Nozzle leakage detection method, system, and semiconductor apparatus

Abstract

This invention provides a nozzle leakage detection method, system, and semiconductor device, comprising: receiving a detection trigger signal output by a control system within a target time window during which the nozzle stops spraying liquid and does not move; responding to the detection trigger signal, acquiring an original full-domain image including the nozzle within the cavity, wherein the original full-domain image is directly output by an image acquisition unit without local cropping processing; inputting the original full-domain image into a target detection model for leakage identification, obtaining an identification result, wherein the target detection model is a deep learning model; and sending the identification result to the control system so that the control system controls the operation of the semiconductor device according to the identification result. This invention abandons the traditional preprocessing method of region of interest cropping, and directly detects the original full-domain image output by the image acquisition unit, completely solving the problems of missing edge nozzle cropping and detection failure caused by equipment installation deviation and structural offset, achieving a detection range without blind spots.

Description

Technical Field

[0001] This invention relates to the field of semiconductor manufacturing and testing technology, and in particular to a nozzle leakage detection method, system and semiconductor equipment. Background Technology

[0002] In semiconductor wafer manufacturing processes, semiconductor equipment such as chemical cleaning, photolithography, wet etching, and packaging cleaning all rely on spray nozzles to precisely spray chemical solutions or pure water. The working condition of the nozzles directly affects the wafer processing yield and equipment lifespan. After spraying stops, problems such as valve core seal wear and residual pipeline pressure often lead to chemical drips at the nozzle or continuous trace leakage. When these drips accumulate and fall onto the wafer surface below, they can cause defects such as localized chemical corrosion, particulate contamination, and pattern failure. In severe cases, they can even corrode internal components of the cavity, increasing equipment maintenance costs.

[0003] Currently, the mainstream automated detection solutions for nozzle dripping and dripping defects are usually traditional machine vision solutions based on threshold segmentation and template matching. These solutions generally use fixed Region of Interest (ROI) clipping preprocessing and only detect nozzles within the preset area. When there are installation deviations, mechanical thermal expansion and contraction, or swing arm structure offsets, edge nozzles are prone to move out of the preset area, directly causing missed detections.

[0004] Therefore, it is necessary to propose a nozzle leakage detection method, system, and semiconductor device to solve the above problems. Summary of the Invention

[0005] The purpose of this invention is to provide a nozzle leakage detection method, system, and semiconductor device to address the problems of missing edge nozzle trimming and detection failure caused by equipment installation deviations and structural offsets.

[0006] This invention provides a nozzle leakage detection method, applied to semiconductor devices with nozzles, the method comprising: Receive the detection trigger signal output by the control system within the target time window when the nozzle stops spraying liquid and does not move; In response to the detection trigger signal, a raw global image containing the nozzle inside the cavity is acquired. The raw global image is directly output by the image acquisition unit without any local cropping processing. The original global image is input into the target detection model for leakage identification to obtain the identification result, wherein the target detection model is a deep learning model; The identification result is sent to the control system so that the control system controls the operation of the semiconductor device based on the identification result.

[0007] In one possible embodiment, the target detection model is a deep learning model trained on a nozzle sample dataset.

[0008] In one possible embodiment, the nozzle sample dataset includes nozzle samples with no dripping and no leakage, nozzle samples with different degrees of dripping, nozzle samples with different degrees of leakage, and samples of scenarios involving cavity reflection, light fluctuation, and liquid residue interference.

[0009] In one possible embodiment, the object detection model is a lightweight object detection model based on the YOLO architecture.

[0010] In one possible embodiment, the target detection model is an inference engine model compiled by an inference acceleration tool.

[0011] In one possible embodiment, after the semiconductor device is powered on, it automatically loads the inference engine model and performs multi-frame sample pre-inference to complete model warm-up.

[0012] In one possible embodiment, during the period when the detection trigger signal is continuously valid, the detection process of image acquisition, model recognition, and result transmission is executed cyclically to achieve multi-frame continuous sampling detection; when the detection trigger signal terminates, the detection process is stopped, and the monitoring standby state of the detection trigger signal is maintained.

[0013] In one possible embodiment, the identification result is sent to the control system so that the control system controls the operation of the device according to the identification result, specifically as follows: If the identification result indicates that the nozzle is normal, the control system controls the semiconductor device to proceed to the next process; If the identification result indicates that there is dripping or leakage in the nozzle, the control system controls the semiconductor device to stop and / or triggers an alarm.

[0014] This invention also provides a nozzle leakage detection system, applied to a semiconductor device with a nozzle. The semiconductor device includes a control system, a cavity, a nozzle, and an image acquisition unit. The nozzle leakage detection system includes a visual algorithm module, a system function module, and a functional function module. The vision algorithm module has a built-in target detection model, which is a deep learning model trained on a nozzle sample dataset. It is used to receive the original global image and perform nozzle leakage identification, and output the identification result. The system function module is communicatively connected to the control system, the image acquisition unit, and the vision algorithm module, and is used to: receive the detection trigger signal output by the control system within the target time window when the nozzle stops spraying liquid and does not move; provide the low-level reading interface of the image acquisition unit; receive the recognition result output by the vision algorithm module and forward the recognition result to the control system. The functional module is communicatively connected to the system functional module and the vision algorithm module, respectively. It is used to respond to the detection trigger signal forwarded by the system functional module, call the underlying reading interface of the system functional module to obtain the original global image containing the nozzle in the cavity. The original global image is directly output by the image acquisition unit without local cropping. The original global image is sent to the vision algorithm module to perform leakage identification, and the identification result output by the vision algorithm module is pushed to the system functional module to complete the detection process scheduling.

[0015] This invention also provides a semiconductor device, characterized in that it includes: Such as the nozzle leakage detection system in the above embodiments; The cavity contains a nozzle. An image acquisition unit, located in the cavity, is used to acquire a raw global image including the nozzles within the cavity; The control system is used to manage the timing of equipment processes, outputs a detection trigger signal within a target time window when the nozzle stops spraying liquid and does not move, and controls the operation of the semiconductor device based on the recognition result. The beneficial effects of this invention are as follows: 1. Full-area cropping-free detection, eliminating missed detections in certain areas: Abandoning the traditional preprocessing method of region of interest cropping, it directly detects the original full-area image output by the image acquisition unit, completely solving the problems of missing edge nozzles and detection failures caused by equipment installation deviations and structural offsets, achieving a detection range without blind spots.

[0016] 2. High detection accuracy and adaptability to complex industrial conditions: This invention adopts a deep learning target detection model trained on a dedicated sample set. It is optimized for small target defects such as tiny drips and residual leakage in semiconductor nozzles and is adapted to complex industrial scenarios such as cavity reflection, light fluctuation, and interference from residual liquid. Compared with traditional vision algorithms, it significantly reduces the false detection rate and false detection rate.

[0017] 3. Fast inference speed, meeting online real-time detection needs: The target detection model is converted into a high-speed inference engine format by the inference acceleration tool. Combined with a lightweight network structure, the inference speed of a single frame full image is greatly improved. With the pre-warm-up mechanism at startup, inference latency is eliminated, which can meet the needs of high-speed automated detection in production lines.

[0018] 4. Precise process linkage and high degree of automation: The detection is triggered by the detection signal of the control system, so as to achieve precise synchronization between the spraying process and the detection process, avoid invalid detection, and realize fully automated closed-loop control of "detection-judgment-alarm-stop" without manual intervention.

[0019] 5. Reduce production losses and save production costs: It can identify defects such as trace amounts of dripping liquid from nozzles in real time, provide early warnings and shutdowns, avoid dripping liquid from contaminating wafers and corroding equipment, significantly improve product yield, and reduce equipment maintenance costs and production scrap costs. Attached Figure Description

[0020] Figure 1 This is a schematic flowchart of the nozzle leakage detection method of the present invention.

[0021] Figure 2 This is a schematic diagram of the training process for the target detection model of the present invention.

[0022] Figure 3 This is a schematic diagram of the model reasoning process of the present invention.

[0023] Figure 4 This is a schematic diagram of the signal interaction between the nozzle leakage detection system and the control system of the present invention.

[0024] Figure 5 This is a schematic diagram of the semiconductor device of the present invention.

[0025] Explanation of reference numerals in the attached figures: 510, Nozzle leakage detection system; 511, Model layer; 5111, Visual algorithm module; 5112, System function module; 5113, Functional function module; 512, Display layer; 513, Scheduling layer; 514, Server; 520, Control system; 530, Image acquisition unit. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, 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.

[0027] To address the problems existing in the above-mentioned technologies, embodiments of the present invention provide a nozzle leakage detection method, applicable to semiconductor devices with nozzles, see [link to relevant documentation]. Figure 1 The method includes: S101: Receives the detection trigger signal output by the control system 520 within the target time window when the nozzle stops spraying liquid and does not move; S102: In response to the detection trigger signal, acquire the original global image containing the nozzle inside the cavity. The original global image is directly output by the image acquisition unit 530 without local cropping processing. S103: Input the original global image into the target detection model to identify leakage and obtain the identification result. The target detection model is a deep learning model. S104: The identification result is sent to the control system 520 so that the control system 520 controls the operation of the semiconductor device according to the identification result.

[0028] In this embodiment, by directly reading the original complete image output by the image acquisition unit 530, the fixed region of interest cropping is eliminated, and the complete scene information is preserved. This enables the target detection model to automatically identify nozzles in the entire image and determine whether there are drips or leaks. This fundamentally eliminates the problem of missed detection of edge nozzles caused by equipment installation deviations and structural offsets, and achieves detection coverage without blind spots.

[0029] In one embodiment, the target detection model is a deep learning model trained on a nozzle sample dataset.

[0030] Furthermore, the nozzle sample dataset includes nozzle samples with no dripping and no leakage, nozzle samples with different degrees of dripping, nozzle samples with different degrees of leakage, as well as samples with interference from cavity reflection, light fluctuation, and liquid residue.

[0031] In this embodiment, a target detection model trained on a nozzle sample dataset is used to identify leaks in the original global image. The sample dataset covers multiple interference scenarios and is optimized for small targets such as tiny drips and trace leaks. This allows the model to learn the morphology, edge, and semantic features of drips and leaks, rather than simple pixel grayscale differences. This effectively distinguishes interference signals such as cavity reflection, light fluctuations, and residual liquid, significantly reducing false detection and false detection rates and improving the accuracy and robustness of nozzle leak detection under complex conditions.

[0032] In one embodiment, the object detection model is a lightweight object detection model based on the YOLO (You Only Look Once, single-stage object detection) architecture, such as the YOLO26m object detection model, the YOLOv5 object detection model, the YOLOv7 object detection model, or the YOLOv13 object detection model.

[0033] See Figure 2 The training process of the lightweight object detection model based on the YOLO architecture will be explained in detail below with reference to specific embodiments.

[0034] We collected nozzle image samples under various working conditions on site, covering normal nozzles without dripping, nozzles with slight dripping, nozzles with heavy dripping, and nozzles with residual leakage after spraying. We also included samples from interfering scenarios such as metal reflection in the cavity, light intensity fluctuations, and residual water stains on the cavity wall. We manually and accurately labeled all samples to mark the location and type of defects, and constructed a nozzle-specific defect dataset. We then divided the dataset into training, validation, and test sets according to the proportions and generated a dataset configuration file.

[0035] The YOLO26m lightweight target detection model was selected as the base network, and training hyperparameters were configured to suit the small target characteristics of tiny droplets: a data augmentation strategy for small targets was enabled, anchor boxes were re-clustered based on the actual size of droplets in the field, the weights of localization loss and confidence loss for tiny drip and leakage samples were increased, and the model’s feature extraction capability for small target defects was enhanced.

[0036] The model is initialized by loading pre-trained weights and multiple training iterations are performed. Each iteration sequentially completes forward propagation, loss function calculation, backpropagation, and weight update. After each training iteration, the model performance is evaluated using a validation set, and core metrics such as precision, recall, and small target detection rate are monitored. After training, the weight file with the best comprehensive performance on the validation set is saved as the dedicated detection model.

[0037] In another embodiment, the object detection model is an inference engine model compiled by an inference acceleration tool. The inference engine model is generated by performing network layer fusion, weight quantization, hardware kernel tuning, and redundant operator pruning on the trained deep learning model by the inference acceleration tool. For example, the YOLO26m-engine object detection model.

[0038] In terms of inference acceleration tools, in addition to TensorRT, OpenVINO, ONNX Runtime, or other model inference acceleration frameworks can also be used, as long as they can convert the trained deep learning model into a high-speed inference engine format that can run efficiently on the target hardware.

[0039] In one specific embodiment, after the semiconductor device is powered on, it automatically loads the inference engine model and performs multi-frame sample pre-inference to complete the model warm-up.

[0040] In this embodiment, by converting the target detection model into a high-speed inference engine format using an inference acceleration tool and selecting a lightweight model structure, the inference time for a single frame of the full-domain image is significantly reduced. At the same time, by eliminating cold start delay through pre-inference warm-up, the detection system can complete multi-frame continuous sampling detection within a short target time window after the nozzle stops spraying liquid, meeting the online real-time rapid detection requirements of the production line.

[0041] See Figure 3The reasoning process of the inference engine model will be explained in detail below with reference to specific embodiments.

[0042] After the device is powered on, the detection system automatically loads the inference engine model and actively performs multi-frame sample pre-inference, completing operations such as memory allocation, operator initialization, and calculation cache establishment. This eliminates the cold start first-frame delay that is common in deep learning models, ensuring a stable and consistent response speed during formal detection.

[0043] Upon receiving the detection trigger signal, the system captures the current original global image and inputs it into the preheated inference engine model to perform forward inference. Post-processing is then performed on the candidate boxes output by the model, filtering by confidence threshold and eliminating invalid interference boxes by non-maximum suppression, distinguishing between real defects and interference from reflections and water stains, and outputting the final recognition result.

[0044] The identification results are transmitted back to the control system 520 in real time for closed-loop management, and simultaneously written to the database to retain the original full-domain images, detection logs and alarm records for production traceability and subsequent model iteration and optimization.

[0045] In one embodiment, see Figure 4 During the period when the detection trigger signal is valid, the detection process of image acquisition, model recognition, and result transmission is executed cyclically to achieve multi-frame continuous sampling detection; when the detection trigger signal is terminated, the detection process is stopped and the system remains in a standby state for listening to the detection trigger signal.

[0046] In this embodiment, by receiving the detection trigger signal output by the control system 520 within the target time window when the nozzle stops spraying liquid and does not move, and cyclically executing the detection process of image reading, model recognition, and result transmission during the continuous validity period of the detection trigger signal, precise timing synchronization between the detection action and the process steps is achieved. The detection results are transmitted back to the control system 520 in real time, and the control system 520 executes process flow, alarm, or emergency shutdown, forming a fully automated closed-loop control of "detection-judgment-alarm-shutdown", without any manual intervention. The role of multi-frame continuous sampling is to offset the instantaneous interference in single-frame images. The sloshing of liquid inside the cavity, the brief obstruction of steam, and the fluctuation of metal reflection may all lead to missed detection or false detection in a single frame. Multi-frame sampling can be matched with multi-frame result fusion rules, such as triggering an alarm only when defects are detected in multiple consecutive frames, thereby improving detection robustness. At the same time, the real-time transmission of results frame by frame ensures that the control system 520 responds immediately after defect detection, without waiting for the detection window to end.

[0047] In some specific embodiments, the detection trigger signal is a continuous high-level signal, a pulse signal, an industrial communication protocol message, or other forms of control signal, as long as it can identify the target time window when "the nozzle stops spraying liquid and does not move".

[0048] In one embodiment, the identification result is sent to the control system 520 so that the control system 520 controls the operation of the equipment according to the identification result. Specifically, if the identification result is that the nozzle is normal, the control system 520 controls the semiconductor equipment to enter the next process; if the identification result is that the nozzle has dripping or leakage, the control system 520 controls the semiconductor equipment to stop and / or triggers an alarm.

[0049] In this embodiment, by triggering shutdown and / or alarm at the initial stage of nozzle dripping and leakage defects, the leakage of liquid can be prevented from contaminating the wafer and corroding the equipment, which helps to improve product yield and reduce equipment operation and maintenance costs and production scrap costs.

[0050] In some embodiments, the capture range of the original global image can be used to cover a single nozzle, multiple nozzles, or all nozzles within the cavity to adapt to the detection station layout of different devices.

[0051] In one embodiment, the identification result includes defect category, location coordinates, and defect confidence level. The defect category includes no abnormality, dripping, and leakage. The location coordinates are the pixel coordinates of the defect box in the original global image. The confidence level represents the degree to which the model judges the actual existence of the defect.

[0052] In one specific embodiment, when the confidence level of a defect is greater than or equal to the confidence level threshold, it is determined to be a real defect and is synchronously fed back to the control system 520; when the confidence level of a defect is less than the confidence level threshold, the defect is filtered out and determined to be interference.

[0053] In addition, the present invention also proposes a nozzle leakage detection system 510, which is applied to semiconductor devices with nozzles, see [link to relevant documentation]. Figure 5 The semiconductor device includes a control system 520, a cavity, a nozzle, and an image acquisition unit 530. The nozzle leakage detection system 510 includes a server 514 for deploying detection software at the hardware level. The control system 520 manages the timing of equipment processes; for example, when the nozzle stops spraying liquid and does not move within a target time window, it outputs a detection trigger signal to the server 514, and upon receiving the identification result, controls the semiconductor device to proceed to the next process, stop, or trigger an alarm. The image acquisition unit 530 continuously outputs a high-definition real-time video stream and transmits the original complete image to the server 514.

[0054] In one embodiment, at the software level, the detection software adopts a Model-View-Controller (MVC) layered architecture, including a model layer 511, a display layer 512, and a scheduling layer 513. The model layer 511 includes a visual algorithm module 5111, a system function module 5112, and a functional function module 5113. The server 514 is responsible for running the visual algorithm module 5111, the system function module 5112, and the functional function module 5113, and for handling data interaction between the image acquisition unit 530, the control system 520, and the external database. The visual algorithm module 5111 has a built-in target detection model, which is a deep learning model trained based on a nozzle sample dataset. This model receives the original global image, performs nozzle leakage identification, and outputs the identification results. Specifically, the visual algorithm module 5111 directly receives the original global image acquired by the server 514, performs forward inference, automatically identifies dripping and leakage defects in all nozzles, and outputs the defect location, category, and confidence level. System function module 5112 is communicatively connected to control system 520, image acquisition unit 530, and vision algorithm module 5111, and is used for: receiving detection trigger signals output by control system 520 within the target time window when the nozzle stops spraying liquid and does not move; providing the low-level reading interface of image acquisition unit 530; managing the loading and preheating of target detection model; receiving the recognition results output by vision algorithm module 5111 and forwarding the recognition results to control system 520. Function module 5113 is communicatively connected to system function module 5112 and vision algorithm module 5111, and is used for responding to the detection trigger signals forwarded by system function module 5112, calling the low-level reading interface of system function module 5112 to obtain the original full-domain image containing the nozzle inside the cavity. The original full-domain image is directly output by image acquisition unit 530 without local cropping processing. The original full-domain image is sent to vision algorithm module 5111 to perform leakage recognition, and the recognition results output by vision algorithm module 5111 are pushed to system function module 5112 to complete the detection process scheduling.

[0055] Furthermore, the display layer 512 is configured with an overview page, video page, detection page, settings page, and alarm page to visualize the detection status, configure parameters, and display alarm information. The scheduling layer 513 undertakes the functions of button control, function call, status update, and data transmission, realizing the decoupled interaction between the display layer 512 and the visual algorithm module 5111.

[0056] In other embodiments of the present invention, a semiconductor device is also disclosed, see [link to relevant documentation] Figure 5The semiconductor device includes: a nozzle leakage detection system 510 as described in the above embodiment, a cavity, an image acquisition unit 530, and a control system 520. A nozzle is disposed inside the cavity. The image acquisition unit 530 is located in the cavity and is used to acquire a raw global image including the nozzle within the cavity. The control system 520 is used to manage the device's process timing and to output a detection trigger signal within a target time window when the nozzle stops spraying liquid and does not move, and to control the operation of the semiconductor device based on the recognition result. In some embodiments, the image acquisition unit 530 is installed on the outside of the cavity and captures the scene inside the cavity through the cavity's sealed window. The image acquisition unit 530 is one of a CCD camera, a CMOS camera, an industrial camera, an area scan camera, or a line scan camera.

[0057] It should be noted that the control system 520 can be any of the following: programmable logic controller (PLC), programmable automation controller (PAC), industrial computer (IPC) with soft PLC real-time system, DCS distributed control system 520, semiconductor equipment dedicated embedded main controller, FPGA motion controller, edge controller. It can output detection trigger signal during the nozzle stationary period and receive visual detection results to realize closed-loop control of the equipment.

[0058] It is understandable that the semiconductor equipment includes various types of semiconductor equipment with spray nozzles, such as single-wafer rotary cleaners, tank wet cleaners, photolithography coating trackers, wet etching equipment, and packaging wafer cleaning machines. The installation position of the image acquisition unit 530 and the model sample are adjusted according to the nozzle shape and cavity structure of the equipment to adapt to the nozzle leakage detection of various semiconductor equipment.

[0059] The flow of the nozzle leakage detection method of the present invention will be described in detail below with specific examples.

[0060] Step 1: System Hardware Deployment and Image Acquisition An image acquisition unit 530 is fixedly installed on the outside of the semiconductor device cavity. The angle and focal length of the image acquisition unit 530 are adjusted so that the captured image completely covers all nozzles in the cavity. The image acquisition unit 530 continuously outputs a high-definition real-time video stream. The detection software directly reads the original video stream image without performing any region of interest cropping or region capture operations, thus preserving complete scene information.

[0061] Step 2: Model training, conversion, and pre-deployment Images of the cleaning machine nozzles under various operating conditions were collected on-site, including samples of normal nozzles without dripping, nozzles with slight dripping, nozzles with heavy dripping, and nozzles with residual leakage after spraying, all with different swing arms and nozzle types. These samples were then manually and precisely labeled to create a dedicated nozzle defect dataset. Based on this dataset, a lightweight YOLO deep learning model was iteratively trained to obtain a dedicated target detection model adapted for nozzle leak detection. After training, the target detection model was converted into an inference engine model using an inference acceleration tool and deployed to the detection software terminal. Upon device startup, the detection software automatically loads the inference engine model and performs multi-frame pre-inference to warm up the model, eliminating initialization delays and ensuring the real-time performance and stability of subsequent detection.

[0062] Step 3: Control system 520 process signal linkage triggering When the semiconductor equipment is running normally, each nozzle completes a single nozzle spraying cleaning process. After the spraying stops and before the nozzle moves, the control system 520 of the semiconductor equipment automatically outputs a continuous detection trigger signal to the detection software. The detection software listens to the signal in real time. During the period when the detection trigger signal is received, it captures the original full-area image from the video stream at the maximum frame rate and starts a single detection process for each captured image.

[0063] Step 4: Real-time image capture and defect inference detection After receiving the trigger signal, the detection software responds to the detection trigger signal by immediately capturing the original full-domain image of the current moment from the real-time video stream of the image acquisition unit 530. The original full-domain image is then input into the preheated inference engine model for forward inference. The model automatically identifies whether there are dripping or leakage defects in each nozzle in the image and outputs the corresponding detection results.

[0064] Step 5: Feedback of test results and closed-loop control of equipment The detection software uploads the detection results to the control system 520 in real time: if the detection result is no dripping or dripping, the control system 520 determines that the equipment is in normal working condition and controls the semiconductor equipment to enter the next production process; if the detection identifies dripping or leakage defects in the nozzle, the control system 520 immediately triggers the equipment's audible and visual alarm and controls the semiconductor equipment to stop urgently, locking the current process to prevent abnormal conditions from continuing production and causing wafer contamination, while also prompting maintenance personnel to check for nozzle malfunctions.

[0065] This invention aims to solve the problems of high manual costs, poor robustness of traditional algorithms, easy omissions during region-of-interest (ROI) clipping, and inability to achieve automated detection linked to equipment processes in existing semiconductor equipment nozzle leak detection. This invention achieves high-precision, high-real-time, full-coverage, and fully automated closed-loop nozzle leak detection through original full-domain image input, a specially trained target detection model, inference acceleration and preheating mechanisms, and 520-process linkage triggering of the control system. It can be adapted to automated operation of process formulations, effectively reducing wafer defect rates and equipment failure risks. Details are as follows: 1. By detecting the original global image without region of interest cropping, the problem of missed detection of edge nozzles caused by equipment installation offset and cavity position deviation is completely solved.

[0066] 2. The deep learning model trained on the nozzle sample dataset is compiled by the inference acceleration tool and deployed at high speed, eliminating inference initialization delay and balancing detection accuracy and industrial real-time performance.

[0067] 3. The control system features precise timing-triggered closed-loop detection, forming a fully automatic closed-loop control system for detection and identification, result feedback, alarm / shutdown / normal operation.

[0068] 4. Specific optimizations were made for small-scale defects such as tiny drips and minor leaks in semiconductor nozzles, addressing the pain points of high false detection and false negative rates in traditional machine vision threshold / template algorithms.

[0069] While embodiments of the present invention have been described in detail above, it will be apparent to those skilled in the art that various modifications and variations can be made to these embodiments. However, it should be understood that such modifications and variations fall within the scope and spirit of the invention as set forth in the claims. Furthermore, the invention described herein may have other embodiments and can be implemented or carried out in various ways. Unless otherwise defined, the technical or scientific terms used herein should be understood in their ordinary sense by one of ordinary skill in the art to which this invention pertains. The terms "comprising" and similar expressions used herein mean that the element or object preceding the word encompasses the element or object listed following the word and its equivalents, but do not exclude other elements or objects.

Claims

1. A method for detecting nozzle leakage, applied to semiconductor devices with nozzles, characterized in that, include: Receive the detection trigger signal output by the control system within the target time window when the nozzle stops spraying liquid and does not move; In response to the detection trigger signal, a raw global image containing the nozzle inside the cavity is acquired. The raw global image is directly output by the image acquisition unit without any local cropping processing. The original global image is input into the target detection model for leakage identification to obtain the identification result, wherein the target detection model is a deep learning model; The identification result is sent to the control system so that the control system controls the operation of the semiconductor device based on the identification result.

2. The nozzle leakage detection method according to claim 1, characterized in that, The target detection model is a deep learning model trained on a nozzle sample dataset.

3. The nozzle leakage detection method according to claim 2, characterized in that, The nozzle sample dataset includes nozzle samples with no dripping and no leakage, nozzle samples with different degrees of dripping, nozzle samples with different degrees of leakage, as well as samples of interference scenarios such as cavity reflection, light fluctuation, and liquid residue.

4. The nozzle leakage detection method according to claim 2, characterized in that, The object detection model is a lightweight object detection model based on the YOLO architecture.

5. The nozzle leakage detection method according to any one of claims 1-4, characterized in that, The target detection model is an inference engine model compiled by an inference acceleration tool.

6. The nozzle leakage detection method according to claim 5, characterized in that, After the semiconductor device is powered on, it automatically loads the inference engine model and performs multi-frame sample pre-inference to complete the model warm-up.

7. The nozzle leakage detection method according to any one of claims 1-4, characterized in that, During the period when the detection trigger signal is valid, the detection process of image acquisition, model recognition, and result transmission is executed cyclically to achieve multi-frame continuous sampling detection; when the detection trigger signal is terminated, the detection process is stopped and the monitoring standby state of the detection trigger signal is maintained.

8. The nozzle leakage detection method according to any one of claims 1-4, characterized in that, The identification result is sent to the control system, so that the control system controls the operation of the device according to the identification result, specifically as follows: If the identification result indicates that the nozzle is normal, the control system controls the semiconductor device to proceed to the next process; If the identification result indicates that there is dripping or leakage in the nozzle, the control system controls the semiconductor device to stop and / or triggers an alarm.

9. A nozzle leakage detection system, applied to a semiconductor device with a nozzle, said semiconductor device comprising a control system, a cavity, a nozzle, and an image acquisition unit, characterized in that, The nozzle leakage detection system includes a vision algorithm module, a system function module, and a function module. The vision algorithm module has a built-in target detection model, which is a deep learning model trained on a nozzle sample dataset. It is used to receive the original global image and perform nozzle leakage identification, and output the identification result. The system function module is communicatively connected to the control system, the image acquisition unit, and the vision algorithm module, and is used to: receive the detection trigger signal output by the control system within the target time window when the nozzle stops spraying liquid and does not move; provide the low-level reading interface of the image acquisition unit; receive the recognition result output by the vision algorithm module and forward the recognition result to the control system. The functional module is communicatively connected to the system functional module and the vision algorithm module, respectively. It is used to respond to the detection trigger signal forwarded by the system functional module, call the underlying reading interface of the system functional module to obtain the original global image containing the nozzle in the cavity. The original global image is directly output by the image acquisition unit without local cropping. The original global image is sent to the vision algorithm module to perform leakage identification, and the identification result output by the vision algorithm module is pushed to the system functional module to complete the detection process scheduling.

10. A semiconductor device, characterized in that, include: The nozzle leakage detection system as described in claim 9; The cavity contains a nozzle. An image acquisition unit, located in the cavity, is used to acquire a raw global image including the nozzles within the cavity; The control system is used to manage the timing of equipment processes, outputs a detection trigger signal within a target time window when the nozzle stops spraying liquid and does not move, and controls the operation of the semiconductor device based on the recognition result.

Images