Target material monitoring method and system based on multi-source data and life prediction

Abstract

The invention discloses a target material monitoring method and system based on multi-source data and life prediction, and relates to the technical field of target material monitoring. Dependence of vulnerable physical reference hardware is eliminated through dynamic self-calibration, the maintenance cost is reduced, meanwhile, the long-term stability and reliability of the measurement process are built, and continuous and credible basic data are provided for follow-up analysis; the limitation of single sensing is broken through from the principle level by means of multi-sensor data fusion, comprehensive coverage of target material state information is achieved, the evaluation result is more accurate, the dimension is more complete, and the misjudgment risk of single data is avoided; passive measurement is upgraded into active prospective decision through intelligent analysis and life prediction, sufficient adjustment windows are reserved for production by means of life prediction, and finally the linkage effect of maximization of the target use efficiency, few interruption of the production process and precise optimization of technological parameters is achieved; and the intelligent level and the operation benefit of the production system are improved from the whole links of measurement, evaluation and decision making.

Description

Technical Field

[0001] This invention relates to the field of target monitoring technology, and in particular to a target monitoring method and system based on multi-source data and lifetime prediction. Background Technology

[0002] In precision manufacturing processes such as vacuum sputtering coating, the target material is one of the core consumables. Its surface morphology and wear level directly affect the stability of the film quality and production efficiency. Therefore, accurate monitoring and evaluation of the target material's condition has become a key technical link to ensure continuous and stable production. Currently, various target material condition measurement solutions have been developed in the industry, but due to limitations in technical approaches, there are still many shortcomings that need to be addressed.

[0003] The core shortcomings of existing technologies are concentrated in the dimensions of measurement reliability, information integrity, and intelligence: First, they rely on vulnerable physical references, often using high-precision physical reference blocks for calibration, which are susceptible to temperature drift and mechanical wear, requiring frequent maintenance, increasing operation and maintenance costs, and easily leading to measurement deviations due to reference inaccuracies; Second, they suffer from the limitation of a single data source, relying solely on single sensors such as laser ranging or 2D vision to collect data, failing to comprehensively cover the complex erosion morphology and surface texture of the target material, resulting in a one-sided condition assessment; Third, they adopt static measurement modes, measuring along fixed paths or at preset points, which lacks adaptability, making it difficult to match different target material types and diverse wear patterns, and limiting data representativeness; Fourth, they lack intelligent analysis capabilities, with measurements only outputting basic data, failing to deeply explore the remaining lifespan of the target material and the impact on the process, thus failing to release the value of the data; Fifth, they rely heavily on manual intervention, requiring experienced engineers to interpret data and make decisions, resulting in low automation and lagging decision-making efficiency; Sixth, they lack a closed-loop optimization mechanism, with measurement and decision-making processes being disconnected, making it impossible for the system to achieve continuous self-optimization and adapt to dynamic changes in operating conditions. Summary of the Invention

[0004] The purpose of this invention is to provide a target monitoring method and system based on multi-source data and lifetime prediction, so as to improve the technical problems of existing methods, such as reliance on vulnerable physical benchmarks, one-sidedness of single data sources, poor adaptability of static measurement, lack of intelligent analysis, excessive manual intervention, and lack of closed-loop optimization.

[0005] To achieve the above-mentioned objectives, the embodiments of the present invention provide the following technical solutions:

[0006] A target monitoring method based on multi-source data and lifetime prediction includes:

[0007] Calculate the pose transformation matrix of the coordinate system of multiple sensors, and combine it with the registration algorithm to calculate the error matrix, and construct a virtual reference formula;

[0008] The laser displacement data, three-dimensional data, and temperature distribution data of the target material are acquired and updated using a virtual reference formula. Combined with three-dimensional reconstruction and feature extraction, a composite three-dimensional topography map and a target material feature vector are generated.

[0009] A dual-branch membrane quality-lifetime prediction model is constructed based on CNN network, LSTM network and attention mechanism;

[0010] The process and quality data are acquired and input into the dual-branch membrane quality-lifetime prediction model along with the composite 3D morphology image and the target material feature vector to generate membrane quality-lifetime prediction results.

[0011] Based on the multi-objective optimization function and the membrane quality-lifetime prediction results, monitoring feedback results are generated through Bayesian optimization algorithm and adaptive decision-making.

[0012] In the above scheme, the core feature of this invention is to construct a virtual reference formula through multi-sensor pose transformation and registration error calculation, which gets rid of the dependence of existing technologies on vulnerable physical references and solves the problems of being susceptible to temperature drift, mechanical wear and frequent maintenance. By fusing laser displacement, three-dimensional and temperature multi-source data and updating through virtual reference, and combining three-dimensional reconstruction and feature extraction to generate composite three-dimensional morphology map and target material feature vector, it overcomes the limitation of a single data source being unable to comprehensively evaluate the complex morphology of the target material. Based on the dual-branch film quality-life prediction model constructed by CNN-LSTM-attention mechanism, it integrates process quality data and multi-dimensional features of the target material to realize film quality-life prediction, making up for the shortcomings of existing technologies that lack intelligent analysis and only output basic data. Finally, through multi-objective optimization function, Bayesian optimization algorithm and adaptive decision generation to generate monitoring feedback results, it solves the problems of frequent manual intervention and lack of closed-loop optimization, realizes full-process automation and continuous self-optimization from measurement to decision, and efficiently and accurately supports target material status assessment and production decision.

[0013] Furthermore, the construction of the virtual benchmark formula includes:

[0014] Based on the coordinate data of multiple sensors, the multiple sensors are calibrated using the hand-eye calibration method, and the pose transformation matrix is ​​calculated.

[0015] Set up a reference object and scan it using a driven measurement system (composed of a robotic arm and sensors) to obtain actual reference point cloud data;

[0016] Based on actual reference point cloud data and theoretical point cloud data of the reference object, the current error matrix is ​​calculated by using a registration algorithm.

[0017] Based on the current error matrix, a virtual benchmark formula is constructed.

[0018] In the above scheme, the present invention obtains the pose transformation matrix of multiple sensors through hand-eye calibration and constructs a virtual reference formula by combining the registration calculation error matrix. This replaces the fragile physical reference block relied upon by the existing technology. It not only solves the problem of inaccuracy caused by temperature drift and mechanical wear of physical reference, but also avoids the maintenance burden of frequent maintenance. At the same time, the virtual reference can be dynamically updated based on the registration error, which ensures the spatial consistency of multi-sensor data fusion. It provides a stable and reliable reference support for the subsequent accurate generation of composite three-dimensional topography map and extraction of target feature vectors, thus overcoming the measurement stability shortcomings of traditional physical reference from the source.

[0019] Furthermore, the generation of the composite three-dimensional topography map and the target feature vector includes:

[0020] Laser displacement data, three-dimensional data, and temperature distribution data of the target material are acquired by multiple sensors and calibrated using a virtual reference formula to obtain laser displacement calibration data, three-dimensional calibration data, and temperature distribution calibration data.

[0021] Based on laser displacement calibration data and 3D calibration data, point cloud fusion and 3D reconstruction are performed for different regions using a weighted fusion algorithm to generate composite 3D topography maps.

[0022] Based on the composite three-dimensional topography map and temperature distribution calibration data, geometric features, topographic features and thermodynamic features are calculated and used as target feature vectors.

[0023] In the above scheme, the present invention uses virtual benchmark formula calibration and weighted fusion algorithm to complete point cloud fusion and 3D reconstruction to generate a composite 3D topography map. The core features obtained not only overcome the limitation of existing technologies that a single data source cannot fully evaluate the complex topography of the target material, but also ensure the consistency and accuracy of multi-source data through virtual benchmark calibration. At the same time, the generated composite 3D topography map can be adapted to different target material types and loss modes, solving the problem of poor adaptability of static measurement. Furthermore, the multi-dimensional target material feature vectors make up for the shortcomings of traditional technologies that only output basic data and are difficult to support in-depth analysis, providing comprehensive and reliable input support for subsequent prediction models.

[0024] Furthermore, the dual-branch membrane quality-lifetime prediction model includes a CNN network, an LSTM network, an attention layer, a first fully connected layer, and a second fully connected layer.

[0025] Furthermore, the generated film quality-lifetime prediction results include:

[0026] Acquire process and quality data and input them into an LSTM network. Use forget gate, input gate and output gate to capture long-term and short-term dependencies and generate a high-dimensional temporal fusion feature vector.

[0027] The composite 3D topography image is input into a CNN network, and local features of the target surface are extracted through multiple convolutional layers to generate deep spatial topography features.

[0028] The high-dimensional temporal fusion feature vector, deep spatial morphology features and target feature vector are input into the attention layer. Different weights of different features on lifetime prediction or quality indicators are calculated to obtain the temporal spatial key feature vector.

[0029] The key feature vectors of the time-series space are input into the first fully connected layer to generate the remaining lifetime prediction results, that is, to obtain the remaining usable lifetime and future performance degradation curve of the target material.

[0030] The key feature vectors of the temporal space are input into the second fully connected layer to generate quality prediction results, that is, to obtain the predicted values ​​of key quality indicators.

[0031] Based on the remaining lifetime prediction results and the quality prediction results, the membrane quality-lifetime prediction results are generated.

[0032] In the above-mentioned scheme, this invention constructs a dual-branch membrane quality-lifetime prediction model, with multi-dimensional feature fusion and targeted prediction as its core features, systematically solving several shortcomings of existing technologies: by using LSTM networks to capture the long-term and short-term temporal dependencies of process and quality data, and CNN networks to extract deep features of the target material's spatial morphology, a comprehensive fusion of temporal and spatial data is achieved, overcoming the problem of one-sided evaluation from a single data source; by dynamically allocating the weights of different features to the prediction task through the attention layer, the contribution of key information is strengthened, and the prediction accuracy is improved; relying on the dual fully connected layers to output the remaining usable life of the target material, the future performance degradation curve, and the predicted values ​​of key quality indicators, the basic data is transformed into deep analysis results, making up for the lack of intelligent analysis in existing technologies. At the same time, it automatically generates comprehensive membrane quality-lifetime prediction results without the need for manual data interpretation and decision-making, greatly reducing the dependence on human intervention and providing accurate and efficient decision-making basis for subsequent optimization and feedback.

[0033] Furthermore, the generation of monitoring feedback results includes:

[0034] Set a remaining lifetime threshold and an erosion depth threshold to determine whether the remaining usable lifetime is less than the remaining lifetime threshold and whether the average erosion depth is greater than the erosion depth threshold. If so, generate a monitoring feedback result through adaptive decision-making, i.e., send the "replace immediately" message to the cloud. Otherwise, determine whether the remaining usable lifetime is less than the remaining lifetime threshold and whether the average erosion depth is less than the erosion depth threshold.

[0035] If so, adaptive decision-making generates monitoring feedback results, i.e., sending "planned maintenance suggestions and recommendations to be executed in the next planned shutdown window" to the cloud; otherwise, a multi-objective optimization function is constructed, and combined with the membrane quality-lifetime prediction results, the current process parameter sequence is optimized using a Bayesian optimization algorithm to generate optimized process parameters, and the optimized process parameters are updated to the current process flow and monitored.

[0036] Furthermore, the formula corresponding to the multi-objective optimization function is:

[0037] ;

[0038] in, Represents a multi-objective optimization function. This represents the risk weighting coefficient. This represents the current sequence of process parameters. This represents the current predicted value of key quality indicators. This represents the current temperature gradient. Indicates the temperature gradient threshold. This represents the maximum value function.

[0039] In the above solution, this invention addresses the shortcomings of existing technologies, such as frequent manual intervention, lack of closed-loop optimization, and lack of risk management, by combining dual-threshold adaptive decision-making with multi-objective Bayesian optimization. First, based on the dual thresholds of remaining lifetime and erosion depth, it automatically outputs precise instructions such as "immediate replacement" and "planned maintenance" through adaptive decision-making, replacing the traditional decision-making mode of manually interpreting data, significantly reducing reliance on manual intervention and avoiding decision lag. Then, it dynamically optimizes the process parameter sequence through a Bayesian optimization algorithm and updates it to the process flow, forming a complete closed loop of "measurement-prediction-decision-optimization-monitoring". This solves the problem of existing technologies lacking continuous self-optimization, ensuring production continuity and further improving film quality and target life through dynamic process adjustments, achieving full-chain optimization of production efficiency.

[0040] A target monitoring system based on multi-source data and lifetime prediction includes:

[0041] The reference construction module is used to calculate the pose transformation matrix of the coordinate system of multiple sensors, and combine it with the registration algorithm to calculate the error matrix and construct the virtual reference formula.

[0042] The calculation module is used to acquire laser displacement data, three-dimensional data and temperature distribution data of the target material and update them through virtual benchmark formula. Combined with three-dimensional reconstruction and feature extraction, it generates a composite three-dimensional topography map and target material feature vector.

[0043] The building blocks are used to construct a dual-branch membrane quality-lifetime prediction model based on CNN networks, LSTM networks, and attention mechanisms.

[0044] The prediction module is used to acquire process and quality data and input them into the dual-branch membrane quality-lifetime prediction model along with the composite 3D morphology map and target material feature vector to generate membrane quality-lifetime prediction results.

[0045] The monitoring module is used to generate monitoring feedback results based on multi-objective optimization functions and membrane quality-lifetime prediction results, through Bayesian optimization algorithms and adaptive decision-making.

[0046] In the above scheme, this system constructs a virtual benchmark formula through a benchmark construction module, and the calculation module integrates multi-source target material data to generate a composite 3D topography map and target material feature vectors. The construction module builds a dual-branch prediction model based on CNN network, LSTM network, and attention mechanism, and the prediction module combines process quality data to output film quality-lifetime prediction results. The monitoring module relies on a multi-objective optimization function and Bayesian optimization algorithm to generate monitoring feedback results through adaptive decision-making. Attached Figure Description

[0047] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0048] Figure 1 This is a flowchart of the method in Embodiment 1 of the present invention;

[0049] Figure 2 This is a structural diagram of the dual-branched membrane quality-lifetime prediction model in Embodiment 1 of the present invention;

[0050] Figure 3 This is a system structure diagram in Embodiment 1 of the present invention. Detailed Implementation

[0051] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0052] Example 1:

[0053] Please see Figure 1 This embodiment provides a target monitoring method based on multi-source data and lifetime prediction. Figure 1 The execution entity of the method shown can be a software and / or hardware device. The execution entity of this application can include, but is not limited to, at least one of the following: user equipment, network equipment, etc. User equipment can include, but is not limited to, computers, smartphones, personal digital assistants (PDAs), and the aforementioned electronic devices. Network equipment can include, but is not limited to, a single network server, a server group consisting of multiple network servers, or a cloud based on cloud computing consisting of a large number of computers or network servers. Cloud computing is a type of distributed computing, consisting of a super virtual computer composed of a group of loosely coupled computers. This embodiment does not impose any limitations on this.

[0054] like Figure 1 As shown, a target monitoring method based on multi-source data and lifetime prediction includes:

[0055] S1. Calculate the pose transformation matrix in the coordinate system of multiple sensors, and combine it with the registration algorithm to calculate the error matrix, constructing a virtual reference formula. In this embodiment, a laser displacement sensor, a 3D vision sensor, and a thermal imager are used. The laser displacement sensor is used to provide high-precision single-point or line-shaped depth information (laser displacement data); the 3D vision sensor is used to provide dense three-dimensional points with rich texture (three-dimensional data); and the thermal imager is used to provide the temperature distribution matrix of the target surface (temperature distribution data).

[0056] S1 includes:

[0057] S1-1. Based on the coordinate data of multiple sensors, calibrate the multiple sensors using the hand-eye calibration method, and calculate the pose transformation matrix. ;

[0058] S1-2. Set up a reference object and scan it using a driven measurement system (composed of a robotic arm and sensors) to acquire actual reference point cloud data. In this embodiment, a ceramic reference sphere is used as the reference object.

[0059] S1-3. Based on the actual reference point cloud data and the theoretical point cloud data of the reference object, the IPC algorithm (registration algorithm) is used to solve for the transformation matrix that minimizes the difference between the actual reference point cloud data and the theoretical point cloud data, i.e., the current error matrix is ​​calculated. ;

[0060] The formula corresponding to S1-3 is:

[0061] ;

[0062] in, Indicates the first A real-world reference point cloud data, Indicates the first A theoretical point cloud dataset, Represents the norm, Represents the spatial transformation matrix. This represents the summation function. Describes the minimum value function. This represents the total number of actual reference point cloud data / theoretical point cloud data.

[0063] Spatial transformation matrix This refers to the transformation operation of registering theoretical point cloud data with actual reference point cloud data, transforming the theoretical point cloud data... Mapped to actual reference point cloud data The best matching position. Through optimization... To minimize the error between the theoretical point cloud and the actual point cloud, the error matrix is ​​obtained. This is used for subsequent virtual benchmark calibration.

[0064] S1-4. Based on the current error matrix, construct a virtual benchmark formula. The corresponding formula is:

[0065] ;

[0066] in, This represents the inverse of the current error matrix. This indicates the calibrated data. This represents the measurement data (data acquired by the sensor in the following section).

[0067] The core idea of ​​a virtual reference is to quickly calibrate the coordinate system and error model of the entire measurement system before each measurement using a fixed, high-precision reference object (such as a ceramic reference sphere). Since the reference object itself is stable and unaffected by wear, this method can eliminate accumulated errors caused by temperature drift, mechanical wear, etc., essentially using a "brand new reference" for each measurement. Therefore, a virtual reference not only replaces fragile physical reference blocks but also enables dynamic self-calibration, improving the accuracy and reliability of long-term measurements.

[0068] The virtual reference formula constructed in this embodiment can improve the problem of errors caused by environmental and mechanical deformation of traditional physical reference blocks. By identifying a fixed, high-precision reference feature (such as a reference sphere installed in a non-wear area), the coordinate system and error model of the current measurement system can be quickly established before each measurement. This is equivalent to using a completely new reference without accumulated errors for each measurement, thereby eliminating systematic drift errors in principle and improving the absolute accuracy and reliability of long-term measurements.

[0069] S2. Acquire laser displacement data, three-dimensional data and temperature distribution data of the target material and update them through virtual benchmark formula. Combine three-dimensional reconstruction and feature extraction to generate composite three-dimensional topography map and target material feature vector.

[0070] Because a single sensor is limited by its own principle and perspective, it is difficult to independently complete a comprehensive and accurate assessment of the target material's condition: while laser displacement sensors can provide high-precision single-point / line depth information, they cannot capture the details and textures of planar defects; 3D vision sensors can output dense 3D point clouds rich in texture, but the accuracy of the depth information is insufficient to support fine measurements; thermal imagers can capture the temperature distribution on the target material surface, but lack spatial dimension information about the geometric shape. To overcome this limitation, a comprehensive characterization of the target material's condition is achieved through the collaborative processing of data from multiple sensors. Therefore, S2 includes:

[0071] S2-1. Laser displacement data, three-dimensional data, and temperature distribution data of the target material are acquired through multiple sensors and calibrated using a virtual reference formula to obtain laser displacement calibration data. 3D calibration data and temperature distribution calibration data ;

[0072] In this embodiment, the laser displacement data, three-dimensional data, and temperature distribution data of the target material are used as measurement data, and calculated with the inverse of the current error matrix to generate laser displacement calibration data, three-dimensional calibration data, and temperature distribution calibration data. , This represents the two-dimensional coordinates measured by the laser displacement sensor and calibrated using a virtual reference formula. , , This represents the coordinates of a 3D point cloud measured by a 3D vision sensor and calibrated using a virtual reference formula. Indicates the first Intensity information of three-dimensional calibration data.

[0073] S2-2. Based on laser displacement calibration data and 3D calibration data, point cloud fusion and 3D reconstruction are performed for different regions using a weighted fusion algorithm to generate composite 3D topography images. ;

[0074] Specifically, a weighted fusion algorithm is used to integrate the registered laser displacement calibration data and 3D calibration data region by region. For regions with overlapping data, weights are assigned based on the performance advantages of the laser displacement sensor and the 3D vision sensor. Since the depth value of the laser displacement sensor has higher measurement accuracy, it is given a higher weight to ensure the accuracy of the spatial geometric information in this region. At the same time, the detailed texture information carried by the point cloud of the 3D vision sensor is called in to supplement the surface micro-morphological features (such as the edge texture of erosion pits and surface unevenness details). Finally, the overlapping region data after weighted fusion is integrated with the non-overlapping effective data of the two types of sensors to construct a high-fidelity composite 3D topography map that combines high-precision spatial dimensional information with rich surface texture details.

[0075] S2-3. Based on the composite three-dimensional topography map and temperature distribution calibration data, calculate the geometric features, topography features and thermodynamic features, and integrate them as the target material feature vector.

[0076] Specifically, the geometric features include the average erosion depth. Maximum erosion depth erosion pit volume 1. Erosion profile uniformity (in standard deviation) (Represented). The morphological feature is the surface roughness morphological feature. Geometric and topographic features are calculated using depth information and detail features from composite 3D topographic maps.

[0077] The thermodynamic characteristics include the highest temperature of the eroded area. Average temperature and temperature gradient First, the erosion area is determined in the composite three-dimensional topography map, and then the corresponding temperature distribution calibration data is calculated.

[0078] In this embodiment, the expression for the target feature vector is: .

[0079] S3. Based on CNN networks, LSTM networks, and attention mechanisms, a dual-branch membrane quality-lifetime prediction model is constructed; the dual-branch membrane quality-lifetime prediction model includes a CNN network, an LSTM network, an attention layer, a first fully connected layer, and a second fully connected layer, as shown below. Figure 2 As shown.

[0080] Traditional methods rely on human experience to establish fuzzy relationships between data and conclusions. Therefore, this approach utilizes machine learning models to automatically learn the complex, non-linear mapping relationships between target morphology characteristics, process parameters, and the final coating quality and target life from a large amount of historical data. This data-driven quantitative model offers predictive accuracy and foresight far exceeding human experience, providing a technological foundation for predictive maintenance and lean manufacturing.

[0081] The core advantage of attention mechanisms is their ability to adaptively assign high weights to key information, filter redundant data, and significantly improve the targeting and efficiency of feature extraction. This adapts to the varying importance of different regions in target material condition assessment; for example, severely eroded areas and regions with large temperature gradients are more valuable for lifetime prediction. Attention mechanisms can automatically focus on these core regions, avoiding indiscriminate processing of invalid information. CNN (Convolutional Neural Network) networks are naturally adept at capturing spatially localized features and can automatically mine geometric details and regional differences in matrix and image-based data. They are highly efficient networks for processing spatial structural data, suitable for handling spatial data sources such as composite 3D topographic maps and temperature distribution maps of target materials. They can accurately extract geometric and thermodynamic spatial features such as erosion pit shape and temperature distribution. LSTM (Long Short-Term Memory) networks can effectively model temporal dependencies, remembering the impact of historical data on the current state. This overcomes the limitation of ordinary networks in handling dynamically evolving data. They are adapted to the dynamic changes in target material conditions over time, such as the cumulative increase in erosion depth and temporal fluctuations in temperature. LSTM networks can accurately capture the patterns of these state evolutions.

[0082] Therefore, in this embodiment, a dual-branch membrane quality-lifespan prediction model (CNN-LSTM-Attention hybrid network) is constructed by combining CNN networks, LSTM networks, and attention mechanisms. This model not only extracts detailed spatial features through CNN networks but also focuses on core regions and reduces redundant interference through attention mechanisms. Furthermore, it captures the temporal evolution logic completely through LSTM networks. This model perfectly covers all the requirements for target material evaluation / lifespan prediction, including "spatial multi-source feature extraction, key information focusing, and temporal trend modeling." Ultimately, this results in more comprehensive and accurate feature representation, significantly improving the reliability of target material remaining lifespan prediction and the foresight of state assessment, providing stronger technical support for subsequent maintenance decisions and process optimization.

[0083] During the training of the bibranched membrane quality-lifetime prediction model, the loss functions (such as mean squared error) of the two output tasks are optimized simultaneously through supervised learning backpropagation, thereby learning to map the correlation between lifetime and quality from the input features.

[0084] The first fully connected layer (FC1) is responsible for predicting the remaining lifetime (RUL). Its output layer is usually a single neuron (for regression tasks), and it can be paired with a time-series output channel to generate performance degradation curves.

[0085] The second fully connected layer (FC2) is responsible for predicting key quality metrics (such as uniformity, adhesion, etc.). The number of neurons in its output layer is consistent with the number of key quality metrics predicted (for multi-output regression or classification tasks).

[0086] Furthermore, the number of neurons, activation functions, and regularization methods of the hidden layers corresponding to FC1 and FC2 can be independently tuned based on historical data from their respective tasks. During training, the two branches share the features extracted by the CNN-LSTM-Attention hybrid network, but are jointly optimized using a multi-task loss function (such as weighted MSE).

[0087] S4. Acquire process and quality data and input them into the dual-branch membrane quality-lifetime prediction model along with the composite 3D morphology diagram and target material feature vector to generate membrane quality-lifetime prediction results.

[0088] The membrane quality-lifetime prediction result includes the remaining lifetime prediction result and the quality prediction result, and S4 includes:

[0089] S4-1. Acquire process and quality data and input them into the LSTM network. Capture long-term and short-term dependencies through forget gate, input gate, and output gate to generate a high-dimensional temporal fusion feature vector. Process and quality data include current and historical process parameter sequences (such as sputtering power and gas pressure) and historical film quality data (such as sheet resistance and uniformity).

[0090] Specifically, the LSTM network, through the synergistic effect of its forget gate, input gate, and output gate, first filters out redundant historical information irrelevant to the current prediction, then updates and stores key temporal features of process parameter fluctuations and quality indicator changes, while capturing the long-term and short-term dependencies between the two types of data (such as the correlation between a continuous increase in sputtering power and a subsequent decrease in film uniformity, and the temporal continuity of historical quality anomaly patterns). Finally, it generates a high-dimensional temporal fusion feature vector that integrates the process-quality temporal evolution logic to supplement the CNN-LSTM-Attention hybrid model with key temporal dimension information. This vector is efficiently fused with the target space features extracted by the CNN network at the attention layer, allowing the subsequent fully connected layer to predict the remaining lifespan of the target and key quality indicators of the film to not only conform to historical data patterns but also have stronger accuracy and foresight. This provides a comprehensive and reliable temporal basis for judging the timing of target maintenance and optimizing coating process parameters.

[0091] S4-2. Input the composite 3D topography image into the CNN network, and extract the local features of the target surface through multiple convolutional layers to generate deep spatial topography features.

[0092] Specifically, the composite 3D topography image undergoes data normalization to unify spatial resolution and numerical range before being input into a CNN network. Multiple convolutional layers with kernels of varying sizes are used to capture local features of the target surface (such as the edge contours of erosion pits, surface texture, and boundary differences between eroded and uneroded areas) by sliding across regions. Pooling layers then compress redundant information and retain key features. Simultaneously, nonlinearity is introduced using activation functions like ReLU to enhance the network's ability to learn complex topography patterns. After multiple rounds of convolution-pooling iterations, the network gradually abstracts from simple geometric features (such as points, lines, and surfaces) to complex semantic features (such as the spatial distribution patterns of erosion pits, regional differences in surface roughness, and the uniformity of the overall erosion contour). Finally, global pooling or fully connected layers are used to integrate and output a deep spatial topography feature vector that encapsulates the core information of the target material's spatial topography.

[0093] S4-3. Input the high-dimensional temporal fusion feature vector, deep spatial morphology features and target feature vector into the attention layer, calculate the different weights of different features on lifetime prediction or quality indicators, and obtain the temporal spatial key feature vector.

[0094] Specifically, the input data is first aligned and normalized to eliminate differences in feature scales, then concatenated to obtain fused data, which is then input into the attention layer. Based on the task objective of lifetime prediction or quality indicators, a self-attention mechanism or task-adaptive attention scoring function is used to quantify the correlation strength between three types of features and the prediction task, as well as the internal dimensions of each feature (e.g., spatial features of severe erosion, temporal features of rapid decay, and thermodynamic features of high temperature gradients are given higher weights, while redundant information is assigned lower weights). Then, the three types of features are dynamically fused through weighted summation, ultimately obtaining a temporal and spatial key feature vector that focuses on core correlation information and highlights key contribution factors of the task. In this embodiment, taking the calculation of different weights for different features in lifetime prediction as an example, the weights... The corresponding formula is:

[0095] ;

[0096] in, This represents an exponential function with base e. This represents the function for calculating attention scores. Indicates the first in the fused data A vector representation of n elements, This represents fused data. The weights of different features on quality indicators are calculated in the same way as the weights for calculating lifetime prediction.

[0097] S4-4. Input the key feature vectors of the time-series space into the first fully connected layer to generate the remaining lifetime prediction results, which include the remaining usable lifetime (RUL) of the target material and the future performance degradation curve.

[0098] Specifically, by using linear transformations of multiple hidden layers and nonlinear activation functions such as ReLU, the intrinsic coupling relationship between temporal evolution patterns, spatial morphology characteristics, and target life decay in feature vectors is deeply explored (such as the impact of the temporal cumulative effect of key erosion features on the remaining lifetime). At the same time, the adaptability to the lifetime prediction task is enhanced by optimizing network parameters. The output channel of a single neuron corresponds to the quantified value of the remaining usable lifetime (RUL) of the target, while the output channels of multiple temporal nodes correspond to the predicted values ​​of performance parameters for different future periods. This generates the remaining usable lifetime (RUL) of the target and a future performance decay curve that reflects the future performance change trend throughout the entire life cycle.

[0099] S4-5. Input the key feature vectors of the temporal space into the second fully connected layer to generate quality prediction results, that is, obtain the predicted values ​​of key quality indicators.

[0100] Specifically, by using linear transformations of multiple hidden layers and nonlinear activation functions such as ReLU, the intrinsic coupling relationship between the temporal evolution patterns, spatial morphology characteristics, and key quality indicators of the membrane (such as uniformity, adhesion, and thickness deviation) in the feature vector is deeply explored. At the same time, the adaptability to the quality prediction task is enhanced by optimizing network parameters. Finally, by using a neuron design that matches the output layer with the dimension of the quality indicators, the predicted values ​​of key quality indicators that can accurately reflect the core performance status of the membrane are generated.

[0101] S5. Based on the multi-objective optimization function and the film quality-lifetime prediction results, monitoring feedback results are generated through Bayesian optimization algorithm and adaptive decision-making. Adaptive decision-making is based on instruction information designed according to historical time; for example, if the remaining usable lifetime is too short, the target material is replaced immediately.

[0102] S5 includes:

[0103] S5-1. Set the remaining lifetime threshold and erosion depth threshold, and determine whether the remaining usable lifetime is less than the remaining lifetime threshold and whether the average erosion depth is greater than the erosion depth threshold. If so, generate monitoring feedback results through adaptive decision-making, that is, send the "replace immediately" message to the cloud; otherwise, proceed to S5-2.

[0104] S5-2. Determine whether the remaining usable lifespan is less than the remaining lifespan threshold and whether the average erosion depth is less than the erosion depth threshold. If so, generate monitoring feedback results through adaptive decision-making, i.e., send a "planned maintenance (such as target rotation) suggestion and recommend execution in the next planned shutdown window" message to the cloud; otherwise, proceed to S5-3.

[0105] S5-3. Construct a multi-objective optimization function and combine it with the membrane quality-lifetime prediction results. Then, use a Bayesian optimization algorithm to optimize the current process parameter sequence and generate optimized process parameters.

[0106] Specifically, the key physical failure model of "suppressing crack risk" (temperature gradient) is directly incorporated as part of the optimization objective. This allows the Bayesian optimization algorithm to proactively avoid high-risk process zones, a feature not previously considered in this field. Physical knowledge (temperature gradients causing cracks) is transformed into a specific algorithmic feature, directly linked to the problem of "extending the lifespan of the target material." Thus, the multi-objective optimization function... The corresponding formula is:

[0107] ;

[0108] in, This represents the risk weighting coefficient (which determines the extent to which safety should be prioritized over simply pursuing quality when the current process may cause cracking). This represents the current sequence of process parameters. This represents the current predicted value of key quality indicators. This represents the current temperature gradient. Indicates the temperature gradient threshold. This represents the maximum value function.

[0109] Using a multi-objective optimization function as the optimization objective, a prior model for Bayesian optimization (such as a Gaussian process) is first initialized. The current sequence of process parameters, the corresponding predicted values ​​of key film quality indicators, and the temperature gradient are input, and the mapping relationship between the process parameters and the multi-objective optimization function is fitted. Then, a data acquisition function (such as the desired improvement in EI) is used to select the next parameter sequence to be evaluated within the process parameter window. This sequence is then input into a bi-branch film quality-lifetime prediction model for simulation, obtaining the corresponding predicted values ​​of key film quality indicators and acquiring the temperature gradient. The optimization objective value is then calculated, and this set of data is used to update the prior model. This iterative process of "parameter selection-evaluation-update" is repeated until the model converges, finally outputting the optimized process parameters (sputtering power, working gas pressure, and gas flow rate) that maximize the optimization objective value. The physical failure knowledge of "temperature gradient-induced cracking" is transformed into algorithmic features, allowing the Bayesian optimization algorithm to proactively avoid high-risk process areas, suppressing crack generation at the source and improving process safety. While optimizing the predicted values ​​of key film quality indicators, the safe use requirements of the sputtering target are also considered, achieving synergy between quality improvement and target risk management. Compared to traditional optimization that focuses solely on quality, this process not only ensures the quality of the film layer but also indirectly extends the life of the target material, allowing process parameter adjustments to better suit the actual needs of production.

[0110] S5-4. Update the optimized process parameters to the current process flow and monitor them.

[0111] In summary, this invention eliminates reliance on fragile physical reference hardware through dynamic self-calibration, reducing maintenance costs while ensuring the long-term stability and reliability of the measurement process, providing a continuous and reliable foundation of data for subsequent analysis. Furthermore, by relying on multi-sensor data fusion, it overcomes the limitations of single sensors at the fundamental level, achieving comprehensive coverage of target material status information, making evaluation results more accurate and complete, and avoiding the risk of misjudgment based on single data. Moreover, through intelligent analysis and lifespan prediction, it upgrades passive measurement to proactive forward-looking decision-making, reserving sufficient adjustment windows for production through lifespan prediction. Ultimately, it achieves a synergistic effect of maximizing target material utilization efficiency, minimizing production interruptions, and precisely optimizing process parameters, improving the intelligence level and operational efficiency of the production system across the entire process from measurement and evaluation to decision-making.

[0112] like Figure 3 As shown, a target monitoring system based on multi-source data and lifetime prediction includes:

[0113] The reference construction module is used to calculate the pose transformation matrix of the coordinate system of multiple sensors, and combine it with the registration algorithm to calculate the error matrix and construct the virtual reference formula.

[0114] The calculation module is used to acquire laser displacement data, three-dimensional data and temperature distribution data of the target material and update them through virtual benchmark formula. Combined with three-dimensional reconstruction and feature extraction, it generates a composite three-dimensional topography map and target material feature vector.

[0115] The building blocks are used to construct a dual-branch membrane quality-lifetime prediction model based on CNN networks, LSTM networks, and attention mechanisms.

[0116] The prediction module is used to acquire process and quality data and input them into the dual-branch membrane quality-lifetime prediction model along with the composite 3D morphology map and target material feature vector to generate membrane quality-lifetime prediction results.

[0117] The monitoring module is used to generate monitoring feedback results based on multi-objective optimization functions and membrane quality-lifetime prediction results, through Bayesian optimization algorithms and adaptive decision-making.

[0118] It should be noted that the specific methods by which each module performs operations in the system described in the above embodiments have been described in detail in the embodiments related to the method, and will not be elaborated here.

[0119] Example 2:

[0120] The approach from Example 1 was applied to predictive maintenance decision-making. Five planar molybdenum targets (500mm × 130mm × 15mm) from the same batch were selected as experimental subjects and designated as targets 1 to 5, respectively. The monitoring cycle was set to perform full-process monitoring every 8 hours of coating, i.e., the approach from Example 1 was executed every 8 hours. The remaining lifetime threshold and erosion depth threshold were set to 48 hours and 10.5mm (70% of nominal thickness), respectively.

[0121] Taking target material No. 3 as an example, the monitoring feedback results generated by the scheme in Example 1 are shown in Table 1. Table 1

[0122]

[0123] Example 3:

[0124] The scheme in Example 1 was used to obtain the optimal operating conditions (real-time optimization of process parameters), with a rotating cylindrical aluminum target (152 mm in diameter and 2500 mm in length) as the experimental subject. Sixty hours after coating, a local temperature gradient of 185°C / cm (safe threshold 200°C / cm) was detected, and microcracks were observed in this area when the composite 3D topography image was fused. Based on the prediction results obtained from the dual-branch film quality-lifetime prediction model, it was determined that the current process parameters would accelerate crack propagation. Therefore, the coating process parameters were adjusted using a Bayesian optimization algorithm to optimize sputtering power, gas pressure, and cooling gas flow rate. The process parameters before and after optimization are shown in Table 2.

[0125] Table 2

[0126]

[0127] As shown in Table 2, the optimized process parameters effectively suppress the risk of cracking while maintaining the quality of the film layer, extending the predicted life of the target material by 15% and reducing the crack propagation rate by 58%.

[0128] Example 4:

[0129] The scheme from Example 1 was applied to optimal usage state 3 (acceptance and documentation of new target materials). Three newly purchased ITO ceramic rotating targets (99.9% density) were named ITO ceramic rotating targets 1 to 3, respectively, and monitored. Their initial three-dimensional morphology, density distribution, and thermal conductivity distribution were collected to generate a unique "digital profile" and calibration matrix. The profile data was then used to fine-tune the bibranched film quality-lifetime prediction model from Example 1, establishing a customized prediction model. Taking ITO ceramic rotating target 1 as an example, the bibranched film quality-lifetime prediction model from Example 1 and the customized prediction model were used to process ITO ceramic rotating target 1, and the results are shown in Table 3.

[0130] Table 3

[0131]

[0132] As shown in Table 3, the customized model based on digital archives improves the lifetime prediction accuracy by more than 55% compared with the general model, and can provide personalized process start-up strategies for specific weak links of the target material.

[0133] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0134] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A target material monitoring method based on multi-source data and lifetime prediction, characterized in that, include: Calculate the pose transformation matrix of the coordinate system of multiple sensors, and combine it with the registration algorithm to calculate the error matrix, and construct a virtual reference formula; The laser displacement data, three-dimensional data, and temperature distribution data of the target material are acquired and updated using a virtual reference formula. Combined with three-dimensional reconstruction and feature extraction, a composite three-dimensional topography map and a target material feature vector are generated. A dual-branch membrane quality-lifetime prediction model is constructed based on CNN network, LSTM network and attention mechanism; The process and quality data are acquired and input into the dual-branch membrane quality-lifetime prediction model along with the composite 3D morphology map and the target material feature vector to generate membrane quality-lifetime prediction results. Based on the multi-objective optimization function and the membrane quality-lifetime prediction results, monitoring feedback results are generated through Bayesian optimization algorithm and adaptive decision-making.

2. The target monitoring method based on multi-source data and lifetime prediction according to claim 1, characterized in that, The formula for constructing the virtual benchmark includes: Based on the coordinate data of multiple sensors, the multiple sensors are calibrated using the hand-eye calibration method, and the pose transformation matrix is ​​calculated. Set up a reference object and scan the reference object using a driving measurement system to obtain actual reference point cloud data; Based on actual reference point cloud data and theoretical point cloud data of the reference object, the current error matrix is ​​calculated by using a registration algorithm. Based on the current error matrix, a virtual benchmark formula is constructed.

3. The target monitoring method based on multi-source data and lifetime prediction according to claim 1, characterized in that, The generation of the composite three-dimensional topography map and the target feature vector includes: Laser displacement data, three-dimensional data, and temperature distribution data of the target material are acquired by multiple sensors and calibrated using a virtual reference formula to obtain laser displacement calibration data, three-dimensional calibration data, and temperature distribution calibration data. Based on laser displacement calibration data and 3D calibration data, point cloud fusion and 3D reconstruction are performed for different regions using a weighted fusion algorithm to generate composite 3D topography maps. Based on the composite three-dimensional topography map and temperature distribution calibration data, geometric features, topographic features and thermodynamic features are calculated and used as target feature vectors.

4. The target monitoring method based on multi-source data and lifetime prediction according to claim 1, characterized in that, The dual-branch membrane quality-lifetime prediction model includes a CNN network, an LSTM network, an attention layer, a first fully connected layer, and a second fully connected layer.

5. The target monitoring method based on multi-source data and lifetime prediction according to claim 4, characterized in that, The generated film quality-lifetime prediction results include: Acquire process and quality data and input them into an LSTM network. Use forget gate, input gate and output gate to capture long-term and short-term dependencies and generate a high-dimensional temporal fusion feature vector. The composite 3D topography image is input into a CNN network, and local features of the target surface are extracted through multiple convolutional layers to generate deep spatial topography features. The high-dimensional temporal fusion feature vector, deep spatial morphology features and target feature vector are input into the attention layer. Different weights of different features on lifetime prediction or quality indicators are calculated to obtain the temporal spatial key feature vector. The key feature vectors of the time-series space are input into the first fully connected layer to generate the remaining lifetime prediction results, that is, to obtain the remaining usable lifetime and future performance degradation curve of the target material. The key feature vectors of the temporal space are input into the second fully connected layer to generate quality prediction results, that is, to obtain the predicted values ​​of key quality indicators. Based on the remaining lifetime prediction results and the quality prediction results, the membrane quality-lifetime prediction results are generated.

6. The target monitoring method based on multi-source data and lifetime prediction according to claim 5, characterized in that, The generation of monitoring feedback results includes: Set a remaining lifetime threshold and an erosion depth threshold to determine whether the remaining usable lifetime is less than the remaining lifetime threshold and whether the average erosion depth is greater than the erosion depth threshold. If so, generate a monitoring feedback result through adaptive decision-making, i.e., send the "replace immediately" message to the cloud. Otherwise, determine whether the remaining usable lifetime is less than the remaining lifetime threshold and whether the average erosion depth is less than the erosion depth threshold. If so, adaptive decision-making generates monitoring feedback results, i.e., sending "planned maintenance suggestions and recommendations to be executed in the next planned shutdown window" to the cloud; otherwise, a multi-objective optimization function is constructed, and combined with the membrane quality-lifetime prediction results, the current process parameter sequence is optimized using a Bayesian optimization algorithm to generate optimized process parameters, and the optimized process parameters are updated to the current process flow and monitored.

7. The target monitoring method based on multi-source data and lifetime prediction according to claim 6, characterized in that, The formula corresponding to the multi-objective optimization function is: ； in, Represents a multi-objective optimization function. This represents the risk weighting coefficient. This represents the current sequence of process parameters. This represents the current predicted value of key quality indicators. This represents the current temperature gradient. Indicates the temperature gradient threshold. This represents the maximum value function.

8. A target material monitoring system based on multi-source data and lifetime prediction, used to implement the target material monitoring method based on multi-source data and lifetime prediction as described in any one of claims 1 to 7, characterized in that, include: The reference construction module is used to calculate the pose transformation matrix of the coordinate system of multiple sensors, and combine it with the registration algorithm to calculate the error matrix and construct the virtual reference formula. The calculation module is used to acquire laser displacement data, three-dimensional data and temperature distribution data of the target material and update them through virtual benchmark formula. Combined with three-dimensional reconstruction and feature extraction, it generates a composite three-dimensional topography map and target material feature vector. The building blocks are used to construct a dual-branch membrane quality-lifetime prediction model based on CNN networks, LSTM networks, and attention mechanisms. The prediction module is used to acquire process and quality data and input them into the dual-branch membrane quality-lifetime prediction model along with the composite 3D morphology map and target material feature vector to generate membrane quality-lifetime prediction results. The monitoring module is used to generate monitoring feedback results based on multi-objective optimization functions and membrane quality-lifetime prediction results, through Bayesian optimization algorithms and adaptive decision-making.

Images