Device state recognition system

Abstract

The application discloses a kind of equipment state identification systems, and its technical scheme main points are: including image acquisition module, including being arranged in industrial field several high-definition photographic equipment, for collecting the image data containing equipment state light;Edge computing module is connected with image acquisition module, with built-in lightweight deep learning model, for receiving image data, the equipment main body in image is positioned and extracts ROI region, and based on dynamic color identification algorithm identifies light color state, and exports state data;Center server module is connected with edge computing module, for receiving and converging each edge computing module uploaded state data, and using time series database carries out data storage and aggregation analysis.The application adopts dynamic HSV threshold processing and time sequence voting mechanism, so that the system can adapt to complex and changeable industrial lighting environment, has very strong resistance to interference such as reflection, shadow, so as to ensure the stability of state recognition accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of equipment monitoring, in particular to an equipment state recognition system. BACKGROUND

[0002] Industrial equipment usually uses red, yellow, green and other colors to show the running state of the equipment, and in the production workshop, there are many processing equipment arranged in order, and the operator can judge the running state of the equipment through the color of the equipment state light. When the equipment fails, the equipment state light changes from green to red, and the operator can distinguish it.

[0003] However, for the operator, although the running state of the equipment can be distinguished through the change of the state light, the efficiency of fault discovery and repair is low when the operator is engaged in other matters and the operation of the equipment, and the vision is easily blocked due to the blocking of the state light by other equipment, so that the equipment is difficult to be discovered in the first time when the equipment fails or other abnormal conditions occur. SUMMARY

[0004] In view of the defects in the prior art, the purpose of the present application is to provide an equipment state recognition system.

[0005] The above technical purpose of the present application is realized by the following technical scheme: an equipment state recognition system, comprising:

[0006] An image acquisition module comprising a plurality of high-definition photographic devices arranged in an industrial site, for acquiring image data containing equipment state lights;

[0007] An edge computing module in communication connection with the image acquisition module, which has a lightweight deep learning model built-in, for receiving image data, positioning the equipment main body in the image and extracting the ROI region, and identifying the light color state based on a dynamic color recognition algorithm, and outputting state data;

[0008] A central server module in communication connection with the edge computing module, for receiving and converging the state data uploaded by each edge computing module, and using a time series database for data storage and aggregation analysis;

[0009] A dynamic threshold processing module, which performs dynamic threshold segmentation based on a dynamic HSV threshold segmentation algorithm, and determines the color of the state light through sliding window filtering and time series voting mechanism;

[0010] The adjustment step of the dynamic threshold processing module comprises:

[0011] S1, initialize threshold range, at first deployment, offline calibration is carried out, the three-color lamp image of the state lamp under normal illumination condition is collected, the HSV value of the lamp area pixel is extracted, the H, S and V channel mean value and standard deviation value of each color are calculated, and the initial threshold range is set;

[0012] S2, real-time dynamic adjustment, frame-by-frame detection is carried out on the real-time collected image data, the lamp area pixel is segmented based on the initial threshold, the statistical value of the H, S and V channels of the current frame is extracted, and then the dynamic average value of the S and V channels in the preset time window is calculated through sliding window filtering, when the abnormal condition exists, the threshold updating process is triggered;

[0013] S3, time series verification, the detection structure of a continuous specified number of frames is voted, and the state of the state lamp is determined according to more than half of the results.

[0014] The application is further provided that: the judgment logic of the dynamic threshold processing module for abnormal conditions in step S2 is that: if more than half of the continuous frames detect that the saturation S or the value of the brightness V deviates from the initial threshold by ± 15%, the threshold updating process is triggered.

[0015] The application is further provided that: the formula of the threshold updating process is: new threshold = historical threshold * 0.7 + statistical value of current frame * 0.3.

[0016] The application is further provided that: the dynamic threshold processing module in step S2 further includes an outlier rejection step: if the fluctuation amplitude of the S or V value of a frame exceeds 3 times the standard deviation, the frame is determined to be an interference frame, the threshold updating process is skipped, and the last effective threshold is maintained.

[0017] The application is further provided that: the edge computing module performs two-stage detection, including:

[0018] The first stage is to locate the device main body through a lightweight model.

[0019] The second stage is to crop and extract the ROI region containing the device main body in the image, and a high-precision model is used for positioning and identifying the three-color lamp region.

[0020] The application is further provided that: the image acquisition module is based on a high-resolution camera, the resolution is greater than or equal to 1920x1080, and the pixel ratio of the three-color lamp region in the image is greater than or equal to 20x20, accounting for 0.5% of the total area of the image.

[0021] The application is further provided with: the edge computing module performs an implementation process, which firstly labels the image data of a specified area acquired by a high-resolution camera as a data set, distinguishes the device main body and the three-color light area during the labeling, trains a lightweight model to locate the device to achieve a recall rate of 99%, and then trains the lightweight model to detect the three-color light area to achieve an average precision of 95%.

[0022] The application is further provided with: when the pixel ratio of the three-color light area in the image is less than 5%, the high-resolution detection head of the lightweight model is increased, and the width-height ratio of the anchor frame is reduced to match the rectangular shape of the three-color light area.

[0023] The application is further provided with: the edge computing module and the center server module are connected through a PTP network port and adopt the IEEE 1588 precision time protocol, and clock configuration is performed.

[0024] The application is further provided with: the edge computing module and the center server module perform clock configuration, and the steps include: the center server module is clock initialized, and the center server module is configured as a master clock source, then the edge computing module is started automatically after booting, and the PTP time is synchronized automatically, the synchronization frequency is less than or equal to 1 minute per day, when the edge computing module and the dynamic threshold processing module perform data uploading, the data carries a PTP timestamp to the center server module, and the center server module sorts the data through the timestamp.

[0025] In summary, the application has the following beneficial effects:

[0026] 1. Through the two-stage lightweight model of locating the device first and then recognizing the three-color light area, combined with the high-resolution detection head optimized for small targets, the problem of small proportion of the light area in the image is effectively solved, and high recall rate (≥99%) and high precision (≥95%) positioning are realized.

[0027] 2. The dynamic HSV threshold processing and time sequence voting mechanism are adopted, so that the system can adapt to complex and changeable industrial lighting environments, and has strong resistance to interference such as reflection and shadow, thereby ensuring the stability of the state recognition accuracy.

[0028] 3. The IEEE 1588 PTP protocol is adopted to realize microsecond-level clock synchronization, which ensures the strict consistency of the data timestamp of all edge nodes in the distributed deployment, lays a solid foundation for centralized monitoring and accurate data analysis across workshops and multiple cameras, and solves the core problem of distributed systems.

[0029] 4. The non-contact visual solution is adopted, and no electrical modification or wiring is needed for the existing equipment, realizing "plug and play", being compatible with most industrial production equipment with status lights on the market, and realizing digital upgrading of old production lines. BRIEF DESCRIPTION OF DRAWINGS

[0030] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed to be used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present application, and therefore should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can also be obtained without creative labor on the basis of these drawings.

[0031] Figure 1 The structure of the present application is shown in the figure. DETAILED DESCRIPTION

[0032] In order to make the purpose, technical solutions and advantages of the embodiments of the present application more clear, the technical solutions in the embodiments of the present application will be described clearly and completely below in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all. The components of the embodiments of the present application described and shown in the drawings can be arranged and designed in various different configurations.

[0033] Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the claimed present application, but only represents selected embodiments of the present application. All other embodiments obtained by those skilled in the art on the basis of the embodiments in the present application without creative labor are within the scope of protection of the present application.

[0034] It should be noted that: similar numbers and letters represent similar items in the following drawings, so once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings.

[0035] In the description of the present application, it should be understood that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, or the orientation or positional relationship commonly used when the product of the present application is used, or the orientation or positional relationship commonly understood by those skilled in the art, and are only for the convenience of describing the present application and simplifying the description, and therefore cannot be understood as indicating or implying that the indicated device or element must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application.

[0036] In addition, the terms "first", "second", "third", and the like are used only to distinguish descriptions, and cannot be understood as indicating or implying relative importance.

[0037] In the description of the present application, it should also be noted that, unless otherwise explicitly specified and limited, the terms "set", "install", "connect", "connect" should be understood broadly, for example, it can be fixedly connected, or it can be detachably connected, or integrally connected; it can be mechanically connected, or it can be electrically connected; it can be directly connected, or it can be indirectly connected through an intermediate medium, or it can be the communication inside two elements. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0038] A device state recognition system, as shown in Figure 1 It comprises:

[0039] An image acquisition module comprising a plurality of high-definition camera devices arranged in an industrial site for acquiring image data containing device status lights;

[0040] An edge computing module in communication connection with the image acquisition module, which has a lightweight deep learning model built-in, for receiving image data, positioning the device main body in the image and extracting the ROI region, and identifying the light color state based on a dynamic color recognition algorithm, and outputting the state data;

[0041] A central server module in communication connection with the edge computing module through a PTP network port and using IEEE 1588 precision time protocol, and performing clock configuration, for receiving and aggregating the state data uploaded by each edge computing module, and using a time series database for data storage and aggregated analysis. In this embodiment, the central server module uses an industrial-grade service, such as Dell PowerEdge R750, with a CPU of ≥16 cores, 64GB of memory, at least 2TB of solid state SSD hard disk, supporting Docker containerized deployment InfluxDB time series database and MQTT message middleware.

[0042] A dynamic threshold processing module based on a dynamic HSV threshold separation algorithm to perform dynamic threshold segmentation, and determine the state light color through a sliding window filtering and a time series voting mechanism.

[0043] In this embodiment, in addition to setting a plurality of image acquisition modules, edge computing modules, central server modules and dynamic threshold processing modules, peripherals such as host computers and API interfaces can also be set, and the host computer is used to display the image data acquired by the image acquisition module in real time, and the API interface is connected with the MES / SCADA / IOT system.

[0044] In this embodiment, the image acquisition module is based on a high-resolution camera with a resolution of ≥1920x1080, and the three-color light area occupies ≥20x20 pixels in the image, accounting for 0.5% of the total image area. In this embodiment, the image acquisition module selects a 1080P (1920x1080 pixels) or higher resolution camera that supports the RTSP protocol and ONVIF standard, such as Hikvision DS-2CD3T26WD-I or Basler acA2040-90um, with a frame rate of ≥30fps, and the camera lens can be equipped with a polarizing lens such as B+W polarizing filter.

[0045] When deploying the image acquisition module on site, first determine the size of the site area and the arrangement of the equipment. If the equipment is concentrated or the number of devices to be monitored is small, use a single camera deployment. Install the camera above or to the side of multiple devices to ensure that the camera covers the three-color light area. If multiple production areas are involved or a single camera cannot cover all device areas, and there is no need for verification between multiple cameras, deploy 4-8 cameras locally according to the requirements. Install the camera at a distance of 1.5-3 meters from the light area, and the installation angle of the camera should be ≤30°. Avoid mirror reflection when installing.

[0046] In this embodiment, the edge computing module uses an embedded NPU computing unit, with the core hardware being a NVIDIA Jetson Xavier NX development kit (or equivalent Advantech UNO-2483G edge computing box). It is equipped with a 6-core ARM CPU, a 384-core Volta GPU, and 48 Tensor Cores, with an INT8 computing power of 21 TOPS, and a built-in 16GB LPDDR4x memory. Each edge computing module is connected to 4-8 cameras through a gigabit Ethernet port, enabling real-time processing of multiple video streams.

[0047] The edge computing module performs two-stage detection, including:

[0048] In the first stage, the light-weight model is used to locate the device body.

[0049] In the second stage, the ROI region containing the device body is extracted and cropped in the image, and the high-precision model is used to locate and identify the three-color light area.

[0050] Wherein, before the three-color light area positioning is performed by using the lightweight model and the high-precision model, the lightweight model and the high-precision model need to be trained: first, a data enhancement strategy is formulated, dynamic light simulation is adopted, random brightness and contrast adjustment are set, and occlusion simulation is performed, and the model data is tested by using enhancement methods such as random noise block covering the light area, so as to improve the robustness of the model to complex working condition scenes. On the other hand, GAN is used to generate three-color light images under extreme lighting conditions, such as strong reflection and shadow interference, to simulate actual working conditions and train the model.

[0051] After the simulation data training is completed, the model is optimized: including lightweight improvement and attention mechanism optimization, wherein, based on L1-norm channel pruning (pruning rate 30%), INT8 quantization precision loss ≤1.5% (FP32: 98.2%→INT8: 96.7%), and CBAM module is inserted in YOLOv7 backbone network, small target detection AP@0.5 is improved by 12%.

[0052] The implementation process of the edge computing module is as follows: first, the image data of the specified area obtained by the high-resolution camera is labeled as a data set, and the device main body and the three-color light area are distinguished during labeling the data set. The lightweight model is trained to locate the device to achieve a recall rate ≥99%, and then the average precision of the three-color light area detection is trained to be ≥95%. When the pixel ratio of the three-color light area in the image is <5%, the high-resolution detection head of the lightweight model is increased, and the width-height ratio of the anchor frame is reduced to match the rectangular shape of the three-color light area.

[0053] Further, the edge computing module and the center server module are clock configured, and the steps include: initializing the clock of the center server module and configuring the center server module as a master clock source, and then the edge computing module automatically synchronizes the PTP time after starting, and the synchronization frequency is ≤1 minute per day. When the edge computing module and the dynamic threshold processing module upload data, the data carries the PTP timestamp to the center server module, and the center server module sorts the data by timestamp.

[0054] Wherein, the adjustment steps of the dynamic threshold processing module include:

[0055] S1, initialize the threshold range, when deployed for the first time, collect the three-color light images of the status light under normal lighting conditions, extract the HSV values of the light area pixels, calculate the H, S and V channel means and standard deviation values of each color, and set the initial threshold range;

[0056] S2, real-time dynamic adjustment, frame-by-frame detection is performed on the real-time collected image data, the lamp area pixels are segmented based on the initial threshold value, the statistical values of the current frame H, S and V channels are extracted, and then the dynamic average values of the S and V channels in a preset time window are calculated through sliding window filtering. When an abnormal condition exists, a threshold value updating process is triggered. The specific abnormal condition judgment logic is: if more than half of the continuous frames detect that the saturation S or the brightness V value deviates from the initial threshold value by ±15%, it is determined that an abnormal condition exists, the threshold value updating process is triggered, and the formula of the threshold value updating process is: new threshold value = historical threshold value * 0.7 + statistical value of current frame * 0.3.

[0057] For example: the saturation S channel of the red lamp, the initial threshold value mean is μ_S_init=50, and the actual measured implementation processing data of 5 continuous frames are:

[0058] The S channel mean of the red lamp area of the first frame is μ_S_curr1=52, the calculation deviation is 4%, which is less than 15%, does not reach the trigger condition, the threshold value is not updated, the effective threshold value is still 50, and the sliding window record value is

[52] .

[0059] When the S channel mean of the red lamp area of the second frame is μ_S_curr1=55, the calculation deviation is 10%, which is less than 15%, does not reach the trigger condition, the threshold value is not updated, the effective threshold value is still 50, and the sliding window record value is 【52, 55】.

[0060] When the S channel mean of the red lamp area of the third frame is μ_S_curr1=58, the calculation deviation is 16%, which is greater than 15%, the judgment result is that a significant deviation first appears, the sliding window record value is 【52, 55, 58】, the value μ in the current sliding window meets the half deviation over-limit, and the threshold value is still 50.

[0061] When the S channel mean of the red lamp area of the fourth frame is μ_S_curr1=61, the calculation deviation is 22%, which is greater than 15%, the judgment result is that a significant deviation appears, the sliding window record value is 【52, 55, 58, 61】, the value in the current sliding window does not meet the half deviation over-limit, and the threshold value is still 50.

[0062] When the S channel mean of the red light area of the 5th frame is μ_S_curr1=63, the calculation deviation is 26%, which is greater than 15%, and the judgment result is that there is a significant deviation. Check the consecutive frames: the deviation of the 3rd, 4th and 5th frames is more than 15% for 3 consecutive frames, which meets the triggering condition, and the threshold updating process is started: new threshold = 50 (historical threshold) * 0.7 + 57.8 (current dynamic average) * 0.3 = 52.34, the system updates the threshold of the red light saturation S channel from 50 to 52.76, and this new threshold will be used for subsequent frame processing, so that the color recognition algorithm can better adapt to the current bright environment light, and avoid misjudging the original red color as other colors due to threshold fixation;

[0063] If after the 4th frame, the next frame is disturbed by instantaneous strong light, resulting in μ_S_curr5=100 (the fluctuation amplitude is far more than 3 times the standard deviation), the system will determine that this frame is an abnormal interference frame, ignore its data, and do not update the threshold, maintaining the threshold at 50, thereby avoiding the misadjustment of the system caused by instantaneous interference.

[0064] In the present embodiment, the dynamic threshold processing module is implemented using a self-defined algorithm based on OpenCV and NumPy libraries. When the system is first deployed, an offline calibration program is started, and static images of each device status light are collected under normal lighting conditions. A pre-trained YOLOv7 light area detection model is used to locate the three-color light areas in the image, and the pixel set of each light area is extracted. The pixels are converted from BGR color space to HSV space, and the H, S, and V channel data of red, yellow, and green pixels are counted respectively. The mean and standard deviation of each color channel are calculated, and the initial threshold range is set, such as saturation S≥40 and brightness V≥80, to exclude low saturation (such as gray interference) and low brightness (such as unlit lights) pixels. In the dynamic adjustment step, when the specific values of the H, S, and V channels of the current frame are extracted, abnormal values are excluded from the numerical range of the statistical values. The specific judgment rule is: if the fluctuation amplitude of S or V value of a frame exceeds 3 times the standard deviation, it is determined that the frame is an interference frame, the threshold updating process is skipped, and the last valid threshold is maintained.

[0065] S3, time series verification, voting for the detection structure of consecutive specified frames to determine the state of the status light with more than half of the results. For time series verification, in the present embodiment, odd frame time series filtering is used. Specifically, a 5-frame time window is used to calculate the dynamic average of saturation S and brightness V. In the verification, the detection structure of the last 5 frames is voted, and at least 3 frames are consistent and confirmed, for example, 3 frames are detected as red, eliminating the influence of single frame noise.

[0066] In the present embodiment, the state mapping rule is set as follows: green light color indicates normal operation of the equipment; yellow light color indicates standby or warning of the equipment; red light color indicates failure of the equipment; yellow and red light colors indicate serious failure of the equipment; and the equipment state definition can be configured by the user according to the type of the equipment when the yellow and red state lights are on at the same time. (See the following table for a comparison table)

[0067] Light color State definition Green Normal operation Yellow Standby / Warning Red Fault Yellow + Red Critical fault (configurable)

[0068] The central server records the changes in the above-mentioned equipment operating state in real time, and checks the following statistical data: the duration of each state; the time point of state switching; the frequency and cumulative duration of failure, which is beneficial to the operation supervision personnel to trace back the equipment failure time point and the maintenance status of the equipment in time.

[0069] To solve the problem of high-concurrency data processing delay in actual operation, edge computing combined with load balancing algorithm is adopted, and a distributed edge computing module is further adopted, which processes 4-8 cameras through a single node to avoid overloading of the central server. The edge node selects a low-delay NPU (such as Jetson NX supporting parallel processing of 2-way 1080P video stream). At the algorithm level, when the load is >80%, the inference frame rate of the non-faulty state camera is automatically reduced, for example, from 30fps to 15fps, or the resolution is reduced from 1080P to 720P, to achieve dynamic load balancing. On the other hand, adjustments are made in the region of interest (ROI) interception. After the equipment is positioned, only the ROI where the light area is located is derived, for example, the 640x480 pixel area where the light area is located, to reduce the amount of calculation and reduce the inference time by about 40%.

[0070] In the configuration, the edge nodes are divided according to the physical partition of the workshop, each node covers a radius ≤50m, and the central server module monitors the load of each node in real time and issues frame rate adjustment instructions through the MQTT protocol. In the non-faulty state, the asynchronous inference mode (process one frame every 00ms) is adopted, and the fault state is switched to synchronous real-time processing (30fps full-frame inference), thereby solving the problem of high-concurrency data processing delay.

[0071] In summary, the application realizes "plug and play" through an external camera, saves the need to arrange sensors on the device, does not need to be connected through a protocol, is suitable for the automatic transformation of old devices, and the ONVIF plug and play architecture supports distributed expansion, can realize single node 4-8 cameras, adapts to complex layouts in workshops, and is more flexible and simple to deploy. In terms of algorithm optimization, the technical means realizes small target detection enhancement, solves the challenge of a light area pixel ratio <5% through a two-stage detection method combined with CBAM attention, and achieves a recall rate of 97.5%. Moreover, through the lightweight deployment of YOLOv7 pruning + INT8 quantization, a delay of ≤50ms is realized, which adapts to edge computing (Jetson NX), meets real-time requirements, and achieves a fault response of ≤2 seconds. Through the innovative architecture of "AI + machine vision + spatial positioning + time sequence analysis", non-contact global monitoring of the state of the workshop-level device is realized, and the pain points of complex wiring and poor scalability of traditional solutions are solved. In the future, multi-modal data (such as infrared thermal imaging and voiceprint detection) can be further integrated to improve the robustness of complex scenes, and AI large models can be used to optimize predictive maintenance accuracy, helping the digital transformation of the manufacturing industry.

[0072] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Those skilled in the art can make various modifications and changes to the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A device status identification system, characterized in that: include The image acquisition module includes several high-definition photography devices installed at the industrial site, used to acquire image data including equipment status lights; The edge computing module communicates with the image acquisition module. It has a built-in lightweight deep learning model to receive image data, locate the main device in the image and extract the ROI region, and identify the color status of the light based on the dynamic color recognition algorithm and output the status data. The central server module communicates with the edge computing module and is used to receive and aggregate status data uploaded by each edge computing module, and use a time-series database for data storage and aggregation analysis. The dynamic threshold processing module performs dynamic threshold segmentation based on the dynamic HSV threshold segmentation algorithm and determines the status light color through sliding window filtering and time series voting mechanism. The adjustment steps of the dynamic threshold processing module include: S1. Initialize the threshold range. During the initial deployment, perform offline calibration, collect the three-color light images of the status lights under normal lighting conditions, extract the HSV values ​​of the pixels in the light area, calculate the mean and standard deviation of the H, S, and V channels of each color, and set the initial threshold range. S2. Real-time dynamic adjustment: Frame-by-frame detection of real-time acquired image data, segmentation of light area pixels based on initial threshold, extraction of statistical values ​​of H, S, and V channels in the current frame, and then calculation of dynamic average values ​​of S and V channels within a preset time window through sliding window filtering. In case of abnormal conditions, the threshold update process is triggered. S3, Time Series Verification: Voting is performed on the detection results for a specified number of consecutive frames, and the state of the status light is determined by a majority vote. The dynamic threshold processing module's judgment logic for abnormal conditions in step S2 is as follows: if more than half of the consecutive frames detect that the value of saturation S or brightness V deviates from the initial threshold by ±15%, the threshold update process is triggered. The dynamic threshold processing module further includes an outlier removal step in step S2: if the fluctuation range of the S or V value of a certain frame exceeds 3 times the standard deviation, the frame is determined to be an interference frame, the threshold update process is skipped, and the previous effective threshold is maintained.

2. The equipment status identification system according to claim 1, characterized in that: The formula for the threshold update process is: New threshold = Historical threshold * 0.7 + Statistical value of the current frame * 0.

3.

3. The equipment status identification system according to claim 1, characterized in that: The edge computing module performs two-stage detection, including: The first phase involves using a lightweight model to locate the main body of the device. In the second stage, the ROI region containing the main body of the device is cropped and extracted from the image, and a high-precision model is used to locate and identify the three-color light region.

4. The equipment status identification system according to claim 3, characterized in that: The image acquisition module is based on a high-resolution camera with a resolution of ≥1920x1080. The three-color light area accounts for ≥20x20 pixels in the image, representing 0.5% of the total image area.

5. The equipment status identification system according to claim 4, characterized in that: The implementation process of the edge computing module is as follows: First, the image data of the specified area acquired by the high-resolution camera is labeled into a dataset. When labeling the dataset, the main body of the device and the three-color light area are distinguished. The lightweight model is first trained to achieve a recall rate of ≥99% for device localization, and then trained to achieve an average accuracy of ≥95% for the detection of the three-color light area.

6. The equipment status identification system according to claim 4, characterized in that: When the pixel ratio of the tri-color light area in the image is less than 5%, a high-resolution detection head of a lightweight model is added, the aspect ratio of the anchor frame is reduced, and the rectangular shape of the tri-color light area is matched.

7. The equipment status identification system according to claim 1, characterized in that: The edge computing module and the central server module are connected via a PTP network port and use the IEEE 1588 precision time protocol for clock configuration.

8. The equipment status identification system according to claim 7, characterized in that: The edge computing module and the central server module are configured to use a clock. The steps include: initializing the clock of the central server module and configuring the central server module as the main clock source. After the edge computing module is powered on, it automatically synchronizes the PTP time with a synchronization frequency of ≤1 minute per day. When the edge computing module and the dynamic threshold processing module upload data, the data carries a PTP timestamp to the central server module, and the central server module sorts the data by timestamp.

Images