not limited to: personal computers, server computers, workstations, desktop computers, laptop computers, notebook computers, mobile computing devices, smart phones, tablet computers, cellular phones, personal digital assistants (PDAs), handheld devices, messaging devices, wearable computing devices, consumer electronic devices, and the like.

[0063] Example 3 This embodiment also provides a machine-readable storage medium storing executable instructions, which, when executed, enable the machine to execute the intelligent noise reduction method for distribution line hidden danger alarm based on multi-source data fusion as described above.

[0064] Specifically, a system or device equipped with a readable storage medium can be provided, on which software program codes that implement the functions of any of the above-mentioned embodiments are stored, and a computer or processor of the system or device can read and execute instructions stored in the readable storage medium.

[0065] In this case, the program code itself read from the machine-readable medium can implement the functions of any one of the above embodiments, and thus the machine-readable code and the machine-readable storage medium storing the machine-readable code constitute part of this specification.

[0066] Examples of readable storage media include floppy disks, hard disks, magneto-optical disks, optical disks (e.g., CD-ROMs, CD-Rs, CD-RWs, DVD-ROMs, DVD-RAMs, DVD-RWs, DVD-RWs), magnetic tapes, non-volatile memory cards, and ROMs. Alternatively, the program code may be downloaded from a server computer or a cloud via a communication network.

[0067] Obviously, the above embodiments of the present invention are merely examples for the purpose of clearly illustrating the technical solutions of the present invention, and are not intended to limit the specific implementation methods of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the claims of the present invention shall be included within the scope of protection of the claims of the present invention.

Claims

1. An intelligent noise reduction method for distribution line hidden danger alarm based on multi-source data fusion, characterized in that: The method comprises: S1. The design uses intelligent sensors and visual monitoring devices to collect time series data and image data on distribution lines; S2. Build a spatiotemporal alignment engine to perform multi-source data fusion on the data collected on the distribution lines; S3. Improve the YOLOv5s model and use the improved YOLOv5s model to count, analyze, and clean the distribution line hidden danger alarm data; S4. Construct label data for distribution line hidden danger alarm features based on the XGBoost algorithm. Specifically, first, randomly extract samples from the collected distribution line hidden danger data to ensure coverage of different distribution line scenarios; then, distribution line operation and maintenance professionals manually label the initial samples, and then use the XGBoost algorithm to train the classification model. The model then predicts new data, and finally, distribution line operation and maintenance professionals perform verification and iterative optimization to obtain the constructed label data. The label data is specifically defined as 0 = invalid alarm and 1 = valid alarm; S5. Use labeled data to train an alarm feature model and construct a multi-dimensional feature vector for hidden danger alarms in distribution line scenarios, using space, time, historical repetition, and object dynamics as dimensions. S6. The extracted multi-dimensional feature vector is used as the input of the CNN-LSTM model to perform the first stage of alarm recognition, identifying the valid alarms and invalid alarm noise data of the hidden dangers of the distribution line; S7. Perform the second stage of filtering on the valid alarms identified in the first stage, that is, perform noise reduction filtering on the repeated alarms: first, for the valid alarms, identify and analyze the alarm potential targets in the focus area of the pictures captured by the intelligent monitoring device of the distribution line, and extract the repeated alarms; then perform noise reduction filtering on the repeated alarms.

2. The intelligent noise reduction method for distribution line hidden danger alarm based on multi-source data fusion according to claim 1 is characterized in that: In step S2, the spatiotemporal alignment engine includes a time synchronization module, a spatial coordinate mapping module, a data indexing and storage module, a calculation engine module, and a data verification module. The specific design is as follows: Time synchronization module: This module uses a GPS / Beidou dual-mode timing device to achieve nanosecond-level time synchronization, eliminating clock drift between devices. Furthermore, the design uses a hardware-level timestamp engine, NTP protocol conversion, and a four-level electromagnetic shielding design to ensure stable timing accuracy within ±15 nanoseconds in strong interference environments. Spatial coordinate mapping module: maps multi-source data to a unified three-dimensional geographic coordinate system. Further design uses spatial grid coding to establish the mapping relationship between pixel coordinates and physical locations with an error of <0.5 meters; Data indexing and storage module: This module builds a spatiotemporal joint index to achieve millisecond-level data retrieval. Further designs use R-tree indexes to accelerate spatial range queries and B+-tree indexes to optimize time series retrieval. The integrated storage and computing architecture design places computing nodes close to storage locations to reduce network transmission latency. Computing engine module: performs spatiotemporal correlation analysis, interpolation compensation and complex calculations, further designs and integrates a streaming computing framework to process real-time data, and a batch processing engine to analyze historical data; Data verification module: Verify the integrity and consistency of spatiotemporal data, further design and detect the benchmark deviation of the coordinate system and dynamically correct it, identify non-stationary data through ADF test and perform differential processing.

3. The intelligent noise reduction method for distribution line hidden danger alarm based on multi-source data fusion according to claim 2 is characterized in that: In step S2, the specific processing steps for multi-source data fusion are as follows: S21. Data preprocessing based on the time synchronization module: removing invalid timestamps from the collected distribution line data and converting data of different coordinate references into the CGCS2000 coordinate system; S22. Perform spatiotemporal grid division based on the spatial coordinate mapping module: Divide the monitoring area into 100m×100m grid cells, bind each cell to a unique spatiotemporal ID, slice the time dimension at 1-minute intervals, and establish a spatiotemporal cube data model; S23, extracting and associating features based on the data index and storage module: performing a spatiotemporal connection operation to associate the second-level current data collected by the smart sensor with the minute-level data collected by the visual monitoring device within the spatiotemporal grid; S24. Dynamic weight allocation based on the calculation engine module and data verification module: Build a feature importance evaluation model. When the image recognition confidence level is greater than 90%, visual features are preferentially used for spatial positioning. S25. Output fused data based on the data verification module: Generate spatiotemporal feature vectors in the format of [timestamp, longitude, latitude, current value, image feature code] and output them through Kafka streaming.

4. The intelligent noise reduction method for distribution line hidden danger alarm based on multi-source data fusion according to claim 1 is characterized in that: In step S3, the improvement of the YOLOv5s model refers to adding a small target detection weight to the loss function, which is divided into classification loss weighting and bounding box regression loss optimization, as follows: The classification loss weighting is designed as follows: In terms of dynamic weight allocation, a category weight factor is introduced into the original binary cross entropy loss. The calculation formula is: （1） in, is the classification loss, is the weight factor of category C, ,in, and is the adjustment coefficient, the default value =0.7, =0.3, is the category frequency; Indicates the area ratio of hidden danger targets in the image; represents the true label of category C, Represents the probability that the model predicts category C; The bounding box regression loss optimization design is as follows: Perform weighted IoU loss design and use dynamic IoU weights for special targets: （2） in, is the bounding box regression loss, is a constant factor, Indicates the number of target objects, represents the weighted intersection-over-union of the i-th target, is the category weight of the i-th target; The improved formula is as follows: （3） Among them, WIoU is the weighted intersection over union, A is the target box, B is the real box, is the intersection area of the target box A and the real box B, is the union area of the target box A and the real box B, is the scaling factor (the default value is 0.5), is the average object size of the dataset.

5. The intelligent noise reduction method for distribution line hidden danger alarm based on multi-source data fusion according to claim 4 is characterized in that: According to the preferred embodiment of the present invention, in step S3, the improved YOLOv5s model is used to perform statistics and analysis on the distribution line hidden danger alarm data as follows: Statistics and analysis of hidden danger data of distribution lines, and design of category distribution statistics and target size distribution statistics; The specific design of category distribution statistics is as follows: Perform category frequency analysis on the labeled data set to identify special targets that need to be weighted. The frequency calculation formula for each category is: （4） in, is the number of samples of category C, K is the total number of categories, is the category frequency, is the number of samples in the K-th category; The specific design of target size distribution statistics is as follows: Count the area ratio of hidden danger targets in the image and identify small targets. The calculation formula is as follows: （5） in, is the target box size, is the image resolution, is the area ratio of the hidden danger target in the image.

6. The intelligent noise reduction method for distribution line hidden danger alarm based on multi-source data fusion according to claim 1 is characterized in that: In step S5, the alarm feature model training is specifically as follows: First, an incremental training strategy is performed, followed by feature importance screening, and finally adversarial verification; The incremental training strategy is designed as follows: Initial stage: The baseline model is trained based on 10,000 samples annotated by distribution line operation and maintenance professionals; Iteration phase: a. Use the model to predict unlabeled data and select samples with prediction probabilities in [0.3, 0.7]; b. Manually label 500 items per batch and update the training set; c. Update the model in each round of iteration until the PR-AUC improvement is less than 0.5%; The feature importance screening is designed to screen the top 20 features by SHAP value; The adversarial validation design is to build a two-classification model to distinguish the data distribution of the training set and the validation set, and eliminate features with large distribution differences, that is, PSI>0.

1.

7. The intelligent noise reduction method for distribution line hidden danger alarm based on multi-source data fusion according to claim 1 is characterized in that: The S7 is specifically as follows: Analyze the hidden danger targets identified in the continuous capture scene and design a filtering strategy based on the three factors of the type and number of hidden danger objects remaining unchanged and the location not changing significantly. When all the above three factors are met, the hidden danger alarm is judged as a repeated alarm. When a repeated alarm is first detected, noise reduction and filtering are performed on this type of hidden danger alarm as follows: First, the intersection-and-union ratio of the hidden danger target in the current captured image and the hidden danger target in the previous warning image is calculated. It is designed that when the calculated intersection-and-union ratio is greater than the preset judgment threshold, the hidden danger target is judged to have moved; otherwise, it is judged that the hidden danger target has not moved. The intersection-and-union ratio of the hidden danger target is defined as the intersection area of the two rectangular boxes divided by the union area, which is used to describe the overlap of the two hidden danger target identification frames. The judgment threshold of the intersection-and-union ratio can be further adjusted according to actual conditions.

8. An electronic device, characterized in that: The electronic device comprises: at least one processor; and A memory storing instructions, which, when executed by the at least one processor, causes the at least one processor to execute the intelligent noise reduction method for distribution line hidden danger alarm based on multi-source data fusion as described in any one of claims 1 to 7.

9. A machine-readable storage medium, characterized in that The machine-readable storage medium stores executable instructions, which, when executed, enable the machine to execute the intelligent noise reduction method for distribution line hidden danger alarm based on multi-source data fusion according to any one of claims 1 to 7.

Images