Megasonic cleaning control system and method

Abstract

The invention provides a megasonic cleaning control system and a megasonic cleaning control method, which relate to the field of intelligent control and are characterized in that a time queue of real-time temperature of cleaning fluid is acquired through a temperature sensor, and a time queue of temperature gradient values is extracted from the time queue of the real-time temperature; the time sequence correlation analysis of the real-time temperature and the temperature gradient is carried out by adopting a data analysis and processing algorithm based on artificial intelligence; and intelligently recommending the power value of the input megasonic wave at the next time point according to the significant interactive fusion representation between the real-time temperature absolute quantity time sequence characteristic and the real-time temperature fluctuation quantity time sequence characteristic obtained by the real-time temperature and temperature gradient time sequence association. Therefore, the change trend of the temperature can be predicted in advance so as to control the temperature more finely by adjusting the megasonic power, so that the great fluctuation of the temperature is avoided, the dynamic and intelligent control of the temperature of the cleaning fluid is realized, and the defects caused by a traditional threshold triggering mode are avoided.

Description

Technical Field

[0001] The present application relates to the field of intelligent control, and more specifically, to a megasonic cleaning control system and method. Background Art

[0002] Megasonic cleaning technology has been widely used in semiconductor manufacturing, precision machining and other industries due to its high efficiency and environmental protection. Megasonic cleaning can effectively remove pollutants on the surface of objects through the tiny bubble bursting effect generated by high-frequency vibration. In the cleaning process, temperature control is crucial to ensure the cleaning effect. Too high temperature may cause the properties of the wafer material to change or even damage the wafer; too low temperature may reduce the cleaning efficiency.

[0003] Patent CN115739886A proposes a method for controlling the cleaning capacity of megasonic waves, which introduces bubbles of inert gas into the cleaning liquid during the wafer cleaning process, and inputs megasonic waves to refine the bubbles into fine bubbles with smaller diameters. At the same time, the temperature of the cleaning liquid is monitored in real time, and once the temperature is higher than the set value, the temperature is lowered.

[0004] In the above patent, the control method of the cleaning liquid temperature is mainly based on the threshold-triggered cooling process, that is, the cooling is performed only when the temperature exceeds the set value. However, this processing method will cause a response lag. In actual applications, the cleaning liquid temperature may rise rapidly in a short period of time, and the cooling measures take a certain amount of time to take effect, which may cause the cleaning liquid temperature to exceed the safety threshold in a short period of time, affecting the cleaning effect. In addition, since action is only taken when the temperature exceeds the threshold, this method may lead to frequent responses within the normal temperature range, and at the same time cause the temperature to fluctuate frequently around the set value. This fluctuation not only affects the consistency and stability of cleaning, but may also cause the chemical reaction rate in the cleaning liquid to be unstable, thereby affecting the cleaning effect and product quality.

[0005] Therefore, an optimized megasonic cleaning control scheme is desired. Summary of the invention

[0006] In view of the shortcomings of the prior art, the present application provides a megasonic cleaning control system and method.

[0007] According to one aspect of the present application, a megasonic cleaning control method is provided, which includes: immersing a wafer in a cleaning solution in a cleaning tank and cleaning it; the cleaning step includes: introducing bubbles of an inert gas into the cleaning solution, and refining the bubbles of the inert gas by inputting megasonic waves; and monitoring and controlling the temperature of the cleaning solution in real time; wherein, monitoring and controlling the temperature of the cleaning solution in real time includes: A time queue for obtaining the real-time temperature of the cleaning fluid collected by the temperature sensor; After calculating the temperature gradient value of the real-time temperature at each predetermined time point in the time queue of the real-time temperature to obtain the time queue of the temperature gradient value, the time queue of the real-time temperature and the time queue of the temperature gradient value are sequence-encoded to obtain a real-time temperature absolute quantity time series associated feature vector and a real-time temperature fluctuation quantity time series associated feature vector; The real-time temperature absolute quantity time series correlation feature vector and the real-time temperature fluctuation quantity time series correlation feature vector are subjected to significant interaction fusion between temperature features to obtain a temperature absolute quantity-fluctuation quantity time series significant interaction fusion representation, including: extracting shared implicit features from the real-time temperature absolute quantity time series correlation feature vector and the real-time temperature fluctuation quantity time series correlation feature vector to obtain a real-time temperature condition feature; taking the real-time temperature condition feature as condition information, performing feature fusion on the real-time temperature absolute quantity time series correlation feature vector and the real-time temperature fluctuation quantity time series correlation feature vector to obtain the temperature absolute quantity-fluctuation quantity time series significant interaction fusion representation; Based on the significant interaction fusion representation of the temperature absolute quantity and the temperature fluctuation quantity, a planning instruction is obtained, and the planning instruction includes a decoded value of the recommended input megasonic power at the next time point.

[0008] In the above-mentioned megasonic cleaning control method, after calculating the temperature gradient value of the real-time temperature at each predetermined time point in the time queue of the real-time temperature to obtain the time queue of the temperature gradient value, the time queue of the real-time temperature and the time queue of the temperature gradient value are respectively sequence-encoded to obtain the real-time temperature absolute quantity time series associated feature vector and the real-time temperature fluctuation quantity time series associated feature vector, including: calculating the temperature gradient value of the real-time temperature at each predetermined time point in the time queue of the real-time temperature to obtain the time queue of the temperature gradient value; inputting the time queue of the real-time temperature and the time queue of the temperature gradient value into a sequence encoder based on a forward LSTM model respectively to obtain the real-time temperature absolute quantity time series associated feature vector and the real-time temperature fluctuation quantity time series associated feature vector.

[0009] In the above-mentioned megasonic cleaning control method, shared implicit feature extraction is performed on the real-time temperature absolute quantity time series correlation feature vector and the real-time temperature fluctuation quantity time series correlation feature vector to obtain the real-time temperature condition feature, including: inputting the real-time temperature absolute quantity time series correlation feature vector and the real-time temperature fluctuation quantity time series correlation feature vector into a joint implicit feature capture network to obtain a real-time temperature absolute quantity-fluctuation quantity joint implicit feature vector; and performing feature activation based on a Sigmoid function on the real-time temperature absolute quantity-fluctuation quantity joint implicit feature vector to obtain a real-time temperature condition feature vector as the real-time temperature condition feature.

[0010] In the above-mentioned megasonic cleaning control method, the real-time temperature absolute quantity time series associated feature vector and the real-time temperature fluctuation quantity time series associated feature vector are input into a joint implicit feature capture network to obtain a real-time temperature absolute quantity-fluctuation quantity joint implicit feature vector, including: adding the real-time temperature absolute quantity time series associated feature vector and the real-time temperature fluctuation quantity time series associated feature vector by position point, multiplying the obtained real-time temperature absolute quantity-fluctuation quantity sum vector with a weight matrix and then adding it by position with a bias vector to obtain a real-time temperature absolute quantity-fluctuation quantity joint interaction vector; using a tanh function to process the real-time temperature absolute quantity-fluctuation quantity joint interaction vector to obtain the real-time temperature absolute quantity-fluctuation quantity joint implicit feature vector.

[0011] In the above-mentioned megasonic cleaning control method, the real-time temperature condition feature is used as condition information, and the real-time temperature absolute quantity time series associated feature vector and the real-time temperature fluctuation quantity time series associated feature vector are feature fused to obtain the temperature absolute quantity-fluctuation quantity time series significant interaction fusion representation, including: calculating the real-time temperature absolute quantity semantic contribution of the real-time temperature absolute quantity time series associated feature vector relative to the real-time temperature condition feature vector; calculating the real-time temperature fluctuation quantity semantic contribution of the real-time temperature fluctuation quantity time series associated feature vector relative to the real-time temperature condition feature vector; normalizing the real-time temperature absolute quantity semantic contribution and the real-time temperature fluctuation quantity semantic contribution, and using the normalized real-time temperature absolute quantity semantic contribution and the normalized real-time temperature fluctuation quantity semantic contribution to the real-time temperature condition feature vector. The real-time temperature absolute quantity time series association feature vector and the real-time temperature fluctuation quantity time series association feature vector are weighted modulated to obtain the modulated real-time temperature absolute quantity time series association feature vector and the modulated real-time temperature fluctuation quantity time series association feature vector; the modulated real-time temperature absolute quantity time series association feature vector is used as the query vector, the modulated real-time temperature fluctuation quantity time series association feature vector is used as the key vector and the real-time temperature condition feature vector is used as the value vector, the modulated real-time temperature absolute quantity time series association feature vector, the modulated real-time temperature fluctuation quantity time series association feature vector and the real-time temperature condition feature vector are input into the feature-to-feature significant guided interaction module based on the converter structure to obtain the temperature absolute quantity-fluctuation quantity time series significant interaction fusion representation vector as the temperature absolute quantity-fluctuation quantity time series significant interaction fusion representation.

[0012] In the above-mentioned megasonic cleaning control method, the real-time temperature absolute quantity semantic contribution of the real-time temperature absolute quantity time series associated feature vector relative to the real-time temperature condition feature vector is calculated, including: calculating the positional division of the real-time temperature absolute quantity time series associated feature vector and the real-time temperature condition feature vector to obtain the real-time temperature absolute quantity semantic contribution vector; calculating the base-two logarithmic function value of the absolute value of each eigenvalue of the real-time temperature absolute quantity semantic contribution vector to obtain the real-time temperature absolute quantity semantic contribution logarithmic vector; calculating the positional dot product of the real-time temperature absolute quantity time series associated feature vector and the real-time temperature absolute quantity semantic contribution logarithmic vector, and performing positional multiplication on the obtained dot product vector to obtain the real-time temperature absolute quantity semantic contribution value; calculating an exponential function with the natural constant e as the base and the real-time temperature absolute quantity semantic contribution value as the exponent to obtain the real-time temperature absolute quantity semantic contribution degree.

[0013] In the above-mentioned megasonic cleaning control method, the real-time temperature absolute quantity semantic contribution and the real-time temperature fluctuation quantity semantic contribution are normalized, and the normalized real-time temperature absolute quantity semantic contribution and the normalized real-time temperature fluctuation quantity semantic contribution are used to weighted modulate the real-time temperature absolute quantity time series association feature vector and the real-time temperature fluctuation quantity time series association feature vector to obtain the modulated real-time temperature absolute quantity time series association feature vector and the modulated real-time temperature fluctuation quantity time series association feature vector, including: calculating the sum of the real-time temperature absolute quantity semantic contribution and the real-time temperature fluctuation quantity semantic contribution to obtain the total sum of real-time temperature semantic contributions; dividing The real-time temperature absolute quantity semantic contribution and the real-time temperature fluctuation quantity semantic contribution are divided by the total sum of the real-time temperature semantic contributions to obtain the normalized real-time temperature absolute quantity semantic contribution and the normalized real-time temperature fluctuation quantity semantic contribution; the real-time temperature absolute quantity time series association feature vector is multiplied by the normalized real-time temperature absolute quantity semantic contribution at a position point to obtain the modulated real-time temperature absolute quantity time series association feature vector; the real-time temperature fluctuation quantity time series association feature vector is multiplied by the normalized real-time temperature fluctuation quantity semantic contribution at a position point to obtain the modulated real-time temperature fluctuation quantity time series association feature vector.

[0014] In the above-mentioned megasonic cleaning control method, the real-time temperature absolute quantity time series associated feature vector after modulation is used as a query vector, the real-time temperature fluctuation quantity time series associated feature vector after modulation is used as a key vector and the real-time temperature condition feature vector is used as a value vector, the real-time temperature absolute quantity time series associated feature vector after modulation, the real-time temperature fluctuation quantity time series associated feature vector after modulation and the real-time temperature condition feature vector are input into a feature-to-feature significant guided interaction module based on a converter structure to obtain a temperature absolute quantity-fluctuation quantity time series significant interaction fusion representation vector, including: the real-time temperature absolute quantity time series after modulation is used as a query vector, the real-time temperature fluctuation quantity time series associated feature vector after modulation is used as a key vector and the real-time temperature condition feature vector is used as a value vector. The associated feature vector is vector-multiplied with the transposed vector of the associated feature vector of the real-time temperature fluctuation quantity time series after modulation, and the obtained real-time temperature association matrix after modulation is divided by the square root of the length of the real-time temperature fluctuation quantity time series associated feature vector by position to obtain the real-time temperature association scaling matrix after modulation; the real-time temperature association scaling matrix after modulation is activated using a softmax function, and the obtained real-time temperature association scaling activation matrix after modulation is matrix-vector multiplied with the real-time temperature condition feature vector to obtain the temperature absolute quantity-fluctuation quantity time series significant interaction fusion representation vector.

[0015] In the above-mentioned megasonic cleaning control method, a planning instruction is obtained based on the temperature absolute quantity-fluctuation quantity time series significant interaction fusion representation, including: inputting the temperature absolute quantity-fluctuation quantity time series significant interaction fusion representation vector into a decoder-based power dynamic planner to obtain a planning instruction, and the planning instruction includes a decoded value of the recommended input megasonic power at the next time point.

[0016] According to another aspect of the present application, a megasonic cleaning control system is provided, comprising: A cleaning fluid temperature data acquisition module is used to obtain a time queue of the real-time temperature of the cleaning fluid collected by the temperature sensor; A real-time temperature absolute quantity fluctuation quantity time series feature encoding module is used to calculate the temperature gradient value of the real-time temperature at each predetermined time point in the time queue of the real-time temperature to obtain the time queue of the temperature gradient value, and then perform sequence encoding on the time queue of the real-time temperature and the time queue of the temperature gradient value to obtain a real-time temperature absolute quantity time series associated feature vector and a real-time temperature fluctuation quantity time series associated feature vector; A significant interactive fusion module for real-time temperature absolute quantity fluctuation quantity time series features is used to perform significant interactive fusion between temperature features on the real-time temperature absolute quantity time series associated feature vector and the real-time temperature fluctuation quantity time series associated feature vector to obtain a significant interactive fusion representation of temperature absolute quantity-fluctuation quantity time series, wherein the real-time temperature absolute quantity fluctuation quantity time series features significant interactive fusion module includes: a real-time temperature condition feature generation unit, used to perform shared implicit feature extraction on the real-time temperature absolute quantity time series associated feature vector and the real-time temperature fluctuation quantity time series associated feature vector to obtain a real-time temperature condition feature; a temperature absolute quantity-fluctuation quantity time series significant interactive fusion representation generation unit, used to perform feature fusion on the real-time temperature absolute quantity time series associated feature vector and the real-time temperature fluctuation quantity time series associated feature vector using the real-time temperature condition feature as condition information to obtain the temperature absolute quantity-fluctuation quantity time series significant interactive fusion representation; The input megasonic power decoding module is used to obtain a planning instruction based on the significant interaction fusion representation of the temperature absolute value-fluctuation value time series, and the planning instruction includes a decoded value of the recommended input megasonic power at the next time point.

[0017] This application has significant technical effects due to the adoption of the above technical solutions: The megasonic cleaning control system and method provided by the present application collects the time queue of the real-time temperature of the cleaning liquid through the temperature sensor, extracts the time queue of the temperature gradient value from the time queue of the real-time temperature, and uses the data analysis and processing algorithm based on artificial intelligence to perform the time series correlation analysis of the real-time temperature and the temperature gradient, so as to intelligently recommend the power value of the input megasonic wave at the next time point based on the significant interactive fusion expression between the time series characteristics of the real-time temperature absolute quantity and the time series characteristics of the real-time temperature fluctuation quantity obtained by the time series correlation of the real-time temperature and the temperature gradient. In this way, the temperature change trend can be predicted in advance to control the temperature more finely by adjusting the megasonic wave power, thereby avoiding large temperature fluctuations, realizing dynamic and intelligent control of the cleaning liquid temperature, and avoiding the defects caused by the traditional threshold triggering method. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] By describing the embodiments of the present application in more detail in conjunction with the accompanying drawings, the above and other purposes, features and advantages of the present application will become more apparent. The accompanying drawings are used to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the accompanying drawings, the same reference numerals generally represent the same components or steps.

[0019] Figure 1 The present invention is a flowchart of real-time monitoring and controlling of the temperature of the cleaning liquid in the megasonic cleaning control method according to an embodiment of the present application.

[0020] Figure 2 It is a data flow diagram for real-time monitoring and controlling the temperature of the cleaning liquid in the megasonic cleaning control method according to an embodiment of the present application.

[0021] Figure 3 In the megasonic cleaning control method according to an embodiment of the present application, the temperature features are significantly interactively fused between the real-time temperature absolute quantity time series correlation feature vector and the real-time temperature fluctuation quantity time series correlation feature vector to obtain a flow chart representing the significant interactive fusion of the temperature absolute quantity-fluctuation quantity time series.

[0022] Figure 4 The present invention is a flow chart of extracting shared implicit features of the real-time temperature absolute quantity time series correlation feature vector and the real-time temperature fluctuation quantity time series correlation feature vector in the megasonic cleaning control method according to an embodiment of the present application to obtain real-time temperature condition features.

[0023] Figure 5 In the megasonic cleaning control method according to an embodiment of the present application, the real-time temperature condition characteristics are used as conditional information, and the real-time temperature absolute quantity time series associated feature vector and the real-time temperature fluctuation quantity time series associated feature vector are feature fused to obtain a flowchart representing the significant interactive fusion of the temperature absolute quantity-fluctuation quantity time series.

[0024] Figure 6 4 is a system block diagram of a megasonic cleaning control system according to an embodiment of the present application. DETAILED DESCRIPTION

[0025] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described here.

[0026] Megasonic cleaning technology is widely used in fields such as semiconductor manufacturing and precision machining due to its high efficiency and environmental protection characteristics. This technology uses the tiny bubble bursting effect generated by high-frequency sound waves to remove pollutants from the surface of objects. Patent CN115739886A proposes a method for controlling the cleaning capacity of megasonic waves, which injects inert gas bubbles into the cleaning liquid and uses megasonic waves to refine the bubbles, while monitoring the temperature of the cleaning liquid. Once it exceeds the set value, it will cool down. However, this threshold-based temperature control method may cause response lag, affecting the cleaning effect and product quality. In addition, it only takes action when the temperature exceeds the threshold, which may cause frequent temperature fluctuations and affect the consistency and stability of cleaning.

[0027] Based on this, the present application proposes a megasonic cleaning control method, comprising: immersing a wafer in a cleaning liquid in a cleaning tank and cleaning it; the cleaning step comprises: introducing bubbles of an inert gas into the cleaning liquid, and refining the bubbles of the inert gas by inputting megasonic waves; and monitoring and controlling the temperature of the cleaning liquid in real time.

[0028] It should be noted that the megasonic wave generator generates high-frequency sound waves that propagate in the liquid, forming compression and rarefaction areas of the sound waves. When the power of the megasonic wave generator increases, the sound energy input into the cleaning liquid increases, and this sound energy is converted into heat energy, causing the temperature of the cleaning liquid to rise. At the same time, the increase in power leads to the generation of higher intensity sound waves, which will cause more frequent and more violent vibrations of bubbles, increasing the frequency of bubble formation and collapse. The rapid formation and collapse of such bubbles will release a large amount of energy, further heating the cleaning liquid. When the power of the megasonic wave decreases, the sound energy transmitted into the cleaning liquid decreases, resulting in a decrease in the rate of heat energy generation in the cleaning liquid, thereby reducing the temperature of the cleaning liquid, and the power reduction will also lead to a decrease in the frequency of bubble formation and collapse, which means that the released energy is reduced, thereby reducing the heating of the liquid and achieving a temperature reduction.

[0029] Accordingly, in real-time monitoring of the temperature of the cleaning liquid and adjusting the temperature of the cleaning liquid in real time by controlling the megasonic wave, the technical concept of the present application is to collect the time queue of the real-time temperature of the cleaning liquid by the temperature sensor, extract the time queue of the temperature gradient value from the time queue of the real-time temperature, and use the data analysis and processing algorithm based on artificial intelligence to perform the time series correlation analysis of the real-time temperature and the temperature gradient, so as to intelligently recommend the power value of the input megasonic wave at the next time point according to the significant interactive fusion representation between the real-time temperature absolute quantity time series characteristics and the real-time temperature fluctuation quantity time series characteristics obtained by the time series correlation of the real-time temperature and the temperature gradient. In this way, the temperature change trend can be predicted in advance to more finely control the temperature by adjusting the megasonic wave power, thereby avoiding large temperature fluctuations, realizing dynamic and intelligent control of the cleaning liquid temperature, and avoiding the defects brought by the traditional threshold triggering method.

[0030] Figure 1 The present invention is a flowchart of real-time monitoring and controlling of the temperature of the cleaning liquid in the megasonic cleaning control method according to an embodiment of the present application. Figure 2 FIG. 1 is a schematic diagram of data flow for real-time monitoring and control of the temperature of the cleaning liquid in the megasonic cleaning control method according to an embodiment of the present application. Figure 1 and Figure 2As shown, the temperature of the cleaning liquid is monitored and controlled in real time, including: S110, obtaining a time queue of the real-time temperature of the cleaning liquid collected by a temperature sensor; S120, calculating the temperature gradient value of the real-time temperature at each predetermined time point in the time queue of the real-time temperature to obtain a time queue of the temperature gradient value, and then performing sequence encoding on the time queue of the real-time temperature and the time queue of the temperature gradient value to obtain a real-time temperature absolute quantity time series associated feature vector and a real-time temperature fluctuation quantity time series associated feature vector; S130, performing significant interaction fusion between temperature features on the real-time temperature absolute quantity time series associated feature vector and the real-time temperature fluctuation quantity time series associated feature vector to obtain a temperature absolute quantity-fluctuation quantity time series significant interaction fusion representation; S140, obtaining a planning instruction based on the temperature absolute quantity-fluctuation quantity time series significant interaction fusion representation, the planning instruction including a recommended decoded value of the input megasonic power at the next time point.

[0031] In step S110, the time queue of the real-time temperature of the cleaning liquid collected by the temperature sensor is obtained. It should be understood that the time queue data of the real-time temperature of the cleaning liquid records the dynamic process of the temperature change of the cleaning liquid. By analyzing the real-time temperature time queue, the rising, falling or stable trend information of the temperature can be obtained, which helps the system to deeply understand the temperature change pattern during the cleaning process, and then adjust the input power of the megasonic wave during the cleaning process based on the change of the cleaning liquid temperature, so as to realize the dynamic and intelligent control of the cleaning liquid temperature.

[0032] In step S120, after calculating the temperature gradient value of the real-time temperature at each predetermined time point in the time queue of the real-time temperature to obtain the time queue of the temperature gradient value, the time queue of the real-time temperature and the time queue of the temperature gradient value are respectively sequence-encoded to obtain the real-time temperature absolute quantity time series associated feature vector and the real-time temperature fluctuation quantity time series associated feature vector. Specifically, in an embodiment of the present application, after calculating the temperature gradient value of the real-time temperature at each predetermined time point in the time queue of the real-time temperature to obtain the time queue of the temperature gradient value, the time queue of the real-time temperature and the time queue of the temperature gradient value are respectively sequence-encoded to obtain the real-time temperature absolute quantity time series associated feature vector and the real-time temperature fluctuation quantity time series associated feature vector, including: calculating the temperature gradient value of the real-time temperature at each predetermined time point in the time queue of the real-time temperature to obtain the time queue of the temperature gradient value; inputting the time queue of the real-time temperature and the time queue of the temperature gradient value into a sequence encoder based on a forward LSTM model to obtain the real-time temperature absolute quantity time series associated feature vector and the real-time temperature fluctuation quantity time series associated feature vector.

[0033] It should be understood that in the process of cleaning, the speed information of temperature change over time is also crucial for understanding and controlling the temperature in the cleaning process. Based on this, in order to more accurately identify whether the temperature rises rapidly or slowly, or tends to be stable in a time series dynamic change, so as to prevent the adverse effects of cleaning caused by sudden temperature changes, in the technical solution of the present application, the temperature gradient value of the real-time temperature at each predetermined time point in the time queue of the real-time temperature is calculated to obtain the time queue of the temperature gradient value. That is, the temperature gradient reflects the rate of change of temperature over time. By calculating the temperature gradient, the temperature of the cleaning liquid can be more accurately controlled to ensure that it remains within an optimal operating range. Accordingly, considering that there are different correlation effects between different time scales in the time queue of the real-time temperature and the time queue of the temperature gradient value, and the information at the current time point is affected by the previous time point. Therefore, in the technical solution of the present application, the time queue of the real-time temperature and the time queue of the temperature gradient value are respectively input into the sequence encoder based on the forward LSTM model to capture and mine the temperature correlation relationship between different time spans, and obtain the real-time temperature absolute quantity time series correlation feature vector and the real-time temperature fluctuation quantity time series correlation feature vector.

[0034] In step S130, the real-time temperature absolute quantity time series correlation feature vector and the real-time temperature fluctuation quantity time series correlation feature vector are subjected to significant interaction fusion between temperature features to obtain a significant interaction fusion representation of the temperature absolute quantity-fluctuation quantity time series. Specifically, Figure 3 The present invention is a flowchart of performing significant interaction fusion between temperature features on the real-time temperature absolute quantity time series correlation feature vector and the real-time temperature fluctuation quantity time series correlation feature vector in the megasonic cleaning control method according to the embodiment of the present application to obtain a significant interaction fusion representation of the temperature absolute quantity-fluctuation quantity time series. Figure 3 As shown, the temperature features of the real-time temperature absolute quantity time series correlation feature vector and the real-time temperature fluctuation quantity time series correlation feature vector are significantly interactively fused to obtain a temperature absolute quantity-fluctuation quantity time series significant interactive fusion representation, including: S131, extracting shared implicit features of the real-time temperature absolute quantity time series correlation feature vector and the real-time temperature fluctuation quantity time series correlation feature vector to obtain real-time temperature condition features; S132, taking the real-time temperature condition features as conditional information, feature fusion of the real-time temperature absolute quantity time series correlation feature vector and the real-time temperature fluctuation quantity time series correlation feature vector is performed to obtain the temperature absolute quantity-fluctuation quantity time series significant interactive fusion representation.

[0035] Accordingly, considering that the real-time temperature absolute quantity time series correlation feature vector reflects the actual value of temperature changing over time and its changing trend, it is possible to capture the pattern of temperature changing over time. The real-time temperature fluctuation quantity time series correlation feature vector reflects the speed and direction of temperature change and expresses the rate of change of temperature over time. These two features respectively reflect two aspects of temperature information, but at the same time, the two pieces of information have a mutual correlation and interaction relationship in time series. Therefore, in order to comprehensively analyze the interaction between the two feature vectors to obtain a deeper understanding and more effective feature representation, in the technical solution of the present application, the real-time temperature absolute quantity time series correlation feature vector and the real-time temperature fluctuation quantity time series correlation feature vector are subjected to significant interaction fusion between temperature features to obtain a temperature absolute quantity-fluctuation quantity time series significant interaction fusion representation vector as a temperature absolute quantity-fluctuation quantity time series significant interaction fusion representation.

[0036] Specifically, first, the shared implicit features of the real-time temperature absolute quantity time series correlation feature vector and the real-time temperature fluctuation quantity time series correlation feature vector are extracted to obtain the real-time temperature condition feature vector as the real-time temperature condition feature. That is, by integrating the common information of the two, the dynamic characteristics of temperature can be more efficiently represented. Specifically, by inputting the real-time temperature absolute quantity time series correlation feature vector and the real-time temperature fluctuation quantity time series correlation feature vector into the joint implicit feature capture network to capture and refine the shared implicit features and inherent relationships in the original features of the two, and at the same time enhancing the implicit expression ability in the original feature vector to obtain the real-time temperature absolute quantity-fluctuation quantity joint implicit feature vector. Then, by using a Sigmoid function to activate the joint implicit feature to highlight and characterize the most critical and important parts of the joint feature, suppress and reduce the influence of irrelevant details, and obtain the real-time temperature condition feature vector.

[0037] Then, taking the real-time temperature condition feature as condition information, the real-time temperature absolute quantity time series associated feature vector and the real-time temperature fluctuation quantity time series associated feature vector are feature fused to obtain the temperature absolute quantity-fluctuation quantity time series significant interaction fusion representation. Specifically, the semantic contribution of the real-time temperature absolute quantity time series associated feature vector and the real-time temperature fluctuation quantity time series associated feature vector relative to the real-time temperature condition feature vector are calculated respectively to obtain the real-time temperature absolute quantity semantic contribution and the real-time temperature fluctuation quantity semantic contribution. That is, by calculating the semantic contribution, the importance of the real-time temperature absolute quantity feature and the real-time temperature fluctuation quantity feature to the real-time temperature condition feature vector can be evaluated, which helps to determine which features are more critical to temperature control. Subsequently, in order to enable the two contributions to be compared and combined on the same scale, the real-time temperature absolute quantity semantic contribution and the real-time temperature fluctuation quantity semantic contribution are normalized to obtain the normalized real-time temperature absolute quantity semantic contribution and the normalized real-time temperature fluctuation quantity semantic contribution, which helps to reasonably weight each feature according to its importance in the subsequent feature fusion process. Furthermore, the original features are weighted modulated based on the normalized semantic contribution to enhance the influence of important features and weaken the influence of minor features, ensuring that those features that have a greater impact on the conditional feature vector occupy a more important position in the comprehensive representation, and the modulated real-time temperature absolute quantity time series associated feature vector and the modulated real-time temperature fluctuation quantity time series associated feature vector are obtained. Finally, in order to further capture and strengthen the relationship between different features and generate a more comprehensive representation with significant interactive features, the modulated real-time temperature absolute quantity time series associated feature vector is used as the query vector, the modulated real-time temperature fluctuation quantity time series associated feature vector is used as the key vector, and the real-time temperature conditional feature vector is used as the value vector, which is input into the feature-to-feature significant guided interaction module based on the transformer structure for processing. In this way, the complex interactive relationship between the absolute temperature and the temperature fluctuation quantity can be captured from different perspectives, and the multi-head attention mechanism of the Transformer can be used to highlight the most relevant feature combinations at different time points, thereby strengthening the significant features and making the final more representative temperature absolute quantity-fluctuation quantity time series significant interaction fusion representation vector.

[0038] More specifically, Figure 4 The present invention is a flow chart of extracting shared implicit features from the real-time temperature absolute quantity time series correlation feature vector and the real-time temperature fluctuation quantity time series correlation feature vector to obtain real-time temperature condition features in the megasonic cleaning control method according to an embodiment of the present application. Figure 4As shown, shared implicit feature extraction is performed on the real-time temperature absolute quantity time series correlation feature vector and the real-time temperature fluctuation quantity time series correlation feature vector to obtain the real-time temperature condition feature, including: S1311, inputting the real-time temperature absolute quantity time series correlation feature vector and the real-time temperature fluctuation quantity time series correlation feature vector into a joint implicit feature capture network to obtain a real-time temperature absolute quantity-fluctuation quantity joint implicit feature vector; S1312, performing feature activation based on a Sigmoid function on the real-time temperature absolute quantity-fluctuation quantity joint implicit feature vector to obtain a real-time temperature condition feature vector as the real-time temperature condition feature.

[0039] More specifically, in an embodiment of the present application, the real-time temperature absolute quantity time-series associated feature vector and the real-time temperature fluctuation quantity time-series associated feature vector are input into a joint implicit feature capture network to obtain a real-time temperature absolute quantity-fluctuation quantity joint implicit feature vector, including: adding the real-time temperature absolute quantity time-series associated feature vector and the real-time temperature fluctuation quantity time-series associated feature vector by position point, multiplying the obtained real-time temperature absolute quantity-fluctuation quantity sum vector with a weight matrix, and then adding it by position with a bias vector to obtain a real-time temperature absolute quantity-fluctuation quantity joint interaction vector; using a tanh function to process the real-time temperature absolute quantity-fluctuation quantity joint interaction vector to obtain the real-time temperature absolute quantity-fluctuation quantity joint implicit feature vector.

[0040] More specifically, Figure 5 In the megasonic cleaning control method according to the embodiment of the present application, the real-time temperature condition feature is used as condition information to perform feature fusion on the real-time temperature absolute quantity time series associated feature vector and the real-time temperature fluctuation quantity time series associated feature vector to obtain a flow chart of the significant interaction fusion representation of the temperature absolute quantity-fluctuation quantity time series. Figure 5As shown, taking the real-time temperature condition feature as condition information, the real-time temperature absolute quantity time series associated feature vector and the real-time temperature fluctuation quantity time series associated feature vector are feature fused to obtain the temperature absolute quantity-fluctuation quantity time series significant interaction fusion representation, including: S1321, calculating the real-time temperature absolute quantity semantic contribution of the real-time temperature absolute quantity time series associated feature vector relative to the real-time temperature condition feature vector; S1322, calculating the real-time temperature fluctuation quantity semantic contribution of the real-time temperature fluctuation quantity time series associated feature vector relative to the real-time temperature condition feature vector; S1323, normalizing the real-time temperature absolute quantity semantic contribution and the real-time temperature fluctuation quantity semantic contribution, and using the normalized real-time temperature absolute quantity semantic contribution and the normalized real-time temperature fluctuation quantity semantic contribution to the real-time temperature condition feature vector. The real-time temperature absolute quantity time series association feature vector and the real-time temperature fluctuation quantity time series association feature vector are weighted modulated to obtain the modulated real-time temperature absolute quantity time series association feature vector and the modulated real-time temperature fluctuation quantity time series association feature vector; S1324, using the modulated real-time temperature absolute quantity time series association feature vector as the query vector, the modulated real-time temperature fluctuation quantity time series association feature vector as the key vector and the real-time temperature condition feature vector as the value vector, the modulated real-time temperature absolute quantity time series association feature vector, the modulated real-time temperature fluctuation quantity time series association feature vector and the real-time temperature condition feature vector are input into the feature-to-feature significant guided interaction module based on the converter structure to obtain the temperature absolute quantity-fluctuation quantity time series significant interaction fusion representation vector as the temperature absolute quantity-fluctuation quantity time series significant interaction fusion representation.

[0041] More specifically, in an embodiment of the present application, the real-time temperature absolute quantity semantic contribution of the real-time temperature absolute quantity time-series associated feature vector relative to the real-time temperature condition feature vector is calculated, including: calculating the positional division of the real-time temperature absolute quantity time-series associated feature vector and the real-time temperature condition feature vector to obtain the real-time temperature absolute quantity semantic contribution vector; calculating the base-two logarithmic function value of the absolute value of each eigenvalue of the real-time temperature absolute quantity semantic contribution vector to obtain the real-time temperature absolute quantity semantic contribution logarithmic vector; calculating the positional dot product of the real-time temperature absolute quantity time-series associated feature vector and the real-time temperature absolute quantity semantic contribution logarithmic vector, and performing positional multiplication on the obtained dot product vector to obtain the real-time temperature absolute quantity semantic contribution value; calculating an exponential function with the natural constant e as the base and the real-time temperature absolute quantity semantic contribution value as the exponent to obtain the real-time temperature absolute quantity semantic contribution degree.

[0042] More specifically, in an embodiment of the present application, the real-time temperature absolute quantity semantic contribution and the real-time temperature fluctuation quantity semantic contribution are normalized, and the normalized real-time temperature absolute quantity semantic contribution and the normalized real-time temperature fluctuation quantity semantic contribution are used to weighted modulate the real-time temperature absolute quantity time series association feature vector and the real-time temperature fluctuation quantity time series association feature vector to obtain the modulated real-time temperature absolute quantity time series association feature vector and the modulated real-time temperature fluctuation quantity time series association feature vector, including: calculating the sum of the real-time temperature absolute quantity semantic contribution and the real-time temperature fluctuation quantity semantic contribution to obtain the total sum of real-time temperature semantic contributions; dividing The real-time temperature absolute quantity semantic contribution and the real-time temperature fluctuation quantity semantic contribution are divided by the total sum of the real-time temperature semantic contributions to obtain the normalized real-time temperature absolute quantity semantic contribution and the normalized real-time temperature fluctuation quantity semantic contribution; the real-time temperature absolute quantity time series association feature vector is multiplied by the normalized real-time temperature absolute quantity semantic contribution at a position point to obtain the modulated real-time temperature absolute quantity time series association feature vector; the real-time temperature fluctuation quantity time series association feature vector is multiplied by the normalized real-time temperature fluctuation quantity semantic contribution at a position point to obtain the modulated real-time temperature fluctuation quantity time series association feature vector.

[0043] More specifically, in an embodiment of the present application, the post-modulation real-time temperature absolute quantity time series associated feature vector is used as a query vector, the post-modulation real-time temperature fluctuation quantity time series associated feature vector is used as a key vector, and the real-time temperature condition feature vector is used as a value vector. The post-modulation real-time temperature absolute quantity time series associated feature vector, the post-modulation real-time temperature fluctuation quantity time series associated feature vector, and the real-time temperature condition feature vector are input into a feature-to-feature significant guided interaction module based on a converter structure to obtain a temperature absolute quantity-fluctuation quantity time series significant interaction fusion representation vector, including: the post-modulation real-time temperature absolute quantity time series associated feature vector is used as a query vector, the post-modulation real-time temperature fluctuation quantity time series associated feature vector is used as a key vector, and the real-time temperature condition feature vector is used as a value vector. The associated feature vector is vector-multiplied with the transposed vector of the associated feature vector of the real-time temperature fluctuation quantity time series after modulation, and the obtained real-time temperature association matrix after modulation is divided by the square root of the length of the real-time temperature fluctuation quantity time series associated feature vector by position to obtain the real-time temperature association scaling matrix after modulation; the real-time temperature association scaling matrix after modulation is activated using a softmax function, and the obtained real-time temperature association scaling activation matrix after modulation is matrix-vector multiplied with the real-time temperature condition feature vector to obtain the temperature absolute quantity-fluctuation quantity time series significant interaction fusion representation vector.

[0044] In the embodiment of the present application, specifically, the temperature features are significantly interactively fused for the real-time temperature absolute value time series correlation feature vector and the real-time temperature fluctuation value time series correlation feature vector, which can be expressed by the formula: in, and are respectively the time series correlation feature vector of the real-time temperature absolute quantity and the time series correlation feature vector of the real-time temperature fluctuation quantity, To add by location, and are the weight matrix and bias vector respectively, for h function, is the real-time temperature absolute value-fluctuation value joint implicit eigenvector, for function, represents the real-time temperature condition feature vector, is the characteristic value of each position in the time series associated characteristic vector of the real-time temperature absolute quantity, is the characteristic value of each position in the time series associated characteristic vector of the real-time temperature fluctuation quantity, is the eigenvalue of each position in the real-time temperature condition eigenvector, is the number of eigenvalues ​​in the time series associated eigenvector of the real-time temperature absolute quantity and the time series associated eigenvector of the real-time temperature fluctuation quantity, represents the logarithmic function value with base 2, represents an exponential function with the natural constant e as the base, is the semantic contribution of the real-time temperature absolute value, The semantic contribution of real-time temperature fluctuation. and They are respectively the normalized semantic contribution of the real-time temperature absolute value and the normalized semantic contribution of the real-time temperature fluctuation value. and They are the time series correlation feature vector of the real-time temperature absolute quantity after modulation and the time series correlation feature vector of the real-time temperature fluctuation quantity after modulation, yes The transposed vector of yes Length, represents vector multiplication, yes function, It is the representation vector of the significant interaction fusion of the temperature absolute quantity and the temperature fluctuation quantity time series.

[0045] In step S140, based on the significant interaction fusion representation of the absolute temperature quantity-fluctuation quantity time series, a planning instruction is obtained, and the planning instruction includes the decoded value of the recommended input megasonic power at the next time point. Specifically, in an embodiment of the present application, based on the significant interaction fusion representation of the absolute temperature quantity-fluctuation quantity time series, a planning instruction is obtained, including: inputting the absolute temperature quantity-fluctuation quantity time series significant interaction fusion representation vector into a power dynamic planner based on a decoder to obtain a planning instruction, and the planning instruction includes the decoded value of the recommended input megasonic power at the next time point. That is, the temperature absolute quantity-fluctuation quantity time series significant interaction fusion representation obtained by significantly interacting and fusion of the real-time absolute temperature quantity time series associated feature vector and the real-time temperature fluctuation quantity time series associated feature vector is decoded and processed, so as to intelligently obtain the recommended input megasonic power value at the next time point. In this way, the temperature change trend can be predicted in advance to more finely control the temperature by adjusting the megasonic power, thereby avoiding large temperature fluctuations, realizing dynamic and intelligent control of the cleaning liquid temperature, and avoiding the defects caused by the traditional threshold triggering method.

[0046] In particular, when the real-time temperature absolute quantity time-series correlation feature vector and the real-time temperature fluctuation quantity time-series correlation feature vector respectively represent the time-series correlation features of the real-time temperature and the temperature gradient value, when performing significant interaction fusion between the temperature features, the difference in the distribution of implicit time-series pattern features will lead to a difference in feature significance based on the joint condition, causing the temperature absolute quantity-fluctuation quantity time-series significant interaction fusion representation vector obtained by interaction to have a complex interaction aggregation space structure. Therefore, when the temperature absolute quantity-fluctuation quantity time-series significant interaction fusion representation vector is input into a decoder-based power dynamic planner to obtain a planning instruction, it is expected to improve the decoding regression convergence and generalization effect of the temperature absolute quantity-fluctuation quantity time-series significant interaction fusion representation vector under the complex interaction aggregation space structure.

[0047] Therefore, the present application considers optimizing the temperature absolute quantity-fluctuation time series significant interaction fusion representation vector when the temperature absolute quantity-fluctuation time series significant interaction fusion representation vector is input into the decoder-based power dynamic planner, and the optimization process includes: calculating the absolute value sum of each eigenvalue of the temperature absolute quantity-fluctuation time series significant interaction fusion representation vector to obtain a first temperature absolute quantity-fluctuation time series significant interaction fusion spatial structure value, and calculating the square root of the sum of squares of each eigenvalue of the temperature absolute quantity-fluctuation time series significant interaction fusion representation vector to obtain a second temperature absolute quantity-fluctuation time series significant interaction fusion spatial structure value: wherein, the process can be expressed by a formula as follows: , , in, is the significant interaction fusion representation vector of the temperature absolute value-fluctuation value time series, express The number of eigenvalues ​​in , Indicates that the temperature absolute value-fluctuation value time series significant interaction fusion representation vector The eigenvalues ​​at the positions, is the spatial structure value of the significant interaction fusion of the first temperature absolute value-fluctuation value time series, The spatial structure value of the significant interaction fusion of the second temperature absolute value-fluctuation value time series; Each eigenvalue of the temperature absolute quantity-fluctuation quantity time series significant interaction fusion representation vector is multiplied by the first temperature absolute quantity-fluctuation quantity time series significant interaction fusion space structure value and the second temperature absolute quantity-fluctuation quantity time series significant interaction fusion space structure value to obtain the first temperature absolute quantity-fluctuation quantity time series significant interaction fusion structure reference value and the second temperature absolute quantity-fluctuation quantity time series significant interaction fusion structure reference value corresponding to each eigenvalue; each eigenvalue of the temperature absolute quantity-fluctuation quantity time series significant interaction fusion representation vector is multiplied by the length of the temperature absolute quantity-fluctuation quantity time series significant interaction fusion representation vector and the square root of the length to obtain the first temperature absolute quantity-fluctuation quantity time series significant interaction fusion scale transformation value and the second temperature absolute quantity-fluctuation quantity time series significant interaction fusion scale transformation value corresponding to each eigenvalue; the first temperature absolute quantity-fluctuation quantity time series significant interaction The fusion structure reference value is divided by the difference between the first temperature absolute quantity-fluctuation quantity time series significant interaction fusion space structure value and the first temperature absolute quantity-fluctuation quantity time series significant interaction fusion scale transformation value to obtain the first temperature absolute quantity-fluctuation quantity time series significant interaction fusion transformation adjustment value; the second temperature absolute quantity-fluctuation quantity time series significant interaction fusion structure reference value is divided by the difference between the second temperature absolute quantity-fluctuation quantity time series significant interaction fusion space structure value and the second temperature absolute quantity-fluctuation quantity time series significant interaction fusion scale transformation value to obtain the second temperature absolute quantity-fluctuation quantity time series significant interaction fusion transformation adjustment value; the weighted sum of the first temperature absolute quantity-fluctuation quantity time series significant interaction fusion transformation adjustment value and the second temperature absolute quantity-fluctuation quantity time series significant interaction fusion transformation adjustment value is calculated to obtain each eigenvalue of the optimized temperature absolute quantity-fluctuation quantity time series significant interaction fusion representation vector.

[0048] Here, the optimization expression of the temperature absolute value-fluctuation value time series significant interaction fusion representation vector is: in, Indicates the number of eigenvalues ​​in the temperature absolute value-fluctuation time series significant interaction fusion representation vector, Indicates that the temperature absolute value-fluctuation value time series significant interaction fusion representation vector The eigenvalues ​​at the positions, is the spatial structure value of the significant interaction fusion of the first temperature absolute value-fluctuation value time series, The spatial structure value of the second temperature absolute value-fluctuation value time series significant interaction fusion, Indicates the first temperature absolute value-fluctuation value time series significant interaction fusion transformation adjustment vector The absolute value of the first temperature at each position - the fluctuation value time series significantly interacts and fuses the transformation adjustment value, Indicates the second temperature absolute value-fluctuation value time series significant interaction fusion transformation adjustment vector The absolute value of the second temperature at each position - the fluctuation value time series significantly interacts and fuses the transformation adjustment value, is the first temperature absolute quantity-fluctuation quantity time series significant interaction fusion transformation adjustment vector, is the second temperature absolute quantity-fluctuation quantity time series significant interactive fusion transformation adjustment vector, It means adding by position. represents the weighted hyperparameter, It means point multiplication by position. A representation vector representing the optimized temperature absolute quantity-fluctuation quantity time series significant interaction fusion is shown.

[0049] That is, for the spatial structure information of the feature set of the temperature absolute quantity-fluctuation time series significant interaction fusion representation vector in the high-dimensional space, the scale-based frame transformation of each eigenvalue of the temperature absolute quantity-fluctuation time series significant interaction fusion representation vector is performed by taking the class norm space structured representation of the temperature absolute quantity-fluctuation time series significant interaction fusion representation vector as the reference window, and the spatial structure-based frame attention weight adjustment of each eigenvalue of the temperature absolute quantity-fluctuation time series significant interaction fusion representation vector is realized to ensure the spatial transformation invariance of the temperature absolute quantity-fluctuation time series significant interaction fusion representation vector under the feature space interaction, thereby improving the convergence and generalization effect of the decoding regression of the feature set of the temperature absolute quantity-fluctuation time series significant interaction fusion representation vector under the complex spatial structure representation, and improving the accuracy of the planning instruction obtained by inputting it into the decoder-based power dynamic planner. In this way, the temperature change trend can be predicted in advance to control the temperature more finely by adjusting the megasonic power, thereby avoiding large temperature fluctuations, realizing dynamic and intelligent control of the cleaning liquid temperature, and avoiding the defects brought by the traditional threshold triggering method.

[0050] In summary, the megasonic cleaning control method based on the embodiment of the present application is explained, by collecting the time queue of the real-time temperature of the cleaning liquid by the temperature sensor, and extracting the time queue of the temperature gradient value from the time queue of the real-time temperature, and using the data analysis and processing algorithm based on artificial intelligence to perform the time series correlation analysis of the real-time temperature and the temperature gradient, so as to intelligently recommend the input megasonic power value at the next time point according to the significant interactive fusion representation between the real-time temperature absolute quantity time series characteristics and the real-time temperature fluctuation quantity time series characteristics obtained by the time series correlation of the real-time temperature and the temperature gradient. In this way, the temperature change trend can be predicted in advance to more finely control the temperature by adjusting the megasonic power, thereby avoiding large temperature fluctuations, realizing dynamic and intelligent control of the cleaning liquid temperature, and avoiding the defects brought by the traditional threshold triggering method.

[0051] Figure 6 FIG. 1 is a system block diagram of a megasonic cleaning control system according to an embodiment of the present application. Figure 6 As shown, according to the embodiment of the present application, the megasonic cleaning control system 100 includes: a cleaning liquid temperature data acquisition module 110, which is used to obtain a time queue of the real-time temperature of the cleaning liquid collected by the temperature sensor; a real-time temperature absolute quantity fluctuation quantity time series feature encoding module 120, which is used to calculate the temperature gradient value of the real-time temperature at each predetermined time point in the time queue of the real-time temperature to obtain the time queue of the temperature gradient value, and then perform sequence encoding on the time queue of the real-time temperature and the time queue of the temperature gradient value to obtain a real-time temperature absolute quantity time series associated feature vector and a real-time temperature fluctuation quantity time series associated feature vector; a real-time temperature absolute quantity fluctuation quantity time series feature significant interaction fusion module 130, which is used to perform significant interaction fusion between temperature features on the real-time temperature absolute quantity time series associated feature vector and the real-time temperature fluctuation quantity time series associated feature vector to obtain a temperature absolute quantity-fluctuation quantity time series associated feature vector. Significant interaction fusion representation, wherein the real-time temperature absolute quantity fluctuation quantity time series feature significant interaction fusion module includes: a real-time temperature condition feature generation unit, used for extracting shared implicit features of the real-time temperature absolute quantity time series associated feature vector and the real-time temperature fluctuation quantity time series associated feature vector to obtain real-time temperature condition features; a temperature absolute quantity-fluctuation quantity time series significant interaction fusion representation generation unit, used for taking the real-time temperature condition features as conditional information, performing feature fusion on the real-time temperature absolute quantity time series associated feature vector and the real-time temperature fluctuation quantity time series associated feature vector to obtain the temperature absolute quantity-fluctuation quantity time series significant interaction fusion representation; an input megasonic power decoding module 140, used for obtaining a planning instruction based on the temperature absolute quantity-fluctuation quantity time series significant interaction fusion representation, the planning instruction including a recommended decoded value of the input megasonic power at the next time point.

[0052] Here, those skilled in the art will appreciate that the specific functions and operations of the various units and modules in the above-mentioned megasonic cleaning control system 100 have been described in detail above. Figures 1 to 5 The description of the megasonic cleaning control method has been described in detail, and therefore, its repeated description will be omitted.

[0053] In summary, the megasonic cleaning control system 100 based on the embodiment of the present application is explained, by collecting the time queue of the real-time temperature of the cleaning liquid by the temperature sensor, and extracting the time queue of the temperature gradient value from the time queue of the real-time temperature, and using the data analysis and processing algorithm based on artificial intelligence to perform the time series correlation analysis of the real-time temperature and the temperature gradient, so as to intelligently recommend the input megasonic power value at the next time point according to the significant interactive fusion representation between the real-time temperature absolute quantity time series characteristics and the real-time temperature fluctuation quantity time series characteristics obtained by the time series correlation of the real-time temperature and the temperature gradient. In this way, the temperature change trend can be predicted in advance to more finely control the temperature by adjusting the megasonic power, thereby avoiding large temperature fluctuations, realizing dynamic and intelligent control of the cleaning liquid temperature, and avoiding the defects brought by the traditional threshold triggering method.

Claims

1. A megasonic cleaning control method, comprising: Immersing the wafer in a cleaning solution in a cleaning tank and cleaning it; The cleaning step comprises: introducing bubbles of inert gas into the cleaning liquid, and refining the bubbles of inert gas by inputting megasonic waves; and monitoring and controlling the temperature of the cleaning liquid in real time; characterized in that the temperature of the cleaning liquid is monitored and controlled in real time, including: A time queue for obtaining the real-time temperature of the cleaning fluid collected by the temperature sensor; After calculating the temperature gradient value of the real-time temperature at each predetermined time point in the time queue of the real-time temperature to obtain the time queue of the temperature gradient value, the time queue of the real-time temperature and the time queue of the temperature gradient value are sequence-encoded to obtain a real-time temperature absolute quantity time series associated feature vector and a real-time temperature fluctuation quantity time series associated feature vector; The real-time temperature absolute quantity time series correlation feature vector and the real-time temperature fluctuation quantity time series correlation feature vector are subjected to significant interaction fusion between temperature features to obtain a temperature absolute quantity-fluctuation quantity time series significant interaction fusion representation, including: extracting shared implicit features from the real-time temperature absolute quantity time series correlation feature vector and the real-time temperature fluctuation quantity time series correlation feature vector to obtain a real-time temperature condition feature; taking the real-time temperature condition feature as condition information, performing feature fusion on the real-time temperature absolute quantity time series correlation feature vector and the real-time temperature fluctuation quantity time series correlation feature vector to obtain the temperature absolute quantity-fluctuation quantity time series significant interaction fusion representation; Based on the significant interaction fusion representation of the temperature absolute quantity and the temperature fluctuation quantity, a planning instruction is obtained, and the planning instruction includes a decoded value of the recommended input megasonic power at the next time point.

2. The megasonic cleaning control method according to claim 1, characterized in that: After calculating the temperature gradient value of the real-time temperature at each predetermined time point in the time queue of the real-time temperature to obtain the time queue of the temperature gradient value, the time queue of the real-time temperature and the time queue of the temperature gradient value are sequence-encoded respectively to obtain a real-time temperature absolute quantity time series associated feature vector and a real-time temperature fluctuation quantity time series associated feature vector, including: Calculate the temperature gradient value of the real-time temperature at each predetermined time point in the time queue of the real-time temperature to obtain the time queue of the temperature gradient value; The time queue of the real-time temperature and the time queue of the temperature gradient value are respectively input into a sequence encoder based on a forward LSTM model to obtain a time series correlation feature vector of the real-time temperature absolute quantity and a time series correlation feature vector of the real-time temperature fluctuation quantity.

3. The megasonic cleaning control method according to claim 2, characterized in that: The shared implicit feature extraction is performed on the real-time temperature absolute quantity time series associated feature vector and the real-time temperature fluctuation quantity time series associated feature vector to obtain the real-time temperature condition feature, including: Inputting the real-time temperature absolute quantity time series associated feature vector and the real-time temperature fluctuation quantity time series associated feature vector into a joint implicit feature capture network to obtain a real-time temperature absolute quantity-fluctuation quantity joint implicit feature vector; The real-time temperature absolute value-fluctuation value joint implicit feature vector is subjected to feature activation based on a Sigmoid function to obtain a real-time temperature condition feature vector as the real-time temperature condition feature.

4. The megasonic cleaning control method according to claim 3, characterized in that: Inputting the real-time temperature absolute quantity time series associated feature vector and the real-time temperature fluctuation quantity time series associated feature vector into a joint implicit feature capture network to obtain a real-time temperature absolute quantity-fluctuation quantity joint implicit feature vector, including: After adding the time series correlation feature vector of the real-time temperature absolute quantity and the time series correlation feature vector of the real-time temperature fluctuation quantity by position point, the obtained real-time temperature absolute quantity-fluctuation quantity sum vector is multiplied by the weight matrix and then added by position with the bias vector to obtain the real-time temperature absolute quantity-fluctuation quantity joint interaction vector; The real-time temperature absolute quantity-fluctuation quantity joint interaction vector is processed using a tanh function to obtain the real-time temperature absolute quantity-fluctuation quantity joint implicit feature vector.

5. The megasonic cleaning control method according to claim 4, characterized in that: Taking the real-time temperature condition feature as condition information, feature fusion is performed on the real-time temperature absolute quantity time series associated feature vector and the real-time temperature fluctuation quantity time series associated feature vector to obtain the temperature absolute quantity-fluctuation quantity time series significant interaction fusion representation, including: Calculating the real-time temperature absolute quantity semantic contribution of the real-time temperature absolute quantity time series association feature vector relative to the real-time temperature condition feature vector; Calculating the real-time temperature fluctuation semantic contribution of the real-time temperature fluctuation time series associated feature vector relative to the real-time temperature condition feature vector; Normalizing the semantic contribution of the real-time absolute temperature quantity and the semantic contribution of the real-time temperature fluctuation quantity, and using the normalized semantic contribution of the real-time absolute temperature quantity and the normalized semantic contribution of the real-time temperature fluctuation quantity to weighted modulate the real-time absolute temperature quantity time series association feature vector and the real-time temperature fluctuation quantity time series association feature vector to obtain a modulated real-time absolute temperature quantity time series association feature vector and a modulated real-time temperature fluctuation quantity time series association feature vector; Taking the time series associated feature vector of real-time temperature absolute quantity after modulation as the query vector, the time series associated feature vector of real-time temperature fluctuation quantity after modulation as the key vector and the real-time temperature condition feature vector as the value vector, the time series associated feature vector of real-time temperature absolute quantity after modulation, the time series associated feature vector of real-time temperature fluctuation quantity after modulation and the real-time temperature condition feature vector are input into a feature-to-feature significant guided interaction module based on a converter structure to obtain a temperature absolute quantity-fluctuation quantity time series significant interaction fusion representation vector as the temperature absolute quantity-fluctuation quantity time series significant interaction fusion representation.

6. The megasonic cleaning control method according to claim 5, characterized in that: Calculating the semantic contribution of the real-time temperature absolute quantity time series associated feature vector to the real-time temperature condition feature vector includes: Calculating the positional division of the real-time temperature absolute quantity time series association feature vector and the real-time temperature condition feature vector to obtain a real-time temperature absolute quantity semantic contribution vector; Calculating a logarithmic function value with base 2 of the absolute value of each eigenvalue of the real-time temperature absolute quantity semantic contribution vector to obtain a real-time temperature absolute quantity semantic contribution logarithmic vector; Calculate the positional dot product of the time series correlation feature vector of the real-time absolute temperature quantity and the logarithmic vector of the semantic contribution of the real-time absolute temperature quantity, and perform positional multiplication on the obtained dot product vector to obtain the semantic contribution value of the real-time absolute temperature quantity; An exponential function with the natural constant e as the base and the semantic contribution value of the real-time absolute temperature as the exponent is calculated to obtain the semantic contribution degree of the real-time absolute temperature.

7. The megasonic cleaning control method according to claim 6, characterized in that: The real-time temperature absolute quantity semantic contribution and the real-time temperature fluctuation quantity semantic contribution are normalized, and the real-time temperature absolute quantity time series association feature vector and the real-time temperature fluctuation quantity time series association feature vector are weighted modulated using the normalized real-time temperature absolute quantity semantic contribution and the normalized real-time temperature fluctuation quantity semantic contribution to obtain a modulated real-time temperature absolute quantity time series association feature vector and a modulated real-time temperature fluctuation quantity time series association feature vector, including: Calculating the sum of the real-time temperature absolute value semantic contribution and the real-time temperature fluctuation value semantic contribution to obtain a total sum of real-time temperature semantic contributions; Respectively dividing the real-time temperature absolute quantity semantic contribution and the real-time temperature fluctuation quantity semantic contribution by the total sum of the real-time temperature semantic contributions to obtain the normalized real-time temperature absolute quantity semantic contribution and the normalized real-time temperature fluctuation quantity semantic contribution; Multiplying the real-time temperature absolute quantity time series associated feature vector and the normalized real-time temperature absolute quantity semantic contribution degree by position points to obtain the modulated real-time temperature absolute quantity time series associated feature vector; The real-time temperature fluctuation amount time series associated feature vector and the normalized real-time temperature fluctuation amount semantic contribution are multiplied by position point to obtain the modulated real-time temperature fluctuation amount time series associated feature vector.

8. The megasonic cleaning control method according to claim 7, characterized in that: The post-modulation real-time temperature absolute quantity time series associated feature vector is used as a query vector, the post-modulation real-time temperature fluctuation quantity time series associated feature vector is used as a key vector, and the real-time temperature condition feature vector is used as a value vector. The post-modulation real-time temperature absolute quantity time series associated feature vector, the post-modulation real-time temperature fluctuation quantity time series associated feature vector, and the real-time temperature condition feature vector are input into a feature-to-feature significant guided interaction module based on a converter structure to obtain a temperature absolute quantity-fluctuation quantity time series significant interaction fusion representation vector, including: Performing vector multiplication of the post-modulation real-time temperature absolute quantity time series correlation feature vector and the transposed vector of the post-modulation real-time temperature fluctuation quantity time series correlation feature vector, and dividing the obtained post-modulation real-time temperature correlation matrix by the square root of the length of the post-modulation real-time temperature fluctuation quantity time series correlation feature vector by position to obtain a post-modulation real-time temperature correlation scaling matrix; The post-modulation real-time temperature association scaling matrix is ​​activated using a softmax function, and the obtained post-modulation real-time temperature association scaling activation matrix is ​​matrix-vector multiplied with the real-time temperature condition feature vector to obtain the temperature absolute value-fluctuation value time series significant interaction fusion representation vector.

9. The megasonic cleaning control method according to claim 8, characterized in that: Based on the temperature absolute quantity-fluctuation quantity time series significant interaction fusion representation, a planning instruction is obtained, including: inputting the temperature absolute quantity-fluctuation quantity time series significant interaction fusion representation vector into a decoder-based power dynamic planner to obtain a planning instruction, wherein the planning instruction includes a decoded value of the recommended input megasonic power at the next time point.

10. A megasonic cleaning control system, characterized in that: include: A cleaning fluid temperature data acquisition module is used to obtain a time queue of the real-time temperature of the cleaning fluid collected by the temperature sensor; A real-time temperature absolute quantity fluctuation quantity time series feature encoding module is used to calculate the temperature gradient value of the real-time temperature at each predetermined time point in the time queue of the real-time temperature to obtain the time queue of the temperature gradient value, and then perform sequence encoding on the time queue of the real-time temperature and the time queue of the temperature gradient value to obtain a real-time temperature absolute quantity time series associated feature vector and a real-time temperature fluctuation quantity time series associated feature vector; A significant interactive fusion module for real-time temperature absolute quantity fluctuation quantity time series features is used to perform significant interactive fusion between temperature features on the real-time temperature absolute quantity time series associated feature vector and the real-time temperature fluctuation quantity time series associated feature vector to obtain a significant interactive fusion representation of temperature absolute quantity-fluctuation quantity time series, wherein the real-time temperature absolute quantity fluctuation quantity time series features significant interactive fusion module includes: a real-time temperature condition feature generation unit, used to perform shared implicit feature extraction on the real-time temperature absolute quantity time series associated feature vector and the real-time temperature fluctuation quantity time series associated feature vector to obtain a real-time temperature condition feature; a temperature absolute quantity-fluctuation quantity time series significant interactive fusion representation generation unit, used to perform feature fusion on the real-time temperature absolute quantity time series associated feature vector and the real-time temperature fluctuation quantity time series associated feature vector using the real-time temperature condition feature as condition information to obtain the temperature absolute quantity-fluctuation quantity time series significant interactive fusion representation; The input megasonic power decoding module is used to obtain a planning instruction based on the significant interaction fusion representation of the temperature absolute value-fluctuation value time series, and the planning instruction includes a decoded value of the recommended input megasonic power at the next time point.

Images