Double-path fusion sensing live monitoring model and device

Abstract

The invention discloses a double-channel fusion perception live monitoring model and device, the model comprises a double-channel convolution module, and the double-channel convolution module comprises an input layer, an upper channel convolution layer and a lower channel convolution layer; the feature fusion module is used for carrying out feature fusion on the output data of the upper channel convolution layer and the output data of the lower channel convolution layer; the global average pooling-softmax module is used for processing the fused features and outputting an equipment state recognition result; the device comprises a multi-core control board, a front-end monitoring module, an upper computer module and a wireless communication module. According to the model and the device provided by the invention, efficient and accurate monitoring and diagnosis of the state of the power equipment are realized through multi-parameter data fusion and double-path collaborative identification.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of multi-physical field monitoring of power equipment, and in particular to a dual-path fusion perception live monitoring model and device. BACKGROUND

[0002] It is the key of power equipment operation and maintenance management to master the operation state and operation environment of power transmission and transformation equipment and to find the hidden danger of power grid equipment in time. At present, due to the large number of equipment, complex operation environment and limitations of state detection technology, power equipment state monitoring still faces many challenges, which are specifically shown in the following aspects:

[0003] It is difficult to mine features: the traditional convolutional neural network has limited feature extraction capability for multi-parameter long sequence data, which easily leads to blocked or lost information transmission and makes it difficult to effectively mine key modal information in the data;

[0004] Weak anti-interference ability: there are noise and other irrelevant factors in the actual collection environment, and the traditional algorithm is difficult to focus on key feature information, which affects the accuracy of state recognition;

[0005] Insufficient time sequence feature capture: the one-way network structure cannot adapt to the time sequence features of complex multi-parameter data, which limits the state research and judgment accuracy.

[0006] At the same time, the existing power equipment online monitoring system also has the problems of low use efficiency, poor economy, insufficient stability and reliability, etc., which is easy to cause false alarm or omission; the traditional patrol and live detection depends on fixed cycle and cannot realize whole process monitoring, which is difficult to find equipment hidden danger in time; and the existing system only collects single data such as current and voltage, lacks multi-source data support, and is difficult to realize multi-dimensional comprehensive diagnosis.

[0007] Therefore, it becomes an urgent problem in the industry to provide a live monitoring technology that can fuse multi-parameter data, efficiently extract features and adapt to complex environments to improve the accuracy and real-time performance of power equipment state monitoring. SUMMARY

[0008] In view of the above-mentioned deficiencies in the current multi-physical field monitoring technology field of power equipment, the present application provides a dual-path fusion perception live monitoring technology, which realizes efficient and accurate monitoring and diagnosis of the state of power equipment through multi-parameter data fusion and dual-path collaborative recognition.

[0009] To achieve the above-mentioned purpose, the first aspect of the present application provides a dual-path fusion perception live monitoring model, which comprises:

[0010] A dual-channel convolution module, the dual-channel convolution module comprises an input layer, an upper channel convolution layer and a lower channel convolution layer;

[0011] a feature fusion module configured to fuse the output data of the upper channel convolutional layer and the output data of the lower channel convolutional layer;

[0012] a global average pooling-softmax module configured to process the fused features and output a device state recognition result.

[0013] In some embodiments of the first aspect of the application, the input layer is configured to receive two types of charged parameter time domain data, and the dimension of the charged parameter time domain data is 1*50*2.

[0014] The upper channel convolutional layer comprises two sub-modules and one pooling layer.

[0015] The lower channel convolutional layer comprises four sub-modules and two pooling layers.

[0016] In some embodiments of the first aspect of the application, each sub-module of the upper channel convolutional layer comprises a convolutional layer, an AdaBN layer and an activation layer.

[0017] The first sub-module inputs the time domain data of the input layer and outputs data with a dimension of 1*30*2.

[0018] The second sub-module inputs the output data of the first sub-module and outputs data with a dimension of 1*10*2.

[0019] The pooling layer inputs the output data of the second sub-module and outputs data with a dimension of 1*2*2.

[0020] In some embodiments of the first aspect of the application, each sub-module of the lower channel convolutional layer comprises a convolutional layer, an AdaBN layer and an activation layer.

[0021] The first sub-module inputs the time domain data of the input layer and outputs data with a dimension of 1*40*2.

[0022] The second sub-module inputs the output data of the first sub-module and outputs data with a dimension of 1*20*2.

[0023] The first pooling layer inputs the output data of the second sub-module and outputs data with a dimension of 1*2*2.

[0024] The third sub-module inputs the output data of the first pooling layer and outputs data with a dimension of 1*30*2.

[0025] The fourth sub-module inputs the output data of the third sub-module and outputs data with a dimension of 1*20*2.

[0026] The second pooling layer inputs the output data of the fourth sub-module and outputs data with a dimension of 1*2*2.

[0027] In some embodiments of the first aspect of the application, the specific way of performing feature fusion on the output data of the upper channel convolutional layer and the output data of the lower channel convolutional layer is to perform element-wise multiplication operation on the two same-dimension feature data to form comprehensive features.

[0028] In some embodiments of the first aspect of the application, the specific process of the global average pooling-softmax module for processing the fused features and outputting the device state recognition result includes:

[0029] The global average pooling operation is used to process the fused features, convert the high-dimension feature matrix into a fixed-length feature vector, realize feature dimension reduction and retain key information;

[0030] The softmax function is used to process the reduced feature vector, map the feature vector to the probability distribution of the device state, and output the recognition result of the device state.

[0031] To achieve the above-mentioned purpose, the second aspect of the application provides a dual-path fusion perception live-line monitoring device, which comprises a multi-core control board, a front-end monitoring module, an upper computer module and a wireless communication module.

[0032] The multi-core control board carries the dual-path fusion perception live-line monitoring model, which comprises a data sorting unit and a communication unit, the data sorting unit is used to sort two live-line parameter time domain data input into the dual-path fusion perception live-line monitoring model, and the communication unit is used to realize communication with other modules.

[0033] The front-end monitoring module is connected with the multi-core control board through a communication bus, and comprises a vibration monitoring module, a humidity monitoring module, a temperature monitoring module and a partial discharge monitoring module.

[0034] The upper computer module comprises a liquid crystal display module and an analysis module connected with the liquid crystal display module, the analysis module is connected with the multi-core control board through a network port, and the upper computer module is provided with a TYPE-C power supply interface for supplying power to the liquid crystal display module and the analysis module.

[0035] The wireless communication module is a 5G functional module.

[0036] In some embodiments of the second aspect of the application, the vibration monitoring module comprises an RS422 interface and an analysis unit connected with the RS422 interface; the analysis unit is connected with the multi-core control board through a network port; and the vibration monitoring module is provided with a three-core power supply interface for providing DC 5V voltage for the RS422 interface and DC 3.3V voltage for the analysis unit.

[0037] In some embodiments of the second aspect of the application, the humidity monitoring module comprises a humidity sensor, a collection unit and an analysis unit connected in sequence; the analysis unit is connected to the multi-core control board through a network port; the humidity monitoring module is provided with a TYPE-C power supply interface for supplying power to the humidity sensor, the collection unit and the analysis unit.

[0038] In some embodiments of the second aspect of the application, the temperature monitoring module comprises a temperature sensor, a collection unit and an analysis unit connected in sequence; the analysis unit is connected to the multi-core control board through a network port; the temperature monitoring module is provided with a TYPE-C power supply interface for supplying power to the temperature sensor, the collection unit and the analysis unit.

[0039] In some embodiments of the second aspect of the application, the partial discharge monitoring module comprises a partial discharge sensor, a collection unit and an analysis unit connected in sequence; the analysis unit is connected to the multi-core control board through a network port; the partial discharge monitoring module is provided with a TYPE-C power supply interface for supplying power to the partial discharge sensor, the collection unit and the analysis unit.

[0040] In some embodiments of the second aspect of the application, the communication bus is a MODBUS bus.

[0041] The multi-core control board is integrated with an AC-DC module and a power management module, the AC-DC module is used for power supply and voltage conversion, and the power management module is used for power distribution and management.

[0042] The 5G function module comprises a 5G communication module, an analysis unit and a power management unit, the analysis unit is connected to the multi-core control board through a USB interface and performs data interaction, and the power management unit is powered through a TYPE-C power supply interface and outputs a DC-3.3V voltage to the analysis unit

[0043] The advantages of the embodiment of the present application are as follows: firstly, different features of multi-parameter data are learned through the double-channel convolution module, and feature complementation is realized through element-by-element multiplication of the feature fusion module, so that the robustness and detail extraction capability of the model are enhanced, and the focusing capability of the model on key information is improved, solving the problem of insufficient feature mining of traditional models; secondly, by integrating vibration, temperature, humidity and partial discharge four types of monitoring data, multi-parameter fusion analysis and multi-dimensional comprehensive diagnosis are realized, avoiding the limitations of single parameter monitoring; thirdly, the multi-core control board of the present application has edge technology function, can analyze and judge the data on site, reduces the data transmission pressure, and improves the response speed; fourthly, modular design is adopted, which reduces the deployment cost of similar functions, and supports 5G remote communication and local interaction, and can be flexibly adapted to different application scenarios; fifthly, through multi-module cooperation and anti-interference design, the false alarm and missed alarm problems are reduced, the long-term running performance is stable, and it is suitable for complex power environment. BRIEF DESCRIPTION OF DRAWINGS

[0044] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed to be used in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0045] Figure 1 The structure diagram of the charged monitoring model of the double-path fusion perception according to the present application;

[0046] Figure 2 The structure diagram of the charged monitoring device of the double-path fusion perception according to the present application;

[0047] Figure 3 The structure diagram of the 5G function module according to the present application;

[0048] Figure 4 The structure diagram of the multi-core control board according to the present application;

[0049] Figure 5 The structure diagram of the vibration detection module according to the present application;

[0050] Figure 6 The structure diagram of the humidity monitoring module according to the present application;

[0051] Figure 7 The structure diagram of the temperature monitoring module according to the present application;

[0052] Figure 8 The structure diagram of the partial discharge monitoring module according to the present application;

[0053] Figure 9A structure diagram of the host computer module according to the present application. DETAILED DESCRIPTION

[0054] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative effort fall within the protection scope of the present application.

[0055] Figure 1 A structure diagram of a live-line monitoring model of a dual-channel fusion perception according to the present application is shown. Figure 1 As shown in the figure, the model of the present application includes a dual-channel convolution module, a feature fusion module and a global average pooling-softmax module.

[0056] In the embodiments of the present application, the dual-channel convolution module specifically includes an input layer, an upper channel convolution layer and a lower channel convolution layer.

[0057] The input layer is used to receive input of feature data. In the present model, the input feature data is artificially sorted two kinds of live-line parameter time domain data, and the dimension is 1*50*2. This dimension design provides a basic input format for the dual-channel convolution module.

[0058] It should be noted that the two kinds of live-line parameter time domain data described above refer to vibration parameters and partial discharge parameters. In the dimension representation of time domain data, the first field represents the number of single input samples, the second field represents the length of single parameter time domain sequence, and the third field represents the number of double parameter channels (representing vibration signals and partial discharge signals respectively). In the embodiments of the present application, the specific meaning of the input data with the dimension of 1*50*2 is as follows: one sample (representing a single processing unit to ensure real-time response), 50 time sequence points (covering a 50ms sampling window to capture transient characteristics), and two parameter channels (channel 1 corresponds to vibration parameters, and channel 2 corresponds to partial discharge parameters to realize dual-channel feature complementation).

[0059] The upper channel convolution layer is composed of two sub-modules and a pooling layer. The two sub-modules each include a convolution layer (in which: the convolution kernel size is 3*3, the number of convolution kernels is 2, and the stride is 1), an AdaBN layer and an activation layer. The input of the first sub-module is the time domain data of the input layer, and the output dimension is 1*30*2. The input of the second sub-module is the output data of the first sub-module, and the output dimension is 1*10*2. The input of the pooling layer (average pooling) is the output data of the second sub-module, and the output dimension is 1*2*2.

[0060] Here, the pooling layer in the upper channel convolution layer adopts a flat pooling operation.

[0061] The lower channel convolutional layer is composed of four sub-modules and two pooling layers, each of the four sub-modules is composed of a convolutional layer (in which the convolution kernel size is 3*3, the number of convolution kernels is 2, and the stride is 1), an AdaBN layer, and an activation layer, the input of the first sub-module is the time domain data of the input layer, and the output dimension is 1*40*2; the input of the second sub-module is the output data of the first sub-module, and the output dimension is 1*20*2; the input of the first pooling layer is the output data of the second sub-module, and the output dimension is 1*2*2; the input of the third sub-module is the output of the first pooling layer, and the output dimension is 1*30*2; the input of the fourth sub-module is the output of the third sub-module, and the output dimension is 1*20*2; the input of the second pooling layer is the output data of the fourth sub-module, and the output dimension is 1*2*2.

[0062] Here, the first pooling layer and the second pooling layer in the lower channel convolutional layer both adopt the average pooling operation.

[0063] It should be noted that the AdaBN (Adaptive Batch Normalization) layer is a deep learning technology specially used for domain adaptation and cross-domain learning. The core idea is to dynamically adjust the statistics (mean and variance) of the batch normalization (BN) layer, so that the model can quickly adapt to the data distribution of the target domain.

[0064] In the embodiment of the application, the feature fusion module realizes feature fusion in the form of element-by-element multiplication.

[0065] The feature fusion module is the core unit for realizing feature complementation in the dual-path fusion perception live-line monitoring model, and its core function is to efficiently integrate the output data of the upper and lower two channel convolutional layers in the dual-channel convolutional module. The module first receives the feature data with a dimension of 1*2*2 output by the upper channel convolutional layer after being processed by two sub-modules and once pooling, and the feature data with a dimension of 1*2*2 output by the lower channel convolutional layer after being processed by four sub-modules and twice pooling. Since the dimensions of the two paths of feature data are completely consistent, the module performs deep fusion on the two paths of feature data in the form of element-by-element multiplication. This fusion method can fully combine the data local key detail extraction capability of the upper channel convolutional layer and the data global trend capturing capability of the lower channel convolutional layer, so that the two paths of feature data realize complementary advantages at the numerical level, effectively enhance the information richness and representativeness of the integrated feature after fusion, and further improve the robustness of the subsequent module for equipment state recognition, while solving the problems of insufficient focusing on key information and weak anti-interference capability of the traditional single-channel model due to single feature, thereby laying a feature foundation for the model to output accurate recognition results.

[0066] In the embodiment of the present application, the global average pooling-softmax module receives the fusion features output by the feature fusion module, processes them and outputs the device state recognition result.

[0067] The global average pooling-softmax module undertakes the key conversion task of the model from the fusion features to the device state recognition result output. Its function implementation is divided into two core stages: the first stage is the global average pooling processing. This module first receives the comprehensive features output by the feature fusion module, and performs dimension reduction processing on the high-dimensional feature matrix through the global average pooling operation, converting the originally structured feature matrix into a fixed-length feature vector. In this process, not only can the subsequent calculation complexity be effectively simplified, but also the key information in the fusion features can be preserved to the greatest extent, avoiding the information redundancy or loss problem that may be caused by the traditional fully connected layer. The second stage is the softmax function processing. The module inputs the fixed-length feature vector after dimension reduction into the softmax function, and through the function, maps the feature vector into a numerical set conforming to the probability distribution law. Each number corresponds to the probability of a device state (such as normal, abnormal). Finally, according to the probability distribution result, an explicit device state recognition result is output, ensuring that the recognition result not only has the reliability of probability basis, but also can directly serve the power equipment state monitoring and diagnosis scene, providing clear and accurate device operation state judgment basis for operation and maintenance personnel.

[0068] Figure 2 The structure of the device is shown in the figure. Figure 2 As shown in the figure, the system of the present application includes a multi-core control board, a front-end monitoring module, an upper computer module and a wireless communication module. Among them, the front-end monitoring module has multiple front-end monitoring modules, each of which is connected with the multi-core control board through a communication bus.

[0069] Among them, the communication bus adopts MODBUS bus, but is not limited to MODBUS bus, and the wireless communication module preferably adopts a remote communication 5G function module as a core component.

[0070] Figure 3 The specific structure of the remote communication 5G function module is shown in the figure. Figure 3 As shown in the figure, the 5G function module mainly consists of a 5G communication module, an analysis unit and a power management unit. Among them, the power management unit supplies power to the system through the integrated TYPE-C power supply interface and outputs a DC-3.3V stable voltage to the analysis unit; the analysis unit realizes physical connection and data interaction with the multi-core control board through the USB interface.

[0071] In the embodiment of the present application, the number of cores of the multi-core control board includes but is not limited to 4, and the first core is denoted as A core, and the remaining cores are denoted as B core. At the same time, the multi-core control board loads a live-line monitoring model with double-path fusion perception and includes a data sorting unit and a communication unit for communication with other modules. Among them, the A core is used for communication with other modules; the B core is used for running the live-line monitoring model with double-path fusion perception; and the data sorting unit is used for sorting two kinds of live-line parameter time domain data input to the live-line monitoring model with double-path fusion perception.

[0072] At the same time, the multi-core control board is also integrated with an AC-DC module and a power management module, the AC-DC module is used for power supply and voltage conversion of the multi-core control board, and the power management module is used for power distribution and management.

[0073] In the embodiment of the present application, the front-end monitoring module includes a vibration monitoring module, a humidity monitoring module, a temperature monitoring module and a partial discharge monitoring module, and each of these front-end monitoring modules is provided with a power supply interface for power supply.

[0074] Figure 5 The specific structure of the vibration monitoring module is shown. As shown in Figure 5 The vibration monitoring module includes an RS422 interface and an analysis unit connected with the RS422 interface. Among them, the analysis unit is connected with the multi-core control board through a network port; in addition, the vibration monitoring module is also provided with a three-core power supply interface for power supply of the RS422 interface and the analysis unit, wherein the power supply voltage of the RS422 interface is DC-5V, and the power supply voltage of the analysis unit is DC-3.3V.

[0075] Figure 6 The specific structure of the humidity monitoring module is shown. As shown in Figure 6 The humidity monitoring module includes a humidity sensor, a collection unit and an analysis unit connected in sequence. Among them, the analysis unit is connected with the multi-core control board through a network port; in addition, the humidity monitoring module also includes a TYPE-C power supply interface for power supply of the temperature sensor, the collection unit and the analysis unit.

[0076] Figure 7 The specific structure of the temperature monitoring module is shown. As shown in Figure 7 The temperature monitoring module includes a temperature sensor, a collection unit and an analysis unit connected in sequence. Among them, the analysis unit is connected with the multi-core control board through a network port; in addition, the temperature monitoring module also includes a TYPE-C power supply interface for power supply of the temperature sensor, the collection unit and the analysis unit.

[0077] Figure 8 The specific structure of the partial discharge monitoring module is shown. As shown in Figure 8As shown, the partial discharge monitoring module includes a partial discharge sensor, an acquisition unit and an analysis unit connected in sequence. The analysis unit is connected to the multi-core control board through a network port. In addition, the partial discharge monitoring module further includes a TYPE-C power supply interface for supplying power to the partial discharge sensor, the acquisition unit and the analysis unit.

[0078] Figure 9 The specific structure of the host computer module is shown. As shown in the figure, Figure 9 The host computer module includes a liquid crystal display module and an analysis module connected to the liquid crystal display module. The analysis module is connected to the multi-core control board through a network port. In addition, the host computer module further includes a TYPE-C power supply interface for supplying power to the liquid crystal display module and the analysis unit.

[0079] To verify the effectiveness and reliability of the charged monitoring model and device of the dual-path fusion perception of the present application, the embodiment of the present application designs a systematic test scheme. The test scheme covers multiple aspects such as test environment construction, data acquisition process, model performance verification and long-term stability test. The test scheme can comprehensively evaluate its application value in actual live monitoring scenarios.

[0080] The test environment simulates typical high-voltage electrical equipment operation scenarios. 10kV switch cabinet, 35kV cable terminal and 110kV transformer are selected as test objects, and a comprehensive test platform is built, which includes temperature, humidity, vibration and adjustable partial discharge parameters. The platform is mainly composed of a closed test cabin, which integrates a temperature control system, a humidity adjustment system, a vibration excitation system and a partial discharge simulation system. The temperature control system has a regulation range of 20-80℃ and an accuracy of ±0.5℃; the humidity adjustment system has a regulation range of 20-95% RH and an accuracy of ±2% RH; the vibration excitation system has a vibration frequency of 10-500HZ and an amplitude of 0-5mm; the partial discharge simulation system can simulate corona discharge, surface discharge and internal discharge, with a discharge amount range of 10-10000PC. An independent control console is arranged outside the test cabin, which is connected with the equipment in the cabin through MODBUS bus, realizing remote control and real-time monitoring of each environmental parameter. At the same time, front-end monitoring modules are arranged at the key monitoring points of the test objects: the vibration monitoring module is installed at the vibration sensitive part of the equipment shell, using an acceleration sensor to collect vibration signals; the humidity monitoring module and the temperature monitoring module are installed on the surface of the equipment to ensure sufficient contact with the measured environment; the partial discharge monitoring module is connected to the high-voltage end of the equipment through a coupling capacitor to obtain accurate partial discharge signals. All front-end monitoring modules are connected to the multi-core control board through a network port, and the multi-core control board is connected to the remote host computer through a 5G function module to realize real-time data transmission and remote monitoring.

[0081] The test procedure is divided into four stages: system calibration, static performance test, dynamic response test and long-term stability test. In the system calibration stage, the parameters of each front-end monitoring module need to be calibrated: the vibration monitoring module generates a known frequency and amplitude vibration signal through a standard vibration table, compares the deviation between the module output value and the standard value, and corrects the error through the data sorting unit of the multi-core control board; the humidity and temperature monitoring modules are placed in a standard humidity chamber and a constant temperature tank respectively, and the module readings are recorded at different temperature and humidity points to establish a correction curve; the partial discharge monitoring module inputs a standard pulse signal through a partial discharge calibration instrument to adjust the gain and threshold of the module to ensure that the discharge measurement error is controlled within ±5%. After calibration, the multi-core control board is loaded into the collaborative recognition model of double-path convolution and feature fusion, the input parameters of the model are set as the time domain data of vibration and partial discharge, the input layer dimension is 1*50*2, the parameters of the upper and lower channel convolution layers are fixed according to the document, the feature fusion adopts the element-by-element multiplication method, and the output layer realizes the classification of equipment state through the softmax function, including normal and abnormal states.

[0082] In the static performance test stage, several groups of stable environmental parameter combinations are set in the test cabin, such as temperature 25℃, humidity 50%RH, no vibration, no partial discharge normal state; temperature 60℃, humidity 80%RH, vibration frequency 50HZ, partial discharge quantity 500PC abnormal state. Each group of parameters is maintained for 30 minutes, during which the front-end monitoring module continuously collects data at a sampling frequency of 1kHZ, and the data is processed by the multi-core control board and uploaded to the host computer through the 5G module. 1000 groups of samples are collected under each state, the recognition accuracy of the model for the equipment state is calculated, and the time consumption from data collection to recognition of a single group of data is recorded, which includes the total time of data transmission, feature extraction and classification calculation. To verify the advantages of double-path fusion, a single parameter monitoring comparison test is carried out simultaneously, and only vibration parameters or partial discharge parameters are used to input the model, and the recognition accuracy is calculated and analyzed with the double-path fusion result.

[0083] In the dynamic response test stage, the parameter mutation process in the equipment running is simulated, the temperature is linearly increased from 30℃ to 70℃ in 5 minutes, the partial discharge quantity is stepped from 0PC to 800PC, the vibration frequency is randomly fluctuated from 10Hz to 200Hz, and the humidity remains at 60%RH. In this process, data is continuously collected and the state recognition result output by the model is recorded, the response delay time of the model from the occurrence of abnormal equipment state to accurate recognition is analyzed, and the real-time monitoring ability of the device to dynamic changes is evaluated. In addition, random noise is introduced in the data transmission path to simulate data packet loss in a weak 5G signal environment, test the anti-interference performance of the wireless communication module, observe whether the communication unit of the multi-core control board can ensure the integrity of the data through the retransmission mechanism, and the recognition stability of the model under noise interference.

[0084] In the long-term stability test phase, the device is placed in a continuously running test environment with a temperature of 40°C, a humidity of 70% RH, intermittent vibration and partial discharge, and is continuously monitored for 30 days. The power supply stability of the front-end monitoring module is recorded every day, and the power supply stability is obtained through the log data of the power management module. At the same time, the CPU occupancy rate and memory usage of the multi-core control board are recorded. Static performance retesting is performed regularly, with a retesting interval of 24 hours. The identification accuracy trend at different time points is compared. At the same time, the physical state of each module is observed, the interface connection is checked for looseness, and the sensor is checked for drift to evaluate the reliability of the device in long-term live monitoring scenarios.

[0085] In the data processing and result analysis stage, the data analysis software of the upper computer module is used to process the test data offline. First, the quality of the original data of the front-end monitoring module is verified through time-domain waveform analysis to ensure that there is no obvious distortion or packet loss. Second, the intermediate features of the double-path convolution module are extracted, such as the feature dimension of 1*2*2 output by the upper channel pooling layer and the feature dimension of 1*2*2 output by the second pooling layer of the lower channel. The differences before and after feature fusion are displayed through a visualization tool to verify the improvement effect of the element-by-element multiplication method on feature complementarity. Finally, the precision, recall and F1 score of the model are calculated using a confusion matrix, and compared with traditional single convolution models, including models using only the upper channel or lower channel structure, to quantify the performance advantages of the double-path fusion model. For the response delay time recorded in dynamic testing, the physical delay of data transmission needs to be deducted, which is obtained from the communication log of the 5G module. The actual recognition delay of the model is obtained to evaluate whether it meets the requirements of real-time monitoring, which usually requires an actual recognition delay ≤100ms.

[0086] Static performance test results:

[0087] Parameter Type Accuracy in Normal State Accuracy in Abnormal State Average Accuracy Time Consumption (ms) for Single Group of Data Accuracy Standard Deviation Dual-Channel Fusion (Vibration + Partial Discharge) 99.2% 98.8% 99.0% 45 0.3% Single Parameter (Vibration Only) 87.5% 86.3% 86.9% 32 1.2% Single Parameter (Partial Discharge Only) 90.1% 99.7% 89.9% 38 0.8%

[0088] From the static performance test results, the accuracy of the dual-path fusion (vibration + partial discharge) monitoring method under normal and abnormal conditions was 99.2% and 98.8%, respectively, with an average accuracy of 99.0%, which was significantly higher than that of single parameter monitoring. Among them, the average accuracy of only vibration parameters was 86.9%, and the average accuracy of only partial discharge parameters was 89.9%, which was 12.1 and 9.1 percentage points higher than that of the two, respectively. In terms of time consumption, the time consumption of dual-path fusion for processing a single group of data was 45 ms, which was slightly higher than that of single parameter but still at a low level, fully meeting the basic requirements of real-time monitoring. In addition, the standard deviation of the accuracy of dual-path fusion was only 0.3%, which was much lower than that of single parameter monitoring, indicating that the recognition result was more stable, fully embodying the superiority of dual-path fusion in static scene for equipment state recognition.

[0089] Dynamic response test results:

[0090] Parameter Type Average Response Delay (ms) Accuracy in Noise Environment Accuracy in Different Noise Intensities (Weak Noise) Accuracy in Different Noise Intensities (Medium Noise) Accuracy in Different Noise Intensities (Strong Noise) Dual-Channel Fusion (Vibration + Partial Discharge) 85 96.5% 99.1% 97.3% 94.2% Single Parameter (Vibration Only) 120 82.3% 88.5% 83.2% 75.6% Single Parameter (Partial Discharge Only) 105 88.7% 92.3% 89.1% 82.5%

[0091] The dynamic response test results showed that the average response delay of dual-path fusion monitoring was 85 ms, meeting the real-time monitoring requirement (≤100 ms), while the response delays of single parameter monitoring were all more than 100 ms, among which only vibration parameters were 120 ms and only partial discharge parameters were 105 ms. In the simulated weak signal environment with different intensity random noise, the accuracy of dual-path fusion remained at a high level, even in the strong noise environment, the accuracy was still 94.2%, which was much higher than that of single parameter, 75.6% (only vibration) and 82.5% (only partial discharge), indicating that dual-path fusion effectively improved the anti-interference ability and response speed in dynamic scene through feature complementation.

[0092] Long-term stability test results:

[0093] Average Accuracy CPU Occupancy Rate Memory Usage Rate Power Supply Stability Accuracy for 10kV Switch Cabinet Accuracy for 35kV Cable Terminal Accuracy for 110kV Transformer Day 1 99.0% 35% 28% 100% 99.2% 98.9% 98.8% Day 5 98.9% 35% 28% 100% 99.1% 98.8% 98.7% Day 10 98.9% 36% 29% 100% 99.0% 98.8% 98.6% Day 15 98.8% 36% 29% 100% 98.9% 98.7% 98.5% Day 20 98.7% 37% 30% 100% 98.8% 98.6% 98.5% Day 25 98.6% 37% 30% 99.9% 98.7% 98.5% 98.4% Day 30 98.5% 38% 31% 99.8% 98.6% 98.4% 98.3%

[0094] The long-term stability test results showed that during the 30-day continuous operation of the dual-path fusion model, the average accuracy decreased from the initial 99.0% to only 98.5%, with a decrease of only 0.5 percentage points, and remained stable overall. In terms of device type, the accuracy of 10kV switchgear, 35kV cable terminal and 110kV transformer decreased by less than 1 percentage point, indicating that the device had good long-term stability for different types of high-voltage equipment. During this period, the CPU occupancy rate of the multi-core control board increased from 35% to 38%, and the memory usage rate increased from 28% to 31%, both of which were within a reasonable range and did not appear resource overload. The power supply stability remained above 99.8%, the interface connection of each module was firm, and the sensor had no obvious drift, fully verifying the reliable operation ability of the device in the long-term live monitoring scene.

[0095] Model performance comparison results:

[0096] Model Type Precision Recall F1 Score Missed Detection Rate in Normal State False Alarm Rate in Abnormal State Element-wise Addition Fusion F1 Score Concatenate Fusion F1 Score SVM Model F1 Score Random Forest Model F1 Score Dual-Channel Fusion Model 99.1% 98.9% 99.0% 0.3% 0.5% 97.5% 98.2% 85.3% 87.6% Only Upper Channel Model 88.6% 87.9% 88.2% 5.2% 6.8% - - - - Only Lower Channel Model 91.2% 90.8% 91.0% 3.8% 5.1% - - - -

[0097] The model performance comparison results show that the precision, recall and F1 score of the dual-path fusion model are 99.1%, 98.9% and 99.0% respectively, which are significantly better than the upper channel model (88.6%, 87.9%, 88.2%) and the lower channel model (91.2%, 90.8%, 91.0%). Among them, F1 score as a comprehensive evaluation index, dual-path fusion model is 10.8 percentage points higher than the upper channel model, and 8.0 percentage points higher than the lower channel model. Compared with other feature fusion methods, the F1 score of element-wise product fusion is higher than that of element-wise addition fusion (97.5%) and concatenate fusion (98.2%), which fully reflects the superiority of element-wise product fusion.

[0098] At the same time, the false negative rate of the dual-path fusion model under normal state is only 0.3%, and the false positive rate under abnormal state is only 0.5%, which is much lower than the single channel model, which can effectively reduce the loss caused by false detection and false alarm. Compared with traditional machine learning models such as SVM and random forest, the F1 score of the dual-path fusion model is 13.7 and 11.4 percentage points higher than that of the single channel model, indicating that its performance advantage in device state recognition task is obvious.

[0099] The model performance comparison data of different feature dimensions are as follows:

[0100] Feature Dimension Dual-Channel Fusion F1 Score Only Upper Channel F1 Score Only Lower Channel F1 Score Calculation Time Consumption (ms) 1102 97.2% 85.3% 88.5% 32 1302 98.5% 87.1% 90.2% 38 1502 99.0% 88.2% 91.0% 45 11002 99.1% 88.5% 91.3% 62

[0101] From the feature dimension, dual-path fusion is better than single channel in each dimension, and the performance improves obviously with the increase of dimension, but when the dimension exceeds 1*50*2, the improvement rate tends to be slow, while the calculation time increases significantly. Therefore, 1*50*2 is selected as the input dimension, which can balance the performance and efficiency.

[0102] The model performance data under different training iteration numbers are as follows:

[0103] Iteration Number Dual-Channel Fusion F1 Score Only Upper Channel F1 Score Only Lower Channel F1 Score Overfitting Risk (Validation Set Error) 100 95.3% 82.1% 86.5% 1.2% 500 98.2% 86.3% 89.7% 1.5% 1000 99.0% 88.2% 91.0% 1.8% 2000 99.1% 88.4% 91.2% 3.2%

[0104] With the increase of training iteration number, the performance of dual-path fusion model gradually improves, and the F1 score reaches 99.0% at 1000 iterations. Continue to iterate the performance improvement is limited, but the risk of overfitting increases. Therefore, 1000 iterations is the optimal training number, which takes into account the performance and generalization ability.

[0105] The model performance data under different noise filtering algorithms are as follows:

[0106] Filtering Algorithm Dual-Channel Fusion F1 Score Accuracy in Noise Environment Proportion of Increased Calculation Time No Filtering 96.5% 92.3% 0% Mean Filtering 97.8% 94.5% 10% Wavelet Denoising 98.5% 96.2% 25% Kalman Filtering 99.0% 97.5% 30%

[0107] After introducing the noise filtering algorithm, the performance of the dual-path fusion model is further improved, among which the Kalman filter has the best effect, with an F1 score of 99.0%, but the calculation time increases by 30%. In practical applications, appropriate algorithms can be selected according to real-time requirements, and Kalman filter is preferred in strong noise environment.

[0108] To avoid the influence of electromagnetic interference on monitoring data, shielded cables are used for the signal lines of the front-end monitoring module, and electromagnetic shielding covers are installed at the connection parts of the multi-core control board and the 5G module to ensure the data acquisition accuracy in strong electromagnetic environment.

[0109] In terms of the specific composition of the test device, the material of the closed test cabin is selected to be an alloy material that is resistant to high temperature and corrosion, with a thickness of 5 mm to ensure good sealing and structural strength, preventing external interference on the cabin environment parameters. The temperature control system uses PID regulation, and the temperature is raised and lowered through heating pipes and refrigeration compressors. The temperature sensor is a platinum resistance installed at different positions in the cabin to ensure uniformity of temperature monitoring. The humidity adjustment system consists of an ultrasonic humidifier and a dehumidifier, and the humidity sensor is a capacitive type with a response time of ≤2 s, ensuring the rapidity and accuracy of humidity adjustment. The vibration excitation system uses an electromagnetic vibration table, which is fixed to the test object through a rigid connection to ensure effective transmission of vibration energy. The partial discharge simulation system includes a high-voltage power supply and a discharge model, which can be replaced to simulate different types of partial discharge phenomena.

[0110] The acceleration sensor used in the vibration monitoring module of the front-end monitoring module is piezoelectric, with a sensitivity of 100 mV / g and a frequency response range of 1 Hz to 10 kHz, meeting the requirements of equipment vibration signal acquisition. The sensor is installed using a magnetic attraction fixing method, which ensures the firmness of the installation and facilitates the disassembly and replacement of the monitoring points. The humidity and temperature monitoring module is integrated in the same package, using a patch design, which is small in size and can be installed in the narrow space of the equipment. The sensor is attached to the surface of the equipment through heat-conducting silicone, improving the response speed of temperature and humidity measurement. The coupling capacitance of the partial discharge monitoring module is 100 pF, and the withstand voltage value is 1.5 times the maximum working voltage of the equipment, ensuring safe operation in a high-voltage environment. The module has a filter circuit inside, which can effectively suppress power frequency interference.

[0111] The multi-core control board adopts an industrial-grade processor with 4 cores and a main frequency of 1.8 GHz, and has powerful data processing capability, and can simultaneously run data acquisition, feature extraction and model inference tasks. The control board is integrated with a 16-bit A / D converter, and the sampling rate can reach 1 MHz, ensuring the accuracy and speed of analog signal conversion. The 5G functional module supports SA / NSA dual-mode networking, the data transmission rate is ≥100 Mbps, the delay is ≤20 ms, which meets the requirements of real-time data transmission, and the module is built-in with a SIM card slot, which can realize wide-area network connection through an Internet of Things card.

[0112] In terms of storage and management of test data, the host computer adopts a server-level computer, which is configured with an Intel Xeon processor, 32 GB of memory, and a 2 TB solid state disk, and is installed with a database management system, which can realize real-time storage and query of test data. The data analysis software is developed based on Python, integrates signal processing, machine learning algorithms and visualization tools, and can perform filtering, noise reduction and other preprocessing on raw data, dimensionality reduction, clustering analysis on feature data, and statistical analysis and evaluation on model output results.

[0113] Different test objects adopt differentiated monitoring point arrangement strategies: the vibration monitoring module of the 10kV switch cabinet is mainly installed at the connection between the circuit breaker operating mechanism and the bus, and the partial discharge monitoring module is connected through the bushing coupling capacitor; the temperature and humidity monitoring module of the 35kV cable terminal is installed at the transition part between the terminal surface and the cable body, and the vibration monitoring module is fixed to the metal shielding layer of the cable terminal through a clamp; the partial discharge monitoring module of the 110kV transformer is connected to the high-voltage side and low-voltage side bushings, respectively, and the vibration monitoring module is arranged at the four corners of the oil tank wall and the core grounding, ensuring comprehensive capture of state signals at different parts of the equipment. The correspondence between environmental parameters and monitoring data of each type of equipment needs to be recorded separately, and an associated database of equipment type-parameter combination-identification result is established, which provides a basis for subsequent model adaptation and optimization on different equipment.

[0114] The advantages of the embodiment of the present application are as follows: firstly, different features of multi-parameter data are learned through the double-channel convolution module, and the feature complementation is realized through the element-by-element multiplication of the feature fusion module, so that the robustness and detail extraction capability of the model are enhanced, the focusing capability of the model on key information is improved, and the problem of insufficient feature mining of the traditional model is solved; secondly, through the integration of four types of monitoring data of vibration, temperature, humidity and partial discharge, multi-parameter fusion analysis and multi-dimensional comprehensive diagnosis are realized, and the limitation of single parameter monitoring is avoided; thirdly, the multi-core control board of the present application has edge technology function, can analyze and judge the data on site, reduces the data transmission pressure, and improves the response speed; fourthly, the modular design is adopted, the deployment cost of similar functions is reduced, 5G remote communication and local interaction are supported, and different application scenarios can be flexibly adapted; fifthly, through multi-module cooperation and anti-interference design, the false alarm and missed alarm problems are reduced, the long-term operation performance is stable, and the present application is suitable for complex power environment. In summary, the live monitoring model and device of the double-path fusion perception of the present application realize efficient and accurate monitoring and diagnosis of the state of power equipment through multi-parameter data fusion and double-path cooperative recognition, and have important application value.

[0115] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A live-line monitoring model based on dual-path fusion sensing, characterized in that, The live-line monitoring model based on dual-path fusion sensing includes: A dual-channel convolutional module, comprising an input layer, an upper channel convolutional layer, and a lower channel convolutional layer; The feature fusion module is used to perform feature fusion on the output data of the upper channel convolutional layer and the output data of the lower channel convolutional layer; The global average pooling-softmax module is used to process the fused features and output the device status recognition results.

2. The live-line monitoring model based on dual-path fusion sensing according to claim 1, characterized in that, The input layer is used to receive two types of time-domain data of charged parameters, and the dimension of the time-domain data of charged parameters is 1*50*2; The upper channel convolutional layer includes two sub-modules and one pooling layer; The lower channel convolutional layer comprises four sub-modules and two pooling layers.

3. The live-line monitoring model based on dual-path fusion sensing according to claim 2, characterized in that, Each sub-module of the upper channel convolutional layer consists of a convolutional layer, an AdaBN layer, and an activation layer. The first submodule takes the time-domain data from the input layer as input and outputs 1*30*2 dimensions. The second submodule takes the output data of the first submodule as input, and the output dimension is 1*10*2; The input to the pooling layer is the output data of the second submodule, and the output dimension is 1*2*2.

4. The live-line monitoring model based on dual-path fusion sensing according to claim 2, characterized in that, Each sub-module of the lower channel convolutional layer consists of a convolutional layer, an AdaBN layer, and an activation layer. The first submodule takes the time-domain data from the input layer as input and outputs 1*40*2 dimensions. The second submodule takes the output data of the first submodule as input, and the output dimension is 1*20*2; The input to the first pooling layer is the output data of the second submodule, and the output dimension is 1*2*2; The third submodule takes the output data of the first pooling layer as input and outputs data with a dimension of 1*30*2. The fourth submodule takes the output data of the third submodule as input, and the output dimension is 1*20*2. The input to the second pooling layer is the output data of the fourth sub-module, and the output dimension is 1*2*2.

5. The live-line monitoring model based on dual-path fusion sensing according to claim 1, characterized in that, The specific method for feature fusion of the output data of the upper channel convolutional layer and the output data of the lower channel convolutional layer is as follows: perform element-wise multiplication of the two feature data of the same dimension to form a comprehensive feature.

6. The live-line monitoring model based on dual-path fusion sensing according to any one of claims 1 to 5, characterized in that, The specific process by which the global average pooling-softmax module processes the fused features and outputs the device status recognition result includes: The fused features are processed by global average pooling, which transforms the high-dimensional feature matrix into a fixed-length feature vector, thereby achieving feature dimensionality reduction while retaining key information. The feature vector after dimensionality reduction is processed by the softmax function, which maps the feature vector to the probability distribution of device status and outputs the recognition result of device status.

7. A live-line monitoring device with dual-channel fusion sensing, characterized in that, The dual-path fusion sensing live-line monitoring device includes a multi-core control board, a front-end monitoring module, a host computer module, and a wireless communication module. The multi-core control board carries a dual-path fusion sensing live-line monitoring model according to any one of claims 1 to 6, including a data sorting unit and a communication unit. The data sorting unit is used to sort out the two types of live-line parameter time-domain data input to the dual-path fusion sensing live-line monitoring model, and the communication unit is used to enable communication with other modules. The front-end monitoring module is connected to the multi-core control board via a communication bus and includes a vibration monitoring module, a humidity monitoring module, a temperature monitoring module, and a partial discharge monitoring module. Each front-end monitoring module is equipped with a power supply interface. The host computer module includes an LCD display module and a parsing module connected to the LCD display module. The parsing module is connected to the multi-core control board via a network port. The host computer module is equipped with a TYPE-C power supply interface to power the LCD display module and the parsing module. The wireless communication module is a 5G functional module.

8. The live-line monitoring device with dual-channel fusion sensing according to claim 7, characterized in that, The vibration monitoring module includes an RS422 interface and a parsing unit connected to the RS422 interface; the parsing unit is connected to the multi-core control board via a network port; the vibration monitoring module is provided with a three-pin power supply interface, providing DC 5V voltage to the RS422 interface and DC 3.3V voltage to the parsing unit.

9. The live-line monitoring device with dual-channel fusion sensing according to claim 7, characterized in that, The humidity monitoring module includes a humidity sensor, a data acquisition unit, and a data analysis unit connected in sequence; the data analysis unit is connected to the multi-core control board via a network port; the humidity monitoring module is equipped with a TYPE-C power supply interface to power the humidity sensor, data acquisition unit, and data analysis unit.

10. The live-line monitoring device with dual-channel fusion sensing according to claim 7, characterized in that, The temperature monitoring module includes a temperature sensor, a data acquisition unit, and a data analysis unit connected in sequence; the data analysis unit is connected to the multi-core control board via a network port; the temperature monitoring module is equipped with a TYPE-C power supply interface to power the temperature sensor, data acquisition unit, and data analysis unit.

11. The live-line monitoring device with dual-channel fusion sensing according to any one of claims 7 to 10, characterized in that, The partial discharge monitoring module includes a partial discharge sensor, a data acquisition unit, and a data analysis unit connected in sequence; the data analysis unit is connected to the multi-core control board via a network port; the partial discharge monitoring module is equipped with a TYPE-C power supply interface to power the partial discharge sensor, the data acquisition unit, and the data analysis unit.

12. The live-line monitoring device with dual-channel fusion sensing according to claim 7, characterized in that, The communication bus is the MODBUS bus; The multi-core control board integrates an AC-DC module and a power management module. The AC-DC module is used for power supply and voltage conversion, and the power management module is used for power distribution and management. The 5G functional module includes a 5G communication module, a parsing unit, and a power management unit. The parsing unit is connected to the multi-core control board via a USB interface and performs data interaction. The power management unit is powered via a TYPE-C power supply interface and outputs DC-3.3V voltage to the parsing unit.

Images