Semiconductor machine data anomaly detection method, device and computer equipment

Through computer equipment and computer programs, the target loss function is constructed using the distance similarity matrix and attention weight matrix, and the abnormal detection model is trained, which solves the problem of low abnormal detection accuracy of semiconductor machines and achieves higher abnormal detection accuracy and Production line stability.

CN118656764BActive Publication Date: 2024-11-19SEMITRONIX
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411134809.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-08-19
Publication Date
2024-11-19
Estimated Expiration
2044-08-19

AI Technical Summary

Technical Problem

In the prior art, the accuracy of abnormal detection of semiconductor machines is low, and abnormal data cannot be effectively identified, which affects the stability and reliability of the production line.

Method used

By obtaining machine sample data and initial anomaly detection model, calculate the distance similarity matrix and attention weight matrix, construct the target loss function, train the anomaly detection model until the loss function converges, and then is used for abnormal detection of the data to be detected.

Benefits of technology

It improves the accuracy and reliability of abnormal detection, can promptly identify abnormal data in semiconductor machines, avoid potential failures and downtime, and enhances the stability and reliability of the production line.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118656764B_ABST
    Figure CN118656764B_ABST
Patent Text Reader

Abstract

The present application relates to a semiconductor machine data anomaly detection method, device and computer equipment. According to machine sample data, a distance similarity matrix corresponding to the machine sample data is calculated; the machine sample data is input into an initial anomaly detection model to determine the attention weight matrix corresponding to the machine sample data; according to the distance similarity matrix and the attention weight matrix, a target loss function of the initial anomaly detection model is determined; according to the machine sample data and the target loss function, the initial anomaly detection model is trained until the target loss function converges to a preset numerical range to obtain a trained anomaly detection model; the machine data to be detected is obtained; the machine data to be detected is input into the trained anomaly detection model to determine the abnormal data in the machine data to be detected; based on the target loss function, the performance of the initial anomaly detection model can be improved; based on the trained anomaly detection model, the accuracy and reliability of anomaly detection can be effectively improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of semiconductor manufacturing, and in particular to a method, device and computer equipment for detecting data anomalies in a semiconductor machine. Background Art

[0002] Semiconductor chip manufacturing is a complex, expensive and lengthy process involving hundreds of process steps. It requires real-time monitoring of corresponding process parameter changes based on sensors and other monitoring equipment to improve the yield of chip generation. At the same time, with the improvement of equipment precision and complexity, higher requirements are placed on the quality of fault detection in the semiconductor chip manufacturing process. Timely discovery and identification of potential anomalies is crucial to improving the stability and reliability of the production line.

[0003] In the prior art, rule-based or statistical anomaly detection methods are usually used, such as calculating the mean and standard deviation, making box plots, etc., to achieve anomaly detection of semiconductor machines; however, in practical applications, rule-based or statistical anomaly detection methods often cannot meet the accuracy requirements, resulting in the inability to accurately identify abnormal data in semiconductor machines, and thus unable to ensure the stability and reliability of the production line.

[0004] There is currently no effective solution to the problem of low accuracy in abnormality detection of semiconductor tools in the prior art. Summary of the invention

[0005] Based on this, it is necessary to provide a semiconductor machine data anomaly detection method, device and computer equipment to address the above technical problems.

[0006] In a first aspect, the present application provides a method for detecting data anomalies in a semiconductor machine, the method comprising:

[0007] Obtain machine sample data and initial anomaly detection model;

[0008] According to the machine sample data, calculating the distance similarity matrix corresponding to the machine sample data;

[0009] Inputting the machine sample data into the initial anomaly detection model to determine the attention weight matrix corresponding to the machine sample data;

[0010] Determining a target loss function of the initial anomaly detection model according to the distance similarity matrix and the attention weight matrix;

[0011] According to the machine sample data and the target loss function, the initial anomaly detection model is trained until the target loss function converges to a preset value interval to obtain a trained anomaly detection model;

[0012] Obtain the data of the machine to be tested;

[0013] The machine data to be detected is input into the trained anomaly detection model to determine the abnormal data in the machine data to be detected.

[0014] In one embodiment, obtaining machine sample data includes:

[0015] Obtain the raw sensor data of the machine;

[0016] The data length of the original sensor data is adjusted to a preset data length to obtain machine sample data.

[0017] In one embodiment, before adjusting the data length of the original sensor data to a preset data length to obtain the machine sample data, the method further includes:

[0018] Obtaining a timestamp for each data point in the raw sensor data;

[0019] Determine the time interval based on the difference between the timestamps of adjacent data points;

[0020] Determining data missing conditions of the raw sensor data according to the duration of the raw sensor data, the time interval, and the data length corresponding to the raw sensor data;

[0021] According to the data missing situation, the raw sensor data is preprocessed.

[0022] In one embodiment, determining the data missing condition of the raw sensor data according to the duration of the raw sensor data, the time interval, and the data length corresponding to the raw sensor data includes:

[0023] If the ratio of the duration of the raw sensor data to the time interval is not equal to the data length corresponding to the raw sensor data, it is determined that there is a data missing point in the raw sensor data;

[0024] If the ratio of the duration of the raw sensor data to the time interval is equal to the data length corresponding to the raw sensor data, it is determined that there is no data missing point in the raw sensor data.

[0025] In one embodiment, preprocessing the raw sensor data according to the data missing situation includes:

[0026] If there are data missing points in the original sensor data, determining the data missing position corresponding to the data missing point according to a preset detection rule based on the timestamp of each data point in the original sensor data and the time interval corresponding to the original sensor data;

[0027] According to the preset completion rules, the data missing position is completed.

[0028] In one embodiment, determining the target loss function of the initial anomaly detection model according to the distance similarity matrix and the attention weight matrix includes:

[0029] Determining sample reconstruction data according to the machine sample data and the initial anomaly detection model;

[0030] Determining a first difference between the machine sample data and the sample reconstruction data according to the machine sample data and the sample reconstruction data;

[0031] Determining, according to the distance similarity matrix and the attention weight matrix, a second difference between the distance similarity matrix and the attention weight matrix;

[0032] An objective loss function of the initial anomaly detection model is determined according to the first difference and the second difference.

[0033] In one embodiment, the step of inputting the to-be-detected machine data into the trained anomaly detection model to determine the anomaly data in the to-be-detected machine data includes:

[0034] Inputting the machine data to be detected into the trained anomaly detection model to determine the target reconstructed data corresponding to the machine data to be detected;

[0035] Determine, according to the machine data to be detected and the target reconstruction data, a difference score sequence corresponding to the machine data to be detected and the target reconstruction data;

[0036] Determining an abnormal data detection range according to the difference score sequence;

[0037] Abnormal data in the to-be-detected sensor data is determined according to the difference score sequence and the abnormal data detection range.

[0038] In one embodiment, determining the abnormal data detection range according to the difference score sequence includes:

[0039] According to the difference score sequence, determining a mean and a standard deviation corresponding to the difference score sequence;

[0040] Determining a first detection threshold and a second detection threshold according to the mean and the standard deviation; the first detection threshold is less than the second detection threshold;

[0041] An abnormal data detection range is determined according to the first detection threshold and the second detection threshold.

[0042] In one embodiment, the determining the abnormal data in the to-be-detected machine data according to the difference score sequence and the abnormal data detection range includes: the difference score sequence includes a plurality of difference scores;

[0043] If the difference score is not within the abnormal data detection range, the to-be-detected machine data corresponding to the difference score is taken as abnormal data.

[0044] In a second aspect, the present application further provides a semiconductor machine data anomaly detection device, the device comprising:

[0045] A data acquisition module is used to acquire machine sample data and machine data to be tested;

[0046] A distance similarity calculation module, used to calculate the distance similarity matrix corresponding to the machine sample data according to the machine sample data;

[0047] A model training module, used to input the machine sample data into an initial anomaly detection model, determine the attention weight matrix corresponding to the machine sample data; determine the target loss function of the initial anomaly detection model according to the distance similarity matrix and the attention weight matrix, and obtain a trained anomaly detection model based on the target loss function;

[0048] The abnormality judgment module inputs the machine data to be detected into the trained abnormality detection model to determine the abnormal data in the machine data to be detected.

[0049] In a third aspect, the present application further provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of the method described in any one of the embodiments of the first aspect are implemented.

[0050] In a fourth aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps of the method described in any one of the embodiments of the first aspect above are implemented.

[0051] In a fifth aspect, the present application further provides a computer program product, including a computer program, which, when executed by a processor, implements the steps of the method described in any one of the embodiments of the first aspect above.

[0052] The above-mentioned semiconductor machine data anomaly detection method, device and computer equipment, for model training of the initial anomaly detection model, can accurately determine the relationship between the distance and similarity corresponding to each data point in the machine sample data based on the machine sample data, and obtain the corresponding distance similarity matrix; based on the initial anomaly detection model, it can determine the relationship between the distance and attention score corresponding to each data point in the machine sample data, and obtain the corresponding attention weight matrix; then, the target loss function of the initial anomaly detection model is constructed through the distance similarity matrix and the attention weight matrix, and the initial anomaly detection model can be trained based on the relationship between the similarity and attention score corresponding to each data point to obtain a trained anomaly detection model, thereby improving the performance of the trained anomaly detection model; then, based on the trained anomaly detection model, anomaly detection of the machine data to be detected is realized, which can effectively improve the accuracy and reliability of anomaly detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the drawings required for use in the embodiments of the present application or related technical descriptions will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0054] Figure 1 A diagram of an application environment of a semiconductor machine data anomaly detection method in one embodiment;

[0055] Figure 2 A schematic diagram of a flow chart of a method for detecting abnormal data of a semiconductor machine in one embodiment;

[0056] Figure 3 A schematic diagram of raw sensor data completion in one embodiment;

[0057] Figure 4 is a schematic diagram of a process for preprocessing raw sensor data in one embodiment;

[0058] Figure 5 A schematic diagram of a process for determining a target loss function in one embodiment;

[0059] Figure 6 A schematic diagram of a flow chart of a step of determining abnormal data in the machine data to be detected in one embodiment;

[0060] Figure 7 It is a flowchart of a method for detecting abnormal data of a semiconductor machine in a specific embodiment;

[0061] Figure 8is a structural block diagram of a semiconductor machine data anomaly detection device in one embodiment;

[0062] Fig. 9 FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0063] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0064] The semiconductor machine data anomaly detection method provided by the embodiment of the present application can be applied to Figure 1 In the application environment shown. Among them, the terminal 102 communicates with the server 104 through the network. The data storage system can store the data that the server 104 needs to process. The data storage system can be integrated on the server 104, or it can be placed on the cloud or other network servers. The server 104 can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides cloud computing services.

[0065] In an exemplary embodiment, Figure 2 As shown, Figure 2 A semiconductor machine data anomaly detection method is provided. The method is applied to Figure 1 The terminal in the example is used for explanation. It can be understood that the method can also be applied to a server, and can also be applied to a system including a terminal and a server, and is implemented through the interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0066] Step S201, obtaining machine sample data and an initial anomaly detection model.

[0067] Among them, the machine sample data refers to the data collected from the semiconductor machine for training the initial anomaly detection model. The machine sample data may include, but is not limited to, various types of sensor data; for example, the machine sample data may be temperature data collected based on a temperature sensor, or pressure data collected based on a pressure sensor, which is not specifically limited here. Optionally, the machine sample data may be data composed of trace, label, and timestamp; wherein trace represents data based on real-time monitoring by sensors; label represents the anomaly type corresponding to historical data; anomaly types include normal and abnormal; normal is represented by "0", and abnormal is represented by "1"; timestamp represents the timestamp of each data point; it can be understood that the machine sample data includes multiple data points; each data point has a corresponding timestamp.

[0068] The initial anomaly detection model refers to an untrained anomaly detection model or an anomaly detection model that processes data of the same type; the initial anomaly detection model can be, but is not limited to, a Transformer model, which is not specifically limited here. It should be noted that based on the initial anomaly detection model, the machine sample data can be reconstructed to obtain sample reconstructed data corresponding to the machine sample data, and then based on the difference between the sample reconstructed data and the machine sample data, the initial anomaly detection model is optimized to meet the data reconstruction requirements.

[0069] Step S202: Calculate the distance similarity matrix corresponding to the machine sample data according to the machine sample data.

[0070] Among them, the distance similarity matrix is ​​a two-dimensional symmetric matrix, which is used to characterize the relationship between the distance and similarity corresponding to each pair of data points in the machine sample data; each data in the distance similarity matrix represents the similarity value between each pair of data points in the distance dimension; the closer the distance between two data points, the greater the similarity value, which means that the similarity between the two data points is higher.

[0071] Optionally, based on the machine sample data, the method for calculating the distance similarity matrix corresponding to the machine sample data may be: based on the machine sample data, determining the distance between two data points in sequence; based on the distance between two data points, determining the similarity value corresponding to two data points; based on the similarity values ​​corresponding to all two data points, determining the distance similarity matrix corresponding to the machine sample data; wherein the calculation formula for the similarity value is as shown in formula (1).

[0072] (1)

[0073] In formula (1), n ​​represents the data length of the machine sample data; i represents the position of the data point in the machine sample data; j represents the position of the data point in the machine sample data; , represents the similarity value between the data point at the i-th position and the data point at the j-th position. For example, assuming that the machine sample data is [1, 3, 4], it can be concluded that the data length of the machine sample data is 3, so i=j=0, 1, 2; based on formula (1), it can be calculated that m[0][0]=0, m[0][1]=-1, m[0][2]=-4, and so on, the similarity values ​​corresponding to all two data points can be calculated; based on the similarity values ​​corresponding to all two data points, it can be concluded that the distance similarity matrix corresponding to the machine sample data is [[0, -1, -4], [-1, 0, -1], [-4, -1, 0]]; based on the above distance similarity matrix, it can be seen that the closer the distance between two data points, the greater the similarity value, which proves that the two data points are more similar.

[0074] Step S203: input the machine sample data into the initial anomaly detection model to determine the attention weight matrix corresponding to the machine sample data.

[0075] The attention weight matrix is ​​a two-dimensional matrix of n×n, where n is equal to the data length of the machine sample data. Each data in the attention weight matrix represents the attention score between two data points. The attention score is used to quantify the importance of each data point in the input machine sample data. The higher the attention score, the more important the corresponding data point is, and the greater its contribution to the reconstruction of the input data.

[0076] Exemplarily, the Transformer model is selected as the initial anomaly detection model. Based on the Transformer model, the machine sample data is linearly transformed to obtain three matrices Q, K and V. Then, the Q matrix and the K matrix are multiplied to obtain the corresponding attention weight matrix. Among them, Q represents the query vector; K represents the key vector; V represents the numerical vector. For example, assuming that the data length of the machine sample data is 128, that is, the machine sample data is one-dimensional data of 1×128; based on the Transformer model, three trainable two-dimensional parameter matrices of size 128×128 are generated; wherein, the two-dimensional parameter matrix needs to be determined according to the training of the model, and is not specifically limited here; the 1×128 machine sample data is multiplied by three trainable two-dimensional parameter matrices of size 128×128 respectively, and the machine sample data is linearly changed, and the corresponding Q, K and V matrices of size 1×128 can be obtained; further, the 1×128 Q matrix and the 1×128 K matrix are matrix multiplied to obtain an attention weight matrix of size 1281×128; wherein, any row in the attention weight matrix represents the attention score between a data point and all other data points, and any data represents the attention score between two data points, for example, the data in the i-th row and j-th column represents the attention score between the data point at the i-th position and the data point at the j-th position.

[0077] Step S204, determining the target loss function of the initial anomaly detection model according to the distance similarity matrix and the attention weight matrix.

[0078] Among them, the target loss function can be used to evaluate the performance of the initial anomaly detection model during the training process. In an exemplary embodiment, when the Transformer model is selected as the initial anomaly detection model to reconstruct the machine sample data, the reconstruction effect of the initial anomaly detection model can be determined based on the loss value of the target loss function.

[0079] Optionally, the method for determining the target loss function of the initial anomaly detection model based on the distance similarity matrix and the attention weight matrix may be: determining the target loss function of the initial anomaly detection model based on the difference between the distance similarity matrix and the attention weight matrix.

[0080] It should be noted that, under normal circumstances, the data points near the abnormal data points are also likely to be abnormal data points; therefore, when reconstructing the machine sample data based on the initial anomaly detection model, if there are abnormal data points in the machine sample data, it can be concluded based on the distance similarity matrix that the closer the data point is to the abnormal data point, the higher the similarity between the two. At this time, when reconstructing the abnormal data point based on the attention weight matrix, if the corresponding attention score between the abnormal data point and the data points closer to it is higher, the reconstruction will pay more attention to the data points closer to the abnormal data point, making it more likely that the reconstructed data is also abnormal data, which will affect the accuracy of abnormal data detection.

[0081] To avoid the above problems, when reconstructing abnormal data points, it is necessary to ensure that more attention can be paid to data points that are far away from the abnormal data points, and reconstruct the abnormal data points. Based on this, it can be ensured that the abnormal data points are corrected to normal data points, laying the foundation for improving the accuracy of anomaly detection. It can be understood that the greater the difference between the attention weight matrix and the distance similarity matrix, the stronger the model's ability to correct abnormal data points to normal data points. Therefore, based on the difference between the distance similarity matrix and the attention weight matrix, constructing the target loss function of the initial anomaly detection model can effectively improve the performance of the initial anomaly detection model.

[0082] Step S205 , training the initial anomaly detection model according to the machine sample data and the target loss function until the target loss function converges to a preset value range, thereby obtaining a trained anomaly detection model.

[0083] Among them, the preset numerical range needs to be set according to the training requirements of the initial anomaly detection model and is not specifically limited here.

[0084] Exemplarily, a large amount of machine sample data of the same type is obtained to construct a machine sample data set; the machine sample data set is divided into a training set and a test set according to a preset division ratio; based on the training set, the initial anomaly detection model is trained; and based on the stochastic gradient descent optimization algorithm, the target loss function is converged to a preset numerical range; based on the test set, the performance of the trained initial anomaly detection is evaluated; if the trained initial anomaly detection model has not improved in the index on the test set for multiple consecutive times, it means that the current trained initial anomaly detection is close to convergence, and the training can be stopped to obtain a trained anomaly detection model. It should be noted that the preset division ratio needs to be set according to the training requirements of the initial anomaly detection model, and is not specifically limited here.

[0085] Step S206, obtaining the data of the machine to be detected.

[0086] The data of the testing machine to be tested may be, but is not limited to, sensor data collected from a semiconductor machine; for example, the data of the testing machine to be tested may be temperature data or pressure data, which is not specifically limited here.

[0087] Step S207 , inputting the machine data to be detected into the trained anomaly detection model to determine the abnormal data in the machine data to be detected.

[0088] Exemplarily, the machine data to be detected is input into the trained anomaly detection model, and the target reconstruction data corresponding to the machine data to be detected can be obtained; then, based on the machine data to be detected and the target reconstruction data, the abnormal data points in the machine data to be detected can be determined. It can be understood that, based on the trained anomaly detection model, the abnormal data points in the machine data to be detected can be corrected to normal data points to obtain the corresponding target reconstruction data. Therefore, based on the difference between the machine data to be detected and the target reconstruction data, the abnormal data points in the machine data to be detected can be accurately determined, and then the accurate detection of abnormal data in the machine data to be detected can be achieved.

[0089] In the above-mentioned semiconductor machine data anomaly detection method, based on the machine sample data, the relationship between the distance and similarity corresponding to each other data points in the machine sample data can be accurately determined to obtain the corresponding distance similarity matrix; based on the initial anomaly detection model, the relationship between the distance and attention score corresponding to each other data points in the machine sample data can be determined to obtain the corresponding attention weight matrix; when the difference between the attention weight matrix and the distance similarity matrix is ​​greater, the model's ability to correct abnormal data points to normal data points is stronger. Therefore, based on the difference between the distance similarity matrix and the attention weight matrix, constructing the target loss function of the initial anomaly detection model can effectively improve the performance of the initial anomaly detection model; and then based on the trained anomaly detection model, realizing anomaly detection of the machine data to be detected can effectively improve the accuracy and reliability of anomaly detection, which helps to avoid potential failures, shutdowns or safety issues in semiconductor machines, and lays the foundation for improving the stability and reliability of production lines.

[0090] In one embodiment, obtaining machine sample data includes the following steps:

[0091] Step 1: Obtain the original sensor data of the machine.

[0092] The raw sensor data refers to sensor data collected from a semiconductor machine. For example, the raw sensor data may be temperature data collected by a temperature sensor, which is not specifically limited here.

[0093] Step 2: Adjust the data length of the original sensor data to a preset data length to obtain machine sample data.

[0094] The preset data length needs to be set according to the data input requirements corresponding to the initial anomaly detection model, and is not specifically limited here. Optionally, the preset data length can be the minimum data length trace_len allowed by the initial anomaly detection model. min .

[0095] Exemplarily, the preset data length is the minimum data length trace_len allowed by the initial anomaly detection model. min Take the example to illustrate that a large amount of raw sensor data of the same type is obtained to construct a raw sensor data set, and the data length of all raw sensor data in the raw sensor data set is adjusted to the preset data length, that is, the minimum data length trace_len min , obtain the corresponding machine sample data, and then build a machine sample data set based on a large amount of machine sample data of the same type.

[0096] Optionally, adjust the data length of the original sensor data to the minimum data length trace_len min The method can be: if the data length of the original sensor data is greater than the preset data length, the original sensor data is truncated, that is, the data corresponding to the original sensor that exceeds the preset length is truncated and discarded to obtain the machine sample data; if the data length of the original sensor data is less than the preset data length, the original sensor data is completed to obtain the machine sample data.

[0097] Optionally, since the terminal data of most sensor data are basically the same in real scenarios, the method for completing the original sensor data can be: based on the average of the data at the corresponding positions of the terminal of multiple machine sample data, the missing data at the terminal of the original sensor data is completed. The completion diagram is as follows: Figure 3 shown.

[0098] In this embodiment, by adjusting the original sensor data to a preset data length to obtain corresponding machine sample data, the computing power of the initial anomaly detection model can be guaranteed, laying a foundation for improving the performance of the model.

[0099] In one embodiment, Figure 4 As shown, Figure 4 The figure is a flow chart of preprocessing raw sensor data in one embodiment; before adjusting the data length of the raw sensor data to a preset data length and obtaining machine sample data, the method further includes the following steps:

[0100] Step S401, obtaining the timestamp of each data point in the original sensor data of the machine.

[0101] Among them, the timestamp is the time information corresponding to when the data point is collected or generated.

[0102] Step S402: determining the time interval according to the difference between the timestamps of adjacent data points.

[0103] Optionally, since there are multiple data points in the original sensor data, each data point has a corresponding timestamp, and therefore the time difference between all adjacent timestamps can be calculated in turn by traversing the timestamp of each data point to obtain multiple corresponding time difference values; further, the mode of the multiple time difference values ​​is used as the time interval corresponding to the timestamps of adjacent data points to ensure the accuracy of the time interval.

[0104] Step S403: determining the data missing status of the original sensor data according to the duration and time interval of the original sensor data and the data length corresponding to the original sensor data.

[0105] The duration of the raw sensor data is equal to the difference between the timestamp corresponding to the last data point and the timestamp corresponding to the first data point in the raw sensor data; and the data length corresponding to the raw sensor data is equal to the total amount of all data points in the raw sensor data.

[0106] It should be noted that based on the duration and time interval of the original sensor data, the total number of timestamps expected within the duration of the original sensor data can be determined; since each data point has a corresponding timestamp, the number of data points and timestamps is equal; further, based on the total number of timestamps expected within the duration of the original sensor data, the total number of data points expected within the duration of the original sensor data can be determined; if there is a difference between the total number of expected data points and the actual data length corresponding to the original sensor data, it can be determined that there are data missing points in the original sensor data.

[0107] Step S404: pre-processing the original sensor data according to the data missing situation.

[0108] Preferably, step S403 determines the data missing status of the original sensor data according to the duration and time interval of the original sensor data and the data length corresponding to the original sensor data, and includes the following steps:

[0109] Step 1: determine whether the ratio of the duration of the original sensor data to the time interval is equal to the data length corresponding to the original sensor data. If yes, proceed to step 3; if not, proceed to step 2.

[0110] Step 2: If the ratio of the duration of the original sensor data to the time interval is not equal to the data length corresponding to the original sensor data, it is determined that there are data missing points in the original sensor data.

[0111] Step 3: If the ratio of the duration of the original sensor data to the time interval is equal to the data length corresponding to the original sensor data, it is determined that there is no data missing point in the original sensor data.

[0112] For example, the duration of the original sensor data is recorded as T, T=time n -time 1 , where time n , is the timestamp of the last data point in the raw sensor data; time 1 , is the timestamp of the first data point in the raw sensor data; the time interval is recorded as t; the data length corresponding to the raw sensor is recorded as L; further, determine whether the ratio of the time length T of the raw sensor data to the time interval t is equal to the data length L corresponding to the raw sensor data; if the ratio of the time length T of the raw sensor data to the time interval t is not equal to the data length L corresponding to the raw sensor data, that is, T / t≠L, then it is determined that there is no data missing point in the raw sensor data. If the ratio of the time length T of the raw sensor data to the time interval t is equal to the data length L corresponding to the raw sensor data, that is, T / t=L, then it is determined that there is a data missing point in the raw sensor data, and the raw sensor data needs to be preprocessed.

[0113] Preferably, step S404 pre-processes the raw sensor data according to the data missing situation, including the following steps:

[0114] Step 1: If there are data missing points in the original sensor data, the data missing position corresponding to the data missing point is determined according to the timestamp of each data point in the original sensor data and the time interval corresponding to the original sensor data according to the preset detection rules.

[0115] Among them, the preset detection rule needs to be set according to the data missing position detection requirements, and is not specifically limited here; optionally, the preset detection rule can be to determine whether there is a data missing point between two timestamps based on the ratio of the difference between the two timestamps to the time interval, and thus traverse all data points in the original sensor data to determine the data missing positions corresponding to all data missing points; for example, for any two timestamps time1 and time2, time2>time1; if (time2-time1) / t=1, it indicates that there is no data missing point between the two timestamps; if (time2-time1) / t=n(n>1), it indicates that there is a data missing point between the two timestamps.

[0116] Optionally, in another exemplary embodiment, in order to further reduce the time complexity and processing time of data processing, a detection rule is preset, and the duration of the original sensor data can be divided into multiple time periods with a preset time step; for each of the multiple time periods, the data missing situation of each time period is determined according to the duration of each time period, the time interval and the data length corresponding to the time period; if the ratio of the duration of the time period to the time interval is not equal to the data length corresponding to the time period, it is determined that there is a data missing point in the time period; according to the ratio of the difference between the two timestamps in the time period to the time interval, it is determined whether there is a data missing point between the two timestamps in the time period, and all data points in each time period are traversed in this way. Among them, according to the ratio of the difference between the two timestamps and the time interval, determining whether there is a data missing point between the two timestamps can be achieved by the method recorded in the above embodiment, which will not be repeated here.

[0117] Step 2: Complete the data at the missing locations according to the preset completion rules.

[0118] Among them, the preset completion rules need to be set according to the actual data completion requirements. For example, the data at the data missing position can be completed based on the previous normal data point of the data missing position; or the data at the data missing position can be completed based on the next normal data point of the data missing position. The preset completion rules are not specifically limited here.

[0119] In this embodiment, the data missing situation of the original sensor data is determined based on the duration, time interval and data length corresponding to the original sensor data, which is conducive to determining the data missing situation in scenarios with few data missing points. Based on this, the time complexity and processing time of the original sensor data processing can be greatly reduced, and the processing efficiency of the original sensor data can be effectively improved. Based on the preset detection rules, the accuracy of the data missing position detection can be improved; and then the data missing position is supplemented based on the preset completion rules, which can ensure the integrity and reliability of the original sensor data and improve the integrity and reliability of the machine sample data.

[0120] In one embodiment, Figure 5 As shown, Figure 5 The flowchart of the step of determining the target loss function in one embodiment is as follows; the target loss function of the initial anomaly detection model is determined according to the distance similarity matrix and the attention weight matrix, including the following steps:

[0121] Step S501, determining sample reconstruction data according to machine sample data and an initial anomaly detection model.

[0122] Optionally, the initial anomaly detection model is a Transformer model; it should be noted that based on the Transformer model, the machine sample data can be reconstructed to obtain sample reconstructed data corresponding to the machine sample data, and the specific reconstruction process will not be described in detail here.

[0123] Step S502 , determining a first difference between the machine sample data and the sample reconstruction data according to the machine sample data and the sample reconstruction data.

[0124] The first difference is determined by the square difference between the machine sample data and the sample reconstruction data, and is used to characterize the reconstruction capability of the initial anomaly detection model.

[0125] Step S503: Determine a second difference between the distance similarity matrix and the attention weight matrix according to the distance similarity matrix and the attention weight matrix.

[0126] Among them, the second difference is determined by the KL divergence (Kullback-Leibler Divergence) corresponding to the distance similarity matrix and the attention weight matrix, which is used to characterize the ability of the initial anomaly detection model to correct abnormal data points in the input data into normal data points.

[0127] Among them, KL divergence can measure the difference between two distributions; during the initial anomaly detection model training process, the input of KL divergence is the distance similarity matrix and the attention weight matrix, and the output is the KL divergence value corresponding to the distance similarity matrix and the attention weight matrix; among them, the range of KL divergence value is 0 to positive infinity; the smaller the KL divergence value, the more similar the distance similarity matrix and the attention weight matrix are; the larger the KL divergence value, the less similar the distance similarity matrix and the attention weight matrix are.

[0128] It should be noted that, for an abnormal data point, the probability that the data point that is closer to it is also an abnormal data point is relatively high, and the probability that the data point that is farther away from it is a normal data point is relatively high; because, in order to ensure that the abnormal data point can be corrected to a normal data point when reconstructing the abnormal data point, more attention needs to be paid to the data points that are farther away from the abnormal data point; therefore, in the initial anomaly detection model training process, it is necessary to ensure that the KL divergence values ​​corresponding to the distance similarity matrix and the attention weight matrix are as large as possible, that is, to ensure that the distance similarity matrix and the attention weight matrix are as dissimilar as possible, based on which it can be ensured that the model can correct the abnormal data point to a normal data point.

[0129] It can be understood that the larger the KL divergence value corresponding to the distance similarity matrix and the attention weight matrix is, the more dissimilar the distance similarity matrix and the attention weight matrix are; because each data in the distance similarity matrix represents the similarity value between two data points, and each data in the attention weight matrix represents the attention score between two data points; when the distance similarity matrix and the attention weight matrix are more dissimilar, for any two data points, the difference between their corresponding similarity values ​​and attention scores is greater; for example, the reference data point in the machine sample data is recorded as i, and the target data point is recorded as j; among them, when the reference data i is fixed, the target data point j represents all data points in the machine sample data; the attention scores corresponding to the reference data point i and the target data point j are recorded as n[i][j]; the similarity values ​​corresponding to the reference data point i and the target data point j are recorded as m[i][j]; based on the distance similarity matrix corresponding to the machine sample data, it can be seen that the greater the distance between the reference data point i and the target data point j, the smaller the similarity value m[i][j], and the lower the similarity between the two; when the distance similarity matrix and the attention weight matrix are not similar, the greater the distance between the reference data point i and the target data point j, the greater the attention score n[i][j], and the more important the target data point j is to the reference data point i. It can be understood that the role of the distance similarity matrix is ​​to obtain an attention weight matrix that is not similar to itself, so as to ensure that the reference data point i pays more attention to the target data point j that is farther away from it during the reconstruction process.

[0130] Step S504: determining a target loss function of the initial anomaly detection model according to the first difference and the second difference.

[0131] Optionally, a target loss function of the initial anomaly detection model is constructed according to the first difference and the second difference, and the obtained target loss function is shown in formula (2).

[0132] (2)

[0133] In formula (2), , represents the target loss function; x, represents the input data of the initial anomaly detection model; the input data includes the machine sample data; y, represents the output data of the initial anomaly detection model; the output data includes the sample reconstruction data; X, represents the distance similarity matrix; Y, represents the attention weight matrix; , represents the KL divergence value corresponding to the distance similarity matrix X and the attention weight matrix Y; it can be understood that, , represents the first degree of difference; , indicating the second degree of difference; , indicating the importance of the first difference; , indicating the importance of the second difference; it should be noted that and The two parameters need to be set according to the actual anomaly detection requirements and are not specifically limited here; for example, Set to 0.5; It is set to 1.5 to ensure that the initial anomaly detection model has the ability to reconstruct while constraining the model to reconstruct data points with a longer reference distance through the KL divergence value.

[0134] It should be noted that when training the initial anomaly detection model based on the target loss function, when the target loss function does not converge to the preset numerical range, it is necessary to optimize the parameters in the initial anomaly detection model to achieve optimal adjustment of the attention weight matrix until the target loss function converges to the preset numerical range. At this time, the attention weight matrix has been optimized, that is, the dissimilarity between the attention weight matrix and the distance similarity matrix has reached the expected level.

[0135] In this embodiment, a first difference is determined based on machine sample data and sample reconstruction data; a second difference is determined based on a distance similarity matrix and the attention weight matrix; and a target loss function of an initial anomaly detection model is constructed based on the first difference and the second difference, which can ensure that the initial anomaly detection model has reconstruction capability while reconstructing data points with a longer reference distance through the KL divergence value constraint model; and training the initial anomaly detection model based on the target loss function can improve the performance of the trained anomaly detection model.

[0136] In one embodiment, Figure 6 As shown, Figure 6 The flowchart of the step of determining abnormal data in the machine data to be detected in one embodiment is as follows; the machine data to be detected is input into the trained abnormality detection model, and the abnormal data in the machine data to be detected is determined, including the following steps:

[0137] Step S601 : input the machine data to be detected into the trained anomaly detection model to determine the target reconstructed data corresponding to the machine data to be detected.

[0138] It should be noted that the performance of the trained initial anomaly detection is evaluated based on the test set in the machine sample data set. If the indicators of the trained initial anomaly detection model on the test set have not been improved for multiple consecutive times, it means that the current trained initial anomaly detection is close to convergence, and the training can be stopped. At this time, the trained initial anomaly detection can be used as the trained anomaly detection model. In an actual production environment, after the machine data to be detected undergoes the same data processing process as the machine sample data, the target reconstructed data corresponding to the machine data to be detected can be determined by the trained anomaly detection model; it should be noted that the method of detecting and processing missing points of the machine data to be detected, and adjusting the data length of the machine data to be detected can be implemented by the method described in the above embodiment, and the specific details will not be repeated here.

[0139] Step S602 : determining a difference score sequence corresponding to the machine data to be detected and the target reconstructed data according to the machine data to be detected and the target reconstructed data.

[0140] It should be noted that before inputting the machine data to be detected into the trained anomaly detection model, the data length of the machine data to be detected needs to be adjusted to the preset data length to obtain the corresponding target detection machine data; based on the trained anomaly detection model, the target detection machine data is reconstructed to obtain the target reconstructed data corresponding to the target detection machine data; wherein, the data length of the target reconstructed data is equal to the preset data length.

[0141] Optionally, according to the machine data to be detected and the target reconstructed data, the difference score sequence corresponding to the machine data to be detected and the target reconstructed data can be determined as follows: taking the data length of the machine data to be detected being equal to the preset data length as an example, each data point in the machine data to be detected is subtracted from the corresponding data point in the target reconstructed data, and the difference score sequence corresponding to the machine data to be detected and the target reconstructed data can be obtained; for example, assuming that the machine data to be detected is: 0, 0, 0, 100, 0, 0; the target reconstructed data is: 0, 0, 0, 0, 0, 0; it can be determined that the difference score sequence corresponding to the machine data to be detected and the target reconstructed data is: 0, 0, 0, 100, 0, 0.

[0142] Step S603: determine the abnormal data detection range according to the difference score sequence.

[0143] The abnormal data detection range is used to determine the abnormal difference scores in the difference score sequence, and then determine the abnormal data in the sensor data to be detected based on the abnormal difference scores.

[0144] Preferably, step S603, determining the abnormal data detection range according to the difference score sequence, includes the following steps:

[0145] Step 1: According to the difference score sequence, determine the mean and standard deviation corresponding to the difference score sequence.

[0146] Step 2: Determine a first detection threshold and a second detection threshold according to the mean and the standard deviation; the first detection threshold is smaller than the second detection threshold.

[0147] Step 3: Determine an abnormal data detection range based on the first detection threshold and the second detection threshold.

[0148] Exemplarily, based on the difference score sequence, the mean and standard deviation corresponding to the difference score sequence can be calculated; the mean corresponding to the difference score sequence is denoted as μ, and the standard deviation is denoted as σ; based on the mean μ and the standard deviation denoted as σ, the first detection threshold and the second detection threshold can be determined respectively; wherein the first detection threshold is less than the second detection threshold; for example, based on the mean μ and the standard deviation denoted as σ, μ-3σ can be used as the first detection threshold, and μ+3σ can be used as the second detection threshold, wherein μ-3σ is less than μ+3σ, and thus the abnormal data detection range can be determined as (μ-3σ, μ+3σ).

[0149] Step S604: determining abnormal data in the sensor data to be detected according to the difference score sequence and the abnormal data detection range.

[0150] It should be noted that the difference score sequence includes multiple difference scores.

[0151] Preferably, step S604 determines the abnormal data in the machine data to be detected according to the difference score sequence and the abnormal data detection range, including the following steps:

[0152] Step 1: Determine whether the difference score is within the abnormal data detection range. If yes, proceed to step 3; if no, proceed to step 2.

[0153] Step 2: If the difference score is not within the abnormal data detection range, the machine data to be detected corresponding to the difference score is taken as abnormal data.

[0154] Step 3: If the difference score is within the abnormal data detection range, the machine data to be detected corresponding to the difference score is determined to be normal data.

[0155] Exemplarily, the machine data to be detected is input into a trained anomaly detection model to determine the target reconstructed data corresponding to the machine data to be detected; then, each data point in the machine data to be detected is subtracted from the corresponding data point in the target reconstructed data to obtain a corresponding plurality of difference scores; based on the plurality of difference scores, a difference score sequence corresponding to the machine data to be detected and the target reconstructed data is obtained; based on the difference score sequence, a mean μ and a standard deviation σ corresponding to the difference score sequence are calculated; then, based on the mean μ and the standard deviation σ, an abnormal data detection range is determined to be (μ-3σ, μ+3σ), wherein μ-3σ is less than μ+3σ; further, each difference score in the difference score sequence is judged to be within the abnormal data detection range (μ-3σ, μ+3σ); if the difference score is not within the abnormal data detection range, the data point corresponding to the difference score is taken as an abnormal data point; if the difference score is within the abnormal data detection range, the data point corresponding to the difference score is determined to be a normal data point.

[0156] In this embodiment, based on the machine data to be detected and the target reconstructed data, the corresponding difference score sequence can be determined; further based on the difference score sequence, the abnormal data detection range can be accurately determined; based on the abnormal data detection range, it can be accurately judged whether the machine data to be detected corresponding to each difference score in the difference score sequence is abnormal data, based on which accurate detection of abnormal data is achieved, thereby improving the accuracy and reliability of abnormal data detection.

[0157] In a specific embodiment, during the initial anomaly detection model training process, in order to further reduce the initial anomaly detection model's ability to reconstruct abnormal data points, optionally, a random Mask method can be used to mask the machine sample data in training; wherein, the data masking ratio parameter needs to be set according to actual reconstruction requirements, for example, it can be taken as 0.1, and no specific limitation is made here; masking the machine sample data in training based on a random Mask method can force the model to learn in the absence of some information, which can further improve the generalization ability of the model.

[0158] It should be noted that, because most of the data points in the sensor data to be detected are normal and only a small number of data points are abnormal, the initial anomaly detection model during the training process will be more inclined to reconstruct most of the data points to be normal, so that the loss value of the target loss function is smaller, and based on this, a trained anomaly detection model can be obtained. For the input of an abnormal data point, the trained anomaly detection model will correct it to a normal data point to achieve the purpose of abnormal data detection; among them, the reason why the trained anomaly detection model can correct the abnormal data point to a normal data point is that the mean and standard deviation of the normal data point and the abnormal data point in the intermediate state are consistent, so the final output is also consistent. Therefore, in order to reconstruct most of the normal data points to be normal, the trained anomaly detection model will ignore a small number of abnormal data points.

[0159] In a specific embodiment, see Figure 7 , taking the semiconductor machine data anomaly detection method applied to the actual production environment of the semiconductor machine as an example for explanation, the semiconductor machine data anomaly detection method includes the following steps:

[0160] Step S701, obtaining the data of the machine to be detected.

[0161] The machine data to be detected may be sensor data.

[0162] Step S702, determine whether there are data missing points in the machine data to be detected. If yes, execute step S703; if no, execute step S704.

[0163] Specifically, the method of determining whether there are data missing points in the data of the machine to be detected can be implemented by the method described in the above embodiment, which will not be described in detail here.

[0164] Step S703, if there are data missing points in the machine data to be detected, determine the data missing position corresponding to the data missing point according to the timestamp of each data point in the machine data to be detected and the time interval corresponding to the machine data to be detected according to the preset detection rules; and complete the data missing position according to the preset completion rules.

[0165] Among them, the preset detection rules and the preset completion rules can be set in the manner described in the above embodiments, which will not be described in detail here.

[0166] Step S704 , if there is no data missing point in the machine data to be detected, the data length of the machine data to be detected is adjusted to a preset data length to obtain the target detection machine data.

[0167] Among them, the preset data length needs to be set according to the requirements of the trained anomaly detection model for the corresponding input data, and no specific limitation is made here.

[0168] Step S705 , input the target detection machine data into the trained anomaly detection model to obtain target reconstruction data corresponding to the target detection machine data.

[0169] Step S706, subtracting the target detection machine data from the target reconstruction data to determine a corresponding difference score sequence.

[0170] The difference score sequence includes multiple difference scores.

[0171] Step S707: for each difference score in the difference score sequence, determine whether each difference score is within the abnormal data detection range. If yes, execute step S708; if no, execute step S709.

[0172] Preferably, the abnormal data detection range is (μ-3σ, μ+3σ); wherein μ represents the mean of the difference score sequence corresponding to the target detection machine data and the target reconstruction data; σ represents the standard deviation of the difference score sequence corresponding to the target detection machine data and the target reconstruction data; wherein μ-3σ is less than μ+3σ.

[0173] Step S708: If the difference score is within the abnormal data detection range, the to-be-detected machine data corresponding to the difference score is determined to be a normal data point.

[0174] Step S709: if the difference score is not within the abnormal data detection range, the to-be-detected machine data corresponding to the difference score is taken as an abnormal data point.

[0175] Based on the above-mentioned semiconductor machine data anomaly detection method, abnormal data points in the semiconductor machine can be detected accurately and timely, which helps to avoid potential failures, shutdowns or safety issues in the semiconductor machine, and lays the foundation for improving the stability and reliability of the production line.

[0176] The above-mentioned semiconductor machine data anomaly detection method, in the first aspect, based on the machine sample data, can accurately determine the relationship between the distance and similarity corresponding to each data point in the machine sample data, and obtain the corresponding distance similarity matrix; based on the initial anomaly detection model, can determine the relationship between the distance and attention score corresponding to each data point in the machine sample data, and obtain the corresponding attention weight matrix; when the difference between the attention weight matrix and the distance similarity matrix is ​​greater, the model's ability to correct abnormal data points to normal data points is stronger, therefore, based on the difference between the distance similarity matrix and the attention weight matrix, constructing the target loss function of the initial anomaly detection model can effectively improve the performance of the initial anomaly detection model; then based on the trained anomaly detection model, the anomaly detection of the machine data to be detected is realized, which can effectively improve the accuracy and reliability of anomaly detection, which helps to avoid potential failures, shutdowns or safety issues in semiconductor machines, and lays the foundation for improving the stability and reliability of the production line. In the second aspect, by adjusting the original sensor data to the preset data length to obtain the corresponding machine sample data, the computing power of the initial anomaly detection model can be guaranteed, laying the foundation for improving the performance of the model. Thirdly, according to the duration, time interval and data length of the original sensor data, the data missing situation of the original sensor data is determined, which is conducive to the determination of data missing situation in the scenario with few data missing points. Based on this, the time complexity and processing time of the original sensor data processing can be greatly reduced, and the processing efficiency of the original sensor data can be effectively improved. Based on the preset detection rules, the accuracy of the data missing position detection can be improved; and then the data missing position is supplemented based on the preset completion rules, which can ensure the integrity and reliability of the original sensor data. Fourthly, the target loss function of the initial anomaly detection model is constructed based on the first difference and the second difference, which can ensure that the initial anomaly detection model has the reconstruction capability and reconstructs the data points with a long reference distance through the KL divergence value constraint model; based on the target loss function, the initial anomaly detection model is trained, which can improve the performance of the trained anomaly detection model. Fifthly, based on the difference score sequence corresponding to the machine data to be detected and the target reconstructed data and the abnormal data detection range, it can accurately determine whether the machine data to be detected corresponding to each difference score in the difference score sequence is abnormal data, based on which the accurate detection of abnormal data is achieved, and the accuracy and reliability of abnormal data detection are improved.

[0177] It should be understood that, although the various steps in the flowcharts involved in the above-mentioned embodiments are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps does not have a strict order restriction, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-mentioned embodiments can include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the steps or stages in other steps.

[0178] Based on the same inventive concept, the embodiment of the present application also provides a semiconductor machine data anomaly detection device for implementing the semiconductor machine data anomaly detection method involved above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above method, so the specific limitations in one or more semiconductor machine data anomaly detection device embodiments provided below can refer to the limitations of the semiconductor machine data anomaly detection method above, and will not be repeated here.

[0179] In an exemplary embodiment, Figure 8 As shown, Figure 8 8 is a structural block diagram of a semiconductor machine data anomaly detection device in one embodiment; the semiconductor machine data anomaly detection device includes: a data acquisition module 801, a distance similarity calculation module 802, a model training module 803 and an anomaly judgment module 804, wherein:

[0180] The data acquisition module 801 is used to acquire machine sample data and machine data to be detected.

[0181] The data acquisition module 801 is also used to acquire an initial anomaly detection model.

[0182] The distance similarity calculation module 802 is used to calculate the distance similarity matrix corresponding to the machine sample data according to the machine sample data.

[0183] The model training module 803 is used to input the machine sample data into the initial anomaly detection model to determine the attention weight matrix corresponding to the machine sample data; determine the target loss function of the initial anomaly detection model based on the distance similarity matrix and the attention weight matrix, and obtain the trained anomaly detection model based on the target loss function.

[0184] Specifically, the model training module 803 is also used to train the initial anomaly detection model according to the machine sample data and the target loss function, until the target loss function converges to a preset numerical range, thereby obtaining a trained anomaly detection model.

[0185] The abnormality determination module 804 inputs the machine data to be detected into the trained abnormality detection model to determine the abnormal data in the machine data to be detected.

[0186] The above-mentioned semiconductor machine data anomaly detection device can accurately determine the relationship between the distance and similarity corresponding to each data point in the machine sample data based on the machine sample data, and obtain the corresponding distance similarity matrix; based on the initial anomaly detection model, it can determine the relationship between the distance and attention score corresponding to each data point in the machine sample data, and obtain the corresponding attention weight matrix; when the difference between the attention weight matrix and the distance similarity matrix is ​​greater, the model's ability to correct abnormal data points to normal data points is stronger, therefore, based on the difference between the distance similarity matrix and the attention weight matrix, constructing the target loss function of the initial anomaly detection model can effectively improve the performance of the initial anomaly detection model; and then based on the trained anomaly detection model, realizing anomaly detection of the machine data to be detected, can effectively improve the accuracy and reliability of anomaly detection, which helps to avoid potential failures, shutdowns or safety issues in semiconductor machines and improve the stability and reliability of the production line.

[0187] In one embodiment, the data acquisition module 801 is also used to

[0188] Obtain the raw sensor data of the machine;

[0189] The data length of the original sensor data is adjusted to the preset data length to obtain the machine sample data.

[0190] In one embodiment, the data acquisition module 801 is also used to

[0191] Get the timestamp of each data point in the machine's raw sensor data;

[0192] Determine the time interval based on the difference between the timestamps of adjacent data points;

[0193] Determine the data missing status of the original sensor data according to the duration and time interval of the original sensor data and the data length corresponding to the original sensor data;

[0194] According to the data missing situation, the raw sensor data is preprocessed.

[0195] In one embodiment, the data acquisition module 801 is also used to

[0196] If the ratio of the duration of the original sensor data to the time interval is not equal to the data length corresponding to the original sensor data, it is determined that there are data missing points in the original sensor data;

[0197] If the ratio of the duration of the original sensor data to the time interval is equal to the data length corresponding to the original sensor data, it is determined that there is no data missing point in the original sensor data.

[0198] In one embodiment, the data acquisition module 801 is also used to

[0199] If there are data missing points in the original sensor data, the data missing position corresponding to the data missing point is determined according to the timestamp of each data point in the original sensor data and the time interval corresponding to the original sensor data according to the preset detection rules;

[0200] According to the preset completion rules, data is completed at the missing locations.

[0201] In one embodiment, the model training module 803 is also used to

[0202] Determine sample reconstruction data based on machine sample data and initial anomaly detection model;

[0203] Determine a first difference between the machine sample data and the sample reconstruction data according to the machine sample data and the sample reconstruction data;

[0204] Determine, according to the distance similarity matrix and the attention weight matrix, a second difference corresponding to the distance similarity matrix and the attention weight matrix;

[0205] According to the first difference and the second difference, a target loss function of the initial anomaly detection model is determined.

[0206] In one embodiment, the abnormality determination module 804 is also used to

[0207] Input the machine data to be detected into the trained anomaly detection model to determine the target reconstructed data corresponding to the machine data to be detected;

[0208] Determine the difference score sequence corresponding to the machine data to be tested and the target reconstruction data according to the machine data to be tested and the target reconstruction data;

[0209] Determine the abnormal data detection range based on the difference score sequence;

[0210] The abnormal data in the sensor data to be detected is determined according to the difference score sequence and the abnormal data detection range.

[0211] In one embodiment, the abnormality determination module 804 is also used to

[0212] According to the difference score sequence, determine the mean and standard deviation corresponding to the difference score sequence;

[0213] Determine a first detection threshold and a second detection threshold according to the mean and the standard deviation; the first detection threshold is less than the second detection threshold;

[0214] An abnormal data detection range is determined according to the first detection threshold and the second detection threshold.

[0215] In one embodiment, the difference score sequence includes multiple difference scores; the abnormality judgment module 804 is also used to

[0216] If the difference score is not within the abnormal data detection range, the to-be-detected machine data corresponding to the difference score will be taken as abnormal data.

[0217] Each module in the semiconductor machine data anomaly detection device can be implemented in whole or in part by software, hardware, or a combination thereof. Each module can be embedded in or independent of a processor in a computer device in the form of hardware, or can be stored in a memory in a computer device in the form of software, so that the processor can call and execute the operations corresponding to each module.

[0218] In an exemplary embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as shown in FIG. Fig. 9 As shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, referred to as I / O) and a communication interface. Among them, the processor, the memory and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store semiconductor machine data anomaly detection related data. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a semiconductor machine data anomaly detection method is implemented.

[0219] Those skilled in the art will understand that Fig. 9 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0220] In one embodiment, a computer device is further provided, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the steps in the above-mentioned method embodiments when executing the computer program.

[0221] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0222] In one embodiment, a computer program product is provided, including a computer program, which implements the steps in the above method embodiments when executed by a processor.

[0223] A person of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment method can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in the present application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. Non-relational databases may include distributed databases based on blockchains, etc., but are not limited to this. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic unit, a data processing logic unit based on quantum computing, an artificial intelligence (AI) processor, etc., but are not limited to this.

[0224] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0225] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A method for detecting abnormal data of a semiconductor machine; It is characterized in that The method comprises: Obtain machine sample data and initial anomaly detection model; According to the machine sample data, calculating the distance similarity matrix corresponding to the machine sample data; Input the machine sample data into the initial anomaly detection model to determine the attention weight matrix corresponding to the machine sample data; each data in the attention weight matrix represents the attention score between two data points; wherein the attention score is used to quantify the importance of each data point in the input machine sample data; Determining a target loss function of the initial anomaly detection model according to the distance similarity matrix and the attention weight matrix; According to the machine sample data and the target loss function, the initial anomaly detection model is trained until the target loss function converges to a preset value interval to obtain a trained anomaly detection model; Obtain the data of the machine to be tested; Inputting the machine data to be detected into the trained anomaly detection model to obtain target reconstruction data corresponding to the machine data to be detected; Determining abnormal data in the machine data to be detected based on the difference between the machine data to be detected and the target reconstructed data; Determining the target loss function of the initial anomaly detection model according to the distance similarity matrix and the attention weight matrix includes: Determining sample reconstruction data according to the machine sample data and the initial anomaly detection model; Determining a first difference between the machine sample data and the sample reconstruction data according to the machine sample data and the sample reconstruction data; Determining, according to the distance similarity matrix and the attention weight matrix, a second difference between the distance similarity matrix and the attention weight matrix; An objective loss function of the initial anomaly detection model is determined according to the first difference and the second difference.

2. The method according to claim 1, It is characterized in that The obtaining of machine sample data comprises: Obtain the raw sensor data of the machine; The data length of the original sensor data is adjusted to a preset data length to obtain machine sample data.

3. The method according to claim 2, It is characterized in that Before adjusting the data length of the original sensor data to a preset data length to obtain machine sample data, the method further includes: Obtaining a timestamp for each data point in the raw sensor data; Determine the time interval based on the difference between the timestamps of adjacent data points; Determining data missing conditions of the raw sensor data according to the duration of the raw sensor data, the time interval, and the data length corresponding to the raw sensor data; According to the data missing situation, the raw sensor data is preprocessed.

4. The method according to claim 3, It is characterized in that The determining, according to the duration of the raw sensor data, the time interval, and the data length corresponding to the raw sensor data, the data missing condition of the raw sensor data includes: If the ratio of the duration of the raw sensor data to the time interval is not equal to the data length corresponding to the raw sensor data, it is determined that there is a data missing point in the raw sensor data; If the ratio of the duration of the raw sensor data to the time interval is equal to the data length corresponding to the raw sensor data, it is determined that there is no data missing point in the raw sensor data.

5. The method according to claim 4, It is characterized in that Preprocessing the raw sensor data according to the data missing situation includes: If there are data missing points in the original sensor data, determining the data missing position corresponding to the data missing point according to a preset detection rule based on the timestamp of each data point in the original sensor data and the time interval corresponding to the original sensor data; According to the preset completion rules, the data missing position is completed.

6. The method according to claim 1, It is characterized in that The determining of abnormal data in the machine data to be detected based on the difference between the machine data to be detected and the target reconstructed data includes: Determine, according to the machine data to be detected and the target reconstruction data, a difference score sequence corresponding to the machine data to be detected and the target reconstruction data; Determining an abnormal data detection range according to the difference score sequence; Abnormal data in the to-be-detected sensor data is determined according to the difference score sequence and the abnormal data detection range.

7. The method according to claim 6, It is characterized in that The determining of the abnormal data detection range according to the difference score sequence includes: According to the difference score sequence, determining a mean and a standard deviation corresponding to the difference score sequence; Determining a first detection threshold and a second detection threshold according to the mean and the standard deviation; the first detection threshold is less than the second detection threshold; An abnormal data detection range is determined according to the first detection threshold and the second detection threshold.

8. The method according to claim 6, It is characterized in that The determining the abnormal data in the to-be-detected machine data according to the difference score sequence and the abnormal data detection range includes: the difference score sequence includes a plurality of difference scores; If the difference score is not within the abnormal data detection range, the to-be-detected machine data corresponding to the difference score is taken as abnormal data.

9. A semiconductor machine data anomaly detection device, It is characterized in that The device comprises: A data acquisition module is used to acquire machine sample data and machine data to be tested; A distance similarity calculation module, used to calculate the distance similarity matrix corresponding to the machine sample data according to the machine sample data; A model training module is used to input the machine sample data into the initial anomaly detection model, determine the attention weight matrix corresponding to the machine sample data; determine the target loss function of the initial anomaly detection model according to the distance similarity matrix and the attention weight matrix, and obtain a trained anomaly detection model based on the target loss function; each data in the attention weight matrix represents the attention score between two data points; wherein the attention score is used to quantify the importance of each data point in the input machine sample data; An abnormality judgment module inputs the machine data to be detected into the trained abnormality detection model to obtain target reconstruction data corresponding to the machine data to be detected; based on the difference between the machine data to be detected and the target reconstruction data, determines the abnormal data in the machine data to be detected; The model training module is also used to determine sample reconstruction data based on the machine sample data and the initial anomaly detection model; determine a first degree of difference between the machine sample data and the sample reconstruction data based on the machine sample data and the sample reconstruction data; determine a second degree of difference between the distance similarity matrix and the attention weight matrix based on the distance similarity matrix and the attention weight matrix; and determine a target loss function of the initial anomaly detection model based on the first degree of difference and the second degree of difference.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program. It is characterized in that When the processor executes the computer program, the steps of the method according to any one of claims 1 to 8 are implemented.

11. A computer-readable storage medium having a computer program stored thereon, It is characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.

12. A computer program product comprising a computer program, It is characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.