Method and system for evaluating health condition of spring needle

Abstract

The invention relates to a method and a system for evaluating the health condition of a spring needle in real time by utilizing artificial intelligence (AI). The method comprises the following steps: extracting one or more features from a plurality of tested spring needles; using a feature selection module to construct a health index module according to the extracted features; and applying a pre-trained deep learning module to evaluate the health condition of the spring needle in real time. The system comprises: a receiving module for receiving one or more data inputs collected from a plurality of tested spring needles; the feature selection module is used for selecting the most important features from the input and constructing a health index module; and the pre-training deep learning module is used for evaluating the health condition of the spring needle in real time in the form of the health index percentage. According to the method and the system, the spring needle can be evaluated at the operation site during the operation period, the accuracy of an existing deep learning model in evaluating the spring needle on a production line is improved, and the health condition of the spring needle can be evaluated in real time.

Description

Technical Field

[0001] This invention relates to the field of semiconductor and electronic device manufacturing. More specifically, this invention relates to methods and systems for assessing the health status of spring pins. Background Technology

[0002] A contact probe is a probe that brings a target electronic component into contact with a test point. It is used for continuous electrical testing of microelectronic components in manufacturing. Detecting any defects using a contact probe ensures product reliability. Using a defective contact probe may result in error measurements that exceed normal values. If consecutive measurements deviate from the normal range, it indicates a potential defect in the contact probe. Spring pins, also known as spring contact probes, are a common type of contact probe widely used in semiconductors and electronic equipment. Defective spring pins can lead to unreliable test results and incorrect product rejection, resulting in wasted costs and production output.

[0003] In traditional methods, engineers select features sensitive to spring pin failure modes based on their experience. Then, through iterative testing, upper control limits (UCLs) and lower control limits (LCLs) are defined for each feature. Maintenance action is triggered once a measured feature exceeds the upper or lower control limits. However, the significant human factors involved in triggering maintenance actions present two serious problems: (1) UCLs and LCLs are not calculated based on rigorous mathematical or statistical foundations; and (2) significant feature fluctuations can easily lead to false alarms or missed alarms.

[0004] While some systems exist for assessing the health of spring pins, their accuracy remains far from satisfactory. Existing deep learning models suffer from shortcomings in feature selection when assessing spring pins on a production line, and progress in this area is limited. The range of features available for indicating spring pin health remains restricted. Examples of relevant references are provided below:

[0005] A patent document published by the World Intellectual Property Organization (WIPO) with publication number WO2022227018A1 discloses a power semiconductor device testing system, a cloud server, and a testing machine. The system includes a cloud server and at least one testing machine. The cloud server is used to pre-train an artificial intelligence model using historical data from multiple power semiconductor devices; the testing machine is used to acquire data from the power semiconductor device under test; and the cloud server or testing machine is used to input the data from the power semiconductor device under test into the artificial intelligence model, the output of which is a diagnostic result for the power semiconductor device under test. However, the aforementioned system and apparatus can only display the diagnostic results obtained by the testing machine and cannot display or visualize the health status of the device under test. Furthermore, the aforementioned prior art lacks a technical solution for providing a system suitable for real-time detection of the health status of spring pins, because the system is not suitable for detecting the health status of spring pins during operation, but only for when the spring pin is removed from its operating location and tested using the testing machine. The aforementioned invention may have the ability to extract physical correlation features related to the health status of a target component based on human experience and knowledge. However, identifying hidden features specifically related to the health status of the target component may pose a challenge to users of the aforementioned invention. Therefore, it is necessary to address the problem of selecting features that have a hidden correlation with health status, which may involve unfavorable certainty when monitoring health status in real time.

[0006] US Patent Publication No. US11493548 discloses a method for predicting failure parameters of semiconductor devices. This method includes receiving a set of data, including characteristics of a sample semiconductor device and parameters characterizing stress conditions. The method further includes extracting multiple feature values ​​from the dataset and inputting these feature values ​​into a trained model executed on one or more processors. The trained model is configured according to an artificial intelligence (AI) algorithm based on the previously obtained feature values, and outputs a failure prediction based on the multiple feature values. Furthermore, the method includes generating failure prediction parameters of the sample semiconductor device due to stress conditions using the trained model. While the above invention may employ statistical procedures such as principal component analysis (PCA), it may still lack the ability to accurately and selectively select important features from the data pool, particularly in determining the health status of spring pins in real time. This conventional method lacks any technical components, and the training model does not include steps for handling resistive and non-resistive related features, demonstrating its lack of technical capability.

[0007] US Patent Publication No. US10281519 discloses an anomaly measurement method and a device for measuring device anomalies. The anomaly measurement method includes the following steps: acquiring a feature sequence corresponding to the lifecycle based on formulation information and sensor information, wherein the feature sequence includes multiple feature subset sequences, and the lifecycle corresponds to multiple process operations; repeatedly performing a lifespan analysis process to obtain multiple lifespans of the lifecycle and each feature subset sequence in multiple feature subset sequences corresponding to one of the multiple lifespans; establishing a trend distribution corresponding to each lifespan in the multiple lifespans based on the feature subset sequences corresponding to the lifespans; and determining whether to send an alarm message based on the multiple trend distributions. However, the above invention only discloses how to identify whether a semiconductor device is normal or abnormal when the measurement parameters are out of range. Considering that degraded spring pins have a high probability of abnormal (high) measured resistance, but do not always show abnormal values, statistical results often fluctuate, thus failing to accurately indicate the health status. Therefore, there is an urgent need for a more definitive system and method to monitor the health status of spring pins using fluctuating data to improve the reliability of statistical conclusions.

[0008] Chinese patent document CN114913119A discloses an automatic probe status identification method during semiconductor wafer inspection. The system includes an RPA server, a KVM switch, an image recognition engine, and a manufacturing machine. An interface connecting the RPA server to the image recognition engine is added, allowing the RPA and image recognition engine to connect and exchange data. A wafer chip pad is placed below the test probe. The RPA obtains product images captured by the manufacturing machine via the KVM switch and then sends a recognition task to the image recognition engine. Based on the product images obtained by the RPA, the image recognition engine uses an automatic image recognition algorithm to perform target location detection and category identification of the pad position, crosshair position, and test marks left by the probe on the pad in the image, and then returns the recognition results to the RPA. After receiving the recognition results, the RPA decides whether to issue an alarm to the user based on pre-set alarm rules. This invention can effectively improve the automation level of the production process. However, the aforementioned prior art lacks the function of displaying the probe's health status.

[0009] Aye et al. disclosed a method for testing integrated circuit (IC) products in the semiconductor manufacturing industry, typically using socket spring pins that contact the IC product. These spring pins can degrade due to repeated insertion and removal from the socket. This degradation significantly impacts the accuracy of final testing in semiconductor manufacturing. To quickly and accurately detect this degradation, data mining techniques are needed to automatically identify it from test data. However, the disclosed method can only predict degraded spring pins and, like contemporary systems and methods, cannot determine the spring pin's health status in real time. Existing analyses only reflect a linear correlation between two parameters. The relationship between parameters and spring pin health status is not a simple linear correlation. This method lacks a system for selecting important features from a large number of parameters with complex correlations, and does not provide a suitable technical solution for managing the relationship between parameters and spring pin health status. Furthermore, this existing technology only uses time-domain features selected from signal energy as input, resulting in biases in signal energy characteristics and a technical limitation in comprehensively summarizing spring pin conditions.

[0010] The aforementioned references and other existing systems do not help to provide comprehensive feature selection to indicate health status, improve the accuracy of detecting spring pin abnormalities, and provide real-time health status assessments, while the objectives and features of the present invention attempt to address these problems.

[0011] It should also be noted that those skilled in the art would not be aware of the techniques and / or additional advantages of combining these features together (even if the prior art provides individual uses / advantages for these features), or that the prior art cannot overcome the technical obstacles in the technical problem.

[0012] Therefore, there is still a need in the field to provide a system and method to address the problems described herein. Summary of the Invention

[0013] This invention relates to a method for assessing the condition of a spring needle using health indices and deep learning. The method trains a deep learning model with a Long Short-Term Model (LSTM) architecture, with the model input including multiple features, including surface resistance-related and non-surface resistance-related features, suitable for detecting the current health status of the spring needle.

[0014] Therefore, the purpose of this invention is to provide a feature selection method that constructs a health index and model input from over a hundred measurement parameters through (1) expert experience; (2) failure physics analysis; and (3) unsupervised learning. This invention first selects features through expert experience and failure physics analysis to identify important features as input to a deep learning model. Then, an unsupervised learning method—Principal Component Analysis (PCA)—is applied to the selected features to determine the most important features for constructing the health index. This feature selection method helps reduce training parameters, thereby improving training and inference efficiency. Simultaneously, any irrelevant information is eliminated to improve model accuracy. Furthermore, feature selection through unsupervised learning can identify the most important features for monitoring the health status of spring needles.

[0015] One object of this invention is to construct a health index based on the statistical results of selected features. The health status (i.e., "healthy" or "unhealthy") is defined based on the statistical results of the most important feature within its minimum fluctuation range over the time domain. The health index is then defined as the percentage of abnormal spring needles within a specific time range. This method helps reduce the influence of human factors in defining the health status of spring needles, as the health status is defined scientifically with statistical support. Furthermore, this method allows for the labeling of spring needle health status, thereby constructing a health index for model labeling. Defining the health index as a percentage of abnormalities more accurately reflects the condition of the spring needles because the noise influence of the measured features is reduced, and the error / bias of the health index is also reduced. Moreover, in this method, the health index can also reflect the temporal performance (trend) of the spring needle's state.

[0016] Another objective of this invention is to construct a deep learning model with multiple feature inputs, including resistance-related features and non-resistance-related features. Resistance-related features can directly indicate the health status of the spring pin, while non-resistance-related features can indirectly reflect the health status of the spring pin. Monitoring the condition of the spring pin from different aspects is a more comprehensive approach.

[0017] Another objective of this invention is to provide a deep learning model employing a Long Short-Term Memory (LSTM) architecture for spring needle condition assessment. The LSTM architecture is specifically designed for time-series data, and the time-specific characteristics of spring needles (such as degradation trends) can be identified and learned by the model. This improves the model's accuracy in predicting and determining spring needle health in real time. Furthermore, the deep learning LSTM model can also be used to learn and extract intrinsic features of input parameters for better learning performance. This also contributes to improving the accuracy of spring needle health prediction. To prevent overfitting, an early stopping method is employed during training, and at least six months of long-term data is used in the model training. Long-term data helps learn information and features that vary considerably, thereby reducing model bias under specific time conditions.

[0018] Therefore, these objectives can be achieved through the teachings of this invention, which provides a more intelligent method for automatically assessing the condition of spring needles during production line operation. This invention relates to a method for assessing the health status of spring needles. The method includes extracting one or more features from a plurality of tested spring needles; constructing a health index module based on the extracted features using a feature selection module; and applying a pre-trained deep learning module to assess the health status of the spring needles in real time as a percentage of the health index.

[0019] The present invention further relates to a system for assessing the health status of spring needles. The system includes: a receiving module for receiving one or more data inputs collected from a plurality of tested spring needles; a feature selection module for selecting the most important features from the data inputs to construct a health index module; and a pre-trained deep learning module for assessing the health status of the spring needles in real time as a percentage of the health index.

[0020] The beneficial effects of this invention are as follows:

[0021] The method and system of the present invention help to evaluate spring needles at their operating location during operation, rather than measuring them at another location, and improve the accuracy of existing deep learning models when evaluating spring needles on a production line.

[0022] The method and system of the present invention provide a more comprehensive feature selection model and deep learning model for evaluating spring needles on a production line.

[0023] The method and system of the present invention can also assess the health status of the spring needle in real time.

[0024] A careful reading of the detailed description provided below, and with appropriate reference to the accompanying drawings, will enable a better understanding of the above and other objects, features, aspects, and advantages of the present invention. Attached Figure Description

[0025] The features of the invention will be more readily understood and recognized when reading the following detailed description in conjunction with the accompanying drawings of preferred embodiments of the invention, wherein:

[0026] Figure 1 A flowchart outlining the invention is shown;

[0027] Figure 2 The Long Short-Term Model (LSTM) cell structure is shown, where C t ,C' t, h t, x t ,f t ,,i t , and o t These represent the current unit stage, the update time unit stage, the hidden layer output, the current input, the forget gate, the input gate, and the output gate, respectively; and

[0028] Figure 3 An LSTM structure with encoder and decoder layers is shown. Detailed Implementation

[0029] Specific embodiments of the present invention are disclosed herein as requested. However, it should be understood that the disclosed embodiments are merely examples of the invention, which may be implemented in many different forms. Therefore, the specific structural and functional details disclosed herein should not be construed as limiting, but rather serve as the basis for the claims. It should be understood that the drawings and their detailed description are not intended to limit the invention to the specific forms disclosed herein; rather, the invention covers all modifications, equivalents, and alternatives falling within the scope defined by the claims. As used throughout this application, the word “may” indicates optional (i.e., possible) rather than mandatory (i.e., required). Similarly, the words “comprising” and “including” mean including but not limited to. Furthermore, unless otherwise mentioned, the word “a” means “at least one” and the word “a plurality” means one or more. When using abbreviations or technical terms, these refer to their generally accepted meanings known in the art.

[0030] Reference Figures 1 to 3 The invention will now be described in more detail, based on the accompanying drawings.

[0031] This invention discloses a method 100 for assessing the health status of spring needles. The method 100 includes: extracting one or more features from a plurality of tested spring needles; constructing a health index module 104 based on the extracted features using a feature selection module 102; and applying a pre-trained deep learning module 105 to assess the health status of the spring needles in real time in the form of a health index percentage. Figure 1 This has been further explained.

[0032] Figure 1 This is a flowchart of a method for assessing the health status of spring needles, which includes the interaction between a primary feature selection system, a secondary feature selection system, a health index construction system, a model training system, and a spring needle condition assessment system.

[0033] In one embodiment of the invention, the step of extracting features from a plurality of tested spring pins includes: testing the performance of each semi-finished semiconductor after the manufacturing process using a tester; and extracting at least one feature from each test. The spring pins are inserted into the semiconductor tester as a component of the tester.

[0034] After the manufacturing process is completed, the performance of each semi-finished product is tested. During the operation of the testing equipment and the spring pin, more than one hundred parameters on the spring pin are measured. All parameters measured on the spring pin are collected from the production line.

[0035] In one embodiment of the present invention, the step of using the feature selection module 102 to construct the health index module 104 based on the extracted features further includes: receiving one or more extracted features; selecting multiple first important features through a first-level selection unit; selecting the most important feature from the multiple first important features through a second-level selection unit; classifying the selected most important feature through statistical analysis; and constructing the health index module 104 in the form of a percentage based on the classified most important feature to indicate the health status of the spring needle within a specific time range.

[0036] In a preferred embodiment of the present invention, the step of selecting multiple first important features by a primary selection unit includes: selecting at least one feature related to health status based on the domain knowledge of an expert; and analyzing the root cause of the failure by performing a fault physics analysis on the selected feature.

[0037] In one embodiment of the present invention, the step of selecting the most important feature from a plurality of first important features by means of a secondary selection unit includes: applying unsupervised machine learning to the selected features obtained by means of a primary selection unit; and selecting at least one most important feature for constructing a health index module 104.

[0038] In another embodiment of the invention, the step of selecting an important feature from a plurality of first important features by means of a secondary selection unit further includes selecting an important feature that directly or indirectly indicates the health status of the spring needle.

[0039] Based on the above embodiments, the step of selecting important characteristics that directly or indirectly indicate the health status of the spring needle further includes selecting spring needle measurement data related to resistance or non-resistance within a predetermined time range of at least 6 months or longer.

[0040] In model training, the recorded data spans a long period, at least six months, allowing the model to learn from highly variable data and more diverse patterns. Furthermore, the key features selected for model input include both those directly related to resistance and those unrelated to resistance. The model can learn features from different perspectives. By utilizing information with significant variations, the model can achieve generalization.

[0041] In one embodiment of the invention, the step of selecting at least one feature related to health status based on an expert's domain knowledge unit further includes filtering out one or more incomplete features. A primary feature selection system is used to select important features sensitive to the health status of the spring needle from a large number of measurement features. Features closely related to spring needle failure prediction are then selected from the remaining features.

[0042] In one embodiment of the invention, the step of classifying the selected most important features through statistical analysis further includes: calculating the mean and standard deviation of the most important features selected by the secondary selection unit; and classifying the most important features selected by the secondary selection unit according to a health index threshold. In unsupervised learning analysis, among the measured surface resistance-related features, at least one feature has the highest weight, and the resistance fluctuates significantly when using a defective spring needle; these features are selected as the most important features.

[0043] In a preferred embodiment of the present invention, the step of classifying the selected most important feature according to a health index threshold includes: if the statistical value of the most important feature is equal to or lower than the health index threshold, then the most important feature is classified as "healthy".

[0044] In another preferred embodiment of the invention, the step of classifying the selected most important feature according to a health index threshold includes: if the statistical value of the most important feature exceeds the health index threshold, then the selected most important feature is classified as "unhealthy".

[0045] Figure 2 A Long Short-Term Memory (LSTM) cell structure is shown, where C t ,C' t, h t, x t ,f t ,,i t , and o t These represent the current unit stage, the update time unit stage, the hidden layer output, the current input, the forget gate, the input gate, and the output gate, respectively.

[0046] The model training module in the system takes the health index labeled in the health index module 104 and the important features selected in the main feature selection module as input to train a deep learning model to evaluate the condition of the spring needle. The selected important features are the input to the model, while the health index defined in each time step is the label to be learned by the model.

[0047] Then, the aforementioned data with input and labels will be used for training the LSTM architecture model. LSTM unit 200 includes three gates: a forget gate, an input gate, and an output gate. Applying gates with unit stage information in LSTM unit 200 is beneficial for the transmission of important information over a longer time range.

[0048] Within each time step, the hidden layer output (h) of the previous time step t-1 ) and current input data (x) t They are sent to the forgetting gate ft and processed according to the following known formula:

[0049] f t =σ(W f ·[h t-1 ,x t ]+b f ),

[0050]

[0051] Among them, W f and b f The weights and biases of the forget gate are represented, with values ​​in the range [0,1]. Therefore, the cell stage information (C) from the previous time step is first... t-1 Input the forget gate to determine the proportion of old information to retain.

[0052] Furthermore, the forget gate is immediately followed by the input gate. The input is [h] t-1 ,x t The hyperbolic tangent layer is used to update the state C′. t :

[0053] C′ t =tanh(W C ·[h t-1 ,x t ]+b C ),

[0054]

[0055] Among them, W C and b C These represent the weights and biases of the current unit stage, respectively.

[0056] C′ tThe value is used to update the current unit phase (C) t ), by input gate i t :control:

[0057] i t =(W i ·[h t-1 ,x t ]+b i ),

[0058] C t =f t ⊙C t-1 +i t ⊙C′ t ,

[0059] Among them W i and b i These represent the weights and biases of the input gate, respectively. The multiplication operation for each element is represented as ⊙.

[0060] Current Unit Phase (C) t The current hidden layer output (h) t ) by output gate (o t )control:

[0061] o t =(W o ·[h t-1 ,x t ]+b o )

[0062] h t =o t ⊙tanh(C t ),

[0063] Among them W o and b o These represent the weights and biases of the output gate, respectively.

[0064] The current cell stage and the current hidden layer output are the two outputs of each memory cell. The propagation of the cell stage along the memory cell can solve the long dependency problem in traditional neural network (RNN) models. RNN models face difficulties in learning and utilizing the received information from the input because the information may decrease or explode at each time step.

[0065] When time-series data is input into the LSTM model, data from each time step is fed into the corresponding LSTM unit 200. The hidden layer output from the first LSTM layer is passed as input to the second layer. This process is repeated up to the last LSTM layer, whose output is then fed into the dense layer to produce the model's predictions. The predictions are compared to the provided ground truth values ​​using loss functions such as mean squared error and mean absolute error. Then, the weights of the units in the model are updated using backpropagation based on the loss function. This process is repeated to minimize the loss function. Once the loss function converges to a stable value, the optimization algorithm terminates, and the model is ready for prediction.

[0066] In one embodiment of the invention, the step of applying a pre-trained deep learning module to evaluate the health status of the spring needle in real time further includes training the deep learning module 105. Figure 3 This has been further explained.

[0067] Figure 3 An LSTM structure with encoder and decoder layers is shown, applied to sequence-to-sequence prediction, with a desired output length of a health index. The encoder-decoder LSTM model allows users to choose a suitable combination between encoder 302 and LSTM decoder 308 to extract features with different properties. The encoder is an example of an encoder in conjunction with convolutional neural networks (CNNs) and LSTMs, while... Figure 3 The present invention will use LSTM as the decoder. CNN encoders are capable of extracting short-term dependency features and correlations between features; while LSTM encoders are suitable for capturing long-term temporal features. Therefore, this invention selects stacked LSTM as the decoder.

[0068] Furthermore, the steps for training the deep learning module 105 include: receiving inputs 105a and b from the initial selected feature units and the health index module; training sequence-to-sequence prediction using the long short-term memory (LSTM) encoder-decoder architecture of the deep learning module, and outputting a health index of a preset length for real-time prediction; stopping training once a preset duration is reached or no further improvement is achieved. More specifically, the early stopping function helps the model training stop at its optimal state without overfitting.

[0069] In addition, input 105a from the main feature selection unit is the important feature, while input 105b from the health index module 104 is the label of the most important feature that has been classified.

[0070] This invention further teaches a system for assessing the health status of spring needles. The system includes: a receiving module for receiving one or more data inputs collected from a plurality of tested spring needles; a feature selection module for selecting the most important features from the data inputs to construct a health index module; and a pre-trained deep learning module for real-time assessment of the health status of the spring needles. Figure 1 As shown.

[0071] Furthermore, the feature selection module in the system includes a primary selection unit and a secondary selection unit. The primary selection unit includes various data filtering methods such as expert experience and fault physics analysis, while the secondary selection unit includes unsupervised machine learning methods.

[0072] Based on expert experience, feature selection is performed on the collected data from a large number of parameters. In this step, some irrelevant parameters are filtered out. Then, feature selection is performed on this filtered data through fault physics analysis. Fault physics analysis analyzes the root cause of spring needle failure; therefore, after the above two feature selection methods, important features related to the health status of the spring needle are selected. These important features then proceed to the next feature selection step in the secondary feature selection system. Furthermore, the features selected in the primary feature selection module are used in the model training module.

[0073] Important features selected from the two feature selection steps of the first-level feature selection module are passed to the second-level feature selection module to determine the most important features related to the health status of the spring needle. Unsupervised machine learning methods in the second-level feature selection module include, but are not limited to, Principal Component Analysis (PCA) for dimensionality reduction and feature extraction via linear transformation. PCA is applied to identify the correlation between measurement data and the health status of the spring needle through feature clustering.

[0074] PCA results from time steps before and after effective maintenance actions were selected to analyze performance differences between healthy and unhealthy states. PCA identified the most sensitive parameters for detecting spring needle anomalies. Further analysis of the data clustering behavior at the selected time steps and the weight of each feature in the principal components was performed to derive the most important features—the most sensitive features related to the health status of the spring needle. These most important features were selected after the two feature selection modules described above.

[0075] Furthermore, the pre-trained deep learning module employs an LSTM architecture, comprising an encoder 302 and a decoder 308. The pre-trained deep learning model is further configured to evaluate the trend and temporal characteristics of the spring needle. The encoder portion 302 of the LSTM architecture is constructed from at least one layer of LSTM unit 200.

[0076] More specifically, the health index module takes the most important features selected by the secondary feature selection module as input to construct a health index for the model training module. The data input to the receiving module is measurement data, including but not limited to the performance of the magnetic writer, sensor, and power supply. First, the region with the least data fluctuation is determined; then, the statistical results of the most important features from the secondary feature selection module are calculated within this region. The calculated statistical results are used to set thresholds to identify the healthy and unhealthy states of the spring needle, thereby achieving the labeling purpose.

[0077] Data at each time step is labeled as healthy or unhealthy. Using these health and / or unhealthy status labels, a health index is defined as the percentage of unhealthy springpins within a specific time scale. Each time step's data is labeled with a corresponding health index.

[0078] Therefore, the method 100 and system of the present invention provide a system and method that facilitates the evaluation of spring needles at their operating location during operation, rather than measuring them at another location, and improves the accuracy of existing deep learning models in evaluating spring needles on a production line. Furthermore, the method 100 and system of the present invention provide a more comprehensive feature selection model and deep learning model when evaluating spring needles on a production line. In addition, the method 100 and system of the present invention can also assess the health status of spring needles in real time.

[0079] The above explanation of the present invention is not limited to the foregoing embodiments and drawings, and it will be apparent to those skilled in the art that various substitutions, modifications and alterations can be made without departing from the scope of the present invention.

Claims

1. A method for assessing the health status of a spring needle, characterized in that, Includes the following steps: Extract one or more features from multiple tested spring pins; The health index module is constructed based on the extracted features using the feature selection module. as well as A pre-trained deep learning module is used to assess the health status of the spring needle in real time as a percentage of the health index.

2. As described in claim 1 A method for assessing the health status of a spring needle, characterized in that... Extracting one or more features from a plurality of tested spring pins includes the following steps: The performance of each semi-finished semiconductor was tested using a testing instrument after the manufacturing process; and Extract at least one feature from each test.

3. The method for assessing the health status of a spring needle according to claim 1, characterized in that, The step of using the feature selection module to construct the health index module based on the extracted features further includes the following steps: Receive one or more extracted features; Multiple primary features are selected through a first-level selection unit; The most important feature is selected from multiple primary features using a two-level selection unit; The most important selected features are classified through statistical analysis; and A health index module is constructed as a percentage based on the most important characteristics after classification, to indicate the health status of the spring needle within a specific time range.

4. The method for assessing the health status of a spring needle according to claim 3, characterized in that, The step of selecting multiple primary important features through a primary selection unit includes the following steps: Select at least one feature related to health status based on the expert's domain knowledge; and By performing physical analysis on the selected features, the root cause of the failure can be analyzed.

5. The method for assessing the health status of a spring needle according to claim 3, characterized in that, The step of selecting the most important feature from multiple primary important features through a secondary selection unit includes the following steps: Unsupervised machine learning is applied to the selected features obtained through the first-level selection unit; and Select at least one of the most important features to construct the health index module.

6. The method for assessing the health status of a spring needle according to claim 3 or 5, characterized in that, The step of selecting the most important feature from multiple first-important features through a secondary selection unit further includes the following steps: Choose to directly or indirectly display important characteristics of the spring needle's health status.

7. The method for assessing the health status of a spring needle according to claim 6, characterized in that, The selection of key features that directly or indirectly display the health status of the spring needle includes the following steps: Select spring needle measurement data related to resistance or non-resistance within a predetermined time range of at least 6 months or longer.

8. The method for assessing the health status of a spring needle according to claim 4, characterized in that, The step of selecting at least one feature related to health status based on the expert's domain knowledge further includes the following steps: Filter out one or more incomplete features.

9. The method for assessing the health status of a spring needle according to claim 3, characterized in that, The classification of the most important selected features through statistical analysis further includes the following steps: Calculate the mean and standard deviation of the most important features selected by the secondary selection units; and The most important features selected by the secondary selection unit are classified according to the health index threshold.

10. The method for assessing the health status of a spring needle according to claim 9, characterized in that, The process of classifying the most important features selected by the secondary selection unit based on the health index threshold includes the following steps: If the statistical value of the most important characteristic is equal to or lower than the health index threshold, then the most important characteristic is classified as healthy.

11. The method for assessing the health status of a spring needle according to claim 9, characterized in that, The process of classifying the most important features selected by the secondary selection unit based on the health index threshold includes the following steps: If the statistical value of the most important feature exceeds the health index threshold, the selected most important feature will be classified as unhealthy.

12. The method for assessing the health status of a spring needle according to claim 3, characterized in that, The application of a pre-trained deep learning module to assess the health status of the spring pin in real time as a health index percentage further includes the following steps: Train the deep learning module.

13. The method for assessing the health status of a spring needle according to claim 12, characterized in that, The training of the deep learning module includes the following steps: Receives input from the primary selection unit and the health index module; The deep learning module uses a long short-term memory encoder-decoder architecture to train sequence-to-sequence prediction, outputting a health index of a preset length for real-time prediction; and Training stops once the preset duration is reached or no further improvement is achieved.

14. The method for assessing the health status of a spring needle according to claim 13, characterized in that, The input from the first-level selection unit is the important feature, while the input from the health index module is the label of the most important feature that has been classified.

15. A system for assessing the health status of a spring needle, characterized in that, include: A receiving module for receiving one or more data inputs collected from multiple tested spring pins; A feature selection module is used to select the most important features from the data input to construct a health index module; as well as A pre-trained deep learning module is used to assess the health status of the spring needle in real time.

16. The system for assessing the health status of a spring needle according to claim 15, characterized in that, The feature selection module includes a primary selection unit and a secondary selection unit.

17. The system for assessing the health status of a spring needle according to claim 16, characterized in that, The first-level selection unit includes expert experience method and / or fault physics analysis method, and the second-level selection unit includes unsupervised machine learning method.

18. The system for assessing the health status of a spring needle according to claim 16, characterized in that, The unsupervised machine learning method includes principal component analysis for dimensionality reduction and feature extraction through linear transformation.

19. The system for assessing the health status of a spring needle according to claim 18, characterized in that, The principal component analysis was applied to identify the correlation between measurement data and the health status of spring needles through feature clustering.

20. The system for assessing the health status of a spring needle according to claim 15, characterized in that, The pre-trained deep learning module adopts an LSTM architecture and includes an encoder and a decoder.

21. The system for assessing the health status of a spring needle according to claim 15, characterized in that, The pre-trained deep learning module is further configured to evaluate the trend and temporal characteristics of the spring needle.

22. The system for assessing the health status of a spring needle according to claim 20, characterized in that, The encoder portion 302 of the LSTM architecture is constructed from at least one layer of LSTM cells.

23. The system for assessing the health status of a spring needle according to claim 16, characterized in that, The data input to the receiving module is measurement data, including the performance of the magnetic writer, sensor, and power supply.

Images