A data analysis-based method and system for monitoring the condition of thin film evaporation equipment

Abstract

This application relates to the field of data processing technology, and in particular to a method and system for monitoring the condition of thin film evaporation equipment based on data analysis. The method includes: extracting the first principal component direction and the maximum eigenvalue of the monitoring matrix of the monitoring window and adjacent monitoring windows respectively, and constructing a cooperative stability index; constructing a penalty factor by combining the projection residuals of multidimensional monitoring data along the first principal component direction; constructing a segmentation descriptor to divide the multidimensional monitoring data within the separable nodes into a first subset and a second subset; constructing a segmentation factor based on the difference in penalty factors between the multidimensional monitoring data in the first subset and the second subset; selecting the largest segmentation descriptor to construct an isolated forest model; inputting real-time multidimensional monitoring data into the isolated forest model; outputting anomaly scores; and monitoring the equipment condition. The technical solution of this application can improve the accuracy of the isolated forest model in monitoring the condition of thin film evaporation equipment.

Description

Technical Field

[0001] This application relates to the field of data processing technology. More specifically, this application relates to a method and system for monitoring the condition of thin film evaporation equipment based on data analysis. Background Technology

[0002] Thin-film evaporation equipment is a core process device that uses a heated evaporation source under vacuum conditions to evaporate specific materials and condense them on a substrate surface to form nanoscale or microscale thin films. It has applications in high-precision technology fields such as semiconductor manufacturing, OLED display panels, optical coatings, and solar cells. The operation of this equipment is precise and complex; even minute fluctuations in hundreds of parameters, such as the vacuum level, evaporation source temperature, substrate temperature, and cooling water flow rate, can directly affect key quality indicators such as the thickness, uniformity, and purity of the final film. Therefore, real-time and accurate status monitoring of thin-film evaporation equipment, early warning of potential anomalies, and diagnosis of the root cause of failures are crucial for ensuring product yield and improving production efficiency.

[0003] In existing technologies, the Isolation Forest model, which does not require complex distance or density calculations when processing high-dimensional data, is commonly used for condition monitoring of thin-film evaporation equipment. This algorithm isolates outliers by randomly constructing isolated trees. Because outliers are relatively few in number and deviate from normal points, they are quickly isolated within the isolated trees. However, during the operation of thin-film evaporation equipment, different components cooperate to complete tasks, resulting in relatively stable cooperative patterns in monitoring data across different dimensions. When the equipment experiences early failures or performance degradation, a single dimension may not show deviations from normal points, thus disrupting these cooperative patterns. Traditional Isolation Forest models split based on a single random dimension, ignoring the correlation structure between different dimensions and failing to detect whether the cooperative constraints between parameters have been broken. Therefore, when an early failure occurs, even if the data in a single dimension has not yet shown significant deviations, the disruption of the cooperative pattern has already occurred, which traditional models cannot identify. Directly using the traditional Isolation Forest model to monitor the operating status of thin-film evaporation equipment is insensitive to early failures that lack significant single-dimensional anomalies and only involve disruption of cooperative patterns, leading to reduced accuracy of monitoring results. Summary of the Invention

[0004] This application provides a data analysis-based method and system for monitoring the condition of thin film evaporation equipment. It aims to solve the problem that the traditional isolated forest model ignores the collaborative correlation structure between multidimensional monitoring data and cannot detect the violation of collaborative constraints, thus making it insensitive to early failures of equipment without significant single-dimensional anomalies and reducing the accuracy of condition monitoring.

[0005] In a first aspect, this application provides a data analysis-based method for monitoring the condition of a thin-film evaporation equipment. The method includes: constructing a monitoring window for arbitrary multidimensional monitoring data; using principal component analysis (PCA) to extract the first principal component direction and the maximum eigenvalue of the multidimensional monitoring data within the monitoring window and the previous adjacent monitoring window, respectively; using the product of the similarity between the two first principal component directions and the sum of the two maximum eigenvalues ​​as the co-stability index of the multidimensional monitoring data; calculating the projection residual of the multidimensional monitoring data along the first principal component direction; constructing a penalty factor for the multidimensional monitoring data, wherein the penalty factor is negatively correlated with the co-stability index and positively correlated with the projection residual; and constructing isolated... The forest model includes: randomly selecting monitoring data and split points within any divisible node of an isolated tree to construct a split descriptor; dividing the multidimensional monitoring data within the divisible node into a first subset and a second subset based on the split descriptor; constructing a splitting factor for the split descriptor based on the difference in penalty factors between the multidimensional monitoring data in the first subset and the second subset; splitting the divisible node based on the split descriptor corresponding to the largest splitting factor until the maximum depth is reached or the number of multidimensional monitoring data within the divisible node reaches the minimum, thus obtaining the isolated forest model; inputting the current multidimensional monitoring data into the isolated forest model, outputting an anomaly score, and determining that the device status is abnormal when the anomaly score is greater than the anomaly threshold.

[0006] The stability index, constructed by summing the first principal component direction and the maximum eigenvalue of multidimensional monitoring data within the monitoring window and adjacent monitoring windows, is used to evaluate the stability of the co-evolution pattern of multidimensional monitoring data across different dimensions. A penalty factor, constructed based on the co-stability index and projection residuals, assesses the degree of deviation of multidimensional monitoring data along the first principal component direction. A segmentation factor, constructed by the difference in penalty factors between multidimensional monitoring data within the first and second subsets, evaluates the segmentation effect of the segmentation descriptor on multidimensional monitoring data within separable nodes. Based on the segmentation descriptor corresponding to the largest segmentation factor, an isolated forest model with good identification of healthy and unhealthy states of thin-film evaporation equipment is established, improving the accuracy of the isolated forest model in monitoring the state of thin-film evaporation equipment.

[0007] Furthermore, the collaborative stability index is calculated as follows: the cosine similarity of the first principal component direction of the covariance matrix corresponding to the monitoring window and the adjacent monitoring window is calculated, the sum of the maximum eigenvalues ​​of the covariance matrix corresponding to the monitoring window and the adjacent monitoring window is calculated, and the product of the absolute value of the cosine similarity and the sum of the maximum eigenvalues ​​is used as the collaborative stability index of the multidimensional monitoring data.

[0008] By constructing a co-stability index by multiplying the absolute value of cosine similarity with the sum of two maximum eigenvalues, the cosine similarity is used to measure the degree of change in the direction of the multidimensional monitoring data manifold within adjacent monitoring windows, while the sum of the maximum eigenvalues ​​is used to measure the dominance strength in the direction of the first principal component. The product of the two can simultaneously reflect the stability and strength of the co-evolution direction, making the co-stability index more sensitive to subtle changes in the co-evolution pattern of multidimensional monitoring data, thereby improving the ability to detect early faults.

[0009] Furthermore, the penalty factor is calculated as follows: the ratio of the projection residual of the multidimensional monitoring data in the direction of the first principal component to the distance of the multidimensional monitoring data to the origin is calculated as the normalized projection residual; the cooperative stability index is subjected to a hyperbolic tangent transformation and subtracted from 1 to obtain the cooperative offset factor; the product of the normalized projection residual and the cooperative offset factor is used as the penalty factor of the multidimensional monitoring data.

[0010] By constructing a penalty factor by multiplying the normalized projection residual with the cooperative offset factor, the normalized projection residual eliminates the amplitude interference caused by the different distances of the monitoring data from the origin, reflecting the relative offset of the multidimensional monitoring data in the direction of the first principal component. The cooperative offset factor introduces the cooperative stability index into the calculation of the penalty factor, enabling the penalty factor to simultaneously perceive the destruction of single-point deviation and the overall cooperative pattern. Compared with the traditional method that only relies on single-dimensional deviation, it can more comprehensively characterize the degree of anomaly of multidimensional monitoring data.

[0011] Furthermore, the projection residual is calculated as follows: the projection point of the multidimensional monitoring data on the first principal component direction of the covariance matrix corresponding to the monitoring window is calculated, and the distance between the multidimensional monitoring data and its projection point is taken as the projection residual.

[0012] Furthermore, the segmentation factor is calculated as follows: the mean of the penalty factors of all multidimensional monitoring data in the first subset and the second subset are calculated respectively, and the absolute value of the difference between the two means is taken as the segmentation factor of the segmentation descriptor.

[0013] By calculating the absolute value of the difference between the mean values ​​of the penalty factors of the multidimensional monitoring data in the first and second subsets as the segmentation factor, the segmentation descriptor can be directly quantified to distinguish between healthy and unhealthy data within the separable nodes. The larger the segmentation factor, the better the segmentation descriptor can separate data with significant differences in penalty factors into different subsets, thus providing an effective evaluation basis for subsequent optimal segmentation.

[0014] Furthermore, the step of randomly selecting monitoring data and segmentation points to construct segmentation descriptors includes: randomly generating a preset number of candidate segmentation descriptors within any divisible node of the isolated tree, calculating the segmentation factor of each candidate segmentation descriptor, and selecting the segmentation descriptor with the largest segmentation factor to segment the divisible node.

[0015] By randomly generating multiple candidate segmentation descriptors within the divisible nodes and selecting the one with the largest segmentation factor for segmentation, the traditional random single-segmentation method of isolated forests is changed. This ensures that each split of the isolated tree proceeds in a direction that can distinguish between healthy and unhealthy states to the greatest extent possible. As a result, the constructed isolated forest model is more sensitive to anomalies caused by the disruption of the collaborative mode of multidimensional monitoring data, thereby improving the model's accuracy in identifying early faults in thin film evaporation equipment.

[0016] Furthermore, the adjacent monitoring windows are adjacent to each other and do not overlap; by dividing the windows without overlap, the multidimensional monitoring data within the adjacent monitoring windows are completely independent, avoiding the decrease in the sensitivity of the collaborative stability index to changes in the collaborative mode caused by high overlap of window data.

[0017] Furthermore, the method also includes: constructing the isolated forest model using multi-dimensional monitoring data of the equipment's historical normal operating status; during the real-time monitoring phase, inputting the penalty factor of the real-time collected multi-dimensional monitoring data into the isolated forest model and outputting anomaly scores; by separating the model construction phase from the real-time monitoring phase, ensuring that the isolated forest model learns the collaborative mode benchmark when the equipment is operating normally.

[0018] Furthermore, the monitoring method also includes: preprocessing the multidimensional monitoring data before constructing the monitoring window. The preprocessing includes: interpolating and filling missing data to eliminate missing values ​​caused by sensor interruption; normalizing the monitoring data of each dimension to eliminate the dimensional differences of monitoring quantities in different physical domains, so that the data of each dimension are comparable when constructing the covariance matrix, and ensuring the effectiveness of the extraction of the first principal component direction and the calculation of the projection residual.

[0019] In a second aspect, this application also provides a data analysis-based thin film evaporation equipment status monitoring system, including a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, a data analysis-based thin film evaporation equipment status monitoring method according to the first aspect of this application is implemented.

[0020] This application has the following technical effects: The cooperative stability index, constructed by summing the first principal component direction and the maximum eigenvalue of multidimensional monitoring data within the monitoring window and adjacent monitoring windows, can quantify the cooperative evolution pattern of multidimensional monitoring data from two dimensions: manifold direction and dominance strength. This enables the model to perceive cooperative constraint violations that traditional methods cannot identify. The penalty factor, constructed based on the cooperative stability index and projected residuals, can integrate cooperative stability information into the splitting decision of the isolation tree, guiding it to preferentially split in directions that can distinguish between healthy and unhealthy cooperative patterns. This results in stronger sensitivity to early faults that show no significant single-dimensional anomalies but only changes in cooperative correlation structure. The multidimensional monitoring data within separable nodes is divided into a first subset and a second subset using a segmentation descriptor. A segmentation factor, constructed based on the difference in penalty factors between the first and second subsets, can evaluate the segmentation effect of the segmentation descriptor on the multidimensional monitoring data within separable nodes. By segmenting the separable nodes based on the segmentation descriptor corresponding to the largest segmentation factor, an isolation forest model with good identification of healthy and unhealthy states of thin-film evaporation equipment is established, improving the accuracy of the isolation forest model in monitoring the condition of thin-film evaporation equipment. Attached Figure Description

[0021] Figure 1 This is a flowchart of a data analysis-based method for monitoring the condition of a thin film evaporation equipment according to an embodiment of this application.

[0022] Figure 2 This is a structural block diagram of a thin film evaporation equipment status monitoring system based on data analysis, according to an embodiment of this application. Detailed Implementation

[0023] Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of this application.

[0024] The first aspect of this application provides a method for monitoring the condition of thin film evaporation equipment based on data analysis. Figure 1 This is a flowchart of a data analysis-based method for monitoring the status of a thin film evaporation equipment according to an embodiment of this application. The specific implementation process of this method will be described in detail below.

[0025] S101: Acquire multidimensional monitoring data.

[0026] In this embodiment, to analyze the operating status of the thin film evaporation equipment, it is necessary to collect multi-dimensional monitoring data of historical operating status, including: temperature data of the evaporation source and substrate surface collected by a temperature sensor, heating current and voltage data of the evaporation source collected by a Hall effect sensor, and flow rate data of the cooling water circuit collected by a flow meter. In this embodiment, the data collection interval is 1 second.

[0027] It should be noted that during the data collection process, there may be interference with the sensors or communication interruptions, which may lead to a decrease in data quality. Therefore, the above data needs to be preprocessed, including: using Lagrange interpolation to fill in missing values; and using the maximum-minimum normalization method to map the multidimensional monitoring data to the range of [0,1] to remove the influence of different dimensions.

[0028] Thus, multi-dimensional monitoring data of the pre-treated thin film evaporation equipment were obtained.

[0029] S102: Construct a monitoring window for arbitrary multidimensional monitoring data; use principal component analysis algorithm to extract the first principal component direction and the maximum eigenvalue of the multidimensional monitoring data in the monitoring window and the previous adjacent monitoring window respectively, and use the product of the similarity of the two first principal component directions and the sum of the two maximum eigenvalues ​​as the co-stability index of the multidimensional monitoring data.

[0030] In this embodiment, when the thin-film evaporation equipment is in a healthy operating state at a certain process stage, such as during a stable evaporation period, the monitoring data from different dimensions will exhibit a manifold structure. This manifold structure reflects the inherent and stable cooperative constraint relationship of the monitoring data from different dimensions. When the equipment is in a healthy state, the main direction of this manifold structure should remain highly stable for a short period of time; however, when the equipment experiences a potential fault, such as heater aging leading to a slower thermal response, it will first cause a slight change in the cooperative relationship between heating current and temperature, which will then manifest as unstable oscillation or drift of the manifold structure.

[0031] Therefore, this step constructs a collaborative stability index to reflect the stability of collaborative evolution patterns of any multidimensional monitoring data across different dimensions.

[0032] Specifically, for any multidimensional monitoring data, a monitoring window of length N is constructed, with the multidimensional monitoring data as the endpoint. Adjacent, non-overlapping monitoring windows before this monitoring window are designated as adjacent monitoring windows. The matrix formed by the multidimensional monitoring data within the monitoring window is designated as the monitoring matrix. The covariance matrix of the monitoring matrix is ​​obtained. This covariance matrix is ​​used as input to a principal component analysis (PCA) algorithm, which outputs the eigenvalues ​​and corresponding eigenvectors of the monitoring matrix. The eigenvectors represent the manifold changes of the monitoring data across multiple dimensions within the monitoring window. The eigenvector corresponding to the largest eigenvalue is taken as the first principal component direction, i.e., the main manifold change direction of the monitoring data across multiple dimensions within the monitoring window. The same steps are performed on adjacent monitoring windows to obtain the first principal component direction and the largest eigenvalue of the monitoring matrix corresponding to the adjacent monitoring windows. In this embodiment, the length N is 60.

[0033] Based on the above characteristics, a collaborative stability index for the multidimensional monitoring data is constructed, and the calculation method is as follows: In the formula This represents the collaborative stability index of the multidimensional monitoring data. and These represent the directions of the first principal components of the covariance matrices corresponding to the monitoring window and adjacent monitoring windows, respectively. This represents the cosine similarity along the direction of the first principal component of the monitoring matrix corresponding to the monitoring window and adjacent monitoring windows. and These represent the largest eigenvalues ​​of the covariance matrices corresponding to the monitoring window and the adjacent monitoring window, respectively.

[0034] In the formula This reflects the degree of change in the manifold direction of multidimensional monitoring data within the monitoring window and adjacent monitoring windows. This reflects the strength of the first principal component direction, i.e., the dominant strength of the multidimensional monitoring data along the manifold direction. When the equipment is operating stably, the collaborative pattern of the multidimensional monitoring data within the monitoring window and adjacent monitoring windows does not change significantly, and the corresponding manifold structure remains basically unchanged. The absolute value of the cosine similarity calculated from the first principal component direction of the covariance matrix of the monitoring window and adjacent monitoring windows is large, close to 1, and the dominant strength of the multidimensional monitoring data along the manifold direction is strong, with a large corresponding eigenvalue, resulting in a large calculated collaborative stability index. When the collaborative pattern changes, the manifold structure of the monitoring window and adjacent monitoring windows differs, and the absolute value of the cosine similarity calculated from the first principal component direction of the covariance matrix of the monitoring window and adjacent monitoring windows is small. This also leads to a weakening of the dominant strength of the multidimensional monitoring data along the manifold direction, a smaller corresponding eigenvalue, and a smaller calculated collaborative stability index.

[0035] S103: Calculate the projection residual of the multidimensional monitoring data in the direction of the first principal component, and construct the penalty factor of the multidimensional monitoring data. The penalty factor is negatively correlated with the co-stability index and positively correlated with the projection residual.

[0036] In this embodiment, when the thin-film evaporation equipment is in a healthy and stable process stage, based on the heat conduction equation and power control logic theory, the multidimensional monitoring data not only has a stable manifold structure but also a certain constraint relationship. When the equipment experiences an early failure, such as poor contact between the evaporation source and the crucible leading to an increase in local thermal resistance, or a minor leak in the vacuum system, this local constraint will be disrupted. This will result in single or a small number of continuous multidimensional monitoring data deviating from the direction of the first principal component. Based on the reconstruction error theory of principal component analysis, the degree of deviation from the direction of the first principal component can be measured by the distance between the projection points of the multidimensional monitoring data in the direction of the first principal component.

[0037] Therefore, this step constructs a penalty factor for the multidimensional monitoring data to reflect the degree of deviation of the multidimensional monitoring data in the principal component direction.

[0038] Specifically, for the monitoring matrix obtained from the multidimensional monitoring data, the distance between the multidimensional monitoring data and its projection points in the first principal component direction of the covariance matrix of the monitoring matrix is ​​calculated, i.e., the projection residual.

[0039] Based on the above characteristics, a penalty factor for the multidimensional monitoring data is constructed, and the calculation method is as follows: In the formula This represents the penalty factor for the multidimensional monitoring data. This represents the projection residuals of the multidimensional monitoring data along the principal component directions. This represents the distance from the multidimensional monitoring data point to the origin. This represents a preset hyperparameter used to avoid calculation failures due to a denominator of zero. In this embodiment, the value is 1. This represents the normalized projection residual. This represents the hyperbolic tangent function, used to normalize the data within the parentheses. This represents the collaborative stability index of multidimensional monitoring data.

[0040] In the formula This reflects the relative shift of the multidimensional monitoring data along the direction of the first principal component. This reflects the overall offset of the multidimensional monitoring data within the monitoring window, based on and The constructed penalty factor can reflect the degree of deviation of the multidimensional monitoring data in the direction of the principal components.

[0041] When the equipment exhibits unhealthy deviation characteristics, the relative and overall deviation of the multidimensional monitoring data increases, resulting in a larger calculated penalty factor. When the equipment is in a healthy operating state, the relative and overall deviation of the multidimensional monitoring data are both small, resulting in a smaller calculated penalty factor.

[0042] S104: Within any separable node of the isolated tree, randomly select monitoring data and split points to construct a split descriptor. Based on the split descriptor, divide the multidimensional monitoring data within the separable node into a first subset and a second subset. Construct the splitting factor of the split descriptor according to the difference in the penalty factor of the multidimensional monitoring data within the first subset and the second subset.

[0043] In this embodiment, in the traditional isolated forest model, the isolated forest model consists of multiple isolated trees. When constructing any isolated tree, within any divisible node of the isolated tree, the divisible node is divided by randomly selecting dimensions and split points until the maximum depth of the isolated tree is reached or further division is impossible, thus forming an isolated tree. To improve the accuracy of isolated trees in monitoring the status of thin film evaporation equipment, it is necessary to guide the division process of the divisible nodes of the isolated tree, enabling the isolated tree to learn the unhealthy state characteristics of the equipment.

[0044] Therefore, within any divisible node of the isolated tree, dimensions and split points are randomly selected to construct a segmentation descriptor. The segmentation factor of the segmentation descriptor constructed in this step is used to reflect the segmentation effect of the segmentation descriptor on the multidimensional monitoring data within the divisible node.

[0045] Specifically, within any separable node of the isolated tree, the segmentation descriptor constructed by randomly selected monitoring data and the segmentation point is denoted as... ,in This indicates a randomly selected dimension, such as the heating current data of the evaporation source. A randomly selected split point, such as 0.8, can be used to divide the multidimensional monitoring data within the divisible node into two subsets, namely a first subset and a second subset, using a segmentation descriptor. The upper part is less than or equal to the dividing point The set of multidimensional monitoring data constitutes the first subset. In dimension The upper part is greater than the dividing point. The set of multidimensional monitoring data constitutes the second subset. .

[0046] The segmentation factor of the segmentation descriptor is constructed based on the above features, and the calculation method is as follows: In the formula This represents the segmentation factor of the segmentation descriptor. and They represent the first subset and the second subset, respectively. and They represent the first set of elements. The multidimensional monitoring data and the second subset A penalty factor for multidimensional monitoring data. and These represent the number of multidimensional monitoring data in the first subset and the second subset, respectively.

[0047] If the segmentation descriptor can effectively distinguish between the healthy and unhealthy states of the thin-film evaporation equipment, then the penalty factors of the multidimensional monitoring data in the first subset and the second subset obtained after segmentation should show a significant difference, i.e. and The differences are large, resulting in a large calculated segmentation factor; if the segmentation descriptor is poor at distinguishing between the healthy and unhealthy states of the thin film evaporation equipment, the difference in the penalty factor between the multidimensional monitoring data in the first subset and the second subset after segmentation is weak, resulting in a small calculated segmentation factor.

[0048] S105: Segment the separable nodes based on the segmentation descriptor corresponding to the largest segmentation factor until the maximum depth is reached or the number of multidimensional monitoring data in the separable node reaches the minimum, thus obtaining an isolated forest model; input the current multidimensional monitoring data into the isolated forest model, output the anomaly score, and determine that the equipment status is abnormal when the anomaly score is greater than the anomaly threshold.

[0049] In this embodiment, for the divisible node, the segmentation factor of all segmentation descriptors is calculated, and the divisible node is segmented based on the segmentation descriptor corresponding to the largest segmentation factor to obtain the two leaf nodes of the divisible node.

[0050] Repeat the segmentation process for the divisible nodes described above until no further segmentation is possible or the maximum depth of the isolated tree is reached, thus completing the construction of an isolated tree; repeat the construction process for the isolated trees described above until the preset number of isolated trees is reached, forming an isolated forest. In this embodiment, the maximum depth of the isolated trees is set to 10, and the preset number of isolated trees is set to 100.

[0051] The isolated forest model is constructed through the above steps. During the monitoring of the thin film evaporation equipment status, the real-time collected multidimensional monitoring data is input into the isolated forest model, and the anomaly score of the real-time multidimensional monitoring data is output. If the anomaly score is greater than the anomaly threshold, the real-time multidimensional monitoring data is determined to be abnormal; if the anomaly score is less than or equal to the anomaly threshold, the real-time multidimensional monitoring data is determined to be normal, and the determination result of the multidimensional monitoring data is used as the status monitoring result. In this embodiment, the anomaly threshold is set to 0.7.

[0052] According to a second aspect of this application, this application also provides a data analysis-based thin film evaporation equipment condition monitoring system. Figure 2 This is a structural block diagram of a data analysis-based condition monitoring system for thin film evaporation equipment according to an embodiment of this application. Figure 2 As shown, the system 50 includes a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement a data analysis-based method for monitoring the status of a thin-film evaporation equipment according to the first aspect of this application. The system also includes other components well-known to those skilled in the art, such as a communication bus and a communication interface; their configurations and functions are known in the art and will not be described further here.

Claims

1. A method for monitoring the status of thin film evaporation equipment based on data analysis, characterized in that, The monitoring method includes: constructing a monitoring window for arbitrary multidimensional monitoring data; using principal component analysis (PCA) to extract the first principal component direction and the maximum eigenvalue of the multidimensional monitoring data within the monitoring window and the previous adjacent monitoring window, respectively; using the product of the similarity between the two first principal component directions and the sum of the two maximum eigenvalues ​​as the co-stability index of the multidimensional monitoring data; calculating the projection residual of the multidimensional monitoring data along the first principal component direction; constructing a penalty factor for the multidimensional monitoring data, wherein the penalty factor is negatively correlated with the co-stability index and positively correlated with the projection residual; and constructing an isolated forest model based on multiple multidimensional monitoring data, including: in any separable segment of the isolated tree... Within a node, monitoring data and segmentation points are randomly selected to construct a segmentation descriptor. Based on the segmentation descriptor, the multidimensional monitoring data within the separable node is divided into a first subset and a second subset. The segmentation factor of the segmentation descriptor is constructed according to the difference in the penalty factor of the multidimensional monitoring data within the first subset and the second subset. The separable node is segmented based on the segmentation descriptor corresponding to the largest segmentation factor until the maximum depth is reached or the number of multidimensional monitoring data within the separable node reaches the minimum, thus obtaining an isolated forest model. The current multidimensional monitoring data is input into the isolated forest model, and an anomaly score is output. In response to an anomaly score greater than an anomaly threshold, it is determined that the device status is abnormal.

2. The method for monitoring the status of thin film evaporation equipment based on data analysis according to claim 1, characterized in that, The collaborative stability index is calculated as follows: the cosine similarity of the first principal component direction of the covariance matrix corresponding to the monitoring window and the adjacent monitoring window is calculated, the sum of the maximum eigenvalues ​​of the covariance matrix corresponding to the monitoring window and the adjacent monitoring window is calculated, and the product of the absolute value of the cosine similarity and the sum of the maximum eigenvalues ​​is used as the collaborative stability index of the multidimensional monitoring data.

3. The method for monitoring the status of thin film evaporation equipment based on data analysis according to claim 1, characterized in that, The penalty factor is calculated as follows: the ratio of the projection residual of the multidimensional monitoring data in the direction of the first principal component to the distance of the multidimensional monitoring data to the origin is calculated, and the ratio is used as the normalized projection residual. The cooperative stability index is subjected to a hyperbolic tangent transformation and subtracted from 1 to obtain the cooperative offset factor; the product of the normalized projection residual and the cooperative offset factor is used as the penalty factor for the multidimensional monitoring data.

4. The method for monitoring the status of thin film evaporation equipment based on data analysis according to claim 1, characterized in that, The projection residual is calculated as follows: the projection point of the multidimensional monitoring data on the first principal component direction of the covariance matrix corresponding to the monitoring window is calculated, and the distance between the multidimensional monitoring data and its projection point is taken as the projection residual.

5. The method for monitoring the status of thin film evaporation equipment based on data analysis according to claim 1, characterized in that, The segmentation factor is calculated as follows: the mean of the penalty factors of all multidimensional monitoring data in the first subset and the second subset are calculated respectively, and the absolute value of the difference between the two means is taken as the segmentation factor of the segmentation descriptor.

6. The method for monitoring the status of thin film evaporation equipment based on data analysis according to claim 1, characterized in that, The process of randomly selecting monitoring data and segmentation points to construct segmentation descriptors includes: randomly generating a preset number of candidate segmentation descriptors within any divisible node of an isolated tree, calculating the segmentation factor of each candidate segmentation descriptor, and selecting the segmentation descriptor with the largest segmentation factor to segment the divisible node.

7. The method for monitoring the status of thin film evaporation equipment based on data analysis according to claim 1, characterized in that, The adjacent monitoring windows are adjacent to each other and do not overlap; by dividing the windows without overlap, the multi-dimensional monitoring data in the adjacent monitoring windows are completely independent, avoiding the decrease in the sensitivity of the collaborative stability index to changes in the collaborative mode due to high overlap of window data.

8. The method for monitoring the status of thin film evaporation equipment based on data analysis according to claim 1, characterized in that, The method further includes: constructing the isolated forest model using multidimensional monitoring data of the equipment's historical normal operating status; inputting the real-time collected multidimensional monitoring data into the isolated forest model during the real-time monitoring phase and outputting anomaly scores; and ensuring that the isolated forest model learns the collaborative mode benchmark when the equipment is operating normally by separating the model construction phase from the real-time monitoring phase.

9. The method for monitoring the status of thin film evaporation equipment based on data analysis according to claim 1, characterized in that, The monitoring method further includes: preprocessing the multidimensional monitoring data before constructing the monitoring window. The preprocessing includes: interpolating and filling missing data to eliminate missing values ​​caused by sensor interruption; normalizing the monitoring data of each dimension to eliminate the dimensional differences of the monitoring quantities in different physical domains, so that the data of each dimension are comparable when constructing the covariance matrix, and ensuring the effectiveness of the extraction of the first principal component direction and the calculation of the projection residual.

10. A data analysis-based condition monitoring system for thin film evaporation equipment, characterized in that, include: The device includes a processor, a memory, and a communication interface. The memory stores a computer program, which, when executed by the processor, implements a data analysis-based method for monitoring the status of a thin-film evaporation equipment as described in any one of claims 1 to 9.

Images