A substation mechanical instrument low-power intelligent identification and remote monitoring method

Abstract

The application discloses a kind of low-power intelligent identification and remote monitoring method of mechanical instrument of transformer substation, belong to the technical field of transformer substation instrument monitoring, the method includes: after power-on initialization, hardware self-check is executed;Collect ambient light, vibration intensity and equipment posture parameters;According to environmental parameter, camera exposure, light compensation lamp and polarization filter are executed anti-interference image acquisition according to adaptive adjustment;Original image is carried out illumination correction, vibration blur recovery, defogging and contrast enhancement preprocessing;The dial area is detected by improving YOLOv8 network;Actual reading is converted by using improved ResNet50 network to identify pointer angle;The effectiveness of reading is checked;Data is uploaded to remote monitoring platform by BLE5.0 or NB-IoT dual-mode communication.This application can realize the automatic identification and remote monitoring of mechanical instrument under outdoor complex environment, the identification accuracy is high, the endurance time reaches 3 years, reduces the transformation cost.

Description

Technical Field

[0001] This invention belongs to the field of electrical engineering technology, specifically, it relates to a method for low-power intelligent identification and remote monitoring of mechanical instruments in substations. Background Technology

[0002] As a critical node in power grid operation, substations rely heavily on various mechanical monitoring instruments to monitor equipment status. Taking the State Grid Tianshui Power Supply Company as an example, its jurisdiction includes 34 110kV and above substations, with over 1,000 mechanical instruments installed, accounting for more than 70% of all monitoring instruments. These instruments are primarily used to monitor key operating parameters in real time, such as SF6 gas pressure, transformer oil temperature and level, hydraulic mechanism pressure, and oil pressure in oil-filled cables, forming the first line of defense for ensuring the safe and stable operation of power equipment.

[0003] However, the readings of these mechanical instruments still heavily rely on regular on-site inspections by maintenance personnel, manual visual interpretation, and manual recording. This traditional maintenance model is increasingly showing its limitations in today's digital and intelligent power grid development: First, manual inspections are prone to subjective errors, omissions, and misrecording, making it difficult to effectively guarantee the authenticity and consistency of data, resulting in a lack of reliable data support for equipment operating status assessment; second, on-site inspections are limited by personnel allocation and inspection cycles, making real-time data collection and transmission impossible, and timely warnings of abnormal changes in equipment status are difficult, especially during extreme weather or at night, when inspection blind spots are large, posing safety hazards; third, manual recording methods are not conducive to structured data storage and subsequent analysis and utilization, failing to meet the requirements of intelligent maintenance for data integrity and timeliness; finally, the maintenance model relying solely on manual inspections is labor-intensive and inefficient, incompatible with the overall trend of power grid digital transformation.

[0004] Furthermore, traditional smart meter replacement solutions involve modifications to primary and secondary equipment, resulting in long construction periods, high costs, and the need for power outages, making them unsuitable for the large-scale digital transformation of existing meters. Therefore, exploring an efficient, economical, and reliable solution for the digital transformation of mechanical meters has become an urgent task for power grid companies to promote digital transformation and innovate their operation and maintenance models. Summary of the Invention

[0005] The purpose of this invention is to provide a low-power intelligent identification and remote monitoring method for mechanical instruments in substations, which mainly solves the technical problems of large subjective errors in data acquisition, inability to monitor in real time, low inspection efficiency, and high cost of digital transformation of existing mechanical instruments in the traditional manual inspection mode.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A method for low-power intelligent identification and remote monitoring of mechanical instruments in substations includes the following steps:

[0008] S1: After the device is powered on, it completes the initialization configuration of the embedded processor and provides a software running environment for hardware self-test;

[0009] S2 executes the whole machine hardware self-test process, generates a hardware status report, and triggers an alarm mechanism when an abnormality is detected;

[0010] S3 collects the current ambient light information, obtains the device installation attitude angle, calculates adaptive image acquisition parameters, and provides environmental reference data for subsequent anti-interference image acquisition;

[0011] S4, adjust the acquisition parameters based on environmental baseline data, perform adaptive anti-interference image acquisition, and obtain the original acquisition image containing the target dial;

[0012] S5 performs preprocessing on the original acquired image, outputs the enhanced target image, and extracts the coordinates and rotation angle of the outer rectangle of the dial based on the target image, and crops the dial image.

[0013] S6: Obtain the actual physical reading based on the dial image and perform validity verification, then store the valid results;

[0014] S7: Transmit the valid results remotely with low power and determine whether the device is in continuous operation. If yes, return to step S3 to perform the acquisition and recognition task in a loop. If no, enter the low-power sleep mode and wait for the next wake-up timer trigger.

[0015] Furthermore, in this invention, the specific method of step S2 is as follows:

[0016] S21 sends a self-test command to the industrial camera module through the GPIO interface, reads the camera's chip temperature, exposure parameter register value and image sensor dead pixel detection results, and verifies whether the camera is working properly.

[0017] S22 reads the illuminance lux value and color temperature K value of the ambient light sensor through the I2C bus, reads the raw triaxial acceleration data of the accelerometer and calculates the composite acceleration amplitude, and determines whether the sensor is in a working ready state.

[0018] S23, query the RF power calibration parameters of the Bluetooth Low Energy BLE5.0 module and the signal strength RSRP value of the NB-IoT communication module through the SPI interface to verify the link quality of the wireless communication module;

[0019] S24 monitors the terminal voltage and remaining power percentage of the lithium battery through the ADC sampling channel. When the power is lower than a preset threshold, a low power alarm is triggered and the subsequent sampling frequency strategy is adjusted.

[0020] Furthermore, in this invention, the acquisition of environmental reference data in step S3 includes: continuously sampling and averaging ambient light intensity using an ambient light sensor; sampling and calculating the root mean square value of triaxial acceleration using a triaxial accelerometer as a vibration intensity index; and obtaining the heading angle of the device relative to the geomagnetic north pole using a magnetometer to provide an attitude reference for subsequent image rotation correction.

[0021] Furthermore, in this invention, in step S4, the adjustment of acquisition parameters includes adjusting the exposure time, gain value, and white balance parameters of the industrial camera, activating the coordinated control of the polarizing filter and the anti-glare shield, and triggering the LED fill light group to dynamically adjust the illumination brightness according to the ambient darkness; wherein, the polarizing filter dynamically adjusts the rotation angle according to the polarization state of the incident light to minimize glare interference; the angle between the normal direction of the anti-glare shield and the direction of solar incidence is controlled within the range of 75° to 105°.

[0022] Furthermore, in this invention, in step S5, the image preprocessing includes illumination correction based on histogram equalization, vibration blur restoration based on Wiener filtering, dehazing based on dark channel prior, and image sharpening based on adaptive contrast enhancement. Illumination correction employs an adaptive histogram equalization algorithm, dividing the image into 16 sub-regions (4×4) and performing histogram stretching on each, eliminating the difference in brightness in the dial area caused by uneven illumination. Vibration blur restoration uses a Wiener filtering algorithm, with the noise power spectral density parameter of the Wiener filter dynamically adjusted according to the collected vibration intensity index. Dehazing uses a dark channel prior algorithm to estimate atmospheric light values ​​and calculate the transmittance map, restoring a clear image. Contrast enhancement uses an adaptive CLAHE algorithm.

[0023] Furthermore, in this invention, in step S5, a dial target detection algorithm based on an improved YOLOv8 network is used to locate the dial region and extract the coordinates and rotation angle of the dial's bounding rectangle. The improved YOLOv8 network uses a CSPDarknet structure, includes three downsampling stages, and outputs feature maps with resolutions of 1 / 8, 1 / 16, and 1 / 32 of the original image. The detection head employs a decoupled design, with independent classification and regression branches. During training, Mosaic data augmentation and label assignment strategies are used, and the loss function includes CIoU Loss for bounding box regression and BCE Loss for class classification.

[0024] Furthermore, in this invention, in step S6, the actual physical reading adopts a reading recognition model based on an improved ResNet50 network. This model connects an attention mechanism module SE-Net after the feature map of ResNet50, and adaptively adjusts the channel weights through a Squeeze-and-Excitation structure. The final fully connected layer of the model outputs a 128-dimensional feature vector, and outputs the predicted value of the pointer angle through a regression head. The root mean square error of the angle regression is no greater than 1.5°. The network output is the actual reading after range conversion.

[0025] Furthermore, in this invention, the method for data validity verification in step S6 is as follows:

[0026] Establish a database of reasonable ranges for meter readings, storing the upper and lower limits of the range, alarm thresholds, and historical statistical characteristics of various meters;

[0027] For the identified reading, it is determined whether it meets the set upper and lower limit basic constraints. If the reading is outside the reasonable range, the deviation between the reading and the historical average of the same type of meter is calculated. When the deviation exceeds 3 times the historical standard deviation, it is judged as abnormal data. Abnormal data is stored together with the corresponding original image and marked with an abnormality marker.

[0028] Furthermore, in this invention, in step S7, the low-power remote transmission adopts a BLE5.0+NB-IoT dual-mode communication architecture: when the device is within a 50-meter radius of the substation, the BLE5.0 communication mode is used first to establish a connection with the local operation and maintenance terminal; when the device is outside the BLE coverage area, it switches to the NB-IoT communication mode; the data packet is sent using the CoAP protocol, and the payload is encoded in JSON format.

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] (1) This invention breaks through the technical bottleneck of traditional visual meter reading devices being only suitable for stable indoor working conditions by constructing an integrated hardware and software anti-interference perception system. By integrating multiple optimization methods such as polarization filtering, intelligent supplementary lighting, illumination correction, vibration compensation and attention mechanism algorithms, the accuracy of meter recognition in complex outdoor environments is improved. Compared with traditional indoor meter reading solutions, the recognition stability is greatly improved, and the technical problems such as outdoor illumination changes, vibration interference and reflection blockage are effectively solved.

[0031] (2) This invention adopts a passive power supply solution that combines a wide-temperature-range long-life energy storage battery with high-efficiency solar charging. With the help of a tiered energy scheduling and power consumption control strategy, the device can achieve long-term stable operation in scenarios without external power supply. Under the condition of collecting data 12 times a day, the device can last for more than 3 years, which completely solves the power supply bottleneck of digital transformation of existing meters under outdoor power supply conditions and is suitable for the long-term operation needs of unattended substations.

[0032] (3) This invention adopts a dual-mode low-power communication architecture of BLE5.0 + NB-IoT, combined with a dynamic sleep mechanism to achieve extreme power consumption optimization. The device automatically switches the communication mode according to the communication distance. In the local operation and maintenance scenario, BLE is used to achieve fast data interaction, and in the remote transmission scenario, NB-IoT is used to achieve low-power wide area network coverage, so that the average operating current of the whole machine is reduced to below 50μA, and the device's battery life is extended to the maximum extent while ensuring communication stability.

[0033] (4) This invention adopts an external, non-destructive, lightweight retrofit mode, which abandons the high-cost solution of replacing the entire smart meter in the traditional way. The equipment is directly installed on the housing of the existing mechanical meter through stainless steel clamps. There is no need to modify the primary equipment or secondary circuit, and no need to lay power lines and communication lines. The construction cycle is shortened to less than 15 minutes per meter, which greatly reduces the retrofit cost of digital upgrading of substation meters and the number of households affected by power outages. It is suitable for the batch digital retrofit of a large number of existing meters in the entire station. Attached Figure Description

[0034] Figure 1 This is a schematic diagram of the method flow of the present invention.

[0035] Figure 2 This is a curve comparing the accuracy of instrument recognition under different ambient light intensities in an embodiment of the present invention. Detailed Implementation

[0036] The present invention will be further described below with reference to the accompanying drawings and embodiments. The embodiments of the present invention include, but are not limited to, the following embodiments.

[0037] This invention discloses a low-power intelligent identification and remote monitoring method for mechanical instruments in substations. This method is implemented based on an embedded hardware system. The core hardware components of this system include an embedded processing unit, an industrial camera module, an environmental sensing sensor array, a passive long-term power supply module, and a low-power communication module. The embedded processing unit uses an ARM Cortex-A series processor as its core, with a main frequency of up to 1.5GHz, and is equipped with an 8GB eMMC storage module for storing system firmware, neural network model parameters, and historical acquisition data. The industrial camera module uses a global shutter CMOS image sensor with 5 million effective pixels, providing high-speed image acquisition capabilities. The lens focal length is set to 6mm, and the field of view covers 60 degrees, suitable for imaging and acquiring various mechanical instruments both indoors and outdoors in substations. The environmental sensing sensor array includes an ambient light sensor, an acceleration and vibration sensor, and a magnetometer, used to collect ambient light intensity, vibration intensity, and equipment installation attitude angle, respectively, providing environmental reference data for adaptive image acquisition.

[0038] The passive long-lasting power supply module adopts a combination of a wide-temperature-range lithium thionyl chloride battery pack and a high-efficiency monocrystalline silicon solar charging panel. The battery pack has a nominal capacity of 30Ah, a discharge cut-off voltage of 3.0V, and an operating temperature range of -40℃ to +85℃, enabling it to adapt to the harsh outdoor environment of substations. The solar charging panel has a peak power of 5W and uses the MPPT (Maximum Power Point Tracking) algorithm, achieving a charging conversion efficiency of no less than 22%, ensuring effective energy replenishment even under low-light conditions. The energy management unit dynamically adjusts the device's operating mode based on the remaining battery power, achieving long-term stable operation.

[0039] The low-power communication module supports a dual-mode communication architecture of BLE5.0 and NB-IoT. The BLE5.0 module has a transmit power of 0dBm, a connection interval of 100ms, and a data transmission rate of at least 1Mbps, suitable for rapid data exchange with local maintenance terminals. The NB-IoT module uses the Banding3 band, with an upload rate of at least 250kbps and a transmit power of 23dBm, suitable for remote data transmission. The communication module automatically switches communication modes based on the distance between the device and the monitoring platform, achieving a balance between power consumption and communication quality.

[0040] like Figure 1 As shown, after the system starts, the processor first loads the pre-set system firmware and neural network model parameters, and initializes the various peripheral driver modules, including the industrial camera driver, environmental sensor driver, storage driver, and communication driver. After initialization, the system enters the hardware self-test phase to perform a comprehensive check on the overall hardware status.

[0041] The hardware self-test process includes the following specific operations: Sending a self-test command to the industrial camera module via the GPIO interface to read the camera's chip temperature, exposure parameter register values, and image sensor dead pixel detection results to verify whether the camera is functioning normally; reading the illuminance (lux) and color temperature (K) values ​​from the ambient light sensor via the I2C bus, reading the raw triaxial acceleration data from the accelerometer and calculating the composite acceleration amplitude to determine if the sensor is in a ready-to-work state; querying the RF power calibration parameters of the Bluetooth Low Energy (BLE) 5.0 module and the signal strength (RSRP) value of the NB-IoT communication module via the SPI interface to verify the link quality of the wireless communication module; monitoring the terminal voltage and remaining battery percentage of the lithium battery via the ADC sampling channel, triggering a low-power alarm and adjusting the subsequent sampling frequency strategy when the battery level is below a preset threshold.

[0042] During the environmental parameter acquisition phase, the system continuously samples and averages ambient light intensity using an ambient light sensor, with a sampling frequency set to 10Hz. The vibration intensity index is obtained by collecting 32 samples within a 100ms time window using a triaxial accelerometer, and calculating the root mean square value of the triaxial acceleration. The calculation formula is as follows:

[0043]

[0044] In the formula, , , These are the j-th sample values ​​of the triaxial acceleration. This represents the number of samples.

[0045] The equipment's installation attitude angle is obtained by measuring the heading angle relative to the Earth's magnetic north pole using a magnetometer. This provides an attitude reference for subsequent image rotation correction. Heading angle The calculation formula is:

[0046]

[0047] In the formula, These represent the geomagnetic components of the magnetometer along the X and Y axes, respectively. During the adaptive anti-interference image acquisition phase, the system adjusts the image acquisition based on ambient light intensity. The camera exposure time can be obtained by looking up the table. The initial value is adjustable from 0.1ms to 100ms. Exposure time With ambient light intensity The relationship is obtained by looking up a table, as shown in Table 1.

[0048] (lux) (ms) ＜50 80-100 50-200 20-80 200-500 5-20 500-1000 2-5 ＞1000 0.1-2

[0049] When the ambient light intensity is below 200 lux, the LED fill light group is triggered to turn on the fill light with PWM duty cycle.

[0050] The polarizing filter dynamically adjusts its rotation angle according to the polarization state of the incident light to minimize glare interference. The angle between the normal direction of the anti-glare shield and the direction of sunlight incidence is controlled within the range of 75° to 105°, effectively blocking direct sunlight from interfering with the dial.

[0051] In the image preprocessing stage, the system performs multiple enhancement processes on the original acquired image. Illumination correction uses an adaptive histogram equalization algorithm, dividing the image into 16 sub-regions of 4×4 and stretching the histograms of each sub-region to eliminate the differences in brightness and darkness in the dial area caused by uneven illumination.

[0052] Vibration fuzzy reconstruction employs the Wiener filtering algorithm, with the noise power spectral density parameter dynamically adjusted based on the vibration intensity index. The formula for calculating the noise power spectral density parameter is as follows:

[0053]

[0054] In the formula, The preset scaling factor is set to 0.5 in this implementation. The transfer function of the Wiener filter... as follows:

[0055]

[0056] In the formula, Represents a fuzzy transfer function. For its conjugate, This represents the signal power spectral density.

[0057] Dehazing employs a dark channel prior algorithm to estimate atmospheric light values ​​and calculate a transmittance map, thereby restoring a clear image. Transmittance The calculation formula is as follows:

[0058]

[0059] In the formula, This represents the intensity value of the image in channel c. Atmospheric light value, The atomization factor is set to 0.95 in this embodiment. For The local area centered on.

[0060] Contrast enhancement employs an adaptive CLAHE algorithm, with the clipLimit parameter set to 2.5 and the tileGridSize set to 8×8 pixels. The contrast limiting formula for the CLAHE algorithm is as follows:

[0061]

[0062] In the formula, The number of pixels per tile, represents the number of gray levels in the histogram.

[0063] In the dial target detection stage, the system uses an improved YOLOv8 network model to locate the dial region. The backbone network of this model adopts a CSPDarknet structure, containing three downsampling stages, with output feature map resolutions of 1 / 8, 1 / 16, and 1 / 32 of the original image, respectively. The detection head employs a decoupled design, with independent classification and regression branches. The network input image size is 640×640 pixels, the inference frame rate is no less than 25 FPS, and the dial detection confidence threshold is set to 0.7.

[0064] During the training phase, a Mosaic data augmentation strategy is employed, randomly stitching four images together to form a single training sample, thereby enhancing the model's generalization ability across different scenarios. The loss functions include CIoU Loss for bounding box regression and BCE Loss for class classification.

[0065] In the reading recognition stage, the system employs an improved ResNet50 network architecture for dial reading recognition. This network integrates an attention mechanism module, SE-Net, after the ResNet50 feature maps, adaptively adjusting channel weights through a Squeeze-and-Excitation structure. The final fully connected layer outputs a 128-dimensional feature vector, which, through a regression head, predicts the pointer angle, with a root mean square error (RMSE) of no more than 1.5°. The network output is then converted into the actual physical reading using a range conversion formula. The range conversion formula is as follows:

[0066]

[0067] In the formula, and The upper and lower limits of the meter's measurement range, and This refers to the physical range of the pointer angle.

[0068] During the data validity verification phase, the system establishes a database of reasonable ranges for meter readings, storing the upper and lower limits of the range, alarm thresholds, and historical statistical characteristics of various meters. For the identified readings... First, determine whether it satisfies The basic constraints are as follows: If the reading exceeds the reasonable range, the deviation of the reading from the historical average of similar meters is calculated. When the deviation exceeds three times the historical standard deviation, it is judged as abnormal data, and the abnormal data is stored together with the corresponding original image and marked with an anomaly flag.

[0069] During the remote data upload phase, the system encapsulates the verified reading data, timestamps, and device status information into a JSON data packet via a low-power communication module. The data packet contains the fields device_id, timestamp, readings, status, and battery_level.

[0070] When the device is within a 50-meter radius of the substation, it prioritizes using BLE 5.0 communication mode to establish a connection with the local maintenance terminal; when the device is outside the BLE coverage area, it switches to NB-IoT communication mode. Data packets are sent using the CoAP protocol to ensure low-power transmission.

[0071] The system determines whether the device is in a continuous operating state. If so, it returns to the environmental parameter acquisition phase and repeatedly executes the acquisition and identification tasks. Otherwise, it enters a low-power sleep mode and waits for the next wake-up timer. In sleep mode, the system power consumption is reduced to below 50μA, and only the RTC real-time clock and timer are maintained.

[0072] Taking an SF6 gas pressure gauge in a substation as an example, in actual operation, the SF6 gas pressure gauge typically has a range of 0 to 0.8 MPa and a pointer angle range of -60° to +60°. Before installation, maintenance personnel first use the arc-shaped clamps of an external non-destructive mounting bracket to fix the equipment to the instrument housing, and adjust the dual-axis universal joint to ensure the camera's line of sight is aligned with the center of the dial. After the system is powered on, it automatically completes a hardware self-test to verify that each module is functioning normally.

[0073] Once the system is operational, the ambient light sensor collects real-time data on the ambient lighting conditions. At midday in summer, when the light intensity can reach 80,000 lux, the system automatically adjusts the exposure time to 0.5 ms, turns off the LED fill light, and adjusts the polarizing filter angle to eliminate glare from the glass cover. In the late afternoon in winter, when the light intensity drops to 150 lux, the system automatically extends the exposure time to 50 ms and turns on the LED fill light group at 50% brightness. Figure 2 As shown, the curves display the comparison of instrument recognition accuracy under different ambient light intensities.

[0074] After image acquisition, the system first eliminates uneven illumination through adaptive histogram equalization, and then dynamically adjusts the Wiener filter parameters based on vibration data collected by the accelerometer. When strong vibrations are detected (such as vibrations generated by substation equipment startup), the system automatically increases the number of iterations in the deblurring process to ensure image clarity.

[0075] The preprocessed image is input into the YOLOv8 detection network. The network inference time is approximately 40ms, successfully locating the dial area and outputting the coordinates of the bounding rectangle. The system then crops the dial image based on the detection results and inputs it into an improved ResNet50 network for reading recognition. The network outputs a pointer angle of 12.3°, and the actual reading is calculated to be 0.65MPa using the range conversion formula.

[0076] The system query confirmed that 0.65 MPa is within the normal range (0.4-0.7 MPa) of the SF6 gas pressure gauge, and the deviation from the historical average is less than 2 standard deviations, thus determining it as valid data. The data is uploaded to the remote monitoring platform via the NB-IoT module. After receiving the data, the platform updates its database and triggers the corresponding operation and maintenance procedures.

[0077] During long-term operation, the system dynamically adjusts the sampling frequency based on the remaining battery power (SOC). When the SOC is above 80%, a high-frequency sampling mode is used with a sampling period of 15 minutes; when the SOC is between 50% and 80%, a standard sampling mode is used with a sampling period of 30 minutes; and when the SOC is below 50%, a low-power sampling mode is used, with the sampling period extended to 60 minutes. This adaptive energy management strategy ensures that the device can operate continuously for more than 3 years without external power supply.

[0078] In terms of spatial layout, the embedded processing unit serves as the core control hub, connected to the industrial camera module, environmental sensing sensor array, and storage module via a high-speed bus, forming the system's data acquisition and processing link. A passive, long-lasting power supply module connects to each hardware component through a power management circuit, providing a stable operating voltage for the entire unit. The communication module establishes a wireless connection with an external network via an antenna, enabling remote data transmission. An external, non-destructive mounting bracket secures the entire device to the existing mechanical instrument housing, eliminating the need to modify primary equipment or secondary circuits, achieving truly non-destructive and lightweight retrofitting.

[0079] This invention departs from traditional manual inspection as the sole means of data acquisition, instead deeply integrating it with intelligent visual recognition technology. By introducing adaptive anti-interference image acquisition, multi-level image enhancement processing, and deep learning target detection and reading recognition algorithms, this invention successfully incorporates existing mechanical instruments into the scope of digital monitoring. This not only solves the technical bottlenecks of low efficiency and poor data quality in manual inspection but also provides a practical and feasible solution for upgrading existing instruments in the digital transformation of the power grid. In actual industrial applications, the system demonstrates strong environmental adaptability, operating stably in the high-temperature, humid, and electromagnetically interference-prone outdoor environments of substations, ensuring the reliability and timeliness of data acquisition.

[0080] The above embodiments are merely one of the preferred embodiments of the present invention and should not be used to limit the scope of protection of the present invention. Any modifications or refinements made to the main design concept and spirit of the present invention that are not of substantial significance, but solve the same technical problem as the present invention, should be included within the scope of protection of the present invention.

Claims

1. A method for low-power intelligent identification and remote monitoring of mechanical instruments in substations, characterized in that, Includes the following steps: S1: After the device is powered on, it completes the initialization configuration of the embedded processor and provides a software running environment for hardware self-test; S2 executes the whole machine hardware self-test process, generates a hardware status report, and triggers an alarm mechanism when an abnormality is detected; S3 collects the current ambient light information, obtains the device installation attitude angle, calculates adaptive image acquisition parameters, and provides environmental reference data for subsequent anti-interference image acquisition; S4, adjust the acquisition parameters based on environmental baseline data, perform adaptive anti-interference image acquisition, and obtain the original acquisition image containing the target dial; S5 performs preprocessing on the original acquired image, outputs the enhanced target image, and extracts the coordinates and rotation angle of the outer rectangle of the dial based on the target image, and crops the dial image. S6: Obtain the actual physical reading based on the dial image and perform validity verification, then store the valid results; S7: Transmit the valid results remotely with low power and determine whether the device is in continuous operation. If yes, return to step S3 to perform the acquisition and recognition task in a loop. If no, enter the low-power sleep mode and wait for the next wake-up timer trigger.

2. The method for low-power intelligent identification and remote monitoring of mechanical instruments in substations according to claim 1, characterized in that, The specific method for step S2 is as follows: S21 sends a self-test command to the industrial camera module through the GPIO interface, reads the camera's chip temperature, exposure parameter register value and image sensor dead pixel detection results, and verifies whether the camera is working properly. S22 reads the illuminance lux value and color temperature K value of the ambient light sensor through the I2C bus, reads the raw triaxial acceleration data of the accelerometer and calculates the composite acceleration amplitude, and determines whether the sensor is in a working ready state. S23, query the RF power calibration parameters of the Bluetooth Low Energy BLE5.0 module and the signal strength RSRP value of the NB-IoT communication module through the SPI interface to verify the link quality of the wireless communication module; S24 monitors the terminal voltage and remaining power percentage of the lithium battery through the ADC sampling channel. When the power is lower than a preset threshold, a low power alarm is triggered and the subsequent sampling frequency strategy is adjusted.

3. The method for low-power intelligent identification and remote monitoring of mechanical instruments in substations according to claim 2, characterized in that, The acquisition of environmental reference data in step S3 includes: continuously sampling and averaging ambient light intensity using an ambient light sensor; sampling and calculating the root mean square value of triaxial acceleration using a triaxial accelerometer as a vibration intensity index; and obtaining the heading angle of the device relative to the geomagnetic north pole using a magnetometer to provide an attitude reference for subsequent image rotation correction.

4. The method for low-power intelligent identification and remote monitoring of mechanical instruments in substations according to claim 3, characterized in that, In step S4, the adjustment of acquisition parameters includes adjusting the exposure time, gain value, and white balance parameters of the industrial camera, activating the coordinated control of the polarizing filter and the anti-glare shield, and triggering the LED fill light group to dynamically adjust the illumination brightness according to the ambient darkness. Among them, the polarizing filter dynamically adjusts the rotation angle according to the polarization state of the incident light to minimize glare interference; the angle between the normal direction of the anti-glare shield and the direction of sunlight incidence is controlled within the range of 75° to 105°.

5. The method for low-power intelligent identification and remote monitoring of mechanical instruments in substations according to claim 4, characterized in that, In step S5, the image preprocessing includes illumination correction based on histogram equalization, vibration blur restoration based on Wiener filtering, dehazing based on dark channel prior, and image sharpening based on adaptive contrast enhancement. Illumination correction employs an adaptive histogram equalization algorithm, dividing the image into 16 4×4 sub-regions and performing histogram stretching on each, eliminating brightness differences in the dial area caused by uneven illumination. Vibration blur restoration uses a Wiener filtering algorithm, with the noise power spectral density parameter dynamically adjusted based on the collected vibration intensity index. Dehazing uses a dark channel prior algorithm to estimate atmospheric light values ​​and calculate the transmittance map, restoring a clear image. Contrast enhancement uses an adaptive CLAHE algorithm.

6. The method for low-power intelligent identification and remote monitoring of mechanical instruments in substations according to claim 5, characterized in that, In step S5, a dial target detection algorithm based on an improved YOLOv8 network is used to locate the dial region and extract the coordinates and rotation angle of the dial's bounding rectangle. The improved YOLOv8 network adopts a CSPDarknet structure, which includes three downsampling stages, and the output feature map resolutions are 1 / 8, 1 / 16, and 1 / 32 of the original image, respectively. The detection head adopts a decoupled head design, with the classification branch and regression branch set independently. During training, a Mosaic data augmentation and label assignment strategy is used, and the loss function includes CIoU Loss for bounding box regression and BCE Loss for class classification.

7. A method for low-power intelligent identification and remote monitoring of mechanical instruments in substations according to claim 6, characterized in that, In step S6, the actual physical reading adopts a reading recognition model based on an improved ResNet50 network. This model connects the attention mechanism module SE-Net after the feature map of ResNet50 and adaptively adjusts the channel weights through a Squeeze-and-Excitation structure. The final fully connected layer of the model outputs a 128-dimensional feature vector, and the predicted value of the pointer angle is output through a regression head. The root mean square error of the angle regression is no greater than 1.5°. The network output is the actual reading after range conversion.

8. The method for low-power intelligent identification and remote monitoring of mechanical instruments in substations according to claim 7, characterized in that, In step S6, the method for data validity verification is as follows: Establish a database of reasonable ranges for meter readings, storing the upper and lower limits of the range, alarm thresholds, and historical statistical characteristics of various meters; For the identified reading, it is determined whether it meets the set upper and lower limit basic constraints. If the reading is outside the reasonable range, the deviation between the reading and the historical average of the same type of meter is calculated. When the deviation exceeds 3 times the historical standard deviation, it is judged as abnormal data. Abnormal data is stored together with the corresponding original image and marked with an abnormality marker.

9. A method for low-power intelligent identification and remote monitoring of mechanical instruments in substations according to claim 8, characterized in that, In step S7, the low-power remote transmission adopts a BLE5.0+NB-IoT dual-mode communication architecture: when the device is within a 50-meter radius of the substation, the BLE5.0 communication mode is used first to establish a connection with the local operation and maintenance terminal; when the device is outside the BLE coverage area, it switches to the NB-IoT communication mode; the data packet is sent using the CoAP protocol, and the payload is encoded in JSON format.

Images