Chip mounter component position self-correction system and method based on deep learning

Abstract

The invention belongs to the technical field of electronic manufacturing, and particularly discloses a chip mounter element position self-correction system and method based on deep learning. The system comprises a data acquisition module, a data processing and deep learning module and a control and correction module. The method comprises the following steps: synchronously acquiring image, pressure and displacement data through a data acquisition module, inputting the preprocessed data into a trained convolutional neural network model, identifying element position deviation in combination with multi-dimensional data, generating a control signal to adjust the position of a mounting head, and iteratively optimizing the model through an incremental learning algorithm. The device can correct the position deviation of the element in real time, improve the mounting precision and stability, reduce the reject ratio, adapt to element and environment changes in the production process, and improve the electronic manufacturing efficiency and quality.

Description

Technical Field

[0001] This invention relates to the field of electronic manufacturing technology, specifically to a self-calibration system and method for component position in a chip mounter based on deep learning. Background Technology

[0002] In the electronics manufacturing industry, pick-and-place machines are crucial equipment for accurately placing surface-mount components onto printed circuit board (PCB) pads. Their placement accuracy directly impacts the quality and performance of electronic products. As electronic products continue to evolve towards miniaturization and high density, the requirements for component positioning accuracy in pick-and-place machines are becoming increasingly stringent. Traditional pick-and-place machines primarily rely on pre-set mechanical parameters and simple visual recognition algorithms for component positioning. However, in actual production, due to factors such as mechanical wear, environmental temperature variations, and differences in incoming components, the actual placement position of components often deviates from the preset position after a period of operation. This leads to an increased defect rate, and in severe cases, requires extensive manual rework, reducing production efficiency and increasing costs. While some existing methods attempt to address the positional deviation problem through periodic calibration and manual intervention, these methods suffer from drawbacks such as low efficiency and difficulty in consistently maintaining accuracy. For example, traditional template-matching-based visual positioning methods are prone to misjudging components with similar shapes and sizes and cannot adapt to subtle changes in component appearance; periodic calibration requires downtime, affecting the continuity of the production line. Therefore, developing a system and method capable of real-time and accurate self-calibration of component positions in a pick-and-place machine is of significant practical importance. Summary of the Invention

[0003] The purpose of this invention is to provide a component position self-correction system and method for a pick-and-place machine based on deep learning. By using deep learning algorithms to analyze and process various data during the operation of the pick-and-place machine in real time, the system automatically detects and corrects component position deviations, improves the placement accuracy and stability of the pick-and-place machine, reduces the placement defect rate, and enhances the production efficiency and product quality of electronic manufacturing.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a self-calibration system for component position of a chip mounter based on deep learning, comprising a data acquisition module, a data processing and deep learning module, and a control and calibration module.

[0005] The data acquisition module includes a high-precision industrial camera, a pressure sensor, and a displacement sensor. The high-precision industrial camera is used to acquire image information of the component gripping and placement process, the pressure sensor is used to detect the gripping and placement pressure, and the displacement sensor is used to acquire the movement trajectory data of the placement head. The collaborative acquisition of multi-dimensional data (visual, pressure, and displacement) by multiple sensors provides a comprehensive information foundation for deviation identification, improving detection accuracy and reliability.

[0006] The data processing and deep learning module receives the output data from the data acquisition module, preprocesses the image data, and inputs it into the trained convolutional neural network model. This model then combines pressure and displacement data to identify component position deviations. The deep learning model can automatically extract complex features and, through comprehensive analysis of multi-source data, significantly improve the accuracy and adaptability of deviation identification.

[0007] The control and correction module generates placement head movement adjustment commands based on the position deviation results output by the data processing and deep learning module to achieve real-time component position correction. This achieves millisecond-level dynamic response, ensuring placement accuracy down to the micrometer level, while reducing manual intervention and improving production efficiency.

[0008] As a preferred embodiment of the present invention, the high-precision industrial camera is equipped with an adjustable light source, which includes a ring light source and a strip light source. This light source can automatically switch between light source types and adjust the light intensity and angle according to the size, color, and surface reflectivity of the components, thereby optimizing the clarity of component edges and feature points in the image. The adaptive light source system can eliminate interference such as reflections and shadows, ensuring stable image quality and providing high signal-to-noise ratio input data for deep learning models.

[0009] As a preferred embodiment of the present invention, the convolutional neural network model of the data processing and deep learning module includes an input layer, multiple convolutional layers, multiple pooling layers, a fully connected layer, and an output layer. The input layer receives preprocessed image data; the convolutional layers extract the shape, texture, and contour features of elements in the image using convolutional kernels of different sizes; the pooling layers use max pooling to reduce the dimensionality of the feature maps; the fully connected layers fuse the extracted features; and the output layer outputs the positional and rotational deviations of the elements in the X and Y axes. This multi-level network structure balances local features with global correlations, max pooling reduces computational cost, and the fully connected layers achieve end-to-end output of multi-dimensional deviations.

[0010] In a preferred embodiment of the present invention, the control and correction module includes a motion control unit and a drive unit. The motion control unit has a built-in deviation compensation algorithm, which can generate motion control commands containing velocity and acceleration parameters based on the magnitude and direction of the position deviation. The drive unit includes a servo motor and a driver. After receiving the motion control command, the driver drives the servo motor to coordinate the X-axis, Y-axis, and Z-axis moving and rotating components of the mounting head. Dynamic parameterized control commands enable multi-axis coordinated precision motion, the servo system ensures execution accuracy, and the overall error compensation capability is superior to traditional PID control.

[0011] This invention also proposes a deep learning-based self-calibration method for component position in a chip mounter, comprising the following steps: S1. Data Acquisition: Images, pressure, and displacement data of the component placement process are simultaneously acquired using industrial cameras, pressure sensors, and displacement sensors. Simultaneous acquisition of multi-source data provides a spatiotemporal consistency basis for subsequent analysis, avoiding misjudgments caused by asynchronous data.

[0012] S2. Deviation identification: After denoising and enhancing the image data, it is input into the trained deep learning model, which calculates the component position deviation by combining pressure and displacement data. Preprocessing improves data quality, and multimodal data fusion analysis reduces single sensor errors, making deviation identification more robust.

[0013] S3. Real-time correction: Based on the deviation, a motion control signal for the placement head is generated to adjust the spatial coordinates and rotation angle of the placement head. Closed-loop real-time correction shortens the response delay, ensuring a continuous and efficient placement process, suitable for high-speed production line requirements.

[0014] S4. Model Iteration: After accumulating a preset number of mounting data, the parameters of the deep learning model are updated using an incremental learning algorithm. The model is continuously optimized to adapt to new component types and operating conditions, avoiding the performance degradation caused by data drift in traditional models.

[0015] As a preferred embodiment of the present invention, during the data acquisition process in step S1, the acquired image, pressure, and displacement data are timestamped with a timestamp accuracy at the microsecond level. A data synchronization verification mechanism is used to eliminate data with timestamp deviations exceeding a preset threshold, ensuring the consistency of image, pressure, and displacement data corresponding to the same mounting action in the time dimension. Microsecond-level time synchronization eliminates phase differences during multi-sensor data fusion, ensuring the spatiotemporal accuracy of the analysis results.

[0016] As a preferred embodiment of the present invention, the training samples of the deep learning model in step S2 cover components of various packaging types, including chip resistors, chip capacitors, and integrated circuit chips; the samples contain scene data of components under different postures (tilt, flip) and different environmental conditions (lighting changes, dust interference), and each scene data includes the standard position parameters of the component and the corresponding deviation annotation information. Diverse training samples enhance the model's generalization ability, enabling it to adapt to the changing scenarios in complex industrial environments.

[0017] In a preferred embodiment of the present invention, the placement head motion control signal in step S3 is generated through a closed-loop control algorithm, which includes a position feedback loop and a speed feedback loop. The position feedback loop collects the actual position information of the placement head in real time and compares it with the target position information. The speed feedback loop collects the actual movement speed of the placement head and compares it with the commanded speed. Based on the deviation between the two, the pulse frequency and pulse number of the control signal are adjusted in real time. The dual closed-loop control (position + speed) achieves dynamic error suppression, improving the accuracy by more than 30% compared to open-loop control.

[0018] As a preferred embodiment of the present invention, the incremental learning algorithm in step S4 employs knowledge distillation technology, by constructing a teacher model and a student model. The teacher model is the original deep learning model that has been trained, while the student model uses newly collected mounting data and soft labels output by the teacher model as training samples. This process retains the teacher model's ability to identify historical component types while learning new component types and positional deviation characteristics under new operating conditions. Knowledge distillation avoids catastrophic forgetting, and the fusion of old and new knowledge enables the model to continuously evolve.

[0019] As a preferred embodiment of the present invention, the method further includes a temperature compensation unit, which comprises a temperature sensor and a compensation algorithm module. The temperature sensor is distributed on the mounting head's motion guide rail, the displacement sensor's mounting position, and the PCB board placement platform, collecting temperature data from multiple points in real time. The compensation algorithm module performs linear or nonlinear correction on the displacement sensor's measurement data based on the temperature data and a preset temperature-error mapping relationship, thereby eliminating the impact of mechanical deformation caused by temperature changes on measurement accuracy. Multi-point temperature drift compensation offsets thermal deformation errors, enabling the system to maintain stability in a wide temperature range environment, with accuracy fluctuations less than ±1μm / ℃.

[0020] Compared with the prior art, the beneficial effects of the present invention are: This invention uses deep learning algorithms to analyze and process various types of data in real time, which can accurately detect and correct component position deviations. Compared with traditional methods, it greatly reduces the placement defect rate and improves the quality of electronic products.

[0021] The system can monitor and correct component positions in real time during the operation of the pick-and-place machine, without the need for manual calibration or periodic mechanical calibration, thus improving production efficiency and ensuring the continuity of the production line.

[0022] Deep learning models can adapt to the effects of different types of components, different batches of PCBs, and changes in the production environment by continuously learning new data samples. They have strong versatility and adaptability, reducing the workload of frequently adjusting the parameters of the pick-and-place machine due to changes in components or production conditions.

[0023] By improving mounting accuracy and production efficiency, reducing defect rates and labor rework costs, and minimizing production losses due to equipment downtime for calibration, the overall production cost of electronic manufacturing has been reduced. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the system of the present invention; Figure 2 This is a schematic diagram of the method of the present invention. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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.

[0026] Example 1 Please see Figure 1 This embodiment provides a technical solution: a self-calibration system for component position of a chip mounter based on deep learning, including a data acquisition module, a data processing and deep learning module, and a control and calibration module.

[0027] The data acquisition module includes a high-precision industrial camera, a pressure sensor, and a displacement sensor. The high-precision industrial camera is used to acquire image information of the component gripping and placement process, the pressure sensor is used to detect the gripping and placement pressure, and the displacement sensor is used to acquire the movement trajectory data of the placement head. The collaborative acquisition of multi-dimensional data (visual, pressure, and displacement) by multiple sensors provides a comprehensive information foundation for deviation identification, improving detection accuracy and reliability.

[0028] The data processing and deep learning module receives the output data from the data acquisition module, preprocesses the image data, and inputs it into the trained convolutional neural network model. This model then combines pressure and displacement data to identify component position deviations. The deep learning model can automatically extract complex features and, through comprehensive analysis of multi-source data, significantly improve the accuracy and adaptability of deviation identification.

[0029] The control and correction module generates placement head movement adjustment commands based on the position deviation results output by the data processing and deep learning module to achieve real-time component position correction. This achieves millisecond-level dynamic response, ensuring placement accuracy down to the micrometer level, while reducing manual intervention and improving production efficiency.

[0030] The high-precision industrial camera is equipped with an adjustable light source, including a ring light source and a strip light source. This light source can automatically switch between different types and adjust the light intensity and angle based on the size, color, and surface reflectivity of the components, thereby optimizing the clarity of component edges and feature points in the image. The adaptive light source system eliminates interference such as reflections and shadows, ensuring stable image quality and providing high signal-to-noise ratio input data for deep learning models.

[0031] The convolutional neural network model of the data processing and deep learning module includes an input layer, multiple convolutional layers, multiple pooling layers, a fully connected layer, and an output layer. The input layer receives preprocessed image data. The convolutional layers extract the shape, texture, and contour features of elements in the image using convolutional kernels of different sizes. The pooling layers use max pooling to reduce the dimensionality of the feature maps. The fully connected layer fuses the extracted features. The output layer outputs the positional and rotational deviations of the elements along the X and Y axes. This multi-level network structure balances local features with global correlations, max pooling reduces computational cost, and the fully connected layers provide end-to-end output of multi-dimensional deviations.

[0032] The control and correction module includes a motion control unit and a drive unit. The motion control unit has a built-in deviation compensation algorithm, which can generate motion control commands containing velocity and acceleration parameters based on the magnitude and direction of the position deviation. The drive unit includes a servo motor and a driver. After receiving the motion control command, the driver drives the servo motor to coordinate the X, Y, and Z axis moving and rotating parts of the mounting head. Dynamic parameterized control commands enable multi-axis coordinated precision motion, the servo system ensures execution accuracy, and the overall error compensation capability is superior to traditional PID control.

[0033] The deep learning-based component position self-calibration system for a pick-and-place machine consists of three core modules. The data acquisition module uses a high-precision industrial camera, pressure sensor, and displacement sensor to collect image information (including component posture, position, and PCB markers), placement pressure data (to determine component gripping and placement status), and placement head motion trajectory data during the component picking and placement process. The data processing and deep learning module uses a GPU server as its computing core. After preprocessing the image data, such as enhancement and denoising, it inputs the data into a convolutional neural network (CNN) model trained on a large number of samples. This model combines pressure and displacement data to complete component feature recognition and position deviation judgment. The control and calibration module generates control commands based on the analysis results. When the position deviation exceeds the limit, it automatically adjusts the coordinate position, angle, and other motion parameters of the placement head to achieve real-time correction. The calibration information is then fed back to the pick-and-place machine control system to ensure subsequent placement accuracy.

[0034] Example 2 Please see Figure 2 This embodiment proposes a deep learning-based self-calibration method for component position in a chip mounter, comprising the following steps: S1. Data Acquisition: Images, pressure, and displacement data of the component placement process are simultaneously acquired using industrial cameras, pressure sensors, and displacement sensors. Simultaneous acquisition of multi-source data provides a spatiotemporal consistency basis for subsequent analysis, avoiding misjudgments caused by asynchronous data.

[0035] S2. Deviation identification: After denoising and enhancing the image data, it is input into the trained deep learning model, which calculates the component position deviation by combining pressure and displacement data. Preprocessing improves data quality, and multimodal data fusion analysis reduces single sensor errors, making deviation identification more robust.

[0036] S3. Real-time correction: Based on the deviation, a motion control signal for the placement head is generated to adjust the spatial coordinates and rotation angle of the placement head. Closed-loop real-time correction shortens the response delay, ensuring a continuous and efficient placement process, suitable for high-speed production line requirements.

[0037] S4. Model Iteration: After accumulating a preset number of mounting data, the parameters of the deep learning model are updated using an incremental learning algorithm. The model is continuously optimized to adapt to new component types and operating conditions, avoiding the performance degradation caused by data drift in traditional models.

[0038] In step S1, during the data acquisition process, the acquired image, pressure, and displacement data are timestamped with a precision of microseconds. A data synchronization verification mechanism is used to eliminate data with timestamp deviations exceeding a preset threshold, ensuring the consistency of image, pressure, and displacement data corresponding to the same mounting action across the time dimension. Microsecond-level time synchronization eliminates phase differences during multi-sensor data fusion, ensuring the spatiotemporal accuracy of the analysis results.

[0039] The training samples for the deep learning model described in step S2 cover components of various package types, including chip resistors, chip capacitors, and integrated circuit chips. The samples contain scene data of components under different orientations (tilt, flip) and environmental conditions (lighting variations, dust interference), and each scene data includes the component's standard position parameters and corresponding deviation annotation information. This diverse training sample enhances the model's generalization ability, enabling it to adapt to the changing scenarios in complex industrial environments.

[0040] The placement head motion control signal mentioned in step S3 is generated through a closed-loop control algorithm, which includes a position feedback loop and a speed feedback loop. The position feedback loop collects the actual position information of the placement head in real time and compares it with the target position information. The speed feedback loop collects the actual movement speed of the placement head and compares it with the commanded speed. Based on the deviation between the two, the pulse frequency and pulse number of the control signal are adjusted in real time. The dual closed-loop control (position + speed) achieves dynamic error suppression, improving the accuracy by more than 30% compared to open-loop control.

[0041] The incremental learning algorithm described in step S4 employs knowledge distillation, constructing a teacher model and a student model. The teacher model is the original, fully trained deep learning model, while the student model uses newly collected mounting data and soft labels output by the teacher model as training samples. This process retains the teacher model's ability to identify historical component types while learning new component types and positional deviation characteristics under new operating conditions. Knowledge distillation avoids catastrophic forgetting, and the fusion of old and new knowledge enables the model to continuously evolve.

[0042] The method further includes a temperature compensation unit, which comprises a temperature sensor and a compensation algorithm module. The temperature sensor is distributed along the mounting head's motion guide rail, the displacement sensor's mounting position, and the PCB board placement platform, collecting temperature data from multiple points in real time. The compensation algorithm module performs linear or nonlinear correction on the displacement sensor's measurement data based on the temperature data and a preset temperature-error mapping relationship, eliminating the impact of mechanical deformation caused by temperature changes on measurement accuracy. Multi-point temperature drift compensation offsets thermal deformation errors, enabling the system to maintain stability over a wide temperature range, with accuracy fluctuations less than ±1μm / ℃.

[0043] The deep learning-based component position self-calibration method for pick-and-place machines comprises four stages. In the data acquisition stage, the data acquisition module operates continuously, with an industrial camera capturing images at a certain frame rate, and pressure and displacement sensors acquiring corresponding data in real time, all transmitted to the data processing module via a high-speed interface. In the deep learning analysis stage, images are preprocessed and then input into the trained model. Multi-dimensional data is combined to identify the component position status and calculate the magnitude and direction of deviations (e.g., by comparing the vertex position of a rectangular component with a standard position to determine the deviation). In the position calibration stage, based on the deviation results, the motion and rotation motors of the placement head along the X, Y, and Z axes are adjusted to achieve coordinate or angle correction, and the placement parameters are updated. In the model update stage, new placement data (including correct and incorrect samples) is continuously collected, and the model is periodically retrained to adapt to component differences and environmental changes during the production process, maintaining the model's accuracy and adaptability.

[0044] In summary, the technical solution of the present invention will be further described in detail with reference to specific embodiments.

[0045] Implementation Environment Setup: A Yamaha YSM40R pick-and-place machine was selected and modified on the SMT production line of an electronics manufacturing company. A data acquisition module was installed, including two 5-megapixel Baslerace series high-precision industrial cameras to capture images during component gripping and placement; an HBM1-C9C / 10N pressure sensor with an accuracy of ±0.1g was installed at the contact point between the placement head and the component; and a Renishaw XL-80 laser interferometer displacement sensor with an accuracy of ±0.5μm was installed on the motion axis of the placement head. The data acquisition module was connected to a Dell PowerEdge R750 server equipped with an NVIDIA Tesla V100 GPU to run the data processing and deep learning modules. The control and calibration module was connected to the pick-and-place machine's control system via PROFINET industrial Ethernet. Deep learning model training: A large amount of image data, along with corresponding pressure and displacement data, was collected for various component types (such as common package types like 0402, 0603, and SOT-23) under correct and incorrect mounting conditions. This data was labeled, including the component type, correct position coordinates, angle, and actual deviation. The labeled data was divided into training, validation, and test sets in a 7:2:1 ratio. A deep learning model based on a convolutional neural network was built using the TensorFlow deep learning framework. The model structure included 5 convolutional layers, 3 pooling layers, and 2 fully connected layers. The model was trained iteratively for 50 rounds on the training set, adjusting the model parameters to achieve an accuracy of over 95% on the validation set. Finally, the model was tested on the test set to verify its performance. System Operation and Effect Verification: During normal production of the pick-and-place machine, a deep learning-based component position self-calibration system was activated. The data acquisition module collected real-time image, pressure, and displacement data, while an industrial camera captured images at 30 frames per second. The data processing and deep learning module analyzed the data, and the control and calibration module calibrated the placement head position based on the analysis results. After a week of continuous operation and testing, the placement defect rate of the pick-and-place machine was statistically analyzed. Compared to before the system was installed, the placement defect rate decreased from 3% to 0.5%, and production efficiency increased by 20%, effectively verifying the effectiveness and practicality of the invention.

[0046] It is worth noting that the entire device is controlled by a master control button. Since the device matched with the control button is a common device and belongs to existing mature technology, its electrical connection relationship and specific circuit structure will not be described in detail here.

[0047] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A component position self-calibration system for a chip mounter based on deep learning, characterized in that: It includes a data acquisition module, a data processing and deep learning module, and a control and correction module; The data acquisition module includes a high-precision industrial camera, a pressure sensor, and a displacement sensor. The high-precision industrial camera is used to acquire image information of the component gripping and placement process, the pressure sensor is used to detect the component gripping and placement pressure, and the displacement sensor is used to acquire the motion trajectory data of the placement head. The data processing and deep learning module is used to receive the output data of the data acquisition module, preprocess the image data and input it into the trained convolutional neural network model, and combine pressure and displacement data to identify the position deviation of the component. The control and correction module is used to generate placement head motion adjustment commands based on the position deviation results output by the data processing and deep learning module to achieve real-time correction of component positions.

2. The self-calibration system for component position of a chip mounter based on deep learning according to claim 1, characterized in that: The high-precision industrial camera is equipped with an adjustable light source, which includes a ring light source and a strip light source. It can automatically switch the light source type and adjust the light intensity and light angle according to the size, color and surface reflectivity of the component to optimize the clarity of the component edges and feature points in the image.

3. The self-calibration system for component position of a chip mounter based on deep learning according to claim 1, characterized in that: The convolutional neural network model of the data processing and deep learning module includes an input layer, multiple convolutional layers, multiple pooling layers, a fully connected layer, and an output layer. The input layer receives preprocessed image data. The convolutional layers extract the shape, texture, and contour features of elements in the image using convolutional kernels of different sizes. The pooling layers reduce the dimension of the feature map using max pooling. The fully connected layer fuses the extracted features. The output layer outputs the positional deviation and rotational angle deviation of the elements in the X and Y axes.

4. The component position self-calibration system for a chip mounter based on deep learning according to claim 1, characterized in that: The control and correction module includes a motion control unit and a drive unit. The motion control unit has a built-in deviation compensation algorithm, which can generate motion control commands including speed parameters and acceleration parameters according to the magnitude and direction of the position deviation. The drive unit includes a servo motor and a driver. After receiving the motion control command, the driver drives the servo motor to drive the X-axis, Y-axis, Z-axis moving parts and rotating parts of the mounting head to perform coordinated actions.

5. A self-calibration method for component position in a pick-and-place machine based on deep learning, characterized in that, Includes the following steps: S1. Data acquisition: Simultaneously acquire images, pressure and displacement data of the component placement process through industrial cameras, pressure sensors and displacement sensors; S2, Deviation Recognition: After denoising and enhancing the image data, it is input into the trained deep learning model, and the deviation of the component position is calculated by combining the pressure and displacement data. S3. Real-time correction: Generate a placement head motion control signal based on the deviation amount to adjust the spatial coordinates and rotation angle of the placement head. S4. Model Iteration: After accumulating a preset number of mounting data, the parameters of the deep learning model are updated using an incremental learning algorithm.

6. The component position self-calibration system for a chip mounter based on deep learning according to claim 5, characterized in that: In the data acquisition process described in step S1, the acquired image, pressure, and displacement data are timestamped with a timestamp accuracy of microseconds. Through a data synchronization verification mechanism, data with timestamp deviations exceeding a preset threshold are removed to ensure the consistency of image, pressure, and displacement data corresponding to the same mounting action in the time dimension.

7. The component position self-calibration system for a chip mounter based on deep learning according to claim 5, characterized in that: The training samples of the deep learning model in step S2 cover a variety of components with different packaging types, including chip resistors, chip capacitors, and integrated circuit chips; the samples contain scene data of components under different postures and environmental conditions, and each scene data contains the standard position parameters of the component and the corresponding deviation annotation information.

8. A component position self-calibration system for a chip mounter based on deep learning according to claim 5, characterized in that: The placement head motion control signal mentioned in step S3 is generated by a closed-loop control algorithm, which includes a position feedback loop and a speed feedback loop. The position feedback loop collects the actual position information of the placement head in real time and compares it with the target position information. The speed feedback loop collects the actual movement speed of the placement head and compares it with the command speed. Based on the deviation between the two, the pulse frequency and pulse number of the control signal are adjusted in real time.

9. A component position self-calibration system for a chip mounter based on deep learning according to claim 5, characterized in that: The incremental learning algorithm described in step S4 employs knowledge distillation technology by constructing a teacher model and a student model. The teacher model is the original deep learning model that has been trained, while the student model uses newly collected mounting data and soft labels output by the teacher model as training samples. While retaining the teacher model's ability to identify historical component types, it learns new component types and positional deviation characteristics under new operating conditions.

10. A component position self-calibration system for a chip mounter based on deep learning according to claim 5, characterized in that: The method further includes a temperature compensation unit, which includes a temperature sensor and a compensation algorithm module; the temperature sensor is distributed on the mounting head motion guide rail, the displacement sensor mounting position and the PCB board placement platform to collect temperature data at multiple points in real time; The compensation algorithm module performs linear or nonlinear corrections on the displacement sensor's measurement data based on the temperature data and the preset temperature-error mapping relationship, in order to eliminate the impact of mechanical deformation caused by temperature changes on measurement accuracy.

Images