An oil mist generator oil level identification method

Abstract

The application discloses an oil mist generator oil level identification method and belongs to the technical field of hump signal equipment monitoring. The oil level image of the oil cup of the oil mist generator is shot by the camera fixedly installed on the inner wall of the case and provided with a light source, is transmitted to the system platform through the parallel multi-mode communication module, is processed based on the YOLO segmentation model, and the oil level percentage is calculated. The current oil level and the historical curve are used to judge the abnormality and generate a pre-warning information. The oil level identification result and the pre-warning are displayed. The application can completely replace the manual case opening inspection, realizes the intelligent monitoring and the preventive early warning of the oil mist generator oil level of the hump signal equipment.

Description

Technical Field

[0001] This invention relates to the field of hump signal equipment monitoring technology, specifically to a method for identifying the oil level in an oil mist lubricator. Background Technology

[0002] Hump ​​signaling equipment consists of various types and complex components, including route control interlocking devices, speed control devices, related detection equipment, and mechanical power equipment. The core of the air source treatment element in the electro-pneumatic switch machine is a pneumatic triplet (filter + pressure reducing valve + oil mist lubricator), responsible for purifying compressed air, stabilizing output pressure, and atomizing the lubricating medium. The oil mist lubricator is a key lubrication component of the hump vehicle's reducer; it atomizes specialized lubricating oil and delivers it to moving parts via airflow, forming an oil film to reduce wear and prevent corrosion.

[0003] The relevant operating standards for oil mist lubricators include: daily inspection of the oil level (not lower than 1 / 3 and not higher than 2 / 3), and observation of the atomization status; regular maintenance: replenish oil every 1-2 weeks under normal operating conditions, and perform a deep cleaning of the oil cup and nozzle every 1-3 months; shorten the maintenance cycle under special conditions. All of the above operations must be manually performed by the signalman during the maintenance window. Daily inspections, in particular, require daily unpacking and checking, and the work is complex; visual observation can only provide a rough assessment and cannot provide accurate identification.

[0004] Therefore, there is an urgent need for an efficient and intelligent automated method to accurately identify and judge the oil level and dripping status of the oil mist lubricator. Summary of the Invention

[0005] This invention provides a method for identifying the oil level in an oil mist lubricator, to solve the problems of low efficiency and poor accuracy of manual inspection in the prior art.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A method for identifying the oil level in an oil mist lubricator includes the following steps: An image of the oil level in the oil cup of the oil mist lubricator is captured by a camera fixedly installed on the inner wall of the chassis and equipped with its own light source; the oil level image is transmitted to the system platform through a communication module; the oil level image is processed based on a deep learning image recognition model to identify and calculate the oil level percentage; the presence of anomalies is determined based on the current oil level percentage and historical oil level change curves, and if an anomaly is found, a pre-alarm message is generated; the oil level percentage and the pre-alarm message are displayed.

[0007] Furthermore, the camera has a built-in white light source, and its fill light intensity and timing are automatically adjusted via code. The camera also supports serial port or Ethernet port communication.

[0008] Furthermore, the communication module adopts a parallel multi-mode communication scheme, integrating at least two types of communication modules from LORA, star flash, and high-speed carrier HPLC, supporting wired and wireless multi-link parallel transmission and reception; the communication module is compatible with DC12V and AC220V power supply and adopts DIN rail mounting.

[0009] Furthermore, the deep learning image recognition model is built based on the YOLO real-time object detection model, specifically including: collecting oil cup photos at different oil level heights to construct an oil level image dataset; labeling the images in the dataset, at least labeling all liquid level areas and the current liquid level area, and also labeling markers for the reducer oil mist lubricator; performing data augmentation on the labeled images to expand the dataset; training the YOLO segmentation model using the expanded dataset to obtain the oil level recognition model; and calculating the oil level percentage using the following formula: Calculate the oil level percentage using the following formula: ; in Indicates the current liquid level height. This indicates the total liquid level height.

[0010] Furthermore, the data enhancement includes rotation transformation and brightness transformation; the rotation transformation is a change of 0.5° to 2° each time, and the brightness transformation is a change of 5% to 10% each time.

[0011] Furthermore, the YOLOv8X-SEG model was used for training, and the following parameters were set: the initial size of the training image was (640, 480) and multi-scale training was enabled; the number of warm-up training rounds was 5; anchor boxes were recalculated; the bounding box weight was set to 1.2; the IOU threshold was set to 0.7; the mosaic intensity was set to 0.3; the copypaste ratio was set to 0.3; the scaling range was set to 0.3; the perspective transformation parameter was set to 0.001; and the mask downsampling ratio was set to 2.

[0012] Furthermore, the method also includes an oil level identification error processing step: in the operating mode, the oil level percentage identified last time is recorded; if the current identification result is higher than the previous identification result and the difference is not greater than 1%, then the previous oil level percentage is maintained as the current identification result.

[0013] Furthermore, the anomalies include abnormal oil level exceeding limits and abnormal oil level change trends; abnormal oil level exceeding limits refers to an oil level higher than 66.6% or lower than 33.3%, at which point a real-time alarm is generated; abnormal oil level change trends refer to: analyzing the historical oil level curves of the past 15 days, if the oil level change on two consecutive days is lower than the daily lower limit or higher than the daily upper limit, at which point a real-time warning is generated; the initial value of the daily lower limit is 0.05%, and the initial value of the daily upper limit is 3.0%, and the daily lower limit and daily upper limit can be dynamically adjusted according to the number of equipment operations.

[0014] Furthermore, the real-time alarm or real-time warning is accompanied by voice broadcast; the historical oil level change curve is displayed in the form of a bar chart.

[0015] Furthermore, the method also includes a manual triggering step: in response to the user's manual triggering command, the camera is immediately controlled to capture an image of the current oil level, and the deep learning image recognition model is invoked to return the oil level recognition result.

[0016] Compared with the prior art, the present invention has the following beneficial effects: This invention utilizes a self-developed camera with a white light source to automatically acquire photos of the oil level in the oil mist lubricator via timed or polling methods. It also supports manual triggering for immediate acquisition, thus solving the problems of unclear images and glare in dark environments.

[0017] After acquiring a photo of the oil level in the oil mist lubricator, this invention can automatically identify the percentage of oil level (error <1%) and display the inference result photo, which is clear and intuitive. The accuracy of identification is far higher than that of manual observation, and the average error accuracy is improved by more than 200%.

[0018] This invention triggers a real-time alarm and voice broadcast when the oil level in the oil mist lubricator exceeds the limit, and triggers a real-time warning and voice broadcast when the oil level change trend is abnormal, thus realizing prevention, pre-control, early warning, and prediction.

[0019] This invention can display the current oil level information and historical oil level change curves of the oil mist lubricator, with a clear and intuitive display effect, making it convenient for users to view and analyze.

[0020] This invention can completely replace the manual operation of opening the box to check the oil level of the oil mist lubricator, greatly improving the intelligence and automation level of the oil mist lubricator operation, effectively improving the efficiency of hump yard operations and reducing labor costs.

[0021] This invention has good application scalability. For liquid level recognition projects, only the corresponding oil cup level dataset needs to be updated to apply this method to achieve accurate liquid level recognition and early warning. Attached Figure Description

[0022] Figure 1 Air source processing components for switch machines.

[0023] Figure 2 Vehicle speed reducer oil mist lubricator.

[0024] Figure 3 Framework flowchart.

[0025] Figure 4 A self-developed camera with a white light source.

[0026] Figure 5 Image of the oil level markings on the oil mist lubricator of a switch machine.

[0027] Figure 6 Oil level marking on the gearbox oil mist lubricator (with metal cover).

[0028] Figure 7 Oil level marking on the gearbox oil mist lubricator (without metal cover).

[0029] Figure 8 Model evaluation results.

[0030] Figure 9 Oil level identification results for the switch machine's oil mist lubricator. Figure 10 Oil level identification results for the gearbox oil mist lubricator with metal protective cover. Figure 11 shows the oil level identification results of the gearbox oil mist lubricator without a metal protective cover. Detailed Implementation

[0031] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] Example 1 This embodiment provides a method for identifying the oil level of an oil mist lubricator, which is applied to the oil mist lubricator of the switch machine and the oil mist lubricator of the reducer in a hump yard signaling system. Figure 1 The oil mist lubricator in the air source treatment element of the switch machine is shown. Figure 2 The vehicle reducer lubricator is shown. Figure 3 This is a flowchart illustrating the framework of the method of the present invention.

[0033] This embodiment uses a self-developed, fixable, high-definition camera with its own light source (such as...). Figure 4As shown in the image, the camera is installed on the inner wall of the enclosure housing the oil mist lubricator. The interior of the enclosure is normally closed, creating a dark environment. The camera has a built-in white light source and supports serial or Ethernet communication. For the dark environment inside the enclosure, the fill light intensity and timing are automatically adjusted via code: the light source is triggered 500ms before shooting and automatically turned off after exposure; the fill light intensity is adjustable in 5 levels, controlled by feedback from the average grayscale value G of the ROI region in the center of the image—if G≤80, it increases by one level; if G>180, it decreases by one level; if 80≤G≤180, it maintains the current level, ensuring that the captured photos are clear and reflection-free, meeting the requirements of deep learning models for image datasets.

[0034] Specifically, a large number of oil cup photos with different oil levels were taken for the switch machine oil mist lubricator and the reducer oil mist lubricator, respectively, to construct two datasets.

[0035] For hump monitoring scenarios, this embodiment features a customized parallel multi-mode communication solution. The communication module is equipped with a combined Ethernet and RS485 communication interface and integrates three types of communication modules: LORA, SLE / SLB, and high-speed carrier HPLC, supporting wired and wireless multi-link parallel transmission and reception. It can achieve 4 / 6-channel parallel transmission and reception on a single device, which can be expanded to 8 channels through stacking. The channel frequency planning interval is greater than 2MHz, supporting flexible channel allocation to adapt to the differentiated transmission needs of images and sensors. The various communication modules can also be mixed and matched as needed to adapt to complex networking environments. The device is compatible with DC12V and AC220V power supplies and adopts a DIN rail mounting method.

[0036] The captured oil level images are transmitted back to the system platform in real time or periodically via the aforementioned communication module. The system platform is a server or edge computing gateway deployed at the monitoring center. Its database uses a conventional relational database or time-series database to store oil level percentage and historical curve data, all of which are conventional backend technologies in this field. The communication module supports parallel transmission and reception of wired and wireless multi-links: normally, image data is transmitted via a high-speed carrier HPLC wired channel, while control signaling and sensor data are transmitted in parallel via LORA or StarFlash wireless channels; when the signal-to-noise ratio of any link falls below a threshold or is interrupted, automatic primary / backup redundancy switching is initiated, migrating the corresponding data stream to other available links, with a switching time of less than 500ms.

[0037] This embodiment uses the YOLO (YouOnlyLookOnce) real-time object detection model as an example. It builds an algorithm environment based on the PyTorch deep learning training framework and the OpenCV vision processing framework, and completes the construction and deployment of the oil level recognition model for oil mist generators through a standardized process. Other deep learning models with instance segmentation capabilities (such as MaskR-CNN, SOLO series, etc.) can also use the same dataset construction and post-processing process to calculate the oil level percentage. The specific steps are as follows: We collected a large number (over 400) of photos of oil cups at different oil levels by using cameras installed on-site and cameras on experimental equipment.

[0038] The collected images were divided into training, validation, and test sets in an 8:1:1 ratio to ensure that each dataset contained images of different oil levels. The oil level photos in the test dataset required actual measurements of the corresponding oil level percentages for subsequent error calculation.

[0039] Images in the training and validation sets are labeled using the LabelMe tool. All liquid levels (aol) and the current liquid level (col) are labeled separately. For the reducer oil mist lubricator, a symbol label is also required. The symbol label is a feature point on the bottom edge of the reducer oil mist lubricator's metal protective cover or a fixed scale mark on the outer wall of the oil cup. It is used to locate the bottom center point of the oil level mark line during the inference phase and does not participate in the numerical calculation of the oil level percentage.

[0040] Figure 5 The image shows a diagram indicating the oil level in the oil mist lubricator of a switch machine. Figure 6 (with metal casing) and Figure 7 (Without metal casing) A picture showing the oil level markings for the gearbox lubricator.

[0041] The image data is subjected to rotation and brightness transformations to generate images with varying degrees of rotation (0.5°–2° each time) and gradual brightness changes (5%–10% each time). For labeled image data, corresponding labels are generated simultaneously to expand the dataset. The rotation and brightness transformations are different for the training and validation sets to improve the model's generalization performance.

[0042] The YOLOv8X-SEG model is used as the base model. The model training parameters and callback functions are designed. The training parameters are set as follows: The initial size of the training images is (640, 480), and multi-scale training is enabled (multiscale=True), with the size randomly selected in each iteration; Enable warmup training (warmupepochs=5) to allow the model to smoothly adapt to new data and avoid training oscillations; Recalculate anchor frames and automatically calculate anchor frames that fit the 4:3 aspect ratio (anchor='auto'). Increase the bounding box weight (box=1.2) to optimize small target localization, and reduce the IOU threshold (iou=0.7) to better suit small target matching; Reduce the intensity of the mosaic (mosaic=0.3) to preserve the features of small targets, increase the copypaste ratio (copypaste=0.3) to increase the number of small target samples; Increase the scaling range (scale=0.3) to simulate different sizes, and set a slight perspective transformation (perspective=0.001) to enhance robustness; Reduce the mask downsampling ratio to preserve the outline of small targets (maskratio=2).

[0043] The custom callback function is used to: track the detection mAP50, mAP50-95, and segmentation mAP50 metrics in real time after each round of validation, record and update the best mAP50-95 value and the corresponding round; synchronously print the current system time and cumulative training time; automatically trigger the optimization mechanism when mAP50 continuously reaches the specified number of rounds, decay the learning rate, and slightly increase the loss weight of the detection box; and enhance the feature learning of the segmentation mask through fine-tuning of parameters.

[0044] Training is stopped when the model's overall performance is optimal or when the maximum number of training epochs is reached, and the optimal parameters are saved.

[0045] The optimal parameter model was evaluated on the test set, achieving precision, recall, and mAP all above 90%. The average error between the model's calculated percentages and the actual percentages on the test set images was less than 1%. Specific model evaluation results are as follows: Figure 8 As shown, the accuracy rate of oil level recognition for the switch machine oil mist lubricator is 99.6%, and the accuracy rate of oil level recognition for the reducer oil mist lubricator is 99.5%. The maximum errors are 0.84% ​​and 0.95%, respectively, and the average errors are 0.65% and 0.76%, respectively. The accuracy improvement percentages are 223% and 360.5%, respectively.

[0046] Add visual auxiliary lines, oil level percentage values, and photo capture time to the inference result image. The oil level percentage calculation formula is: Calculate the oil level percentage using the following formula: ; in Indicates the current liquid level height. This represents the total liquid level height. Specifically, from the pixel-level masks of the current liquid level region and the entire liquid level region output by the YOLO segmentation model, the pixel heights of the bounding rectangles of the two masks in the vertical direction of the image are extracted as follows. and If a marker label is detected, the pixel height is normalized and corrected using the vertical pixel coordinates of the marker label's center point as the reference zero point to eliminate system errors caused by the camera's installation tilt angle. The marker is used to locate the bottom center point of the oil level mark line.

[0047] Figure 9 An example of the inference result for the oil level identification of the switch machine's oil mist lubricator is shown. Figure 10 (with metal protective cover) and Figure 11 (Without metal protective cover) An example of the inference result for the gearbox oil mist lubricator oil level identification is shown.

[0048] Add an HTTP interface for model inference, take in image data, and return inference result information.

[0049] In operating mode, record the previously identified oil level percentage. If the current identification result is higher than the previous one, but the difference is no more than 1% (considering that the oil level will not rise automatically when the oil cup is not operated, such a small increase is likely due to identification error), then maintain the previous oil level percentage as the current identification result.

[0050] The system takes photos of the oil level in the oil mist lubricator at regular intervals each day (e.g., at 7:00 AM, in front of the sunroof), and calls the model inference interface to obtain the oil level percentage recognition result. The inference result image and percentage data are saved to the database. Based on the database records, a historical oil level change curve and bar chart are plotted for monitoring analysts to view and compare.

[0051] The pre-alarm business logic is as follows: Oil level abnormality alarm: When the oil level is higher than 66.6% or lower than 33.3%, a real-time alarm is triggered and a voice broadcast is given to inform the user to take immediate action. The voice broadcast is achieved by calling an existing TTS speech synthesis engine or triggering a preset audio broadcast module, which is a conventional human-computer interaction technology in this field.

[0052] Oil level change trend abnormality warning: Analyze the historical oil level curve data of the past 15 days. If the oil level change of two consecutive days is lower than the daily lower limit of decline (initial value 0.05%, which can be dynamically adjusted based on the number of equipment operations) or higher than the daily upper limit of decline (initial value 3.0%, which can be dynamically adjusted based on the number of equipment operations), a real-time warning will be triggered and a voice broadcast will be given to inform the user to analyze and deal with the situation in time (such as the oil cup may be leaking or not dripping). The historical oil level change curve is displayed in the form of a bar chart, which is generated by an existing data visualization chart library and is a conventional data display method.

[0053] The system supports user-initiated manual capture of the latest oil level image from the oil mist lubricator. In response to the user's manual trigger command, the system immediately sends a capture command to the camera module and receives the latest oil level image returned by the camera (at this point, the image resolution can be appropriately reduced to improve responsiveness). Simultaneously, it calls the inference model interface and returns the oil level identification inference result. Based on the result, the user can take measures such as adding oil, cleaning, or stopping the machine to replace the oil.

[0054] Example 2 This embodiment is basically the same as Embodiment 1, except that the communication module integrates only two types of communication modules: LORA and high-speed carrier HPLC. It is suitable for scenarios where wired network coverage is insufficient but power line communication conditions are good. The other steps are the same as in Embodiment 1.

[0055] Example 3 This embodiment is basically the same as Embodiment 1, except that the threshold for abnormal oil level exceeding the limit can be adjusted according to the actual oil cup capacity and operating requirements, for example, set to be higher than 70% or lower than 30%. The dynamic adjustment coefficients of the daily lower and upper limits of descent change linearly according to the number of equipment actions: the more actions, the larger the daily upper limit of descent, to reflect normal lubrication consumption.

[0056] Specifically, let N be the number of daily equipment operations, Tmax = 3.0% be the initial upper limit of the daily decrease, and Tmin = 0.05% be the initial lower limit of the daily decrease. Then, the dynamic adjustment formula is: Tmax′ = Tmax + k1 × NTmin′ = Tmin + k2 × N, where k1 is the upper limit adjustment coefficient, ranging from 0.0005% / operation to 0.002% / operation, and k2 is the lower limit adjustment coefficient, ranging from 0.00005% / operation to 0.0002% / operation. For example, when k1 is 0.001% / operation, k2 is 0.0001% / operation, and N = 100, the adjusted upper limit of the daily decrease is 3.1%, and the lower limit of the daily decrease is 0.06%. Other steps are the same as in Example 1.

[0057] Using the method of this invention, continuous monitoring of the oil mist lubricators of 10 switch machines and 10 reducers was conducted at the hump yard signaling equipment site for three months. The results showed that the average error in oil level identification was 0.65% (switch machine) and 0.76% (reducer), while the average errors of manual visual observation were 2.1% and 3.5%, respectively. This invention improves the identification accuracy by more than 200%. Simultaneously, the system successfully issued warnings for three instances of excessively rapid oil level drops (oil leakage) and two instances of excessively slow oil level drops (oil dripping and blockage), preventing equipment failures.

[0058] This invention completely replaces the manual operation of opening and checking the oil level of the lubricator, improving the intelligence and automation level of hump yard signaling equipment. It is applicable to railway hump marshalling yards and other industrial scenarios requiring lubrication monitoring of lubricators. For other liquid level identification projects, this method can be applied simply by updating the corresponding oil cup level dataset, demonstrating good scalability.

[0059] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural modifications made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

[0060] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for identifying the oil level in an oil mist lubricator, characterized in that, include: The oil level image of the oil cup of the oil mist lubricator is captured by a camera that is fixedly installed on the inner wall of the chassis and has its own light source. The oil level image is transmitted to the system platform via a communication module; The oil level image is processed based on a deep learning image recognition model to identify and calculate the oil level percentage; Based on the current oil level percentage and historical oil level change curves, determine if there are any abnormalities; if so, generate a pre-alarm message. Displays the oil level percentage and early warning information.

2. The method for identifying the oil level in an oil mist lubricator according to claim 1, characterized in that, The camera has a built-in white light source, and its fill light intensity and timing are automatically adjusted by code. The camera also supports serial port or network port communication.

3. The method for identifying the oil level in an oil mist lubricator according to claim 1, characterized in that, The communication module adopts a parallel multi-mode communication scheme, integrating at least two types of communication modules from LORA, star flash, and high-speed carrier HPLC, and supports wired and wireless multi-link parallel transmission and reception; the communication module is compatible with DC12V and AC220V power supply and adopts DIN rail mounting.

4. The method for identifying the oil level in an oil mist lubricator according to claim 3, characterized in that, The deep learning image recognition model is built based on the YOLO real-time object detection model and specifically includes: Collect oil cup photos at different oil level heights to construct an oil level image dataset; The images in the dataset are labeled, including at least all liquid level areas and the current liquid level area. For reducer lubricators, additional marker labels are also added. Data augmentation was performed on the labeled images to expand the dataset; The YOLO segmentation model was trained using the expanded dataset to obtain the oil level recognition model; Calculate the oil level percentage using the following formula: ； in Indicates the current liquid level height. This indicates the total liquid level height.

5. The method for identifying the oil level in an oil mist lubricator according to claim 4, characterized in that, The data enhancement includes rotation transformation and brightness transformation; the rotation transformation changes by 0.5° to 2° each time, and the brightness transformation changes by 5% to 10% each time.

6. The method for identifying the oil level in an oil mist lubricator according to claim 4, characterized in that, The YOLOv8X-SEG model was used for training, with the following parameters set: initial training image size (640, 480) and multi-scale training enabled; 5 warm-up training epochs; recalculated anchor boxes; bounding box weights set to 1.2; IOU threshold set to 0.7; mosaic intensity set to 0.3; copypaste ratio set to 0.3; scaling range set to 0.3; perspective transformation parameter set to 0.001; and mask downsampling ratio set to 2.

7. The method for identifying the oil level in an oil mist lubricator according to claim 3, characterized in that, It also includes a step to handle oil level identification errors: In operating mode, the percentage of oil level identified last time is recorded. If the current identification result is higher than the previous identification result and the difference is not greater than 1%, the previous oil level percentage is maintained as the current identification result.

8. The method for identifying the oil level in an oil mist lubricator according to claim 5, characterized in that, The anomalies include abnormally high oil levels and abnormal oil level change trends; The abnormal oil level refers to an oil level higher than 66.6% or lower than 33.3%, at which point a real-time alarm will be generated; The abnormal oil level change trend refers to the following: when analyzing the historical oil level curve of the past 15 days, if the oil level change of two consecutive days is lower than the lower limit of the daily decrease or higher than the upper limit of the daily decrease, a real-time warning will be generated. The initial value of the daily lower limit is 0.05%, and the initial value of the daily upper limit is 3.0%. The daily lower limit and the daily upper limit can be dynamically adjusted according to the number of device actions.

9. The method for identifying the oil level in an oil mist lubricator according to claim 8, characterized in that, The real-time alarm or real-time warning is accompanied by voice broadcast; the historical oil level change curve is displayed in the form of a bar chart.

10. The method for identifying the oil level in an oil mist lubricator according to claim 1, characterized in that, It also includes a manual triggering step: in response to the user's manual triggering command, the camera is immediately controlled to capture an image of the current oil level, and the deep learning image recognition model is invoked to return the oil level recognition result.

Images