Search

CN-121980178-A - Device abnormality early warning method for advanced packaging of semiconductor

CN121980178ACN 121980178 ACN121980178 ACN 121980178ACN-121980178-A

Abstract

The invention is applicable to the technical field of device abnormality early warning, and discloses a device abnormality early warning method for advanced packaging of semiconductors, which comprises the following steps: and synchronously acquiring temperature, pressure and time data of a bonding head in a hot-press bonding process, synchronously acquiring a current production situation label, binding the current production situation label with the acquired data in real time to generate a data packet, storing the data packet into a historical database, inputting the data packet into a pre-trained LSTM time sequence prediction model to obtain a future process parameter predicted value, and comparing the acquired process parameter actual value with the process parameter predicted value to obtain a comprehensive deviation value. According to the invention, the temperature, pressure and time parameters in the hot-press bonding process are synchronously acquired and analyzed in real time, the future process state is predicted by utilizing the pre-trained LSTM time sequence prediction model, the risk is identified at the initial stage of abnormality occurrence, and the batch defect is prevented from being detected after being formed, so that the waste of materials and working hours is reduced, and the production efficiency is further improved.

Inventors

  • ZHANG JIAN

Assignees

  • 无锡芯享信息科技有限公司

Dates

Publication Date
20260505
Application Date
20260119

Claims (8)

  1. 1. The device abnormality early warning method for advanced packaging of a semiconductor comprises the steps of synchronously acquiring temperature, pressure and time data of a bonding head in a hot-press bonding process, synchronously acquiring a current production situation label, binding the current production situation label with the acquired data in real time to generate a data packet, and storing the data packet into a historical database, and is characterized by further comprising the following steps: inputting the data packet into a pre-trained LSTM time sequence prediction model to obtain a future technological parameter predicted value, comparing the obtained technological parameter actual value with the technological parameter predicted value to obtain a comprehensive deviation value, and generating a first risk index based on the comprehensive deviation value; searching related abnormal events from a historical database according to the situation label of the current production product, and acquiring comprehensive historical influence coefficients of all the related abnormal events on the current production event; and dynamically adjusting the first risk index through the comprehensive history influence coefficient to obtain a second risk index, and carrying out grading early warning according to the second risk index.
  2. 2. The device anomaly pre-warning method for advanced semiconductor packaging according to claim 1, wherein the data packet is input into a pre-trained LSTM timing prediction model, and the specific steps of obtaining future process parameter predicted values are as follows: Reading a data packet, analyzing the content of the data packet, extracting acquisition data in the data packet, and extracting acquisition data in N continuous front ID data packets of the current ID data packet based on the ID of the current data packet, wherein N is an integer greater than 1; Carrying out data preprocessing on the extracted acquisition data, and converting the preprocessed data into a three-dimensional tensor for identification by an LSTM time sequence prediction model; And inputting the three-dimensional tensor into a pre-trained LSTM model, and outputting temperature predicted values and pressure predicted values of m time points in the future by the LSTM model, wherein m is an integer greater than 1.
  3. 3. The device anomaly pre-warning method for advanced semiconductor packaging according to claim 2, wherein the specific steps of obtaining the integrated deviation value are as follows: establishing a two-dimensional reference plane by taking temperature as a horizontal axis and pressure as a vertical axis; marking the temperature predicted values and the pressure predicted values of the future m time points output by the LSTM model in the reference plane respectively, and sequentially connecting the temperature predicted values and the pressure predicted values to form an expected path Marking the temperature value and the actual value of the pressure acquired in the same time period in the reference plane respectively, and sequentially connecting to form an actual path Respectively calculating the difference between two adjacent time points to obtain a predicted direction vector Actual direction vector By the formula Calculating the included angle between each predicted direction vector and the actual direction vector Obtaining an included angle set Setting corresponding weight for each parameter in the included angle set Weighting and summing each included angle and the weight corresponding to each included angle to obtain a comprehensive angle deviation value ; For each sampling point i, calculating the Euclidean distance between the actual point A (i) and the predicted point F (i) on the two-dimensional plane Finally, a distance set is obtained And corresponding weight is allocated to each parameter in the distance set Weighting and summing the distances and the weights corresponding to the distances to obtain a comprehensive absolute position deviation D; Setting up And D, the upper limit of the expected normal range of And By the formula Obtaining normalized data And Wherein X represents Or D; Respectively is And D1 sets the corresponding weight And And generating a comprehensive deviation value B through weighted fusion.
  4. 4. The device anomaly early warning method for advanced semiconductor packaging according to claim 3, wherein the specific obtaining step of the first risk index is: The mean u and standard deviation of the integrated deviation value B are counted from the historical normal data Obtaining B obtained by real-time calculation, subtracting the mean value u from the current comprehensive deviation value B, and dividing the current comprehensive deviation value B by the standard deviation The obtained result is a standard fraction Z value; By the formula Calculating to obtain a first risk index Wherein Z0 represents the offset and k represents the steepness of the curve.
  5. 5. The device anomaly early warning method for advanced packaging of semiconductors according to claim 4, wherein the acquiring step of the related anomaly event is as follows: inquiring historical events with abnormal production state identification in a historical database, and marking the inquired historical events as similar events to form a similar event set; Screening historical events with process flow identifiers completely consistent with current production events from the similar event sets, and marking the screened historical events as candidate events to form candidate event sets; calculating the same ratio R1 of product identifiers between each historical event and the current production event in the candidate event set and the same ratio R2 of equipment and tool identifiers respectively, and setting corresponding weights C1 and C2 for the R1 and the R2 respectively; The same ratio of the product identifiers and the same ratio of the tool identifiers are respectively weighted and fused with the corresponding weights to obtain the identifier coincidence degree E, and the identifier coincidence degree E is compared with a preset threshold E1 to screen out As related abnormal events.
  6. 6. The device anomaly early warning method for advanced semiconductor packaging according to claim 5, wherein the specific obtaining step of the comprehensive history influence coefficient is as follows: Obtaining severity of each associated anomaly Ratio of overlap ratio Time of occurrence Where i represents the index of the relevant anomaly event, which is the severity Ratio of overlap ratio Setting corresponding weights C3 and C4 respectively; By the formula A single event history impact value I1 of each related abnormal event on the current production event is calculated, wherein, Is a time decay coefficient; And summing and normalizing the single event history influence values I1 of all related abnormal events to obtain a comprehensive history influence coefficient H.
  7. 7. The device anomaly early warning method for advanced packaging of a semiconductor according to claim 6, wherein the specific obtaining step of the second risk index is: Acquiring a first risk index And a comprehensive history influence coefficient H, by the formula A second risk index S2 is calculated.
  8. 8. The device abnormality pre-warning method for advanced semiconductor packaging according to claim 7, wherein the specific steps of performing hierarchical pre-warning according to the second risk index are as follows: Presetting a normal threshold Z1 and an early warning threshold Z2, comparing the second risk index S2 with the threshold, and when At the same time, for low risk levels, when In order to treat stroke risk, when And is a high risk level.

Description

Device abnormality early warning method for advanced packaging of semiconductor Technical Field The invention relates to the technical field of device abnormality early warning, in particular to a device abnormality early warning method for advanced packaging of semiconductors. Background In the advanced packaging field of semiconductors, along with the development of technologies such as heterogeneous integration and high-density interconnection, the complexity and the precision of a packaging process are exponentially improved. In advanced packaging manufacture of semiconductors, core processes such as thermocompression bonding are key to achieving high-density, high-reliability interconnection. Such processes are highly dependent on precise coordinated and dynamic control of multiple parameters such as bond head temperature, applied pressure, and applied time on the microsecond to second time scale. Transient fluctuations of process parameters, progressive drift of equipment or batch differences of material characteristics may cause defects such as incomplete bonding, voids, chip warpage or stress damage at a microscopic interface, which cannot be recovered when a subsequent test link is detected. The existing method is mostly dependent on data query and defect analysis after production is completed, belongs to a post-remediation mode, cannot identify risk symptoms before abnormality occurs or at an initial stage due to lack of real-time monitoring and analysis of technological parameters, and when batch abnormality is finally found in a test link, defects are formed, so that serious waste of materials and working hours is caused. Disclosure of Invention The invention aims to provide a device abnormality early warning method for advanced packaging of semiconductors, which aims to solve the problem that real-time monitoring and analysis of technological parameters are lacking at present, and risk symptoms cannot be identified before abnormality occurs or at the beginning of the abnormality. The invention aims at realizing the technical scheme that the device abnormality early warning method for advanced packaging of the semiconductor comprises the following steps: synchronously acquiring temperature, pressure and time data of a bonding head in a hot-press bonding process, synchronously acquiring a current production situation label, binding the current production situation label with the acquired data in real time to generate a data packet, and storing the data packet into a historical database; Inputting the data packet into a pre-trained LSTM time sequence prediction model to obtain a future technological parameter prediction value, comparing the obtained technological parameter actual value with the technological parameter prediction value to obtain a comprehensive deviation value, and generating a first risk index based on the comprehensive deviation value; searching related abnormal events from a historical database according to the situation label of the current production product, and acquiring comprehensive historical influence coefficients of all the related abnormal events on the current production event; and dynamically adjusting the first risk index through the comprehensive history influence coefficient to obtain a second risk index, and carrying out grading early warning according to the second risk index. Preferably, the specific step of inputting the data packet into a pre-trained LSTM timing prediction model to obtain the predicted value of the future process parameter is: Reading a data packet, analyzing the content of the data packet, extracting acquisition data in the data packet, and extracting acquisition data in N continuous forward ID data packets of the current ID data packet based on the ID of the current data packet, wherein N is an integer greater than 1; Carrying out data preprocessing on the extracted acquisition data, and converting the preprocessed data into a three-dimensional tensor for identification by an LSTM time sequence prediction model; and inputting the three-dimensional tensor into a pre-trained LSTM model, and outputting temperature predicted values and pressure predicted values of m time points in the future by the model, wherein m is an integer greater than 1. Preferably, the specific steps of the comprehensive deviation value are as follows: Establishing a two-dimensional reference plane by taking temperature as a horizontal axis and pressure as a vertical axis; marking the temperature predicted values and the pressure predicted values of the future m time points output by the LSTM model in the reference plane respectively, and sequentially connecting the temperature predicted values and the pressure predicted values to form an expected path Marking the temperature value and the actual value of the pressure acquired in the same time period in the reference plane respectively, and sequentially connecting to form an actual pathRespectively calculating the diffe