Return Anomaly Prediction System

Abstract

The return anomaly prediction system has a machine-learned model that has a relationship between a data set containing sensor data output during a past substrate return from each of a plurality of sensors provided in a substrate return unit and the degree of return anomaly during said substrate return, and has an estimation unit that takes a data set containing sensor data output during a new substrate return from each of the plurality of sensors as input and estimates and outputs the degree of return anomaly during said new substrate return.

Description

Technology Field

[0001] The present disclosure relates to a return anomaly prediction system. Background Technology

[0002] In semiconductor manufacturing equipment, handling substrates is essential. However, when equipment operates for extended periods, handling errors may occur due to component wear or misalignment. If handling errors occur, the equipment may stop, leading to reduced productivity, or there is a possibility that the substrate itself may be damaged.

[0003] Japanese Patent Publication No. 6325325 proposes a system that compares the output of a sensor during the operation of a wafer handling machine with the output of a sensor stored during position alignment before the operation of the wafer handling machine, and determines that the machine is out of position alignment if the difference exceeds a certain threshold.

[0004] However, in the system disclosed in Japanese Patent Publication No. 6325325, abnormalities are detected by a simple method in which an abnormality is determined if the difference between the sensor output during the operation of the wafer handling machine and the sensor output stored during position alignment before operation exceeds a certain threshold, and thus the detection probability is not high.

[0005] It is desired to provide a return anomaly prediction system capable of improving the probability of detecting return anomalies.

[0006] A return anomaly prediction system according to one aspect of the present disclosure is,

[0007] An estimation unit having a fully trained model that machine-learns the relationship between a data set containing sensor data output during a past substrate transport from each of a plurality of sensors provided in a substrate transport unit and the transport anomaly degree during said substrate transport, and taking a data set containing sensor data output during a new substrate transport from each of the plurality of sensors as input to estimate and output the transport anomaly degree during said new substrate transport.

[0008] It is equipped with. Brief explanation of the drawing

[0009] FIG. 1 is a diagram illustrating the schematic configuration of a substrate processing apparatus according to one embodiment. FIG. 2 is a block diagram illustrating the configuration of a return abnormality prediction system according to one embodiment. Figure 3 is a diagram illustrating an example of sensor data during normal substrate transport. Figure 4 is a drawing for explaining the substrate transport section in the event of a transport abnormality. Figure 5 is a diagram illustrating an example of sensor data when a return abnormality occurs. FIG. 6 is a schematic diagram illustrating the configuration of a learning completion model in the first embodiment. FIG. 7 is a schematic diagram illustrating the configuration of a learning completion model in a second embodiment. FIG. 8 is a schematic diagram illustrating the configuration of a learning completion model in a third embodiment. FIG. 9 is a schematic diagram illustrating the configuration of a learning completion model in the fourth embodiment. FIG. 10 is a schematic diagram illustrating the configuration of a learning completion model in the fifth embodiment. FIG. 11 is a schematic diagram illustrating the configuration of a learning completion model in the sixth embodiment. FIG. 12 is a schematic diagram for explaining a learning completion model in the seventh embodiment. FIG. 13 is a schematic diagram illustrating the configuration of a learning completion model in the eighth embodiment. FIG. 14 is a flowchart for explaining an example of a return abnormality prediction method regarding one embodiment. Specific details for implementing the invention

[0010] A return anomaly prediction system regarding the first aspect of an embodiment is,

[0011] An estimation unit having a fully trained model that machine-learns the relationship between a data set containing sensor data output during a past substrate transport from each of a plurality of sensors provided in a substrate transport unit and the transport anomaly degree during said substrate transport, and taking a data set containing sensor data output during a new substrate transport from each of the plurality of sensors as input to estimate and output the transport anomaly degree during said new substrate transport.

[0012] It is equipped with.

[0013] According to this embodiment, the estimation unit can estimate and output the degree of transport anomaly during a new substrate transport by using a fully trained model that machine-learns the relationship between a data set containing sensor data from a past substrate transport and the degree of transport anomaly during said substrate transport. This is achieved by synthesizing multiple indicator data for a data set containing sensor data during a new substrate transport. As a result, the probability of detecting a transport anomaly can be improved compared to a conventional embodiment in which an anomaly is determined if the difference between the sensor output during substrate transport and the sensor output stored during position alignment before substrate transport exceeds a certain threshold. Furthermore, by using a fully trained model, it is possible to utilize device vibration, sound, image data, etc., which were difficult to handle in conventional embodiments, as sensor data.

[0014] The return anomaly prediction system according to the second embodiment of the embodiment is a return anomaly prediction system according to the first embodiment, and

[0015] The plurality of sensors described above consists of one or more types among a vibration sensor, a sound sensor, an image sensor, a video sensor, a temperature sensor, a device movement speed sensor, a device operation torque sensor, and a device parallelism sensor.

[0016] The return anomaly prediction system according to the third embodiment is a return anomaly prediction system according to the first or second embodiment, and

[0017] An output signal transmitting unit that compares the return abnormality output by the above estimation unit with a predetermined threshold, and if the return abnormality exceeds the threshold, transmits an output signal to an output device for outputting a maintenance notice and / or alarm.

[0018] Provide more of

[0019] The return anomaly prediction system according to the fourth embodiment is a return anomaly prediction system according to any of the first to third embodiments, and

[0020] The above estimation unit takes as input a data set including sensor data from the start of transport during the transport of a new substrate to the present time, and estimates and outputs the transport anomaly degree during the transport of the new substrate.

[0021] The return anomaly prediction system according to the fifth embodiment is a return anomaly prediction system according to any of the first to third embodiments, and

[0022] The above estimation unit takes as input a data set including sensor data from the start of transport to the end of transport during the transport of a new substrate, and estimates and outputs the transport anomaly degree during the transport of the new substrate.

[0023] The return anomaly prediction system according to the sixth embodiment is a return anomaly prediction system according to any of the first to third embodiments, and

[0024] The above estimation unit takes as input a data set including sensor data from the start of transport of the first substrate to the end of transport of the last substrate during the transport of multiple new substrates, and estimates and outputs the transport anomaly degree during the transport of multiple new substrates.

[0025] The return anomaly prediction system according to the seventh embodiment is a return anomaly prediction system according to any of the first to sixth embodiments, and

[0026] A retraining unit that retrains the completed learning model using a data set including sensor data output during the return of the new substrate as teacher data.

[0027] Provide more of

[0028] The return anomaly prediction system according to the eighth embodiment is a return anomaly prediction system according to any of the first to seventh embodiments, and

[0029] The above data set further includes at least one time information among the device operation time during substrate return, the elapsed time after maintenance, and the component usage time of the substrate handling unit.

[0030] The return anomaly prediction system according to the ninth embodiment is a return anomaly prediction system according to any of the first to eighth embodiments, and

[0031] The above-mentioned completed learning model is machine-learned from a data set containing sensor data from past substrate returns, wherein the remaining time or the number of returns from the time of the substrate return until the occurrence of a return anomaly is associated with the teacher data. The above-mentioned estimation unit takes a data set containing sensor data from a new substrate return as input, estimates the degree of return anomaly based on the remaining time or the number of returns predicted by the above-mentioned completed learning model, and outputs the estimate of the return anomaly.

[0032] The return anomaly prediction system according to the 10th embodiment is a return anomaly prediction system according to any of the 1st to 8th embodiments, and

[0033] The above-mentioned completed model is machine-learned using the k-neighbor method with a dataset containing sensor data from past normal substrate transport as teacher data, and the above-mentioned estimation unit takes a dataset containing sensor data from a new substrate transport as input and estimates and outputs the transport anomaly based on the distance to k-neighbor calculated by the above-mentioned completed model.

[0034] The return anomaly prediction system according to the 11th embodiment is a return anomaly prediction system according to any of the 1st to 8th embodiments, and

[0035] The above-mentioned learning completed model is machine-learned using a Long Short-Term Memory (LSTM) with a dataset containing sensor data from a past normal substrate return as teacher data, and the above-mentioned estimation unit takes a dataset containing actual sensor data up to the time of new substrate return as input, calculates the discrepancy between the dataset containing sensor data at the time of new substrate return predicted by the above-mentioned learning completed model from the dataset containing actual sensor data up to the time immediately before the time of new substrate return, and the dataset containing actual sensor data at the time of new substrate return, and estimates and outputs the return anomaly based on the discrepancy.

[0036] The return anomaly prediction system according to the 12th embodiment is a return anomaly prediction system according to any of the 1st to 8th embodiments, and

[0037] The above-mentioned completed model is machine-learned from a dataset containing sensor data from past substrate returns, with teacher data labeled as to whether the substrate return time corresponds to a return anomaly. The above-mentioned estimation unit takes a dataset containing sensor data from a new substrate return as input, estimates the return anomaly level based on the probability of a return anomaly occurring time predicted by the above-mentioned completed model, and outputs the return anomaly level.

[0038] The return anomaly prediction system according to the 13th embodiment is a return anomaly prediction system according to the 12th embodiment, and

[0039] The above-mentioned completed model is machine-learned from teacher data in which a data set containing sensor data from past substrate returns is labeled as to whether the substrate return is a time when a return anomaly occurs, and if it is a time when a return anomaly occurs, the cause of the return anomaly is labeled. The above-mentioned estimation unit takes a data set containing sensor data from a new substrate return as input, and estimates and outputs the degree of return anomaly for each cause of return anomaly based on the probability of a time when a return anomaly occurs predicted by the above-mentioned completed model for each cause of return anomaly.

[0040] The return anomaly prediction system according to the 14th embodiment is a return anomaly prediction system according to any of the 1st to 13th embodiments, and

[0041] The above estimation unit has a plurality of trained models and estimates and outputs the return anomaly based on a combination of predictions by the plurality of trained models.

[0042] A substrate processing apparatus according to the 15th aspect of an embodiment is

[0043] A substrate carrier, and

[0044] Return anomaly prediction system regarding any of the first to fourth embodiments

[0045] Equipped with

[0046] The method for predicting return abnormalities according to the 16th aspect of the embodiment is,

[0047] It is a return anomaly prediction method executed by a computer, and

[0048] A step of estimating and outputting the transport anomaly during a new substrate transport by using a fully trained model that machine-learns the relationship between a data set containing sensor data output during a past substrate transport from each of a plurality of sensors provided in a substrate transport unit and the transport anomaly during said substrate transport, by taking a data set containing sensor data output during a new substrate transport from each of the plurality of sensors as input.

[0049] Includes

[0050] The return anomaly prediction program regarding the 17th aspect of the embodiment is

[0051] on the computer,

[0052] A step of estimating and outputting the transport anomaly during a new substrate transport by using a fully trained model that machine-learns the relationship between a data set containing sensor data output during a past substrate transport from each of a plurality of sensors provided in a substrate transport unit and the transport anomaly during said substrate transport, by taking a data set containing sensor data output during a new substrate transport from each of the plurality of sensors as input.

[0053] Execute.

[0054] A computer-readable recording medium relating to the 18th aspect of an embodiment is,

[0055] on the computer,

[0056] A step of estimating and outputting the transport anomaly during a new substrate transport by using a fully trained model that machine-learns the relationship between a data set containing sensor data output during a past substrate transport from each of a plurality of sensors provided in a substrate transport unit and the transport anomaly during said substrate transport, by taking a data set containing sensor data output during a new substrate transport from each of the plurality of sensors as input.

[0057] The return anomaly prediction program for executing is being recorded non-transitory.

[0058] The learning completion model regarding the 19th embodiment is a tuned neural network system, and

[0059] The relationship between the data set containing sensor data output during past substrate transport and the transport anomaly during the substrate transport is machine learned by inputting a data set containing sensor data output during past substrate transport and the data set containing sensor data output during past substrate transport, each of a plurality of sensors provided in a substrate transport unit, inputting the data set containing sensor data output during past substrate transport from each of a plurality of sensors provided in the substrate transport unit to the input layer, comparing the output result output from the output layer with the transport anomaly degree during the substrate transport and updating the parameters of each node according to the error, and repeating this process for a data set containing sensor data from multiple past substrate transports.

[0060] When a data set including sensor data output from each of the plurality of sensors during new substrate transport is input to the input layer, the computer is functioned to estimate the transport abnormality during said new substrate transport and output it from the output layer.

[0061] Hereinafter, specific embodiments of the embodiment will be described in detail with reference to the attached drawings. In addition, in the following description and the drawings used therein, parts that can be configured identically are denoted by the same reference numerals, and redundant descriptions are omitted.

[0062] FIG. 1 is a diagram illustrating the schematic configuration of a substrate processing device (1) according to one embodiment.

[0063] As illustrated in FIG. 1, the substrate processing device (1) has a top ring (6), a substrate transport unit (2), nozzles (3a, 3b), a transport abnormality prediction system (10), and an output device (4).

[0064] The top ring (6) is configured to have an airbag on its lower surface and to support the substrate (W) downward, and to move the substrate (W) between a substrate transport position (a position opposite to the substrate transport unit (2)) and a substrate processing position (e.g., a position opposite a polishing table not shown).

[0065] The nozzles (3a, 3b) are positioned on the side of the top ring (6) at the substrate transport position and are configured to spray air and shower between the airbag of the top ring (6) and the substrate (W) to detach the substrate (W) from the top ring (6).

[0066] The substrate transport unit (2) is positioned below the top ring (6) at the substrate transport position and is configured to receive the substrate (W) that is detached from and falls from the top ring (6). For example, a lifter or a linear transporter (LTP) is used as the substrate transport unit (2).

[0067] As illustrated in FIG. 1, the substrate transport unit (2) is provided with a plurality of sensors (51 to 53) (three in the illustrated example). In the illustrated example, the plurality of sensors (51 to 53) are all vibration sensors (acceleration sensors) that measure vibrations of the substrate transport unit (2), and are installed at locations where vibrations during the transport of the substrate (W) are directly transmitted, such as the arm portion or the shaft portion of the substrate transport unit (2).

[0068] In addition, the plurality of sensors (51 to 53) are not limited to vibration sensors, and may consist of one or more types among vibration sensors, sound sensors, image sensors, video sensors, temperature sensors, device movement speed sensors, device operation torque sensors, and device parallelism sensors.

[0069] FIG. 3 is a diagram illustrating an example of sensor data output from sensors (51 to 53) during normal substrate transport. In FIG. 3, the area enclosed by the solid rectangle represents sensor data from the start of transport to the end of transport during the transport of one substrate (1 cycle). As shown in FIG. 3, during normal transport, sensor data with the same waveform is repeatedly output from sensors (51 to 53) every cycle.

[0070] FIG. 4 is a drawing for explaining the substrate return unit (2) when a return abnormality occurs, and FIG. 5 is a drawing illustrating an example of sensor data when a return abnormality occurs.

[0071] As illustrated in FIG. 4, a typical sign of a conveying malfunction is that the substrate (W) does not detach from the top ring (6) symmetrically, and the substrate (W) falls in an obliquely tilted position. In this case, the impact when the substrate (W) is seated on the substrate conveying unit (2) is concentrated at one point on the substrate (W), making the substrate (W) prone to damage. If the symptoms worsen further, the substrate (W) falls off the substrate conveying unit (2). In the example illustrated in FIG. 4, two causes (types of malfunction) of the conveying malfunction are considered: (1) the position where the air and shower sprayed from the nozzles (3a, 3b) come into contact is misaligned, and (2) the axis of the top ring (6) and the axis of the substrate conveying unit (2) are misaligned.

[0072] As illustrated in FIG. 5, when a return abnormality occurs, sensor data with a waveform different from the normal waveform is output from the sensors (51 to 53). In FIG. 5, the area enclosed by the solid rectangle represents the sensor data from the start of return to the end of return when a return abnormality occurs.

[0073] The output device (4) is an interface that outputs various information to a user (e.g., an operator of the substrate processing device (1)), and uses, for example, an image display means (display) such as a liquid crystal display, a lamp, a speaker, etc.

[0074] As illustrated in FIG. 1, the return anomaly prediction system (10) is connected to each of a plurality of sensors (51 to 53) and an output device (4) so ​​as to be communicable.

[0075] Next, the configuration of the return anomaly prediction system (10) will be described. FIG. 2 is a block diagram illustrating the configuration of the return anomaly prediction system (10). At least part of the return anomaly prediction system (10) is configured by a single computer or quantum computing system, or a plurality of computers or quantum computing systems connected to each other via a network.

[0076] As illustrated in FIG. 2, the return anomaly prediction system (10) has an input unit (11), a control unit (12), a memory unit (13), and an output unit (14). Each unit (11 to 14) is connected to communicate via a bus or network.

[0077] Among these, the input unit (11) is a communication interface for a plurality of sensors (51 to 5N) provided in the substrate transport unit (2). The input unit (11) may be connected to each output terminal of the plurality of sensors (51 to 5N) via a wired connection or via a wireless connection.

[0078] The output unit (14) is a communication interface for the output device (4). The output unit (14) may be connected to the input terminal of the output device (4) via a wired connection or via a wireless connection.

[0079] The memory unit (13) is a non-volatile data storage, such as flash memory. Various data handled by the control unit (12) is stored in the memory unit (13). For example, the memory unit (13) stores a data set (151) referenced by the estimation unit (121) described later, and a threshold (152) referenced by the output signal transmission unit (122) described later.

[0080] The data set (151) includes sensor data obtained through the input unit (11) which is output from each of the plurality of sensors (51 to 5N) provided in the substrate transport unit (2) when a new substrate is transported. The data set (151) may further include at least one time information among the device operation time during the substrate transport, the elapsed time after maintenance, and the usage time of the part of the substrate handling unit (a part that comes into contact with the substrate (W) and gradually wears out).

[0081] As illustrated in FIG. 2, the control unit (12) has an estimation unit (121), an output signal transmission unit (122), and a relearning unit (123). Each of these units may be realized by a processor within the return anomaly prediction system (10) executing a predetermined program, or may be implemented in hardware.

[0082] The estimation unit (121) has a machine-learned model (120) (e.g., a tuned neural network system, see FIGS. 6 to 13) that machine-learns the relationship between a data set containing sensor data output during past substrate transport from each of the plurality of sensors (51 to 5N) provided in the substrate transport unit (2) and the transport abnormality during said substrate transport, and takes as input a data set containing sensor data output during new substrate transport from each of the plurality of sensors (51 to 5N) (i.e., a data set (131) stored in the memory unit (13)) to estimate and output the transport abnormality during said new substrate transport.

[0083] As a first example of the timing of processing in the estimation unit (121), the estimation unit (121) may process in real time, that is, by inputting a data set (131) containing sensor data from the start of transport during the new substrate transport to the present time, the estimation unit may estimate and output the transport abnormality during the new substrate transport.

[0084] As a second example of the timing of processing in the estimation unit (121), the estimation unit (121) may process for each substrate (W), that is, it may take a data set (131) containing sensor data from the start of transport to the end of transport during the transport of a new substrate as input, and estimate and output the transport abnormality during the transport of the new substrate.

[0085] As a third example of the timing of processing in the estimation unit (121), the estimation unit (121) may process for each lot of substrates (W) (e.g., 25 substrates per lot), that is, the estimation unit (121) may input a data set (131) containing sensor data from the start of the first substrate transport to the end of the last substrate transport during the transport of multiple substrates (1 lot) and estimate and output the transport abnormality during the transport of multiple substrates (1 lot).

[0086] The estimation unit (121) may use the sensor data output from each of the plurality of sensors (51 to 5N) as is as input, or may use the intensity of a predetermined frequency range extracted (preprocessed) by FFT (Fast Fourier Transform) as input.

[0087] FIG. 6 is a schematic diagram for explaining the configuration of a learning completion model (120) in a first embodiment. In the example illustrated in FIG. 6, the learning completion model (120) is a tuned neural network system and includes a hierarchical neural network or a quantum neural network (QNN) having an input layer, one or more intermediate layers connected to the input layer, and an output layer connected to the intermediate layer. In FIG. 6, a feedforward neural network is illustrated as the hierarchical neural network, but various types of neural networks such as a convolutional neural network (CNN) or a recurrent neural network (RNN) may be used. The learning completion model (120) may include a neural network with two or more intermediate layers, i.e., deep learning.

[0088] To explain an example of a method for generating a learning completion model (120) in the first embodiment, teacher data is prepared in which sensor data output from each of the plurality of sensors (51 to 5N) provided in the substrate transport unit (2) during a past substrate transport is labeled with whether the substrate transport time is a time when a transport abnormality occurs (e.g., 0 for normal substrate transport, 1 for transport abnormality). Then, as shown in FIG. 6, one sensor data of a substrate transport time included in the teacher data is input to an input layer, and the output result output from the output layer is compared with the label of whether the substrate transport time included in the teacher data is a time when a transport abnormality occurs, and the parameter of each node (weight, threshold, etc.) is updated according to the error. This process is repeated for each of the plurality of substrate transport time data included in the teacher data. Accordingly, a learning completed model (120) (tuned neural network system) is generated based on sensor data output from each of the plurality of sensors (51 to 5N) provided in the substrate transport unit (2) during past substrate transport, which predicts the probability that the substrate transport time is the time when a transport abnormality occurs (transport abnormality probability).

[0089] In the first embodiment, the estimation unit (121) takes sensor data output from each of the plurality of sensors (51 to 5N) at the time of new substrate transport as input, and estimates and outputs the degree of transport abnormality at the time of new substrate transport based on the probability of a time of transport abnormality (transport abnormality probability) predicted by the learning completed model (120). In addition, the degree of transport abnormality may be the probability of a time of transport abnormality (transport abnormality probability) predicted by the learning completed model (120) itself, or may be a value obtained by uniquely converting the said transport abnormality probability into a predetermined function.

[0090] FIG. 7 is a schematic diagram illustrating the configuration of a learning completion model (120) in a second embodiment. In the example illustrated in FIG. 7, the learning completion model (120) is a tuned neural network system and includes a hierarchical neural network or a quantum neural network (QNN) having an input layer, one or more intermediate layers connected to the input layer, and an output layer connected to the intermediate layer. In FIG. 7, a feedforward neural network is illustrated as the hierarchical neural network, but various types of neural networks, such as a convolutional neural network (CNN) or a recurrent neural network (RNN), may be used. The learning completion model (120) may include a neural network with two or more intermediate layers, i.e., deep learning.

[0091] To explain an example of a method for generating a learning completion model (120) in the second embodiment, a teacher data set is prepared that includes sensor data output from each of the plurality of sensors (51 to 5N) provided in the substrate transport unit (2) during a past substrate transport, and time information during the substrate transport (i.e., at least one of the device operation time, elapsed time after maintenance, and component usage time of the substrate handling unit). The teacher data is labeled as indicating whether the substrate transport time is a time when a transport abnormality occurs (e.g., 0 for normal substrate transport, 1 for transport abnormality). Then, as shown in FIG. 7, one data set of substrate transport time included in the teacher data is input to an input layer, and the output result output from the output layer is compared with the label indicating whether the substrate transport time is a time when a transport abnormality occurs included in the teacher data. The teacher data is then used to update the parameters (weights, thresholds, etc.) of each node according to the error. This is repeated for each of the data for the multiple substrate transports included. By doing so, a learning completed model (120) (tuned neural network system) is generated based on a data set containing sensor data output from each of the multiple sensors (51 to 5N) provided in the substrate transport unit (2) during a past substrate transport and time information during the substrate transport, which predicts the probability that the substrate transport time is the time when a transport abnormality occurs (transport abnormality probability).

[0092] In the second embodiment, the estimation unit (121) takes as input a data set including sensor data output from each of the plurality of sensors (51 to 5N) at the time of new substrate transport and time information at the time of new substrate transport, and estimates and outputs the degree of transport abnormality at the time of new substrate transport based on the probability of a time of transport abnormality (transport abnormality probability) predicted by the learning completed model (120). In addition, the degree of transport abnormality may be the probability of a time of transport abnormality (transport abnormality probability) predicted by the learning completed model (120) itself, or may be a value obtained by uniquely converting the transport abnormality probability into a predetermined function.

[0093] FIG. 8 is a schematic diagram for explaining the configuration of a learning completion model (120) in a third embodiment. In the example illustrated in FIG. 8, the learning completion model (120) is a tuned neural network system and includes a hierarchical neural network or a quantum neural network (QNN) having an input layer, one or more intermediate layers connected to the input layer, and an output layer connected to the intermediate layer. In FIG. 8, a feedforward neural network is illustrated as the hierarchical neural network, but various types of neural networks, such as a convolutional neural network (CNN) or a recurrent neural network (RNN), may be used. The learning completion model (120) may include a neural network with two or more intermediate layers, i.e., deep learning.

[0094] To explain an example of a method for generating a learning completion model (120) in the third embodiment, sensor data output from each of the plurality of sensors (51 to 5N) provided in the substrate transport unit (2) during a past substrate transport is labeled with whether the substrate transport time is a time when a transport abnormality occurs (e.g., 0 for normal substrate transport, 1 for transport abnormality), and teacher data is prepared in which the cause of the transport abnormality (type of abnormality) is labeled in the case of a transport abnormality. Then, as shown in FIG. 8, one sensor data of a substrate transport time included in the teacher data is input to an input layer, and the output result output from the output layer is compared with the label of whether the substrate transport time is a time when a transport abnormality occurs for each cause of the transport abnormality (type of abnormality) included in the teacher data, and the parameter (weight, threshold, etc.) of each node is updated according to the error. This process is repeated for each of the plurality of substrate transport time data included in the teacher data. Accordingly, a learning completed model (120) (tuned neural network system) is generated based on sensor data output from each of the plurality of sensors (51 to 5N) provided in the substrate transport unit (2) during past substrate transport, which predicts the probability (transport abnormality probability) that the substrate transport time is the time when a transport abnormality occurs for each cause (type of abnormality) of the transport abnormality.

[0095] In the third embodiment, the estimation unit (121) takes sensor data output from each of the plurality of sensors (51 to 5N) at the time of new substrate transport as input, and estimates and outputs the degree of transport abnormality at the time of new substrate transport for each cause of transport abnormality (type of abnormality) predicted by the learning completed model (120) based on the probability of the time of transport abnormality (transport abnormality probability). In addition, the degree of transport abnormality may be the probability of the time of transport abnormality (transport abnormality probability) predicted by the learning completed model (120) itself, or may be a value obtained by uniquely converting the said transport abnormality probability into a predetermined function.

[0096] FIG. 9 is a schematic diagram illustrating the configuration of a learning completion model (120) in a fourth embodiment. In the example illustrated in FIG. 9, the learning completion model (120) is a tuned neural network system and includes a hierarchical neural network or a quantum neural network (QNN) having an input layer, one or more intermediate layers connected to the input layer, and an output layer connected to the intermediate layer. In FIG. 9, a feedforward neural network is illustrated as the hierarchical neural network, but various types of neural networks, such as a convolutional neural network (CNN) or a recurrent neural network (RNN), may be used. The learning completion model (120) may include a neural network with two or more intermediate layers, i.e., deep learning.

[0097] To explain an example of a method for generating a learning completion model (120) in the fourth embodiment, a data set is prepared that includes sensor data output from each of the plurality of sensors (51 to 5N) provided in the substrate transport unit (2) during a past substrate transport and time information during the substrate transport (i.e., at least one of the device operation time, elapsed time after maintenance, and component usage time of the substrate handling unit). In this data set, whether the substrate transport time is a time when a transport abnormality occurs (e.g., 0 for normal substrate transport, 1 for transport abnormality) is labeled, and in the case of a transport abnormality, the cause of the transport abnormality (type of abnormality) is labeled. Then, as shown in FIG. 9, one data set of substrate transport time included in the training data is input to an input layer, and the output result output from the output layer and the substrate transport time for each cause of the transport abnormality (type of abnormality) included in the training data is a time when a transport abnormality occurs A process of comparing labels indicating whether it is a time or not, and updating parameters (weights or thresholds, etc.) of each node according to the error, is repeated for each of the data on multiple substrate transport times included in the teacher data. By doing so, a learning completed model (120) (tuned neural network system) is generated based on a data set containing sensor data output from each of the multiple sensors (51 to 5N) provided in the substrate transport unit (2) during a past substrate transport time and time information at the time of the said substrate transport time, which predicts the probability (transport abnormality probability) that the time of said substrate transport is the time when a transport abnormality occurs for each cause (type of abnormality) of the transport abnormality.

[0098] In the fourth embodiment, the estimation unit (121) takes as input a data set including sensor data output from each of the plurality of sensors (51 to 5N) at the time of new substrate transport and time information at the time of new substrate transport, and estimates and outputs the degree of transport abnormality at the time of new substrate transport for each cause of transport abnormality (type of abnormality) predicted by the learning completed model (120), based on the probability of the time of transport abnormality (transport abnormality probability) predicted by the learning completed model (120). In addition, the degree of transport abnormality may be the probability of the time of transport abnormality (transport abnormality probability) predicted by the learning completed model (120) itself, or may be a value obtained by uniquely converting the said transport abnormality probability into a predetermined function.

[0099] FIG. 10 is a schematic diagram illustrating the configuration of a learning completion model (120) in a fifth embodiment. In the example illustrated in FIG. 10, the learning completion model (120) is a tuned neural network system and includes a hierarchical neural network or a quantum neural network (QNN) having an input layer, one or more intermediate layers connected to the input layer, and an output layer connected to the intermediate layer. In FIG. 10, a feedforward neural network is illustrated as the hierarchical neural network, but various types of neural networks, such as a convolutional neural network (CNN) or a recurrent neural network (RNN), may be used. The learning completion model (120) may include a neural network with two or more intermediate layers, i.e., deep learning.

[0100] To explain an example of a method for generating a learning completion model (120) in the fifth embodiment, a teacher data is prepared in which the remaining time or the number of remaining returns from the time of return to the time of return abnormality is associated with the sensor data output from each of the plurality of sensors (51 to 5N) provided in the substrate return unit (2) during a past substrate return. Then, as shown in FIG. 10, the sensor data of one substrate return included in the teacher data is input to an input layer, and the output result output from the output layer is compared with the remaining time or the number of remaining returns from the time of return to the time of return abnormality included in the teacher data, and the parameter (weight or threshold, etc.) of each node is updated according to the error. This process is repeated for each of the plurality of substrate return data included in the teacher data. Accordingly, a learning completed model (120) (tuned neural network system) is generated to predict the remaining time or the number of remaining returns from the time of the said return until the return abnormality occurs, based on sensor data output from each of the plurality of sensors (51 to 5N) provided in the substrate return unit (2) during the past return of the substrate.

[0101] In the fifth embodiment, the estimation unit (121) takes sensor data output from each of the plurality of sensors (51 to 5N) at the time of new substrate transport as input, and estimates and outputs the transport abnormality at the time of new substrate transport based on the remaining time or remaining number of transports predicted by the learning completion model (120). Furthermore, the transport abnormality may be the remaining time or remaining number of transports predicted by the learning completion model (120) itself, or it may be a value obtained by uniquely converting the remaining time or remaining number of transports into a predetermined function. For example, the transport abnormality may be a value obtained by dividing the remaining time predicted by the learning completion model (120) by the average time from the time of maintenance until the time of transport abnormality, or a value obtained by dividing the remaining number of transports predicted by the learning completion model (120) by the average number of transports from the time of maintenance until the time of transport abnormality.

[0102] FIG. 11 is a schematic diagram for explaining the configuration of a learning completion model (120) in the sixth embodiment. In the example illustrated in FIG. 11, the learning completion model (120) is a tuned neural network system and includes a hierarchical neural network or a quantum neural network (QNN) having an input layer, one or more intermediate layers connected to the input layer, and an output layer connected to the intermediate layer. In FIG. 11, a feedforward neural network is illustrated as the hierarchical neural network, but various types of neural networks, such as a convolutional neural network (CNN) or a recurrent neural network (RNN), may be used. The learning completion model (120) may include a neural network with two or more intermediate layers, i.e., deep learning.

[0103] To explain an example of a method for generating a learning completion model (120) in the sixth embodiment, a teacher data is prepared in which the remaining time or the number of remaining returns from the time of the return anomaly occurs is associated with a data set containing sensor data output during a past return of a substrate from each of the plurality of sensors (51 to 5N) provided in the substrate return unit (2) and time information during the return of the substrate (i.e., at least one of the device operation time during the return of the substrate, the elapsed time after maintenance, and the usage time of the component of the substrate handling unit). Then, as shown in FIG. 10, a data set of one return of the substrate included in the teacher data is input to an input layer, and the output result output from the output layer is compared with the remaining time or the number of remaining returns from the time of the return anomaly occurs from the time of the return of the substrate included in the teacher data, and the parameters (weights or thresholds, etc.) of each node are updated according to the error. This is repeated for each of the data. Accordingly, a learning completed model (120) (tuned neural network system) is generated based on a data set containing sensor data output from each of the plurality of sensors (51 to 5N) provided in the substrate transport unit (2) during a past substrate transport and time information during the substrate transport. This model predicts the remaining time or the number of remaining transports from the time of substrate transport until the occurrence of a transport anomaly.

[0104] In the sixth embodiment, the estimation unit (121) takes as input a data set including sensor data output from each of the plurality of sensors (51 to 5N) at the time of new substrate transport and time information at the time of new substrate transport, and estimates and outputs the transport abnormality at the time of new substrate transport based on the remaining time or remaining number of transports predicted by the learning completion model (120). In addition, the transport abnormality may be the remaining time or remaining number of transports predicted by the learning completion model (120) itself, or may be a value obtained by uniquely converting the remaining time or remaining number of transports into a predetermined function. For example, the return anomaly value may be a value obtained by dividing the remaining time predicted by the learning completion model (120) by the average time from the time of maintenance until the return anomaly occurs, or a value obtained by dividing the remaining number of returns predicted by the learning completion model (120) by the average number of returns from the time of maintenance until the return anomaly occurs.

[0105] FIG. 12 is a schematic diagram for explaining the configuration of a learning completion model (120) in the seventh embodiment. In the example illustrated in FIG. 12, the learning completion model (120) is machine learned using a k-neighbor method with a data set containing sensor data from a past normal substrate transport as teacher data, and calculates the distance to k-neighbors of the data set based on a data set containing sensor data from a new substrate transport. In the example illustrated in FIG. 12, a plurality of white circles ("○") each represent a position in the feature space of the sensor data from a past normal substrate transport, a white triangle ("△") represents a position in the feature space of the sensor data from a new substrate transport, and a dashed line represents the distance to k-neighbors in the case of k=3.

[0106] In the seventh embodiment, the estimation unit (121) takes a data set containing sensor data at the time of new substrate transport as input and estimates and outputs a transport anomaly based on the distance to k calculated by the learning completion model (121). Additionally, the transport anomaly may be the distance to k calculated by the learning completion model (120) itself, or it may be a value obtained by uniquely converting the distance to k using a predetermined function.

[0107] FIG. 13 is a schematic diagram for explaining the configuration of a learning completion model (120) in the eighth embodiment. In the example illustrated in FIG. 13, the learning completion model (120) is a tuned neural network system and includes a hierarchical neural network or a quantum neural network (QNN) having an input layer, one or more intermediate layers connected to the input layer, and an output layer connected to the intermediate layer. The learning completion model (120) may include a neural network with two or more intermediate layers, i.e., deep learning.

[0108] The learning completion model (120) in the eighth embodiment is machine learned using LSTM (Long Short-Term Memory) with a data set containing sensor data from a past normal substrate transport as teacher data. In the example shown in FIG. 13, the learning completion model (120) predicts a data set containing sensor data for the new substrate transport (predicted value of sensor data for the nth cycle) based on a data set containing actual sensor data (sensor data of the first to n-1 cycles) from, for example, during maintenance until just before the new substrate transport.

[0109] In the eighth embodiment, the estimation unit (121) takes as input a data set containing actual sensor data (sensor data of the first to n cycles) from the time of maintenance until the time of new substrate transport, calculates a discrepancy between a data set containing sensor data at the time of new substrate transport (predicted value of sensor data of the nth cycle) predicted by the learning completion model (120) from a data set containing actual sensor data (sensor data of the first to n-1 cycles) up to the time immediately before the time of new substrate transport, and a data set containing actual sensor data at the time of new substrate transport (sensor data of the nth cycle), and estimates and outputs a transport anomaly based on the discrepancy. Additionally, the transport anomaly may be the discrepancy itself, or a value obtained by uniquely transforming the discrepancy using a predetermined function.

[0110] The estimation unit (121) may have a plurality of completed learning models (120) (for example, two or more of the learning models (120) in the first to eighth embodiments) and may estimate and output the return anomaly based on a combination of predictions by the plurality of completed learning models (120) (i.e., ensemble learning).

[0111] The output signal transmitting unit (122) compares the return abnormality output by the estimation unit (121) with a predetermined threshold (152), and if the return abnormality exceeds the threshold (152), transmits an output signal to the output device (4) through the output unit (14) to output a maintenance notification and / or alarm.

[0112] The retraining unit (123) retrains the learning completion model (120) using a data set (131) containing sensor data output when a new substrate is returned as teacher data.

[0113] Next, an example of a return anomaly prediction method by a return anomaly prediction system (10) configured in this way will be described. FIG. 14 is a flowchart illustrating an example of a return anomaly prediction method.

[0114] As shown in FIG. 14, the estimation unit (121) acquires a data set (131) containing sensor data output at the time of new substrate transport from each of the plurality of sensors (51 to 5N) provided in the substrate transport unit (2) (step S11). The data set (131) acquired by the estimation unit (121) may include at least one time information among the device operation time, the elapsed time after maintenance, and the component usage time of the substrate handling unit at the time of substrate transport 2. The data set (131) acquired by the estimation unit (121) is stored in the memory unit (13).

[0115] Then, the estimation unit (121) uses a machine-learned model (120) that machine-learns the relationship between a data set containing sensor data output from each of the plurality of sensors (51 to 5N) during a past substrate transport and the transport abnormality during said substrate transport, and takes a data set (131) containing sensor data output from each of the plurality of sensors (51 to 5N) during a new substrate transport as input, and estimates and outputs the transport abnormality during said new substrate transport (step S12).

[0116] In step S12, the estimation unit (121) may use a machine-learned model (120) (see FIG. 8 and FIG. 9) that has learned the relationship between the degree of return abnormality at the time of return abnormality for each cause of return abnormality (type of abnormality) and the sensor data output from each of the plurality of sensors (51 to 5N) at the time of return abnormality, and input a data set (131) that has received sensor data output from each of the plurality of sensors (51 to 5N) at the time of new return abnormality, and estimate and output the degree of return abnormality at the time of new return abnormality for each cause of return abnormality (type of abnormality).

[0117] Next, the output signal transmitting unit (122) compares the return abnormality output by the estimation unit (121) with a predetermined threshold (152) (step S13).

[0118] If the return abnormality output by the estimation unit (121) exceeds the threshold (152) (Step S13: "Yes"), the output signal transmission unit (122) transmits an output signal to the output device (4) through the output unit (14) to output a maintenance notice and / or alarm (Step S14).

[0119] Meanwhile, if the return abnormality output by the estimation unit (121) does not exceed the threshold (152) (Step S13: "No"), the retraining unit (123) retrains the learning completion model (120) by using the data set (131) containing the sensor data output during the return of the new substrate as the normal labeled teacher data (Step S15).

[0120] According to the present embodiment as described above, the estimation unit (121) can estimate and output the degree of transport abnormality during the new substrate transport by using a machine learning completed model (120) that machine learns the relationship between a data set containing sensor data during a past substrate transport and the degree of transport abnormality during the substrate transport. This is achieved by synthesizing from a plurality of indicator data for a data set (131) containing sensor data during the new substrate transport. By doing so, the probability of detecting a transport abnormality can be improved compared to a conventional embodiment in which a difference between the output of a sensor during substrate transport and the output of a sensor stored during position alignment before substrate transport exceeds a certain threshold. Furthermore, by using the learned completed model (120), it is possible to use vibration, sound, image data, etc. of a device, which were difficult to handle in the conventional embodiment, as sensor data.

[0121] In addition, according to the present embodiment, when the return abnormality estimated by the estimation unit (121) exceeds the threshold (152), the output signal transmission unit (122) transmits an output signal to the output device (4) to output a maintenance notification and / or alarm, so that the user (e.g., the operator of the substrate processing unit (1)) can perform maintenance by using the output from the output device (4) as a trigger, thereby preventing handling errors in advance.

[0122] In addition, according to the present embodiment, the relearning unit (123) relearns the learning completion model (120) using a data set (131) containing sensor data output when a new substrate is returned as teacher data, thereby obtaining a system that follows the change in the operating status of the device.

[0123] In addition, according to the embodiment shown in FIGS. 8 and 9, the estimation unit (121) takes a data set containing sensor data during a past substrate transport and a data set containing sensor data during a new substrate transport based on the relationship between the degree of transport abnormality during the substrate transport for each cause of the transport abnormality (type of abnormality) and the degree of transport abnormality during the new substrate transport, and estimates and outputs the degree of transport abnormality during the new substrate transport for each cause of the transport abnormality (type of abnormality), thereby allowing the maintenance time of the device to be shortened by investigating the cause of the transport abnormality (type of abnormality) estimated to have a high degree of transport abnormality.

[0124] Additionally, the return anomaly prediction system (10) according to the present embodiment may be configured by a single computer or quantum computing system, or by a plurality of computers or quantum computing systems connected to each other via a network, but a program for realizing the return anomaly prediction system (10) on one or a plurality of computers or quantum computing systems and a computer-readable recording medium that records said program non-transitory are also subject to protection of the present invention.

[0125] Although embodiments and variations have been described by way of example above, the scope of the present technology is not limited thereto, and modifications and variations are possible within the scope described in the claims according to the purpose. Furthermore, each embodiment and variation may be appropriately combined within a scope that does not contradict the processing content.

Claims

Claim 1 A transport anomaly prediction system characterized by having a machine-learned model that machine-learns the relationship between a data set containing sensor data output during a past transport of a substrate from each of a plurality of sensors provided in a substrate transport unit and the transport anomaly degree during said substrate transport, and having an estimation unit that takes a data set containing sensor data output during a new transport of a substrate from each of the plurality of sensors as input and estimates and outputs the transport anomaly degree during said new substrate transport, wherein the machine-learned model is machine-learned using a Long Short-Term Memory (LSTM) with a data set containing sensor data during a past normal transport of a substrate as teacher data, and the estimation unit takes a data set containing actual sensor data up to said new substrate transport as input, calculates the discrepancy between a data set containing sensor data at said new substrate transport predicted by the machine-learned model from a data set containing actual sensor data up to said new substrate transport and a data set containing actual sensor data at said new substrate transport, and estimates and outputs the transport anomaly degree based on said discrepancy. Claim 2 A transport anomaly prediction system characterized by having a machine-learned model that machine-learns the relationship between a data set containing sensor data output during a past transport of a substrate from each of a plurality of sensors provided in a substrate transport unit and the transport anomaly degree during said substrate transport, and having an estimation unit that takes a data set containing sensor data output during a new transport of a substrate from each of the plurality of sensors as input and estimates and outputs the transport anomaly degree during said new transport of a substrate, wherein the machine-learned model is a machine-learned model that machine-learns a data set containing sensor data during a past transport of a substrate and teacher data in which the said substrate transport time is labeled as a time when a transport anomaly occurs, and the estimation unit takes a data set containing sensor data during a new transport of a substrate as input and estimates and outputs the transport anomaly degree based on the probability of a time when a transport anomaly occurs predicted by the machine-learned model. Claim 3 A return anomaly prediction system according to paragraph 2, wherein the learning completion model is machine-learned from teacher data in which a data set containing sensor data from a past return of a substrate is labeled as to whether the return of the substrate is a time when a return anomaly occurs, and if a return anomaly occurs, the cause of the return anomaly is labeled; and the estimation unit takes a data set containing sensor data from a new return of a substrate as input and estimates and outputs the degree of return anomaly for each cause of return anomaly based on the probability of a return anomaly occurring time predicted by the learning completion model for each cause of return anomaly. Claim 4 A conveyance anomaly prediction system according to any one of claims 1 to 3, characterized in that the plurality of sensors comprises one or more types among a vibration sensor, a sound sensor, an image sensor, a video sensor, a temperature sensor, a device movement speed sensor, a device operation torque sensor, and a device parallelism sensor. Claim 5 A return abnormality prediction system characterized by further comprising, in any one of claims 1 to 3, an output signal transmitting unit that compares the return abnormality output by the estimation unit with a predetermined threshold, and if the return abnormality exceeds the threshold, transmits an output signal to an output device for outputting a maintenance notice and / or an alarm. Claim 6 A transport anomaly prediction system according to any one of claims 1 to 3, wherein the estimation unit takes as input a data set including sensor data from the start of transport to the present time during transport of a new substrate, and estimates and outputs the transport anomaly degree during transport of the new substrate. Claim 7 A transport anomaly prediction system according to any one of claims 1 to 3, wherein the estimation unit takes as input a data set including sensor data from the start of transport to the end of transport during transport of a new substrate, and estimates and outputs the transport anomaly degree during transport of the new substrate. Claim 8 A transport anomaly prediction system according to any one of claims 1 to 3, wherein the estimation unit takes as input a data set including sensor data from the start of transport of the first substrate to the end of transport of the last substrate during transport of a plurality of new substrates, and estimates and outputs the transport anomaly degree during transport of the plurality of new substrates. Claim 9 A return anomaly prediction system according to any one of claims 1 to 3, further comprising a retraining unit that retrains the completed learning model using a data set including sensor data output during the return of the new substrate as teacher data. Claim 10 A transport anomaly prediction system according to any one of claims 1 to 3, wherein the data set further includes at least one time information among the device operation time during transport of the substrate, the elapsed time after maintenance, and the component usage time of the substrate handling unit. Claim 11 A return anomaly prediction system according to any one of claims 1 to 3, wherein the estimation unit has a plurality of learned models and estimates and outputs the return anomaly based on a combination of predictions by the plurality of learned models. Claim 12 A substrate processing device characterized by having a substrate transport unit and a transport abnormality prediction system described in any one of claims 1 to 3. Claim 13 A method for predicting return anomalies executed by a computer, comprising a step of using a fully trained model that machine-learns the relationship between a data set containing sensor data output during a past return from each of a plurality of sensors provided in a substrate return section and the degree of return anomaly during the return, and taking a data set containing sensor data output during a new return from each of the plurality of sensors as input to estimate and output the degree of return anomaly during the return of the new substrate, wherein the fully trained model is machine-learned using a Long Short-Term Memory (LSTM) with a data set containing sensor data during a past normal return as teacher data, and taking a data set containing actual sensor data up to the return of the new substrate as input to calculate the discrepancy between a data set containing sensor data at the return of the new substrate predicted by the fully trained model from a data set containing actual sensor data up to the return of the new substrate and a data set containing actual sensor data at the return of the new substrate, and estimating and outputting the degree of return anomaly based on said discrepancy. method. Claim 14 A method for predicting return anomalies executed by a computer, wherein the method comprises a step of using a machine-learned model that machine-learns the relationship between a data set containing sensor data output from each of a plurality of sensors provided in a substrate transport unit during a past substrate transport and the degree of return anomaly during said substrate transport, and inputting a data set containing sensor data output from each of the plurality of sensors during a new substrate transport, and estimating and outputting the degree of return anomaly during said new substrate transport, wherein the machine-learned model is a data set containing sensor data during a past substrate transport and teacher data in which whether said substrate transport is a time when a return anomaly occurs is labeled, and inputting a data set containing sensor data during a new substrate transport, and estimating and outputting the degree of return anomaly based on the probability of a time when a return anomaly occurs predicted by the machine-learned model. Claim 15 A transport anomaly prediction program stored in a non-transient recording medium for executing a step of using a machine-learned model that machine-learns the relationship between a data set containing sensor data output during a past transport of a substrate from each of a plurality of sensors provided in a substrate transport section and the transport anomaly degree during said substrate transport, taking a data set containing sensor data output during a new substrate transport from each of the plurality of sensors as input, and estimating and outputting the transport anomaly degree during said new substrate transport, wherein the machine-learned model is machine-learned using a Long Short-Term Memory (LSTM) with a data set containing sensor data during a past normal substrate transport as teacher data, taking a data set containing actual sensor data up to said new substrate transport as input, calculating the discrepancy between a data set containing sensor data at said new substrate transport predicted by the machine-learned model from a data set containing actual sensor data up to said new substrate transport and a data set containing actual sensor data at said new substrate transport, and estimating and outputting the transport anomaly degree based on said discrepancy. Claim 16 A transport anomaly prediction program stored in a non-transient recording medium for executing a step of estimating and outputting the transport anomaly at a new substrate transport time by using a machine-learned model that machine-learns the relationship between a data set containing sensor data output at a past substrate transport time from each of a plurality of sensors provided in a substrate transport section and the transport anomaly at the time of said substrate transport, by inputting a data set containing sensor data output at a new substrate transport time from each of the plurality of sensors, wherein the machine-learned model is a data set containing sensor data at a past substrate transport time and teacher data in which whether the time of said substrate transport time is a time when a transport anomaly occurs is labeled, and the step of estimating and outputting the transport anomaly based on the probability of a time when a transport anomaly occurs predicted by the machine-learned model, by inputting a data set containing sensor data at the time of said substrate transport time. Claim 17 A computer-readable, non-transient device storing a transport anomaly prediction program for executing a step in which a transport anomaly is estimated and outputted by using a fully trained model that machine-learns the relationship between a data set containing sensor data output during a past substrate transport from each of a plurality of sensors provided in a substrate transport unit and the transport anomaly during said substrate transport, and inputs a data set containing sensor data output during a new substrate transport from each of the plurality of sensors, wherein the fully trained model is machine-trained using LSTM (Long Short-Term Memory) with a data set containing sensor data during a past normal substrate transport as teacher data, and inputs a data set containing actual sensor data up to said new substrate transport, calculates the discrepancy between a data set containing sensor data at said new substrate transport predicted by the fully trained model from a data set containing actual sensor data up to said new substrate transport and a data set containing actual sensor data at said new substrate transport, and estimates and outputs the transport anomaly based on said discrepancy. Recording media. Claim 18 A computer-readable, non-transient recording medium having a transport anomaly prediction program for executing a step of estimating and outputting the transport anomaly at a new substrate transport by using a machine-learned model that machine-learns the relationship between a data set containing sensor data output at a past substrate transport from each of a plurality of sensors provided in a substrate transport section and the transport anomaly at the time of said substrate transport, by inputting a data set containing sensor data output at a new substrate transport from each of the plurality of sensors, wherein the machine-learned model is a machine-learned model that machine-learns a data set containing sensor data at a past substrate transport and teacher data in which whether the time of said substrate transport is a time when a transport anomaly occurs is labeled, and the step of estimating and outputting the transport anomaly based on the probability of a time when a transport anomaly occurs predicted by the machine-learned model by inputting a data set containing sensor data at the time of said substrate transport. Claim 19 A program implementing a learned model used in an estimation unit that takes as input a data set containing actual sensor data output from each of a plurality of sensors provided in a substrate transport unit until the time of new substrate transport, calculates a discrepancy between a data set containing sensor data at the time of new substrate transport predicted by a learned model from a data set containing actual sensor data up to the time immediately before the time of new substrate transport and a data set containing actual sensor data at the time of new substrate transport, and estimates and outputs the transport anomaly degree at the time of new substrate transport based on the discrepancy, and has an input layer, one or more intermediate layers connected to the input layer, and an output layer connected to the intermediate layer, machine learns using a Long Short-Term Memory (LSTM) with a data set containing sensor data at the time of normal substrate transport in the past as teacher data, and when a data set containing actual sensor data up to the time immediately before the time of new substrate transport is input to the input layer, predicts sensor data at the time of new substrate transport and outputs it from the output layer. Claim 20 A program that implements a learning completion model used in an estimation unit, which takes as input a data set including sensor data output from each of a plurality of sensors provided in a substrate transport unit at the time of new substrate transport, and estimates and outputs the degree of transport abnormality at the time of new substrate transport based on the probability of a transport abnormality occurring at the time predicted by the learning completion model, and has an input layer, one or more intermediate layers connected to the input layer, and an output layer connected to the intermediate layer, and machine learns a training data in which whether the time of said substrate transport is a time of transport abnormality is labeled on a data set including sensor data at the time of past substrate transport, and when a data set including sensor data at the time of new substrate transport is input to the input layer, predicts the probability that the time of said new substrate transport is a time of transport abnormality and outputs from the output layer.

Images