Semiconductor automated testing method and testing equipment based on artificial intelligence

Abstract

The present invention relates to the field of data processing, and specifically provides a semiconductor automation detection method and detection equipment based on artificial intelligence, by obtaining semiconductor equipment operation monitoring information including s monitoring sequence information arranged in time sequence, s≥2; obtaining exclusive operation state characterization arrays corresponding to the s monitoring sequence information respectively; constructing an uninterrupted operation state characterization array based on the exclusive operation state characterization arrays corresponding to the first r monitoring sequence information in the semiconductor equipment operation monitoring information; constructing a target operation state characterization array corresponding to the s monitoring sequence information based on the uninterrupted operation state characterization array and the exclusive operation state characterization array corresponding to the s monitoring sequence information; obtaining the equipment operation state corresponding to the s monitoring sequence information based on the target operation state characterization array corresponding to the s monitoring sequence information. The present invention can improve the understandability of determining the equipment operation state and the reliability of the equipment operation state.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to an artificial intelligence-based semiconductor automated detection method and detection equipment. Background Art

[0002] Semiconductors are the cornerstone of the modern electronic information industry and are widely used in many fields. With the advancement of technology, the complexity and integration of semiconductor equipment have increased, and its operating status has a significant impact on product quality and production efficiency. Minor faults or fluctuations may lead to increased costs and reduced competitiveness. Therefore, real-time and accurate detection of equipment operating status is a key requirement for industrial development. Traditional detection methods have shortcomings. Single parameter detection based on sensors only focuses on single parameters such as temperature and pressure, which is difficult to fully reflect the status of the equipment, and cannot capture the correlation between parameters, and is prone to miss hidden dangers; in terms of regular maintenance and post-fault repair, regular maintenance lacks dynamic evaluation, which may be excessive or insufficient maintenance, and post-fault repair will cause production interruption and economic losses; manual judgment based on experience is highly subjective and has different standards, making it difficult to deal with new fault modes, and it is impossible to monitor and warn in real time. At present, intelligent detection methods also have shortcomings. The data processing capacity is limited, and it is difficult to efficiently process the increasing number of high-dimensional complex monitoring data; there is a lack of comprehensive representation of the operating status of the equipment, ignoring the correlation between data, and the state identification is prone to misjudgment and omission; the artificial intelligence model has poor comprehensibility, and its black box characteristics make it difficult for enterprises to trust and apply the detection results. Summary of the invention

[0003] In view of this, an embodiment of the present invention provides a semiconductor automatic detection method and detection equipment based on artificial intelligence. The technical solution of the present invention is implemented as follows:

[0004] On the one hand, the present invention provides a semiconductor automation detection method based on artificial intelligence, the method comprising: obtaining semiconductor equipment operation monitoring information, the semiconductor equipment operation monitoring information comprising s monitoring sequence information arranged in time sequence, wherein s≥2; obtaining exclusive operation state characterization arrays corresponding to the s monitoring sequence information in the semiconductor equipment operation monitoring information, wherein the exclusive operation state characterization array is used to characterize the monitoring sequence information; based on the exclusive operation state characterization arrays corresponding to the first r monitoring sequence information in the semiconductor equipment operation monitoring information, constructing an uninterrupted operation state characterization array, wherein the uninterrupted operation state characterization array is used to characterize the semiconductor equipment operation monitoring information, wherein 1≤r<s; based on the uninterrupted operation state characterization array and the exclusive operation state characterization array corresponding to the s-th monitoring sequence information, constructing a target operation state characterization array corresponding to the s-th monitoring sequence information; based on the target operation state characterization array corresponding to the s-th monitoring sequence information, performing state identification on the s-th monitoring sequence information to obtain the equipment operation state corresponding to the s-th monitoring sequence information.

[0005] On the other hand, the present invention provides a detection device, including a memory and a processor, wherein the memory stores a computer program that can be run on the processor, and the processor implements the steps in the above method when executing the program.

[0006] Technical effect of the present invention: The present invention determines the uninterrupted operation state characterization array corresponding to the semiconductor equipment operation monitoring information based on the exclusive operation state characterization arrays corresponding to the first r monitoring sequence information, and then obtains the target operation state characterization array corresponding to the s-th monitoring sequence information based on the uninterrupted operation state characterization array and the exclusive operation state characterization array corresponding to the s-th monitoring sequence information. The target operation state characterization array corresponding to the s-th monitoring sequence information can not only characterize the characteristics of the s-th monitoring sequence information, but also characterize the influence of the first r monitoring sequence information on the s-th monitoring sequence information, and the overall characteristics of the s-th monitoring sequence information in the semiconductor equipment operation monitoring information, thereby increasing the understandability and reliability of the target operation state characterization array, thereby increasing the understandability of the monitoring sequence information state identification and the reliability of the equipment operation state.

[0007] It should be understood that the above general description and the following detailed description are merely exemplary and explanatory, and are not intended to limit the technical solutions of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] The accompanying drawings herein are incorporated into and constitute a part of the specification. These drawings illustrate embodiments consistent with the present invention and, together with the specification, are used to explain the technical solutions of the present invention.

[0009] Figure 1 A schematic diagram of the implementation flow of an artificial intelligence-based semiconductor automated detection method provided in an embodiment of the present invention.

[0010] Figure 2 A hardware entity schematic diagram of a detection device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0011] In order to make the purpose, technical solutions and advantages of the present invention clearer, the technical solutions of the present invention are further elaborated in detail below in conjunction with the drawings and embodiments. The described embodiments should not be regarded as limiting the present invention. All other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present invention.

[0012] The embodiment of the present invention provides a semiconductor automated detection method based on artificial intelligence, which can be executed by a processor of a detection device, wherein the detection device can refer to a device with data processing capabilities such as a server, a laptop, a tablet computer, and a desktop computer.

[0013] Figure 1 A schematic diagram of an implementation flow of a semiconductor automated detection method based on artificial intelligence provided by an embodiment of the present invention is shown in FIG. Figure 1 As shown, the method includes:

[0014] Step S100: obtaining semiconductor equipment operation monitoring information, where the semiconductor equipment operation monitoring information includes s monitoring sequence information arranged in time sequence, where s≥2.

[0015] Semiconductor equipment operation monitoring information covers various relevant data generated by semiconductor equipment during operation. These data can come from different components and different operation stages of the equipment, and are time-series, that is, arranged in chronological order. Monitoring sequence information is a subset of the operation monitoring information, and each monitoring sequence information represents the operation of the semiconductor equipment in a specific time period. For example, for a semiconductor chip manufacturing equipment, its operation monitoring information may include the changes in the equipment's temperature, pressure, current, voltage and other parameters over time. And each monitoring sequence information may be a specific numerical record of these parameters of the equipment within a certain hour.

[0016] There are many ways for the testing equipment to obtain the operation monitoring information of the semiconductor equipment. One technical means is to collect data through sensors. The testing equipment can install various types of sensors at the key parts of the semiconductor equipment, such as temperature sensors, pressure sensors, current sensors, etc. These sensors can monitor the operating parameters of the equipment in real time and transmit the collected data to the testing equipment. The testing equipment can receive this data by wired or wireless means, and store and process it. Another technical means is to use the monitoring system that comes with the equipment. Many semiconductor devices are equipped with their own monitoring systems, which can record the operating status and related parameters of the equipment in real time. The testing equipment can interact with these monitoring systems to obtain the required operation monitoring information. For example, the testing equipment can be connected to the monitoring system of the semiconductor equipment through a network interface and read the operation data stored in the system according to a certain protocol.

[0017] The semiconductor equipment operation monitoring information obtained by the detection equipment can be represented by a mathematical model. Assume that the monitoring sequence information is represented by a vector, and each monitoring sequence information vector contains multiple elements, each element representing the value of a monitoring parameter. For example, a monitoring sequence information vector can be represented as: ;in represents the i-th monitoring sequence information vector, represents the jth element in the i-th monitoring sequence information vector, and n represents the number of monitoring parameters.

[0018] The semiconductor equipment operation monitoring information can be expressed as a matrix M, where each row of the matrix represents a monitoring sequence information vector, and the number of rows in the matrix is ​​s and the number of columns is n. That is: ;

[0019] After obtaining the operation monitoring information, the detection equipment needs to store and manage the information. Database technology can be used to store the operation monitoring information in the database to facilitate subsequent query and analysis. The database can be organized according to the temporal nature of the monitoring sequence information, for example, the monitoring sequence information can be sorted and stored in chronological order to facilitate rapid retrieval and statistical analysis.

[0020] Step S200: obtaining exclusive operation state characterization arrays corresponding to s monitoring sequence information in the semiconductor device operation monitoring information, wherein the exclusive operation state characterization array is used to characterize the monitoring sequence information.

[0021] The exclusive operation status characterization array is an array that can represent the characteristics of the monitoring sequence information. It can simplify and abstract the complex monitoring sequence information and extract the key characteristic information, so as to more conveniently describe and analyze the operation status of the semiconductor equipment. For example, for the monitoring sequence information of a semiconductor chip manufacturing equipment, it contains the changes of multiple parameters such as the temperature, pressure, and current of the equipment over time. By constructing an exclusive operation status characterization array, the detection equipment can convert these complex parameter information into a representative array that can reflect the overall operation status of the equipment during the time period represented by the monitoring sequence.

[0022] The process of obtaining a dedicated operating status characterization array for a detection device involves multiple steps. First, the detection device needs to preprocess the semiconductor device operation monitoring information to improve the quality and availability of the data. The purpose of preprocessing is to remove noise, outliers, missing values, etc. in the data to make the data more accurate and complete. One preprocessing technique is convolution median filtering, which processes the data through a convolution window and uses the median of the data in the window as the processed data, thereby effectively removing noise and outliers. Assuming that the data acquisition timestamp corresponding to the convolution window is A (A is an odd number), for each data acquisition timestamp in the semiconductor device operation monitoring information, the timestamp is used as the center, and the monitoring information corresponding to the A data acquisition timestamps is obtained in the monitoring information, and the information to be processed is constructed, and then the median monitoring information in the information to be processed is used as the preprocessed information corresponding to the data acquisition timestamp. By performing such processing on all data acquisition timestamps, the detection device can obtain the preprocessed semiconductor device operation monitoring information.

[0023] Next, the detection device obtains the coded array representations corresponding to the s monitoring sequence information based on the preprocessed semiconductor device operation monitoring information. The coded array representation is the result of digitally encoding the monitoring sequence information. It can convert various characteristic information in the monitoring sequence information into digital form to facilitate subsequent calculations and analysis. The detection device can construct a corresponding coded array based on the data characteristics and time sequence in the monitoring sequence information through a specific coding algorithm. For example, for a monitoring sequence information containing temperature and pressure data, the detection device can encode it into a two-dimensional array based on the numerical range and change trend of the temperature and pressure. Each element in the array represents the characteristic values ​​of temperature and pressure at different times.

[0024] After obtaining the coded array representation, the detection device needs to further process it to construct an exclusive operating status representation array. The detection device can consider the importance of each part in the monitoring sequence information and perform weighted integration of the coded array representation. By calculating the importance coefficient of each part, the detection device can perform weighted summation of the coded array representation to obtain a comprehensive exclusive operating status representation array. For example, for a monitoring sequence information, in which the temperature data has a greater impact on the operating status of the equipment, the detection device can assign a higher importance coefficient to the coded array representation corresponding to the temperature data, so that when constructing the exclusive operating status representation array, the characteristic information of the temperature data can be more fully reflected.

[0025] The detection device can also consider the relationship between the various parts of the monitoring sequence information and further integrate the coded array representation. For example, the detection device can calculate the element-by-element product between the coded array representation of the sub-monitoring sequence information and the coded array representation of other related sub-monitoring sequence information, and then integrate these products to obtain an array that can reflect the relationship between the sub-monitoring sequence information. This array is further integrated with the array obtained by weighted integration to finally obtain an exclusive operating status representation array. This integration method can enable the exclusive operating status representation array to more comprehensively reflect the characteristics of the monitoring sequence information and improve the ability to characterize the operating status of the equipment.

[0026] Step S300: constructing an uninterrupted operation state characterization array based on the exclusive operation state characterization arrays corresponding to the first r monitoring sequence information in the semiconductor device operation monitoring information, wherein 1≤r<s.

[0027] The exclusive operation status characterization array is a unique abstract representation of the characteristics of each monitoring sequence information. It contains the operating characteristics of the semiconductor equipment reflected by the monitoring sequence information in a specific time period. The uninterrupted operation status characterization array is a synthesis and integration of the characteristics of the first r monitoring sequence information. It can reflect the continuous operation status of the semiconductor equipment in a relatively long period of time. For example, for a semiconductor packaging equipment, the first r monitoring sequence information may record the operating parameters of the equipment in different batches of production, such as temperature, pressure, speed, etc. The exclusive operation status characterization array corresponding to each monitoring sequence information reflects the operating characteristics of the equipment in the batch production process, and the uninterrupted operation status characterization array integrates the overall operation characteristics of the equipment in these r batches of production processes.

[0028] The process of constructing an uninterrupted operating status characterization array for the detection equipment involves multiple key links. First, the detection equipment needs to determine the importance coefficients corresponding to the first r monitoring sequence information. The importance coefficient reflects the relative importance of each monitoring sequence information in constructing the uninterrupted operating status characterization array. Different monitoring sequence information may have different degrees of influence on the overall operating status of the semiconductor equipment, so it is necessary to assign a corresponding importance coefficient to each monitoring sequence information according to the actual situation. The detection equipment can use a variety of methods to determine the importance coefficient, such as a time-based weighted method. The recent monitoring sequence information may have a greater impact on the current equipment operating status, so a higher importance coefficient can be assigned; a machine learning algorithm can also be used to learn the relationship between each monitoring sequence information and the overall operating status of the equipment through a training model, thereby automatically determining the importance coefficient.

[0029] Assume that the exclusive operation status representation array corresponding to the first r monitoring sequence information is , the corresponding importance coefficient is , and satisfies The importance coefficient of the detection equipment can be calculated by the following formula: ;in, Is a dedicated running state representation array A i It can be defined according to specific needs. For example, you can define A i The weighted sum of some key features in A i Characterizes the array's similarity to known normal operating conditions.

[0030] After determining the importance coefficient, the detection device needs to map the exclusive operating status representation arrays corresponding to the first r monitoring sequence information. The purpose of the mapping transformation is to convert the exclusive operating status representation arrays of different dimensions and ranges into a unified form for subsequent integration operations. The detection device can use linear mapping, nonlinear mapping and other methods for transformation. For example, for a high-dimensional exclusive operating status representation array, the detection device can reduce its dimension to a lower dimension through the principal component analysis (PCA) method while retaining the main feature information.

[0031] Assume that after the mapping transformation, the mapping transformation arrays corresponding to the first r monitoring sequence information are B1, B2,…, B r The detection device can express the mapping transformation process through the following formula: i = T(A i ), where B i To map the transformation array, T is a mapping transformation function that can be defined according to the specific transformation method.

[0032] Finally, the detection device integrates the mapping transformation array based on the importance coefficients corresponding to the first r monitoring sequence information to obtain the uninterrupted operation state representation array. The integration method can adopt a weighted summation method, and each mapping transformation array is weighted according to its corresponding importance coefficient and then added to obtain the final uninterrupted operation state representation array.

[0033] Assuming that the uninterrupted operation state characterization array is U, it can be calculated by the following formula: .

[0034] In practical applications, the detection equipment also needs to consider the real-time and accuracy of the data. Since the operating status of the semiconductor device may change at any time, the detection equipment needs to update the previous r monitoring sequence information in a timely manner and recalculate the uninterrupted operating status characterization array. At the same time, the detection equipment also needs to verify and evaluate the constructed uninterrupted operating status characterization array to ensure that it can accurately reflect the overall operating status of the semiconductor device. The detection equipment can compare the constructed uninterrupted operating status characterization array with the known normal operating status characterization array and calculate the similarity or error rate between the two. If the similarity is high or the error rate is low, it means that the constructed uninterrupted operating status characterization array can better characterize the operating status of the semiconductor device; otherwise, the construction process needs to be adjusted and optimized.

[0035] Step S400: constructing a target running state representation array corresponding to the s-th monitoring sequence information based on the uninterrupted running state representation array and the exclusive running state representation array corresponding to the s-th monitoring sequence information.

[0036] The uninterrupted operation status characterization array is constructed by the detection equipment based on the exclusive operation status characterization arrays corresponding to the first r monitoring sequence information in the semiconductor equipment operation monitoring information. It represents the continuous operation status characteristics of the semiconductor equipment in a relatively long period of time. The exclusive operation status characterization array corresponding to the s-th monitoring sequence information focuses on the characteristics of the s-th monitoring sequence information itself, reflecting the operation status of the semiconductor equipment in this specific time period. The target operation status characterization array combines these two aspects of information, taking into account the historical operation status of the equipment and highlighting the characteristics of the current s-th monitoring sequence.

[0037] The detection device can use a variety of technical means to construct the target operation status representation array. One way is to combine two arrays. The combination method can be selected according to specific needs, such as splicing, that is, connecting the uninterrupted operation status representation array and the exclusive operation status representation array corresponding to the sth monitoring sequence information in sequence.

[0038] Another technical means is weighted combination. The detection device can assign different weights to the uninterrupted operation status representation array and the exclusive operation status representation array corresponding to the sth monitoring sequence information according to different application scenarios and needs, and then perform weighted summation of the two arrays. The detection device can automatically learn and determine the optimal values ​​of these two weights through machine learning algorithms, such as neural networks. By training the neural network model, the target operation status representation array constructed under different weight combinations is used as input, and the actual operation status of the equipment is used as output. The model learns the relationship between weights and operation status, so as to find the weight combination that most accurately reflects the operation status of the equipment.

[0039] Step S500: Based on the target operating state representation array corresponding to the s-th monitoring sequence information, the state of the s-th monitoring sequence information is identified to obtain the device operating state corresponding to the s-th monitoring sequence information.

[0040] The device operating status refers to the working condition of the semiconductor device at a specific moment or time period, such as normal operation, abnormal operation, failure, etc. The detection device analyzes and processes the target operating status representation array and compares it with various known device operating status patterns to determine the device operating status corresponding to the sth monitoring sequence information.

[0041] For example, for a semiconductor chip test device, the target operating state characterization array contains comprehensive information of multiple parameters such as the device's temperature, voltage, and test pass rate. Based on this information, the detection device determines whether the device is in a normal and efficient test state during the test process corresponding to the sth monitoring sequence, or is in an abnormal test state due to factors such as excessive temperature and unstable voltage, or whether a device failure has occurred and the test cannot be performed normally.

[0042] There are many technical means that can be used for state recognition of detection equipment. One method is a classification algorithm based on machine learning. The detection equipment can use a pre-trained classification model and take the target operation state representation array as input. The model will output the corresponding equipment operation state category. Feasible classification algorithms include support vector machine (SVM), decision tree, neural network, etc. Taking support vector machine as an example, the basic idea of ​​support vector machine is to find an optimal hyperplane in the feature space to separate samples of different categories. Another technical means is a rule-based recognition method. The detection equipment can formulate a series of rules to judge the operation state of the equipment based on the working principle and empirical knowledge of semiconductor equipment. For example, for a semiconductor etching equipment, if the etching rate in the target operation state representation array is lower than a certain set threshold, and the etching temperature is higher than another set threshold, then the detection equipment can judge that the equipment is in an abnormal operation state according to the preset rules. The detection equipment can also adopt a recognition method based on similarity. The detection equipment pre-stores standard target operation state representation arrays under various known equipment operation states. After obtaining the target operation state representation array corresponding to the sth monitoring sequence information, it calculates the similarity between it and each standard array. The feasible similarity calculation methods include cosine similarity, Euclidean distance, etc. The detection device will select the device operation state corresponding to the standard array with the highest similarity to T_s as the device operation state corresponding to the sth monitoring sequence information.

[0043] When performing state recognition, the detection equipment needs to take into account the uncertainty of the data and the influence of noise. Due to factors such as sensor errors and environmental interference, there may be certain noise and errors in the target operation state representation array. The detection equipment can use methods such as data filtering and noise reduction to pre-process the target operation state representation array to improve the quality of the data. For example, median filtering, mean filtering and other methods are used to remove noise.

[0044] The detection equipment can also use a multi-model fusion method to improve the accuracy of state recognition. The detection equipment can use multiple different recognition models at the same time, such as support vector machines, decision trees, and neural networks, to perform state recognition on the sth monitoring sequence information. Then, a comprehensive judgment is made based on the output results of each model, such as voting, and the device operation state with the most votes is selected as the final result.

[0045] The detection equipment needs to verify and evaluate the results of state recognition. The detection equipment can compare the results of state recognition with the actual equipment operation status and calculate the recognition accuracy, recall rate and other indicators. If the recognition accuracy is low, the detection equipment needs to adjust and optimize the recognition model, such as adding training samples and adjusting model parameters.

[0046] As an implementation manner, step S200, obtaining exclusive operation status representation arrays corresponding to s monitoring sequence information in the semiconductor device operation monitoring information, includes:

[0047] Step S210: preprocessing the semiconductor device operation monitoring information to obtain preprocessed semiconductor device operation monitoring information, where the preprocessed semiconductor device operation monitoring information is information after data cleaning is completed;

[0048] Step S220: based on the pre-processed semiconductor device operation monitoring information, obtaining the encoding array representations corresponding to the s monitoring sequence information respectively;

[0049] Step S230: based on the encoding array representations respectively corresponding to the s monitoring sequence information, obtaining the exclusive operation status representation arrays respectively corresponding to the s monitoring sequence information.

[0050] Step S210 requires the detection device to pre-process the semiconductor device operation monitoring information to obtain the pre-processed semiconductor device operation monitoring information, and the pre-processed semiconductor device operation monitoring information is the information after data cleaning. The original semiconductor device operation monitoring information may contain noise, outliers and missing values, etc. These factors will affect the accuracy and reliability of subsequent analysis, so pre-processing is required to improve data quality. The detection device can use the convolution median filtering method for data cleaning. Specifically, the detection device first obtains the convolution window corresponding to the semiconductor device operation monitoring information, and the data acquisition timestamp corresponding to the convolution window is A, and A is an odd number. For each data acquisition timestamp in the semiconductor device operation monitoring information, the timestamp is taken as the center, and the monitoring information corresponding to the A data acquisition timestamps is obtained in the monitoring information, and the information to be processed is constructed, and then the median monitoring information in the information to be processed is used as the pre-processed information corresponding to the data acquisition timestamp. By performing such processing on all data acquisition timestamps, the detection device can obtain the pre-processed semiconductor device operation monitoring information. For example, for the operation monitoring information of a semiconductor lithography device, which contains the light intensity data of the device, due to the interference of the sensor, the light intensity data at certain moments may have abnormal peaks or valleys. The detection device can effectively remove these abnormal values ​​through the convolution median filtering method, making the light intensity data smoother and more accurate.

[0051] Step S220 requires the detection device to obtain the coded array representations corresponding to the s monitoring sequence information based on the pre-processed semiconductor device operation monitoring information. The coded array representation is the result of digitally encoding the monitoring sequence information, which can convert various characteristic information in the monitoring sequence information into digital form, which is convenient for subsequent calculation and analysis. The detection device can construct the corresponding coded array using a specific coding algorithm according to the data characteristics and time sequence in the monitoring sequence information. For example, for a monitoring sequence information containing temperature and pressure data, the detection device can encode it into a two-dimensional array according to the numerical range and change trend of the temperature and pressure, and each element in the array represents the characteristic value of the temperature and pressure at different times. The detection device can also consider the key time points in the monitoring sequence information, such as the time point corresponding to the peak value of the monitoring information in the data acquisition timestamp, and encode it as a reference point. This can highlight the important features in the monitoring sequence information and make the coded array more representative.

[0052] Step S230 requires the detection device to obtain the exclusive operating state characterization arrays corresponding to the s monitoring sequence information based on the encoding array representations corresponding to the s monitoring sequence information. The exclusive operating state characterization array is a further abstraction and integration of the characteristics of the monitoring sequence information, which can more comprehensively and accurately reflect the operating state of the semiconductor device represented by the monitoring sequence information. The detection device can consider the importance of each part in the monitoring sequence information and perform weighted integration on the encoding array representation. By calculating the importance coefficient of each part, the detection device can perform weighted summation on the encoding array representation to obtain a comprehensive exclusive operating state characterization array. For example, for a monitoring sequence information, in which the temperature data has a greater impact on the operating state of the device, the detection device can assign a higher importance coefficient to the encoding array representation corresponding to the temperature data, so that when constructing the exclusive operating state characterization array, the characteristic information of the temperature data can be more fully reflected. The detection device can also consider the relationship between the various parts in the monitoring sequence information and further integrate the encoding array representation. For example, the detection device can calculate the element-by-element product between the encoded array representation of the sub-monitoring sequence information and the encoded array representation of other related sub-monitoring sequence information, and then integrate these products to obtain an array that can reflect the relationship between the sub-monitoring sequence information. This array is further integrated with the array obtained by weighted integration to finally obtain an exclusive operating status representation array. This integration method can make the exclusive operating status representation array more comprehensively reflect the characteristics of the monitoring sequence information and improve the ability to characterize the operating status of the equipment.

[0053] When implementing step S210, the detection device determines the size A of the convolution window. The value of A will affect the effect of data cleaning. If the value of A is too small, noise and outliers may not be effectively removed; if the value of A is too large, some normal fluctuation information may be smoothed out. The detection device can determine the appropriate value of A through experiments and experience. Assume that the original semiconductor device operation monitoring information is a one-dimensional array , the convolution window size is A, and for the i-th data acquisition timestamp, the information to be processed is , after preprocessing information x i 'For W i By performing this process on all i, we get the preprocessed array .

[0054] When implementing step S220, the detection device needs to select a suitable encoding algorithm. Different monitoring sequence information may require different encoding methods. For example, for monitoring sequence information with periodic changes, Fourier transform can be used for encoding, converting the time domain signal into a frequency domain signal, and extracting its frequency characteristics. For monitoring sequence information with obvious trend changes, a polynomial fitting method can be used for encoding, and the coefficients of the polynomial are used to represent the trend characteristics of the monitoring sequence.

[0055] When implementing step S230, it is critical that the detection device determines the importance coefficient. The detection device can use a machine learning algorithm to calculate the importance coefficient. For example, by training a neural network model, taking the monitoring sequence information as input and the operating status of the device as output, the model learns the relationship between each part of the monitoring sequence information and the operating status of the device, thereby automatically determining the importance coefficient of each part. Assuming that the monitoring sequence information contains k sub-monitoring sequence information, the corresponding encoding array representations are respectively The importance coefficients are , then the first integrated coding array obtained by weighted integration is It can be expressed as For the integration of the relationship between sub-monitoring sequence information, it is assumed that the sub-monitoring sequence information Monitor sequence information with other targets The element-wise products between are , these products are integrated to obtain the sub-monitoring sequence information The corresponding second integrated encoding array , and then integrate the second integrated coding arrays of all sub-monitoring sequence information to obtain the third integrated coding array I2, and finally integrate the first integrated coding array I1 and the third integrated coding array I2 to obtain the exclusive operation status representation array R.

[0056] As an implementation mode, the monitoring sequence information includes monitoring information corresponding to a plurality of uninterrupted data acquisition timestamps respectively; step S220, based on the pre-processed semiconductor device operation monitoring information, obtains the encoding array representation corresponding to the s monitoring sequence information respectively, including:

[0057] Step S221: based on the pre-processed semiconductor device operation monitoring information, obtaining target data acquisition timestamps corresponding to the s monitoring sequence information respectively, where the target data acquisition timestamp is a data acquisition timestamp corresponding to the peak value of the monitoring information among the multiple data acquisition timestamps corresponding to the monitoring sequence information;

[0058] Step S222: for each monitoring sequence information, using the target data collection timestamp corresponding to the monitoring sequence information as a reference timestamp, forwardly acquiring the data collection timestamp of the first number, and reversely acquiring the data collection timestamp of the second number, to obtain a data collection timestamp set corresponding to the monitoring sequence information;

[0059] Step S223: based on the monitoring information corresponding to each data collection timestamp in the data collection timestamp set corresponding to the monitoring sequence information, obtain the encoding array representation corresponding to the monitoring sequence information.

[0060] In step S221, the detection device obtains the target data acquisition timestamps corresponding to the s monitoring sequence information respectively based on the pre-processed semiconductor device operation monitoring information, and the target data acquisition timestamp is the data acquisition timestamp corresponding to the peak value of the monitoring information in the multiple data acquisition timestamps corresponding to the monitoring sequence information. During the operation of the semiconductor device, the peak value of the monitoring information often represents the key change point of the equipment operation state, which may correspond to a special operation mode, abnormal situation or important production stage of the equipment. By determining the target data acquisition timestamp, the detection device can focus on the key information in the monitoring sequence, so as to more accurately extract the features of the monitoring sequence. For example, for the operation monitoring information of a semiconductor etching device, which includes the parameter of etching rate, when the etching rate peaks, it may mean that the etching process has entered an efficient or abnormal stage. The detection device determines it as the target data acquisition timestamp by finding the timestamp corresponding to the peak value in the etching rate data. The detection device can use a traversal method to find the peak value in the monitoring information. For a monitoring parameter in a monitoring sequence information, the detection device starts from the first data collection timestamp and compares the monitoring values ​​corresponding to adjacent timestamps in sequence. When it is found that the monitoring value corresponding to a timestamp is greater than the monitoring values ​​corresponding to the previous and next timestamps, the timestamp is determined as the timestamp corresponding to the peak value. The detection device can record all such timestamps, and then select the appropriate one as the target data collection timestamp according to specific needs, such as selecting the timestamp corresponding to the peak value with the largest monitoring value.

[0061] In step S222, the detection device uses the target data acquisition timestamp corresponding to the monitoring sequence information as the reference timestamp, forwardly acquires the data acquisition timestamp of the first number, and reversely acquires the data acquisition timestamp of the second number for each monitoring sequence information, and obtains the data acquisition timestamp set corresponding to the monitoring sequence information. The purpose of this step is to select timestamps within a certain range and their corresponding monitoring information around the target data acquisition timestamp, so as to construct a data subset containing key information. The number of forward acquisitions and reverse acquisitions (i.e., the first number and the second number) can be set according to actual conditions, and their values ​​will affect the range and information richness of the data acquisition timestamp set. For example, for the monitoring sequence information of the above-mentioned semiconductor etching equipment, using the timestamp corresponding to the etching rate peak as the reference timestamp, the detection device forwardly acquires 3 timestamps and reversely acquires 2 timestamps, so as to obtain a data acquisition timestamp set containing 6 timestamps. In this way, the detection device can focus on the information around the target data acquisition timestamp and better capture the local features and change trends in the monitoring sequence information.

[0062] In step S223, the detection device obtains the coded array representation corresponding to the monitoring sequence information based on the monitoring information corresponding to each data acquisition timestamp in the data acquisition timestamp set corresponding to the monitoring sequence information. The coded array representation is the result of digitally encoding the monitoring information in the data acquisition timestamp set, which can convert complex monitoring information into a form that is convenient for computer processing and analysis. The detection device can adopt a variety of encoding methods, such as binary encoding, hash encoding, etc., and select a suitable encoding method according to the characteristics of the monitoring information and the needs of subsequent analysis. For example, for a data acquisition timestamp set containing two monitoring parameters, temperature and pressure, the detection device can quantize the values ​​of temperature and pressure, and then binary encode the quantized results to obtain a binary coded array. Assume that the data acquisition timestamp set T set The corresponding monitoring information matrix is ​​S, where each row represents a monitoring information vector corresponding to a timestamp, and each column represents a monitoring parameter. The detection device can define a coding function f to map the monitoring information matrix S to a coding array representation C, that is, C=f(S). The coding function f can be defined according to a specific coding method. For example, for binary coding, f converts each element in the monitoring information matrix into a corresponding binary number, and then combines these binary arrays into one array.

[0063] As an implementation mode, step S223, based on the monitoring information corresponding to each data collection timestamp in the data collection timestamp set corresponding to the monitoring sequence information, obtains the encoding array representation corresponding to the monitoring sequence information, including:

[0064] Step S2231: dividing the monitoring information corresponding to the data collection timestamp set to obtain a plurality of sub-monitoring sequence information, wherein the sub-monitoring sequence information includes the monitoring information corresponding to the data collection timestamp of the third number;

[0065] Step S2232: for each sub-monitoring sequence information, based on the monitoring information corresponding to the sub-monitoring sequence information, obtaining the encoding array representation corresponding to the sub-monitoring sequence information;

[0066] Step S2233: based on the coded array representations respectively corresponding to the plurality of sub-monitoring sequence information, obtain the coded array representation corresponding to the monitoring sequence information.

[0067] In step S2231, the detection device divides the monitoring information corresponding to the data acquisition timestamp set to obtain multiple sub-monitoring sequence information, and the sub-monitoring sequence information includes the monitoring information corresponding to the data acquisition timestamp of the third number. The purpose of dividing the monitoring information into sub-monitoring sequence information is to analyze the local characteristics of the monitoring data more carefully, because the monitoring data in different time periods may have different change patterns and characteristics. By dividing into sub-monitoring sequence information, the detection device can encode and analyze each sub-sequence separately, so as to more accurately capture the overall characteristics of the monitoring sequence information. For example, for the monitoring information corresponding to the data acquisition timestamp set of a semiconductor chip testing device, the parameters such as the test voltage, current and temperature of the chip change over time. The detection device can divide this monitoring information into multiple sub-monitoring sequence information in chronological order, and each sub-monitoring sequence information contains a certain number (i.e., the third number) of monitoring information corresponding to the data acquisition timestamp set. Assume that the monitoring information matrix corresponding to the data acquisition timestamp set is S, in which each row represents a monitoring information vector corresponding to a timestamp, and each column represents a monitoring parameter. The detection device can divide the matrix S into multiple sub-matrices S1, S2,…, S according to the third number k. n , each submatrix S i Contains k rows of data, i.e., monitoring information corresponding to k data collection timestamps. The partitioning method can be sequential partitioning, i.e., starting from the first row of the matrix S, k rows are selected in sequence as a submatrix until the entire matrix S is traversed.

[0068] In step S2232, the detection device obtains the encoding array representation corresponding to each sub-monitoring sequence information based on the monitoring information corresponding to the sub-monitoring sequence information. This step is to encode each sub-monitoring sequence information separately to extract its unique features.

[0069] The detection device can select a suitable encoding method according to the characteristics of the sub-monitoring sequence information. For example, for sub-monitoring sequence information containing continuous monitoring parameters (such as voltage and temperature), the detection device can use a quantization encoding method. First, determine the value range of each monitoring parameter, and then divide the range into several intervals, each interval corresponding to a coding value. For each monitoring value in the sub-monitoring sequence information, the detection device maps it to the corresponding interval and takes the coding value of the interval as the encoding of the monitoring value. Assume that the value range of a monitoring parameter x in the sub-monitoring sequence information is , divide the range into m intervals , each interval corresponds to a code value For a specific value x of the monitoring parameter in the sub-monitoring sequence information i ,if , then x i The encoding value is c j By encoding each monitoring parameter in this way, the detection device can obtain the encoding array representation corresponding to the sub-monitoring sequence information.

[0070] In step S2233, the detection device obtains the coded array representation corresponding to the monitoring sequence information based on the coded array representations corresponding to the multiple sub-monitoring sequence information. This step is to integrate the coded array representations of each sub-monitoring sequence information to obtain a coded array representation that can represent the entire monitoring sequence information. The detection device can use a variety of integration methods, such as splicing, weighted summation, etc.

[0071] As an implementation manner, step S230, based on the encoding array representations corresponding to the s monitoring sequence information, obtains the exclusive operation status representation arrays corresponding to the s monitoring sequence information, including:

[0072] Step S231: for each monitoring sequence information, obtaining the importance coefficient corresponding to each sub-monitoring sequence information in the monitoring sequence information;

[0073] Step S232: based on the importance coefficients corresponding to the sub-monitoring sequence information in the monitoring sequence information, respectively, the encoding array representations corresponding to the sub-monitoring sequence information in the monitoring sequence information are integrated to obtain a first integrated encoding array;

[0074] Step S233: for each sub-monitoring sequence information in the monitoring sequence information, obtaining the element-by-element product between the encoding array representation of the sub-monitoring sequence information and the encoding array representation of each target sub-monitoring sequence information corresponding to the sub-monitoring sequence information; wherein the target sub-monitoring sequence information is the sub-monitoring sequence information after the sub-monitoring sequence information in the monitoring sequence information;

[0075] Step S234: integrating the element-by-element product between the sub-monitoring sequence information and each target sub-monitoring sequence information corresponding to the sub-monitoring sequence information to obtain a second integrated coding array corresponding to the sub-monitoring sequence information;

[0076] Step S235: integrating the second integrated coding arrays corresponding to each sub-monitoring sequence information in the monitoring sequence information to obtain a third integrated coding array;

[0077] Step S236: Integrate the first integrated coding array and the third integrated coding array to obtain an exclusive operation status representation array corresponding to the monitoring sequence information.

[0078] In step S231, the detection device obtains the importance coefficients corresponding to each sub-monitoring sequence information in the monitoring sequence information for each monitoring sequence information. The importance coefficient reflects the relative importance of each sub-monitoring sequence information in the entire monitoring sequence information for characterizing the operating state of the equipment. Different sub-monitoring sequence information may have different degrees of influence on the operating state of the equipment due to factors such as the time period in which it is located and the monitoring parameters contained. For example, for the monitoring sequence information of a semiconductor wafer lithography machine, it contains sub-monitoring sequence information such as the exposure stage and the development stage. The parameters of the exposure stage are crucial to the quality of lithography, so the importance coefficient of the sub-monitoring sequence information corresponding to the exposure stage may be higher; and the development stage may have a smaller direct impact on the final lithography effect than the exposure stage, and the importance coefficient of the corresponding sub-monitoring sequence information may be relatively low. The detection device can use a variety of methods to determine the importance coefficient. One method is based on experience and field knowledge, and professionals assign corresponding importance coefficients to each sub-monitoring sequence information based on the working principle of semiconductor equipment and past experience. Another method is to use machine learning algorithms, such as training a neural network model, taking the monitoring sequence information as input and the actual operating status of the equipment as output, and letting the model learn the relationship between each sub-monitoring sequence information and the equipment operating status, thereby automatically determining the importance coefficient.

[0079] In step S232, the detection device integrates the encoding array representations corresponding to each sub-monitoring sequence information in the monitoring sequence information based on the importance coefficients corresponding to each sub-monitoring sequence information in the monitoring sequence information to obtain a first integrated encoding array. This step is to perform weighted summation on the encoding array representations of each sub-monitoring sequence information to comprehensively consider the importance of each sub-monitoring sequence information. Through weighted integration, the characteristics of important sub-monitoring sequence information can be highlighted while weakening the influence of unimportant sub-monitoring sequence information.

[0080] In step S233, the detection device obtains the element-by-element product between the coded array representation of the sub-monitoring sequence information and the coded array representation of each target sub-monitoring sequence information corresponding to the sub-monitoring sequence information for each sub-monitoring sequence information in the monitoring sequence information. The target sub-monitoring sequence information is the sub-monitoring sequence information following the sub-monitoring sequence information in the monitoring sequence information. The purpose of calculating the element-by-element product is to explore the relationship between the sub-monitoring sequence information, because there may be a certain correlation between different sub-monitoring sequence information, and this correlation may be of great significance to the characterization of the operating status of the equipment. For example, in the semiconductor chip packaging process, the parameters of the previous packaging step may affect the effect of the next packaging step. By calculating the element-by-element product, the interaction between the previous and next steps can be captured.

[0081] In step S234, the detection device integrates the element-by-element product between the sub-monitoring sequence information and each target sub-monitoring sequence information corresponding to the sub-monitoring sequence information to obtain a second integrated coding array corresponding to the sub-monitoring sequence information. The purpose of integrating the element-by-element product is to comprehensively express the relationship between the sub-monitoring sequence information and the subsequent target sub-monitoring sequence information. The detection device can integrate by summing, that is, adding the element-by-element product of the sub-monitoring sequence information and each target sub-monitoring sequence information.

[0082] In step S235, the detection device integrates the second integrated coding arrays corresponding to each sub-monitoring sequence information in the monitoring sequence information to obtain a third integrated coding array. This step is to globally integrate the relationship between each sub-monitoring sequence information and the subsequent sub-monitoring sequence information to obtain an array that comprehensively reflects the relationship between all sub-monitoring sequence information in the monitoring sequence information. The detection device can integrate by summing, that is, adding the second integrated coding arrays corresponding to each sub-monitoring sequence information.

[0083] In step S236, the detection device integrates the first integrated coding array and the third integrated coding array to obtain an exclusive operation status representation array corresponding to the monitoring sequence information. This step is to merge the first integrated coding array that integrates the importance of each sub-monitoring sequence information and the third integrated coding array that reflects the relationship between the sub-monitoring sequence information, so as to obtain an exclusive operation status representation array that takes into account both the characteristics of each sub-monitoring sequence information and the relationship between them. The detection device can be integrated by splicing or weighted summation.

[0084] When implementing step S231, when the detection device uses a machine learning algorithm to determine the importance coefficient, a large amount of training data is required. These training data should contain monitoring sequence information under different equipment operating states and the corresponding actual operating state labels. The detection device can be trained using algorithms such as random forests and support vector machines. For example, when using the random forest algorithm, multiple decision tree models are constructed by sampling and feature selection of the training data multiple times. Each decision tree model evaluates the importance of each sub-monitoring sequence information, and finally the results of all decision trees are combined to obtain the final importance coefficient. At the same time, the detection device also needs to evaluate and verify the trained model to ensure the accuracy of the determined importance coefficient. The cross-validation method can be used to divide the training data into multiple subsets, one part for training the model and the other for verifying the performance of the model.

[0085] When implementing step S232, the detection device needs to ensure that the dimensions of the coding array representations corresponding to each sub-monitoring sequence information are consistent. If the dimensions are inconsistent, the weighted summation result may be inaccurate. The detection device can adopt a dimension adjustment method, such as filling the coding array representation with a lower dimension to make its dimension consistent with other coding array representations. At the same time, the accuracy of the importance coefficient will also affect the quality of the first integrated coding array. If the importance coefficient is determined inaccurately, it may cause important sub-monitoring sequence information features to be weakened, while unimportant sub-monitoring sequence information features are over-emphasized. Therefore, the detection device needs to continuously optimize the method for determining the importance coefficient to improve its accuracy.

[0086] When implementing step S233, the detection device needs to pay attention to the element type and value range of the coded array representation when calculating the element-by-element product. If the elements in the coded array representation contain negative numbers or zeros, the element-by-element product may cause the positive and negative values ​​of the result to change. The detection device can pre-process the coded array representation before calculation, such as normalizing it to unify the value range of the elements to the interval [0, 1] to avoid the impact of this situation. At the same time, for longer monitoring sequence information, the amount of calculation for calculating the element-by-element product will be large, and the detection device can use parallel computing methods to improve computing efficiency.

[0087] When implementing step S234, the detection device may be affected by the larger or smaller individual product results when integrating the element-by-element products by summing. In order to avoid this situation, the detection device can standardize the element-by-element products before summing, for example, calculating the mean and standard deviation of each element-by-element product, and then performing a standardized transformation. This can make the integrated second integrated coding array more stable and reliable. At the same time, the detection device can also consider giving different weights to different element-by-element products to highlight certain important relationships.

[0088] When implementing step S235, when the detection device sums the second integrated coding array corresponding to each sub-monitoring sequence information, it is also necessary to pay attention to the consistency of the dimensions. If the dimensions are inconsistent, the detection device needs to adjust the dimensions. In addition, the detection device can further process the summation result, such as smoothing, to reduce the impact of noise and fluctuations. Smoothing can use methods such as moving average to smooth the elements of the third integrated coding array.

[0089] As an implementation manner, step S210, preprocessing the semiconductor device operation monitoring information to obtain the preprocessed semiconductor device operation monitoring information, includes:

[0090] Step S211: obtaining a convolution window corresponding to the semiconductor device operation monitoring information, wherein the number of data acquisition timestamps corresponding to the convolution window is A, where A is an odd number;

[0091] Step S212: for the first data collection timestamp in the semiconductor device operation monitoring information, taking the first data collection timestamp as the center, obtaining monitoring information corresponding to A data collection timestamps in the semiconductor device operation monitoring information, and constructing information to be processed corresponding to the first data collection timestamp;

[0092] Step S213: using the median monitoring information in the to-be-processed information corresponding to the first data collection timestamp as the pre-processed information corresponding to the first data collection timestamp;

[0093] Step S214: obtaining pre-processed semiconductor device operation monitoring information based on the pre-processed information corresponding to each data acquisition timestamp in the semiconductor device operation monitoring information.

[0094] In step S211, the detection device obtains the convolution window corresponding to the semiconductor equipment operation monitoring information, and the data acquisition timestamp corresponding to the convolution window is A, where A is an odd number. The convolution window is a sliding window used for data processing. In the processing of semiconductor equipment operation monitoring information, it is used to select data within a certain range for specific operations. The selection of the A value will affect the effect of preprocessing. If the A value is too small, the noise may not be effectively removed; if the A value is too large, some normal data fluctuations may be smoothed out. For example, for the operation monitoring information of a semiconductor etching device, it contains data on the variation of the etching rate over time. The detection device can determine the size A of the convolution window based on experience or experiments. Assuming that A=5 is determined, this means that the etching rate data corresponding to 5 data acquisition timestamps will be covered each time the convolution window slides. The detection device can adjust the value of A by presetting or dynamically adjusting it according to the characteristics of the data to achieve the best preprocessing effect.

[0095] In step S212, the detection device acquires monitoring information corresponding to A data acquisition timestamps in the semiconductor device operation monitoring information with the first data acquisition timestamp as the center, and constructs the information to be processed corresponding to the first data acquisition timestamp. The first data acquisition timestamp here can be any timestamp in the monitoring information, and the data of A timestamps is selected with it as the center in order to use the surrounding data information to process the data of the timestamp. For example, for the monitoring information of the etching equipment mentioned above, when a specific data acquisition timestamp t i When the detection device is centered, it will obtain t i-2 , t i-1 , t i , t i+1 , t i+2 (Because A=5) The etching rate data corresponding to these five timestamps are combined into an array, namely t i In this way, the detection device can comprehensively consider the data around the timestamp and reduce the impact of noise and outliers of a single data point.

[0096] In step S213, the detection device uses the median monitoring information in the information to be processed corresponding to the first data acquisition timestamp as the pre-processed information corresponding to the first data acquisition timestamp. The median is the value in the middle of a group of data after being sorted from small to large. Using the median as the pre-processed information can effectively resist abnormal values ​​in the data, because abnormal values ​​are usually values ​​that deviate greatly from the normal data range, while the median is relatively less affected by abnormal values. For example, for the above-mentioned array of information to be processed for etching rate [r i-2 , r i-1 , r i , r i+1 , r i+2 ], the detection device sorts them from small to large. If A=5, the value in the middle is the third value, which is used as t i Corresponding post-processing etching rate. Assuming that the array of information to be processed is [10, 12, 15, 18, 20], after sorting it is still [10, 12, 15, 18, 20], and the median is 15, then the detection device will use 15 as the post-processing etching rate corresponding to the data acquisition timestamp. This median filtering method can make the data smoother and reduce noise interference.

[0097] In step S214, the detection device obtains the preprocessed semiconductor device operation monitoring information based on the preprocessed information corresponding to each data acquisition timestamp in the semiconductor device operation monitoring information. The detection device processes each data acquisition timestamp according to steps S212 and S213 to obtain the preprocessed information corresponding to each timestamp, and then combines these preprocessed information in chronological order to form the preprocessed semiconductor device operation monitoring information. For example, for etching equipment operation monitoring information containing n data acquisition timestamps, the detection device processes each timestamp in turn to obtain n preprocessed etching rate values, and arranges these values ​​in chronological order into a new array, which is the preprocessed etching rate monitoring information.

[0098] When implementing step S211, the detection device may use an experimental method to determine a suitable A value. For example, the detection device may select different A values, preprocess the same set of semiconductor device operation monitoring information, and then select the best A value by evaluating the quality of the preprocessed data. The evaluation index may include the smoothness of the data, the similarity with known normal data, etc.

[0099] When implementing step S212, the detection device needs to handle boundary conditions. When the first data acquisition timestamp is close to the start or end position of the monitoring information, it may not be possible to obtain the monitoring information corresponding to the complete A data acquisition timestamps. For example, when the first data acquisition timestamp is centered, there may not be enough timestamps in front. The detection device can use zero padding or copying boundary values ​​to handle this situation. The detection device can select a suitable processing method according to the specific situation to ensure the preprocessing effect of the boundary data.

[0100] When implementing step S214, the detection device needs to ensure that the order of the preprocessed information is consistent with the time order of the original monitoring information. The detection device can use an index array to record the position of the preprocessed information corresponding to each data acquisition timestamp, and then combine the preprocessed information into new monitoring information according to the order of the index array.

[0101] As an implementation mode, step S300, based on the exclusive operation status characterization arrays corresponding to the first r monitoring sequence information in the semiconductor equipment operation monitoring information, constructs an uninterrupted operation status characterization array, including: step S310: based on the exclusive operation status characterization arrays corresponding to the first r monitoring sequence information, obtaining the importance coefficients corresponding to the first r monitoring sequence information; step S320: mapping and transforming the exclusive operation status characterization arrays corresponding to the first r monitoring sequence information, to obtain the mapping transformation arrays corresponding to the first r monitoring sequence information; step S330: based on the importance coefficients corresponding to the first r monitoring sequence information, integrating the mapping transformation arrays corresponding to the first r monitoring sequence information, to obtain the uninterrupted operation status characterization array.

[0102] In step S310, the detection device obtains the importance coefficients corresponding to the first r monitoring sequence information based on the exclusive operating state characterization arrays corresponding to the first r monitoring sequence information. The importance coefficient reflects the relative importance of each monitoring sequence information in constructing an uninterrupted operating state characterization array. Different monitoring sequence information may have different degrees of influence on the overall operating state of the semiconductor device due to factors such as their time position and the key information they contain. For example, for a semiconductor wafer manufacturing equipment, recent monitoring sequence information may better reflect the current operating state of the equipment, so its importance coefficient may be higher; and although the earlier monitoring sequence information also has a certain reference value, it may be relatively less important. The detection device can use a variety of methods to determine the importance coefficient. One method is a time-based weighting method, which assigns higher weights to newer monitoring sequence information and lower weights to older monitoring sequence information according to the time sequence of the monitoring sequence information.

[0103] In step S320, the detection device performs mapping transformation on the exclusive operating state characterization arrays corresponding to the first r monitoring sequence information, and obtains the mapping transformation arrays corresponding to the first r monitoring sequence information. The purpose of the mapping transformation is to convert the exclusive operating state characterization arrays of different dimensions and ranges into a unified form for subsequent integration operations. The exclusive operating state characterization arrays corresponding to different monitoring sequence information may have different characteristics and scales, and direct integration may cause some features to be overemphasized or ignored. Through the mapping transformation, each array can be compared and integrated in the same feature space. The detection device can use linear mapping or nonlinear mapping methods for transformation. For example, linear mapping, for example, can use a linear function T(x)=ax+b to transform each element in the exclusive operating state characterization array, where a and b are constants. Nonlinear mapping can use more complex functions, such as logarithmic functions, exponential functions or neural networks. Taking a neural network as an example, the detection device can train a neural network model, take the exclusive operating state characterization array as input, perform nonlinear transformation through the hidden layer of the network, and output a mapping transformation array.

[0104] In step S330, the detection device integrates the mapping transformation arrays corresponding to the first r monitoring sequence information based on the importance coefficients corresponding to the first r monitoring sequence information to obtain an uninterrupted operation status characterization array. This step is to weighted sum the arrays after mapping transformation according to their corresponding importance coefficients to obtain a comprehensive array, namely, the uninterrupted operation status characterization array. Through weighted summation, the characteristics and importance of each monitoring sequence information can be comprehensively considered, so that the constructed uninterrupted operation status characterization array can more accurately reflect the overall operation status of the semiconductor device.

[0105] As an implementation mode, step S400, based on the uninterrupted operation state characterization array and the exclusive operation state characterization array corresponding to the s-th monitoring sequence information, constructs a target operation state characterization array corresponding to the s-th monitoring sequence information, including: step S410: combining the uninterrupted operation state characterization array and the exclusive operation state characterization array corresponding to the s-th monitoring sequence information to construct a target operation state characterization array corresponding to the s-th monitoring sequence information.

[0106] Step S410 combines the overall operating characteristics reflected by the first r monitoring sequence information of the semiconductor device with the characteristics of the s-th monitoring sequence information itself, thereby obtaining an array that more comprehensively and accurately reflects the characteristics of the s-th monitoring sequence information during the entire operating process, providing a basis for subsequent more reliable identification of the device operating status corresponding to the s-th monitoring sequence information.

[0107] The uninterrupted operation status characterization array is constructed by the detection device based on the exclusive operation status characterization arrays corresponding to the first r monitoring sequence information in the semiconductor device operation monitoring information. It combines the characteristics of the first r monitoring sequence information and reflects the continuous operation status of the semiconductor device over a relatively long period of time. The exclusive operation status characterization array corresponding to the s-th monitoring sequence information focuses on the characteristics of the s-th monitoring sequence information itself, reflecting the operation status of the semiconductor device in this specific time period. By combining these two arrays, the detection device can obtain a target operation status characterization array that contains both the historical operation information of the equipment and the current specific time period information.

[0108] The detection device can combine these two arrays in a variety of ways. One of the ways is splicing combination. Through splicing combination, the target operating state characterization array contains complete parameter feature information from the previous batches to the current batch. In addition to splicing combination, the detection device can also use weighted combination. This method takes into account the relative importance of the uninterrupted operating state characterization array and the exclusive operating state characterization array corresponding to the sth monitoring sequence information when constructing the target operating state characterization array.

[0109] As an implementation method, the semiconductor equipment operation monitoring information is obtained by sampling from the initial monitoring information; the method further includes:

[0110] Step S600: clearing the first r monitoring sequence information in the initial monitoring information to obtain cleared initial monitoring information;

[0111] Step S700: obtaining the device operation status corresponding to each remaining monitoring sequence information in the cleared initial monitoring information;

[0112] Step S800: based on the equipment operation states respectively corresponding to the remaining monitoring sequence information, each remaining monitoring sequence information is divided to obtain a plurality of information clusters, each information cluster corresponding to an equipment operation state;

[0113] Step S900: extracting a remaining monitoring sequence information from multiple information clusters for multiple times and combining them to obtain multiple information matrices;

[0114] Step S1000: for multiple transparency models, obtain the significance index of the transparency model under the restriction of each information matrix, the significance index is used to measure the positive influencing factor of the first remaining monitoring sequence information in the information matrix, and the proportion of the positive influencing factor in the information matrix;

[0115] Step S1100: based on the saliency indexes of the transparency model under the constraints of each information matrix, obtaining a target saliency index corresponding to the transparency model;

[0116] Step S1200: Determine a transparency model suitable for initial monitoring information based on target significance indicators corresponding to a plurality of transparency models.

[0117] In step S600, the initial monitoring information is a set of raw data obtained by the detection equipment to monitor the operation of the semiconductor equipment, which contains multiple monitoring sequence information arranged in time sequence. These monitoring sequence information reflect the operating status of the semiconductor equipment in different time periods. The first r monitoring sequence information, where 1≤r<s, s is the total number of monitoring sequence information in the initial monitoring information, may interfere with subsequent analysis due to instability in the equipment startup phase, errors in the initial data collection, etc. Therefore, the detection equipment needs to clear the first r monitoring sequence information from the initial monitoring information.

[0118] In step S700, after obtaining the cleaned initial monitoring information, the detection device needs to determine the device operation status corresponding to each of the remaining monitoring sequence information. The device operation status is a description of the operation of the semiconductor device in a certain period of time, such as normal operation, fault warning, fault occurrence, etc.

[0119] In step S800, after obtaining the device operation states corresponding to the remaining monitoring sequence information in the cleaned initial monitoring information, the detection device divides the remaining monitoring sequence information according to the device operation states to form multiple information clusters. The monitoring sequence information in each information cluster corresponds to the same device operation state, and such division helps the detection device to centrally analyze and manage the monitoring data under different operation states.

[0120] For example, in an example containing 7 monitoring sequence information, after state identification, the detection device finds that the device operation state corresponding to 4 monitoring sequence information is normal operation, the device operation state corresponding to 2 monitoring sequence information is fault warning, and the device operation state corresponding to 1 monitoring sequence information is fault occurrence. Then the detection device divides the 7 monitoring sequence information into 3 information clusters. The first information cluster contains 4 monitoring sequence information corresponding to normal operation state, the second information cluster contains 2 monitoring sequence information corresponding to fault warning state, and the third information cluster contains 1 monitoring sequence information corresponding to fault occurrence state.

[0121] In step S600, the detection device can dynamically determine the value of r based on the characteristics of the device and historical data to ensure that the monitoring sequence information that is most likely to interfere with the analysis is cleared. In step S700, the detection device can use a more complex state recognition algorithm or model to improve the accuracy of state recognition. In step S800, the detection device can perform further statistical analysis and feature extraction on each information cluster to dig out the inherent laws and characteristics of the monitoring data under different operating conditions, and provide more valuable information for fault prediction, maintenance decision-making, etc. of semiconductor equipment.

[0122] In step S900, the detection device extracts a remaining monitoring sequence information from multiple information clusters for multiple times, and combines to obtain multiple information matrices. The information clusters here are obtained by dividing the remaining monitoring sequence information based on the equipment operation state in step S800, and each information cluster corresponds to an equipment operation state. The detection device can construct an information matrix reflecting the combination of different equipment operation states by selecting and combining monitoring sequence information from different information clusters. For example, suppose there are three information clusters, corresponding to three equipment operation states of normal operation, fault warning and fault occurrence, and there are several remaining monitoring sequence information in each information cluster. The detection device selects a monitoring sequence information from the normal operation information cluster for the first time, selects a monitoring sequence information from the fault warning information cluster, and selects a monitoring sequence information from the fault occurrence information cluster, and combines these three monitoring sequence information into an information matrix. Then, the detection device repeats this process again, selecting different monitoring sequence information each time, so as to obtain multiple information matrices. The technical means to implement this step can be to use a random sampling method, and the detection device randomly selects a monitoring sequence information in each information cluster for combination, and repeats this process multiple times to obtain multiple information matrices.

[0123] Step S1000 is for a variety of transparency models, and the detection device needs to obtain the significance index of these transparency models under the constraints of each information matrix. Transparency model, also known as interpretability model, is used to explain the decision-making process and results of the model, such as Gradient, Integrated Gradients, CAM, Grad-CAM, etc. The significance index is used to measure the proportion of the positive influencing factors of the first remaining monitoring sequence information in the information matrix in the positive influencing factors of the information matrix. The positive influencing factor refers to the positive attribution value generated by each monitoring information in the monitoring sequence information on the judgment of the equipment operation status. For example, for a transparency model and an information matrix, the detection device needs to calculate the positive influencing factors between each remaining monitoring sequence information in the information matrix and the transparency model, and then determine the proportion of the positive influencing factors of the first remaining monitoring sequence information in the sum of all positive influencing factors. This proportion is the significance index of the transparency model under the constraints of this information matrix. The detection device can determine the positive influencing factor by calculating the contribution value of each monitoring sequence information to the judgment of the equipment operation status, and then obtain the significance index according to the corresponding calculation rules.

[0124] In step S1100, the detection device obtains the target significance index corresponding to the transparency model based on the significance index of the transparency model under the constraints of each information matrix. Since the detection device constructs multiple information matrices, each transparency model has different significance indexes under different information matrices. In order to comprehensively evaluate the performance of the transparency model, it is necessary to calculate a target significance index. A feasible method is to average the significance index of the transparency model under all information matrices, that is, the target significance index = (the significance index of the transparency model under information matrix 1 + the significance index of the transparency model under information matrix 2 + ... + the significance index of the transparency model under information matrix n) / n, where n is the number of information matrices. In this way, the detection device can obtain an indicator that comprehensively reflects the performance of the transparency model under different information matrices.

[0125] In step S1200, the detection device determines a transparency model suitable for the initial monitoring information based on the target significance indicators corresponding to the multiple transparency models. The detection device compares the target significance indicators of different transparency models and selects the transparency model whose target significance indicator best meets the specific requirements as the model suitable for the initial monitoring information. For example, if the detection device wants to select a transparency model that can maximize the positive impact of the first remaining monitoring sequence information, then the transparency model with the largest target significance indicator is selected.

[0126] In actual operation, when the detection device executes step S900, it is necessary to consider the randomness and comprehensiveness of sampling. The randomness of sampling can ensure that the constructed information matrix can represent different combinations of equipment operating states to avoid deviations. Comprehensiveness requires the detection device to construct as many different information matrices as possible to cover various possible situations. To achieve this, the detection device can set a sufficiently large number of sampling times to ensure that the monitoring sequence information in each information cluster can be fully utilized. At the same time, the detection device can record and manage the constructed information matrix for subsequent analysis.

[0127] In actual application scenarios, when the detection device constructs the information matrix in step S900, it may face the problem of unbalanced number of monitoring sequence information in the information cluster. For example, the number of monitoring sequence information in the normal operation information cluster is much larger than the number of monitoring sequence information in the fault occurrence information cluster. In order to solve this problem, the detection device can adopt oversampling or undersampling methods. Oversampling refers to copying or constructing new similar information in a smaller number of information clusters to increase their number; undersampling refers to randomly deleting the monitoring sequence information in a larger number of information clusters to reduce their number. Through these methods, the detection device can make the number of monitoring sequence information in each information cluster more balanced, thereby improving the quality of constructing the information matrix.

[0128] For step S1000 to calculate the significance index of the transparency model under the information matrix constraint, the detection device can be optimized in combination with domain knowledge. For example, in the field of semiconductor equipment detection, certain monitoring information may be more important for determining the operating status of the equipment. The detection device can assign different weights to different monitoring information based on this domain knowledge, and consider these weights when calculating the positive impact factor, so that the significance index can better reflect the actual situation.

[0129] When calculating the target significance index in step S1100, the detection device can adopt a method of dynamically adjusting the weight. If it is found during the processing that a certain information matrix has a greater impact on the final result, the detection device can appropriately increase the weight of the significance index corresponding to the information matrix; conversely, if the impact of a certain information matrix is ​​small, its weight can be reduced. In this way, the target significance index can more flexibly reflect the performance of the transparency model in different situations.

[0130] When determining the applicable transparency model in step S1200, the detection device may consider the degree of matching with the actual business needs. For example, if the production process of semiconductor equipment has high requirements for the timeliness of fault warning, the detection device may select a transparency model with a higher target significance index under the fault warning related information matrix. At the same time, the detection device may also consider whether the interpretability of the transparency model is consistent with the operator's understanding ability, so as to enable better decision-making and control in actual applications.

[0131] As an implementation mode, step S1000, obtaining the significance index of the transparency model under the constraints of each information matrix, includes:

[0132] Step S1001: for each remaining monitoring sequence information in the information matrix, obtain the total positive impact factor between the remaining monitoring sequence information and the transparency model, the total positive impact factor being the sum of the positive impact factors between each monitoring information in the remaining monitoring sequence information and the transparency model under the restriction of the equipment operating state corresponding to the remaining monitoring sequence information;

[0133] Step S1002: adding the total positive impact factors between each remaining monitoring sequence information in the information matrix and the transparency model to obtain a first addition result;

[0134] Step S1003: Taking the total positive impact factor between the remaining monitoring sequence information and the transparency model as the quotient and the first addition result, a significance index of the transparency model under the information matrix restriction is obtained.

[0135] In step S1001, the detection device obtains the total positive impact factor between the remaining monitoring sequence information and the transparency model for each remaining monitoring sequence information in the information matrix, where the total positive impact factor is the sum of the positive impact factors between each monitoring information in the remaining monitoring sequence information and the transparency model under the restriction of the equipment operation state corresponding to the remaining monitoring sequence information. The positive impact factor can be understood as the positive attribution value generated by the monitoring information for the judgment of the equipment operation state, that is, the quantitative representation of the positive role played by the monitoring information in judging the current operation state of the equipment. The transparency model is an interpretable model used to explain the model decision process and results, such as Gradient, Integrated Gradients, CAM, Grad-CAM, etc. For example, for an information matrix, it contains three remaining monitoring sequence information, corresponding to three equipment operation states of normal operation, fault warning and fault occurrence. For the remaining monitoring sequence information corresponding to the normal operation state, the sequence information contains multiple monitoring information, such as the temperature, pressure, current, etc. of the equipment. The detection device uses a transparency model (such as the Gradient model) to calculate the positive impact factor of each monitoring information on judging whether the device is in normal operation, and then adds these positive impact factors to obtain the total positive impact factor between the remaining monitoring sequence information and the transparency model. The technical means to achieve this step can be to calculate the positive impact factor based on the specific principles of the transparency model. Taking the Gradient model as an example, the detection device can calculate the gradient of the monitoring information to the model output. The size of the gradient value can indicate the degree of influence of the monitoring information on the model decision, and a positive gradient can be regarded as a positive impact factor.

[0136] In step S1002, the detection device needs to add the total positive impact factors between each remaining monitoring sequence information in the information matrix and the transparency model to obtain a first addition result. Continuing with the above information matrix as an example, the detection device has calculated the total positive impact factors between the remaining monitoring sequence information corresponding to the three equipment operating states of normal operation, fault warning and fault occurrence and the transparency model, and the three total positive impact factors are added to obtain the first addition result. This first addition result represents the sum of all positive impact factors of the entire information matrix under the transparency model, which reflects the overall positive impact of all the remaining monitoring sequence information in the information matrix on the judgment of the equipment operating state.

[0137] Step S1003 requires the detection device to take the total positive impact factor between the remaining monitoring sequence information and the transparency model as a quotient with the first addition result to obtain the significance index of the transparency model under the information matrix restriction. The significance index is used to measure the proportion of the positive impact factor of the first remaining monitoring sequence information in the information matrix to the positive impact factor of the information matrix. This significance index can help the detection device evaluate the importance of the remaining monitoring sequence information in the entire information matrix, that is, the proportion of the positive impact of the sequence information on the judgment of the equipment operation status in all sequence information.

[0138] When calculating the positive impact factor, the detection equipment can also introduce weight factors. Different monitoring information may have different importance for the judgment of the equipment's operating status. The detection equipment can assign different weights to each monitoring information based on domain knowledge or historical data, and consider these weights when calculating the positive impact factor, so that the total positive impact factor can better reflect the actual situation. For example, in semiconductor equipment testing, the key parameters of the equipment (such as the temperature during chip manufacturing) may be more important for the judgment of the equipment's operating status, and a higher weight can be assigned to it.

[0139] In step S1001, the detection device can use a cross-validation method to improve the reliability of the total positive impact factor calculation. Specifically, the detection device can divide the remaining monitoring sequence information into multiple subsets, and then calculate the positive impact factors on different subsets respectively, and finally combine these results to obtain a more accurate total positive impact factor. This method can reduce the impact of sampling errors on the results and improve the credibility of the total positive impact factor.

[0140] For step S1002, the detection device can adopt a dynamic adjustment strategy when adding the total positive impact factor. If it is found during the calculation process that the total positive impact factor of a remaining monitoring sequence information is too different from other sequence information, it may be caused by data anomalies or calculation errors. The detection device can further check and correct this abnormal total positive impact factor, or exclude it from the addition range to avoid excessive impact on the first addition result.

[0141] In step S1003, the detection device can visualize the calculated significance index. By drawing charts (such as bar charts, line charts), etc., the detection device can intuitively compare the significance indexes of transparency models under different information matrices, so as to more conveniently select a suitable transparency model. At the same time, the detection device can predict and evaluate the performance of the transparency model according to the changing trend of the significance index.

[0142] During the entire processing, the detection device needs to be effectively connected with the previous steps. For example, the remaining monitoring sequence information used in step S1001 is the data in the information matrix constructed in step S900. The detection device needs to ensure that the construction process of the information matrix is ​​accurate, otherwise it will affect the subsequent calculation of positive impact factors and significance indicators. At the same time, the detection device needs to record and log the entire processing process so that it can be traced and checked when problems arise.

[0143] In terms of data storage, the detection equipment needs to reasonably store the calculated positive impact factor, total positive impact factor, first addition result, and significance index. These data can be stored in a database to facilitate subsequent query and analysis. At the same time, in order to ensure the security of the data, the detection equipment needs to encrypt the stored data to prevent data leakage.

[0144] When the detection device calculates the positive impact factor in step S1001, it can also combine the feature importance evaluation method in machine learning. For example, models such as random forests can be used to evaluate the importance of each monitoring information to the judgment of the equipment operation status, and these importance scores can be considered as part of the positive impact factor. In this way, the advantages of multiple methods can be comprehensively utilized to improve the accuracy of the calculation of positive impact factors. At the same time, the detection device can compare and analyze different transparency models to understand their advantages and disadvantages in calculating positive impact factors, and select the most appropriate model according to the specific situation.

[0145] For step S1002, the detection device can use a layered addition method when adding the total positive impact factor. If the remaining monitoring sequence information in the information matrix can be layered according to a certain rule (such as layered according to the severity of the equipment operation status), the detection device can first add the total positive impact factor in each layer, and then add the results of each layer to obtain the first addition result. This method can more clearly analyze the contribution of the remaining monitoring sequence information at different levels to the overall positive impact.

[0146] In step S1003, the detection device may normalize the significance index. Since the scale and composition of different information matrices may be different, direct comparison of the significance index may be biased. Through normalization, the significance index can be mapped to a unified range, which is convenient for comparison and analysis. For example, the minimum-maximum normalization method can be used to normalize the significance index to the interval [0, 1].

[0147] The detection equipment can also introduce uncertainty analysis in the entire calculation process. Since the monitoring information may contain noise and errors, the calculated positive impact factors and significance indicators may also have certain uncertainties. The detection equipment can use methods such as Monte Carlo simulation to evaluate the uncertainty in the calculation process and provide confidence intervals for significance indicators, thereby more accurately describing the performance of the transparency model.

[0148] In addition, the testing equipment can correlate the calculated significance index with the actual equipment operation status. By comparing the significance index of different transparency models with the actual failure of the equipment, the testing equipment can verify the validity of the significance index and further optimize the calculation method and model selection strategy. At the same time, the testing equipment can feed back these analysis results to the entire testing system to continuously improve the level of automated testing of semiconductor equipment.

[0149] As an embodiment, the method is performed by a semiconductor device state recognition network, and the method also includes a debugging process of the semiconductor device state recognition network, including:

[0150] Step S10: Obtain a debugging sample, the debugging sample includes s monitoring sequence information arranged in time sequence, each monitoring sequence information corresponds to a priori equipment operation status, s≥2;

[0151] Step S20: acquiring exclusive operation state characterization arrays corresponding to s monitoring sequence information in the debugging sample based on the semiconductor device state recognition network, wherein the exclusive operation state characterization array is used to characterize the monitoring sequence information;

[0152] Step S30: constructing an uninterrupted operation state characterization array based on the exclusive operation state characterization arrays corresponding to the first r monitoring sequence information in the debugging sample based on the semiconductor device state recognition network, and the uninterrupted operation state characterization array is used to characterize the debugging sample, 1≤r<s;

[0153] Step S40: constructing a target operating state representation array corresponding to the s-th monitoring sequence information based on the uninterrupted operating state representation array and the exclusive operating state representation array corresponding to the s-th monitoring sequence information;

[0154] Step S50: Based on the semiconductor device state recognition network and the target operation state representation array corresponding to the s-th monitoring sequence information, the s-th monitoring sequence information is state recognized to obtain the predicted device operation state corresponding to the s-th monitoring sequence information;

[0155] Step S60: Based on the predicted device operating state and the prior device operating state corresponding to the sth monitoring sequence information, the semiconductor device state recognition network is debugged to obtain a debugged semiconductor device state recognition network, and the debugged semiconductor device state recognition network is used to perform state recognition on the monitoring sequence information in the semiconductor device operation monitoring information.

[0156] In step S10, the detection device obtains a debugging sample, and the debugging sample contains s monitoring sequence information arranged in time sequence, each monitoring sequence information corresponds to a priori equipment operation state, and s≥2. The debugging sample is a data set used to train and adjust the semiconductor equipment state recognition network. The prior equipment operation state is a known correct equipment operation state, which can be used as a reference standard for subsequent evaluation of network prediction results. For example, the detection device monitors a semiconductor manufacturing equipment, collects 10 monitoring sequence information arranged in time sequence (i.e., s=10) within a period of time, and determines the equipment operation state corresponding to each monitoring sequence information through manual inspection or other reliable means, such as normal operation, fault warning, fault occurrence, etc. These known equipment operation states are the prior equipment operation states. The detection device can select representative monitoring sequence information from historical monitoring data as debugging samples, or it can collect data in real time during actual operation and mark the prior equipment operation state.

[0157] In step S20, the detection device obtains the exclusive operation state characterization arrays corresponding to the s monitoring sequence information in the debugging sample based on the semiconductor device state recognition network, and the exclusive operation state characterization array is used to characterize the monitoring sequence information. The detection device first preprocesses the debugging sample, which is similar to step S210. The median filtering method can be used to obtain the preprocessed information corresponding to each data acquisition timestamp through the convolution window to obtain the preprocessed debugging sample. Then, according to the method of step S220, the encoding array representation corresponding to each monitoring sequence information is obtained based on the preprocessed information, including determining the target data acquisition timestamp, obtaining the data acquisition timestamp set, and obtaining the encoding array representation based on the monitoring information corresponding to each data acquisition timestamp in the set. Finally, according to the method of step S230, the exclusive operation state characterization array corresponding to each monitoring sequence information is obtained based on the encoding array representation. For example, for the first monitoring sequence information in the debugging sample, the detection device obtains an exclusive operation state characterization array that can reflect the characteristics of the monitoring sequence information after the above series of processing steps.

[0158] Step S30 is that the detection device builds an uninterrupted operation state characterization array based on the semiconductor device state recognition network according to the exclusive operation state characterization arrays corresponding to the first r monitoring sequence information in the debugging sample, and the array is used to characterize the debugging sample, where 1≤r<s. The detection device first obtains the importance coefficients corresponding to the first r monitoring sequence information based on the exclusive operation state characterization arrays corresponding to the first r monitoring sequence information, which can be determined by analyzing the time sequence of the monitoring sequence information, data stability and other factors. Then, the exclusive operation state characterization arrays corresponding to the first r monitoring sequence information are mapped and transformed, for example, the array is mapped to another space using a linear transformation or a nonlinear transformation to extract more valuable features. Finally, based on the importance coefficients corresponding to the first r monitoring sequence information, the mapping transformation array is integrated, and a weighted summation method can be used, that is, the uninterrupted operation state characterization array = ∑ (importance coefficient of the i-th monitoring sequence information × mapping transformation array of the i-th monitoring sequence information), where i ranges from 1 to r, thereby obtaining an uninterrupted operation state characterization array that can comprehensively reflect the characteristics of the first r monitoring sequence information.

[0159] Step S40 requires the detection device to construct a target operation state representation array corresponding to the s-th monitoring sequence information based on the uninterrupted operation state representation array and the exclusive operation state representation array corresponding to the s-th monitoring sequence information. The detection device can combine the two arrays, for example, splicing them together, or adopt a more complex fusion method, such as weighted fusion, assigning different weights according to the importance of different arrays, and then weighted adding them to obtain the target operation state representation array corresponding to the s-th monitoring sequence information, which contains both the characteristics of the s-th monitoring sequence information itself and the influence of the first r monitoring sequence information on it.

[0160] In step S50, the detection device performs state recognition on the s-th monitoring sequence information based on the target operating state representation array corresponding to the s-th monitoring sequence information based on the semiconductor device state recognition network, and obtains the predicted device operating state corresponding to the s-th monitoring sequence information. The detection device can use a classification algorithm, such as a support vector machine, a neural network, etc., to take the target operating state representation array as input and output the predicted device operating state. For example, the target operating state representation array is input into a trained neural network, and a probability distribution is obtained through forward propagation calculation of the network. The detection device selects the category with the highest probability as the predicted device operating state.

[0161] Step S60 is that the detection device debugs the semiconductor device state recognition network based on the predicted device operating state and the prior device operating state corresponding to the s-th monitoring sequence information, and obtains a debugged semiconductor device state recognition network, which is used to identify the state of the monitoring sequence information in the semiconductor device operation monitoring information. The detection device can use a loss function to measure the difference between the predicted device operating state and the prior device operating state. The feasible loss function includes a cross entropy loss function, etc. The detection device uses an optimization algorithm (such as stochastic gradient descent, Adam optimizer, etc.) to adjust the parameters of the semiconductor device state recognition network according to the value of the loss function, so that the value of the loss function gradually decreases, thereby improving the prediction accuracy of the network. After multiple iterative training, when the value of the loss function converges to a smaller value or reaches a preset number of training rounds, the detection device considers that the network has been trained and obtains a debugged semiconductor device state recognition network.

[0162] In actual operation, the detection device needs to ensure the quality and representativeness of the debugging samples when executing step S10. The debugging samples should cover all possible operating states of the semiconductor device, and the data should be accurate. The detection device can perform quality checks on the collected monitoring sequence information, remove noise data and outliers, and adopt data enhancement methods, such as translation, scaling and other transformations on the monitoring sequence information, to increase the diversity of debugging samples and improve the generalization ability of the network.

[0163] For step S20, when the detection device obtains the exclusive operation status characterization array, it is necessary to pay attention to the parameter selection in the preprocessing process. For example, in the convolution window selection of the median filter, the window size will affect the filtering effect, and the detection device needs to select a suitable window size according to the characteristics of the data and the noise level. In the process of obtaining the encoding array representation and the exclusive operation status characterization array, the detection device can continuously adjust the calculation method and parameters to improve the array's ability to characterize the monitoring sequence information.

[0164] In step S30, when the detection device determines the importance coefficient of the first r monitoring sequence information, it can use feature selection methods in machine learning, such as recursive feature elimination, random forest feature importance, etc., to determine the importance coefficient by analyzing the correlation between the monitoring sequence information and the device operating status. During the mapping transformation process, the detection device can try different transformation functions, such as ReLU, Sigmoid, etc., to find the most suitable transformation method.

[0165] In step S40, when combining the uninterrupted operation state characterization array and the exclusive operation state characterization array corresponding to the s-th monitoring sequence information, the detection device can select a suitable combination method according to different application scenarios and data characteristics. For example, if the feature dimensions of the two arrays are quite different, they can be first subjected to dimensionality reduction processing and then combined. At the same time, the detection device can compare the effects of different combination methods through experiments and select the optimal combination method.

[0166] In step S50, when the detection device selects a classification algorithm, it needs to consider the complexity and accuracy of the algorithm. For simple data sets, a support vector machine can be used; for complex data sets, a deep neural network may be more advantageous. The detection device can experimentally compare different classification algorithms and select the most suitable algorithm. At the same time, the detection device can tune the parameters of the classification algorithm, such as the learning rate of the neural network, the number of hidden layer nodes, etc., to improve the accuracy of the classification.

[0167] In step S60, the detection device needs to select the loss function and optimization algorithm according to the specific situation. Different loss functions are suitable for different tasks. For example, for multi-classification problems, the cross entropy loss function is usually a better choice; for regression problems, the mean square error loss function may be more appropriate. The choice of optimization algorithm will also affect the training speed and performance of the network. The detection device can try different optimization algorithms, such as stochastic gradient descent, Adagrad, Adadelta, etc., to find the most suitable optimization algorithm. During the training process, the detection device can adopt an early stopping strategy. When the loss function value on the validation set no longer decreases, the training is stopped in advance to avoid overfitting.

[0168] During the entire debugging process, the testing equipment also needs to divide and verify the data. The testing equipment can divide the debugging samples into training sets, verification sets, and test sets. The training set is used to train the network, the verification set is used to adjust the network parameters and select the optimal model, and the test set is used to evaluate the final performance of the network. The testing equipment can use the cross-validation method to divide the data set multiple times for training and verification to improve the reliability of the evaluation results.

[0169] Figure 2 A hardware entity diagram of a detection device provided by an embodiment of the present invention, such as Figure 2 As shown, the hardware entity of the detection device 1000 includes: a processor 1001 and a memory 1002, wherein the memory 1002 stores a computer program that can be run on the processor 1001, and the processor 1001 implements the steps in the method of any of the above embodiments when executing the program.

[0170] The above description is only an implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A semiconductor automated detection method based on artificial intelligence, characterized in that: The method comprises: Acquire semiconductor device operation monitoring information, wherein the semiconductor device operation monitoring information includes s monitoring sequence information arranged in time sequence, wherein s≥2; the semiconductor device operation monitoring information includes time sequence data generated by the semiconductor device during operation; Preprocessing the semiconductor device operation monitoring information to obtain preprocessed semiconductor device operation monitoring information, wherein the preprocessed semiconductor device operation monitoring information is information after data cleaning; based on the preprocessed semiconductor device operation monitoring information, obtaining the encoding array representations corresponding to the s monitoring sequence information respectively; based on the encoding array representations corresponding to the s monitoring sequence information respectively, obtaining the exclusive operation state representation arrays corresponding to the s monitoring sequence information respectively, wherein the exclusive operation state representation array is used to represent the monitoring sequence information; Based on the exclusive operation state characterization arrays corresponding to the first r monitoring sequence information, the importance coefficients corresponding to the first r monitoring sequence information are obtained; the exclusive operation state characterization arrays corresponding to the first r monitoring sequence information are mapped and transformed to obtain the mapping transformation arrays corresponding to the first r monitoring sequence information; based on the importance coefficients corresponding to the first r monitoring sequence information, the mapping transformation arrays corresponding to the first r monitoring sequence information are integrated to obtain the uninterrupted operation state characterization array, which is used to characterize the semiconductor device operation monitoring information, and is an integration of the features of the first r monitoring sequence information, and is used to reflect the continuous operation state of the semiconductor device within a time period, wherein 1≤r<s; Combining the uninterrupted running state representation array and the exclusive running state representation array corresponding to the s-th monitoring sequence information to construct a target running state representation array corresponding to the s-th monitoring sequence information; Through the semiconductor device state recognition network, based on the target operating state representation array corresponding to the sth monitoring sequence information as input, the state of the sth monitoring sequence information is recognized, and the device operating state corresponding to the sth monitoring sequence information is output.

2. The method according to claim 1, characterized in that The monitoring sequence information includes monitoring information corresponding to a plurality of uninterrupted data acquisition timestamps respectively; The obtaining, based on the preprocessed semiconductor device operation monitoring information, encoding array representations corresponding to the s monitoring sequence information respectively includes: Based on the preprocessed semiconductor equipment operation monitoring information, obtaining target data acquisition timestamps corresponding to the s monitoring sequence information respectively, wherein the target data acquisition timestamp is a data acquisition timestamp corresponding to a peak value of the monitoring information among a plurality of data acquisition timestamps corresponding to the monitoring sequence information; For each of the monitoring sequence information, using the target data collection timestamp corresponding to the monitoring sequence information as a reference timestamp, forwardly acquiring the data collection timestamp of the first number, and reversely acquiring the data collection timestamp of the second number, to obtain a data collection timestamp set corresponding to the monitoring sequence information; Based on the monitoring information corresponding to each data collection timestamp in the data collection timestamp set corresponding to the monitoring sequence information, the encoding array representation corresponding to the monitoring sequence information is obtained.

3. The method according to claim 2, characterized in that The acquiring, based on the monitoring information corresponding to each data collection timestamp in the data collection timestamp set corresponding to the monitoring sequence information, the encoding array representation corresponding to the monitoring sequence information comprises: Dividing the monitoring information corresponding to the data collection timestamp set to obtain a plurality of sub-monitoring sequence information, wherein the sub-monitoring sequence information includes the monitoring information corresponding to the data collection timestamp of the third number; For each of the sub-monitoring sequence information, based on the monitoring information corresponding to the sub-monitoring sequence information, obtaining a coding array representation corresponding to the sub-monitoring sequence information; Based on the encoding array representations respectively corresponding to the multiple sub-monitoring sequence information, the encoding array representation corresponding to the monitoring sequence information is obtained.

4. The method according to claim 3, characterized in that The obtaining, based on the encoding array representations respectively corresponding to the s monitoring sequence information, the exclusive operation status representation arrays respectively corresponding to the s monitoring sequence information includes: For each of the monitoring sequence information, obtaining the importance coefficients corresponding to each of the sub-monitoring sequence information in the monitoring sequence information; Based on the importance coefficients respectively corresponding to the sub-monitoring sequence information in the monitoring sequence information, the encoding array representations respectively corresponding to the sub-monitoring sequence information in the monitoring sequence information are integrated to obtain a first integrated encoding array; For each sub-monitoring sequence information in the monitoring sequence information, obtaining the element-by-element product between the encoding array representation of the sub-monitoring sequence information and the encoding array representation of each target sub-monitoring sequence information corresponding to the sub-monitoring sequence information; wherein the target sub-monitoring sequence information is the sub-monitoring sequence information after the sub-monitoring sequence information in the monitoring sequence information; Integrate the element-by-element product between the sub-monitoring sequence information and each target sub-monitoring sequence information corresponding to the sub-monitoring sequence information to obtain a second integrated coding array corresponding to the sub-monitoring sequence information; Integrate the second integrated coding arrays corresponding to each sub-monitoring sequence information in the monitoring sequence information to obtain a third integrated coding array; The first integrated coding array and the third integrated coding array are integrated to obtain an exclusive operation status representation array corresponding to the monitoring sequence information.

5. The method according to claim 1, characterized in that The preprocessing of the semiconductor device operation monitoring information to obtain the preprocessed semiconductor device operation monitoring information includes: Obtaining a convolution window corresponding to the semiconductor device operation monitoring information, wherein the convolution window corresponds to A data acquisition timestamps, where A is an odd number; For a first data collection timestamp in the semiconductor device operation monitoring information, taking the first data collection timestamp as the center, obtaining monitoring information corresponding to A data collection timestamps in the semiconductor device operation monitoring information, and constructing information to be processed corresponding to the first data collection timestamp; Using the median monitoring information in the to-be-processed information corresponding to the first data collection timestamp as the pre-processed information corresponding to the first data collection timestamp; Based on the preprocessed information corresponding to each data acquisition timestamp in the semiconductor device operation monitoring information, the preprocessed semiconductor device operation monitoring information is obtained.

6. The method according to claim 1, characterized in that The semiconductor equipment operation monitoring information is obtained by sampling from the initial monitoring information; the method further comprises: Cleaning the first r monitoring sequence information in the initial monitoring information to obtain cleaned initial monitoring information; Obtaining the device operation status corresponding to each remaining monitoring sequence information in the cleaned initial monitoring information; Based on the equipment operation states respectively corresponding to the remaining monitoring sequence information, the remaining monitoring sequence information is divided to obtain a plurality of information clusters, each of the information clusters corresponding to one of the equipment operation states; extracting one of the remaining monitoring sequence information from the multiple information clusters for multiple times respectively, and combining them to obtain multiple information matrices; For multiple transparency models, obtain the significance index of the transparency model under the constraints of each information matrix, the significance index is used to measure the positive influencing factor of the first remaining monitoring sequence information in the information matrix, and the proportion of the positive influencing factor in the information matrix; wherein the transparency model is an interpretable model, which is used to explain the decision-making process and results of the model; Based on the significance indexes of the transparency model under the constraints of each of the information matrices, obtaining a target significance index corresponding to the transparency model; Based on the target significance indicators respectively corresponding to the plurality of transparency models, a transparency model suitable for the initial monitoring information is determined.

7. The method according to claim 6, characterized in that The obtaining of the significance index of the transparency model under the constraints of each information matrix includes: For each remaining monitoring sequence information in the information matrix, a total positive impact factor between the remaining monitoring sequence information and the transparency model is obtained, wherein the total positive impact factor is the sum of the positive impact factors between each monitoring information in the remaining monitoring sequence information and the transparency model under the restriction of the equipment operating state corresponding to the remaining monitoring sequence information; Adding the total positive impact factors between each remaining monitoring sequence information in the information matrix and the transparency model to obtain a first addition result; The total positive impact factor between the remaining monitoring sequence information and the transparency model is divided by the first addition result to obtain a significance index of the transparency model under the restriction of the information matrix.

8. The method according to any one of claims 1 to 7, characterized in that The method also includes a debugging process of the semiconductor device status recognition network, including: Obtain a debugging sample, wherein the debugging sample includes s monitoring sequence information arranged in time sequence, each monitoring sequence information corresponds to carrying a priori device operation status, s≥2; Acquire, based on the semiconductor device state recognition network, exclusive operation state characterization arrays respectively corresponding to the s monitoring sequence information in the debugging sample, wherein the exclusive operation state characterization array is used to characterize the monitoring sequence information; Based on the semiconductor device state recognition network, based on the exclusive operation state characterization arrays corresponding to the first r monitoring sequence information in the debugging sample, an uninterrupted operation state characterization array is constructed, and the uninterrupted operation state characterization array is used to characterize the debugging sample, 1≤r<s; Based on the uninterrupted running state representation array and the exclusive running state representation array corresponding to the s-th monitoring sequence information, construct a target running state representation array corresponding to the s-th monitoring sequence information; Based on the semiconductor device state recognition network and the target operation state representation array corresponding to the s-th monitoring sequence information, the s-th monitoring sequence information is subjected to state recognition to obtain a predicted device operation state corresponding to the s-th monitoring sequence information; Based on the predicted device operating state and the prior device operating state corresponding to the sth monitoring sequence information, the semiconductor device state identification network is debugged to obtain a debugged semiconductor device state identification network, and the debugged semiconductor device state identification network is used to perform state identification on the monitoring sequence information in the semiconductor device operation monitoring information.

9. A detection device, comprising a memory and a processor, wherein the memory stores a computer program that can be run on the processor, characterized in that: When the processor executes the program, the steps in the method according to any one of claims 1 to 8 are implemented.

Images