A high-precision placement head adaptive thermal compensation system and method for chip placement machines

Abstract

The present invention relates to the technical field of electronic manufacturing equipment, and more specifically, to an adaptive thermal compensation system and method for a high-precision placement head of a chip placement machine. The system comprises a temperature-deformation sensing unit, a thermal-mechanical coupling analysis unit, a dynamic compensation control unit, and a closed-loop execution unit. The system uses the temperature-deformation sensing unit to collect temperature and deformation data of multiple components of the placement head in real time. The thermal-mechanical coupling analysis unit reconstructs a three-dimensional temperature field based on a finite element algorithm and dynamically calculates the thermal expansion distribution, solving the model coarseness problem of traditional single-point temperature measurement. The dynamic compensation control unit uses online parameter identification and a long-short-term memory network model to predict the amount of thermal drift within the next 5 milliseconds, and adaptively adjusts the compensation strategy based on motion conditions. The closed-loop execution unit decomposes the compensation amount into displacement and torsion correction instructions to achieve precise offset of thermal deformation, construct a full-process thermal compensation closed loop, improve the precision and stability of the placement head in complex thermal environments, and enhance production efficiency.

Description

Technical Field

[0001] The present invention relates to the technical field of electronic manufacturing equipment, and in particular to a self-adaptive thermal compensation system and method for a high-precision placement head of a chip placement machine. Background Art

[0002] Electronic manufacturing equipment is an important technology, among which placement machines are very important in electronic manufacturing. The accuracy of the placement head directly affects the accuracy of component placement. However, in actual production, when the current placement machine operates continuously at high frequency, due to the coupling of mechanical movement friction, ambient temperature fluctuations and heating of electronic components, key components of the placement head, such as the Z-axis screw and rotating spindle, will produce nonlinear temperature drift.

[0003] Traditional systems rely solely on single-point temperature sensors to monitor the local temperature of the placement head, failing to capture the three-dimensional temperature gradient distribution of the Z-axis lead screw, rotating spindle, and vacuum generator. When the Z-axis lead screw moves at high speed, frictional heating in the threaded pair causes a non-uniform increase in the axial temperature gradient. Traditional models simplify this as a uniform temperature rise, ignoring the differential thermal expansion along the length of the lead screw, leading to errors in the compensation calculation. Furthermore, the thermal expansion coefficient of the placement head components is significantly affected by ambient humidity and load pressure. Traditional systems use a fixed-parameter heat-deformation transfer function (HTTF) and are unable to dynamically adapt to changing operating conditions. When the placement head switches from low-load chip placement to high-load component crimping, changes in the spindle's radial stiffness lead to increased differential thermal expansion. At high speeds, the placement head can experience accelerations exceeding 2g. The coupled effects of mechanical vibration and thermal expansion result in high-frequency nonlinear thermal drift. Traditional static model-based prediction methods are unable to capture short-term thermal drift trends on the order of 5ms, causing compensation commands to lag behind actual thermal deformation. To address this issue, we provide a high-precision adaptive thermal compensation system and method for placement heads in placement machines. Summary of the Invention

[0004] The object of the present invention is to provide a high-precision placement head adaptive thermal compensation system and method for a placement machine, so as to solve the problems raised in the above background technology.

[0005] 1. Because traditional systems cannot capture three-dimensional temperature gradient distribution, resulting in errors in compensation calculation, this case uses a temperature-deformation sensing unit to obtain multi-component data, and a thermal-mechanical coupling analysis unit to fuse the data and dynamically calculate the three-dimensional thermal expansion distribution. This can accurately calculate the compensation amount and improve placement accuracy.

[0006] 2. Because traditional static models cannot capture high-frequency nonlinear thermal drift trends, resulting in compensation lag, this case introduces a long-short-term memory network model through a dynamic compensation control unit to output a predicted value of thermal drift within the next 5ms. This can predict thermal drift in advance, achieve high-frequency dynamic compensation, and reduce position error.

[0007] To achieve the above purpose, a high-precision placement head adaptive thermal compensation system for a placement machine is provided, comprising the following units:

[0008] The temperature-deformation sensing unit includes a temperature sensor and a deformation sensor, which are used to obtain the temperature sensing data, axial displacement deviation and radial offset of the placement head, rotating spindle and vacuum generator respectively;

[0009] The thermal-mechanical coupling analysis unit has a built-in temperature field reconstruction algorithm based on the finite element heat conduction equation. By integrating temperature sensor data, axial displacement deviation, radial offset and material thermal expansion coefficient database, it dynamically calculates the three-dimensional thermal expansion distribution of the placement head Z-axis screw, rotating spindle and vacuum generator.

[0010] The dynamic compensation control unit introduces an online recursive least squares parameter identifier to receive the three-dimensional thermal expansion distribution. It then updates the thermal-deformation transfer function coefficient matrix in real time based on the current placement head acceleration and load pressure. This matrix, along with the process timing data set, serves as the input to the long-short-term memory network model. The three-dimensional thermal expansion distribution is then time-series expanded to output the thermal drift within the next 5 ms.

[0011] The closed-loop execution unit uses the thermal drift as the compensation amount and decomposes it into displacement command and radial torsion correction amount.

[0012] As a further improvement of the present technical solution, the execution of the temperature field reconstruction algorithm includes the following steps:

[0013] The adaptive wedge-shaped grid is divided based on the thread pitch of the Z-axis lead screw of the placement head. The grid density matches the layout position of the temperature sensor. The temperature sensor data is input into the finite element heat conduction equation as the boundary condition, and the temperature gradient field is solved by the Newton-Raphson iteration method. The thermal expansion coefficient of the adaptive wedge-shaped grid nodes is corrected in combination with the axial displacement deviation. When outputting the three-dimensional thermal expansion distribution, a dynamic compensation weight is applied to the axial strain of the Z-axis lead screw.

[0014] As a further improvement of this technical solution, the method for updating the material thermal expansion coefficient database is:

[0015] During the startup phase of the placement machine, the heating module is controlled to apply a calibration reward of step temperature rise to the Z-axis lead screw of the placement head. The actual deformation at each temperature rise step is measured by a laser displacement meter and compared with the theoretical deformation. When the error exceeds 5%, the dynamic correction value is calculated according to the set mathematical formula, and the dynamic correction value is associated with the current ambient humidity and load pressure to establish a four-dimensional lookup table.

[0016] As a further improvement of the present technical solution, the dynamic calculation of the three-dimensional thermal expansion distribution includes an abnormal data cleaning step:

[0017] The temperature sensor data is filtered using a sliding window. The window length is synchronized with the placement head movement frequency. After time alignment of the axial displacement deviation and the radial offset, sensor failure data is eliminated through cross-correlation analysis. When the temperature difference of the piston cavity of the vacuum generator is detected to be greater than the preset temperature difference threshold, the infrared thermal imaging module is activated for secondary verification. The cleaned data stream is transmitted to the thermal-mechanical coupling analysis unit through the message queue priority.

[0018] As a further improvement of this technical solution, a method for suppressing the coupling of thermal expansion and mechanical vibration is also included:

[0019] The frequency domain characteristics of vibration acceleration are introduced into the heat-deformation transfer function. When a sudden change in the axial strain of the placement head lead screw is detected, accompanied by high-frequency vibration above 200 Hz, a reverse compensation waveform is generated to offset the thermal shock effect. The convergence of the coupling model is verified by the Lyapunov stability criterion.

[0020] As a further improvement of the present technical solution, the operating method of the online recursive least squares parameter identifier includes:

[0021] A time-varying parameter vector is constructed, including the axial stiffness of the screw, the radial stiffness of the spindle, and the sealing loss coefficient of the vacuum chamber. A recursive update rule with a forgetting factor is set. When the parameter mutation exceeds three times the standard deviation of the historical mean, the Bayesian probability model is triggered to perform root cause analysis of the fault.

[0022] As a further improvement of this technical solution, the training method of the long short-term memory network model is:

[0023] A 50ms time series segment is intercepted with a step size of 5ms as the time axis, and the temperature gradient, thermal expansion strain, motion acceleration, and ambient humidity are extracted as feature dimensions respectively. A spatiotemporal feature matrix is ​​constructed based on the time axis and feature dimensions.

[0024] A multi-head attention mechanism is used to enhance the key feature extraction capability, and a composite loss function is defined. Finally, model quantization and compression are implemented on the embedded tensor processing unit chip, and the inference delay is controlled within the preset threshold.

[0025] As a further improvement of the present technical solution, the method for generating the process time series data set includes:

[0026] During the operation of the placement machine, motion parameters, process parameters and environmental parameters are synchronously recorded. The motion parameters include X / Y / Z axis position, velocity and acceleration; the process parameters include nozzle vacuum, component quality and placement pressure; and the environmental parameters include temperature, humidity and vibration spectrum.

[0027] The multi-sensor timestamps of the original data are aligned based on the dynamic time warping algorithm, and the adversarial generative network is used to expand the small sample operating condition data. Then, the dimensionality reduction visualization is used to detect the data clustering characteristics through t-distributed random neighborhood embedding. Finally, a data-physical joint verification mechanism is established. When the deviation between the predicted compensation amount and the theoretical value exceeds three times the standard deviation, the abnormal operating condition is automatically marked and active learning is initiated.

[0028] As a further improvement of the present technical solution, the dynamic adjustment strategy of the compensation amount of the thermal drift amount is:

[0029] Define the axial compensation priority factor. When it is detected that the placement head is in a rapid acceleration or deceleration state, increase the compensation output frequency from 200Hz to 500Hz. When the absolute value of the acceleration is greater than 2 times the acceleration of gravity, it is determined to be a rapid acceleration or deceleration state. The positioning accuracy after compensation is verified through the closed-loop node of the Lissajous figure. If the residual error is greater than 1.5μm, switch to the reinforcement learning compensation mode.

[0030] A second object of the present invention is to provide a method for implementing the above-mentioned high-precision placement head adaptive thermal compensation system of a placement machine, comprising the following steps:

[0031] S1. Use temperature sensors to monitor the temperature data of the Z-axis screw, rotating spindle and vacuum generator of the placement head in real time, and use deformation sensors to synchronously obtain the axial displacement deviation and radial offset;

[0032] S2, based on the thread pitch of the placement head, divides the adaptive wedge grid, inputs the temperature sensor data into the finite element heat conduction equation to iteratively solve the three-dimensional temperature gradient field, and dynamically corrects the thermal-mechanical coupling parameters of the grid nodes based on the material thermal expansion coefficient;

[0033] S3. Use the recursive least squares method with a forgetting factor to update the thermal-deformation transfer function coefficient matrix in real time, and generate a predicted value of the thermal drift within the next 5 ms by fusing the process timing data with the long short-term memory neural network model;

[0034] S4, decomposing the predicted thermal drift into a displacement command of the piezoelectric ceramic micro-motion platform and a radial torsion correction of the servo motor, and adaptively adjusting the compensation frequency based on the acceleration state;

[0035] S5. Verify the positioning accuracy after compensation through Lissajous graph nodes. If the residual error exceeds the threshold, switch to reinforcement learning compensation mode. When the sensor confidence decreases, enable the digital twin simulation module and federated learning to achieve system self-healing upgrade.

[0036] Compared with the prior art, the present invention has the following beneficial effects:

[0037] The thermal-mechanical coupling analysis unit integrates multi-source sensor data with the material thermal expansion coefficient database, and dynamically calculates the three-dimensional thermal expansion distribution based on the finite element heat conduction equation. It can accurately capture the non-uniform temperature field and thermal deformation difference of the Z-axis lead screw and the rotating spindle of the placement head, solve the problem of rough compensation model caused by traditional single-point temperature measurement, and provide a reliable data basis for high-precision compensation. The dynamic compensation control unit updates the thermal-deformation transfer function in real time through an online recursive least squares parameter identifier, and uses a long-short-term memory network model to perform time series expansion on the thermal expansion data. It can predict the thermal drift within the next 5ms, and dynamically adjust the compensation strategy by combining motion acceleration and load pressure, thereby improving the system's response speed to nonlinear thermal drift in high-frequency continuous operations, shortening the compensation lag time, and maintaining placement positioning accuracy under complex working conditions of rapid acceleration and deceleration, effectively solving the problem of compensation lag of traditional static models. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is an overall block diagram of the present invention;

[0039] Figure 2 It is the overall flow chart of the present invention.

[0040] The meaning of each number in the figure is:

[0041] 1. Temperature-deformation sensing unit; 2. Thermal-mechanical coupling analysis unit; 3. Dynamic compensation control unit; 4. Closed-loop execution unit. DETAILED DESCRIPTION

[0042] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0043] Example 1

[0044] The present invention provides a high-precision placement head adaptive thermal compensation system for chip placement machines. Figure 1-Figure 2 As shown, it includes the following units:

[0045] The temperature-deformation sensing unit 1 includes a temperature sensor and a deformation sensor, which are used to obtain temperature sensing data, axial displacement deviation and radial offset of the placement head, rotating spindle and vacuum generator respectively;

[0046] The thermal-mechanical coupling analysis unit 2 has a built-in temperature field reconstruction algorithm based on the finite element heat conduction equation. By integrating temperature sensor data, axial displacement deviation, radial offset and material thermal expansion coefficient database, it dynamically calculates the three-dimensional thermal expansion distribution of the placement head Z-axis screw, rotating spindle and vacuum generator.

[0047] The execution of the temperature field reconstruction algorithm includes the following steps:

[0048] The traditional uniform grid cannot accurately capture the temperature changes in key areas. The grid density needs to be optimized according to the thread structure and sensor position to adjust the thread pitch of the Z-axis screw of the placement head. Generate an adaptive wedge mesh for the benchmark and define the base mesh size , a local encryption strategy is used in the thread top and bottom areas, and the mesh size , the grid density matches the temperature sensor layout, and the sensor position is calculated To the nearest grid node distance ,when When Generate radius for center The mesh size in this area is , while ensuring the calculation accuracy, reduce the number of grids and improve the calculation efficiency. The heat conduction process of the placement head has nonlinear characteristics, and an efficient iterative method is required to solve the heat conduction equation. The temperature sensor data is input as the boundary condition into the finite element heat conduction equation, and the Newton-Raphson iteration method is used to solve the temperature gradient field. The finite element heat conduction equation is ;in, is the material density, is the specific heat capacity, is the thermal conductivity, is the internal heat source, is the temperature value, For time, Expressed as Hamiltonian operator, in the heat conduction equation, Indicates the divergence of heat flux density, reflecting the net outflow or inflow of heat per unit volume. Axial displacement deviation will affect the thermal expansion characteristics of the material. The thermal expansion coefficient needs to be dynamically corrected to improve the reconstruction accuracy. A laser displacement sensor is used to measure the axial displacement of the Z-axis screw in real time. , calculate the displacement deviation ,in If it is the theoretical displacement value, then the thermal expansion coefficient of the adaptive wedge grid node is corrected in combination with the axial displacement deviation to set the basic thermal expansion coefficient , the corrected formula is ;in is the correction factor, To compensate for the influence of axial stress on thermal expansion after correction, improve the accuracy of thermal deformation prediction, and apply dynamic compensation weight to the axial strain of the Z-axis screw when outputting the three-dimensional thermal expansion distribution. The influence of axial strain on placement accuracy varies under different working conditions, so it is necessary to dynamically adjust the compensation strategy and define the working condition parameters, including feed speed. ,load and the temperature change rate , the dynamic weight is ;in, is the shape parameter, then the axial strain compensation ; The original length of the Z-axis screw is used to adjust the Z-axis position command in real time through the servo control system. ;in is the actual position control signal output by the control system to the servo motor, The theoretical target position can achieve optimal compensation under different working conditions and improve the robustness of the system.

[0049] The update method of the material thermal expansion coefficient database is:

[0050] Traditional single-point calibration cannot cover the full temperature range characteristics, and the thermal expansion characteristic curve needs to be obtained through step temperature rise. During the startup phase of the placement machine, the heating module is controlled to apply a step temperature rise calibration reward to the Z-axis lead screw of the placement head. The initial temperature, end temperature, temperature rise step and PID controller parameters are set. Data is collected after each temperature point is stable for 15 minutes. The sampling frequency is 10Hz, and 100 data points are collected in each group. The laser displacement meter has a measurement range of ±2mm and a resolution of 0.1μm. Zero point calibration is performed three times before measurement to cover the actual operating temperature range of the placement machine and obtain a continuous thermal expansion characteristic curve. Environmental factors will cause deviations between the theoretical model and the actual thermal expansion behavior, which need to be corrected in real time. The actual deformation variable under each temperature rise step is measured by the laser displacement meter and compared with the theoretical deformation variable. When the error exceeds 5%, the dynamic correction value is calculated according to the set mathematical formula. The theoretical deformation variable ; is the temperature rise step, the actual deformation ; Relative error ;when When the correction value ,in In order to correct the coefficient, adaptively compensate for the influence of environmental factors, and improve the accuracy of thermal expansion prediction, a two-dimensional table cannot express the influence of multiple factors coupling. It is necessary to establish a multi-dimensional mapping relationship, define the humidity dimension, temperature dimension, load dimension and correction value dimension respectively, associate the dynamic correction value with the current ambient humidity and load pressure, establish a four-dimensional lookup table, and reduce database storage space.

[0051] The dynamic calculation of the three-dimensional thermal expansion distribution includes the following steps to clean up abnormal data:

[0052] Traditional fixed window filtering cannot adapt to the variable speed movement of the placement head, resulting in dynamic response lag. Sliding window filtering is used for temperature sensor data. The window length is synchronized with the movement frequency of the placement head, and a mapping relationship between the movement frequency and the window length is established. The movement frequency is obtained in real time through the servo motor encoder of the placement head. When the movement frequency changes by more than ±20%, the window length is smoothly adjusted within 3 sampling cycles to avoid filter oscillation caused by mutations. The data smoothness is maintained at high-frequency movement, and the mutation characteristics are retained at low frequency. The asynchronous sampling frequency of the sensor will cause data timing disorder, affecting the accuracy of thermal expansion calculation. After the axial displacement deviation and the radial offset are time-aligned, the sensor failure data is eliminated through cross-correlation analysis. The GPS timing module is used to unify the clocks of each sensor with a synchronization accuracy of ±10μs. Linear interpolation is used for asynchronous data alignment, and the cross-correlation function of the axial displacement and radial offset is calculated. When the maximum correlation coefficient is less than 0.6 and the duration is greater than 2 seconds, the corresponding sensor is marked as failed, eliminating the sensor asynchronous error and automatically identifying hardware faults. Abnormal temperature differences in key components may indicate potential failures and require high-precision verification. When the temperature difference in the piston cavity of the vacuum generator is detected to be greater than the preset temperature difference threshold, the infrared thermal imaging module is activated for secondary verification. A piston cavity temperature difference threshold is set. When the real-time temperature difference exceeds the piston cavity temperature difference threshold, the infrared module is triggered. The infrared thermal imager startup time is less than 300ms, the spatial resolution is 0.1mm, and the temperature accuracy is ±0.3°C. A temperature field mapping function is established and Kalman filter fusion is used. Double verification improves the reliability of anomaly detection and accurately locates hotspots. Real-time data streams must ensure that key data is processed first to avoid wasting computing resources. The cleaned data stream is transmitted to the thermal-mechanical coupling analysis unit 2 through the message queue priority. Three levels of priority are defined: piston cavity temperature difference data, Z-axis thermal expansion data, and ambient temperature and humidity data. When the queue is full, data is discarded in order. The most recent 50 piston cavity temperature difference data items are retained, and the ID of the most recently processed message is recorded. After system recovery, processing continues from the breakpoint, ensuring zero loss of key data and improving the system's overload resistance.

[0053] Methods for suppressing the coupling of thermal expansion and mechanical vibration:

[0054] The traditional transfer function ignores the influence of vibration on heat transfer, and the frequency domain features need to be introduced to improve the accuracy of the model. The frequency domain features of vibration acceleration are introduced into the heat-deformation transfer function to obtain the corrected transfer function, and the frequency domain sweep method is used to obtain the vibration characteristic parameters. The sweep range is 10-500Hz, and the step size is 5Hz. The transfer function parameters are fitted by the least squares method, and the fitting error is less than 5%. The superposition of high-frequency vibration and thermal shock will aggravate the positioning error, and the offset waveform needs to be generated in real time. When the axial strain mutation of the Z-axis screw of the placement head is detected and accompanied by high-frequency vibration above 200Hz, a reverse compensation waveform is generated to offset the thermal shock effect. The trigger condition is judged as the axial strain mutation rate And high-frequency vibration energy ;in It is the spectral density function of vibration acceleration. The nonlinear characteristics of the coupled system may lead to divergence, which requires mathematical verification of stability. The convergence of the coupled model is verified by the Lyapunov stability criterion. The coupled model is composed of multiple interrelated parts. Convergence means that as time goes by or the number of iterations increases, the output of the model can tend to a certain stable value or stable state. Verifying the convergence of the coupled model by the Lyapunov stability criterion is to use the relevant theories and methods of the criterion to determine whether the coupled model will stably tend to a specific result during operation without irregular fluctuations or divergence.

[0055] The dynamic compensation control unit 3 introduces an online recursive least squares parameter identifier to receive the three-dimensional thermal expansion distribution, and updates the heat-deformation transfer function coefficient matrix in real time in combination with the current motion acceleration and load pressure of the placement head. It then uses it and the process timing data set as the input of the long-short-term memory network model, performs time series expansion on the three-dimensional thermal expansion distribution, and outputs the thermal drift within the next 5ms.

[0056] The method for running the online recursive least squares parameter identifier includes:

[0057] The traditional fixed parameter model cannot reflect the performance degradation of the equipment. It is necessary to track the changes of key parameters in real time and construct a time-varying parameter vector, including the axial stiffness of the screw, the radial stiffness of the spindle and the sealing loss coefficient of the vacuum chamber. The initial value of the axial stiffness of the screw is calculated based on the elastic modulus of the material, the radial stiffness of the spindle is determined by modal testing, and the sealing loss coefficient of the vacuum chamber is set according to the design indicators. It covers the key performance parameters of the equipment and realizes multi-dimensional status monitoring. Historical data will mask the current trend of change. The weight of old data needs to be reduced to quickly respond to parameter changes. A recursive update rule with a forgetting factor is set. Parameter mutations may indicate equipment failure. The root cause of the problem needs to be quickly located to reduce downtime. When the parameter mutation exceeds three times the standard deviation of the historical mean, the Bayesian probability model is triggered to perform root cause analysis of the fault and calculate the historical mean and standard deviation of the parameters. When The alarm is triggered when is the parameter estimation vector, is the historical average, is the standard deviation, define the fault hypothesis set ,in The screw is worn. Bearing damage, Calculate the posterior probability for seal aging. Calculating the posterior probability means calculating the probability of a hypothesis or event occurring given observed data or evidence. It is based on Bayes' theorem and is a method for updating the probability estimate of an event based on new information. In practical applications, the posterior probability is obtained by first determining the prior probability (that is, the initial estimate of the probability of an event when there is no additional information), and then calculating the likelihood (the probability of observing this data when the hypothesis is true) and the evidence factor (the probability associated with the observed data) based on the observed data. The normal fluctuation range of parameters varies under different working conditions, and fixed thresholds are prone to false alarms. According to the spindle speed, and feed speed The working condition intervals are divided into low-speed zone, medium-speed zone and high-speed zone. The mean and standard deviation are calculated independently for each working condition zone. The threshold coefficient is adjusted according to the working condition stability to maintain high detection sensitivity during high-speed processing.

[0058] The training method of the long short-term memory network model is:

[0059] The original sensor data is high in dimension and redundant, so a structured feature matrix needs to be constructed to improve the efficiency of model training. 50ms time series segments are intercepted from the continuous data stream with a step size of 5ms as the time axis. Each segment contains 10 time points to form a time axis. Temperature gradient, thermal expansion strain, motion acceleration and ambient humidity are extracted as feature dimensions respectively. A spatiotemporal feature matrix is ​​constructed based on the time axis and feature dimensions, while retaining the association between time series and multi-physical field features. Traditional long short-term memory neural networks pay insufficient attention to key features in long sequences, and the weights of important features need to be strengthened. The hidden state of the long short-term memory neural network is used as the query vector. The query vector is A vector representation for retrieving or matching relevant information. It can be understood as a digital encoding of the query content. By comparing with other vectors, the result that best meets the query requirements is found. The row vectors of the time series feature matrix are used as keys and values. The key is used to uniquely identify the identifier of the data, and the value is the data content corresponding to the key. It can be various types of data, such as numbers, strings, and lists. The multi-head attention mechanism is used to enhance the key feature extraction capability and define a composite loss function. A single loss function cannot balance multiple prediction targets and needs to comprehensively consider errors in different dimensions. The original model parameters are large in scale and need to be compressed to adapt to the resource limitations of embedded devices. Finally, the model quantization compression is implemented on the embedded tensor processing unit chip, including weight quantization and activation quantization. Among them, weight quantization is a technology that converts the weight values ​​in the neural network from high-precision values ​​to low-precision values. In deep learning, the weights of the neural network are usually represented by high-precision floating-point numbers, which takes up a lot of memory space and consumes more computing resources during calculation. Through weight quantization, these high-precision weight values ​​can be mapped to a limited set of discrete values, such as from 32-bit floating-point numbers to 8-bit integers or even lower-bit representations. The main purpose of this is to reduce the storage requirements of the model, speed up the reasoning speed of the model, and activate Activation quantization refers to the process of quantizing activation values ​​in deep learning models. Activation values ​​are the output results obtained by calculating neurons in each layer when the neural network processes input data. Activation quantization aims to map these continuous activation values ​​to a finite number of discrete values ​​to reduce the number of bits required for data representation. Based on weight amplitude pruning, connections less than the threshold are removed. The pruning ratio of each layer is 20% for the input layer, 30% for the hidden layer, and 10% for the output layer. The inference delay is controlled within the preset threshold. The deep learning inference optimizer is used for model optimization. The batch size is set to 16, the inference delay threshold is set to 10ms, and load balancing is achieved through dynamic batching.

[0060] The method for generating a process timing data set includes:

[0061] During the operation of the placement machine, motion parameters, process parameters, and environmental parameters are synchronously recorded. Motion parameters include X / Y / Z axis position, velocity, and acceleration; process parameters include nozzle vacuum, component quality, and placement pressure; and environmental parameters include temperature, humidity, and vibration spectrum. This eliminates time deviations between sensors and ensures strict alignment of multi-source data on the time axis, providing a reliable foundation for subsequent joint analysis.

[0062] The sampling frequencies of different sensors may be different. Directly using the original data will cause the time axis to be misaligned. It is necessary to use an algorithm to align the time series data of different frequencies into a unified time series, organize the motion parameters, process parameters, and environmental parameters into independent time series, construct a time series matrix, and align the multi-sensor timestamps of the original data based on the dynamic time warping algorithm. Taking the high-frequency motion parameter sequence as the benchmark, the low-frequency process parameter and environmental parameter sequences are aligned, and the dynamic time warping distance between different sequences is calculated. The optimal time warping path is found through dynamic planning, and the low-frequency data is interpolated onto the high-frequency time axis. It can effectively process sensor data with different sampling frequencies and achieve precise alignment while retaining data features. It can better reflect the real time series relationship of the data than the traditional interpolation method. The data samples of the placement machine under high-speed rapid acceleration and deceleration are relatively small. Direct use for model training will lead to overfitting. It is necessary to expand the small sample data through data enhancement technology to improve the generalization ability of the model. The adversarial generative network is used to expand the small sample working condition data. The Wasserstein distance is used to generate the adversarial network combined with the gradient penalty term architecture. The Wasserstein distance is an important concept in optimal transmission theory, which is used to measure the distance between two probability distributions. The generator input is a 100-dimensional random noise vector, and the multi-layer fully connected network outputs samples with the same dimension as the real data. The discriminator is a convolutional neural network, which is used to determine whether the input data is a real sample or a generated sample. Normal working conditions and known anomalies are used. A small number of samples of common working conditions, such as 500 high-speed rapid acceleration and deceleration samples, are used to train the model. The batch size is set to 64 and the training is performed for 10,000 rounds. After the generator learns the distribution characteristics of the real data, it generates virtual samples with high similarity to the real data, such as thermal expansion strain data simulated in a high humidity environment. Then, the t-distributed random neighbor embedding is used to reduce the dimension and visualize the clustering characteristics of the detection data. The aligned multi-source data are intercepted according to the time window to construct a spatiotemporal feature matrix. Each row corresponds to the 12-dimensional features of a time point. The t-distributed random neighbor embedding algorithm is used to reduce the 12-dimensional features to 2 dimensions. The perplexity is set to 30, the learning rate is 200, and the iteration is performed 1,000 times. The data points after dimensionality reduction are visualized on a two-dimensional plane, and different working conditions are marked by color. For example, under normal, high temperature, and rapid acceleration and deceleration conditions, a density-based spatial clustering application noise algorithm is used to automatically identify the clustering center of data points and detect abnormal working conditions. For example, data points under normal working conditions are clustered into one category, and abnormal data points under high temperature form independent clusters, which helps to quickly discover abnormal data points and their distribution patterns. The theoretical thermal expansion amount is calculated according to the material thermal expansion theory formula as a benchmark, and the relative deviation between the predicted compensation amount and the theoretical value is calculated. When the deviation exceeds three times the standard deviation, it is determined to be an abnormal working condition, and the data sample is automatically marked and added to the abnormal data set. After the abnormal data is marked, the system automatically includes it in the training set and triggers incremental training of the model, so that the model can learn the characteristics of abnormal working conditions and improve the prediction ability of similar abnormalities.

[0063] The dynamic adjustment strategy for thermal drift compensation is:

[0064] The thermal drift of the placement head changes dramatically during rapid acceleration and deceleration. It is necessary to prioritize the real-time performance of axial compensation to avoid a decrease in placement accuracy due to compensation lag. Define the axial compensation priority factor. , value range , the default value is 0.6, when it is detected that the placement head is in a state of rapid acceleration and deceleration, Increased to 0.9, ensuring that the control system prioritizes the allocation of computing resources to the axial compensation path, increasing the compensation output frequency from 200Hz to 500Hz, and collecting acceleration data in real time through the three-axis accelerometer built into the placement head to calculate the absolute value of acceleration. When the absolute value of acceleration is greater than 2 times the acceleration of gravity, it is determined to be a sudden acceleration and deceleration state, triggering the compensation frequency increase mechanism to avoid resource waste caused by global compensation. The compensation output module is designed to support 200Hz and 500Hz dual-frequency modes. It runs in 200Hz mode under normal working conditions. When it is detected In the case of rapid acceleration and deceleration, it switches to 500Hz mode within one sampling cycle (5ms) to output more intensive compensation instructions. The linear interpolation algorithm is used to smooth the compensation amount at the switching moment to avoid control signal oscillation caused by frequency mutation. When switching from 200Hz to 500Hz, the output value of the first three compensation cycles is the weighted average of the compensation amount under the new and old frequencies, which improves the dynamic response capability and makes a smooth transition to avoid control shock. A laser displacement sensor is installed at the end of the placement head to monitor the positioning error of the X-axis and Y-axis in real time, which is used as a Lissajous figure. The input signal is used to draw a trajectory curve. During normal compensation, the trajectory should approach the origin (error < 1 μm). If there is a residual error, the trajectory will take on a specific shape. The positioning accuracy after compensation is verified using the closed-loop Lissajous figure. The trajectory radius threshold of the Lissajous figure is set to 1.5 μm. When the trajectory radius exceeds the threshold for five consecutive cycles, it is determined that a significant residual error still exists after compensation, triggering the switch to reinforcement learning compensation mode. This avoids the limitations of relying on a single numerical judgment and improves the reliability of error detection. The state space is defined as the current thermal drift, acceleration, and compensation history. The action space is the adjustment step size of the compensation amount. The reward function is the negative inverse of the positioning error. When the Lissajous figure verification finds a residual error of 1.5 μm, the system switches to reinforcement learning mode within 10 ms. The intelligent agent iteratively optimizes the compensation strategy through real-time interaction with the placement head control system. After completing 100 compensation actions, the neural network parameters are updated using the experience replay mechanism to improve compensation accuracy and achieve autonomous optimization of the compensation strategy. The closed-loop execution unit 4 uses the thermal drift as the compensation amount and decomposes it into a displacement command and a radial torsion correction.

[0065] The present invention uses the temperature-deformation sensing unit 1 to collect the temperature and deformation data of multiple components of the placement head in real time. The thermal-mechanical coupling analysis unit 2 reconstructs the three-dimensional temperature field based on the finite element algorithm, dynamically calculates the thermal expansion distribution, and solves the model roughness problem of traditional single-point temperature measurement. The dynamic compensation control unit 3 predicts the thermal drift within the next 5ms through online parameter identification and long-short-term memory network model, and adaptively adjusts the compensation strategy in combination with the motion working condition. The closed-loop execution unit 4 decomposes the compensation amount into displacement and torsion correction instructions to achieve accurate offset of thermal deformation, build a full-process thermal compensation closed loop, improve the accuracy and stability of the placement head in complex thermal environments, and improve production efficiency.

[0066] Example 2

[0067] A second object of the present invention is to provide a method for implementing a high-precision placement head adaptive thermal compensation system for a placement machine, comprising the following steps:

[0068] S1. Use temperature sensors to monitor the temperature data of the Z-axis screw, rotating spindle and vacuum generator of the placement head in real time, and use deformation sensors to synchronously obtain the axial displacement deviation and radial offset;

[0069] S2, based on the thread pitch of the placement head, divides the adaptive wedge grid, inputs the temperature sensor data into the finite element heat conduction equation to iteratively solve the three-dimensional temperature gradient field, and dynamically corrects the thermal-mechanical coupling parameters of the grid nodes based on the material thermal expansion coefficient;

[0070] S3. Use the recursive least squares method with a forgetting factor to update the thermal-deformation transfer function coefficient matrix in real time, and generate a predicted value of the thermal drift within the next 5 ms by fusing the process timing data with the long short-term memory neural network model;

[0071] S4, decomposing the predicted thermal drift into a displacement command of the piezoelectric ceramic micro-motion platform and a radial torsion correction of the servo motor, and adaptively adjusting the compensation frequency based on the acceleration state;

[0072] S5. Verify the positioning accuracy after compensation through Lissajous graph nodes. If the residual error exceeds the threshold, switch to reinforcement learning compensation mode. When the sensor confidence decreases, enable the digital twin simulation module and federated learning to achieve system self-healing upgrade.

[0073] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are merely preferred examples of the present invention and are not intended to limit the present invention. Various changes and improvements may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and improvements fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A high-precision placement head adaptive thermal compensation system for a chip placement machine, characterized in that: The following units are included: The temperature-deformation sensing unit (1) includes a temperature sensor and a deformation sensor, which are used to obtain temperature sensing data, axial displacement deviation and radial offset of the placement head, the rotating spindle and the vacuum generator respectively; The thermal-mechanical coupling analysis unit (2) has a built-in temperature field reconstruction algorithm based on the finite element heat conduction equation. By integrating temperature sensor data, axial displacement deviation, radial offset and material thermal expansion coefficient database, it dynamically calculates the three-dimensional thermal expansion distribution of the Z-axis screw of the placement head, the rotating spindle and the vacuum generator; The dynamic compensation control unit (3) introduces an online recursive least squares parameter identifier to receive the three-dimensional thermal expansion distribution, and updates the heat-deformation transfer function coefficient matrix in real time in combination with the current motion acceleration and load pressure of the placement head. It then uses this matrix and the process timing data set as the input of the long-short-term memory network model to perform time series expansion on the three-dimensional thermal expansion distribution and output the thermal drift within the next 5ms. Methods for suppressing the coupling of thermal expansion and mechanical vibration: The frequency domain characteristics of vibration acceleration are introduced into the thermal-deformation transfer function. When a sudden change in the axial strain of the placement head lead screw is detected, accompanied by high-frequency vibration above 200 Hz, an inverse compensation waveform is generated to offset the thermal shock effect. The convergence of the coupled model is verified using the Lyapunov stability criterion. The closed-loop execution unit (4) uses the thermal drift as a compensation amount and decomposes it into a displacement instruction and a radial torsion correction amount.

2. The high-precision placement head adaptive thermal compensation system of a chip placement machine according to claim 1, characterized in that: The execution of the temperature field reconstruction algorithm includes the following steps: The adaptive wedge-shaped grid is divided based on the thread pitch of the Z-axis lead screw of the placement head. The grid density matches the layout position of the temperature sensor. The temperature sensor data is input into the finite element heat conduction equation as the boundary condition, and the temperature gradient field is solved by the Newton-Raphson iteration method. The thermal expansion coefficient of the adaptive wedge-shaped grid nodes is corrected in combination with the axial displacement deviation. When outputting the three-dimensional thermal expansion distribution, a dynamic compensation weight is applied to the axial strain of the Z-axis lead screw.

3. The high-precision placement head adaptive thermal compensation system of a chip placement machine according to claim 2, characterized in that: The updating method of the material thermal expansion coefficient database is: During the startup phase of the placement machine, the heating module is controlled to apply a calibration reward of step temperature rise to the Z-axis lead screw of the placement head. The actual deformation at each temperature rise step is measured by a laser displacement meter and compared with the theoretical deformation. When the error exceeds 5%, the dynamic correction value is calculated according to the set mathematical formula, and the dynamic correction value is associated with the current ambient humidity and load pressure to establish a four-dimensional lookup table.

4. The high-precision placement head adaptive thermal compensation system of a chip placement machine according to claim 1, characterized in that: The dynamic calculation of the three-dimensional thermal expansion distribution includes the abnormal data cleaning step: The temperature sensor data is filtered using a sliding window. The window length is synchronized with the frequency of the placement head movement. After the axial displacement deviation and the radial offset are time-aligned, the sensor failure data is eliminated through cross-correlation analysis. When the temperature difference of the piston cavity of the vacuum generator is detected to be greater than the preset temperature difference threshold, the infrared thermal imaging module is activated for secondary verification. The cleaned data stream is transmitted to the thermal-mechanical coupling analysis unit (2) through the message queue priority.

5. The high-precision placement head adaptive thermal compensation system of a chip placement machine according to claim 1, characterized in that: The operating method of the online recursive least squares parameter identifier includes: A time-varying parameter vector is constructed, including the axial stiffness of the screw, the radial stiffness of the spindle, and the sealing loss coefficient of the vacuum chamber. A recursive update rule with a forgetting factor is set. When the parameter mutation exceeds three times the standard deviation of the historical mean, the Bayesian probability model is triggered to perform root cause analysis of the fault.

6. The high-precision placement head adaptive thermal compensation system of a chip placement machine according to claim 1, characterized in that: The training method of the long short-term memory network model is: A 50ms time series segment is intercepted with a step size of 5ms as the time axis, and the temperature gradient, thermal expansion strain, motion acceleration, and ambient humidity are extracted as feature dimensions respectively. A spatiotemporal feature matrix is ​​constructed based on the time axis and feature dimensions. A multi-head attention mechanism is used to enhance the key feature extraction capability, and a composite loss function is defined. Finally, model quantization and compression are implemented on the embedded tensor processing unit chip, and the inference delay is controlled within the preset threshold.

7. The high-precision placement head adaptive thermal compensation system of a chip placement machine according to claim 1, characterized in that: The method for generating the process timing data set includes: During the operation of the placement machine, motion parameters, process parameters and environmental parameters are synchronously recorded. The motion parameters include X / Y / Z axis position, velocity and acceleration; the process parameters include nozzle vacuum, component quality and placement pressure; and the environmental parameters include temperature, humidity and vibration spectrum. The multi-sensor timestamps of the original data are aligned based on the dynamic time warping algorithm, and the adversarial generative network is used to expand the small sample operating condition data. Then, the dimensionality reduction visualization is used to detect the data clustering characteristics through t-distributed random neighborhood embedding. Finally, a data-physical joint verification mechanism is established. When the deviation between the predicted compensation amount and the theoretical value exceeds three times the standard deviation, the abnormal operating condition is automatically marked and active learning is initiated.

8. The high-precision placement head adaptive thermal compensation system of a chip placement machine according to claim 1, characterized in that: The dynamic adjustment strategy for the compensation of thermal drift is: Define the axial compensation priority factor. When it is detected that the placement head is in a rapid acceleration or deceleration state, increase the compensation output frequency from 200Hz to 500Hz. When the absolute value of the acceleration is greater than 2 times the acceleration of gravity, it is determined to be a rapid acceleration or deceleration state. The positioning accuracy after compensation is verified through the closed-loop node of the Lissajous figure. If the residual error is greater than 1.5μm, switch to the reinforcement learning compensation mode.

9. A method for implementing the high-precision placement head adaptive thermal compensation system for a chip placement machine according to any one of claims 1 to 8, characterized in that: The following steps are involved: S1. Use temperature sensors to monitor the temperature data of the Z-axis screw, rotating spindle and vacuum generator of the placement head in real time, and use deformation sensors to synchronously obtain the axial displacement deviation and radial offset; S2, based on the thread pitch of the placement head, divides the adaptive wedge grid, inputs the temperature sensor data into the finite element heat conduction equation to iteratively solve the three-dimensional temperature gradient field, and dynamically corrects the thermal-mechanical coupling parameters of the grid nodes based on the material thermal expansion coefficient; S3. Use the recursive least squares method with a forgetting factor to update the thermal-deformation transfer function coefficient matrix in real time, and generate a predicted value of the thermal drift within the next 5 ms by fusing the process timing data with the long short-term memory neural network model; S4, decomposing the predicted thermal drift into a displacement command of the piezoelectric ceramic micro-motion platform and a radial torsion correction of the servo motor, and adaptively adjusting the compensation frequency based on the acceleration state; S5. Verify the positioning accuracy after compensation through Lissajous graph nodes. If the residual error exceeds the threshold, switch to reinforcement learning compensation mode. When the sensor confidence decreases, enable the digital twin simulation module and federated learning to achieve system self-healing upgrade.

Images