Defect feedback based thin film production process closed loop regulation method and system

Abstract

The application discloses a kind of thin film production process closed-loop regulation method and system based on defect feedback, it is related to thin film production regulation field.The scheme is solved the absolute physical coordinates of thin film by real-time acquisition spindle position encoding signal, and the process parameter flow of time dimension and the detection image flow of space dimension are forcedly unified mapping to this physical coordinate system.Under the condition that defect is detected by deep neural network, the system utilizes the physical coordinates of defect to carry out accurate space-time backtracking in process state atlas, retrieves the actual history process parameter of the section of thin film when passing through upstream processing area, to construct accurate defect-history parameter causal association data pair.Based on the causal relationship determined, optimal compensation value is solved using sensitivity matrix model, and feedforward regulation is implemented to actuating mechanism, effectively overcome the control oscillation of lag system, and high-precision closed-loop regulation is realized.

Description

Technical Field

[0001] This application relates to the field of thin film production control, and more specifically, to a closed-loop control method and system for thin film production process based on defect feedback. Background Technology

[0002] In the roll-to-roll mass production of modern thin film materials (such as optical films, lithium battery separators, and functional coatings), with increasingly stringent requirements for product thickness uniformity, surface finish, and physical properties, precise process control has become a core factor determining yield. To reduce production costs and waste generation, the industry urgently needs to build a closed-loop control system based on defect feedback. This system utilizes online detection equipment to capture thin film surface defects in real time and automatically corrects upstream production process parameters accordingly, maintaining production stability and high quality.

[0003] However, existing closed-loop control schemes for thin film production processes based on defect feedback generally face feedback inaccuracies due to spatial-temporal lag in practical applications. A thin film production line is a typical system with large lag and distributed parameters. Automated Optical Inspection (AOI) equipment is deployed downstream or at the end of the production line, while core process actuators (such as deposition sources, coating heads, and heaters) are located upstream. When the vision system detects a defect downstream, the physical location of the film that caused the defect actually passed through the processing area at some point in the past. Existing control technologies are often limited to lumped parameter thinking, lacking a dynamic spatiotemporal mapping mechanism, and directly adjusting current process parameters using PID control or simple feedback based on currently detected defect information. This adjustment method, which uses past errors to correct present and future process states that may not have errors, ignores the physical distance and time delay during film transport, easily leading to system control oscillations and even causing new defects due to erroneous compensation actions.

[0004] Furthermore, existing systems also have significant limitations in data fusion. Data such as temperature, flow rate, and tension collected by process sensors are stored as time series, while visual inspection data is stored as image frames, lacking a unified and high-precision index key (such as absolute physical coordinates) between the two. Especially when there are fluctuations in winding speed or changes in tension on the production line, relying solely on timestamps cannot accurately align the data, making it impossible to precisely trace downstream defects back to the specific moment of abrupt change in upstream process parameters. Due to the lack of high-precision spatiotemporal backtracking and causal correlation capabilities based on absolute physical location, existing closed-loop systems struggle to calculate accurate process compensation amounts, failing to achieve truly effective closed-loop control of the production process. Summary of the Invention

[0005] To address the aforementioned problems in the existing technology, this application provides a closed-loop control method for thin film production process based on defect feedback, comprising: Step 1: Calculating the absolute physical coordinates of the thin film in real time using the position encoding signal from the production line spindle, and mapping and binding the synchronously acquired process parameter stream and detection image stream to the absolute physical coordinates to obtain a process state map and a positioning image stream; Step 2: Inputting the positioning image stream into a preset deep neural network model for defect feature extraction and vectorization to obtain a defect feature vector; Step 3: Based on the defect physical position coordinates in the defect feature vector, performing spatiotemporal backtracking and parameter association based on position index in the process state map to obtain historical process parameters, and combining the historical process parameters with the defect feature vector to obtain a causal correlation data pair; Step 4: Inputting the causal correlation data pair into a preset sensitivity matrix model for multi-parameter sensitivity analysis and compensation value calculation to obtain a control correction command; Step 5: Superimposing the control correction command onto the current production baseline setpoint and driving the production line controller to perform feedforward compensation adjustment on the process actuator.

[0006] This application also provides a closed-loop control system for thin film production based on defect feedback, comprising: a process parameter detection image acquisition module, used to calculate the absolute physical coordinates of the thin film in real time using the position encoding signal from the main shaft of the production line, and to map and bind the synchronously acquired process parameter stream and detection image stream to the absolute physical coordinates to obtain a process state map and a positioning image stream; a defect feature extraction and vectorization module, used to input the positioning image stream into a preset deep neural network model for defect feature extraction and vectorization processing to obtain a defect feature vector; a causal association module, used to perform spatiotemporal backtracking and parameter association based on the defect physical position coordinates in the defect feature vector in the process state map to obtain historical process parameters, and to combine the historical process parameters with the defect feature vector to obtain a causal association data pair; a control correction instruction generation module, used to input the causal association data pair into a preset sensitivity matrix model for multi-parameter sensitivity analysis and compensation value calculation to obtain a control correction instruction; and a feedforward compensation adjustment module, used to superimpose the control correction instruction onto the current production baseline setpoint and drive the production line controller to perform feedforward compensation adjustment on the process actuator.

[0007] Compared with existing technologies, this application provides a closed-loop control method and system for thin film production based on defect feedback, addressing the problems of inaccurate feedback caused by large spatiotemporal lags and difficulties in tracing the source of defects due to heterogeneous multi-source data in thin film production. This scheme calculates the absolute physical coordinates of the thin film by real-time acquisition of the spindle position encoding signal, forcibly mapping the time-dimensional process parameter stream and the spatial-dimensional detection image stream to this physical coordinate system, eliminating the spatiotemporal barriers between data streams. When a deep neural network detects a defect, the system uses the defect's physical coordinates to perform precise spatiotemporal backtracking in the process state map, retrieving the actual historical process parameters that occurred when that section of the thin film passed through the upstream processing area, thereby constructing an accurate defect-historical parameter causal relationship data pair. Based on this determined causal relationship, the optimal compensation value is calculated using a sensitivity matrix model, and feedforward adjustment is implemented on the actuator, ensuring that the control command targets the actual process fluctuations that caused the defect rather than the current state, effectively overcoming control oscillations in the lag system and achieving high-precision closed-loop control. Attached Figure Description

[0008] The above and other objects, features and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings.

[0009] Figure 1 This is a flowchart of a closed-loop control method for thin film production process based on defect feedback according to an embodiment of this application.

[0010] Figure 2 This is a schematic diagram of data flow in a closed-loop control method for thin film production based on defect feedback, according to an embodiment of this application.

[0011] Figure 3 This is a flowchart of step 3 in the closed-loop control method for thin film production process based on defect feedback according to an embodiment of this application.

[0012] Figure 4 This is a block diagram of a closed-loop control system for thin film production based on defect feedback, according to an embodiment of this application. Detailed Implementation

[0013] The embodiments of this application will now be described in more detail with reference to the accompanying drawings. It should be understood that the drawings and embodiments of this application are for illustrative purposes only and are not intended to limit the scope of protection of this application.

[0014] In view of the shortcomings in the above-mentioned technical fields, this application proposes a closed-loop control method for thin film production process based on defect feedback. Figure 1 This is a flowchart of a closed-loop control method for thin film production process based on defect feedback according to an embodiment of this application. Figure 2This is a schematic diagram of data flow in a closed-loop control method for thin film production based on defect feedback, according to an embodiment of this application. Figure 1 and Figure 2 As shown, the closed-loop control method for thin film production process based on defect feedback according to an embodiment of this application includes: Step 1: Calculating the absolute physical coordinates of the thin film in real time using the position encoding signal from the main shaft of the production line, and mapping and binding the synchronously acquired process parameter stream and detection image stream to the absolute physical coordinates to obtain a process state map and a positioning image stream; Step 2: Inputting the positioning image stream into a preset deep neural network model for defect feature extraction and vectorization to obtain a defect feature vector; Step 3: Based on the defect physical position coordinates in the defect feature vector, performing spatiotemporal backtracking and parameter association based on position index in the process state map to obtain historical process parameters, and combining the historical process parameters with the defect feature vector to obtain a causal correlation data pair; Step 4: Inputting the causal correlation data pair into a preset sensitivity matrix model for multi-parameter sensitivity analysis and compensation value calculation to obtain a control correction command; Step 5: Superimposing the control correction command onto the current production baseline setting value, and driving the production line controller to perform feedforward compensation adjustment on the process actuator.

[0015] In step 1, the absolute physical coordinates of the film are calculated in real time using the position encoding signal from the production line spindle. The synchronously acquired process parameter stream and detection image stream are then mapped and bound to the absolute physical coordinates to obtain a process state map and a positioning image stream. It should be understood that in the complex industrial scenario of roll-to-roll film production, due to the extremely long production line span and multiple continuous processes, the film's transmission speed is not constant due to fluctuations in winding tension, mechanical slippage, and slight differences in the rotational speed of each roller. This nonlinear speed fluctuation means that data recording methods based solely on timestamps cannot accurately reflect the film's true physical position, making it difficult to precisely correlate downstream defects with the upstream process state at the time the defect occurred. Traditional time alignment methods accumulate errors when speed changes, leading to misalignment in causal analysis. To overcome this problem of spatiotemporal lag and data heterogeneity, a rigid reference system independent of time needs to be constructed. This system uses the absolute physical length of the film itself as a benchmark, forcibly unifying discrete visual information and continuous process parameters into the same spatial coordinate system. This eliminates interference caused by transmission speed fluctuations, ensuring that each frame of image and each process data point can be accurately anchored to a specific micro-element of the thin film, thus providing a solid and definite spatiotemporal reference for subsequent defect tracing and closed-loop compensation.

[0016] In one embodiment, step 1 includes: step 11, performing absolute physical coordinate calculation based on pulse integral on the position encoding signal to obtain the absolute physical coordinates of the thin film at the current time; step 12, using the absolute physical coordinates as the spatial primary key, performing spatiotemporal alignment and linear interpolation resampling on the process parameter stream to obtain a process state map; step 13, reading the acquisition trigger time of each frame of the detection image stream, matching the corresponding absolute physical coordinates, and writing them into the image metadata to obtain the positioning image stream.

[0017] The specific processing is as follows: First, implement step 11: In the physical architecture of the film production line, the main conveyor roller, as the core component driving the film movement, directly maps the film's transport progress to its rotational state. In this embodiment, a high-precision incremental rotary encoder is rigidly connected to the shaft end of the main conveyor roller. This encoder, acting as a sensor for physical position perception, outputs a high-frequency pulse sequence in real time during the roller's rotation, which is the position encoding signal. First, this pulse sequence is received, and a small time sampling window, such as 1 millisecond or shorter, is set. The pulse increments received within this time window are counted. To convert the discrete number of pulses into physically meaningful motion parameters, calculations are performed based on the geometric characteristics of the main conveyor roller. Let the encoder resolution be the number of pulses per revolution. The physical diameter of the main conveyor roller is At any given time If the measured encoder pulse frequency is This refers to the number of pulses per unit time. At this point, the angular velocity of the roller and the real-time linear velocity of the film are... A definite geometric relationship exists. Since the film adheres tightly to the roller surface without slippage during transmission, the linear velocity can be expressed as the ratio of the product of the pulse frequency and the roller circumference to the resolution. To obtain the absolute position of the film relative to the starting point, i.e., its absolute physical coordinates... The real-time linear velocity is calculated cumulatively using the time integration method. This coordinate... This represents the cumulative length of the film passing through the main conveyor roller (or the reference point below the deposition source) from the start of production (t=0) to the current time t. Its mathematical expression is shown in the following formula:

[0018] This represents the current rotational speed of the roller (revolutions per second), multiplied by... The (roller circumference) gives the instantaneous velocity of the film at the current moment. By integrating this instantaneous velocity over the time domain [0, t], the cumulative physical distance traveled by the film can be accurately calculated. In a digital control system, this integration process is implemented using an accumulator, which calculates the displacement increment in each sampling period. And add it to the total length register. For example, the diameter of the main conveyor roller on the production line. =0.5 meters, encoder resolution =10000 pulses / revolution. At a certain moment t during the production process, the pulse frequency output by the encoder is measured. =2000Hz. At this time, the instantaneous linear velocity of the thin film is calculated as: =2000 / 10000×3.14159×0.5≈0.314m / s. Specifically, here there is no relative slippage between the film and the roller under tension control, or this has been corrected using a slip coefficient. If the cumulative coordinates at the previous moment were 1000.0 meters and the sampling period was 0.01 seconds, then the displacement increment within that period would be 0.314×0.01=0.00314 meters. The updated absolute physical coordinates... That is, 1000.00314 meters.

[0019] Secondly, implement step 12: after obtaining high-precision absolute physical coordinates The next challenge is the heterogeneity of multi-source data. Process parameters, such as vacuum chamber temperature, reactive gas flow rate, and sputtering target voltage, are sampled by PLCs or dedicated sensors at fixed time frequencies, such as 1Hz or 10Hz. However, due to the thin film linear velocity... The data fluctuates, resulting in a non-uniform spatial distribution of data points collected at equal time intervals. For example, at higher speeds, the distance between two sampling points is longer; at lower speeds, the distance is shorter. This characteristic of uniform data in the time domain but non-uniform data in the spatial domain severely hinders accurate location-based analysis. To address this issue, this step uses the timestamp t as a bridge to map the original data in the time domain to the spatial domain and resamples it. First, a real-time buffer is maintained, which records each sampling time. Corresponding original process parameter values and the absolute physical coordinates at that moment The result is calculated in step 11. Subsequently, a standardized spatial grid is constructed, with a fixed spatial step size, such as one data point per 1 mm. This is to obtain data at any location on this spatial grid. process parameter values ​​at the location Find the physical coordinates in the buffer. The two original sampling points before and after, i.e. The parameter values ​​for the target location are calculated using a linear interpolation algorithm. In this way, the originally discrete and unevenly distributed time-series data is reconstructed using absolute physical coordinates. A hash map, i.e., a process state map, is constructed using the reconstructed set of process parameters as values ​​and the index key. For example, to construct a process map with a spatial step size of 1 mm, the original gas flow sensor data is as follows: at time... absolute coordinates =50.000 meters, flow rate =100sccm; at time... absolute coordinates =50.005 meters, flow rate =105 sccm. Now we need to calculate the gas flow rate at the absolute coordinate L=50.002 meters. According to the formula: =100+(105-100) / (50.005-50.000)×(50.002-50.000)=102sccm. This calculation fills the data gaps in the sensor sampling gaps and generates a spatially continuous and uniform description of the process state.

[0020] Finally, step 13 is implemented: The Optical Inspection (AOI) system, located at the end of the production line, is used to capture microscopic images of the film surface. To correlate these images with the aforementioned process status map, each frame of the image needs to be assigned specific spatial coordinates. The AOI camera generates a precise hardware trigger pulse or records a timestamp indicating completion of image acquisition when acquiring an image. The camera's image data stream is monitored in real time. Whenever a new frame of image data is received, the trigger moment of that frame's acquisition is immediately read. Using the time-coordinate mapping relationship maintained in step 11, the absolute physical coordinates of the film at that moment are found or interpolated. It should be noted that there is a fixed physical distance between the camera installation location and the process execution mechanism (such as the deposition source), and the absolute physical coordinates are defined as the length of the film relative to the roll head. For ease of subsequent analysis, the coordinates written here refer to the absolute position of the film micro-element corresponding to the center of the image field of view within the entire roll of film, and then the segment of film captured in the current image is marked. The calculated absolute physical coordinates are serialized and encoded, and written as extended metadata into the header of the image file or an attached metadata file. In this way, each frame is no longer just a pixel matrix, but a localization image stream carrying the key spatial attribute of being at position X meters within the entire roll of thin film. For example, an AOI camera is installed 50 meters downstream of the deposition source. When the camera triggers a capture at a certain moment, the absolute physical coordinates of the thin film below the camera are calculated to be 200.5 meters. This information, "Position: 200.5m," is written into the EXIF ​​data segment of that frame or a dedicated XML description file.

[0021] In step 2, the localized image stream is input into a pre-set deep neural network model for defect feature extraction and vectorization to obtain a defect feature vector. Correspondingly, in the high-speed production of thin film materials, the raw data collected by automated optical inspection equipment is essentially a massive unstructured pixel matrix. Although these image data have been assigned physical coordinate labels through previous steps, the defect information (such as shape, category, and size) contained within them still exists in the form of implicit high-dimensional features, which cannot be directly understood by the logic controller or used for parameter inversion. To achieve precise closed-loop control, this application converts these intuitive visual representations into computer-computable and quantifiable mathematical descriptions. By introducing deep learning algorithms to perform real-time inference on the image stream, tiny defect targets are accurately separated from complex background textures, and their potential process hazards are quantitatively assessed, thus compressing the massive image data stream into a compact and semantically clear feature vector.

[0022] In one embodiment, step 2 includes: step 21, performing metadata parsing and normalization on the localized image stream to separate the absolute physical coordinates and generate a tensor of the image to be detected; step 22, inputting the tensor of the image to be detected into a deep convolutional neural network for pixel-by-pixel semantic segmentation inference to obtain a defect segmentation mask and a defect category label; step 23, performing morphological calculations on the defect segmentation mask to assess the severity of the defect; and step 24, serializing and encapsulating the defect severity, defect category label, and absolute physical coordinates to obtain a defect feature vector.

[0023] The specific processing is as follows: First, implement step 21: Continuing from the positioning image stream generated in the previous step, first read the header information of the data frame. Since the metadata area of ​​each frame image has been written with the absolute physical coordinates calculated based on the spindle encoder in the previous step, parse the metadata and extract the physical position index corresponding to the current image frame, such as 200.5 meters. This coordinate value is temporarily stored in the associated cache area as the spatial primary key for subsequent assembly of feature vectors. At the same time, separate the pure pixel data matrix of the image. Considering that modern production lines mostly use wide-format thin films (up to several meters wide), and the imaging resolution of a single camera is extremely high, such as 4096×1024 pixels, directly inputting such a high-dimensional image into the neural network will cause memory overflow and excessive computational delay. Therefore, a sliding window cropping strategy is adopted, using a fixed-size sliding window, such as 512×512 pixels, to move on the original image at a set step size, dividing it into a series of smaller image patches. For each image patch, in order to accelerate the convergence of the neural network and eliminate the interference caused by differences in illumination intensity, normalization processing is required. Let the grayscale value of a certain pixel in the original image be... The bit depth is 8 bits (0-255), which is mapped to the floating-point range of [0,1] and then normalized. The calculation formula is as follows: ,in, and These are the pixel mean and standard deviation of the training dataset, respectively. The normalized data is reorganized into a tensor format (N, C, H, W) of the image to be detected that meets the input requirements of the deep learning model, where N is the batch size, C is the number of channels (single-channel grayscale or three-channel RGB), and H and W are the height and width of the image.

[0024] Secondly, step 22: This embodiment uses a deep fully convolutional neural network based on the U-Net architecture as the core inference model. The specific architecture of this model consists of an encoder and a decoder. The encoder contains multiple consecutive convolutional layers and max-pooling layers, used to extract abstract features of the image layer by layer and reduce the spatial dimension, thereby capturing the contextual semantic information of the defects. The decoder gradually restores the spatial resolution of the image through upsampling and transposed convolution operations. Specifically, the U-Net architecture establishes skip connections between corresponding layers of the encoder and decoder, directly stitching the high-resolution features of the shallow layers onto the semantic features of the deep layers, thus ensuring that the model can maintain accurate edge localization capabilities when segmenting tiny defects (such as pinholes with a diameter of micrometers). The weight parameters and biases in this model are predetermined through offline supervised learning. During the training phase, a large number of historical thin-film images labeled with defective regions are collected. The difference between the model's prediction results and the true labels is measured using the cross-entropy loss function or the Dice coefficient loss function. The network parameters are iteratively updated using the backpropagation algorithm and gradient descent optimizer until the model can accurately distinguish between the background and the defects. In real-time processing, the tensor of the image to be detected is input into the trained network. The network output layer generates a probability map of the same size as the input image using either a Softmax or Sigmoid activation function. The value of each pixel in the map represents the probability that the point belongs to a certain defect category (such as pinholes, scratches, or uneven coating). A confidence threshold, such as 0.5, is set, and pixels in the probability map with a value greater than this threshold are set to 1, while the rest are set to 0, thus generating a binarized defect segmentation mask. Simultaneously, the network outputs a defect category label to which the defect region belongs.

[0025] Next, step 23 is implemented: after obtaining the binarized defect segmentation mask, the connected component analysis algorithm is used to count the regions with a pixel value of 1 in the mask. For each independent defect connected component, its geometric features are calculated, mainly including the defect area. That is, the total number of pixels and the equivalent diameter within the connected region. The equivalent diameter is defined as the diameter of a circle with the same area as the defect, and the calculation formula is: To quantify the impact of defects on the final product performance, a comprehensive index, defect severity, is introduced. The calculation of this index relies on a pre-defined weighting coefficient system. The weighting coefficients corresponding to the defect category labels are as follows: This is an empirical value set based on a knowledge base of process experts and historical quality data. This coefficient reflects the destructive impact of different types of defects on product function. For example, in lithium-ion battery separator production, pinhole defects can cause short circuits, which are considered fatal defects, and their weighting coefficient might be set as follows: =1.0; while minor scratches only affect appearance and not insulation performance, their weighting factor may be set to 1.0; =0.2. Based on the above parameters, the severity of the defect is... The calculation model is as follows: ,in, The physical resolution calibration factor is a constant determined by calculating the ratio of the physical length of a calibration board to the corresponding number of pixels in the image using a standard calibration board of known size imaged within the camera's field of view. It is used to convert pixel dimensions to actual physical dimensions. For example, if a pinhole defect is identified at 200.5 meters in the previous step, the segmentation mask output by the model shows that the defect area is 80 pixels. Calculate the equivalent diameter (pixels): ≈10.1 pixels, such as the camera's image resolution. =0.01mm / pixel, pinhole weighting coefficient =1.0. Therefore, the severity score for this defect is: =0.01×10.1×1.0=0.101, which intuitively represents the physical scale of the defect and its potential hazards.

[0026] Finally, step 24 is implemented: To facilitate subsequent calls to the control algorithm, all the key information extracted in the above steps is structured and packaged. This process is similar to building a standard data packet. The absolute physical coordinates extracted in step 21, the defect category labels identified in step 22, and the defect severity calculated in step 23 are arranged in a predetermined protocol order to generate a defect feature vector. Continuing the above example, the pinhole defect found at an absolute location of 200.5 meters is finally encapsulated into the following vector: [200.500, {"Pinhole"}, 0.101, 1634527890], where 1634527890 is auxiliary timestamp information.

[0027] In step 3, based on the physical location coordinates of the defects in the defect feature vector, spatiotemporal backtracking and parameter correlation based on location index are performed in the process state map to obtain historical process parameters. These historical process parameters are then combined with the defect feature vector to obtain causal correlation data pairs. It is understandable that in the complex industrial scenario of roll-to-roll film production, the production process is essentially a distributed parameter system with large hysteresis. When an automated optical inspection device at the end of the production line captures a microscopic defect on the film surface, the process anomaly that caused the defect often occurred at a certain historical moment in the upstream process. Due to fluctuations in transmission speed and the existence of mechanical slippage, backtracking based solely on timestamps cannot accurately pinpoint the actual process occurrence point. Furthermore, the film formation quality is governed by thermodynamic inertia and fluid diffusion effects. The physical properties of a micro-element are not solely determined by a single process parameter at the instant the point passes through the processing area, but exhibit significant spatial neighborhood correlation. That is, the heat accumulation or material concentration gradient in the small regions before and after the point jointly shapes the final surface morphology. If we ignore this spatiotemporal misalignment and the neighborhood effect of the physical field, and directly use the process parameters collected at the current moment for feedback adjustment, not only will the deviation not be eliminated, but false causal relationships will also be introduced, causing the control algorithm to oscillate or even diverge. Therefore, this application introduces step 3.

[0028] Figure 3 This is a flowchart of step 3 in the closed-loop control method for thin film production based on defect feedback according to an embodiment of this application. Figure 3 As shown, in one embodiment, step 3 includes: step 31, defining a spatial influence window centered on the physical location coordinates of the defect in the defect feature vector, and retrieving all records falling within the spatial influence window in the process state map to obtain historical process slices; step 32, performing weighted spatiotemporal aggregation and feature dimensionality reduction on the historical process slices based on the Gaussian kernel function to obtain historical process parameters; step 33, logically binding and structurally encapsulating the historical process parameters and defect feature vectors to obtain causal data pairs.

[0029] The specific processing is as follows: First, implement step 31: Receive the defect feature vector and extract the core index key—the physical location coordinates of the defect—from it, denoted as... In continuous roll-to-roll manufacturing processes, defect formation is often not an isolated geometric point event, but rather a field event with physical extent. For example, in vacuum magnetron sputtering, the plasma glow discharge on the target surface has a certain spatial distribution range, and the turbulent disturbance of the reactant gas also has a diffusion radius; in precision coating processes, the viscoelasticity of the fluid causes pressure or thickness changes at a certain point to propagate forward and backward along the coating direction. Therefore, in order to comprehensively capture the effective process background leading to defect formation, the processing unit must define a spatial influence window. The half-width of this window... This is a key process physics parameter, and its value is not arbitrarily set, but rather predetermined based on the process response characteristics and transport dynamics model of the production line. Specifically, process engineers will determine it based on the gas diffusion coefficient within the reaction chamber. Thermal conduction time constant and the average running linear velocity of the thin film The effective distance of the process disturbance on the thin film is calculated. The calculation model is expressed as follows: ,in To account for the safety redundancy factor of boundary effects, a value of 1.5 to 2.0 is used. For example, if the characteristic response time of heat conduction is 0.05 seconds and the average linear velocity is 0.5 m / s, then the physical influence radius is approximately 25 mm, and this is used to set... =25mm. The window range has been determined. Next, the process state map stored in high-speed memory is accessed. This map has been constructed in previous steps as a hash map or B-tree index structure with absolute physical coordinates as keys. Using an efficient range query algorithm, a discretized traversal is performed within a set spatial step size, such as 1 mm, to retrieve all process parameter records falling within this interval. These records are arranged in order of physical spatial location, forming a sequence fragment containing multi-dimensional process parameters (such as deposition temperature, sputtering power, gas flow rate, etc.), i.e., a historical process slice. Continuing with the aforementioned embodiment, the coordinates in the defect feature vector... =200.500 meters, the set space affects half the window width =2 mm (for simplified calculation example), spatial sampling step size is 1 mm. The processing unit will search within the interval [200.498, 200.502]. The retrieved deposition temperature channel data sequence is: L=200.498m:149.8°C; L=200.499m:150.0°C; L=200.500m:152.0°C (time corresponding to the defect center); L=200.501m:150.5°C; L=200.502m:149.9°C. This set of continuous data points constitutes a historical process slice of the defect in the temperature dimension, which completely records the thermal experience of the defect area as it passed through the deposition source.

[0030] Next, in step 32: After obtaining the historical process slices, the problem is how to extract the most representative value from this set of sequence data to characterize the actual process conditions that led to the defect. A simple arithmetic mean would give the edge data of the window the same importance as the center data, which violates the objective laws of physics—obviously, the closer the process fluctuation is to the defect center, the greater its contribution to defect formation and the stronger its correlation. Therefore, this embodiment introduces a Gaussian kernel function model to weight the slice data. The Gaussian kernel function acts as a low-pass filter in the spatial domain, effectively filtering out high-frequency sensor measurement noise while retaining the low-frequency trend components that dominate defect formation. The Gaussian kernel function is a weighting function with a bell-shaped curve distribution, and its mathematical expression is:

[0031] in, The physical coordinates of each sampling point in the slice The coordinates of the defect center are... The standard deviation parameter is used to control the central tendency of the weight distribution. In implementation, The value is set to the same as the window half width. A linear relationship, for example Based on the properties of the Gaussian distribution, this means that approximately 99.7% of the weight energy is concentrated within the defined window range, thus ensuring the physical rationality of the weighting. First, iterate through each coordinate point in the slice. Substitute into the above formula to calculate the corresponding original weight value. To ensure that the dimensions of the aggregated parameters remain unchanged, the weights need to be normalized so that the sum of all weights is 1; that is, the normalized weights are calculated. Subsequently, normalized weights were used to analyze the parameter values ​​in the process slices. The weighted summation is calculated using the following formula: Through this process, the originally spatially dimensional sequence data is reduced in dimension and aggregated into historical process parameters for a single point. These parameters not only reflect the process values ​​at the defect center but also incorporate the process trends of the surrounding neighborhood through a weighted approach, thus more robustly characterizing the physical, thermodynamic, or fluid dynamic state at that location. Continuing with the example above, taking the defect center... Using 200.500 as the baseline, the weights of each point are calculated. Since the exponent term at the center point 200.500 is 0, its original weight is the largest; the weights of the edge points 200.498 and 200.502 decrease exponentially due to their greater distance from the center. After normalization, the weight vectors for the five points are [0.05, 0.20, 0.50, 0.20, 0.05]. The historical deposition temperature after aggregation is then: =0.05×149.8+0.20×150.0+0.50×152.0+0.20×150.5+0.05×149.9=151.085°C. This calculated effective temperature parameter, which is considered to have caused the pinhole defect at that location, reflects the true level of local thermal energy better than simply taking 152.0°C at the center point or a simple average value.

[0032] Finally, step 33: After obtaining the aggregated historical process parameter set, a data cleaning procedure is first executed to detect outliers. Using the 3-Sigma criterion or preset hardware physical range thresholds (e.g., deposition temperature cannot exceed 200°C, flow rate cannot be negative), outliers caused by momentary sensor malfunctions are removed to ensure the data entering the control loop is authentic and reliable, avoiding contamination of the subsequent sensitivity model by dirty data. Subsequently, a logical binding operation is performed to construct a defined causal relationship. Here, the cause is the historical process parameter vector obtained through spatiotemporal backtracking and weighted aggregation (containing multi-dimensional data such as temperature, flow rate, and voltage), and the effect is the defect feature vector extracted in the previous steps (containing defect category and severity). A structured data object containing an input domain and a target domain is created. The input domain is filled with the aggregated historical process parameters, and the target domain is filled with defect feature descriptions. This process essentially constructs a high-confidence supervised learning sample in real time on the production line. The final generated causal relationship data pairs are encapsulated in a standard tensor format or JSON object and stored in a first-in-first-out training buffer, awaiting retrieval by the downstream sensitivity analysis model. For example, the final packaged data pair might look like this: Input (historical process): [Temperature=151.085,Flow=50.2,Voltage=380.5] Target (defect characteristic): [Type="Pinhole",Severity=0.101]. This data pair clearly establishes that a temperature of 151.085°C combined with a flow rate of 50.2 sccm and a voltage of 380.5V resulted in a pinhole defect with a severity of 0.101.

[0033] In a preferred embodiment, step 3 includes: step 3-1, defining a spatial influence window centered on the physical location coordinates of the defect in the defect feature vector, and retrieving all records falling within the spatial influence window from the process state map to obtain historical process slices; step 3-2, performing dynamic weight reconstruction on the historical process slices based on a velocity-dependent asymmetric drift-diffusion kernel function to obtain historical process parameters; step 3-3, logically binding and structurally encapsulating the historical process parameters with the defect feature vector to obtain causal relationship data pairs. Here, the implementation of steps 3-1 and 3-3 is consistent with steps 31 and 33 above, so it will not be repeated. The implementation of step 3-2 will now be described in detail.

[0034] It is understandable that in high-precision thin film production processes such as magnetron sputtering or chemical vapor deposition (CVD), the physical fields (such as temperature and plasma density fields) exhibit significant hysteresis effects and fluid transport delays due to the complex plasma reactions or thermochemical equilibrium involved in the process environment. When the production line is in a high-speed winding state, the fluctuations in process parameters captured by the sensors do not act on the thin film surface instantaneously and symmetrically, but rather produce obvious projection tailing or forward shift phenomena as the thin film moves rapidly. If a standard symmetric Gaussian kernel function is used for data aggregation, its default model of the influence of process parameters on the thin film is a static model that decays uniformly from the geometric center outwards. This ignores the dynamic interference of belt speed fluctuations on the physical field distribution, which can easily lead to causal mismatch between downstream defects and upstream process parameters. Since there is an inherent response time between the sensor reading and the physical quantity actually acting on the thin film surface, the thin film has already moved forward by a certain displacement within this time, resulting in significant physical drift. To eliminate the attribution bias caused by system lag and high-speed motion, this application constructs an asymmetric kernel function that can dynamically adjust the center position and distribution width according to the belt speed, which is of vital importance for accurately reconstructing the real process background that leads to defects.

[0035] Based on this, in one embodiment, step 3-2 involves performing dynamic weight reconstruction on the historical process slices based on a velocity-dependent asymmetric drift-diffusion kernel function to obtain historical process parameters, including: The dynamic weights of each sampling point are obtained by weighting them using an asymmetric drift-diffusion kernel function. In this step, a weighting function capable of simulating convection-diffusion behavior under high-speed winding is constructed by retrieving dynamic variables from the real-time process parameter stream. The mathematical expression for the asymmetric drift-diffusion kernel function is as follows:

[0036] In this calculation formula, the input variables include the physical location coordinates of the sampling points. Physical location coordinates of the defect center and the average belt speed of the thin film within that time slice window The physical meaning and logical definition of each symbol in the formula are as follows: This is the drift direction coefficient, used to indicate the geometric relationship between the process flow direction and the conveyor belt direction. It takes a positive value, such as 1, under the condition of downstream operation, and a negative value under the condition of upstream operation. It is a pre-calibrated system response time constant, representing the inherent delay from the occurrence of fluctuations in process parameters to the complete action of the physical field on the film surface, measured based on thermal inertia or fluid transport experiments, for example, set to 0.05 seconds; The square of the reference influence radius (i.e., the initial diffusion coefficient) under static conditions reflects the physical coverage of the process field itself, for example, by setting... =0.003 meters; This is a velocity sensitivity factor used to describe the stretching effect of high band velocities on the diffusion range of the physical field. It is preset based on fluid boundary layer theory, for example, set to 0.1. The numerator of this formula contains... By introducing a dynamic offset term, the peak center of the weighting function is shifted from a purely geometric location. It moved to the physical center of action, thus compensating for the spatial displacement caused by the response lag; in the denominator term This achieves nonlinear scaling of the diffusion range with velocity, simulating the stretching effect of high-speed motion on the flow field distribution. This generates an asymmetric weight distribution that highly matches the real-time dynamic state of the production line, ensuring the physical accuracy of subsequent aggregation calculations.

[0037] Historical process parameters are obtained by aggregating parameters from historical process slices based on the dynamic weights of each sampling point. Then, the discrete historical process slices are compressed into a single feature value through weighted convolution, and the mathematical calculation formula is as follows:

[0038] The input to this formula is the original process parameter values ​​of each sampling point in the process state diagram. and the dynamic weights calculated in the above steps The formula outputs the aggregated value of the true causal parameters after kinetic correction by multiplying the process value of each sampling point with its corresponding asymmetric weight, summing the results, and then dividing by the sum of all weights (i.e., the normalization factor). In other words, the generation of thin film defects is usually the result of the combined effect of local process fields. Using asymmetric kernel functions for polymerization can effectively filter out irrelevant random noise and retain the most physically relevant feature components.

[0039] Specifically, such as in absolute physical coordinates A pinhole defect was detected at a distance of 200.500 meters, at which point the average belt speed of the film was... =0.6 m / s. Preset system response time. =0.05 seconds, drift direction coefficient =1. First, calculate the spatial offset. = Meters. This means that the process parameter point most affected by this defect is actually located near the sampling record at 200.500 + 0.030 = 200.530 meters. Next, the diffusion radius after velocity stretching is calculated: If =0.003 meters, =0.1, then the dynamic diffusion width Rice. Subsequently, historical process slices were retrieved. Parameter records distributed within a 200.530-meter neighborhood (such as a set of deposition temperature values) were analyzed using a pre-constructed asymmetric kernel function. Calculate the weights for each point. Due to the introduction of an offset term, the parameter located at 200.530 meters will receive the maximum weight. Finally, obtain the final historical process parameters through normalized summation. For example, 155.2℃. This aggregated value can accurately trace back the true physical causes under high-speed motion, providing high-quality causal correlation data pairs for subsequent sensitivity matrix models, and significantly improving the robustness of closed-loop control under all speed range conditions.

[0040] In step 4, the causal correlation data is used to perform multi-parameter sensitivity analysis and compensation value calculation on the preset sensitivity matrix model to obtain control correction commands. It should be understood that after obtaining accurate causal correlation data pairs, the core challenge of closed-loop control shifts to how to deduce the optimal process compensation scheme based on known quality deviations. Thin-film production is a multivariate coupled process involving thermodynamics, fluid mechanics, and plasma physics. Every defect on the production line is often a comprehensive manifestation of multiple process parameters deviating from the ideal operating point. Traditional independent-loop PID control struggles to handle such complex cross-influences, easily leading to conflicting parameter adjustments or systemic overshoot. This step aims to construct a sensitivity matrix model to quantitatively characterize the differential contribution rate of each process variable to a specific defect characteristic, thereby finding the optimal compensation vector in the multidimensional parameter space that can offset the defect. This mathematical analytical method based on the Jacobian matrix can coordinate the linkage between multiple parameters, transforming the originally qualitative quality feedback into quantitative, hardware-constrained, and precise control commands, ensuring the scientific nature of the adjustment actions and system stability.

[0041] In one embodiment, step 4 includes: step 41, extracting measured defect features from causal data pairs and calculating the difference between them and the ideal feature vector to obtain the target feature change; step 42, based on the sensitivity matrix model, and using the regularized least squares method with damping factor, performing inverse parameter solution on the target feature change to obtain the optimal process parameter correction; step 43, performing hardware safety threshold verification and amplitude truncation on the optimal process parameter correction to obtain the control correction command.

[0042] The specific processing is as follows: First, implement step 41: Receive the causal correlation data pairs generated in step 3, and extract the measured defect feature vector representing the current quality state. This vector contains semantic information such as the severity score and category label of the defect at a specific location. To determine the adjustment target, simultaneously construct an ideal feature vector with dimensions completely consistent with the measured feature vector. Under the vision of good product manufacturing, the ideal state is defined as having no physical defects on the film surface, i.e., a severity score of 0, and the code corresponding to the category label is also the baseline value. Perform vector subtraction to calculate the difference vector between the ideal state and the measured state, generating the target feature change. The target feature change represents the amount of suppression that must be achieved in the feature space to reach a zero-defect state. Continuing the aforementioned embodiment, if the measured defect severity output in step 3... =0.101 corresponds to a pinhole defect at a physical location of 200.5 meters, while the ideal severity is 0, so the change in the target feature is -0.101. This negative value clearly indicates the physical orientation of the control algorithm, that is, the defect severity needs to be reduced by 0.101 units through process compensation. If the feature vector contains multiple dimensions (such as thickness deviation, roughness deviation, etc.), then the change in the target feature is a column vector containing multiple components.

[0043] Secondly, step 42: The core of this step lies in using a sensitivity matrix model to guide parameter adjustment. Mathematically, this model is a Jacobian matrix. Its structure is defined as a matrix of partial derivatives of the defect feature vector with respect to the process parameter vector. Each row of the matrix corresponds to a defect feature component, and each column corresponds to a process parameter variable such as temperature, flow rate, or pressure. The elements in the matrix... This describes the rate of change of the i-th defect feature when the j-th process parameter undergoes a unit change. The sensitivity matrix model is obtained as follows: During production line debugging or offline simulation, a series of orthogonal experiments (DOEs) are designed to record defect change data under different process gradients. An initial sensitivity matrix is ​​then obtained through training using linear regression or partial least squares. The weight parameters stored within the model reflect the response gain of each process step. During real-time inference, the matrix is ​​fine-tuned online based on the current operating point (e.g., current base temperature, base flow rate) to ensure the accuracy of the linear approximation. However, in actual thin film production, process parameters are often highly coupled. For example, an increase in deposition temperature and a rise in gas flow rate can both lead to changes in film defects, which mathematically manifests as a Jacobian matrix. Multicollinearity or an ill-conditioned state (rank deficiency) exists. In this case, directly taking the generalized inverse of the matrix will lead to extremely sensitive calculated parameter corrections, where even small measurement noise can cause drastic jumps in the control quantity. To address this stability issue, a damping factor is introduced. Regularization is performed, and the optimal process parameter correction is solved using the damped least squares method (also known as Levenberg-Marquardt logic). The core calculation formula is as follows:

[0044] in, It is the change in the target feature. It is the transpose of the Jacobian matrix; Is with Identity matrices of the same order; It is a preset positive scalar, i.e., a damping factor, preset based on the condition number and singular value distribution of the Jacobian matrix, used to adjust the stability and convergence rate of the solution. Analysis of the formula shows that when... When the value is small, the solution tends to be a standard least squares solution, offering high adjustment accuracy but also high risk; when... When the value is large, the regularization term Dominant, it can effectively suppress singularities in the matrix inversion process, ensuring that the calculated correction vector has reliable direction and appropriate magnitude. For example, currently, for pinhole defects, the sensitivity matrix... The calculated parameter influence rates are: for every 1°C increase in temperature, severity increases by 0.04; for every 1 sccm increase in flow rate, severity increases by 0.01. Therefore, the Jacobian matrix is ​​a row vector [0.04, 0.01]. Target feature change. =-0.101. Set the damping factor. =0.005. Substitute into the formula to calculate:

[0045] Invert the matrix and multiply by and The final optimal process parameter correction amount is obtained. Includes temperature correction ≈-2.42℃ and flow correction ≈-0.19sccm. This set of values ​​represents the optimal parameter adjustment scheme that, from a theoretical mathematical perspective, can eliminate the current severity defect of 0.101.

[0046] Finally, implement step 43: Although the mathematical solution provides the optimal path, in a real industrial environment, the physical limits of the hardware must be considered. Hardware limits refer to the adjustment boundaries that production line actuators such as heaters, mass flow controllers (MFCs), and high-voltage power supplies cannot overcome due to physical structure, response inertia, and production process specifications. These limit parameters are provided in the equipment manual or pre-set in the control unit's configuration table by process experts, including the upper limit of the absolute adjustment value, the lower limit of the absolute adjustment value, and the upper limit of the single-step adjustment amount. Read these safety thresholds. Adjust the optimal process parameters calculated above accordingly. Execute the saturation cutoff logic. If the calculated temperature correction... The temperature drop to -2.42℃ exceeds the maximum allowable step size for the heater within the current control cycle, as set. If the temperature reaches 1.0℃, or exceeds the minimum temperature limit specified in the process window, the processing unit will call the cutoff function. Its mathematical logic is as follows:

[0047] This function ensures that instructions sent to the actuator are within a safe range. If the theoretical correction is within the step size, the original value is retained; if it exceeds the step size, the maximum step size is enforced. This limitation effectively prevents the system from resonating due to large-amplitude, high-frequency parameter jumps, protecting the mechanical life of expensive hardware such as heating wires and pumps, while also maintaining a quasi-steady state in the production environment. The parameter increments, after safety verification and amplitude truncation, are encapsulated into a PLC-recognizable industrial communication protocol (such as Modbus TCP or EtherCAT) message format, ultimately generating a control correction instruction. This instruction is no longer merely a mathematical inverse partial derivative solution, but a final control strategy integrating quality requirements, multivariate coupling relationships, and physical safety boundaries. For example, in the aforementioned case, the final generated instruction is... .

[0048] In step 5, the control correction command is superimposed on the current production baseline setpoint, driving the production line controller to perform feedforward compensation adjustment on the process actuators. That is, during the continuous operation of the thin film production line, the upstream actuators maintain a dynamic balance according to the preset process procedures. After precise spatiotemporal backtracking and sensitivity inverse kinematics calculations, yielding the process compensation amount aimed at eliminating specific defects, these theoretical incremental values ​​must be combined with the physical baseline in the real-time production environment. Thin film production is a dynamic and non-stationary process; any single parameter adjustment cannot be separated from the current operating context. This step is implemented to transform the discrete correction logic for past quality deviations into continuous feedforward actions that can act on future production micro-elements. By linearly superimposing the compensation command with the current setpoint, the control action can smoothly integrate into the current production rhythm, using the feedforward compensation mechanism to offset potential process fluctuations in advance, thereby closing the final link from defect perception to process self-healing at the physical level, achieving continuous online optimization of production quality.

[0049] In one embodiment, step 5 includes: step 51, obtaining the current process parameter stream as a reference setpoint, and superimposing the control correction command on it to obtain the target process setpoint; step 52, sending the target process setpoint to the process actuator to trigger a physical adjustment action.

[0050] The specific processing is as follows: First, implement step 51: Synchronize the current process parameter stream in real time from the underlying controller (such as an embedded PLC or industrial controller) of the running production line via an industrial Ethernet interface. In this embodiment, the reference setpoint... The setpoint (SV) is defined as the current setting value being executed by each process actuator, rather than the measured value (PV). The setpoint is used as the reference to avoid secondary oscillations in control commands caused by instantaneous noise interference from measured values. These setpoints are stored in the holding registers of the underlying controller and include core process parameters such as the target temperature of the heater, the target flow rate of the mass flow controller, and the power setpoint of the sputtering power supply. (The last sentence appears to be incomplete and requires further context.) Then, the control correction command output in step 43 is... Instead of performing algebraic superposition, the target process setpoint is calculated. The superposition operation follows the principle of linear addition, as shown in the following formula: Let's analyze this formula: This represents the steady-state operating point that maintains the current production status; This represents the disturbance compensation amount calculated to eliminate a specific defect. The sum of the two is... This constitutes the physical execution target for the next control cycle. Mathematically, this process maps the deviation space to the execution space, ensuring that the adjustment is a fine-tuning based on the current reference, rather than a blind full-scale overlay. Continuing with the aforementioned embodiment, the current heater reference temperature setpoint is read from the underlying controller. =150.5°C, the reference flow setpoint for the mass flow controller =50.0 sccm. The control correction command generated in step 43 is... The calculated target process setting value is: target temperature. =150.5 + (-1.0) = 149.5°C; Target flow rate =50.0+(-0.19)=49.81sccm. The calculated target process setting value is encapsulated into an industry-standard floating-point array and stored in the buffer to be sent out.

[0051] Secondly, in step 52: after generating the target process setpoint, the industrial communication module is invoked to send data packets to the process actuators distributed at various workstations on the production line via industrial fieldbuses such as EtherCAT, Profinet, or Modbus TCP. In this embodiment, the process actuator consists of execution hardware and its supporting underlying regulators. For example, the heating actuator includes an electric heating rod, a high-power thyristor regulator, and a temperature control instrument with a built-in PID algorithm; the flow actuator includes an electromagnetic proportional valve, a flow sensor, and supporting digital processing circuitry. The specific processing procedure is as follows: the communication module sends the target process setpoint according to a pre-configured hardware mapping table. The command is distributed to the corresponding hardware register address. For example, 149.5°C is written to address 40001 of the temperature controller, and 49.81 sccm is written to address 40005 of the flow controller. Once the register value of the underlying controller is updated, its built-in PID control program immediately senses the deviation between the new setpoint and the current measured value, thereby triggering physical adjustment actions. In the heating stage, the thyristor power regulator adjusts the output duty cycle, reducing the heating power to bring the temperature closer to 149.5°C; in the gas path stage, the mass flow controller changes the opening of the proportional valve, reducing the gas flow rate to 49.81 sccm. Since the thin-film production line is a large lag process, in order to achieve precise closed-loop control, the precise timestamp of the command trigger and the physical position corresponding to the current spindle encoder are recorded at the same time as the command is issued. This recording mechanism aims to mark from which meter of thin film the compensation action takes effect. Combined with the absolute physical coordinate system in step 1, a feedforward compensation record table can be established for subsequent verification of the actual effect of this adjustment action at the downstream detection end. The aforementioned process does not passively respond to defects after they occur, but rather utilizes past defect patterns detected downstream to calculate preventative adjustments for the film entering the processing zone. Through high-frequency synchronization of the industrial bus, command delays are controlled to the millisecond level, while the physical response of the hardware (such as temperature decreasing from 150.5°C to 149.5°C) follows its inherent time constant. Because damping factors and safety cutoffs are incorporated into the calculation process, the physical adjustment process exhibits a smooth exponential convergence curve, effectively avoiding overshoot and ensuring stable operation of the film production process at a new and optimal process equilibrium point.

[0052] In summary, a closed-loop control method for thin film production based on defect feedback, as described in this application, is presented to address the challenges of inaccurate feedback due to large spatiotemporal lags and difficulties in tracing the source of defects caused by heterogeneous multi-source data in thin film production. This scheme calculates the absolute physical coordinates of the thin film by real-time acquisition of the spindle position encoding signal, forcibly mapping the time-dimensional process parameter stream and the spatial-dimensional detection image stream to this physical coordinate system, eliminating the spatiotemporal barriers between data streams. When a deep neural network detects a defect, the system uses the defect's physical coordinates to perform precise spatiotemporal backtracking in the process state map, retrieving the actual historical process parameters that occurred when the thin film passed through the upstream processing area, thereby constructing an accurate defect-historical parameter causal relationship data pair. Based on this determined causal relationship, the optimal compensation value is calculated using a sensitivity matrix model, and feedforward adjustment is implemented on the actuator. This ensures that the control commands target the actual process fluctuations that caused the defect, rather than the current state, effectively overcoming control oscillations caused by lag and achieving high-precision closed-loop control.

[0053] Figure 4This is a block diagram of a closed-loop control system for a thin film production process based on defect feedback, according to an embodiment of this application. Figure 4 As shown, the closed-loop control system 100 for thin film production process based on defect feedback according to an embodiment of this application includes: a process parameter detection image acquisition module 110, used to calculate the absolute physical coordinates of the thin film in real time using the position encoding signal from the main shaft of the production line, and to map and bind the synchronously acquired process parameter stream and detection image stream to the absolute physical coordinates to obtain a process state map and a positioning image stream; a defect feature extraction and vectorization module 120, used to input the positioning image stream into a preset deep neural network model for defect feature extraction and vectorization processing to obtain a defect feature vector; and a causal association module 130, used for... Based on the physical location coordinates of the defect in the defect feature vector, spatiotemporal backtracking and parameter association based on the location index are performed in the process state map to obtain historical process parameters. The historical process parameters are then combined with the defect feature vector to obtain causal correlation data pairs. The control correction instruction generation module 140 is used to perform multi-parameter sensitivity analysis and compensation value calculation on the preset sensitivity matrix model input to the causal correlation data pairs to obtain control correction instructions. The feedforward compensation adjustment module 150 is used to superimpose the control correction instructions onto the current production baseline setting value and drive the production line controller to perform feedforward compensation adjustment on the process actuator.

[0054] Here, those skilled in the art will understand that the specific operations of each step in the above-described closed-loop control system for the thin film production process based on defect feedback have been referenced above. Figures 1 to 3 The description of the closed-loop control method for thin film production based on defect feedback has been detailed, and therefore, its repeated description will be omitted.

Claims

1. A closed-loop control method for thin film production process based on defect feedback, characterized in that, include: Step 1: Calculate the absolute physical coordinates of the thin film in real time using the position encoding signal from the production line spindle, and map and bind the synchronously acquired process parameter stream and detection image stream to the absolute physical coordinates to obtain the process status map and the positioning image stream; Step 2: Input the localization image stream into a preset deep neural network model to extract and vectorize defect features to obtain defect feature vectors; Step 3: Based on the physical location coordinates of the defects in the defect feature vector, perform spatiotemporal backtracking and parameter association based on location index in the process state map to obtain historical process parameters, and combine the historical process parameters with the defect feature vector to obtain causal relationship data pairs; Step 4: Perform multi-parameter sensitivity analysis and compensation value calculation on the input preset sensitivity matrix model using the causal correlation data to obtain control correction instructions; Step 5: Overlay the control correction command onto the current production baseline setting and drive the production line controller to perform feedforward compensation adjustment on the process actuator.

2. The closed-loop control method for thin film production process based on defect feedback according to claim 1, characterized in that, Step 1 includes: The absolute physical coordinates of the thin film at the current moment are obtained by performing pulse integral-based absolute physical coordinate calculation on the position-encoded signal. Based on absolute physical coordinates as the spatial primary key, spatiotemporal alignment and linear interpolation resampling are performed on the process parameter stream to obtain the process state map. Read the acquisition trigger time of each frame in the detection image stream, match the corresponding absolute physical coordinates and write them into the image metadata to obtain the positioning image stream.

3. The closed-loop control method for thin film production process based on defect feedback according to claim 1, characterized in that, Step 2 includes: Metadata parsing and normalization are performed on the localization image stream to separate the absolute physical coordinates and generate the tensor of the image to be detected; The tensor of the image to be detected is input into a deep convolutional neural network to perform pixel-by-pixel semantic segmentation inference to obtain the defect segmentation mask and defect category label; Morphological calculations are performed on the defect segmentation mask to assess the severity of the defect; The defect severity, defect category label, and absolute physical coordinates are serialized and encapsulated to obtain the defect feature vector.

4. The closed-loop control method for thin film production process based on defect feedback according to claim 1, characterized in that, Step 3 includes: A spatial influence window is defined with the physical location coordinates of the defect in the defect feature vector as the center. All records falling within the spatial influence window are retrieved in the process state map to obtain historical process slices. Historical process parameters are obtained by weighted spatiotemporal aggregation and feature dimensionality reduction of historical process slices based on Gaussian kernel function. Historical process parameters and defect feature vectors are logically bound and structured to obtain causal data pairs.

5. The closed-loop control method for thin film production process based on defect feedback according to claim 1, characterized in that, Step 3 includes: defining a spatial influence window centered on the physical location coordinates of the defect in the defect feature vector, and retrieving all records falling within the spatial influence window in the process state map to obtain historical process slices; Historical process slices are dynamically weighted based on a velocity-dependent asymmetric drift-diffusion kernel function to obtain historical process parameters. Historical process parameters and defect feature vectors are logically bound and structured to obtain causal data pairs.

6. The closed-loop control method for thin film production process based on defect feedback according to claim 5, characterized in that, Historical process slices are reconstructed using dynamic weights based on a velocity-dependent asymmetric drift-diffusion kernel function to obtain historical process parameters, including: The dynamic weight of each sampling point is obtained by calculating the weight of each sampling point based on the asymmetric drift-diffusion kernel function. Historical process parameters are obtained by aggregating parameters of historical process slices based on the dynamic weights of each sampling point.

7. The closed-loop control method for thin film production process based on defect feedback according to claim 1, characterized in that, Step 4 includes: The measured defect features are extracted from the causal correlation data pairs, and the difference between them and the ideal feature vector is calculated to obtain the change in the target feature. Based on the sensitivity matrix model, and using the regularized least squares method with damping factor, the parameter inverse solution of the target characteristic change is obtained to obtain the optimal process parameter correction. Hardware safety threshold verification and amplitude truncation are performed on the optimal process parameter correction amount to obtain control correction instructions.

8. The closed-loop control method for thin film production process based on defect feedback according to claim 1, characterized in that, Step 5 includes: The current process parameter stream is obtained as the baseline setpoint, and the control correction command is superimposed on it to obtain the target process setpoint. The target process setpoint is sent to the process actuator to trigger physical adjustment actions.

9. A closed-loop control system for thin film production process based on defect feedback, characterized in that, include: The process parameter detection image acquisition module is used to calculate the absolute physical coordinates of the film in real time using the position encoding signal from the production line spindle, and to map and bind the synchronously acquired process parameter stream and detection image stream to the absolute physical coordinates to obtain the process status map and the positioning image stream. The defect feature extraction and vectorization module is used to input the localization image stream into a preset deep neural network model to extract and vectorize defect features to obtain defect feature vectors. The causal association module is used to perform spatiotemporal backtracking and parameter association based on the physical location coordinates of defects in the defect feature vector in the process state map to obtain historical process parameters, and to combine the historical process parameters with the defect feature vector to obtain causal association data pairs. The control correction instruction generation module is used to perform multi-parameter sensitivity analysis and compensation value calculation on the input preset sensitivity matrix model based on causal correlation data to obtain control correction instructions; The feedforward compensation adjustment module is used to superimpose control correction commands onto the current production baseline setting and drive the production line controller to perform feedforward compensation adjustment on the process actuators.

Images