Deposited object ellipticity control system and method based on a time transformer network

Abstract

The present application belongs to the field of parameter optimization control, and particularly relates to a sediment object ellipticity regulation system and method based on a time series transformer network, comprising: obtaining time series data of the surface temperature of a sediment object during a deposition process, real-time point position rotation speed of the sediment object at different angle positions in the circumferential direction, and deposition duration; processing the temperature time series data, real-time point position rotation speed, and deposition duration through a time series transformer network model, estimating the thickness range of the current formed body in the long axis direction and the short axis direction, and then determining the long axis size and the short axis size and calculating the ellipticity parameter; comparing the ellipticity parameter with a preset target ellipticity to generate a rotation speed dynamic adjustment strategy containing a rotation speed proportionality coefficient between different angle positions in the circumferential direction; adjusting the actual rotation speed of each angle position according to the proportionality coefficient to change the thickness of the sediment at the corresponding position; the present application realizes online prediction and closed-loop regulation of ellipticity.

Description

Technical Field

[0001] This invention belongs to the field of parameter optimization control, and particularly relates to a system and method for controlling the ellipticity of deposited objects based on a time-series transformer network. Background Technology

[0002] Fiber ellipticity control technology is widely used in the fabrication of polarization-maintaining elliptical core fibers. Traditional fabrication methods mainly include grinding and depressurization sintering. Grinding involves grinding the cladding of a circular preform into a symmetrical plane, then using high-temperature extrusion to deform the fiber core into an elliptical shape. However, this method involves multiple high-temperature-low-temperature cycles, resulting in poor consistency between longitudinal geometric birefringence and stress birefringence, making precise control of effective birefringence difficult. Depressurization involves reducing the pressure inside the tube during chemical vapor deposition to shrink the formed body into an ellipse; however, this method suffers from poor process stability and low ellipticity repeatability. All of these methods rely on post-deposition detection and adjustment, failing to predict and control the major and minor axis dimensions and ellipticity parameters of the formed body in real time during deposition. This leads to insufficient stability of birefringence parameters and poor longitudinal uniformity of the fiber. Current technologies lack an intelligent closed-loop control method based on time-series data, making it difficult to utilize multi-source time-series information such as the surface temperature of the deposited object, real-time point rotation speed, and deposition time to predict the ellipticity of the formed body online and simultaneously generate dynamic adjustment strategies. Therefore, how to predict the major and minor axis dimensions and ellipticity parameters of the formed body online based on the time-series data of the surface temperature of the deposited object, the real-time rotation speed and deposition time, and simultaneously generate a dynamic rotation speed adjustment strategy to achieve closed-loop control of ellipticity, so as to solve the problems of poor ellipticity controllability, insufficient stability of birefringence parameters and poor longitudinal uniformity in the existing elliptic core fiber fabrication process, has become a technical problem that urgently needs to be overcome in this field. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention proposes a deposition object ellipticity control system and method based on a time-series transformer network. The system includes: acquiring time-series data of the deposition object's surface temperature, real-time position rotation speed at different angular positions along the circumference, and deposition duration during the deposition process; processing the temperature time-series data, real-time position rotation speed, and deposition duration using a preset time-series transformer network model to estimate a first thickness range corresponding to the major axis and a second thickness range corresponding to the minor axis of the current formed body; determining the major axis dimension based on the first thickness range and the minor axis dimension based on the second thickness range, and calculating the ellipticity of the deposition object using the major and minor axis dimensions. The ellipticity parameter is calculated; the ellipticity parameter is compared with a preset target ellipticity, and a dynamic rotation speed adjustment strategy is generated based on the comparison result. The dynamic rotation speed adjustment strategy includes a rotation speed ratio coefficient between different angular positions of the deposition object in the circumferential direction; based on the rotation speed ratio coefficient, the actual rotation speed of the deposition object at different angular positions in the circumferential direction is adjusted to change the deposition thickness at the corresponding positions, so that the ellipticity of the shaped body after deposition meets the preset target ellipticity. This invention realizes online real-time prediction and closed-loop control of the ellipticity of the shaped body during rotational deposition, which significantly improves the dimensional accuracy and longitudinal uniformity of the shaped body.

[0004] To achieve the above objectives, the present invention provides the following technical solution: A sedimentary object ellipticity control system based on a time-series transformer network includes: The acquisition module is configured to: acquire time-series data of surface temperature of the deposited object during the deposition process, real-time position rotation speed of the deposited object at different angular positions in the circumferential direction, and deposition time; The prediction module is configured to: process the time-series data of the surface temperature of the deposited object, the real-time rotation speed of the point, and the deposition time through a preset time-series transformer network model, and predict the first thickness range corresponding to the long axis direction and the second thickness range corresponding to the short axis direction of the current shaped body; The major and minor axis determination module is configured to: determine the major axis dimension based on the first thickness range, determine the minor axis dimension based on the second thickness range, and calculate the ellipticity parameter from the major axis dimension and the minor axis dimension; The discrimination module is configured to: compare the ellipticity parameter with a preset target ellipticity, and generate a dynamic rotation speed adjustment strategy based on the comparison result. The dynamic rotation speed adjustment strategy includes a rotation speed ratio coefficient of the deposition object at different angular positions in the circumferential direction. The adjustment module is configured to: adjust the actual rotation speed of the deposition object at different angular positions in the circumferential direction according to the rotation speed ratio coefficient and the A3C algorithm, so as to change the thickness of the deposition at the corresponding position, so that the ellipticity of the deposited body meets the preset target ellipticity.

[0005] Specifically, the estimation of the first thickness range corresponding to the long axis direction and the second thickness range corresponding to the short axis direction of the current molded body includes: Obtain the predicted thickness value, prediction uncertainty, and prediction uncertainty in the major axis direction from the output layer of the time-series transformer network model, wherein the prediction uncertainty in the major axis direction represents the radius of the confidence interval of the predicted thickness value in the major axis direction, and the prediction uncertainty in the minor axis direction represents the radius of the confidence interval of the predicted thickness value in the minor axis direction. The lower limit of the thickness in the long axis direction is obtained by subtracting the prediction uncertainty in the long axis direction from the predicted thickness value in the long axis direction. The upper limit of the thickness in the long axis direction is obtained by adding the predicted thickness value in the long axis direction to the predicted uncertainty in the long axis direction.

[0006] Specifically, estimating the first thickness range corresponding to the long axis direction and the second thickness range corresponding to the short axis direction of the current molded body also includes: Using the lower limit value of the thickness in the long axis direction and the upper limit value of the thickness in the long axis direction as boundaries, a first thickness range corresponding to the long axis direction is constructed. The lower limit of the thickness in the minor axis direction is obtained by subtracting the minor axis direction prediction uncertainty from the minor axis direction thickness prediction value. Adding the predicted thickness value in the minor axis direction to the predicted uncertainty in the minor axis direction yields the upper limit value of the thickness in the minor axis direction. Using the lower limit of the thickness in the short axis direction and the upper limit of the thickness in the short axis direction as boundaries, a second thickness range corresponding to the short axis direction is constructed.

[0007] Specifically, determining the major axis dimension based on the first thickness range includes: Calculate the center value of the first thickness range; The weighting coefficient is dynamically set according to the prediction uncertainty of the major axis direction, wherein the weighting coefficient is inversely proportional to the prediction uncertainty of the major axis direction. The predicted thickness along the major axis is weighted and averaged with the center value, and the weight of the weighted average is the weight coefficient to obtain a first weighted result, and the first weighted result is used as the major axis dimension. When the asymmetry of the first thickness range after boundary trimming exceeds a preset asymmetry threshold, the median of the first thickness range is obtained, and the median is used as the major axis dimension. When the width of the first thickness range is less than a preset accuracy threshold, the predicted thickness value in the long axis direction is directly used as the long axis dimension.

[0008] Specifically, the ellipticity parameter is calculated from the major axis dimension and the minor axis dimension, including: Obtain the major axis dimension a minor axis dimensions b Predicted thickness along the major axis P a Predicted thickness along the minor axis P b Major axis direction prediction uncertainty and the prediction uncertainty of the minor axis direction ; The uncertainty is predicted based on the major axis direction. and the predicted uncertainty of the minor axis direction Calculate the correction factor c; According to the correction factor c Regarding the major axis dimension a Perform the first weighted compensation to obtain the compensated major axis component A; According to the correction factor c Regarding the minor axis dimension b Perform a second weighted compensation to obtain the compensated minor axis component B.

[0009] Specifically, the ellipticity parameter is calculated from the major axis dimension and the minor axis dimension, and further includes: Dividing the compensated major axis component A by the compensated minor axis component B yields the ellipticity parameter. or ; Determine the length of the major axis a Whether it originates from boundary clipping or a precision threshold judgment, if so, then the major axis dimension... a Apply a preset first penalty coefficient α The ellipticity parameter Updated to ; Determine the minor axis dimension b Whether it originates from boundary clipping or a precision threshold judgment, if so, then adjust the minor axis dimension. b Apply a preset second penalty coefficient β to update the ellipticity parameter η. .

[0010] Specifically, the ellipticity parameter is calculated from the major axis dimension and the minor axis dimension, and further includes: If the major axis dimension a and the minor axis dimension b If all values ​​originate from boundary clipping or precision threshold judgment, then the ellipticity parameter η is updated to... ; Calculate the combined uncertainty of the ellipticity parameter η. The synthesis uncertainty is: ;in, r The correlation coefficient is a preset value, and r It is a negative value; The combined uncertainty With the preset maximum permissible uncertainty U max If a comparison is made, >U max Then, the ellipticity parameter is output. At the same time, an unreliable flag is attached.

[0011] Specifically, a dynamic speed adjustment strategy is generated based on the comparison results, including: Obtain the ellipticity parameter η and the preset target ellipticity interval, wherein the target ellipticity interval has a target upper limit value η. max and target lower limit η min , where η max >η min >0; The ellipticity parameter η is respectively compared with the target lower limit value η min and the target upper limit value η max The comparison is as follows: If η < η min If the current ellipticity is too small, an adjustment command to increase the rotation speed ratio coefficient is generated; the rotation speed ratio coefficient K is defined as the rotation speed V of the deposition object at the angular position corresponding to the major axis of the circumferential direction. AC Rotational speed V at the angular position corresponding to the minor axis BD The ratio, i.e., K=V AC / V BD ; If η>η max If the current ellipticity is too large, an adjustment command to reduce the rotation speed ratio coefficient is generated. If η min ≤η≤η max If the current ellipticity is deemed acceptable, an adjustment command is generated to maintain the current rotation speed proportionality constant.

[0012] Specifically, the dynamic speed adjustment strategy generated based on the comparison results also includes: Calculate the deviation Δη between the ellipticity parameter η and the target ellipticity interval, wherein the deviation Δη is determined as follows: When η < η min When, Δη=η min -η; When η>η max At that time, Δη = η - η max ; When ηmin ≤η≤η max At that time, Δη=0; The correction value ΔK of the rotational speed proportional coefficient is calculated based on the deviation Δη, and the correction value ΔK satisfies ΔK=K p ×Δη, where K p The preset proportional gain, and K p It is a positive number; Obtain the rotational speed scaling factor K at the current moment. cur The rotational speed scaling factor K at the current moment cur Adding this to the correction value ΔK yields the updated rotational speed scaling factor K. new K new =K cur +ΔK; The updated rotational speed scaling factor K new This is output as the rotational speed proportional coefficient in the aforementioned dynamic speed adjustment strategy.

[0013] A method for controlling the ellipticity of deposited objects based on a time-series transformer network includes: Acquire time-series data of surface temperature of the deposited object during the deposition process, real-time position rotation speed of the deposited object at different angular positions in the circumferential direction, and deposition duration; The time-series data of the surface temperature of the deposited object, the real-time rotation speed of the point, and the deposition time are processed by a preset time-series transformer network model to estimate the first thickness range of the current shaped body in the long axis direction and the second thickness range in the short axis direction. The major axis dimension is determined based on the first thickness range, the minor axis dimension is determined based on the second thickness range, and the ellipticity parameter is calculated from the major axis dimension and the minor axis dimension. The ellipticity parameter is compared with a preset target ellipticity, and a dynamic rotation speed adjustment strategy is generated based on the comparison result. The dynamic rotation speed adjustment strategy includes a rotation speed ratio coefficient of the deposition object at different angular positions in the circumferential direction. Based on the rotation speed proportional coefficient, the actual rotation speed of the deposition object at different angular positions in the circumferential direction is adjusted to change the deposition thickness at the corresponding positions, so that the ellipticity of the deposited body meets the preset target ellipticity.

[0014] Compared with the prior art, the beneficial effects of the present invention are: This invention acquires time-series data of the surface temperature of the deposited object during the deposition process, real-time rotational speed at different angular positions along the circumference of the deposited object, and deposition duration. It then utilizes a time-series transformer network model to perform deep feature extraction and correlation analysis on these multi-source time-series data. This allows for online estimation of the thickness range of the shaped body along the major and minor axes, and real-time calculation of the current ellipticity parameters. The predicted ellipticity is compared with a preset target ellipticity to generate a dynamic adjustment strategy that includes a proportional coefficient of rotational speed between different angular positions along the circumference. Based on this proportional coefficient, the actual rotational speed at each angular position is adjusted, changing the deposition thickness at the corresponding position, ensuring that the ellipticity of the deposited body accurately meets the preset target. Therefore, this invention breaks away from the traditional passive mode of post-deposition detection and offline adjustment, achieving online real-time prediction and closed-loop active control of ellipticity, significantly improving the controllability and repeatability of ellipticity. Simultaneously, because the control process is based on real-time time-series data during deposition, it can effectively compensate for process fluctuations, improve the uniformity of the geometric dimensions along the longitudinal direction and the consistency of internal stress distribution, and reduce process difficulty. This makes it suitable for high-quality preparation of rotationally deposited bodies with strict ellipticity requirements. Attached Figure Description

[0015] Figure 1 This is a block diagram of the deposition object ellipticity control system based on a time-series converter network in Embodiment 1 of the present invention. Figure 2 This is a flowchart of the deposition object ellipticity control method based on a time-series converter network in Embodiment 2 of the present invention. Detailed Implementation

[0016] Example 1 Please see Figure 1 One embodiment of the present invention provides a deposition object ellipticity control system based on a time-series transformer network, comprising: The acquisition module is configured to acquire time-series data of the surface temperature of the deposited object during the deposition process, real-time positional rotation speed of the deposited object at different angular positions in the circumferential direction, and deposition duration. In a specific embodiment, the acquisition module acquires multi-source time-series data during the deposition process in the following manner: the time-series data of the surface temperature of the deposited object is obtained by continuously acquiring the radiation temperature of the surface of the deposited object using a pyrometer installed on the deposition equipment at a preset fixed sampling frequency. The measuring probe of the pyrometer is aligned with the deposition area on the surface of the deposited object, and outputs a real-time temperature analog signal, which is then converted from analog to digital to generate a temperature time-series sequence. In this invention, the preset fixed sampling frequency is determined by those skilled in the art based on the deposition process... The inherent dynamic characteristics of the surface temperature of the deposited object over time, along with the requirements of subsequent signal processing and model input, are comprehensively determined as follows: Specifically, the temperature time series data originates from the radiation temperature continuously collected by the pyrometer. Its rate of change is jointly affected by the rotation period of the deposited object (e.g., the alternating heating areas on the surface of the deposited object cause periodic temperature fluctuations per revolution) and the overall temperature rise trend caused by the accumulation of deposition time. The sampling frequency needs to be high enough to capture the key dynamic components in this temperature evolution without distortion. At the same time, it needs to be strictly aligned with the real-time position rotation speed (synchronously acquired by an incremental encoder) at different angular positions along the circumference of the deposited object and the deposition time variable using a unified time step index to form a multi-channel system. The input data is time-series data. Furthermore, the selection of this frequency must also consider the requirements of the subsequent sliding window filtering process for the effective sample size within the window. Specifically, the window width is set to a preset odd number of sampling points to ensure symmetry on both sides of the window's center point. Within the window, the arithmetic mean and standard deviation of the temperature are calculated to remove outliers deviating from the mean by more than three times the standard deviation, leaving sufficient data for linear interpolation. Therefore, the sampling frequency must ensure that a sufficient number of sampling points are accommodated within a window to meet the statistical removal sample size requirements and interpolation accuracy. Simultaneously, the time-series transformer network model requires the input sequence to cover a preset time window with regularly spaced time steps. The sampling frequency directly determines the number of sampling points included within this time window. The number of time steps affects the network's ability to extract features and predict the timeliness of cross-modal nonlinear relationships between the surface temperature, rotation speed, and duration of the deposited object (such as changes in deposition thickness due to temperature increases accompanied by slower rotation speed). Ultimately, those skilled in the art will also combine the response time of the pyrometer itself, the upper limit of the sampling accuracy of the analog-to-digital converter, and the Nyquist sampling criterion for temperature changes in the deposition process, and comprehensively weigh signal fidelity, data redundancy, and real-time computing load to pre-set this fixed sampling frequency. This provides a regular, synchronous, and information-rich time-series data foundation for subsequent sliding window filtering, embedded layer linear transformation, multi-head self-attention cross-modal processing, and online real-time prediction.

[0017] Real-time rotational speed data of the deposition object at different angular positions along the circumference are acquired by an incremental encoder mounted on the shaft end of the rotating chuck. The encoder rotates synchronously with the deposition object, outputting the actual rotational speed value corresponding to the current angular position at each sampling moment. Based on the preset zero-point position A and the circumferential angular distribution, the rotational speed values ​​at the zero-point position, 90-degree rotation position, 180-degree rotation position, and 270-degree rotation position are recorded respectively. Simultaneously, the rotational speed at intermediate interpolation angular positions is estimated using the angular distance weighted interpolation method, forming a rotational speed time series covering the entire angle. The deposition duration data is accumulated by the system timer from the start of the deposition process, recording the time offset of the current moment relative to the start moment in seconds, and using it as an independent time step index variable. The above three data sources are synchronized and aligned according to a unified time step index, forming multi-channel timing input data, providing raw process parameters for the subsequent prediction module.

[0018] The prediction module is configured to: process the time-series data of the surface temperature of the deposited object, the real-time rotation speed of the point, and the deposition time through a preset time-series transformer network model, and estimate the first thickness range corresponding to the long axis direction and the second thickness range corresponding to the short axis direction of the current shaped body; further, in this embodiment, the deposited object is a porous deposit formed by chemical vapor deposition, the deposited object is a cylindrical rotating substrate for supporting the porous deposit, and the shaped body is a deposit deposited on the rotating substrate and constituting the core functional layer of the final product.

[0019] In a specific embodiment, addressing the technical problem of how to construct a converter network suitable for time-series data of the optical fiber deposition process, enabling it to accurately extract the deep nonlinear mapping relationship between the surface temperature of the deposited object and the rotation speed and duration, this embodiment provides a method for constructing and training a time-series converter network model, specifically as follows: The surface temperature values ​​of the deposited object measured by a pyrometer during the deposition process are continuously acquired at a preset fixed sampling frequency to form an original temperature time series. Simultaneously, the real-time position rotation speed of the deposited object at four preset characteristic angle positions in the circumferential direction are recorded. The four characteristic angle positions are the 0 point position, the 90-degree rotation position, the 180-degree rotation position, and the 270-degree rotation position. Based on the rotation speed values ​​of adjacent characteristic angle positions, the estimated rotation speed values ​​of each intermediate interpolation angle position are calculated using the angle distance weighted interpolation method to form a real-time position rotation speed time series covering the entire angle. At the same time, the cumulative duration of the deposition process is used as an independent time step index variable, and after normalization processing, a normalized duration series or a remaining time ratio series is obtained. The temperature time series, rotation speed time series, and duration series of this embodiment are spliced ​​together according to the same time step index to form multi-channel time series input data.

[0020] In this embodiment, to avoid high-frequency noise and abnormal spikes in the original temperature data, a sliding window filtering process is performed on the original temperature time series. Specifically, the width of the sliding window is set to a preset odd number of sampling points. The arithmetic mean and standard deviation of all temperature values ​​within the window are calculated. Temperature values ​​in the window whose absolute deviation from the mean exceeds three times the standard deviation are identified as outliers and removed. The removed positions are filled by linear interpolation of the effective temperature values ​​within the window to obtain the denoised temperature time series. The rotation speed series and the time series do not require filtering.

[0021] The temperature value at each time step of the denoised temperature time series is input into a first linear embedding layer. This first linear embedding layer has a trainable first weight matrix and a first bias vector. Through linear transformation, the single-valued temperature is mapped to a temperature feature vector of a preset fixed dimension. Similarly, the rotational speed value at each time step of the rotational speed sequence is input into a second linear embedding layer, mapped to a rotational speed feature vector of the same fixed dimension. The duration value at each time step of the duration sequence is input into a third linear embedding layer, mapped to a duration feature vector of the same fixed dimension. In this embodiment, the preset fixed dimension is determined by those skilled in the art based on the structural constraints of the multi-head self-attention mechanism, the representation requirements of cross-modal feature fusion, and the trade-off between the number of model parameters and generalization performance. Specifically, the three individual values ​​of temperature, rotation speed, and duration must fall into the same vector space after being mapped by their respective linear embedding layers. This is because the temperature enhancement vector, rotation speed enhancement vector, and duration enhancement vector at the same time step need to be added or concatenated and fused. This requires that the dimensions of the three feature vectors be completely consistent. At the same time, this dimension must also satisfy the divisibility constraint of the number of attention heads in the multi-head self-attention layer. For example, if the number of attention heads is set to 8, this dimension can be selected as a value that is divisible by 8, such as 128 or 256, to ensure that each attention head is assigned to a subspace of equal length for independent calculation. Furthermore, the value of this dimension must also take into account both feature representation capability and computational load: if the dimension is too low, for example less than 32, it will be difficult to fully characterize the complex nonlinear relationships implied by temperature rise and fall, rotation speed and their evolution over time during the deposition process, thereby affecting the network's ability to capture cross-modal correlations such as "the increase in surface temperature of the deposited object accompanied by the decrease in rotation speed of the deposited object leading to an increase in the deposition thickness of the long axis of the shaped body"; if the dimension is too high, for example more than 512, it will significantly increase the number of trainable parameters of the first linear embedding layer, the second linear embedding layer and the third linear embedding layer, which can easily lead to overfitting under the condition that there are usually only a few hundred sets of training samples, and at the same time reduce the real-time performance of online inference. Therefore, those skilled in the art will initially select values ​​within typical ranges such as 64, 128, or 256, and use a validation set to perform grid search or Bayesian optimization fine-tuning with the goal of minimizing the prediction errors of the major and minor axis thicknesses. Finally, a preset fixed dimension is determined that can ensure the structural integrity of multi-head parallel computing, provide sufficient feature representation capacity, and take into account training efficiency and generalization performance. For example, 128 dimensions are selected as this fixed value. At this time, the three augmentation vectors at each time step are all 128 dimensions, and the position encoding is also of the same dimension, thus providing a regular, unified, and informative vector input basis for the feature extraction of the subsequent encoder.

[0022] Learnable position codes are superimposed on the temperature feature vector, rotational speed feature vector, and duration feature vector at each time step. The dimension of the position code is the same as that of the feature vector, and its value is iteratively updated during the training process of the time-series transformer network. It is used to represent the sequential position of each time step in the entire time sequence. The superposition method is to add the position code to the feature vector element by element to obtain the temperature enhancement vector, rotational speed enhancement vector, and duration enhancement vector carrying the time sequence information, respectively.

[0023] The temperature enhancement vector, rotation speed enhancement vector, and duration enhancement vector at the same time step are fused to obtain the embedded feature vector for that time step. In this embodiment, the fusion process is performed by vector addition, which adds the corresponding dimensional elements of the three enhancement vectors, or by vector concatenation, which connects the three enhancement vectors end to end according to their dimensional order. The embedded feature vectors of all time steps are arranged in ascending order of time step index to form an embedded feature sequence.

[0024] The main body of the timing converter network consists of encoders stacked with a preset number of layers. In this embodiment, the encoders are stacked with 6 layers, where each encoder layer includes a multi-head self-attention layer and a feedforward neural network layer.

[0025] The multi-head self-attention layer employs a preset number of attention heads; in this embodiment, there are eight attention heads. Each attention head independently calculates the attention weights between different time steps in the input sequence and specifically calculates the cross-modal attention weights between three modal features: temperature, rotation speed, and duration. This captures complex nonlinear relationships across variables, such as the increase in deposition thickness along the long axis of the formed body due to a decrease in the rotation speed of the deposited object as the surface temperature of the object increases. The output of the multi-head self-attention layer is fed into the feedforward neural network layer after passing through a first residual connection and a first normalization layer.

[0026] The feedforward neural network layer employs a two-layer fully connected structure, where the dimension of the hidden layer is a preset value; in this embodiment, the hidden layer dimension is 2048, and the activation function is a Gaussian error linear unit. The output of the feedforward neural network layer, after being processed by a second residual connection and a second normalization layer, serves as the output feature of the current encoder layer. Each encoder layer processes these features sequentially, with the final encoder layer outputting the final encoded features.

[0027] The encoded features output from the last encoder layer are input to the output layer, which is a linear transform layer. After linear transformation, it outputs three values: the predicted thickness value in the major axis direction, the predicted thickness value in the minor axis direction, and the prediction uncertainty in both the major and minor axis directions. The prediction uncertainty characterizes the confidence interval of the current predicted value. The first thickness range in the major axis direction is determined based on the predicted thickness value and its prediction uncertainty; that is, the range from the predicted thickness value in the major axis direction minus the prediction uncertainty to the predicted thickness value in the major axis direction plus the prediction uncertainty. Similarly, the second thickness range in the minor axis direction is determined based on the predicted thickness value and its prediction uncertainty.

[0028] In the training phase of this embodiment, multiple sets of historical deposition process data are collected as training sample sets; in this embodiment, more than 500 sets are collected. Each training sample set includes a complete temperature time series, site rotation speed time series, deposition duration series, and the true major axis and true minor axis dimensions obtained through precise measurement after deposition. The true major axis and true minor axis dimensions serve as supervision labels.

[0029] A joint loss function is constructed, comprising a first loss term, a second loss term, and a third loss term. The first loss term is the mean square error between the predicted thickness value and the actual major axis dimension in the major axis direction; the second loss term is the mean square error between the predicted thickness value and the actual minor axis dimension in the minor axis direction; and the third loss term is the log-likelihood loss of the prediction uncertainty, used to constrain the reasonableness of the prediction uncertainty. The first, second, and third loss terms are weighted and summed to obtain the value of the joint loss function. In a specific embodiment, the weight coefficients of the first, second, and third loss terms in the joint loss function are obtained as follows: Based on the validation set, a grid search or Bayesian optimization is performed to minimize the weighted composite index of the major axis thickness prediction error and the minor axis thickness prediction error on the validation set, thereby determining the optimal weight allocation for each loss term. In this embodiment, considering that the major and minor axis dimensions have the same weight on the ellipticity parameter, and that the log-likelihood loss of prediction uncertainty is mainly used to constrain the reasonableness of prediction confidence and avoid excessively large or small uncertainties, the weights of the first and second loss terms are both set to 1, and the weight of the third loss term is set to 0.1. The weight coefficients of the loss terms are preset before training begins and remain fixed throughout the training process, not changing with network parameter updates. Those skilled in the art can adjust the above-mentioned weight coefficients of the loss terms through a small number of experiments according to the actual data distribution and prediction accuracy requirements; such adjustment methods are conventional techniques.

[0030] An adaptive moment estimation optimizer is used as the parameter update algorithm, with an initial learning rate preset to 10. -4The batch size is preset to 16. The gradient of the joint loss function with respect to each trainable parameter is calculated using the backpropagation algorithm, and the parameters of the first linear embedding layer, the second linear embedding layer, the third linear embedding layer, the learnable positional encoder, the encoders of each layer of the temporal transformer network, and the parameters of the output layer are updated. The above process is repeated until the joint loss function converges. In this embodiment, the training rounds are 200, resulting in a trained temporal transformer network.

[0031] After training, the trained time-series transformer network is deployed in the online prediction module. During online operation, the time-series data of the surface temperature of the deposited object, the real-time point rotation speed data of each angle position in the circumferential direction of the deposited object, and the current deposition time are collected within a preset time window before the current deposition time. After the same preprocessing, embedding, and position encoding, the data are input into the time-series transformer network. The network outputs the predicted thickness and prediction uncertainty of the current formed body in the major axis direction and the predicted thickness and prediction uncertainty in the minor axis direction in real time, thereby obtaining the thickness range estimates of the major axis and minor axis respectively. The estimation results of this embodiment serve as the quantitative basis for subsequent ellipticity discrimination and dynamic rotation speed adjustment, realizing online real-time prediction of ellipticity during the deposition process.

[0032] It should be further explained that the step of estimating the first thickness range corresponding to the long axis direction and the second thickness range corresponding to the short axis direction of the current molded body in this embodiment specifically includes: A101. Obtain the predicted thickness value, predicted uncertainty, and predicted thickness value in the short axis direction from the output layer of the time-series converter network model; wherein, the predicted uncertainty in the long axis direction is used to characterize the confidence interval radius of the predicted thickness value in the long axis direction, and the predicted uncertainty in the short axis direction is used to characterize the confidence interval radius of the predicted thickness value in the short axis direction; further, this embodiment obtains the original form of the predicted uncertainty in the long axis direction output by the output layer of the time-series converter network, determines whether the original form is a logarithmic variance, if so, performs an exponential transformation on the predicted uncertainty in the long axis direction to obtain a variance value, and then performs a square root operation on the variance value to obtain the predicted uncertainty in the standard deviation form in the long axis direction; if the original form is a variance, directly performs a square root operation on the variance value to obtain the predicted uncertainty in the standard deviation form in the long axis direction; if the original form is a standard deviation, the original value remains unchanged; the predicted uncertainty in the short axis direction is processed similarly. A102. Perform a first arithmetic operation on the predicted value of the thickness in the long axis direction and the prediction uncertainty in the long axis direction to obtain a lower limit value of the thickness in the long axis direction. The first arithmetic operation is to subtract the prediction uncertainty in the long axis direction from the predicted value of the thickness in the long axis direction. A103. Perform a second arithmetic operation on the predicted value of the thickness in the long axis direction and the prediction uncertainty in the long axis direction to obtain the upper limit value of the thickness in the long axis direction. The second arithmetic operation is to add the predicted value of the thickness in the long axis direction to the prediction uncertainty in the long axis direction. A104. Using the lower limit of the thickness in the long axis direction as the left endpoint of the interval and the upper limit of the thickness in the long axis direction as the right endpoint of the interval, construct the first thickness range corresponding to the long axis direction. A105. Perform a third arithmetic operation on the predicted value of the thickness in the minor axis direction and the prediction uncertainty in the minor axis direction to obtain a lower limit value of the thickness in the minor axis direction. The third arithmetic operation is to subtract the prediction uncertainty in the minor axis direction from the predicted value of the thickness in the minor axis direction. A106. Perform a fourth arithmetic operation on the predicted value of the thickness in the minor axis direction and the prediction uncertainty in the minor axis direction to obtain the upper limit value of the thickness in the minor axis direction. The fourth arithmetic operation is to add the predicted value of the thickness in the minor axis direction to the prediction uncertainty in the minor axis direction. Furthermore, in this embodiment, the lower limit of the thickness in the major axis direction is compared with a preset minimum physical thickness threshold. If the lower limit of the thickness in the major axis direction is less than the minimum physical thickness threshold, then the lower limit of the thickness in the major axis direction is assigned the minimum physical thickness threshold. The upper limit of the thickness in the major axis direction is compared with a preset maximum physical thickness threshold. If the upper limit of the thickness in the major axis direction is greater than the maximum physical thickness threshold, then the upper limit of the thickness in the major axis direction is assigned the maximum physical thickness threshold. Similarly, the same boundary trimming process is performed on the lower limit of the thickness in the minor axis direction and the upper limit of the thickness in the minor axis direction.

[0033] In one specific implementation, the minimum physical thickness threshold and the maximum physical thickness threshold are two pre-set process constraint parameters for physically pruning the thickness prediction range output by the time-series transformer network model. The minimum physical thickness threshold is set based on the minimum thickness required for stable deposition on the surface of the deposition object during the deposition process. This value is determined by the minimum thickness required for the deposition object to form a continuous and stable deposited layer. Below this thickness, the deposition cannot adhere uniformly or is prone to detachment. In this implementation, it is exemplarily set to 0.1 mm. The maximum physical thickness threshold is jointly set by the maximum single-layer deposition thickness allowed by the deposition equipment and the shrinkage rate of the subsequent densification treatment. The maximum single-layer deposition thickness allowed by the equipment is limited by the mechanical clearance distance between the torch and the surface of the deposition object, the deposition efficiency of the torch power, and the upper limit of process stability to prevent collapse or powder shedding during single-layer deposition. Furthermore, the constraint that the thickness of the formed body after densification treatment does not exceed the product design specifications must be inversely calculated to the deposition state thickness using the densification shrinkage rate. Finally, the smaller value between the equipment limitation and the process calculation result is taken as the upper limit. In this implementation, it is exemplarily set to 5 mm.

[0034] When performing boundary clipping, if the calculated lower limit of the major axis thickness output by the sequential transformer network model is 0.05 mm, the system assigns it a value of 0.1 mm because this value is less than the minimum physical thickness threshold of 0.1 mm; if the calculated upper limit of the major axis thickness is 5.2 mm, the system assigns it a value of 5 mm because this value is greater than the maximum physical thickness threshold of 5 mm. Through the above boundary trimming process, it is ensured that the values ​​of the first and second thickness ranges are always within the process-feasible physical range defined by the minimum physical adhesion thickness to the maximum permissible deposition thickness. This avoids abnormalities in subsequent ellipticity parameter calculations and rotation speed control commands due to model predictions exceeding physical boundaries. For the short axis direction, its lower and upper limits can be trimmed using the same physical thresholds as those for the long axis direction. Alternatively, those skilled in the art can set separate minimum and maximum physical thickness thresholds, different from those for the long axis direction, based on the specific process characteristics of the short axis direction, such as slower rotation speed, larger deposition thickness, and significant temperature gradient differences during deposition. This boundary trimming step is executed in real time after each prediction to ensure that the values ​​of the first and second thickness ranges output to the long and short axis determination module are both within a reasonable physical range.

[0035] Furthermore, this embodiment also calculates the ratio of the predicted uncertainty in the major axis direction to the predicted thickness value in the major axis direction to obtain the relative uncertainty in the major axis direction. The relative uncertainty in the major axis direction is compared with a preset reliability threshold. If the relative uncertainty in the major axis direction is greater than the reliability threshold, a flag indicating unreliable prediction in the major axis direction is generated. Similarly, the ratio of the predicted uncertainty in the minor axis direction to the predicted thickness value in the minor axis direction is calculated to obtain the relative uncertainty in the minor axis direction. If the relative uncertainty in the minor axis direction is greater than the reliability threshold, a flag indicating unreliable prediction in the minor axis direction is generated. The unreliable flag is used by the discrimination module to determine whether to suspend speed adjustment. In one specific embodiment, the reliability threshold is set to 0.15 based on experimental experience, meaning that when the relative uncertainty exceeds 15%, the prediction result is considered unreliable. For example, if the predicted thickness in the major axis direction is 2.0 mm and the prediction uncertainty in the major axis direction is 0.4 mm, then the relative uncertainty is 0.2, which is greater than 0.15, generating a flag indicating unreliable prediction in the major axis direction. If the predicted thickness in the minor axis direction is 1.2 mm and the prediction uncertainty in the minor axis direction is 0.1 mm, then the relative uncertainty is approximately 0.083, which is less than 0.15, and no unreliable flag is generated. Upon receiving the flag indicating unreliable prediction in the major axis direction, the discrimination module pauses the current speed adjustment and issues a warning, waiting for the prediction result to become reliable before resuming adjustment. This reliability assessment avoids executing erroneous speed adjustment commands when the prediction confidence is low, improving the system's robustness and safety.

[0036] A107. Using the lower limit of the thickness in the short axis direction as the left endpoint of the interval and the upper limit of the thickness in the short axis direction as the right endpoint of the interval, construct a second thickness range corresponding to the short axis direction.

[0037] The major and minor axis determination module is configured to: determine the major axis dimension based on the first thickness range, determine the minor axis dimension based on the second thickness range, and calculate the ellipticity parameter from the major axis dimension and the minor axis dimension; In a specific embodiment, addressing the technical problem of determining the major axis dimension from the first thickness range—namely, the first thickness range being an interval estimate containing a lower and upper limit value, whose span may exhibit a wide distribution due to large prediction uncertainty, and which may result in an asymmetrical shape after physical boundary trimming—directly selecting the midpoint or endpoint of the interval as the major axis dimension would introduce point estimation bias and sever the physical correlation with the minor axis dimension. This embodiment provides the following specific implementation method for determining the major axis dimension based on the first thickness range: B101. Obtain the lower limit and upper limit of the first thickness range, and obtain the predicted thickness value in the long axis direction output by the timing converter network. B102. Calculate the center value of the first thickness range, where the center value is half the sum of the lower limit value and the upper limit value; B103. Dynamically set a weighting coefficient based on the prediction uncertainty of the major axis direction, wherein the weighting coefficient is inversely proportional to the prediction uncertainty of the major axis direction, and the larger the prediction uncertainty, the smaller the weighting coefficient. B104. The predicted thickness value along the major axis is weighted and averaged with the center value, where the weight of the weighted average is the weight coefficient, to obtain a first weighted result, and the first weighted result is used as the major axis dimension. In this embodiment, the weight coefficient is set for weighting and averaging the predicted thickness value along the major axis output by the time-series converter network model with the center value of the first thickness range. Specifically, the weight coefficient is the weight of the predicted thickness value along the major axis in the weighted average, while the weight of the center value of the first thickness range in the weighted average is 1 minus the weight coefficient, and the sum of the two is one. The specific relationship between the weighting coefficient and the prediction uncertainty in the major axis direction is as follows: those skilled in the art pre-set a basic weight value and a reference uncertainty value. The weighting coefficient is equal to the basic weight value multiplied by the ratio of the reference uncertainty to the current prediction uncertainty in the major axis direction, and the calculation result is limited to between 0 and 1. That is, when the calculated ratio is greater than 1, it is taken as 1, and when it is less than 0, it is taken as 0. Alternatively, the value range of the weighting coefficient can be directly set to between 0.3 and 0.9. When the prediction uncertainty is equal to the reference uncertainty, the weighting coefficient takes the basic weight value. When the prediction uncertainty increases, the weighting coefficient decreases accordingly, and when the prediction uncertainty decreases, the weighting coefficient increases accordingly. Taking specific values ​​as an example, if the basic weight value is set to 0.7 and the reference uncertainty is set to 0.2 mm, when the current prediction uncertainty in the major axis direction is 0.2 mm, the weight coefficient is 0.7. At this time, the weight of the predicted thickness value in the major axis direction is 0.7, and the weight of the center value of the first thickness range is 0.3. If the current prediction uncertainty increases to 0.4 mm, the weight coefficient is adjusted according to the inverse relationship to 0.7 × 0.2 ÷ 0.4 equals 0.35, that is, the weight of the predicted thickness value in the major axis direction is reduced to 0.35, while the weight of the center value of the first thickness range is correspondingly increased to 0.65. If the current prediction uncertainty decreases to 0.1 mm, the weight coefficient is adjusted to 0.7 × 0.2 ÷ 0.1 equals 1.4. Since it exceeds the upper limit of the value range of 1, it is taken as 1. At this time, the predicted thickness value in the major axis direction is used as the major axis dimension. By using the above dynamic setting method, when the prediction uncertainty is large, the reliability of the point prediction value output by the model is low. Therefore, its weight in the final major axis dimension is reduced, and the weight of the center value derived from the thickness range boundary is increased accordingly, thereby suppressing the interference of prediction noise on the dimension determination. When the prediction uncertainty is small, the reliability of the point prediction value output by the model is high. Therefore, its weight is increased, so that the major axis dimension can more sensitively reflect the real changes in the current deposition state.

[0038] B105. When the asymmetry of the first thickness range after boundary trimming exceeds a preset asymmetry threshold, the median of the first thickness range is obtained, and the median is used as the major axis dimension. In this embodiment, the asymmetry is calculated by the upper and lower limits of the first thickness range after boundary trimming. Specifically, it is defined as follows: first, the center value of the first thickness range is calculated, which is half of the sum of the upper and lower limits. Then, the difference between the upper limit and the center value is calculated, and the difference between the center value and the lower limit is calculated. The former is divided by the latter to obtain a ratio, which reflects the degree of deviation of the upper and lower limits of the range from the center value. Then, the absolute value of the difference between the ratio and the value 1 is used as the asymmetry. When the absolute value exceeds the preset asymmetry threshold, it is determined that the asymmetry exceeds the standard. Taking specific values ​​as an example, if the lower limit of the first thickness range after boundary trimming is 0.5 mm and the upper limit is 1.5 mm, then the center value is 1.0 mm. The difference between the upper limit and the center value is 0.5 mm, and the difference between the center value and the lower limit is also 0.5 mm. The ratio of the two is 1, and the absolute value deviation of the ratio is 0, indicating that the interval is completely symmetrical. If the lower limit is 0.6 mm, the upper limit is 1.8 mm, and the center value is 1.2 mm, the difference between the upper limit and the center value is 0.6 mm, and the difference between the center value and the lower limit is... If the lower limit is 0.6 mm, the ratio of the two is 1, the deviation is 0, and it remains symmetrical. If the lower limit is 0.5 mm, the upper limit is 2.1 mm, and the center value is 1.3 mm, the difference between the upper limit and the center value is 0.8 mm, and the difference between the center value and the lower limit is also 0.8 mm, with a ratio of 1, it remains symmetrical. If the lower limit is 0.2 mm, the upper limit is 1.8 mm, and the center value is 1.0 mm, the difference between the upper limit and the center value is also 0.8 mm, with a ratio of 1, it remains symmetrical. It should be noted that since the center value is defined as the arithmetic mean of the upper and lower limits, the difference between the upper limit and the center value must be equal to the difference between the center value and the lower limit, and the ratio between the two is always equal to 1, with an absolute deviation of 0 from 1. Therefore, the degree of asymmetry essentially reflects the asymmetry caused by cutting only one side during physical boundary trimming. For example, if the lower limit of 0.05 mm is trimmed to 0.1 mm while the upper limit of 1.5 mm is not trimmed, the lower limit becomes 0.1 mm, the upper limit becomes 1.5 mm, the center value is 0.8 mm, the difference between the upper limit and the center value is 0.7 mm, the difference between the center value and the lower limit is 0.7 mm, the ratio of the two is 1, and the deviation is 0. The actual situation that produces asymmetry includes: when the center value is no longer the original midpoint after one side is trimmed, that is, when the difference between the upper limit and the center value is not equal to the difference between the center value and the lower limit, the ratio of the two deviates from 1, and the absolute value of the difference between this ratio and 1 is the degree of asymmetry.When the degree of asymmetry exceeds the preset asymmetry threshold, it indicates that the interval has seriously deviated from the symmetrical shape after boundary clipping. At this time, the weighted average method is no longer used, and the median within the first thickness range is instead obtained as the major axis dimension.

[0039] In this embodiment, the asymmetry threshold is preset by those skilled in the art based on the tolerance limit of the center value estimation deviation due to the degree of asymmetry of the interval after boundary clipping. Specifically, since boundary clipping only occurs when the upper limit exceeds the maximum physical thickness threshold or the lower limit is lower than the minimum physical thickness threshold, one side of the thickness range is cut off while the other side remains unchanged after clipping. This causes the deviations of the upper and lower limits relative to the center value to no longer be equal. At this time, using a weighted average method to determine the major axis dimension will introduce a systematic bias. Therefore, an asymmetry threshold needs to be set to determine whether it is necessary to switch to a median-based estimation method. The preset basis for this threshold is: those skilled in the art determine a maximum allowable interval deviation ratio based on the requirements of the deposition process for thickness estimation accuracy and the sensitivity of subsequent ellipticity calculation to major axis dimension errors. For example, setting this threshold to 0.2 or 0.3 means that when the ratio of the deviation of one side of the interval relative to the center value to the deviation of the other side exceeds 20% or 30%, the asymmetry is considered to have reached a level that affects the reliability of point estimation. Taking specific values ​​as an example, if the lower limit of the first thickness range after trimming is 0.4 mm and the upper limit is 1.6 mm, then the center value is 1.0 mm, the lower limit deviation is 0.6 mm, the upper limit deviation is 0.6 mm, the ratio of the two is 1, the deviation is 0, and it does not exceed the threshold. If the lower limit after trimming is 0.1 mm and the upper limit is 1.5 mm, the center value is 0.8 mm, the lower limit deviation is 0.7 mm, the upper limit deviation is 0.7 mm, the ratio of the two is 1, and the deviation is 0. In reality, situations exceeding the threshold may occur, such as when the lower limit after trimming is 0. The maximum value is 2.2 mm, the center value is 1.2 mm, the minimum deviation is 1.0 mm, the maximum deviation is 1.0 mm, and the ratio is 1. Due to the nature of the arithmetic mean, it should be noted that the asymmetric threshold is usually used to determine whether to switch when the extreme of the interval after clipping is biased to one side and the median deviates significantly from the center value. Those skilled in the art can combine the distribution statistics of clipping events in historical deposition data and the tolerance range of the process for thickness deviation to select a specific value between 0.15 and 0.35 as the preset asymmetric threshold.

[0040] B106. When the width of the first thickness range is less than a preset accuracy threshold, the predicted thickness value in the long axis direction is directly used as the long axis dimension. In this embodiment, the accuracy threshold is preset by those skilled in the art based on the upper limit of the estimated accuracy that the time-series transformer network model can achieve for thickness prediction and the actual minimum resolution of the deposition process for thickness control. Specifically, the width of the first thickness range is defined as the upper limit value minus the lower limit value. This width is essentially equal to twice the prediction uncertainty in the long axis direction. The smaller the value, the higher the confidence of the model in predicting the long axis thickness at that time step, and the closer the predicted value is to the true value. The accuracy threshold is set based on the following: those skilled in the art first statistically analyze the distribution of the prediction uncertainty in the long axis direction among all predicted samples in the validation set during the model training phase, and obtain its minimum value, lower quartile, and other statistical quantities to understand the prediction accuracy level that the model can achieve under optimal conditions; at the same time, combined with the noise level of the high-temperature meter measurement in the deposition process, the natural thickness fluctuation amplitude caused by the surface roughness of the deposited object, and the minimum thickness difference that the actual measurement method can distinguish, an empirical lower limit value is jointly determined. For example, if statistics show that the prediction uncertainty for the vast majority of samples is between 0.08 mm and 0.15 mm, while the deposition process itself has a resolution of approximately 0.05 mm for thickness measurement, then the accuracy threshold can be preset to 0.1 mm. This means that when the width of the first thickness range is less than 0.1 mm, the current prediction uncertainty is already below this threshold, and the predicted thickness value in the major axis direction output by the model has a sufficiently high confidence level. In this case, this predicted value can be directly used as the major axis dimension without needing to perform a weighted average with the center value of the first thickness range, thus avoiding unnecessary additional errors introduced by the weighted average due to an excessively narrow thickness range. This accuracy threshold matches the level of absolute uncertainty implied by the previously set reliability threshold (relative uncertainty exceeding 15% is considered unreliable), ensuring that accuracy judgment and reliability judgment remain consistent on the same metric.

[0041] In a specific embodiment, regarding the technical problem encountered in determining the minor axis dimension from the second thickness range, namely, the second thickness range typically has higher prediction uncertainty than the major axis direction due to the slower rotation speed, larger deposition thickness, and significant temperature gradient differences of the deposition object corresponding to the minor axis direction, resulting in a wider range and stronger asymmetry in the second thickness range; simultaneously, the absolute value of the minor axis dimension is smaller, leading to a larger relative error under the same absolute uncertainty, and since the ellipticity parameter is the ratio of the major axis to the minor axis, small deviations in the minor axis dimension will be amplified into significant deviations in ellipticity. If the same extraction method as for the major axis dimension is directly used, it will be difficult to guarantee the stability of the ellipticity calculation. To address the above problems, this embodiment uses the following method to determine the minor axis dimension: B107. Obtain the lower limit and upper limit of the second thickness range, and obtain the predicted thickness value in the short axis direction output by the timing converter network. B108. Calculate the center value of the second thickness range, where the center value is equal to the sum of the lower limit value and the upper limit value divided by two; B109. Obtain the pre-set baseline weighting coefficient. This baseline weighting coefficient is pre-set by those skilled in the art based on prior knowledge that the prediction uncertainty in the minor axis direction is generally higher than that in the major axis direction, and the amplification effect of minor axis dimension error on the ellipticity parameter. It is proportionally reduced from the basic weighting value used in the major axis dimension determination process. Specifically, the basic weighting value used in the major axis dimension determination process is, for example, set to 0.7. Those skilled in the art consider that the prediction uncertainty in the minor axis direction is typically 1.5 to 2 times that of the major axis direction due to the slower deposition rotation speed, larger deposition thickness, and significant temperature gradient differences. Simultaneously, the absolute value of the minor axis dimension is smaller, and the amplification effect on ellipticity after division is more significant under the same absolute error. Therefore, a lower weight needs to be assigned to the thickness prediction value in the minor axis direction to suppress the interference of unreliable predictions on dimension determination. Thus, the baseline weighting coefficient for the minor axis direction is set to... Positive numbers less than or equal to 0.5, such as 0.3 or 0.4, can also be adjusted based on the statistical results of the prediction error distribution in the short axis direction on the validation set. If the validation set statistics show that the root mean square error of the predicted value in the short axis direction is 1.5 times that in the long axis direction, the baseline weight coefficient can be obtained by dividing the basic weight value of the long axis (0.7) by 1.5 and rounding it down or taking an approximate value to 0.47. Then, it can be fine-tuned according to the tolerance of the process for the short axis error and finally determined to be 0.4. If the prediction error in the short axis direction is more significant, it can be further reduced to 0.25 or 0.3. By setting the baseline weight coefficient of the short axis to a value significantly lower than that in the long axis direction, the determination of the short axis dimension relies more on the center value derived from the thickness range boundary rather than the model prediction value at the initial level, thereby fundamentally suppressing the propagation of the higher uncertainty in the short axis direction to the ellipticity parameter. B110. Calculate the short axis weighting coefficient dynamically based on the short axis direction prediction uncertainty. The short axis weighting coefficient is equal to the minimum value of the minimum value obtained by multiplying the benchmark weighting coefficient by the ratio of the short axis direction prediction uncertainty to the reference uncertainty. The short axis weighting coefficient is inversely proportional to the short axis direction prediction uncertainty. B111. The predicted thickness value in the minor axis direction is weighted and summed with the center value of the second thickness range. The weight of the predicted thickness value in the minor axis direction in the weighted sum is the minor axis weight coefficient, and the weight of the center value of the second thickness range is one minus the minor axis weight coefficient. The result of the weighted sum is taken as the minor axis dimension. B112. When the asymmetry of the second thickness range after boundary trimming exceeds a preset minor axis asymmetry threshold, the median of the second thickness range is obtained, and the median is used as the minor axis dimension. The asymmetry is defined as the absolute deviation of the ratio of the difference between the upper limit and the center value to the difference between the center value and the lower limit value from 1. In this embodiment, the minor axis asymmetry threshold is preset by those skilled in the art during the determination of the minor axis dimension based on the same judgment logic as the major axis direction, combined with the uncertainty characteristics and error amplification sensitivity of the minor axis direction itself. Specifically, this threshold has the same mechanism of action as the asymmetry threshold used in the determination of the major axis dimension; both are used to determine whether the deviation of the upper and lower limits from the center value due to unilateral truncation after boundary trimming of the second thickness range is so severe that the weighted average method cannot be continued. If the threshold is exceeded, the median is used as the minor axis dimension. Since the prediction uncertainty in the minor axis direction is generally higher than that in the major axis direction, the second thickness range itself has a wider interval span. After a trimming event, the degree of interval asymmetry is often more severe. At the same time, the minor axis dimension error has a significant amplification effect on the ellipticity parameter. Therefore, those skilled in the art usually set the minor axis asymmetry threshold to a more stringent level than that in the major axis direction. For example, if the major axis asymmetry threshold is set to 0.25, the minor axis asymmetry threshold can be set to 0.15 or 0.2. This allows the minor axis dimension to switch to a median-based estimation method in advance when the degree of asymmetry is low, avoiding the introduction of interference from the already skewed center value when the interval offset is not yet severe. The specific value of this threshold can be fine-tuned based on the above initial value and combined with the statistical analysis of minor axis trimming events in the validation set to achieve a balance between interval asymmetry tolerance and dimension estimation stability. This ensures that the minor axis dimension can still obtain a robust estimation result by the median method after boundary trimming, thereby effectively suppressing the adverse effects of asymmetric intervals on the calculation of ellipticity parameters.

[0042] B113. When the width of the second thickness range is less than the preset short axis accuracy threshold, the predicted thickness value in the short axis direction is directly used as the short axis dimension, wherein the width is defined as the upper limit value minus the lower limit value.

[0043] A specific technical problem exists in calculating the ellipticity parameter from the major and minor axis dimensions: the major and minor axis dimensions are extracted from their respective thickness ranges, both carrying residual errors from prediction uncertainty and interval extraction strategies. Furthermore, the errors in the major and minor axis directions are not independently and identically distributed during deposition due to factors such as differences in deposition object rotation speed and uneven temperature distribution; they often exhibit a negative correlation (i.e., a larger major axis dimension results in a smaller minor axis dimension, and vice versa). When a direct ratio is used… or =a / b When calculating the ellipticity parameter, the aforementioned negative correlation error significantly amplifies the fluctuation of the ratio, resulting in a variance of the ellipticity parameter that is much larger than the simple sum of the variances of the major and minor axes. Furthermore, if there are truncation errors in the major or minor axis dimensions due to boundary clipping or threshold judgment, the direct ratio will also produce a system offset. The underlying technical problem is: how to design a method for calculating the ellipticity parameter that can compensate for the error correlation between the major and minor axis dimension point estimates, suppress the ratio amplification effect, and identify and eliminate unreliable ellipticity calculation results caused by single-axis anomalies while fusing the estimated values ​​of the major and minor axis dimensions. To this end, this embodiment provides an implementation method for calculating the ellipticity parameter from the major and minor axis dimensions, specifically: B114. Obtain the major axis dimension a minor axis dimensions b Predicted thickness along the major axis P a Predicted thickness along the minor axis P b Major axis direction prediction uncertainty and the prediction uncertainty of the minor axis direction ; B115. Predicting uncertainty based on the major axis direction. and the predicted uncertainty of the minor axis direction Calculate the correction factor c Among them, the correction factor c Specifically: ; B116, according to the correction factor c Regarding the major axis dimension a Performing the first weighted compensation yields the compensated major axis component A, which is specifically: ; B117. According to the correction factor c Regarding the minor axis dimension b A second weighted compensation is performed to obtain the compensated minor axis component B, which is specifically: ; B118. Divide the compensated major axis component A by the compensated minor axis component B to obtain the ellipticity parameter. or ; B119. Determine the length of the major axis. a Whether it originates from boundary clipping or a precision threshold judgment, if so, then the major axis dimension... a Apply a preset first penalty coefficient α The ellipticity parameter Updated to the formula: ; In this embodiment, when the major axis dimension a When the ellipticity is determined to originate from boundary clipping or precision threshold judgment, a preset first penalty coefficient α is further applied to the compensated major axis component based on the initial ellipticity obtained by bidirectional compensation of the major and minor axes using the correction factor γ. The final updated ellipticity parameter is equal to... a ×(1+γ)×(1+α) divided by b ×(1-γ); where, a ×(1+γ) is the compensated major axis component. b ×(1-γ) represents the compensated minor axis component, and (1+α) is the additional penalty term applied to the major axis component. This penalty term is greater than one and is used to compensate for the truncation deviation introduced by boundary clipping or accuracy threshold judgment in the major axis dimension. Taking a specific value as an example, if the major axis dimension... a 2.0 mm, short axis dimension b With a major axis length of 1.2 mm, a correction factor γ of 0.1, and a first penalty coefficient α of 0.05, the compensated major axis component is 2.0 × 1.1 = 2.2 mm, the compensated minor axis component is 1.2 × 0.9 = 1.08 mm, the initial ellipticity is 2.2 ÷ 1.08 ≈ 2.037, and the updated ellipticity is 2.037 × 1.05 ≈ 2.139. This result is the final ellipticity parameter output value after penalizing and correcting for the abnormal source of the major axis size. This correction mechanism effectively suppresses the influence of the major axis truncation deviation on the ellipticity calculation, ensuring the accuracy of the ellipticity parameter output when the major axis size estimation is unreliable.

[0044] In this embodiment, the first penalty coefficient α is determined by those skilled in the art for compensating for boundary clipping or accuracy thresholds regarding the major axis dimension. a The systemic bias introduced is replaced by a pre-set empirical correction value. Specifically, when the major axis dimension... a When the boundary is clipped, it indicates that the predicted thickness value or thickness range in the major axis direction output by the time-series transformer network has exceeded the upper or lower limit of the physically feasible interval. The size value determined after clipping is not the true predicted value of the model, but is limited by the truncation value of the physical boundary, which has a constant directional offset relative to the true value; when the major axis dimension aWhen the accuracy threshold is determined, it indicates that the prediction uncertainty is too small, causing the thickness range width to be smaller than the accuracy threshold. The system directly uses the predicted thickness value along the major axis as the major axis dimension. In this case, the dimension is essentially a model point estimate without incorporating interval center information, and in rare cases, it may deviate from the true value due to local model overfitting. To suppress the residual bias caused by the above truncation or direct use, those skilled in the art apply a penalty coefficient α to the major axis dimension in the ellipticity parameter calculation formula. The specific value of this coefficient is determined based on the average deviation direction and magnitude obtained statistically from the boundary clipping events on the validation set: if the average deviation of the major axis dimension corresponding to the clipping event in the validation set relative to the true value is a certain proportion, for example, an average deviation of 5%, then α is set to a positive number so that the major axis component is appropriately reduced after applying the penalty to offset the deviation. In one specific implementation, the value of α is a positive number between 0.01 and 0.1 mm. Those skilled in the art can combine the magnitude of the long axis thickness in the deposition process and the typical deviation statistics of the trimming event to preset α to a fixed constant value, such as 0.05 mm, to ensure that when the long axis size comes from boundary trimming or accuracy threshold judgment, the long axis size is appropriately corrected downward by the penalty coefficient, thereby reducing the ellipticity calculation deviation caused by truncation or direct use.

[0045] B120. Determine the size of the minor axis. b Whether it originates from boundary clipping or a precision threshold judgment, if so, then adjust the minor axis dimension. b Apply a preset second penalty coefficient β to the ellipticity parameter or Updated to the formula: ; In this embodiment, when the minor axis dimension b When the ellipticity is determined to originate from boundary clipping or precision threshold judgment, a preset second penalty coefficient β is further applied to the compensated minor axis component based on the initial ellipticity obtained by bidirectional compensation of the major and minor axes using the correction factor γ. The final updated ellipticity parameter is equal to the major axis dimension. a Multiply by (1+γ) then by (1+β), and divide by the minor axis dimension. b Multiply by (1-γ); where, a Multiply by (1+γ) to get the compensated major axis component. b Multiplying by (1-γ) yields the compensated minor axis component, and (1+β) is the additional penalty term applied to the minor axis component. This penalty term is greater than 1 and is used to compensate for the truncation deviation introduced by boundary clipping or accuracy threshold judgment in the minor axis dimension. For a specific numerical example, if the major axis dimension... a 2.0 mm, short axis dimension bWith a minor axis length of 1.2 mm, a correction factor γ of 0.1, and a second penalty coefficient β of 0.08, the compensated major axis component is 2.0 multiplied by 1.1, which equals 2.2 mm. The compensated minor axis component is 1.2 multiplied by 0.9, which equals 1.08 mm. The initial ellipticity is 2.2 divided by 1.08, which is approximately 2.037. The updated ellipticity is 2.037 multiplied by 1.08, which is approximately 2.200. This result is the final ellipticity parameter output value after penalizing and correcting for the abnormal minor axis size. This correction mechanism effectively suppresses the amplification effect of the minor axis truncation deviation on the ellipticity parameter after division, ensuring the accuracy of the ellipticity parameter output when the minor axis size estimation is unreliable.

[0046] In this embodiment, the second penalty coefficient β is determined by those skilled in the art for compensating for boundary clipping or accuracy thresholds regarding the minor axis dimension. b The pre-set empirical correction value for the introduced systematic deviation follows the same logic as the first penalty coefficient α, but the amplification effect of the minor axis dimension error on the ellipticity parameter needs to be considered additionally. Specifically, when the minor axis dimension... b When the value originates from boundary clipping or accuracy threshold judgment, it indicates that the dimension value is not the true predicted value of the model, but rather a limited estimate constrained by physical boundary truncation or direct use of point prediction values. This truncation or direct use results in a systematic offset relative to the true value. Furthermore, because the absolute value of the minor axis dimension is small, the minute absolute deviation of the minor axis dimension is significantly amplified in the division operation where the ellipticity parameter is the ratio of the major axis dimension to the minor axis dimension. Therefore, those skilled in the art, when setting β, will obtain the average deviation of the minor axis dimension relative to the true value under truncation or direct use conditions based on the statistical results of minor axis direction clipping events and accuracy threshold events in the validation set. This deviation amplitude is then multiplied by an amplification factor determined according to the magnitude of the minor axis, and preset as a positive fixed value as β, for example, β is set to 0.08 or 0.1. Generally, the value of β is greater than α to reflect the greater impact of the minor axis dimension deviation on the ellipticity parameter. By using this second penalty coefficient, when the minor axis dimension originates from boundary clipping or accuracy threshold judgment, a moderate correction can be applied to the minor axis dimension in the ellipticity parameter calculation, thereby suppressing the ellipticity calculation deviation caused by the truncation or direct use of the minor axis dimension, and ensuring the reliability of the output of the ellipticity parameter under uniaxial abnormal conditions.

[0047] In this embodiment, the determination of the second penalty coefficient β needs to consider the amplification effect of the minor axis dimension error on the ellipticity parameter. Those skilled in the art first obtain typical minor axis dimensions from all qualified batches or target designs in historical deposition data, for example, using a minor axis dimension design target value of 1.2 mm as the benchmark for the minor axis magnitude. Since the ellipticity parameter is the ratio of the major axis dimension to the minor axis dimension, the smaller the minor axis magnitude, the greater the relative ellipticity deviation caused by the same absolute deviation. When the minor axis magnitudes are 1.2 mm and 2.0 mm, the impact of the same deviation of 0.05 mm on ellipticity differs by approximately 1.67 times. Therefore, the amplification conversion factor is determined based on the reciprocal relationship of the minor axis magnitude, that is, the reciprocal of the typical value is taken as the conversion benchmark, for example, using 0.833, the reciprocal of 1.2 mm, as the amplification conversion factor. The coefficient is multiplied by the average absolute deviation of the minor axis dimension obtained from the validation set under the conditions of boundary clipping or accuracy threshold judgment. The resulting product is the correction amount that can equivalently reflect the actual impact of the deviation on the ellipticity. This correction amount is directly preset as a fixed value of the second penalty coefficient β. For example, if the average absolute deviation is 0.08 mm, then β is 0.08 multiplied by 0.833 and rounded to 0.07. This reasonably compensates for the amplification effect caused by the minor axis truncation deviation through ratio calculation, ensuring that the ellipticity parameter still has high accuracy when the minor axis estimation is unreliable.

[0048] B121, if the length axis dimension a and the minor axis dimension b All of these originate from boundary clipping or precision threshold judgment, so the ellipticity parameter is... or Updated to the formula: In this embodiment, when the major axis dimension a and short axis dimensions b When both are determined to originate from boundary clipping or precision threshold judgment, based on the initial ellipticity obtained by bidirectional compensation of the major and minor axes using the correction factor γ, a preset first penalty coefficient α is applied to the compensated major axis component, and a preset second penalty coefficient β is applied to the compensated minor axis component. The final updated ellipticity parameter is equal to the major axis dimension. a Multiply by (1+γ) then by (1+α), and divide by the minor axis dimension. b Multiply by (1-γ) and then multiply by (1-β); where, a Multiply by (1+γ) to get the compensated major axis component. b Multiplying by (1-γ) yields the compensated minor axis component, (1+α) is the positive penalty term applied to the major axis component, and (1-β) is the denominator penalty term applied to the minor axis component. Since β is a positive number less than 1, and 1-β is also a positive number less than 1, this factor, located in the denominator, increases the overall ratio, effectively amplifying and correcting the minor axis component. For a specific numerical example, if the major axis dimension... a 2.0 mm, short axis dimensionb With a length of 1.2 mm, a correction factor γ of 0.1, a first penalty coefficient α of 0.05, and a second penalty coefficient β of 0.08, the compensated major axis component is 2.0 multiplied by 1.1, which equals 2.2 mm. The compensated minor axis component is 1.2 multiplied by 0.9, which equals 1.08 mm. The initial ellipticity is 2.2 divided by 1.08, which is approximately 2.037. The updated ellipticity is 2.2 multiplied by 1.05 divided by (1.08 × 0.92), which is 2.31 divided by 0.9936, which is approximately 2.325. This result is the final ellipticity parameter output value after double penalty correction for both major and minor axis dimensions originating from anomalies. This correction mechanism simultaneously suppresses the truncation bias introduced by boundary clipping or accuracy threshold judgment for both major and minor axes, ensuring the output accuracy of the ellipticity parameter when both axis dimension estimates are unreliable.

[0049] B122. Calculate the ellipticity parameter. Combination uncertainty The synthesis uncertainty is: ;in, r The correlation coefficient is a preset value, and r It is a negative value; B123, the combined uncertainty With the preset maximum permissible uncertainty U max If a comparison is made, >U max Then, the ellipticity parameter is output. At the same time, an unreliable flag is attached.

[0050] This process determines the major axis dimension as a weighted average of the predicted thickness along the major axis and the center value of the first thickness range, dynamically adjusting the weights based on the prediction uncertainty. It also switches to the median or directly uses the predicted value under asymmetric or high-precision conditions, effectively suppressing point estimation bias caused by the selection of the midpoint value in the interval estimation. This allows the major axis dimension to possess both the sensitivity of the predicted value and the physical constraint of the interval boundary. Secondly, considering the higher uncertainty, wider interval, and sensitivity to error amplification in the minor axis direction, a benchmark weight coefficient lower than that of the major axis is independently set, and this benchmark weight coefficient is dynamically calculated and truncated. This makes the extraction of the minor axis dimension more dependent on the interval center rather than potentially inaccurate predicted values, thereby suppressing... This invention effectively reduces the point estimation variance of the ellipticity parameter, improves its stability and anti-interference ability under dynamic fluctuations in deposition, and provides a reliable interval and reliability indicator for the discrimination module. Furthermore, it achieves accurate and robust calculation of the ellipticity parameter, providing a high-quality decision-making basis for subsequent closed-loop control. Additionally, it introduces a correction factor based on the uncertainty ratio to weight the major and minor axes in the ellipticity parameter calculation, applies a penalty coefficient to dimensions originating from boundary clipping or accuracy threshold judgment, and calculates the combined uncertainty with a negative correlation coefficient. When the combined uncertainty exceeds the limit, an unreliable flag is added. In summary, this invention significantly reduces the point estimation variance of the ellipticity parameter, improves its stability and anti-interference ability under dynamic fluctuations in deposition, and provides a reliable interval and reliability indicator for the discrimination module. This enables accurate and robust calculation of the ellipticity parameter, providing a high-quality decision-making basis for subsequent closed-loop control.

[0051] The discrimination module is configured to: compare the ellipticity parameter with a preset target ellipticity, and generate a dynamic rotation speed adjustment strategy based on the comparison result. The dynamic rotation speed adjustment strategy includes a rotation speed ratio coefficient of the deposition object at different angular positions in the circumferential direction. It should be further explained that this embodiment generates a dynamic speed adjustment strategy based on the comparison results, including: C101. Obtain the ellipticity parameter η and the preset target ellipticity interval, wherein the target ellipticity interval has a target upper limit value η. max and target lower limit η min , where η max >η min >0; C102, compare the ellipticity parameter η with the target lower limit value η respectively. min and the target upper limit value η max The comparison is as follows: C103, If η < η min If the current ellipticity is too small, an adjustment command to increase the rotation speed ratio coefficient is generated; the rotation speed ratio coefficient K is defined as the rotation speed V of the deposition object at the angular position corresponding to the major axis of the circumferential direction. AC Rotational speed V at the angular position corresponding to the minor axis BD The ratio, i.e., K=V AC / V BD ; C104, If η>η max If the current ellipticity is too large, an adjustment command to reduce the rotation speed ratio coefficient is generated. C105, if η min ≤η≤η max If the current ellipticity is deemed acceptable, an adjustment command is generated to maintain the current rotation speed proportionality constant. C106. Calculate the deviation Δη between the ellipticity parameter η and the target ellipticity interval. The deviation Δη is determined as follows: When η < η min When, Δη=η min -η; When η>η max At that time, Δη = η - η max ; When η min ≤η≤η max At that time, Δη=0; C107. Calculate the correction value ΔK of the rotational speed proportionality coefficient based on the deviation Δη, wherein the correction value ΔK satisfies ΔK=K p ×Δη, where K p The preset proportional gain, and K p It is a positive number; C108. Obtain the rotational speed scaling factor K at the current moment. cur The rotational speed scaling factor K at the current moment cur Adding this to the correction value ΔK yields the updated rotational speed scaling factor K. new K new =K cur +ΔK; C109. The updated rotational speed scaling factor K new This is output as the rotational speed proportional coefficient in the aforementioned dynamic speed adjustment strategy.

[0052] In a specific embodiment, to further optimize the robustness and adaptability of the dynamic speed adjustment strategy and avoid ellipticity overshoot due to single prediction errors or excessive adjustments at the end of deposition, the following extended logic is added to the basic proportional control described above.

[0053] C110, acquire the prediction unreliability flag output by the discrimination module. If either the major axis or minor axis direction unreliability flag is set, pause the current rotation speed adjustment, maintain the current rotation speed proportional coefficient unchanged, and accumulate the number of consecutive unreliable events. When the number of consecutive unreliable events exceeds a preset threshold, trigger a system alarm and prompt to check the stability of the deposition process. In this embodiment, the threshold for the number of consecutive unreliable events is preset by those skilled in the art based on the timeliness requirements of the deposition process for control and the natural probability of the predicted unreliable events. Specifically, when the discrimination module receives a signal that either the major axis or minor axis direction unreliability flag is set, the system pauses the current rotation speed adjustment and maintains the current rotation speed proportional coefficient unchanged, while accumulating the number of consecutive unreliable events. A single prediction unreliability may originate from instantaneous noise of the pyrometer, local anomalies on the surface of the deposited object, or accidental fluctuations of the model under very few operating conditions. However, multiple consecutive unreliable events often indicate a substantial change in the deposition process conditions or a systemic failure of the sensor. In this case, continuing to wait will not only fail to restore prediction reliability but will also delay the control opportunity, causing the ellipticity to continuously deviate from the target. Based on the sampling frequency of the pyrometer, the prediction update cycle of the time-series converter network, and the shortest time interval required for a significant change in the surface temperature of the deposited object during deposition, those skilled in the art can set an empirical threshold that can distinguish between sporadic noise and persistent anomalies. If the sampling cycle is 1 second, and the time scale for changes in deposition process parameters causing measurable deviations in ellipticity is approximately 10 to 30 seconds, then the threshold for consecutive unreliable occurrences can be preset to 5 or 6 times. This means that when the system determines the prediction to be unreliable after 5 or 6 consecutive samplings, the cumulative duration has reached approximately 5 to 6 seconds. At this point, it is still within the time scale of process changes but has significantly exceeded the duration of sporadic noise, triggering a system alarm to prompt the operator to check the stability of the deposition process or the sensor's operating status. Through the above-mentioned preset method, this threshold avoids interrupting the normal control process due to frequent false alarms caused by a single sporadic noise, while ensuring timely warnings when real anomalies occur, thus achieving a balance between system robustness and safety.

[0054] C111, introduce a deposition remaining time scaling factor r, which satisfies r=(T total -t) / T total Where t is the current deposition time, T totalThe preset total deposition time is used. In this embodiment, the deposition remaining time ratio factor r is a dynamic variable calculated in real time by the system based on the preset total deposition time and the current deposition time. The calculation formula is r equal to the difference between the preset total deposition time and the current deposition time, divided by the preset total deposition time. This factor is equal to 1 at the start of the deposition process, gradually decreases as the deposition process progresses, and approaches 0 at the end of the deposition process. It is used to characterize the proportion of the remaining time from the deposition end to the total deposition process time. This proportion value is used by the discrimination module to dynamically correct the ratio gain. Specifically, the basic ratio gain is multiplied by 1 in parentheses and the product of the feedforward coefficient and the remaining time ratio factor is added. Thus, a larger gain is obtained when the remaining time ratio factor is close to 1 in the early stage of deposition to quickly approach the target ellipticity. When the factor is close to 0 in the late stage of deposition, the gain is restored to the basic value to avoid over-adjustment.

[0055] In this embodiment, the preset total deposition time is a fixed value pre-set by those skilled in the art based on process parameters such as the target total deposition thickness of the final product, deposition rate, torch power, and time required for single-layer deposition. Specifically, the preset total deposition time is equal to the target total deposition thickness divided by the average deposition rate, and then multiplied by a time correction coefficient determined based on deposition efficiency. The target total deposition thickness is obtained by inversely calculating the shrinkage rate of the subsequent densification treatment and the final shaped body design dimensions. The average deposition rate is determined by the torch power, the flow rate of the reaction gas, and the rotation speed of the deposition object. This time is determined before the start of the entire deposition process for each deposition object and remains unchanged throughout the entire deposition process. Taking specific values ​​as an example, if the target total deposition thickness is 50 mm and the average deposition rate is 0.5 mm per minute, then the preset total deposition time is 100 minutes. In actual engineering, a certain process margin needs to be added based on factors such as equipment preheating time, torch adjustment delay, and the duration of the deceleration phase at the end of deposition. For example, the final preset is 105 minutes or 110 minutes. This duration serves as the termination reference value of the system timer and is used to calculate the proportion of the current deposition time to the total duration at each sampling moment, thereby providing a fixed time reference for the real-time calculation of the remaining deposition time proportion factor.

[0056] Based on r, the proportional gain K p Dynamic correction is performed, and the corrected proportional gain K p 'Satisfy K p '=K p ×(1+λ×r), where λ is a preset feedforward coefficient, ranging from 0.2 to 0.5. In the early stages of deposition, when r approaches 1, the gain increases to quickly approximate the target ellipticity; in the later stages of deposition, when r approaches 0, the gain returns to its base value to avoid over-adjustment. In this embodiment, the preset feedforward coefficient λ ranges from 0.2 to 0.5, a range determined by those skilled in the art based on the proportional gain K.p The baseline value is determined by considering both the baseline value and the actual needs for adjusting the response speed during the initial deposition phase. Specifically, λ is used to dynamically correct the proportional gain. The corrected proportional gain is equal to the base proportional gain multiplied by 1 (in parentheses) plus the product of λ and the remaining deposition time proportional factor r. In the initial deposition phase, when r is close to 1, the gain amplification factor is 1 plus λ. If λ is too low, for example, less than 0.2, the amplification effect is insufficient, making it difficult to quickly close the gap between the ellipticity and the target value in the initial deposition phase, and failing to effectively utilize the relatively long adjustable time window in the early deposition phase. If λ is too high, for example, greater than 0.5, the corrected gain is too large, which may result in an excessively large adjustment step size of the rotation speed proportional coefficient in the initial deposition phase, causing the ellipticity to exceed the target range and triggering overshoot and oscillation. Those skilled in the art determine the appropriate proportional gain based on the response time constant of the ellipticity to the rotation speed proportional coefficient change during the deposition process and the proportional gain K. p The specific value of λ, combined with the design principles of feedforward compensation in control theory, is limited to an empirical range of 0.2 to 0.5. For example, when K... p When K is large, take a smaller value, such as 0.25, to avoid over-adjustment. p When the value of λ is relatively small, a larger value such as 0.45 is taken to improve the response speed. The grid search optimization can be carried out in the range of 0.2 to 0.5 through deposition process verification experiments. The optimal target is to avoid overshoot during the transition of ellipticity from the initial value to the target range and to minimize the adjustment time. Finally, a specific value of λ is determined as the preset feedforward coefficient, so as to achieve a balance between rapid approximation in the early stage of deposition and avoiding overshoot at the end.

[0057] C112, Set the physical lower limit value K of the rotational speed proportional coefficient K. min and physical upper limit value K max K min K corresponds to the maximum permissible ratio of the longer shaft speed being slower than the shorter shaft speed. max This corresponds to the maximum permissible ratio of the longer shaft speed to the shorter shaft speed. K is updated each time. new Then, perform a limit operation: if K new <K min Let K new =K min If K new >K max Let K new =K max The system records clipping events. When the number of consecutive clipping events exceeds a preset threshold, a prompt indicating that the process parameters need adjustment is issued. In this embodiment, the threshold for consecutive clipping events is preset by those skilled in the art based on the update frequency of the rotation speed ratio coefficient during deposition and the nature of the clipping events. Specifically, when a clipping operation is performed after each update of the rotation speed ratio coefficient K, if K is less than the physical lower limit value K... minOr greater than the physical upper limit value K max The system truncates K to the boundary value and records a limiting event. Single or occasional limiting may be caused by instantaneous fluctuations in the ellipticity parameter or a brief overshoot of the rotation speed ratio coefficient due to prediction noise. When the number of consecutive limiting events exceeds the preset threshold, it indicates that under the current deposition stage and process conditions, the expected rotation speed ratio coefficient continuously calculated by the system has systematically exceeded the physical range allowed by the equipment. This often means that the preset physical lower limit or physical upper limit value itself does not match the current actual deposition state, or the setting of the target ellipticity range has exceeded the range that the rotation speed adjustment capability of the deposition object can reach, rather than a simple instantaneous disturbance. Based on the update cycle of the rotation speed proportional coefficient and the actual response speed of ellipticity control during deposition, those skilled in the art can set an empirical number of occurrences that can distinguish between incidental limiting and systematic design defects. Taking an update cycle of one second as an example, if the time scale for ellipticity deviation from the adjustable range due to temperature drift or sediment rate changes during deposition is approximately ten to twenty seconds, then this threshold can be preset to five or six occurrences. This means that when the system triggers limiting for five or six consecutive updates, the cumulative duration has reached five to six seconds, which is sufficient to rule out the possibility of single-incident noise interference and confirm it as a systematic deviation. When the number of consecutive limiting occurrences exceeds this threshold, the system issues a prompt that the process parameter design needs to be adjusted, prompting the operator to check the physical lower limit value K. min and physical upper limit value K max The setting of the threshold is considered reasonable, or the target ellipticity range is assessed to ensure it falls within the current equipment's capabilities. This preset method avoids frequent false alarms caused by occasional limiting events that could disrupt normal control processes, while also ensuring timely adjustments are made to operators when design flaws in process parameters reflected by continuous limiting occur. This guarantees the long-term stability and maintainability of the ellipticity closed-loop control system.

[0058] C113, To avoid frequent adjustments to ellipticity near the boundaries of the acceptable range, a hysteresis dead zone width Δη is set. dead The qualification condition is extended to: if η min +Δη dead ≤η≤η max -Δη dead If η is within [η], it is considered strictly qualified, and K remains unchanged; min η min +Δη dead ) or (η max -Δη dead η max Within the range, a halved proportional gain K is used. p / 2 is fine-tuned.

[0059] C114 outputs the final determined rotational speed proportionality coefficient K. finalThe adjustment flag, limit flag, and alarm flag are sent to the adjustment module together. This embodiment realizes adaptive adjustment for different stages of the deposition process, different prediction reliability, and physical constraints, which significantly improves the stability and safety of ellipticity closed-loop control.

[0060] This embodiment improves system safety and fault tolerance by introducing a prediction unreliability flag and accumulating consecutive unreliable counts, pausing speed adjustment and triggering an alarm when prediction confidence is insufficient, thus avoiding the execution of erroneous control commands due to network prediction errors. Secondly, it introduces a deposition remaining time scaling factor to dynamically correct the scaling gain, increasing the gain in the early stages of deposition to quickly approach the target ellipticity and restoring the base gain at the end of deposition to avoid overshoot, achieving adaptive matching between control intensity and deposition process, effectively suppressing end-stage overshoot. Thirdly, it sets physical upper and lower limits for the rotational speed scaling factor and performs a limiting operation to prevent the speed ratio from exceeding the equipment's allowable range. Simultaneously, it uses continuous limiting events to indicate process parameter design defects, enhancing the system's physical constraint compliance and diagnostics. Finally, it sets a hysteresis dead zone width and adjusts the gain in stages, using a halved gain for fine-tuning when the ellipticity approaches the acceptable boundary, and maintaining the scaling factor unchanged within the strictly acceptable range, avoiding frequent jitter and ineffective adjustments near the boundary.

[0061] The adjustment module is configured to: adjust the actual rotation speed of the deposition object at different angular positions in the circumferential direction according to the rotation speed ratio coefficient and the A3C algorithm, so as to change the thickness of the deposition at the corresponding position, so that the ellipticity of the deposited body meets the preset target ellipticity.

[0062] In a specific embodiment, the technical problem addressed by adjusting the actual rotational speed of the deposition object at different angular positions along the circumference according to the rotational speed proportional coefficient is that the actual rotational speed of the deposition object needs to change continuously along the circumference and cannot only jump at four characteristic angle points. Furthermore, the rotational drive system suffers from response delay due to mechanical inertia. If the proportional coefficient is directly assigned only to the characteristic angle positions, the deposition thickness transition at intermediate angle positions will be uneven, resulting in local unevenness or wavy distortion at the elliptical edge of the sintered body. Simultaneously, the deposition object undergoes continuous reciprocating rotational motion during deposition, and sudden speed changes can trigger mechanical vibration, affecting the density uniformity and thickness consistency of the deposition. In this embodiment, the generation and execution of the dynamic rotational speed adjustment strategy are globally optimized by the A3C reinforcement learning algorithm. Specifically, the A3C algorithm constructs a policy network and a value network. Both networks adopt a fully connected layer structure and share the parameters of the first few layers to extract common state features. The input to the policy network is the current deposition state vector, which consists of the remaining deposition time proportional factor r, the deviation of the ellipticity parameter η from the upper and lower limits of the preset target ellipticity interval, and the current rotational speed proportional coefficient K. cur The output of the policy network is composed of five components: the numerical value of the prediction, the Boolean value of the unreliable flag for the major axis direction, and the Boolean value of the prediction unreliable flag for the minor axis direction. The output of the policy network is the correction amount ΔK of the rotation speed scaling factor. fineThe input to the value network is the same as that of the policy network, and the output is an estimate of the expected cumulative reward that can be obtained by performing the corrective action in the current state. The reward function consists of a weighted sum of three parts: the first part is an ellipticity deviation penalty, which is the negative of the absolute deviation of the current ellipticity parameter η from the midpoint of the preset target ellipticity interval; the second part is a rotational speed change smoothness penalty, which is the negative of the absolute value of the change in the proportional coefficient of two adjacent rotational speeds; the third part is an overshoot penalty at the end of deposition, which takes effect when the remaining deposition time proportional factor r is less than the preset end-of-period threshold and the ellipticity parameter η exceeds the preset target ellipticity interval, and is the negative of the overshoot multiplied by an amplification factor. During the training phase, the A3C algorithm simultaneously collects state transition samples in multiple deposition process simulation environments through asynchronous parallel interaction. Each simulation environment independently runs a replica of the deposition process and adjusts the rotation speed scaling factor according to the correction amount output by the current policy network. It obtains the deposition state and immediate reward at the next time step and stores the state transition trajectory in a shared experience buffer. During the parameter update phase, the algorithm uses the advantage function to calculate the policy gradient to update the policy network parameters. The advantage function is calculated by adding the immediate reward of the current action to the value network's estimate of the next state and subtracting the value network's estimate of the current state. Simultaneously, it uses temporal difference error to update the value network parameters. The temporal difference error is calculated by adding the immediate reward of the current action to the value network's estimate of the next state and subtracting the value network's estimate of the current state. After training, the A3C algorithm is deployed in the online inference phase. Each time, the discrimination module outputs the rotation speed scaling factor K. new Then, the A3C algorithm uses the current deposition state as input, and the policy network performs forward inference to obtain the correction amount ΔK. fine , will K new With ΔK fine The summation yields the final rotational speed proportionality coefficient K. final K final It serves as the input reference for calculating the speed command in the subsequent D101.

[0063] To address the above problems, this embodiment provides the following detailed implementation steps: D101, obtain the final rotational speed scaling factor K output by the discrimination module, and obtain the preset reference rotational speed of the deposition object at the zero point position. V base The zero point position corresponds to one end of the major axis in the circumferential direction, and the reference rotational speed... V base Determined based on the baseline deposition rate required by the deposition process, specifically: reference rotation speed. V base The speed is determined by the torch deposition efficiency, reactive gas flow rate, deposit density, and the relative speed between the torch and the surface of the object being deposited; in this embodiment, the reference rotation speed is... Vbase The standard rotational speed, defined and preset as the zero point position of the deposition object in the circumferential direction, is used to provide the basic deposition rate required for the deposition process at the corresponding angular position of the major axis, i.e., the thickness increment of the deposited material on the surface of the deposition object per unit time. The basic deposition rate is obtained by dividing the target total deposition thickness by the preset total deposition time. For example, if the target total deposition thickness is 50 mm and the preset total deposition time is 100 minutes, the basic deposition rate is 0.5 mm per minute. Based on this, those skilled in the art, combined with the torch power, reactive gas flow rate, distance between the torch and the surface of the deposition object, and the bulk density of the deposit, determine the specific value of the rotational speed of the deposition object corresponding to this basic deposition rate through experimental calibration. The calibration process involves: conducting multiple sets of deposition experiments at different rotational speeds under fixed torch power and gas flow rate conditions, measuring the average increase in deposition thickness per unit time in each set of experiments, obtaining the correlation curve between rotational speed and deposition rate, and then finding the corresponding rotational speed value from this curve based on the required basic deposition rate as the reference rotational speed. V base Simultaneously, it is necessary to ensure that this speed is within the stable operating speed range of the rotary drive motor and to allow sufficient adjustment margin to accommodate subsequent increases or decreases in the speed proportional coefficient. Through the above mapping relationship, the reference speed... V base The determination of makes the actual rotational speed at the corresponding angular position of the major axis equal to V base The actual rotational speed at the angular position corresponding to the short axis is equal to V. base By dividing by the rotation speed proportionality coefficient K, a precise match and dynamic coordination between the deposition rate and the rotation speed is achieved throughout the deposition process.

[0064] D102, based on the rotational speed proportionality coefficient K and the reference rotational speed V base Calculate the actual rotational speed V of the deposition object at points A and C corresponding to the major axis angle. AC And let V AC = V base ; Calculate the actual rotational speed V of the deposition object at points B and D corresponding to the minor axis angle. BD And let V BD = V base / K. Where point A corresponds to a circumferential angle of 0 degrees, point B corresponds to 90 degrees, point C corresponds to 180 degrees, and point D corresponds to 270 degrees.

[0065] D103, for any intermediate angle θ in the circumferential direction, the actual rotational speed command V at that angle position is determined using a cosine interpolation function. cmd(θ) is used to ensure that the velocity curve is continuous over the entire revolution and that the first derivative is continuous. The expression for the cosine interpolation function is: V cmd (θ)=V AC -(V AC -V BD )×(1-cos(2θ)) / 2, where θ is in radians and ranges from 0 to 2π. This expression equals V when θ=0 and θ=π. AC It equals V when θ = π / 2 and θ = 3π / 2. BD Furthermore, the derivative is zero at each point, thus achieving a smooth transition.

[0066] As another optional interpolation method in this embodiment, a linear interpolation method is used: within the interval from point A to point B, i.e., when θ∈(0, π / 2], V cmd (θ)=V AC -(V AC -V BD )×(2θ / π); In the interval from point B to point C, i.e., when θ∈(π / 2,π], V cmd (θ)=V BD +(V AC -V BD )×(2(θ-π / 2) / π); In the interval from point C to point D, i.e., when θ∈(π, 3π / 2], V cmd (θ)=V AC -(V AC -V BD )×(2(θ-π) / π); In the interval from point D to point A, i.e., when θ∈(3π / 2, 2π], V cmd (θ)=V BD +(V AC -V BD )×(2(θ-3π / 2) / π); The linear interpolation method is simple to implement, has low computational cost, and is suitable for scenarios with high real-time requirements.

[0067] D104 will calculate the actual rotational speed command V at each angular position. cmd (θ) According to the real-time rotation angle θ of the sedimentation object curr The signal is sent to the controller of the rotary drive motor; to avoid mechanical shock caused by sudden changes in speed, a speed change rate limiter is set at the front end of the controller: the speed command V at the previous sampling moment is recorded. prev Calculate the desired rotational speed command V at the current moment. des If |V des -V prev |Exceeding the preset maximum acceleration value A max Multiplying by the sampling period Δt, the actual output speed command V is obtained. outRestricted to V prev ±A max Δt, and sends the limited command to the motor; in this embodiment, the maximum acceleration value A max It is determined in advance through experimental calibration based on the comprehensive constraints of the mechanical characteristics of the deposition object's driving system and the vibration sensitivity of the deposition process. Specifically, the mechanical characteristics of the deposition object drive system include the maximum output torque of the rotary drive motor, the transmission inertia of the reduction mechanism, and the total rotational inertia of the deposition object and its clamping mechanism. These parameters together determine the maximum rate of change of rotational speed that the deposition object can withstand per unit time without causing the drive motor to lose synchronism or excessive wear of the mechanical transmission components. Those skilled in the art can drive the deposition object to rotate under no-load conditions with gradually increasing acceleration values, record the torque current and speed tracking error fed back by the motor driver, and determine the maximum acceleration value corresponding to the condition that does not trigger the driver overcurrent protection and the speed tracking error does not exceed a preset threshold as the upper limit of the mechanical characteristics. The sensitivity of the deposition process to vibration is manifested in the fact that the mechanical vibration generated by the deposition object during the acceleration transition will be transmitted to the deposition area, causing density fluctuations or thickness inhomogeneities on the surface of the deposition object, which in turn affects the ellipticity consistency of the shaped body after deposition. Those skilled in the art can conduct deposition experiments with different acceleration values ​​under the same temperature field and gas flow field environment as the actual deposition conditions, use an acceleration sensor installed on the bearing seat of the deposition object to monitor the vibration amplitude in real time, and simultaneously detect the thickness uniformity of the deposition object surface after deposition, and determine the maximum acceleration value that does not cause measurable thickness fluctuations on the surface of the deposition object as the upper limit of vibration sensitivity. After both sets of constraints are determined, those skilled in the art will take the smaller of the upper limit of mechanical characteristics and the upper limit of vibration sensitivity as A. max The initial calibration value, for example, if the upper limit of the mechanical characteristic is 5 revolutions per second. 2 The upper limit of vibration sensitivity is 3 revolutions per second. 2 Then A max The initial calibration was 3 revolutions per second. 2 Then, based on the tracking response quality of the rotation speed command during the actual deposition process, fine-tuning is performed to finally determine an A value that can both ensure the stable operation of the drive system and avoid vibration interference with the deposition thickness. max Specific numerical values; during system operation, the rate of change limiter utilizes this A. max The product of the value and the sampling period Δt is used as the maximum allowable change in the speed command between adjacent sampling times. The desired speed command is limited and then output to the motor controller, thereby achieving a balance between fast response and smooth operation.

[0068] D105, after each speed adjustment command is issued, uses a high-precision encoder mounted on the motor shaft to collect the actual speed V of the deposition object in real time. actualAnd compare the actual speed with the speed command V output at the current moment. out Compare and calculate the speed tracking deviation ΔV = |V actual -V out |;If ΔV continuously exceeds the preset speed tracking error threshold E th The time exceeded the preset tolerance time T tol If the error occurs, it is determined to be a tracking anomaly, triggering closed-loop compensation: the deviation signal is converted into an additional voltage compensation amount by the proportional-integral controller and superimposed on the motor drive signal, while the tracking anomaly event is recorded; if the number of consecutive tracking anomalies exceeds the preset maximum allowable number, a system alarm is issued, prompting the drive mechanism or encoder to be checked.

[0069] In this embodiment, the rotational speed tracking error threshold E th This is preset by those skilled in the art based on the steady-state speed control accuracy of the driving system for the deposition object and the degree of influence of speed deviation on the uniformity of deposition thickness. Specifically, the speed tracking error of the driving system during steady-state operation originates from the quantization error of the motor encoder, the steady-state static error of the speed loop regulator, and the minute backlashes of transmission components such as gears or belts in the mechanical transmission chain. These factors together constitute the inherent noise base of the actual speed fluctuating around the commanded speed. Those skilled in the art, by running the driving system with a fixed speed command under high-precision encoder feedback, record the statistical distribution of the difference between the actual speed and the commanded speed under stable operating conditions. For example, they obtain the standard deviation or peak fluctuation amplitude of this difference as a quantitative indicator of the system's steady-state tracking accuracy. th The preset value must be higher than the inherent noise floor to avoid frequent triggering of tracking anomaly judgments due to minor fluctuations in the system itself. Simultaneously, the deposition process's requirement for rotational speed accuracy manifests as the actual rotational speed of the deposition object deviating from the commanded value, causing the deposition thickness at various angular positions along the circumference to deviate from the expected value, thus affecting the ellipticity consistency of the deposited body. The amount of deposition thickness deviation corresponding to each unit increase in rotational speed deviation is jointly determined by the deposition rate and the torch deposition efficiency. For example, if the deposition rate is 0.01 mm per revolution and the effective deposition width of the torch on the deposition object surface corresponds to an angular span of 30 degrees, then each deviation of one revolution per minute results in a thickness deviation of approximately 0.01 mm in that angular region. Those skilled in the art can deduce the maximum allowable rotational speed deviation value as E based on the maximum allowable thickness deviation allowed by ellipticity control. th The upper limit reference. After considering the above two sets of constraints, those skilled in the art would take a reasonable value between the noise floor of the system's steady-state tracking accuracy and the maximum allowable rotational speed deviation of the deposition process as E. th The preset value, for example, if the peak fluctuation of the steady-state tracking error is 0.5 revolutions per minute, and the maximum allowable rotational speed deviation of the deposition process is 3 revolutions per minute, then E can be set to... thThe preset speed is 1.5 revolutions per minute or 2 revolutions per minute. During trial operation, the trigger frequency of tracking abnormal events is observed and fine-tuned to ultimately determine an E that can both avoid noise-induced false triggers and promptly detect actual tracking failures. th Specific numerical values; during system operation, when the absolute value of the difference between the actual rotational speed and the output command continuously exceeds E... th The time exceeded the preset tolerance time T tol When this occurs, the system determines that the tracking is abnormal and initiates closed-loop compensation.

[0070] In this embodiment, the tolerance time T tol This is a comprehensive preset by those skilled in the art based on the transient disturbance characteristics of the rotational speed of the driving system driven by the deposition object and the timeliness requirements of the deposition process for responding to tracking anomalies. Specifically, in actual operation, due to factors such as minute backlashes in the mechanical transmission chain, motor commutation pulsation, encoder quantization errors, and load fluctuations, the speed tracking deviation ΔV may experience occasional instantaneous spikes exceeding E within a very short time. th However, these spikes usually recover to the normal range on their own within a few sampling periods and do not indicate a persistent tracking failure in the system. Therefore, a tolerance period needs to be set to filter out these transient disturbances caused by the inherent characteristics of the drive system. This tolerance period must be greater than the duration of a single sporadic spike, usually several times that duration. Meanwhile, the deposition process's response timeliness to tracking anomalies is such that if the actual rotational speed deviation exceeds E... th If the duration of the deviation is too long, the thickness of the deposited material at that angular position will continuously deviate from the expected value during the period of sustained deviation, leading to an increase in the cumulative deviation of the deposition thickness in that area. If this deviation continues until the object being deposited rotates more than half a turn, it will significantly affect the ellipticity of the formed body. Therefore, the tolerance period should be much shorter than the time required for the object to rotate one full turn, typically between tens of milliseconds and one or two seconds. Those skilled in the art monitor the transient characteristics of the rotation speed tracking deviation curve under actual deposition conditions and statistically analyze the situation where the deviation continuously exceeds E during normal operation of the drive system. th The duration distribution, for example, if statistics show that more than 95% of occasional spikes last no more than 0.5 seconds, while the time required for the deposited object to rotate once is, for example, ten seconds, then the tolerance period can be preset to two seconds, indicating that when the deviation continuously exceeds E... th If the deviation persists for two seconds, the system determines it to be a definite tracking anomaly rather than an occasional disturbance; this preset value, in seconds, serves as a timing reference for each deviation exceeding E. th The timer starts when the timer reaches T. tol The previous deviation has fallen back to E th The following will reset the timer and treat it as normal fluctuation. If the timer reaches T... tol If the time deviation still does not decrease, closed-loop compensation or an alarm will be triggered. Through the aforementioned preset method, T tolIt effectively eliminates the inherent transient spike noise of the drive system to avoid false triggering of tracking anomaly determination, and ensures that when a real tracking failure occurs, the compensation mechanism can be activated in time before it causes unacceptable impact on the cumulative deviation of the deposition thickness.

[0071] In this embodiment, the maximum allowable number of times is preset by those skilled in the art based on the compensation convergence characteristics of rotational speed tracking anomalies and the tolerance limit of the deposition process for the duration of anomalies; specifically, when the rotational speed tracking deviation continuously exceeds the rotational speed tracking error threshold E th The time exceeded the tolerance duration T tol When a tracking anomaly is triggered, the system identifies it as a tracking anomaly and initiates closed-loop compensation, while simultaneously accumulating the number of consecutive tracking anomalies. After each tracking anomaly is triggered, the proportional-integral controller converts the deviation signal into an additional voltage compensation amount, which is then superimposed on the motor drive signal to correct the speed. This compensation process requires a certain convergence time for the actual speed to re-track the command value. If a single closed-loop compensation successfully eliminates the deviation, the accumulated count is reset to zero. Therefore, a single anomaly is usually a recoverable, occasional disturbance. When tracking anomalies are triggered multiple times consecutively, it indicates that the closed-loop compensation has failed to effectively eliminate the tracking deviation. This may stem from persistent problems such as encoder malfunction leading to distorted feedback signals, abnormal motor driver output, severe wear of transmission components, or mechanical jamming. In this case, continuing to trigger compensation in a loop will not only fail to restore tracking accuracy but may also cause speed overshoot or system oscillation due to the continuous accumulation of compensation. Furthermore, during this continuous anomaly period, the deposition thickness deviation has accumulated over multiple time cycles, and its impact on ellipticity is no longer negligible. Those skilled in the art, based on the typical number of time cycles required for a single convergence of closed-loop compensation (e.g., compensation convergence typically requires two to three sampling update cycles to eliminate deviation), set a maximum allowable number of convergences slightly higher than this value, such as a preset five or six times. This indicates that if the system fails to recover after triggering tracking anomalies five consecutive times, it means the anomaly has exceeded the compensation capability, further waiting is futile, and the deposition deviation has accumulated to an unacceptable level. Through this preset method, this threshold number provides sufficient opportunities for closed-loop compensation to handle occasional recoverable disturbances, while also ensuring timely system alarms in the event of continuous failures, prompting operators to check the drive mechanism or encoder's operating status, thereby guaranteeing the continuity of the deposition process and the reliability of ellipticity control.

[0072] This embodiment obtains the rotational speed proportionality coefficient K and the reference rotational speed V. baseThe invention employs cosine or linear interpolation to generate continuous rotational speed commands at arbitrary angular positions along the circumference, avoiding speed jumps caused by assigning values ​​only at characteristic angle points. This ensures a smooth transition of sediment thickness along the circumference, effectively suppressing local unevenness and wavy distortion at the elliptical edge of the deposited object, and improving the smoothness and dimensional consistency of the elliptical profile. Secondly, a speed change rate limiter is set to limit the rotational speed command based on the maximum acceleration value, preventing mechanical vibration caused by sudden speed changes and ensuring the stability of the continuous reciprocating rotational motion of the deposited object, thereby improving the density uniformity and thickness consistency of the sediment. Thirdly, the actual rotational speed is fed back in real time by the encoder and compared with the command value. When the tracking deviation exceeds the limit, proportional-integral closed-loop compensation is triggered, and alarms are triggered based on the accumulated number of abnormalities. This solves the problem of response delay and tracking error caused by the mechanical inertia of the drive system, enhancing the accuracy and robustness of speed control. In summary, the adjustment module of this invention realizes the generation and closed-loop tracking of smooth, continuous, and controllable rotational speed commands from proportional coefficients, providing a reliable motion execution basis for ellipticity closed-loop control, and ultimately ensuring that the ellipticity of the deposited object accurately meets the preset target.

[0073] Example 2 Please see Figure 2 Another embodiment of the present invention provides a method for controlling the ellipticity of deposited objects based on a time-series transformer network, comprising: S1. Acquire the time-series data of the surface temperature of the deposited object during the deposition process, the real-time rotation speed of the deposited object at different angular positions in the circumferential direction, and the deposition time; S2. The time-series data of the surface temperature of the deposited object, the real-time rotation speed of the point, and the deposition time are processed by a preset time-series transformer network model to estimate the first thickness range of the current shaped body in the long axis direction and the second thickness range in the short axis direction. S3. Determine the major axis dimension based on the first thickness range, determine the minor axis dimension based on the second thickness range, and calculate the ellipticity parameter from the major axis dimension and the minor axis dimension; S4. Compare the ellipticity parameter with the preset target ellipticity, and generate a dynamic rotation speed adjustment strategy based on the comparison result. The dynamic rotation speed adjustment strategy includes a rotation speed ratio coefficient of the deposition object at different angular positions in the circumferential direction. S5. Based on the rotation speed ratio coefficient and the A3C algorithm, adjust the actual rotation speed of the deposition object at different angular positions in the circumferential direction to change the thickness of the deposition at the corresponding positions, so that the ellipticity of the deposited body meets the preset target ellipticity.

[0074] Example 3 An electronic device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement a deposition object ellipticity control system based on a time-series transformer network.

[0075] A computer-readable storage medium having computer instructions stored thereon, which, when executed, perform a deposition object ellipticity control system based on a time-series transformer network.

[0076] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments under the guidance of the present invention without departing from the spirit and scope of the claims. All of these variations are within the protection scope of the present invention.

Claims

1. A deposition object ellipticity control system based on a time-series transformer network, characterized in that, include: The acquisition module is configured to: acquire time-series data of surface temperature of the deposited object during the deposition process, real-time position rotation speed of the deposited object at different angular positions in the circumferential direction, and deposition time; The prediction module is configured to: process the time-series data of the surface temperature of the deposited object, the real-time rotation speed of the point, and the deposition time through a preset time-series transformer network model, and predict the first thickness range corresponding to the long axis direction and the second thickness range corresponding to the short axis direction of the current shaped body; The major and minor axis determination module is configured to: determine the major axis dimension based on the first thickness range, determine the minor axis dimension based on the second thickness range, and calculate the ellipticity parameter from the major axis dimension and the minor axis dimension; The discrimination module is configured to: compare the ellipticity parameter with a preset target ellipticity, and generate a dynamic rotation speed adjustment strategy based on the comparison result. The dynamic rotation speed adjustment strategy includes a rotation speed ratio coefficient of the deposition object at different angular positions in the circumferential direction. The adjustment module is configured to: adjust the actual rotation speed of the deposition object at different angular positions in the circumferential direction according to the rotation speed ratio coefficient and the A3C algorithm, so as to change the thickness of the deposition at the corresponding position, so that the ellipticity of the deposited body meets the preset target ellipticity.

2. The deposition object ellipticity control system based on time-series transformer network as described in claim 1, characterized in that, The estimated first thickness range along the long axis and the second thickness range along the short axis of the current formed body include: Obtain the predicted thickness value, prediction uncertainty, and prediction uncertainty in the major axis direction from the output layer of the time-series transformer network model, wherein the prediction uncertainty in the major axis direction represents the radius of the confidence interval of the predicted thickness value in the major axis direction, and the prediction uncertainty in the minor axis direction represents the radius of the confidence interval of the predicted thickness value in the minor axis direction. The lower limit of the thickness in the long axis direction is obtained by subtracting the prediction uncertainty in the long axis direction from the predicted thickness value in the long axis direction. The upper limit of the thickness in the long axis direction is obtained by adding the predicted thickness value in the long axis direction to the predicted uncertainty in the long axis direction.

3. The deposition object ellipticity control system based on a time-series transformer network as described in claim 2, characterized in that, The estimation of the first thickness range corresponding to the long axis direction and the second thickness range corresponding to the short axis direction of the current molded body further includes: Using the lower limit value of the thickness in the long axis direction and the upper limit value of the thickness in the long axis direction as boundaries, a first thickness range corresponding to the long axis direction is constructed. The lower limit of the thickness in the minor axis direction is obtained by subtracting the minor axis direction prediction uncertainty from the minor axis direction thickness prediction value. Adding the predicted thickness value in the minor axis direction to the predicted uncertainty in the minor axis direction yields the upper limit value of the thickness in the minor axis direction. Using the lower limit of the thickness in the short axis direction and the upper limit of the thickness in the short axis direction as boundaries, a second thickness range corresponding to the short axis direction is constructed.

4. The deposition object ellipticity control system based on a time-series transformer network as described in claim 3, characterized in that, Determining the major axis dimension based on the first thickness range includes: Calculate the center value of the first thickness range; The weighting coefficient is dynamically set according to the prediction uncertainty of the major axis direction, wherein the weighting coefficient is inversely proportional to the prediction uncertainty of the major axis direction. The predicted thickness along the major axis is weighted and averaged with the center value, and the weight of the weighted average is the weight coefficient to obtain a first weighted result, and the first weighted result is used as the major axis dimension. When the asymmetry of the first thickness range after boundary trimming exceeds a preset asymmetry threshold, the median of the first thickness range is obtained, and the median is used as the major axis dimension. When the width of the first thickness range is less than a preset accuracy threshold, the predicted thickness value in the long axis direction is directly used as the long axis dimension.

5. The deposition object ellipticity control system based on a time-series transformer network as described in claim 4, characterized in that, The ellipticity parameter is calculated from the major axis dimension and the minor axis dimension, including: Obtain the major axis dimension a minor axis dimensions b Predicted thickness along the major axis P a Predicted thickness along the minor axis P b Major axis direction prediction uncertainty and the prediction uncertainty of the minor axis direction ; The uncertainty is predicted based on the major axis direction. and the predicted uncertainty of the minor axis direction Calculate the correction factor γ; According to the correction factor γ Regarding the major axis dimension a Perform the first weighted compensation to obtain the compensated major axis component A; According to the correction factor γ Regarding the minor axis dimension b Perform a second weighted compensation to obtain the compensated minor axis component B.

6. The deposition object ellipticity control system based on a time-series transformer network as described in claim 5, characterized in that, The ellipticity parameter is calculated from the major axis dimension and the minor axis dimension, and further includes: Dividing the compensated major axis component A by the compensated minor axis component B yields the ellipticity parameter. η ; Determine the length of the major axis a Whether it originates from boundary clipping or a precision threshold judgment, if so, then the major axis dimension... a Apply a preset first penalty coefficient α The ellipticity parameter Updated to ; Determine the minor axis dimension b Whether it originates from boundary clipping or a precision threshold judgment, if so, then adjust the minor axis dimension. b Apply a preset second penalty coefficient β to update the ellipticity parameter η. .

7. The deposition object ellipticity control system based on a time-series transformer network as described in claim 6, characterized in that, The ellipticity parameter is calculated from the major axis dimension and the minor axis dimension, and further includes: If the major axis dimension a and the minor axis dimension b If all values ​​originate from boundary clipping or precision threshold judgment, then the ellipticity parameter η will be updated to... ; Calculate the ellipticity parameter Combination uncertainty ; B123, the combined uncertainty With the preset maximum permissible uncertainty U max If a comparison is made, >U max Then, the ellipticity parameter is output. At the same time, an unreliable flag is attached.

8. The deposition object ellipticity control system based on a time-series transformer network as described in claim 6, characterized in that, The process of generating a dynamic speed adjustment strategy based on the comparison results includes: Obtain the ellipticity parameter η and the preset target ellipticity interval, wherein the target ellipticity interval has a target upper limit value η. max and target lower limit η min , where η max >η min >0; The ellipticity parameter η is respectively compared with the target lower limit value η min and the target upper limit value η max The comparison is as follows: If η < η min If the current ellipticity is too small, an adjustment command to increase the rotation speed ratio coefficient is generated; the rotation speed ratio coefficient K is defined as the rotation speed V of the deposition object at the angular position corresponding to the major axis of the circumferential direction. AC Rotational speed V at the angular position corresponding to the minor axis BD The ratio, i.e., K=V AC / V BD ; If η>η max If the current ellipticity is too large, an adjustment command to reduce the rotation speed ratio coefficient is generated. If η min ≤η≤η max If the current ellipticity is deemed acceptable, an adjustment command is generated to maintain the current rotation speed proportionality constant.

9. The deposition object ellipticity control system based on a time-series transformer network as described in claim 8, characterized in that, The method for generating a dynamic speed adjustment strategy based on the comparison results also includes: Calculate the deviation Δη between the ellipticity parameter η and the target ellipticity interval, wherein the deviation Δη is determined as follows: Yes<h min When Δη=η min -or; Current max When Δη=η-η max ; When min ≤η≤η max When Δη=0; The correction value ΔK of the rotational speed proportional coefficient is calculated based on the deviation Δη, and the correction value ΔK satisfies ΔK=K p ×Δη, where K p The preset proportional gain, and K p It is a positive number; Obtain the rotational speed scaling factor K at the current moment. cur The rotational speed scaling factor K at the current moment cur Adding this to the correction value ΔK yields the updated rotational speed scaling factor K. new K new =K cur +ΔK; The updated rotational speed scaling factor K new This is output as the rotational speed proportional coefficient in the aforementioned dynamic speed adjustment strategy.

10. A method for controlling the ellipticity of deposited objects based on a time-transformer network, implemented based on the ellipticity control system for deposited objects based on a time-transformer network as described in any one of claims 1-9, characterized in that, include: Acquire time-series data of surface temperature of the deposited object during the deposition process, real-time position rotation speed of the deposited object at different angular positions in the circumferential direction, and deposition duration; The time-series data of the surface temperature of the deposited object, the real-time rotation speed of the point, and the deposition time are processed by a preset time-series transformer network model to estimate the first thickness range of the current shaped body in the long axis direction and the second thickness range in the short axis direction. The major axis dimension is determined based on the first thickness range, the minor axis dimension is determined based on the second thickness range, and the ellipticity parameter is calculated from the major axis dimension and the minor axis dimension. The ellipticity parameter is compared with a preset target ellipticity, and a dynamic rotation speed adjustment strategy is generated based on the comparison result. The dynamic rotation speed adjustment strategy includes a rotation speed ratio coefficient of the deposition object at different angular positions in the circumferential direction. Based on the rotation speed ratio coefficient and the A3C algorithm, the actual rotation speed of the deposition object at different angular positions in the circumferential direction is adjusted to change the thickness of the deposition at the corresponding positions, so that the ellipticity of the deposited body meets the preset target ellipticity.

Images